Star916
1834
代码
代码
Issues220
Pull Requests321
Actions
项目讨论
Wiki
分析
Star916
1834
已合并
tilingdata重构 #5151
zhangxiyan7创建于 5月22日
tilingdata重构 #5151
已合并
zhangxiyan7创建于 5月22日
从已删除 :master合入到cann/ops-nnmaster
共 65 个文件变更+3083-2872
Mconv/common/op_kernel/arch35/conv_common.h+54-54
@@ -174,13 +174,13 @@ template <class CONV, class CONV_TILING>
174__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::174__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
175 CalcStartAddrCommon(const uint32_t din, const uint32_t dout)175 CalcStartAddrCommon(const uint32_t din, const uint32_t dout)
176{176{
177- hwIn = convTilingData->convRunInfo.hin * convTilingData->convRunInfo.win;177+ hwIn = convTilingData->hin * convTilingData->win;
178- hwOut = convTilingData->convRunInfo.hout * convTilingData->convRunInfo.wout;178+ hwOut = convTilingData->hout * convTilingData->wout;
179 179 
180- convOps->fmapOneBatchSize = convTilingData->convRunInfo.cin * din * hwIn;180+ convOps->fmapOneBatchSize = convTilingData->cin * din * hwIn;
181- convOps->outputOneBatchSize = convTilingData->convRunInfo.cout * dout * hwOut;181+ convOps->outputOneBatchSize = convTilingData->cout * dout * hwOut;
182 182 
183- if (convTilingData->convRunInfo.hasBias) {183+ if (convTilingData->hasBias) {
184 biasStartAddr = convOps->nIdxStart;184 biasStartAddr = convOps->nIdxStart;
185 }185 }
186 186 
@@ -202,32 +202,32 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
202 convOps->mIdxStart;202 convOps->mIdxStart;
203 } else if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC || CONV::C_FORMAT == ConvFormat::NHWC) {203 } else if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC || CONV::C_FORMAT == ConvFormat::NHWC) {
204 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +204 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
205- convOps->mIdxStart * convTilingData->convRunInfo.cout +205+ convOps->mIdxStart * convTilingData->cout +
206 convOps->nIdxStart;206 convOps->nIdxStart;
207 }207 }
208 if constexpr (CONV::C_FORMAT == ConvFormat::NCDHW) {208 if constexpr (CONV::C_FORMAT == ConvFormat::NCDHW) {
209 outputStartAddr += doIdxStart * hwOut;209 outputStartAddr += doIdxStart * hwOut;
210 } else if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC) {210 } else if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC) {
211- outputStartAddr += doIdxStart * hwOut * convTilingData->convRunInfo.cout;211+ outputStartAddr += doIdxStart * hwOut * convTilingData->cout;
212 }212 }
213 } else {213 } else {
214 if constexpr (CONV::C_FORMAT == ConvFormat::NCDHW || CONV::C_FORMAT == ConvFormat::NCHW) {214 if constexpr (CONV::C_FORMAT == ConvFormat::NCDHW || CONV::C_FORMAT == ConvFormat::NCHW) {
215 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +215 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
216 convOps->nIdxStart * dout * hwOut +216 convOps->nIdxStart * dout * hwOut +
217- convOps->hoIdxStart * convTilingData->convRunInfo.wout;217+ convOps->hoIdxStart * convTilingData->wout;
218 } else if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC || CONV::C_FORMAT == ConvFormat::NHWC) {218 } else if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC || CONV::C_FORMAT == ConvFormat::NHWC) {
219 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize + convOps->hoIdxStart *219 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize + convOps->hoIdxStart *
220- convTilingData->convRunInfo.wout * convTilingData->convRunInfo.cout + convOps->nIdxStart;220+ convTilingData->wout * convTilingData->cout + convOps->nIdxStart;
221 }221 }
222 if constexpr (CONV::C_FORMAT == ConvFormat::NCHW) {222 if constexpr (CONV::C_FORMAT == ConvFormat::NCHW) {
223 outputStartAddr += convOps->woIdxStart;223 outputStartAddr += convOps->woIdxStart;
224 } else if constexpr (CONV::C_FORMAT == ConvFormat::NHWC) {224 } else if constexpr (CONV::C_FORMAT == ConvFormat::NHWC) {
225- outputStartAddr += convOps->woIdxStart * convTilingData->convRunInfo.cout;225+ outputStartAddr += convOps->woIdxStart * convTilingData->cout;
226 }226 }
227 if constexpr (CONV::A_FORMAT == ConvFormat::NCDHW) {227 if constexpr (CONV::A_FORMAT == ConvFormat::NCDHW) {
228 outputStartAddr += doIdxStart * hwOut;228 outputStartAddr += doIdxStart * hwOut;
229 } else if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {229 } else if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
230- outputStartAddr += doIdxStart * hwOut * convTilingData->convRunInfo.cout;230+ outputStartAddr += doIdxStart * hwOut * convTilingData->cout;
231 }231 }
232 }232 }
233}233}
@@ -240,7 +240,7 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
240 CalcStartAddrCommon(din, dout);240 CalcStartAddrCommon(din, dout);
241 241 
242 if constexpr (CONV::DIS_CONTINUOUS) {242 if constexpr (CONV::DIS_CONTINUOUS) {
243- fmStartAddr = convOps->batchIdxStart * convTilingData->convRunInfo.cin;243+ fmStartAddr = convOps->batchIdxStart * convTilingData->cin;
244 } else {244 } else {
245 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize;245 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize;
246 }246 }
@@ -250,17 +250,17 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
250 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {250 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {
251 weightStartAddr = convOps->nIdxStart * convOps->k0;251 weightStartAddr = convOps->nIdxStart * convOps->k0;
252 } else {252 } else {
253- weightStartAddr = convOps->nIdxStart * convTilingData->convRunInfo.cin * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;253+ weightStartAddr = convOps->nIdxStart * convTilingData->cin * kd * convTilingData->kh * convTilingData->kw;
254 }254 }
255 CalcOutputStartAddr(dout, doIdxStart);255 CalcOutputStartAddr(dout, doIdxStart);
256 256 
257 if constexpr (CONV::IS_EXTEND_CONV2D) {257 if constexpr (CONV::IS_EXTEND_CONV2D) {
258- if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||258+ if (convOps->convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
259- convOps->convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {259+ convOps->convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
260 scaleStartAddr = convOps->nIdxStart;260 scaleStartAddr = convOps->nIdxStart;
261 }261 }
262- if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||262+ if (convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
263- convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {263+ convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
264 reluWeightStartAddr = convOps->nIdxStart;264 reluWeightStartAddr = convOps->nIdxStart;
265 }265 }
266 }266 }
@@ -273,24 +273,24 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
273{273{
274 CalcStartAddrCommon(din, dout);274 CalcStartAddrCommon(din, dout);
275 275 
276- int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->convRunInfo.strideH) -276+ int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->strideH) -
277- static_cast<int64_t>(convTilingData->convRunInfo.padTop);277+ static_cast<int64_t>(convTilingData->padTop);
278 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;278 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;
279 if constexpr (CONV::DIS_CONTINUOUS) {279 if constexpr (CONV::DIS_CONTINUOUS) {
280- fmStartAddr = convOps->singleCoreHiStartPos * convTilingData->convRunInfo.win *280+ fmStartAddr = convOps->singleCoreHiStartPos * convTilingData->win *
281- convTilingData->convRunInfo.batch * convTilingData->convRunInfo.cin +281+ convTilingData->batch * convTilingData->cin +
282- convOps->batchIdxStart * convTilingData->convRunInfo.cin;282+ convOps->batchIdxStart * convTilingData->cin;
283 } else {283 } else {
284 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *284 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *
285- convTilingData->convRunInfo.win;285+ convTilingData->win;
286 }286 }
287 convOps->singleCoreHiStartPos = hiStartPosTmp;287 convOps->singleCoreHiStartPos = hiStartPosTmp;
288 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {288 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {
289- int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->convRunInfo.strideW) -289+ int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->strideW) -
290- static_cast<int64_t>(convTilingData->convRunInfo.padLeft);290+ static_cast<int64_t>(convTilingData->padLeft);
291 convOps->singleCoreWiStartPos = wiStartPosTmp < 0 ? 0 : wiStartPosTmp;291 convOps->singleCoreWiStartPos = wiStartPosTmp < 0 ? 0 : wiStartPosTmp;
292 if constexpr (CONV::DIS_CONTINUOUS) {292 if constexpr (CONV::DIS_CONTINUOUS) {
293- fmStartAddr += convOps->singleCoreWiStartPos * convTilingData->convRunInfo.batch * convTilingData->convRunInfo.cin;293+ fmStartAddr += convOps->singleCoreWiStartPos * convTilingData->batch * convTilingData->cin;
294 } else {294 } else {
295 fmStartAddr += convOps->singleCoreWiStartPos;295 fmStartAddr += convOps->singleCoreWiStartPos;
296 }296 }
@@ -300,7 +300,7 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
300 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {300 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {
301 weightStartAddr = convOps->nIdxStart * convOps->k0;301 weightStartAddr = convOps->nIdxStart * convOps->k0;
302 } else {302 } else {
303- weightStartAddr = convOps->nIdxStart * convTilingData->convRunInfo.cin * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;303+ weightStartAddr = convOps->nIdxStart * convTilingData->cin * kd * convTilingData->kh * convTilingData->kw;
304 }304 }
305 CalcOutputStartAddr(dout, doIdxStart);305 CalcOutputStartAddr(dout, doIdxStart);
306 306 
@@ -317,7 +317,7 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
317 317 
318 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize;318 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize;
319 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {319 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
320- fmStartAddr += diIdxStart * hwIn * convTilingData->convRunInfo.cin;320+ fmStartAddr += diIdxStart * hwIn * convTilingData->cin;
321 }321 }
322 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {322 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {
323 weightStartAddr = convOps->nIdxStart * convOps->k0;323 weightStartAddr = convOps->nIdxStart * convOps->k0;
@@ -333,17 +333,17 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
333{333{
334 CalcStartAddrCommon(din, dout);334 CalcStartAddrCommon(din, dout);
335 335 
336- int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->convRunInfo.strideH) -336+ int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->strideH) -
337- static_cast<int64_t>(convTilingData->convRunInfo.padTop);337+ static_cast<int64_t>(convTilingData->padTop);
338 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;338 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;
339 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *339 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *
340- convTilingData->convRunInfo.win * convTilingData->convRunInfo.cin;340+ convTilingData->win * convTilingData->cin;
341 convOps->singleCoreHiStartPos = hiStartPosTmp;341 convOps->singleCoreHiStartPos = hiStartPosTmp;
342 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {342 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {
343- int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->convRunInfo.strideW) -343+ int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->strideW) -
344- static_cast<int64_t>(convTilingData->convRunInfo.padLeft);344+ static_cast<int64_t>(convTilingData->padLeft);
345 convOps->singleCoreWiStartPos = wiStartPosTmp;345 convOps->singleCoreWiStartPos = wiStartPosTmp;
346- fmStartAddr += Max(convOps->singleCoreWiStartPos, 0) * convTilingData->convRunInfo.cin;346+ fmStartAddr += Max(convOps->singleCoreWiStartPos, 0) * convTilingData->cin;
347 }347 }
348 348 
349 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {349 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {
@@ -354,7 +354,7 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
354 CalcOutputStartAddr(dout, doIdxStart);354 CalcOutputStartAddr(dout, doIdxStart);
355 355 
356 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {356 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
357- fmStartAddr += diIdxStart * hwIn * convTilingData->convRunInfo.cin;357+ fmStartAddr += diIdxStart * hwIn * convTilingData->cin;
358 }358 }
359}359}
360 360 
@@ -369,21 +369,21 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
369 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + diIdxStart * hwIn;369 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + diIdxStart * hwIn;
370 370 
371 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +371 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
372- convOps->mIdxStart * convTilingData->convRunInfo.cout + convOps->nIdxStart +372+ convOps->mIdxStart * convTilingData->cout + convOps->nIdxStart +
373- doIdxStart * hwOut * convTilingData->convRunInfo.cout;373+ doIdxStart * hwOut * convTilingData->cout;
374 } else {374 } else {
375- int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->convRunInfo.strideH) -375+ int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->strideH) -
376- static_cast<int64_t>(convTilingData->convRunInfo.padTop);376+ static_cast<int64_t>(convTilingData->padTop);
377 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;377 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;
378 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *378 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *
379- convTilingData->convRunInfo.win + diIdxStart * hwIn;379+ convTilingData->win + diIdxStart * hwIn;
380 convOps->singleCoreHiStartPos = hiStartPosTmp;380 convOps->singleCoreHiStartPos = hiStartPosTmp;
381 381 
382 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize + convOps->hoIdxStart *382 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize + convOps->hoIdxStart *
383- convTilingData->convRunInfo.wout * convTilingData->convRunInfo.cout + convOps->nIdxStart +383+ convTilingData->wout * convTilingData->cout + convOps->nIdxStart +
384- doIdxStart * hwOut * convTilingData->convRunInfo.cout;384+ doIdxStart * hwOut * convTilingData->cout;
385 }385 }
386- weightStartAddr = convOps->nIdxStart * convTilingData->convRunInfo.cin * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;386+ weightStartAddr = convOps->nIdxStart * convTilingData->cin * kd * convTilingData->kh * convTilingData->kw;
387}387}
388 388 
389template <class CONV, class CONV_TILING>389template <class CONV, class CONV_TILING>
@@ -393,37 +393,37 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
393 if constexpr (CONV::isQuant || CONV::IS_EXTEND_CONV2D) {393 if constexpr (CONV::isQuant || CONV::IS_EXTEND_CONV2D) {
394 if (hasScale) {394 if (hasScale) {
395 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {395 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {
396- if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {396+ if (convOps->convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
397 convOps->fixpipeParams.scale0.SetGlobalBuffer(397 convOps->fixpipeParams.scale0.SetGlobalBuffer(
398 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale0 + scaleStartAddr *398 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale0 + scaleStartAddr *
399 sizeof(typename CONV::SCALE_T)));399 sizeof(typename CONV::SCALE_T)));
400- } else if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {400+ } else if (convOps->convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {
401 convOps->fixpipeParams.scale0.SetGlobalBuffer(401 convOps->fixpipeParams.scale0.SetGlobalBuffer(
402 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale0));402 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale0));
403 }403 }
404 if constexpr (CONV::IS_EXTEND_CONV2D) {404 if constexpr (CONV::IS_EXTEND_CONV2D) {
405- if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {405+ if (convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
406 convOps->fixpipeParams.reluWeight0.SetGlobalBuffer(406 convOps->fixpipeParams.reluWeight0.SetGlobalBuffer(
407 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight0));407 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight0));
408- } else if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {408+ } else if (convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
409 // copy relu_weight from om to gm409 // copy relu_weight from om to gm
410 convOps->fixpipeParams.reluWeight0.SetGlobalBuffer(410 convOps->fixpipeParams.reluWeight0.SetGlobalBuffer(
411 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight0 +411 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight0 +
412 reluWeightStartAddr * sizeof(typename CONV::RELU_WEIGHT_T)));412 reluWeightStartAddr * sizeof(typename CONV::RELU_WEIGHT_T)));
413 }413 }
414- if (convTilingData->convApiTiling.dualOutput) {414+ if (convTilingData->dualOutput) {
415- if (convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {415+ if (convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
416 convOps->fixpipeParams.scale1.SetGlobalBuffer(416 convOps->fixpipeParams.scale1.SetGlobalBuffer(
417 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale1 + scaleStartAddr *417 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale1 + scaleStartAddr *
418 sizeof(typename CONV::SCALE_T)));418 sizeof(typename CONV::SCALE_T)));
419- } else if (convOps->convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {419+ } else if (convOps->convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {
420 convOps->fixpipeParams.scale1.SetGlobalBuffer(420 convOps->fixpipeParams.scale1.SetGlobalBuffer(
421 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale1));421 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale1));
422 }422 }
423- if (convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {423+ if (convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
424 convOps->fixpipeParams.reluWeight1.SetGlobalBuffer(424 convOps->fixpipeParams.reluWeight1.SetGlobalBuffer(
425 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight1));425 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight1));
426- } else if (convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {426+ } else if (convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
427 convOps->fixpipeParams.reluWeight1.SetGlobalBuffer(427 convOps->fixpipeParams.reluWeight1.SetGlobalBuffer(
428 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight1 +428 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight1 +
429 reluWeightStartAddr * sizeof(typename CONV::RELU_WEIGHT_T)));429 reluWeightStartAddr * sizeof(typename CONV::RELU_WEIGHT_T)));
@@ -450,13 +450,13 @@ __aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
450 convOps->outputGm.SetGlobalBuffer(450 convOps->outputGm.SetGlobalBuffer(
451 reinterpret_cast<__gm__ typename CONV::OUTPUT_T*>(y + outputStartAddr * sizeof(typename CONV::OUTPUT_T)));451 reinterpret_cast<__gm__ typename CONV::OUTPUT_T*>(y + outputStartAddr * sizeof(typename CONV::OUTPUT_T)));
452 if constexpr (CONV::IS_EXTEND_CONV2D) {452 if constexpr (CONV::IS_EXTEND_CONV2D) {
453- if (convTilingData->convApiTiling.dualOutput) {453+ if (convTilingData->dualOutput) {
454 convOps->output1Gm.SetGlobalBuffer(454 convOps->output1Gm.SetGlobalBuffer(
455 reinterpret_cast<__gm__ typename CONV::OUTPUT1_T*>(455 reinterpret_cast<__gm__ typename CONV::OUTPUT1_T*>(
456 extendParams->y1 + outputStartAddr * sizeof(typename CONV::OUTPUT1_T)));456 extendParams->y1 + outputStartAddr * sizeof(typename CONV::OUTPUT1_T)));
457 }457 }
458 }458 }
459- if (convTilingData->convRunInfo.hasBias) {459+ if (convTilingData->hasBias) {
460 convOps->biasGm.SetGlobalBuffer(460 convOps->biasGm.SetGlobalBuffer(
461 reinterpret_cast<__gm__ typename CONV::BIAS_T*>(bias + biasStartAddr * sizeof(typename CONV::BIAS_T)));461 reinterpret_cast<__gm__ typename CONV::BIAS_T*>(bias + biasStartAddr * sizeof(typename CONV::BIAS_T)));
462 }462 }
Mconv/common/op_kernel/arch35/conv_common_func.h+69-69
@@ -104,52 +104,52 @@ struct SetFixpipeParams {
104 static __aicore__ inline void call(Intf *self, const Extendconv2dFixpipeParams<typename Intf::ScaleT, typename Intf::ReluWeightT, typename Intf::ClipValue0T, typename Intf::ClipValue1T>& fixpipeParams)104 static __aicore__ inline void call(Intf *self, const Extendconv2dFixpipeParams<typename Intf::ScaleT, typename Intf::ReluWeightT, typename Intf::ClipValue0T, typename Intf::ClipValue1T>& fixpipeParams)
105 {105 {
106 if ASCEND_IS_AIC_CONV {106 if ASCEND_IS_AIC_CONV {
107- if (self->ctx.convTilingData->convApiTiling.quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {107+ if (self->ctx.convTilingData->quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {
108 self->ctx.scalegm.SetGlobalBuffer(fixpipeParams.scale0.GetPhyAddr(0), fixpipeParams.scale0.GetSize());108 self->ctx.scalegm.SetGlobalBuffer(fixpipeParams.scale0.GetPhyAddr(0), fixpipeParams.scale0.GetSize());
109- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {109+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {
110 self->ctx.deqScalar0 = fixpipeParams.scale0.GetValue(0);110 self->ctx.deqScalar0 = fixpipeParams.scale0.GetValue(0);
111 }111 }
112 }112 }
113- if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {113+ if (self->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
114 float m2 = fixpipeParams.reluWeight0.GetValue(0);114 float m2 = fixpipeParams.reluWeight0.GetValue(0);
115 self->ctx.preReluScalar0 = reinterpret_cast<uint64_t&>(m2);115 self->ctx.preReluScalar0 = reinterpret_cast<uint64_t&>(m2);
116- } else if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {116+ } else if (self->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
117 // copy from fixpipe paramIn to global tensor117 // copy from fixpipe paramIn to global tensor
118 self->ctx.reluWeightGM.SetGlobalBuffer(fixpipeParams.reluWeight0.GetPhyAddr(0),118 self->ctx.reluWeightGM.SetGlobalBuffer(fixpipeParams.reluWeight0.GetPhyAddr(0),
119 fixpipeParams.reluWeight0.GetSize());119 fixpipeParams.reluWeight0.GetSize());
120 }120 }
121 if constexpr (Intf::isExtendConv2d) {121 if constexpr (Intf::isExtendConv2d) {
122- if (self->ctx.convTilingData->convApiTiling.dualOutput &&122+ if (self->ctx.convTilingData->dualOutput &&
123- self->ctx.convTilingData->convApiTiling.quantMode1 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {123+ self->ctx.convTilingData->quantMode1 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {
124 self->ctx.scale1gm.SetGlobalBuffer(fixpipeParams.scale1.GetPhyAddr(0), fixpipeParams.scale1.GetSize());124 self->ctx.scale1gm.SetGlobalBuffer(fixpipeParams.scale1.GetPhyAddr(0), fixpipeParams.scale1.GetSize());
125- if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {125+ if (self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {
126 self->ctx.deqScalar1 = fixpipeParams.scale1.GetValue(0);126 self->ctx.deqScalar1 = fixpipeParams.scale1.GetValue(0);
127 }127 }
128- if (self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {128+ if (self->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
129 float m2 = fixpipeParams.reluWeight1.GetValue(0);129 float m2 = fixpipeParams.reluWeight1.GetValue(0);
130 self->ctx.preReluScalar1 = reinterpret_cast<uint64_t&>(m2);130 self->ctx.preReluScalar1 = reinterpret_cast<uint64_t&>(m2);
131- } else if (self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {131+ } else if (self->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
132 self->ctx.reluWeight1GM.SetGlobalBuffer(fixpipeParams.reluWeight1.GetPhyAddr(0),132 self->ctx.reluWeight1GM.SetGlobalBuffer(fixpipeParams.reluWeight1.GetPhyAddr(0),
133 fixpipeParams.reluWeight1.GetSize());133 fixpipeParams.reluWeight1.GetSize());
134 }134 }
135 }135 }
136- if (self->ctx.convTilingData->convApiTiling.quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT) ||136+ if (self->ctx.convTilingData->quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT) ||
137- self->ctx.convTilingData->convApiTiling.quantMode1 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {137+ self->ctx.convTilingData->quantMode1 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {
138 self->ctx.isFirstIterate = true;138 self->ctx.isFirstIterate = true;
139 }139 }
140- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||140+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
141- self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {141+ self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
142 self->ctx.enableVectorQuant = true;142 self->ctx.enableVectorQuant = true;
143 }143 }
144- if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||144+ if (self->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
145- self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {145+ self->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
146 self->ctx.enableVectorRelu = true;146 self->ctx.enableVectorRelu = true;
147 }147 }
148 } else {148 } else {
149- if (self->ctx.convTilingData->convApiTiling.quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {149+ if (self->ctx.convTilingData->quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {
150 self->ctx.isFirstIterate = true;150 self->ctx.isFirstIterate = true;
151 }151 }
152- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {152+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
153 self->ctx.enableVectorQuant = true;153 self->ctx.enableVectorQuant = true;
154 }154 }
155 }155 }
@@ -175,7 +175,7 @@ struct GetTensorC {
175 if ASCEND_IS_AIC_CONV {175 if ASCEND_IS_AIC_CONV {
176 self->ctx.copyOutIns.template CopyOut<TensorTypeT, config>(output0, ubInfo);176 self->ctx.copyOutIns.template CopyOut<TensorTypeT, config>(output0, ubInfo);
177 if constexpr (Intf::isExtendConv2d) {177 if constexpr (Intf::isExtendConv2d) {
178- if (self->ctx.convTilingData->convApiTiling.dualOutput) {178+ if (self->ctx.convTilingData->dualOutput) {
179 self->ctx.copyOutIns1.template CopyOut<GlobalTensor, config>(output1, ubInfo);179 self->ctx.copyOutIns1.template CopyOut<GlobalTensor, config>(output1, ubInfo);
180 }180 }
181 }181 }
@@ -184,11 +184,11 @@ struct GetTensorC {
184 self->ctx.queueCL0.FreeTensor(self->ctx.cl0);184 self->ctx.queueCL0.FreeTensor(self->ctx.cl0);
185 }185 }
186 186 
187- if (self->ctx.enableVectorQuant && !self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {187+ if (self->ctx.enableVectorQuant && !self->ctx.convTilingData->fixpParamsFullLoadFlag) {
188 self->ctx.queueScaleL1.FreeTensor(self->ctx.scaleL1);188 self->ctx.queueScaleL1.FreeTensor(self->ctx.scaleL1);
189 }189 }
190 190 
191- if (self->ctx.enableVectorRelu && !self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {191+ if (self->ctx.enableVectorRelu && !self->ctx.convTilingData->fixpParamsFullLoadFlag) {
192 self->ctx.queueReluWeightL1.FreeTensor(self->ctx.reluWeightL1);192 self->ctx.queueReluWeightL1.FreeTensor(self->ctx.reluWeightL1);
193 }193 }
194 }194 }
@@ -201,7 +201,7 @@ struct ConvPreProcess {
201 {201 {
202 if ASCEND_IS_AIC_CONV {202 if ASCEND_IS_AIC_CONV {
203 if constexpr (!Intf::isDeQuantFlag) {203 if constexpr (!Intf::isDeQuantFlag) {
204- if (self->ctx.convTilingData->convApiTiling.biasFullLoadFlag && self->ctx.enableBias) {204+ if (self->ctx.convTilingData->biasFullLoadFlag && self->ctx.enableBias) {
205 self->ctx.biasL1 = self->ctx.queueBiasL1.template AllocTensor<typename Intf::BiasT>();205 self->ctx.biasL1 = self->ctx.queueBiasL1.template AllocTensor<typename Intf::BiasT>();
206 uint64_t biasLoadNum = self->ctx.singleCoreCo;206 uint64_t biasLoadNum = self->ctx.singleCoreCo;
207 if constexpr (Intf::groupOptPreloadFlag) {207 if constexpr (Intf::groupOptPreloadFlag) {
@@ -212,7 +212,7 @@ struct ConvPreProcess {
212 self->ctx.queueBiasL1.EnQue(self->ctx.biasL1);212 self->ctx.queueBiasL1.EnQue(self->ctx.biasL1);
213 self->ctx.biasL1 = self->ctx.queueBiasL1.template DeQue<typename Intf::BiasT>();213 self->ctx.biasL1 = self->ctx.queueBiasL1.template DeQue<typename Intf::BiasT>();
214 }214 }
215- if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {215+ if (self->ctx.convTilingData->fixpParamsFullLoadFlag) {
216 if (self->ctx.enableVectorQuant || self->ctx.enableVectorRelu) {216 if (self->ctx.enableVectorQuant || self->ctx.enableVectorRelu) {
217 event_t eventId = static_cast<event_t>(self->ctx.pipe.FetchEventID(HardEvent::FIX_MTE2));217 event_t eventId = static_cast<event_t>(self->ctx.pipe.FetchEventID(HardEvent::FIX_MTE2));
218 SetFlag<HardEvent::FIX_MTE2>(eventId);218 SetFlag<HardEvent::FIX_MTE2>(eventId);
@@ -225,11 +225,11 @@ struct ConvPreProcess {
225 if (self->ctx.enableVectorQuant) {225 if (self->ctx.enableVectorQuant) {
226 self->ctx.scaleL1 = self->ctx.queueScaleL1.template AllocTensor<typename Intf::ScaleT>();226 self->ctx.scaleL1 = self->ctx.queueScaleL1.template AllocTensor<typename Intf::ScaleT>();
227 if constexpr (Intf::isExtendConv2d) {227 if constexpr (Intf::isExtendConv2d) {
228- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {228+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
229 self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1, self->ctx.scalegm,229 self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1, self->ctx.scalegm,
230 scaleLoadNum, 0);230 scaleLoadNum, 0);
231 }231 }
232- if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {232+ if (self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
233 self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1[self->ctx.scale1L1offset],233 self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1[self->ctx.scale1L1offset],
234 self->ctx.scale1gm, scaleLoadNum, 0);234 self->ctx.scale1gm, scaleLoadNum, 0);
235 }235 }
@@ -244,11 +244,11 @@ struct ConvPreProcess {
244 if (self->ctx.enableVectorRelu) {244 if (self->ctx.enableVectorRelu) {
245 self->ctx.reluWeightL1 =245 self->ctx.reluWeightL1 =
246 self->ctx.queueReluWeightL1.template AllocTensor<typename Intf::ReluWeightT>();246 self->ctx.queueReluWeightL1.template AllocTensor<typename Intf::ReluWeightT>();
247- if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {247+ if (self->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
248 self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1FullLoad(self->ctx.reluWeightL1,248 self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1FullLoad(self->ctx.reluWeightL1,
249 self->ctx.reluWeightGM, scaleLoadNum, 0);249 self->ctx.reluWeightGM, scaleLoadNum, 0);
250 }250 }
251- if (self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {251+ if (self->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
252 self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1FullLoad(252 self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1FullLoad(
253 self->ctx.reluWeightL1[self->ctx.reluWeight1L1offset],253 self->ctx.reluWeightL1[self->ctx.reluWeight1L1offset],
254 self->ctx.reluWeight1GM, scaleLoadNum, 0);254 self->ctx.reluWeight1GM, scaleLoadNum, 0);
@@ -274,14 +274,14 @@ struct ConvPostProcess {
274 {274 {
275 if ASCEND_IS_AIC_CONV {275 if ASCEND_IS_AIC_CONV {
276 if constexpr (!Intf::isDeQuantFlag) {276 if constexpr (!Intf::isDeQuantFlag) {
277- if (self->ctx.convTilingData->convApiTiling.biasFullLoadFlag && self->ctx.enableBias) {277+ if (self->ctx.convTilingData->biasFullLoadFlag && self->ctx.enableBias) {
278 self->ctx.queueBiasL1.FreeTensor(self->ctx.biasL1);278 self->ctx.queueBiasL1.FreeTensor(self->ctx.biasL1);
279 }279 }
280- if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag && self->ctx.enableVectorQuant) {280+ if (self->ctx.convTilingData->fixpParamsFullLoadFlag && self->ctx.enableVectorQuant) {
281 self->ctx.queueScaleL1.FreeTensor(self->ctx.scaleL1);281 self->ctx.queueScaleL1.FreeTensor(self->ctx.scaleL1);
282 }282 }
283 }283 }
284- if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag && self->ctx.enableVectorRelu) {284+ if (self->ctx.convTilingData->fixpParamsFullLoadFlag && self->ctx.enableVectorRelu) {
285 self->ctx.queueReluWeightL1.FreeTensor(self->ctx.reluWeightL1);285 self->ctx.queueReluWeightL1.FreeTensor(self->ctx.reluWeightL1);
286 }286 }
287 287 
@@ -354,7 +354,7 @@ struct IterateAll {
354 const GlobalTensor<typename Intf::Output1T> &output1) {354 const GlobalTensor<typename Intf::Output1T> &output1) {
355 if constexpr (Intf::formatOutput == ConvFormat::NHWC) {355 if constexpr (Intf::formatOutput == ConvFormat::NHWC) {
356 if constexpr (Intf::isFixedPoint) {356 if constexpr (Intf::isFixedPoint) {
357- if (self->ctx.convTilingData->convApiTiling.dualOutput) {357+ if (self->ctx.convTilingData->dualOutput) {
358 GetTensorC<Intf, ImplType>::template358 GetTensorC<Intf, ImplType>::template
359 call<GlobalTensor, CFG_ROW_MAJOR_FIXED_POINT, sync>(self, output0, output1);359 call<GlobalTensor, CFG_ROW_MAJOR_FIXED_POINT, sync>(self, output0, output1);
360 } else {360 } else {
@@ -362,7 +362,7 @@ struct IterateAll {
362 call<GlobalTensor, CFG_ROW_MAJOR_FIXED_POINT, sync>(self, output0);362 call<GlobalTensor, CFG_ROW_MAJOR_FIXED_POINT, sync>(self, output0);
363 }363 }
364 } else {364 } else {
365- if (self->ctx.convTilingData->convApiTiling.dualOutput) {365+ if (self->ctx.convTilingData->dualOutput) {
366 GetTensorC<Intf, ImplType>::template366 GetTensorC<Intf, ImplType>::template
367 call<GlobalTensor, CFG_ROW_MAJOR, sync>(self, output0, output1);367 call<GlobalTensor, CFG_ROW_MAJOR, sync>(self, output0, output1);
368 } else {368 } else {
@@ -372,7 +372,7 @@ struct IterateAll {
372 }372 }
373 } else {373 } else {
374 if constexpr (Intf::isFixedPoint) {374 if constexpr (Intf::isFixedPoint) {
375- if (self->ctx.convTilingData->convApiTiling.dualOutput) {375+ if (self->ctx.convTilingData->dualOutput) {
376 GetTensorC<Intf, ImplType>::template376 GetTensorC<Intf, ImplType>::template
377 call<GlobalTensor, CFG_COLUMN_MAJOR_FIXED_POINT, sync>(self, output0, output1);377 call<GlobalTensor, CFG_COLUMN_MAJOR_FIXED_POINT, sync>(self, output0, output1);
378 } else {378 } else {
@@ -380,7 +380,7 @@ struct IterateAll {
380 call<GlobalTensor, CFG_COLUMN_MAJOR_FIXED_POINT, sync>(self, output0);380 call<GlobalTensor, CFG_COLUMN_MAJOR_FIXED_POINT, sync>(self, output0);
381 }381 }
382 } else {382 } else {
383- if (self->ctx.convTilingData->convApiTiling.dualOutput) {383+ if (self->ctx.convTilingData->dualOutput) {
384 GetTensorC<Intf, ImplType>::template384 GetTensorC<Intf, ImplType>::template
385 call<GlobalTensor, CFG_COLUMN_MAJOR, sync>(self, output0, output1);385 call<GlobalTensor, CFG_COLUMN_MAJOR, sync>(self, output0, output1);
386 } else {386 } else {
@@ -430,11 +430,11 @@ __aicore__ inline void InitBufferWithDoubleBuf(Intf *self)
430 self->ctx.wholeAl0Tensor = self->ctx.al0Buf.template Get<typename Intf::FmapT>();430 self->ctx.wholeAl0Tensor = self->ctx.al0Buf.template Get<typename Intf::FmapT>();
431 self->ctx.wholeBl0Tensor = self->ctx.bl0Buf.template Get<typename Intf::WeightT>();431 self->ctx.wholeBl0Tensor = self->ctx.bl0Buf.template Get<typename Intf::WeightT>();
432 432 
433- int8_t cl0db = (self->ctx.convTilingData->convApiTiling.pBufferFlag & 0x04) >> 2;433+ int8_t cl0db = (self->ctx.convTilingData->pBufferFlag & 0x04) >> 2;
434 if constexpr (Intf::isInnerBatchFlag || Intf::isDeQuantFlag) {434 if constexpr (Intf::isInnerBatchFlag || Intf::isDeQuantFlag) {
435- uint64_t cl0Spacesize = self->ctx.convTilingData->convApiTiling.mStep * self->ctx.convTilingData->convApiTiling.nL0;435+ uint64_t cl0Spacesize = self->ctx.convTilingData->mStep * self->ctx.convTilingData->nL0;
436 if constexpr (Intf::isInnerBatchFlag) {436 if constexpr (Intf::isInnerBatchFlag) {
437- cl0Spacesize *= self->ctx.convTilingData->convApiTiling.innerBatch;437+ cl0Spacesize *= self->ctx.convTilingData->innerBatch;
438 }438 }
439 if (!cl0db) {439 if (!cl0db) {
440 self->ctx.pipe.InitBuffer(self->ctx.queueCL0, 1, cl0Spacesize * Intf::sizeOfL0c);440 self->ctx.pipe.InitBuffer(self->ctx.queueCL0, 1, cl0Spacesize * Intf::sizeOfL0c);
@@ -446,10 +446,10 @@ __aicore__ inline void InitBufferWithDoubleBuf(Intf *self)
446 self->ctx.wholeCl0Tensor = self->ctx.l0cBuf.template Get<typename Intf::L0cT>();446 self->ctx.wholeCl0Tensor = self->ctx.l0cBuf.template Get<typename Intf::L0cT>();
447 }447 }
448 448 
449- if (!((self->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET)) {449+ if (!((self->ctx.convTilingData->pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET)) {
450- self->ctx.pipe.InitBuffer(self->ctx.queueAL1, 1, self->ctx.convTilingData->convApiTiling.aL1SpaceSize);450+ self->ctx.pipe.InitBuffer(self->ctx.queueAL1, 1, self->ctx.convTilingData->aL1SpaceSize);
451 } else {451 } else {
452- self->ctx.pipe.InitBuffer(self->ctx.queueAL1, DOUBLE_BUF, self->ctx.convTilingData->convApiTiling.aL1SpaceSize);452+ self->ctx.pipe.InitBuffer(self->ctx.queueAL1, DOUBLE_BUF, self->ctx.convTilingData->aL1SpaceSize);
453 }453 }
454 454 
455 if constexpr (!Intf::bL1DBFlag) {455 if constexpr (!Intf::bL1DBFlag) {
@@ -466,7 +466,7 @@ __aicore__ inline void InitBufferWithDoubleBuf(Intf *self)
466 466 
467 if constexpr (Intf::isDeQuantFlag) {467 if constexpr (Intf::isDeQuantFlag) {
468 self->ctx.pipe.InitBuffer(self->ctx.mmadResUbBuf,468 self->ctx.pipe.InitBuffer(self->ctx.mmadResUbBuf,
469- self->ctx.convTilingData->convApiTiling.mUB * self->ctx.convTilingData->convApiTiling.nUB * Intf::sizeOfL0c);469+ self->ctx.convTilingData->mUB * self->ctx.convTilingData->nUB * Intf::sizeOfL0c);
470 self->ctx.mmadResUbTensor = self->ctx.mmadResUbBuf.template Get<typename Intf::L0cT>();470 self->ctx.mmadResUbTensor = self->ctx.mmadResUbBuf.template Get<typename Intf::L0cT>();
471 }471 }
472}472}
@@ -474,21 +474,21 @@ __aicore__ inline void InitBufferWithDoubleBuf(Intf *self)
474template <class Intf>474template <class Intf>
475__aicore__ inline void InitBuffer(Intf *self)475__aicore__ inline void InitBuffer(Intf *self)
476{476{
477- self->ctx.bL1SpaceSize = self->ctx.convTilingData->convApiTiling.nBL1 * self->ctx.convTilingData->convApiTiling.kBL1;477+ self->ctx.bL1SpaceSize = self->ctx.convTilingData->nBL1 * self->ctx.convTilingData->kBL1;
478 478 
479 InitBufferWithDoubleBuf<Intf>(self);479 InitBufferWithDoubleBuf<Intf>(self);
480 480
481 if constexpr (!Intf::isDeQuantFlag) {481 if constexpr (!Intf::isDeQuantFlag) {
482- if (self->ctx.convTilingData->convApiTiling.hasBias) {482+ if (self->ctx.convTilingData->hasBias) {
483- uint64_t biasl1Spacesize = self->ctx.convTilingData->convApiTiling.biasFullLoadFlag ? AlignB(483+ uint64_t biasl1Spacesize = self->ctx.convTilingData->biasFullLoadFlag ? AlignB(
484 self->ctx.singleCoreCo * Intf::sizeOfBias, BLOCK_L0_N * Intf::sizeOfBias) :484 self->ctx.singleCoreCo * Intf::sizeOfBias, BLOCK_L0_N * Intf::sizeOfBias) :
485- self->ctx.convTilingData->convApiTiling.nL0 * Intf::sizeOfBias;485+ self->ctx.convTilingData->nL0 * Intf::sizeOfBias;
486 if constexpr (Intf::groupOptPreloadFlag) {486 if constexpr (Intf::groupOptPreloadFlag) {
487- if (self->ctx.convTilingData->convApiTiling.biasFullLoadFlag) {487+ if (self->ctx.convTilingData->biasFullLoadFlag) {
488- biasl1Spacesize = AlignB(self->ctx.convTilingData->convApiTiling.orgCo * Intf::sizeOfBias, BLOCK_L0_N * Intf::sizeOfBias);488+ biasl1Spacesize = AlignB(self->ctx.convTilingData->orgCo * Intf::sizeOfBias, BLOCK_L0_N * Intf::sizeOfBias);
489 }489 }
490 }490 }
491- uint64_t biasBTSpacesize = self->ctx.convTilingData->convApiTiling.nL0;491+ uint64_t biasBTSpacesize = self->ctx.convTilingData->nL0;
492 self->ctx.pipe.InitBuffer(self->ctx.queueBiasL1, 1, AlignB(biasl1Spacesize, C0_SIZE));492 self->ctx.pipe.InitBuffer(self->ctx.queueBiasL1, 1, AlignB(biasl1Spacesize, C0_SIZE));
493 self->ctx.pipe.InitBuffer(493 self->ctx.pipe.InitBuffer(
494 self->ctx.queueBiasBT, 1, AlignB(biasBTSpacesize * Intf::sizeOfL0c, BT_SIZE));494 self->ctx.queueBiasBT, 1, AlignB(biasBTSpacesize * Intf::sizeOfL0c, BT_SIZE));
@@ -496,50 +496,50 @@ __aicore__ inline void InitBuffer(Intf *self)
496 }496 }
497 497 
498 if constexpr (Intf::isExtendConv2d) {498 if constexpr (Intf::isExtendConv2d) {
499- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||499+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
500- self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {500+ self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
501 uint64_t scaleL1SpaceSize = 0;501 uint64_t scaleL1SpaceSize = 0;
502- uint64_t scale0L1Size = self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag ? AlignB(502+ uint64_t scale0L1Size = self->ctx.convTilingData->fixpParamsFullLoadFlag ? AlignB(
503 self->ctx.singleCoreCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale) :503 self->ctx.singleCoreCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale) :
504- self->ctx.convTilingData->convApiTiling.nL0 * Intf::sizeOfScale;504+ self->ctx.convTilingData->nL0 * Intf::sizeOfScale;
505 if constexpr (Intf::groupOptPreloadFlag) {505 if constexpr (Intf::groupOptPreloadFlag) {
506- if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {506+ if (self->ctx.convTilingData->fixpParamsFullLoadFlag) {
507- scale0L1Size = AlignB(self->ctx.convTilingData->convApiTiling.orgCo * Intf::sizeOfScale,507+ scale0L1Size = AlignB(self->ctx.convTilingData->orgCo * Intf::sizeOfScale,
508 BLOCK_L0_N * Intf::sizeOfScale);508 BLOCK_L0_N * Intf::sizeOfScale);
509 }509 }
510 }510 }
511- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {511+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
512 scaleL1SpaceSize += scale0L1Size;512 scaleL1SpaceSize += scale0L1Size;
513 }513 }
514- if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {514+ if (self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
515 self->ctx.scale1L1offset = scaleL1SpaceSize / Intf::sizeOfScale;515 self->ctx.scale1L1offset = scaleL1SpaceSize / Intf::sizeOfScale;
516 scaleL1SpaceSize += scale0L1Size;516 scaleL1SpaceSize += scale0L1Size;
517 }517 }
518 self->ctx.pipe.InitBuffer(self->ctx.queueScaleL1, 1, AlignB(scaleL1SpaceSize, C0_SIZE));518 self->ctx.pipe.InitBuffer(self->ctx.queueScaleL1, 1, AlignB(scaleL1SpaceSize, C0_SIZE));
519 }519 }
520- if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||520+ if (self->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
521- self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {521+ self->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
522 uint32_t reluWeightL1SpaceSize = 0;522 uint32_t reluWeightL1SpaceSize = 0;
523- uint32_t reluWeight0L1Size = self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag ? AlignB(523+ uint32_t reluWeight0L1Size = self->ctx.convTilingData->fixpParamsFullLoadFlag ? AlignB(
524 self->ctx.singleCoreCo * Intf::sizeOfReluWeight, BLOCK_L0_N * Intf::sizeOfReluWeight) :524 self->ctx.singleCoreCo * Intf::sizeOfReluWeight, BLOCK_L0_N * Intf::sizeOfReluWeight) :
525- self->ctx.convTilingData->convApiTiling.nL0 * Intf::sizeOfReluWeight;525+ self->ctx.convTilingData->nL0 * Intf::sizeOfReluWeight;
526- if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {526+ if (self->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
527 reluWeightL1SpaceSize += reluWeight0L1Size;527 reluWeightL1SpaceSize += reluWeight0L1Size;
528 }528 }
529- if (self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {529+ if (self->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
530 self->ctx.reluWeight1L1offset = reluWeightL1SpaceSize / Intf::sizeOfReluWeight;530 self->ctx.reluWeight1L1offset = reluWeightL1SpaceSize / Intf::sizeOfReluWeight;
531 reluWeightL1SpaceSize += reluWeight0L1Size;531 reluWeightL1SpaceSize += reluWeight0L1Size;
532 }532 }
533 self->ctx.pipe.InitBuffer(self->ctx.queueReluWeightL1, 1, AlignB(reluWeightL1SpaceSize, C0_SIZE));533 self->ctx.pipe.InitBuffer(self->ctx.queueReluWeightL1, 1, AlignB(reluWeightL1SpaceSize, C0_SIZE));
534 }534 }
535 } else if constexpr (Intf::isQuantScene) {535 } else if constexpr (Intf::isQuantScene) {
536- if (self->ctx.convTilingData->convApiTiling.hasScale != 0) {536+ if (self->ctx.convTilingData->hasScale != 0) {
537- uint64_t scaleL1SpaceSize = self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag ? AlignB(537+ uint64_t scaleL1SpaceSize = self->ctx.convTilingData->fixpParamsFullLoadFlag ? AlignB(
538 self->ctx.singleCoreCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale) :538 self->ctx.singleCoreCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale) :
539- self->ctx.convTilingData->convApiTiling.nL0 * Intf::sizeOfScale;539+ self->ctx.convTilingData->nL0 * Intf::sizeOfScale;
540 if constexpr (Intf::groupOptPreloadFlag) {540 if constexpr (Intf::groupOptPreloadFlag) {
541- if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {541+ if (self->ctx.convTilingData->fixpParamsFullLoadFlag) {
542- scaleL1SpaceSize = AlignB(self->ctx.convTilingData->convApiTiling.orgCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale);542+ scaleL1SpaceSize = AlignB(self->ctx.convTilingData->orgCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale);
543 }543 }
544 }544 }
545 self->ctx.pipe.InitBuffer(self->ctx.queueScaleL1, 1, AlignB(scaleL1SpaceSize, C0_SIZE));545 self->ctx.pipe.InitBuffer(self->ctx.queueScaleL1, 1, AlignB(scaleL1SpaceSize, C0_SIZE));
@@ -550,9 +550,9 @@ __aicore__ inline void InitBuffer(Intf *self)
550template <class Intf>550template <class Intf>
551__aicore__ inline void InitHf32Mode(Intf *self)551__aicore__ inline void InitHf32Mode(Intf *self)
552{552{
553- if (self->ctx.convTilingData->convApiTiling.hf32Enable) {553+ if (self->ctx.convTilingData->hf32Enable) {
554- SetHF32Mode(self->ctx.convTilingData->convApiTiling.hf32Enable);554+ SetHF32Mode(self->ctx.convTilingData->hf32Enable);
555- SetHF32TransMode(self->ctx.convTilingData->convApiTiling.hf32TransMode);555+ SetHF32TransMode(self->ctx.convTilingData->hf32TransMode);
556 } else {556 } else {
557 SetHF32Mode(0);557 SetHF32Mode(0);
558 SetHF32TransMode(0);558 SetHF32TransMode(0);
Mconv/common/op_kernel/arch35/conv_group_common.h+72-72
@@ -70,23 +70,23 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
70 CalcStartAddrFmapHW()70 CalcStartAddrFmapHW()
71{71{
72 if constexpr (!CONV::isMMode) {72 if constexpr (!CONV::isMMode) {
73- int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->convRunInfo.strideH) -73+ int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->strideH) -
74- static_cast<int64_t>(convTilingData->convRunInfo.padTop);74+ static_cast<int64_t>(convTilingData->padTop);
75 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;75 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;
76 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {76 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {
77- fmStartAddr += convOps->singleCoreHiStartPos * convTilingData->convRunInfo.win * convTilingData->convRunInfo.cin;77+ fmStartAddr += convOps->singleCoreHiStartPos * convTilingData->win * convTilingData->cin;
78 } else {78 } else {
79- fmStartAddr += convOps->singleCoreHiStartPos * convTilingData->convRunInfo.win;79+ fmStartAddr += convOps->singleCoreHiStartPos * convTilingData->win;
80 }80 }
81 81
82 // Conv2D Split W82 // Conv2D Split W
83 convOps->singleCoreHiStartPos = hiStartPosTmp;83 convOps->singleCoreHiStartPos = hiStartPosTmp;
84 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {84 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {
85- int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->convRunInfo.strideW) -85+ int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->strideW) -
86- static_cast<int64_t>(convTilingData->convRunInfo.padLeft);86+ static_cast<int64_t>(convTilingData->padLeft);
87 convOps->singleCoreWiStartPos = wiStartPosTmp < 0 ? 0 : wiStartPosTmp;87 convOps->singleCoreWiStartPos = wiStartPosTmp < 0 ? 0 : wiStartPosTmp;
88 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {88 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {
89- fmStartAddr += convOps->singleCoreWiStartPos * convTilingData->convRunInfo.cin;89+ fmStartAddr += convOps->singleCoreWiStartPos * convTilingData->cin;
90 } else {90 } else {
91 fmStartAddr += convOps->singleCoreWiStartPos;91 fmStartAddr += convOps->singleCoreWiStartPos;
92 }92 }
@@ -110,16 +110,16 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
110 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +110 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
111 convOps->groupIdxStart * convOps->coPerGroup + convOps->nIdxStart;111 convOps->groupIdxStart * convOps->coPerGroup + convOps->nIdxStart;
112 if constexpr (CONV::isMMode) {112 if constexpr (CONV::isMMode) {
113- outputStartAddr += convOps->mIdxStart * convTilingData->convRunInfo.cout;113+ outputStartAddr += convOps->mIdxStart * convTilingData->cout;
114 } else {114 } else {
115- outputStartAddr += convOps->hoIdxStart * convTilingData->convRunInfo.wout * convTilingData->convRunInfo.cout;115+ outputStartAddr += convOps->hoIdxStart * convTilingData->wout * convTilingData->cout;
116 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {116 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {
117- outputStartAddr += convOps->woIdxStart * convTilingData->convRunInfo.cout;117+ outputStartAddr += convOps->woIdxStart * convTilingData->cout;
118 }118 }
119 }119 }
120 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {120 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
121- fmStartAddr += diIdxStart * hwIn * convTilingData->convRunInfo.cin;121+ fmStartAddr += diIdxStart * hwIn * convTilingData->cin;
122- outputStartAddr += doIdxStart * hwOut * convTilingData->convRunInfo.cout;122+ outputStartAddr += doIdxStart * hwOut * convTilingData->cout;
123 }123 }
124}124}
125 125 
@@ -141,7 +141,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
141 if constexpr (CONV::isMMode) {141 if constexpr (CONV::isMMode) {
142 outputStartAddr += convOps->mIdxStart;142 outputStartAddr += convOps->mIdxStart;
143 } else {143 } else {
144- outputStartAddr += convOps->hoIdxStart * convTilingData->convRunInfo.wout;144+ outputStartAddr += convOps->hoIdxStart * convTilingData->wout;
145 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {145 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {
146 outputStartAddr += convOps->woIdxStart;146 outputStartAddr += convOps->woIdxStart;
147 }147 }
@@ -158,12 +158,12 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
158 CalcStartAddrOriGroup(const uint64_t din, const uint64_t dout, const uint32_t kd, const uint64_t doIdxStart,158 CalcStartAddrOriGroup(const uint64_t din, const uint64_t dout, const uint32_t kd, const uint64_t doIdxStart,
159 const int64_t diIdxStart)159 const int64_t diIdxStart)
160{160{
161- hwIn = convTilingData->convRunInfo.hin * convTilingData->convRunInfo.win;161+ hwIn = convTilingData->hin * convTilingData->win;
162- hwOut = convTilingData->convRunInfo.hout * convTilingData->convRunInfo.wout;162+ hwOut = convTilingData->hout * convTilingData->wout;
163 dhwOut = dout * hwOut;163 dhwOut = dout * hwOut;
164 dhwIn = din * hwIn;164 dhwIn = din * hwIn;
165- convOps->fmapOneBatchSize = dhwIn * convTilingData->convRunInfo.cin;165+ convOps->fmapOneBatchSize = dhwIn * convTilingData->cin;
166- convOps->outputOneBatchSize = dhwOut * convTilingData->convRunInfo.cout;166+ convOps->outputOneBatchSize = dhwOut * convTilingData->cout;
167 167 
168 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {168 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {
169 fmapOneGroupSize = convOps->ciPerGroup;169 fmapOneGroupSize = convOps->ciPerGroup;
@@ -172,13 +172,13 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
172 CalcStartAddrOriGroupHWC(doIdxStart, diIdxStart);172 CalcStartAddrOriGroupHWC(doIdxStart, diIdxStart);
173 } else {173 } else {
174 fmapOneGroupSize = convOps->ciPerGroup * dhwIn;174 fmapOneGroupSize = convOps->ciPerGroup * dhwIn;
175- weightOneCoSize = convOps->ciPerGroup * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;175+ weightOneCoSize = convOps->ciPerGroup * kd * convTilingData->kh * convTilingData->kw;
176 weightOneGroupSize = convOps->coPerGroup * weightOneCoSize;176 weightOneGroupSize = convOps->coPerGroup * weightOneCoSize;
177 outputOneGroupSize = convOps->coPerGroup * dhwOut;177 outputOneGroupSize = convOps->coPerGroup * dhwOut;
178 CalcStartAddrOriGroupCHW(doIdxStart, diIdxStart);178 CalcStartAddrOriGroupCHW(doIdxStart, diIdxStart);
179 }179 }
180 180 
181- if (convTilingData->convRunInfo.hasBias) {181+ if (convTilingData->hasBias) {
182 biasStartAddr = convOps->groupIdxStart * convOps->coPerGroup + convOps->nIdxStart;182 biasStartAddr = convOps->groupIdxStart * convOps->coPerGroup + convOps->nIdxStart;
183 }183 }
184 184 
@@ -197,30 +197,30 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
197 if (hasScale) {197 if (hasScale) {
198 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {198 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {
199 Extendconv2dFixpipeParams fixpipeParamsCopy(convOps->fixpipeParams);199 Extendconv2dFixpipeParams fixpipeParamsCopy(convOps->fixpipeParams);
200- if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {200+ if (convOps->convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
201 fixpipeParamsCopy.scale0 = convOps->fixpipeParams.scale0[groupIter * coPerGroup];201 fixpipeParamsCopy.scale0 = convOps->fixpipeParams.scale0[groupIter * coPerGroup];
202 }202 }
203 if constexpr (CONV::IS_EXTEND_CONV2D) {203 if constexpr (CONV::IS_EXTEND_CONV2D) {
204- if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {204+ if (convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
205 fixpipeParamsCopy.reluWeight0 = convOps->fixpipeParams.reluWeight0[groupIter * coPerGroup];205 fixpipeParamsCopy.reluWeight0 = convOps->fixpipeParams.reluWeight0[groupIter * coPerGroup];
206 }206 }
207- if (convTilingData->convApiTiling.dualOutput) {207+ if (convTilingData->dualOutput) {
208- if (convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {208+ if (convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
209 fixpipeParamsCopy.scale1 = convOps->fixpipeParams.scale1[groupIter * coPerGroup];209 fixpipeParamsCopy.scale1 = convOps->fixpipeParams.scale1[groupIter * coPerGroup];
210 }210 }
211- if (convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {211+ if (convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
212 fixpipeParamsCopy.reluWeight1 = convOps->fixpipeParams.reluWeight1[groupIter * coPerGroup];212 fixpipeParamsCopy.reluWeight1 = convOps->fixpipeParams.reluWeight1[groupIter * coPerGroup];
213 }213 }
214- if (convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||214+ if (convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
215- convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||215+ convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
216- convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||216+ convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
217- convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {217+ convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
218 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);218 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);
219 }219 }
220- } else if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {220+ } else if (convOps->convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
221 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);221 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);
222 }222 }
223- } else if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {223+ } else if (convOps->convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
224 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);224 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);
225 }225 }
226 } else {226 } else {
@@ -237,14 +237,14 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
237 for (uint64_t groupIter = 0; groupIter < convOps->singleGroups; ++groupIter) {237 for (uint64_t groupIter = 0; groupIter < convOps->singleGroups; ++groupIter) {
238 convOps->conv.SetWeight(convOps->filterGm[groupIter * weightOneGroupSize]);238 convOps->conv.SetWeight(convOps->filterGm[groupIter * weightOneGroupSize]);
239 convOps->conv.SetFmap(convOps->fmapGm[groupIter * fmapOneGroupSize]);239 convOps->conv.SetFmap(convOps->fmapGm[groupIter * fmapOneGroupSize]);
240- if (convTilingData->convRunInfo.hasBias) {240+ if (convTilingData->hasBias) {
241 convOps->conv.SetBias(convOps->biasGm[groupIter * convOps->coPerGroup]);241 convOps->conv.SetBias(convOps->biasGm[groupIter * convOps->coPerGroup]);
242 }242 }
243 243 
244 DealFixpiepParams(groupIter, convOps->coPerGroup);244 DealFixpiepParams(groupIter, convOps->coPerGroup);
245 245 
246 if constexpr (CONV::IS_EXTEND_CONV2D) {246 if constexpr (CONV::IS_EXTEND_CONV2D) {
247- if (this->convTilingData->convApiTiling.dualOutput) {247+ if (this->convTilingData->dualOutput) {
248 convOps->conv.IterateAll(convOps->outputGm[groupIter * outputOneGroupSize],248 convOps->conv.IterateAll(convOps->outputGm[groupIter * outputOneGroupSize],
249 convOps->output1Gm[groupIter * outputOneGroupSize]);249 convOps->output1Gm[groupIter * outputOneGroupSize]);
250 } else {250 } else {
@@ -264,7 +264,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
264 convOps->singleCoreN = convOps->singleCoOpt;264 convOps->singleCoreN = convOps->singleCoOpt;
265 265 
266 // update singleGroups : real groups nums in singlecore266 // update singleGroups : real groups nums in singlecore
267- enlargeTail = convTilingData->convRunInfo.groups % convTilingData->convRunInfo.enlarge;267+ enlargeTail = convTilingData->groups % convTilingData->enlarge;
268 268 
269 if (unlikely(convOps->isGroupDimTail && enlargeTail != 0)) {269 if (unlikely(convOps->isGroupDimTail && enlargeTail != 0)) {
270 uint64_t realCout = enlargeTail * convOps->coPerGroup;270 uint64_t realCout = enlargeTail * convOps->coPerGroup;
@@ -278,7 +278,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
278 }278 }
279 279 
280 if (unlikely(convOps->isGroupDimTail)) {280 if (unlikely(convOps->isGroupDimTail)) {
281- convOps->singleGroups = enlargeTail == 0 ? convTilingData->convRunInfo.enlarge : enlargeTail;281+ convOps->singleGroups = enlargeTail == 0 ? convTilingData->enlarge : enlargeTail;
282 }282 }
283}283}
284 284 
@@ -287,25 +287,25 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
287 CalcOutputStartAddrOptGroup(const uint64_t doIdxStart)287 CalcOutputStartAddrOptGroup(const uint64_t doIdxStart)
288{288{
289 if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC || CONV::C_FORMAT == ConvFormat::NHWC) {289 if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC || CONV::C_FORMAT == ConvFormat::NHWC) {
290- outputOneGroupSize = convTilingData->convRunInfo.coutOpt;290+ outputOneGroupSize = convTilingData->coutOpt;
291 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +291 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
292- convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt +292+ convOps->groupIdxStart * convTilingData->coutOpt +
293 convOps->nIdxStart;293 convOps->nIdxStart;
294 294 
295 if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC) {295 if constexpr (CONV::C_FORMAT == ConvFormat::NDHWC) {
296- outputStartAddr += doIdxStart * hwOut * convTilingData->convRunInfo.cout;296+ outputStartAddr += doIdxStart * hwOut * convTilingData->cout;
297 }297 }
298 298 
299 if constexpr (CONV::isMMode) {299 if constexpr (CONV::isMMode) {
300- outputStartAddr += convOps->mIdxStart * convTilingData->convRunInfo.cout;300+ outputStartAddr += convOps->mIdxStart * convTilingData->cout;
301 } else {301 } else {
302- outputStartAddr += convOps->hoIdxStart * convTilingData->convRunInfo.wout * convTilingData->convRunInfo.cout;302+ outputStartAddr += convOps->hoIdxStart * convTilingData->wout * convTilingData->cout;
303 if constexpr (CONV::C_FORMAT == ConvFormat::NHWC) {303 if constexpr (CONV::C_FORMAT == ConvFormat::NHWC) {
304- outputStartAddr += convOps->woIdxStart * convTilingData->convRunInfo.cout;304+ outputStartAddr += convOps->woIdxStart * convTilingData->cout;
305 }305 }
306 }306 }
307 } else if constexpr (CONV::C_FORMAT == ConvFormat::NCDHW || CONV::C_FORMAT == ConvFormat::NCHW) {307 } else if constexpr (CONV::C_FORMAT == ConvFormat::NCDHW || CONV::C_FORMAT == ConvFormat::NCHW) {
308- outputOneGroupSize = convTilingData->convRunInfo.coutOpt * dhwOut;308+ outputOneGroupSize = convTilingData->coutOpt * dhwOut;
309 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +309 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
310 convOps->groupIdxStart * outputOneGroupSize +310 convOps->groupIdxStart * outputOneGroupSize +
311 convOps->nIdxStart * dhwOut;311 convOps->nIdxStart * dhwOut;
@@ -317,7 +317,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
317 if constexpr (CONV::isMMode) {317 if constexpr (CONV::isMMode) {
318 outputStartAddr += convOps->mIdxStart;318 outputStartAddr += convOps->mIdxStart;
319 } else {319 } else {
320- outputStartAddr += convOps->hoIdxStart * convTilingData->convRunInfo.wout;320+ outputStartAddr += convOps->hoIdxStart * convTilingData->wout;
321 if constexpr (CONV::C_FORMAT == ConvFormat::NCHW) {321 if constexpr (CONV::C_FORMAT == ConvFormat::NCHW) {
322 outputStartAddr += convOps->woIdxStart;322 outputStartAddr += convOps->woIdxStart;
323 }323 }
@@ -336,10 +336,10 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
336 convOps->groupIdxStart * fmapOneGroupSize;336 convOps->groupIdxStart * fmapOneGroupSize;
337 CalcStartAddrFmapHW();337 CalcStartAddrFmapHW();
338 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z) {338 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z) {
339- weightOneGroupSize = convTilingData->convRunInfo.coutOpt * convTilingData->convRunInfo.cinOpt * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;339+ weightOneGroupSize = convTilingData->coutOpt * convTilingData->cinOpt * kd * convTilingData->kh * convTilingData->kw;
340 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize + convOps->nIdxStart * convOps->k0;340 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize + convOps->nIdxStart * convOps->k0;
341 } else {341 } else {
342- weightOneGroupSize = convTilingData->convRunInfo.coutOpt * convOps->ciPerGroup * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;342+ weightOneGroupSize = convTilingData->coutOpt * convOps->ciPerGroup * kd * convTilingData->kh * convTilingData->kw;
343 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize;343 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize;
344 }344 }
345 345 
@@ -358,18 +358,18 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
358 // weight: Kd -> KhKw -> Ci_opt -> Group(enlarge) -> Co_opt358 // weight: Kd -> KhKw -> Ci_opt -> Group(enlarge) -> Co_opt
359 // output: Batch -> Do -> HoWo -> Group(enlarge) -> Co_opt359 // output: Batch -> Do -> HoWo -> Group(enlarge) -> Co_opt
360 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize +360 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize +
361- convOps->groupIdxStart * convTilingData->convRunInfo.cinOpt;361+ convOps->groupIdxStart * convTilingData->cinOpt;
362 CalcStartAddrFmapHW();362 CalcStartAddrFmapHW();
363 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z) {363 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z) {
364- weightOneGroupSize = convTilingData->convRunInfo.coutOpt * convTilingData->convRunInfo.cinOpt * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;364+ weightOneGroupSize = convTilingData->coutOpt * convTilingData->cinOpt * kd * convTilingData->kh * convTilingData->kw;
365 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize + convOps->nIdxStart * convOps->k0;365 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize + convOps->nIdxStart * convOps->k0;
366 } else {366 } else {
367- weightOneGroupSize = convTilingData->convRunInfo.coutOpt;367+ weightOneGroupSize = convTilingData->coutOpt;
368- weightStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt;368+ weightStartAddr = convOps->groupIdxStart * convTilingData->coutOpt;
369 }369 }
370 370 
371 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {371 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
372- fmStartAddr += diIdxStart * hwIn * convTilingData->convRunInfo.cin;372+ fmStartAddr += diIdxStart * hwIn * convTilingData->cin;
373 }373 }
374 374 
375 CalcOutputStartAddrOptGroup(doIdxStart);375 CalcOutputStartAddrOptGroup(doIdxStart);
@@ -380,38 +380,38 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
380 CalcStartAddrOptGroup(const uint64_t din, const uint64_t dout, const uint32_t kd, const uint64_t doIdxStart,380 CalcStartAddrOptGroup(const uint64_t din, const uint64_t dout, const uint32_t kd, const uint64_t doIdxStart,
381 const int64_t diIdxStart)381 const int64_t diIdxStart)
382{382{
383- hwIn = convTilingData->convRunInfo.hin * convTilingData->convRunInfo.win;383+ hwIn = convTilingData->hin * convTilingData->win;
384- hwOut = convTilingData->convRunInfo.hout * convTilingData->convRunInfo.wout;384+ hwOut = convTilingData->hout * convTilingData->wout;
385 dhwOut = dout * hwOut;385 dhwOut = dout * hwOut;
386 dhwIn = din * hwIn;386 dhwIn = din * hwIn;
387- convOps->fmapOneBatchSize = convTilingData->convRunInfo.cin * dhwIn;387+ convOps->fmapOneBatchSize = convTilingData->cin * dhwIn;
388- convOps->outputOneBatchSize = convTilingData->convRunInfo.cout * dhwOut;388+ convOps->outputOneBatchSize = convTilingData->cout * dhwOut;
389 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {389 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {
390- fmapOneGroupSize = convTilingData->convRunInfo.cinOpt;390+ fmapOneGroupSize = convTilingData->cinOpt;
391 CalcStartAddrOptGroupHWC(doIdxStart, diIdxStart, kd);391 CalcStartAddrOptGroupHWC(doIdxStart, diIdxStart, kd);
392 } else {392 } else {
393- fmapOneGroupSize = convTilingData->convRunInfo.cinOpt * dhwIn;393+ fmapOneGroupSize = convTilingData->cinOpt * dhwIn;
394 CalcStartAddrOptGroupCHW(doIdxStart, diIdxStart, kd);394 CalcStartAddrOptGroupCHW(doIdxStart, diIdxStart, kd);
395 }395 }
396 396 
397- if (convTilingData->convRunInfo.hasBias) {397+ if (convTilingData->hasBias) {
398- biasStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt + convOps->nIdxStart;398+ biasStartAddr = convOps->groupIdxStart * convTilingData->coutOpt + convOps->nIdxStart;
399 }399 }
400 400 
401 if constexpr (CONV::isQuant) {401 if constexpr (CONV::isQuant) {
402 if (hasScale) {402 if (hasScale) {
403- scaleStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt + convOps->nIdxStart;403+ scaleStartAddr = convOps->groupIdxStart * convTilingData->coutOpt + convOps->nIdxStart;
404 }404 }
405 }405 }
406 406 
407 if constexpr (CONV::IS_EXTEND_CONV2D) {407 if constexpr (CONV::IS_EXTEND_CONV2D) {
408- if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||408+ if (convOps->convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
409- convOps->convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {409+ convOps->convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
410- scaleStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt + convOps->nIdxStart;410+ scaleStartAddr = convOps->groupIdxStart * convTilingData->coutOpt + convOps->nIdxStart;
411 }411 }
412- if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||412+ if (convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
413- convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {413+ convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
414- reluWeightStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt + convOps->nIdxStart;414+ reluWeightStartAddr = convOps->groupIdxStart * convTilingData->coutOpt + convOps->nIdxStart;
415 }415 }
416 }416 }
417}417}
@@ -439,14 +439,14 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
439 convOps->conv.SetFmap(convOps->fmapGm[groupOptIter * fmapOneGroupSize]);439 convOps->conv.SetFmap(convOps->fmapGm[groupOptIter * fmapOneGroupSize]);
440 convOps->conv.SetWeight(convOps->filterGm[groupOptIter * weightOneGroupSize]);440 convOps->conv.SetWeight(convOps->filterGm[groupOptIter * weightOneGroupSize]);
441 441 
442- if (convTilingData->convRunInfo.hasBias) {442+ if (convTilingData->hasBias) {
443- convOps->conv.SetBias(convOps->biasGm[groupOptIter * convTilingData->convRunInfo.coutOpt]);443+ convOps->conv.SetBias(convOps->biasGm[groupOptIter * convTilingData->coutOpt]);
444 }444 }
445 445 
446- DealFixpiepParams(groupOptIter, convTilingData->convRunInfo.coutOpt);446+ DealFixpiepParams(groupOptIter, convTilingData->coutOpt);
447 447 
448 if constexpr (CONV::IS_EXTEND_CONV2D) {448 if constexpr (CONV::IS_EXTEND_CONV2D) {
449- if (this->convTilingData->convApiTiling.dualOutput) {449+ if (this->convTilingData->dualOutput) {
450 convOps->conv.IterateAll(convOps->outputGm[groupOptIter * outputOneGroupSize],450 convOps->conv.IterateAll(convOps->outputGm[groupOptIter * outputOneGroupSize],
451 convOps->output1Gm[groupOptIter * outputOneGroupSize]);451 convOps->output1Gm[groupOptIter * outputOneGroupSize]);
452 } else {452 } else {
@@ -471,7 +471,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
471 }471 }
472 convOps->conv.SetOptGroupParams(enlargeTail, convOps->singleGroupOpt, convOps->singleCoOpt);472 convOps->conv.SetOptGroupParams(enlargeTail, convOps->singleGroupOpt, convOps->singleCoOpt);
473 } else {473 } else {
474- convOps->conv.SetOptGroupParams(convTilingData->convRunInfo.enlarge, convOps->singleGroupOpt, convOps->singleCoOpt);474+ convOps->conv.SetOptGroupParams(convTilingData->enlarge, convOps->singleGroupOpt, convOps->singleCoOpt);
475 }475 }
476 476 
477 convOps->conv.SetIterIndex(groupOptIter);477 convOps->conv.SetIterIndex(groupOptIter);
@@ -479,12 +479,12 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
479 convOps->conv.SetFmap(convOps->fmapGm);479 convOps->conv.SetFmap(convOps->fmapGm);
480 convOps->conv.SetWeight(convOps->filterGm);480 convOps->conv.SetWeight(convOps->filterGm);
481 481 
482- if (convTilingData->convRunInfo.hasBias) {482+ if (convTilingData->hasBias) {
483 convOps->conv.SetBias(convOps->biasGm);483 convOps->conv.SetBias(convOps->biasGm);
484 }484 }
485 485
486 // quant need, current not care486 // quant need, current not care
487- DealFixpiepParams(groupOptIter, convTilingData->convRunInfo.coutOpt);487+ DealFixpiepParams(groupOptIter, convTilingData->coutOpt);
488 488 
489 if constexpr (CONV::IS_EXTEND_CONV2D) {489 if constexpr (CONV::IS_EXTEND_CONV2D) {
490 if (convOps->dualOutput) {490 if (convOps->dualOutput) {
@@ -509,7 +509,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
509 convOps->singleCoreBatch);509 convOps->singleCoreBatch);
510 } else {510 } else {
511 convOps->conv.SetSingleOutputShape(convOps->singleCoOpt, convOps->singleCoreDout, convOps->singleCoreHo,511 convOps->conv.SetSingleOutputShape(convOps->singleCoOpt, convOps->singleCoreDout, convOps->singleCoreHo,
512- convTilingData->convRunInfo.wout, convOps->singleCoreBatch);512+ convTilingData->wout, convOps->singleCoreBatch);
513 }513 }
514 } else {514 } else {
515 if constexpr (CONV::isMMode) {515 if constexpr (CONV::isMMode) {
Mconv/common/op_kernel/arch35/conv_instr_hw_mode_impl.h+74-74
@@ -32,31 +32,31 @@ public:
32 self_ = self;32 self_ = self;
33 if constexpr (Intf::isConv3D) {33 if constexpr (Intf::isConv3D) {
34 if constexpr (Intf::kPreLoadFlag) {34 if constexpr (Intf::kPreLoadFlag) {
35- channelSize_ = self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ? self_->ctx.convTilingData->convApiTiling.cinAInCore :35+ channelSize_ = self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ? self_->ctx.convTilingData->cinAInCore :
36- self_->ctx.convTilingData->convApiTiling.cinATailInCore;36+ self_->ctx.convTilingData->cinATailInCore;
37 } else {37 } else {
38- channelSize_ = self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ? self_->ctx.convTilingData->convApiTiling.cinAInCore :38+ channelSize_ = self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ? self_->ctx.convTilingData->cinAInCore :
39- self_->ctx.convTilingData->convApiTiling.cinATailInCore;39+ self_->ctx.convTilingData->cinATailInCore;
40 }40 }
41 } else {41 } else {
42 if constexpr (Intf::c04Flag) {42 if constexpr (Intf::c04Flag) {
43 channelSize_ = conv::C04_CIN_SIZE;43 channelSize_ = conv::C04_CIN_SIZE;
44- c04KStepTail = (channelSize_ * self_->ctx.convTilingData->convApiTiling.kernelHxkernelW) % self_->ctx.convTilingData->convApiTiling.kL0;44+ c04KStepTail = (channelSize_ * self_->ctx.convTilingData->kernelHxkernelW) % self_->ctx.convTilingData->kL0;
45- c04KStepTail = c04KStepTail == 0 ? self_->ctx.convTilingData->convApiTiling.kL0 : c04KStepTail;45+ c04KStepTail = c04KStepTail == 0 ? self_->ctx.convTilingData->kL0 : c04KStepTail;
46 } else {46 } else {
47 if constexpr (Intf::kPreLoadFlag) {47 if constexpr (Intf::kPreLoadFlag) {
48 if constexpr (Intf::isKL1NL0FullLoad) {48 if constexpr (Intf::isKL1NL0FullLoad) {
49- channelSize_ = AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0);49+ channelSize_ = AlignB(self_->ctx.convTilingData->cinATailInCore, Intf::k0);
50 } else {50 } else {
51 channelSize_ = self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ?51 channelSize_ = self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ?
52- self_->ctx.convTilingData->convApiTiling.cinAInCore : AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0);52+ self_->ctx.convTilingData->cinAInCore : AlignB(self_->ctx.convTilingData->cinATailInCore, Intf::k0);
53 }53 }
54 } else {54 } else {
55 if constexpr (Intf::isKL1NL0FullLoad) {55 if constexpr (Intf::isKL1NL0FullLoad) {
56- channelSize_ = AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0);56+ channelSize_ = AlignB(self_->ctx.convTilingData->cinATailInCore, Intf::k0);
57 } else {57 } else {
58 channelSize_ = self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ?58 channelSize_ = self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ?
59- self_->ctx.convTilingData->convApiTiling.cinAInCore : AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0);59+ self_->ctx.convTilingData->cinAInCore : AlignB(self_->ctx.convTilingData->cinATailInCore, Intf::k0);
60 }60 }
61 }61 }
62 }62 }
@@ -68,8 +68,8 @@ public:
68 currentML0_ = m;68 currentML0_ = m;
69 if constexpr (!Intf::isDmaFlag) {69 if constexpr (!Intf::isDmaFlag) {
70 currentML0Align_ = mNotAlign;70 currentML0Align_ = mNotAlign;
71- mStartPt_ = self_->ctx.hoL0Iter * self_->ctx.currentWoL1 * self_->ctx.convTilingData->convApiTiling.hoL0 +71+ mStartPt_ = self_->ctx.hoL0Iter * self_->ctx.currentWoL1 * self_->ctx.convTilingData->hoL0 +
72- self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;72+ self_->ctx.woL0Iter * self_->ctx.convTilingData->woL0;
73 SetLoad3dRepeatParams();73 SetLoad3dRepeatParams();
74 }74 }
75 }75 }
@@ -82,14 +82,14 @@ public:
82 uint8_t repeatMode = 0;82 uint8_t repeatMode = 0;
83 mExtension_ = currentML0Align_;83 mExtension_ = currentML0Align_;
84 84 
85- if (self_->ctx.currentWoL0 == self_->ctx.convTilingData->convApiTiling.woL0 &&85+ if (self_->ctx.currentWoL0 == self_->ctx.convTilingData->woL0 &&
86- self_->ctx.currentHoL0 == self_->ctx.convTilingData->convApiTiling.hoL0) {86+ self_->ctx.currentHoL0 == self_->ctx.convTilingData->hoL0) {
87- dstStride = self_->ctx.convTilingData->convApiTiling.fmapKStride;87+ dstStride = self_->ctx.convTilingData->fmapKStride;
88 } else {88 } else {
89 dstStride = currentML0_ / BLOCK_L0_M;89 dstStride = currentML0_ / BLOCK_L0_M;
90 }90 }
91 91 
92- if (unlikely(CeilDiv(self_->ctx.currentWoL1, self_->ctx.convTilingData->convApiTiling.woL0) > 1)) {92+ if (unlikely(CeilDiv(self_->ctx.currentWoL1, self_->ctx.convTilingData->woL0) > 1)) {
93 mExtension_ = self_->ctx.currentWoL0;93 mExtension_ = self_->ctx.currentWoL0;
94 repeatStride = self_->ctx.currentWoL1 / BLOCK_L0_M;94 repeatStride = self_->ctx.currentWoL1 / BLOCK_L0_M;
95 repeatTime = self_->ctx.currentHoL0;95 repeatTime = self_->ctx.currentHoL0;
@@ -108,7 +108,7 @@ public:
108 xm_.bf.mExtension_ = mExtension_ & MASK_16;108 xm_.bf.mExtension_ = mExtension_ & MASK_16;
109 xm_.bf.mStartPt_ = mStartPt_ & MASK_16;109 xm_.bf.mStartPt_ = mStartPt_ & MASK_16;
110 110 
111- xt_.n = static_cast<uint64_t>(self_->ctx.convTilingData->convApiTiling.unionDataXt);111+ xt_.n = static_cast<uint64_t>(self_->ctx.convTilingData->unionDataXt);
112 xt_.bf.channelSize = channelSize_;112 xt_.bf.channelSize = channelSize_;
113 param_.SetConfig1(xt_.n);113 param_.SetConfig1(xt_.n);
114 }114 }
@@ -144,17 +144,17 @@ private:
144 __aicore__ inline void DmaLoad2DImpl(const uint64_t &kIter, const LocalTensor<typename Intf::FmapT> &al0)144 __aicore__ inline void DmaLoad2DImpl(const uint64_t &kIter, const LocalTensor<typename Intf::FmapT> &al0)
145 {145 {
146 uint64_t currentWoL1Align = AlignB(self_->ctx.currentWoL1, BLOCK_L0_M);146 uint64_t currentWoL1Align = AlignB(self_->ctx.currentWoL1, BLOCK_L0_M);
147- uint64_t currentKL0 = IsKL0Tail(kIter) ? self_->ctx.kL0Tail : self_->ctx.convTilingData->convApiTiling.kL0;147+ uint64_t currentKL0 = IsKL0Tail(kIter) ? self_->ctx.kL0Tail : self_->ctx.convTilingData->kL0;
148 148
149 LoadData2DParamsV2 loadParams;149 LoadData2DParamsV2 loadParams;
150- loadParams.kStartPosition = (kIter % self_->ctx.multiKAL1) * self_->ctx.convTilingData->convApiTiling.kL0 / Intf::k0;150+ loadParams.kStartPosition = (kIter % self_->ctx.multiKAL1) * self_->ctx.convTilingData->kL0 / Intf::k0;
151 loadParams.mStep = CeilDiv(self_->ctx.currentWoL0, BLOCK_L0_M);151 loadParams.mStep = CeilDiv(self_->ctx.currentWoL0, BLOCK_L0_M);
152 loadParams.kStep = currentKL0 / Intf::k0;152 loadParams.kStep = currentKL0 / Intf::k0;
153 loadParams.srcStride = self_->ctx.currentHoL1 * currentWoL1Align / BLOCK_L0_M;153 loadParams.srcStride = self_->ctx.currentHoL1 * currentWoL1Align / BLOCK_L0_M;
154 loadParams.dstStride = currentML0_ / BLOCK_L0_M;154 loadParams.dstStride = currentML0_ / BLOCK_L0_M;
155 155
156- uint32_t mStartPosition = self_->ctx.hoL0Iter * self_->ctx.convTilingData->convApiTiling.hoL0 * currentWoL1Align +156+ uint32_t mStartPosition = self_->ctx.hoL0Iter * self_->ctx.convTilingData->hoL0 * currentWoL1Align +
157- self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;157+ self_->ctx.woL0Iter * self_->ctx.convTilingData->woL0;
158 uint32_t dstOffset = 0;158 uint32_t dstOffset = 0;
159 uint32_t dstOffsetStride = AlignB(self_->ctx.currentWoL0, BLOCK_L0_M) * Intf::k0;159 uint32_t dstOffsetStride = AlignB(self_->ctx.currentWoL0, BLOCK_L0_M) * Intf::k0;
160 for (uint16_t hoL0Idx = 0; hoL0Idx < self_->ctx.currentHoL0; ++ hoL0Idx) {160 for (uint16_t hoL0Idx = 0; hoL0Idx < self_->ctx.currentHoL0; ++ hoL0Idx) {
@@ -185,7 +185,7 @@ public:
185 __aicore__ inline void SetParams(Intf *self)185 __aicore__ inline void SetParams(Intf *self)
186 {186 {
187 self_ = self;187 self_ = self;
188- valueHoWo_ = self_->ctx.convTilingData->convApiTiling.orgHo * self_->ctx.convTilingData->convApiTiling.orgWo;188+ valueHoWo_ = self_->ctx.convTilingData->orgHo * self_->ctx.convTilingData->orgWo;
189 }189 }
190 190 
191 __aicore__ inline void SetMN(uint64_t m, uint64_t n)191 __aicore__ inline void SetMN(uint64_t m, uint64_t n)
@@ -202,7 +202,7 @@ public:
202 intriParams.params.dnNum = self_->ctx.currentHoL0;202 intriParams.params.dnNum = self_->ctx.currentHoL0;
203 intriParams.params.dstDnMatrixStride = self_->ctx.singleCoreWo;203 intriParams.params.dstDnMatrixStride = self_->ctx.singleCoreWo;
204 } else {204 } else {
205- if (likely(self_->ctx.convTilingData->convApiTiling.woL0 >= self_->ctx.singleCoreWo)) {205+ if (likely(self_->ctx.convTilingData->woL0 >= self_->ctx.singleCoreWo)) {
206 intriParams.mSize = currentML0_;206 intriParams.mSize = currentML0_;
207 intriParams.params.srcNzMatrixStride = 0;207 intriParams.params.srcNzMatrixStride = 0;
208 intriParams.params.dnNum = 1;208 intriParams.params.dnNum = 1;
@@ -218,7 +218,7 @@ public:
218 intriParams.nSize = currentNL0_;218 intriParams.nSize = currentNL0_;
219 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);219 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);
220 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {220 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {
221- intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgDo * valueHoWo_;221+ intriParams.dstStride = self_->ctx.convTilingData->orgDo * valueHoWo_;
222 } else {222 } else {
223 intriParams.dstStride = valueHoWo_;223 intriParams.dstStride = valueHoWo_;
224 }224 }
@@ -232,8 +232,8 @@ public:
232 intriParams.mSize = self_->ctx.currentWoL0;232 intriParams.mSize = self_->ctx.currentWoL0;
233 intriParams.params.srcNdStride = AlignB(self_->ctx.currentWoL0, BLOCK_L0_M);233 intriParams.params.srcNdStride = AlignB(self_->ctx.currentWoL0, BLOCK_L0_M);
234 intriParams.params.ndNum = self_->ctx.currentHoL0;234 intriParams.params.ndNum = self_->ctx.currentHoL0;
235- intriParams.params.dstNdStride = self_->ctx.singleCoreWo * self_->ctx.convTilingData->convApiTiling.orgCo;235+ intriParams.params.dstNdStride = self_->ctx.singleCoreWo * self_->ctx.convTilingData->orgCo;
236- } else if (likely(self_->ctx.convTilingData->convApiTiling.woL0 >= self_->ctx.singleCoreWo)) {236+ } else if (likely(self_->ctx.convTilingData->woL0 >= self_->ctx.singleCoreWo)) {
237 intriParams.mSize = currentML0_;237 intriParams.mSize = currentML0_;
238 intriParams.params.srcNdStride = 0;238 intriParams.params.srcNdStride = 0;
239 intriParams.params.ndNum = 1;239 intriParams.params.ndNum = 1;
@@ -242,10 +242,10 @@ public:
242 intriParams.mSize = self_->ctx.currentWoL0;242 intriParams.mSize = self_->ctx.currentWoL0;
243 intriParams.params.srcNdStride = self_->ctx.currentWoL0;243 intriParams.params.srcNdStride = self_->ctx.currentWoL0;
244 intriParams.params.ndNum = self_->ctx.currentHoL0;244 intriParams.params.ndNum = self_->ctx.currentHoL0;
245- intriParams.params.dstNdStride = self_->ctx.singleCoreWo * self_->ctx.convTilingData->convApiTiling.orgCo;245+ intriParams.params.dstNdStride = self_->ctx.singleCoreWo * self_->ctx.convTilingData->orgCo;
246 }246 }
247 247 
248- intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgCo;248+ intriParams.dstStride = self_->ctx.convTilingData->orgCo;
249 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);249 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);
250 intriParams.nSize = currentNL0_;250 intriParams.nSize = currentNL0_;
251 SetBaseParams<CO2Layout::ROW_MAJOR>(intriParams);251 SetBaseParams<CO2Layout::ROW_MAJOR>(intriParams);
@@ -267,21 +267,21 @@ public:
267 if constexpr (Intf::isKL1NL0FullLoad) {267 if constexpr (Intf::isKL1NL0FullLoad) {
268 ubInfo->outCIdx = ubInfo->nLoopIdx * ubInfo->nUb;268 ubInfo->outCIdx = ubInfo->nLoopIdx * ubInfo->nUb;
269 } else {269 } else {
270- ubInfo->outCIdx = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +270+ ubInfo->outCIdx = self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
271- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0 + ubInfo->nLoopIdx * ubInfo->nUb;271+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0 + ubInfo->nLoopIdx * ubInfo->nUb;
272 }272 }
273- ubInfo->outHIdx = self_->ctx.hoAL1Iter * self_->ctx.convTilingData->convApiTiling.hoL1 +273+ ubInfo->outHIdx = self_->ctx.hoAL1Iter * self_->ctx.convTilingData->hoL1 +
274- self_->ctx.hoL0Iter * self_->ctx.convTilingData->convApiTiling.hoL0 + ubInfo->mLoopIdx * ubInfo->mUb / self_->ctx.convTilingData->convApiTiling.woL0;274+ self_->ctx.hoL0Iter * self_->ctx.convTilingData->hoL0 + ubInfo->mLoopIdx * ubInfo->mUb / self_->ctx.convTilingData->woL0;
275 if (self_->ctx.woL1SmallTail == 0) {275 if (self_->ctx.woL1SmallTail == 0) {
276- ubInfo->outWIdx = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 +276+ ubInfo->outWIdx = self_->ctx.woAL1Iter * self_->ctx.convTilingData->woL1 +
277- self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->convApiTiling.woL0;277+ self_->ctx.woL0Iter * self_->ctx.convTilingData->woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->woL0;
278 } else {278 } else {
279 if (self_->ctx.woAL1Iter == self_->ctx.maxWoL1Iter) {279 if (self_->ctx.woAL1Iter == self_->ctx.maxWoL1Iter) {
280- ubInfo->outWIdx = ((self_->ctx.woAL1Iter - 1) * self_->ctx.convTilingData->convApiTiling.woL1 + self_->ctx.woAL1Tail) +280+ ubInfo->outWIdx = ((self_->ctx.woAL1Iter - 1) * self_->ctx.convTilingData->woL1 + self_->ctx.woAL1Tail) +
281- self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->convApiTiling.woL0;281+ self_->ctx.woL0Iter * self_->ctx.convTilingData->woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->woL0;
282 } else {282 } else {
283- ubInfo->outWIdx = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 +283+ ubInfo->outWIdx = self_->ctx.woAL1Iter * self_->ctx.convTilingData->woL1 +
284- self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->convApiTiling.woL0;284+ self_->ctx.woL0Iter * self_->ctx.convTilingData->woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->woL0;
285 }285 }
286 }286 }
287 287 
@@ -307,29 +307,29 @@ public:
307 __aicore__ inline void SetBaseParams(FixpipeParamsC310<format> &intriParams)307 __aicore__ inline void SetBaseParams(FixpipeParamsC310<format> &intriParams)
308 {308 {
309 intriParams.quantPre = GetQuantPre<Intf, OutputT, FixpipeIdx>(self_);309 intriParams.quantPre = GetQuantPre<Intf, OutputT, FixpipeIdx>(self_);
310- if (self_->ctx.convTilingData->convApiTiling.hasScale == 0) {310+ if (self_->ctx.convTilingData->hasScale == 0) {
311 intriParams.deqScalar = DEQ_SCALAR_ONE;311 intriParams.deqScalar = DEQ_SCALAR_ONE;
312 }312 }
313 if constexpr (Intf::isExtendConv2d) {313 if constexpr (Intf::isExtendConv2d) {
314 if constexpr (FixpipeIdx == 0) {314 if constexpr (FixpipeIdx == 0) {
315- intriParams.reluEn = self_->ctx.convTilingData->convApiTiling.reluMode0 != 0;315+ intriParams.reluEn = self_->ctx.convTilingData->reluMode0 != 0;
316#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)316#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
317- intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->convApiTiling.reluMode0);317+ intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->reluMode0);
318- if (self_->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {318+ if (self_->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
319 intriParams.reluScalar = self_->ctx.preReluScalar0;319 intriParams.reluScalar = self_->ctx.preReluScalar0;
320- } else if (self_->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {320+ } else if (self_->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
321 intriParams.vectorRelu =321 intriParams.vectorRelu =
322 self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr()].GetPhyAddr();322 self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr()].GetPhyAddr();
323 }323 }
324#endif324#endif
325 intriParams.deqScalar = self_->ctx.deqScalar0;325 intriParams.deqScalar = self_->ctx.deqScalar0;
326 } else {326 } else {
327- intriParams.reluEn = self_->ctx.convTilingData->convApiTiling.reluMode1 != 0;327+ intriParams.reluEn = self_->ctx.convTilingData->reluMode1 != 0;
328#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)328#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
329- intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->convApiTiling.reluMode1);329+ intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->reluMode1);
330- if (self_->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {330+ if (self_->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
331 intriParams.reluScalar = self_->ctx.preReluScalar1;331 intriParams.reluScalar = self_->ctx.preReluScalar1;
332- } else if(self_->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {332+ } else if(self_->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
333 intriParams.vectorRelu =333 intriParams.vectorRelu =
334 self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr() + self_->ctx.reluWeight1L1offset].GetPhyAddr();334 self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr() + self_->ctx.reluWeight1L1offset].GetPhyAddr();
335 }335 }
@@ -341,7 +341,7 @@ public:
341 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;341 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
342 } else {342 } else {
343 if constexpr (Intf::isExtendConv2d) {343 if constexpr (Intf::isExtendConv2d) {
344- if (self_->ctx.convTilingData->convApiTiling.dualOutput) {344+ if (self_->ctx.convTilingData->dualOutput) {
345 intriParams.unitFlag = UNIT_FLAG_ENABLE_ONLY;345 intriParams.unitFlag = UNIT_FLAG_ENABLE_ONLY;
346 } else {346 } else {
347 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;347 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
@@ -355,42 +355,42 @@ public:
355 __aicore__ inline uint64_t CalcFixpipeOffset()355 __aicore__ inline uint64_t CalcFixpipeOffset()
356 {356 {
357 uint64_t offset = self_->ctx.batchIter * self_->ctx.outputOneBatchSize;357 uint64_t offset = self_->ctx.batchIter * self_->ctx.outputOneBatchSize;
358- uint64_t offsetH = self_->ctx.hoAL1Iter * self_->ctx.convTilingData->convApiTiling.hoL1 +358+ uint64_t offsetH = self_->ctx.hoAL1Iter * self_->ctx.convTilingData->hoL1 +
359- self_->ctx.hoL0Iter * self_->ctx.convTilingData->convApiTiling.hoL0;359+ self_->ctx.hoL0Iter * self_->ctx.convTilingData->hoL0;
360 uint64_t offsetW;360 uint64_t offsetW;
361 if (self_->ctx.woL1SmallTail == 0) {361 if (self_->ctx.woL1SmallTail == 0) {
362- offsetW = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 +362+ offsetW = self_->ctx.woAL1Iter * self_->ctx.convTilingData->woL1 +
363- self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;363+ self_->ctx.woL0Iter * self_->ctx.convTilingData->woL0;
364 } else {364 } else {
365 if (self_->ctx.woAL1Iter == self_->ctx.maxWoL1Iter) {365 if (self_->ctx.woAL1Iter == self_->ctx.maxWoL1Iter) {
366- offsetW = ((self_->ctx.woAL1Iter - 1) * self_->ctx.convTilingData->convApiTiling.woL1 + self_->ctx.woAL1Tail) +366+ offsetW = ((self_->ctx.woAL1Iter - 1) * self_->ctx.convTilingData->woL1 + self_->ctx.woAL1Tail) +
367- self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;367+ self_->ctx.woL0Iter * self_->ctx.convTilingData->woL0;
368 } else {368 } else {
369- offsetW = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 +369+ offsetW = self_->ctx.woAL1Iter * self_->ctx.convTilingData->woL1 +
370- self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;370+ self_->ctx.woL0Iter * self_->ctx.convTilingData->woL0;
371 }371 }
372 }372 }
373 373 
374 uint64_t offsetCout = 0U;374 uint64_t offsetCout = 0U;
375 if constexpr (!Intf::isKL1NL0FullLoad) {375 if constexpr (!Intf::isKL1NL0FullLoad) {
376- offsetCout = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +376+ offsetCout = self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
377- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;377+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0;
378 }378 }
379 if constexpr (Intf::groupOptPreloadFlag) {379 if constexpr (Intf::groupOptPreloadFlag) {
380- offsetCout += self_->ctx.groupOptIter * self_->ctx.convTilingData->convApiTiling.orgCo /380+ offsetCout += self_->ctx.groupOptIter * self_->ctx.convTilingData->orgCo /
381- self_->ctx.convTilingData->convApiTiling.groups *381+ self_->ctx.convTilingData->groups *
382- self_->ctx.convTilingData->convApiTiling.enlarge; 382+ self_->ctx.convTilingData->enlarge;
383 }383 }
384 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {384 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {
385- offset += offsetCout * self_->ctx.convTilingData->convApiTiling.orgDo * valueHoWo_ + self_->ctx.dOutIter * valueHoWo_ +385+ offset += offsetCout * self_->ctx.convTilingData->orgDo * valueHoWo_ + self_->ctx.dOutIter * valueHoWo_ +
386- offsetH * self_->ctx.convTilingData->convApiTiling.orgWo + offsetW;386+ offsetH * self_->ctx.convTilingData->orgWo + offsetW;
387 } else if constexpr (Intf::formatOutput == ConvFormat::NDHWC) {387 } else if constexpr (Intf::formatOutput == ConvFormat::NDHWC) {
388- offset += self_->ctx.dOutIter * valueHoWo_ * self_->ctx.convTilingData->convApiTiling.orgCo + offsetH * self_->ctx.convTilingData->convApiTiling.orgWo *388+ offset += self_->ctx.dOutIter * valueHoWo_ * self_->ctx.convTilingData->orgCo + offsetH * self_->ctx.convTilingData->orgWo *
389- self_->ctx.convTilingData->convApiTiling.orgCo + offsetW * self_->ctx.convTilingData->convApiTiling.orgCo + offsetCout;389+ self_->ctx.convTilingData->orgCo + offsetW * self_->ctx.convTilingData->orgCo + offsetCout;
390 } else if constexpr (Intf::formatOutput == ConvFormat::NCHW) {390 } else if constexpr (Intf::formatOutput == ConvFormat::NCHW) {
391- offset += offsetCout * valueHoWo_ + offsetH * self_->ctx.convTilingData->convApiTiling.orgWo + offsetW;391+ offset += offsetCout * valueHoWo_ + offsetH * self_->ctx.convTilingData->orgWo + offsetW;
392 } else {392 } else {
393- offset += offsetH * self_->ctx.convTilingData->convApiTiling.orgWo * self_->ctx.convTilingData->convApiTiling.orgCo + offsetW * self_->ctx.convTilingData->convApiTiling.orgCo + offsetCout;393+ offset += offsetH * self_->ctx.convTilingData->orgWo * self_->ctx.convTilingData->orgCo + offsetW * self_->ctx.convTilingData->orgCo + offsetCout;
394 }394 }
395 395 
396 return offset;396 return offset;
@@ -419,9 +419,9 @@ public:
419 if constexpr (Intf::isKL1NL0FullLoad) {419 if constexpr (Intf::isKL1NL0FullLoad) {
420 return 0;420 return 0;
421 }421 }
422- if (self_->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {422+ if (self_->ctx.convTilingData->fixpParamsFullLoadFlag) {
423- return self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +423+ return self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
424- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;424+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0;
425 }425 }
426 return 0;426 return 0;
427 }427 }
@@ -449,7 +449,7 @@ public:
449 if constexpr (Intf::isExtendConv2d) {449 if constexpr (Intf::isExtendConv2d) {
450 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);450 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);
451 } else if constexpr (Intf::isQuantScene) {451 } else if constexpr (Intf::isQuantScene) {
452- if (self_->ctx.convTilingData->convApiTiling.hasScale != 0) {452+ if (self_->ctx.convTilingData->hasScale != 0) {
453 Fixpipe<OutputT, typename Intf::L0cT, config>(453 Fixpipe<OutputT, typename Intf::L0cT, config>(
454 output[offset], self_->ctx.cl0, self_->ctx.scaleL1[GetScaleL1Addr()], intriParams);454 output[offset], self_->ctx.cl0, self_->ctx.scaleL1[GetScaleL1Addr()], intriParams);
455 } else {455 } else {
@@ -472,9 +472,9 @@ private:
472 return 0;472 return 0;
473 }473 }
474 if constexpr (Intf::isQuantScene) {474 if constexpr (Intf::isQuantScene) {
475- if (self_->ctx.convTilingData->convApiTiling.hasScale != 0 && self_->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {475+ if (self_->ctx.convTilingData->hasScale != 0 && self_->ctx.convTilingData->fixpParamsFullLoadFlag) {
476- return self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +476+ return self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
477- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;477+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0;
478 }478 }
479 }479 }
480 return 0;480 return 0;
Mconv/common/op_kernel/arch35/conv_instr_impl.h+28-28
@@ -58,8 +58,8 @@ public:
58 if constexpr (Intf::isKL1NL0FullLoad) {58 if constexpr (Intf::isKL1NL0FullLoad) {
59 LoadChannelWiseL1FullLoad(tensorL1, tensorGm, currentNL0_, 0);59 LoadChannelWiseL1FullLoad(tensorL1, tensorGm, currentNL0_, 0);
60 } else {60 } else {
61- uint64_t tensorGmOffset = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +61+ uint64_t tensorGmOffset = self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
62- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;62+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0;
63 LoadChannelWiseL1FullLoad(tensorL1, tensorGm, currentNL0_, tensorGmOffset);63 LoadChannelWiseL1FullLoad(tensorL1, tensorGm, currentNL0_, tensorGmOffset);
64 }64 }
65 }65 }
@@ -93,15 +93,15 @@ public:
93 }93 }
94 uint32_t offset = 0;94 uint32_t offset = 0;
95 95
96- if (self_->ctx.convTilingData->convApiTiling.biasFullLoadFlag) {96+ if (self_->ctx.convTilingData->biasFullLoadFlag) {
97 if constexpr (!Intf::isKL1NL0FullLoad) {97 if constexpr (!Intf::isKL1NL0FullLoad) {
98- offset += self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +98+ offset += self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
99- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;99+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0;
100 }100 }
101 if constexpr (Intf::groupOptPreloadFlag) {101 if constexpr (Intf::groupOptPreloadFlag) {
102- offset += self_->ctx.groupOptIter * self_->ctx.convTilingData->convApiTiling.orgCo /102+ offset += self_->ctx.groupOptIter * self_->ctx.convTilingData->orgCo /
103- self_->ctx.convTilingData->convApiTiling.groups *103+ self_->ctx.convTilingData->groups *
104- self_->ctx.convTilingData->convApiTiling.enlarge;104+ self_->ctx.convTilingData->enlarge;
105 }105 }
106 }106 }
107 107 
@@ -124,7 +124,7 @@ public:
124 __aicore__ inline void SetParams(Intf *self)124 __aicore__ inline void SetParams(Intf *self)
125 {125 {
126 self_ = self;126 self_ = self;
127- nStep_ = self_->ctx.convTilingData->convApiTiling.nL0 / BLOCK_L0_N;127+ nStep_ = self_->ctx.convTilingData->nL0 / BLOCK_L0_N;
128 }128 }
129 129 
130 __aicore__ inline void SetN(uint64_t n)130 __aicore__ inline void SetN(uint64_t n)
@@ -133,7 +133,7 @@ public:
133 ratioOfNToN0 = n / BLOCK_L0_N;133 ratioOfNToN0 = n / BLOCK_L0_N;
134 } else {134 } else {
135 ratioOfNToN0 = (self_->ctx.nBL1Iter == self_->ctx.maxNBL1Iter && self_->ctx.nL0Iter == self_->ctx.maxNL0Iter) ?135 ratioOfNToN0 = (self_->ctx.nBL1Iter == self_->ctx.maxNBL1Iter && self_->ctx.nL0Iter == self_->ctx.maxNL0Iter) ?
136- n / BLOCK_L0_N : self_->ctx.convTilingData->convApiTiling.nStep;136+ n / BLOCK_L0_N : self_->ctx.convTilingData->nStep;
137 }137 }
138 }138 }
139 139 
@@ -145,7 +145,7 @@ public:
145 param_.SetMStartPosition(static_cast<uint32_t>(self_->ctx.nL0Iter * nStep_));145 param_.SetMStartPosition(static_cast<uint32_t>(self_->ctx.nL0Iter * nStep_));
146 }146 }
147 param_.SetMStep(static_cast<uint16_t>(ratioOfNToN0));147 param_.SetMStep(static_cast<uint16_t>(ratioOfNToN0));
148- param_.SetSrcStride(static_cast<int32_t>(self_->ctx.convTilingData->convApiTiling.nL1DivBlockSize));148+ param_.SetSrcStride(static_cast<int32_t>(self_->ctx.convTilingData->nL1DivBlockSize));
149 param_.SetDstStride(static_cast<uint16_t>(ratioOfNToN0));149 param_.SetDstStride(static_cast<uint16_t>(ratioOfNToN0));
150 param_.SetIfTranspose(false);150 param_.SetIfTranspose(false);
151 }151 }
@@ -166,9 +166,9 @@ public:
166 isLoaded = true;166 isLoaded = true;
167 param_.SetMStartPosition(0);167 param_.SetMStartPosition(0);
168 param_.SetKStartPosition(0);168 param_.SetKStartPosition(0);
169- param_.SetMStep(static_cast<uint16_t>(self_->ctx.convTilingData->convApiTiling.nL0 / BLOCK_L0_N)); 169+ param_.SetMStep(static_cast<uint16_t>(self_->ctx.convTilingData->nL0 / BLOCK_L0_N));
170- param_.SetKStep(static_cast<uint16_t>(AlignB(self_->ctx.convTilingData->convApiTiling.kBL1, Intf::k0)/Intf::k0));170+ param_.SetKStep(static_cast<uint16_t>(AlignB(self_->ctx.convTilingData->kBL1, Intf::k0)/Intf::k0));
171- uint16_t stride = static_cast<uint16_t>(self_->ctx.convTilingData->convApiTiling.nBL1 / BLOCK_L0_N);171+ uint16_t stride = static_cast<uint16_t>(self_->ctx.convTilingData->nBL1 / BLOCK_L0_N);
172 param_.SetSrcStride(stride);172 param_.SetSrcStride(stride);
173 param_.SetDstStride(stride);173 param_.SetDstStride(stride);
174 param_.SetIfTranspose(false);174 param_.SetIfTranspose(false);
@@ -197,7 +197,7 @@ public:
197 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {197 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {
198 uint32_t srcOffset = 0;198 uint32_t srcOffset = 0;
199 uint32_t dstOffset = 0;199 uint32_t dstOffset = 0;
200- uint32_t srcBatchStride = self_->ctx.currentML0Align * self_->ctx.convTilingData->convApiTiling.kL0;200+ uint32_t srcBatchStride = self_->ctx.currentML0Align * self_->ctx.convTilingData->kL0;
201 uint32_t dstBatchStride = self_->ctx.currentML0Align * self_->ctx.currentNL0Align;201 uint32_t dstBatchStride = self_->ctx.currentML0Align * self_->ctx.currentNL0Align;
202 for (uint32_t batchIdx = 0; batchIdx < self_->ctx.innerBatch; batchIdx++) {202 for (uint32_t batchIdx = 0; batchIdx < self_->ctx.innerBatch; batchIdx++) {
203 Mmad<typename Intf::L0cT, typename Intf::FmapT, typename Intf::WeightT>(203 Mmad<typename Intf::L0cT, typename Intf::FmapT, typename Intf::WeightT>(
@@ -228,7 +228,7 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
228 // extend conv2d: may be scalar or vector quant228 // extend conv2d: may be scalar or vector quant
229 if constexpr (AscendC::IsSameType<OutputT, float>::value) {229 if constexpr (AscendC::IsSameType<OutputT, float>::value) {
230 if constexpr (Intf::isExtendConv2d) {230 if constexpr (Intf::isExtendConv2d) {
231- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {231+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
232 return QuantMode_t::VQF322F32_PRE;232 return QuantMode_t::VQF322F32_PRE;
233 } else {233 } else {
234 return QuantMode_t::QF322F32_PRE;234 return QuantMode_t::QF322F32_PRE;
@@ -240,7 +240,7 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
240 240 
241 if constexpr (AscendC::IsSameType<OutputT, half>::value) {241 if constexpr (AscendC::IsSameType<OutputT, half>::value) {
242 if constexpr (Intf::isExtendConv2d) {242 if constexpr (Intf::isExtendConv2d) {
243- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {243+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
244 return QuantMode_t::VQF322F16_PRE;244 return QuantMode_t::VQF322F16_PRE;
245 } else {245 } else {
246 return QuantMode_t::QF322F16_PRE;246 return QuantMode_t::QF322F16_PRE;
@@ -252,7 +252,7 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
252 252 
253 if constexpr (AscendC::IsSameType<OutputT, bfloat16_t>::value) {253 if constexpr (AscendC::IsSameType<OutputT, bfloat16_t>::value) {
254 if constexpr (Intf::isExtendConv2d) {254 if constexpr (Intf::isExtendConv2d) {
255- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {255+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
256 return QuantMode_t::VQF322BF16_PRE;256 return QuantMode_t::VQF322BF16_PRE;
257 } else {257 } else {
258 return QuantMode_t::QF322BF16_PRE;258 return QuantMode_t::QF322BF16_PRE;
@@ -264,16 +264,16 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
264 264 
265 if constexpr (AscendC::IsSameType<OutputT, hifloat8_t>::value) {265 if constexpr (AscendC::IsSameType<OutputT, hifloat8_t>::value) {
266 if constexpr (Intf::isExtendConv2d) {266 if constexpr (Intf::isExtendConv2d) {
267- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {267+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
268 return QuantMode_t::VQF322HIF8_PRE;268 return QuantMode_t::VQF322HIF8_PRE;
269 } else {269 } else {
270 return QuantMode_t::QF322HIF8_PRE;270 return QuantMode_t::QF322HIF8_PRE;
271 }271 }
272 } else {272 } else {
273- if (self->ctx.convTilingData->convApiTiling.hasScale == 0) {273+ if (self->ctx.convTilingData->hasScale == 0) {
274 // conv2d support hif8 in hif8 out274 // conv2d support hif8 in hif8 out
275 return QuantMode_t::QF322HIF8_PRE;275 return QuantMode_t::QF322HIF8_PRE;
276- } else if (self->ctx.convTilingData->convApiTiling.roundMode == ROUND_MODE_ROUND) {276+ } else if (self->ctx.convTilingData->roundMode == ROUND_MODE_ROUND) {
277 // quantconv2d/quantconv3d277 // quantconv2d/quantconv3d
278 return QuantMode_t::VQF322HIF8_PRE;278 return QuantMode_t::VQF322HIF8_PRE;
279 }279 }
@@ -282,7 +282,7 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
282 282 
283 if constexpr (AscendC::IsSameType<OutputT, fp8_e4m3fn_t>::value) {283 if constexpr (AscendC::IsSameType<OutputT, fp8_e4m3fn_t>::value) {
284 if constexpr (Intf::isExtendConv2d) {284 if constexpr (Intf::isExtendConv2d) {
285- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {285+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
286 return QuantMode_t::VQF322FP8_PRE;286 return QuantMode_t::VQF322FP8_PRE;
287 } else {287 } else {
288 return QuantMode_t::QF322FP8_PRE;288 return QuantMode_t::QF322FP8_PRE;
@@ -302,13 +302,13 @@ __aicore__ inline QuantMode_t GetQuantPreInt32(Intf *self)
302 if constexpr (AscendC::IsSameType<OutputT, half>::value) {302 if constexpr (AscendC::IsSameType<OutputT, half>::value) {
303 if constexpr (Intf::isExtendConv2d) {303 if constexpr (Intf::isExtendConv2d) {
304 if constexpr (FixpipeIdx == 0) {304 if constexpr (FixpipeIdx == 0) {
305- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {305+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
306 return QuantMode_t::VDEQF16;306 return QuantMode_t::VDEQF16;
307 } else {307 } else {
308 return QuantMode_t::DEQF16;308 return QuantMode_t::DEQF16;
309 }309 }
310 } else {310 } else {
311- if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {311+ if (self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
312 return QuantMode_t::VDEQF16;312 return QuantMode_t::VDEQF16;
313 } else {313 } else {
314 return QuantMode_t::DEQF16;314 return QuantMode_t::DEQF16;
@@ -323,13 +323,13 @@ __aicore__ inline QuantMode_t GetQuantPreInt32(Intf *self)
323 } else if constexpr (AscendC::IsSameType<OutputT, int8_t>::value) {323 } else if constexpr (AscendC::IsSameType<OutputT, int8_t>::value) {
324 if constexpr (Intf::isExtendConv2d) {324 if constexpr (Intf::isExtendConv2d) {
325 if constexpr (FixpipeIdx == 0) {325 if constexpr (FixpipeIdx == 0) {
326- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {326+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
327 return QuantMode_t::VREQ8;327 return QuantMode_t::VREQ8;
328 } else {328 } else {
329 return QuantMode_t::REQ8;329 return QuantMode_t::REQ8;
330 }330 }
331 } else {331 } else {
332- if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {332+ if (self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
333 return QuantMode_t::VREQ8;333 return QuantMode_t::VREQ8;
334 } else {334 } else {
335 return QuantMode_t::REQ8;335 return QuantMode_t::REQ8;
@@ -359,13 +359,13 @@ __aicore__ inline QuantMode_t GetQuantPreFp32(Intf *self)
359 } else if constexpr (AscendC::IsSameType<OutputT, int8_t>::value) {359 } else if constexpr (AscendC::IsSameType<OutputT, int8_t>::value) {
360 if constexpr (Intf::isExtendConv2d) {360 if constexpr (Intf::isExtendConv2d) {
361 if constexpr (FixpipeIdx == 0) {361 if constexpr (FixpipeIdx == 0) {
362- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {362+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
363 return QuantMode_t::VQF322B8_PRE;363 return QuantMode_t::VQF322B8_PRE;
364 } else {364 } else {
365 return QuantMode_t::QF322B8_PRE;365 return QuantMode_t::QF322B8_PRE;
366 }366 }
367 } else {367 } else {
368- if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {368+ if (self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
369 return QuantMode_t::VQF322B8_PRE;369 return QuantMode_t::VQF322B8_PRE;
370 } else {370 } else {
371 return QuantMode_t::QF322B8_PRE;371 return QuantMode_t::QF322B8_PRE;
Mconv/common/op_kernel/arch35/conv_instr_m_mode_impl.h+49-49
@@ -32,35 +32,35 @@ public:
32 __aicore__ inline void SetParams(Intf *self)32 __aicore__ inline void SetParams(Intf *self)
33 {33 {
34 self_ = self;34 self_ = self;
35- alignCinATailInCore_ = AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0FmapTail);35+ alignCinATailInCore_ = AlignB(self_->ctx.convTilingData->cinATailInCore, Intf::k0FmapTail);
36 36 
37 if constexpr (Intf::c04Flag) {37 if constexpr (Intf::c04Flag) {
38 channelSize_ = conv::C04_CIN_SIZE;38 channelSize_ = conv::C04_CIN_SIZE;
39- c04KStepTail = (conv::C04_CIN_SIZE * self_->ctx.convTilingData->convApiTiling.kernelHxkernelW) % self_->ctx.convTilingData->convApiTiling.kL0;39+ c04KStepTail = (conv::C04_CIN_SIZE * self_->ctx.convTilingData->kernelHxkernelW) % self_->ctx.convTilingData->kL0;
40- c04KStepTail = c04KStepTail == 0 ? self_->ctx.convTilingData->convApiTiling.kL0 : c04KStepTail;40+ c04KStepTail = c04KStepTail == 0 ? self_->ctx.convTilingData->kL0 : c04KStepTail;
41 } else {41 } else {
42 if constexpr (Intf::kPreLoadFlag) {42 if constexpr (Intf::kPreLoadFlag) {
43 if constexpr (Intf::isKL1NL0FullLoad) {43 if constexpr (Intf::isKL1NL0FullLoad) {
44 channelSize_ = alignCinATailInCore_;44 channelSize_ = alignCinATailInCore_;
45 } else {45 } else {
46 channelSize_ = (self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ?46 channelSize_ = (self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ?
47- self_->ctx.convTilingData->convApiTiling.cinAInCore : alignCinATailInCore_);47+ self_->ctx.convTilingData->cinAInCore : alignCinATailInCore_);
48 }48 }
49 } else {49 } else {
50 if constexpr (Intf::isKL1NL0FullLoad) {50 if constexpr (Intf::isKL1NL0FullLoad) {
51 channelSize_ = alignCinATailInCore_;51 channelSize_ = alignCinATailInCore_;
52 } else {52 } else {
53 channelSize_ = (self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ?53 channelSize_ = (self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ?
54- self_->ctx.convTilingData->convApiTiling.cinAInCore : alignCinATailInCore_);54+ self_->ctx.convTilingData->cinAInCore : alignCinATailInCore_);
55 }55 }
56 }56 }
57 }57 }
58 58 
59 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {59 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {
60- uint64_t hiLoadL1 = (self_->ctx.convTilingData->convApiTiling.orgHo - 1) * self_->ctx.convTilingData->convApiTiling.strideH +60+ uint64_t hiLoadL1 = (self_->ctx.convTilingData->orgHo - 1) * self_->ctx.convTilingData->strideH +
61 self_->ctx.dilatedKernelH;61 self_->ctx.dilatedKernelH;
62- hiLoadL1 = hiLoadL1 > self_->ctx.convTilingData->convApiTiling.orgHi ? self_->ctx.convTilingData->convApiTiling.orgHi : hiLoadL1;62+ hiLoadL1 = hiLoadL1 > self_->ctx.convTilingData->orgHi ? self_->ctx.convTilingData->orgHi : hiLoadL1;
63- realHixWi = hiLoadL1 * self_->ctx.convTilingData->convApiTiling.orgWi;63+ realHixWi = hiLoadL1 * self_->ctx.convTilingData->orgWi;
64 }64 }
65 }65 }
66 66 
@@ -90,11 +90,11 @@ public:
90 __aicore__ inline void SetFirst()90 __aicore__ inline void SetFirst()
91 {91 {
92 uint64_t posM = self_->ctx.mL0Iter * self_->ctx.mL0 +92 uint64_t posM = self_->ctx.mL0Iter * self_->ctx.mL0 +
93- (self_->ctx.mStartPos + self_->ctx.mAL1Iter * self_->ctx.mAL1) % self_->ctx.convTilingData->convApiTiling.orgWo;93+ (self_->ctx.mStartPos + self_->ctx.mAL1Iter * self_->ctx.mAL1) % self_->ctx.convTilingData->orgWo;
94 xm_.bf.mExtension_ = currentML0_ & MASK_16;94 xm_.bf.mExtension_ = currentML0_ & MASK_16;
95 xm_.bf.mStartPt_ = posM & MASK_16;95 xm_.bf.mStartPt_ = posM & MASK_16;
96 96 
97- xt_.n = static_cast<uint64_t>(self_->ctx.convTilingData->convApiTiling.unionDataXt);97+ xt_.n = static_cast<uint64_t>(self_->ctx.convTilingData->unionDataXt);
98 xt_.bf.channelSize = channelSize_;98 xt_.bf.channelSize = channelSize_;
99 param_.SetConfig1(xt_.n);99 param_.SetConfig1(xt_.n);
100 }100 }
@@ -115,8 +115,8 @@ public:
115 115 
116 if constexpr (Intf::preFusionFlag) {116 if constexpr (Intf::preFusionFlag) {
117 uint64_t aL1Offset = 0;117 uint64_t aL1Offset = 0;
118- if ((self_->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {118+ if ((self_->ctx.convTilingData->pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {
119- aL1Offset = (self_->ctx.al1PingPongFlag^1) * self_->ctx.convTilingData->convApiTiling.aL1SpaceSize / Intf::sizeOfFmap;119+ aL1Offset = (self_->ctx.al1PingPongFlag^1) * self_->ctx.convTilingData->aL1SpaceSize / Intf::sizeOfFmap;
120 }120 }
121 LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(al0,121 LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(al0,
122 self_->ctx.al1[aL1Offset], param_);122 self_->ctx.al1[aL1Offset], param_);
@@ -125,12 +125,12 @@ public:
125 uint32_t dstOffset = 0;125 uint32_t dstOffset = 0;
126 uint32_t srcBatchStride = 0;126 uint32_t srcBatchStride = 0;
127 if constexpr (Intf::c04Flag) {127 if constexpr (Intf::c04Flag) {
128- srcBatchStride = AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->convApiTiling.orgHixWi, Intf::k0);128+ srcBatchStride = AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->orgHixWi, Intf::k0);
129 } else {129 } else {
130 srcBatchStride = ((kIter / self_->ctx.multiKAL1) != self_->ctx.maxKAL1Iter ?130 srcBatchStride = ((kIter / self_->ctx.multiKAL1) != self_->ctx.maxKAL1Iter ?
131- self_->ctx.convTilingData->convApiTiling.cinAInCore : alignCinATailInCore_) * realHixWi;131+ self_->ctx.convTilingData->cinAInCore : alignCinATailInCore_) * realHixWi;
132 }132 }
133- uint32_t dstBatchStride = currentML0Align_ * self_->ctx.convTilingData->convApiTiling.kL0;133+ uint32_t dstBatchStride = currentML0Align_ * self_->ctx.convTilingData->kL0;
134 for (uint16_t batchIter = 0; batchIter < self_->ctx.innerBatch; batchIter++) {134 for (uint16_t batchIter = 0; batchIter < self_->ctx.innerBatch; batchIter++) {
135 LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(al0[dstOffset],135 LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(al0[dstOffset],
136 self_->ctx.al1[srcOffset], param_);136 self_->ctx.al1[srcOffset], param_);
@@ -164,7 +164,7 @@ public:
164 __aicore__ inline void SetParams(Intf *self)164 __aicore__ inline void SetParams(Intf *self)
165 {165 {
166 self_ = self;166 self_ = self;
167- valueHoWo_ = self_->ctx.convTilingData->convApiTiling.orgHo * self_->ctx.convTilingData->convApiTiling.orgWo;167+ valueHoWo_ = self_->ctx.convTilingData->orgHo * self_->ctx.convTilingData->orgWo;
168 }168 }
169 169 
170 __aicore__ inline void SetMN(uint64_t m, uint64_t n)170 __aicore__ inline void SetMN(uint64_t m, uint64_t n)
@@ -199,7 +199,7 @@ public:
199 if constexpr (Intf::isExtendConv2d) {199 if constexpr (Intf::isExtendConv2d) {
200 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);200 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);
201 } else if constexpr (Intf::isQuantScene) {201 } else if constexpr (Intf::isQuantScene) {
202- if (self_->ctx.convTilingData->convApiTiling.hasScale == 0) {202+ if (self_->ctx.convTilingData->hasScale == 0) {
203 Fixpipe<OutputT, typename Intf::L0cT, config>(output[offset], self_->ctx.cl0, intriParams);203 Fixpipe<OutputT, typename Intf::L0cT, config>(output[offset], self_->ctx.cl0, intriParams);
204 } else {204 } else {
205 Fixpipe<OutputT, typename Intf::L0cT, config>(205 Fixpipe<OutputT, typename Intf::L0cT, config>(
@@ -231,7 +231,7 @@ private:
231 if constexpr (Intf::isExtendConv2d) {231 if constexpr (Intf::isExtendConv2d) {
232 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);232 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);
233 } else if constexpr (Intf::isQuantScene) {233 } else if constexpr (Intf::isQuantScene) {
234- if (self_->ctx.convTilingData->convApiTiling.hasScale != 0) {234+ if (self_->ctx.convTilingData->hasScale != 0) {
235 Fixpipe<OutputT, typename Intf::L0cT, config>(235 Fixpipe<OutputT, typename Intf::L0cT, config>(
236 output[offset], self_->ctx.cl0, self_->ctx.scaleL1[GetScaleL1Addr()], intriParams);236 output[offset], self_->ctx.cl0, self_->ctx.scaleL1[GetScaleL1Addr()], intriParams);
237 } else {237 } else {
@@ -261,7 +261,7 @@ private:
261 intriParams.params.srcNzMatrixStride = self_->ctx.currentML0Align * CeilDiv(currentNL0_ , BLOCK_L0_N);261 intriParams.params.srcNzMatrixStride = self_->ctx.currentML0Align * CeilDiv(currentNL0_ , BLOCK_L0_N);
262 }262 }
263 intriParams.params.srcNzC0Stride = 1;263 intriParams.params.srcNzC0Stride = 1;
264- intriParams.params.dstDnMatrixStride = self_->ctx.convTilingData->convApiTiling.orgCo * valueHoWo_;264+ intriParams.params.dstDnMatrixStride = self_->ctx.convTilingData->orgCo * valueHoWo_;
265 intriParams.dstStride = valueHoWo_;265 intriParams.dstStride = valueHoWo_;
266 }266 }
267 267 
@@ -277,8 +277,8 @@ private:
277 intriParams.srcStride = self_->ctx.currentML0Align;277 intriParams.srcStride = self_->ctx.currentML0Align;
278 intriParams.params.srcNdStride = self_->ctx.currentML0Align * CeilDiv(currentNL0_ , BLOCK_L0_N);278 intriParams.params.srcNdStride = self_->ctx.currentML0Align * CeilDiv(currentNL0_ , BLOCK_L0_N);
279 }279 }
280- intriParams.params.dstNdStride = self_->ctx.convTilingData->convApiTiling.orgCo * valueHoWo_;280+ intriParams.params.dstNdStride = self_->ctx.convTilingData->orgCo * valueHoWo_;
281- intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgCo;281+ intriParams.dstStride = self_->ctx.convTilingData->orgCo;
282 }282 }
283 283 
284 __aicore__ inline void SetFixpipeIntriParamsHWC(FixpipeParamsC310<CO2Layout::ROW_MAJOR> &intriParams)284 __aicore__ inline void SetFixpipeIntriParamsHWC(FixpipeParamsC310<CO2Layout::ROW_MAJOR> &intriParams)
@@ -286,7 +286,7 @@ private:
286 intriParams.nSize = currentNL0_;286 intriParams.nSize = currentNL0_;
287 intriParams.mSize = currentML0_;287 intriParams.mSize = currentML0_;
288 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);288 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);
289- intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgCo;289+ intriParams.dstStride = self_->ctx.convTilingData->orgCo;
290 intriParams.params.ndNum = 1;290 intriParams.params.ndNum = 1;
291 intriParams.params.dstNdStride = 0;291 intriParams.params.dstNdStride = 0;
292 intriParams.params.srcNdStride = 0;292 intriParams.params.srcNdStride = 0;
@@ -299,7 +299,7 @@ private:
299 intriParams.mSize = currentML0_;299 intriParams.mSize = currentML0_;
300 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);300 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);
301 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {301 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {
302- intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgDo * valueHoWo_;302+ intriParams.dstStride = self_->ctx.convTilingData->orgDo * valueHoWo_;
303 } else {303 } else {
304 intriParams.dstStride = valueHoWo_;304 intriParams.dstStride = valueHoWo_;
305 }305 }
@@ -327,8 +327,8 @@ private:
327 if constexpr (Intf::isKL1NL0FullLoad) {327 if constexpr (Intf::isKL1NL0FullLoad) {
328 ubInfo->outCIdx = ubInfo->nLoopIdx * ubInfo->nUb;328 ubInfo->outCIdx = ubInfo->nLoopIdx * ubInfo->nUb;
329 } else {329 } else {
330- ubInfo->outCIdx = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +330+ ubInfo->outCIdx = self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
331- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0 + ubInfo->nLoopIdx * ubInfo->nUb;331+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0 + ubInfo->nLoopIdx * ubInfo->nUb;
332 }332 }
333 ubInfo->outHIdx = 0;333 ubInfo->outHIdx = 0;
334 ubInfo->outWIdx = self_->ctx.mAL1Iter * self_->ctx.mAL1 +334 ubInfo->outWIdx = self_->ctx.mAL1Iter * self_->ctx.mAL1 +
@@ -339,29 +339,29 @@ private:
339 __aicore__ inline void SetBaseParams(FixpipeParamsC310<format> &intriParams)339 __aicore__ inline void SetBaseParams(FixpipeParamsC310<format> &intriParams)
340 {340 {
341 intriParams.quantPre = GetQuantPre<Intf, OutputT, FixpipeIdx>(self_);341 intriParams.quantPre = GetQuantPre<Intf, OutputT, FixpipeIdx>(self_);
342- if (self_->ctx.convTilingData->convApiTiling.hasScale == 0) {342+ if (self_->ctx.convTilingData->hasScale == 0) {
343 intriParams.deqScalar = DEQ_SCALAR_ONE;343 intriParams.deqScalar = DEQ_SCALAR_ONE;
344 }344 }
345 if constexpr (Intf::isExtendConv2d) {345 if constexpr (Intf::isExtendConv2d) {
346 if constexpr (FixpipeIdx == 0) {346 if constexpr (FixpipeIdx == 0) {
347- intriParams.reluEn = self_->ctx.convTilingData->convApiTiling.reluMode0 != 0;347+ intriParams.reluEn = self_->ctx.convTilingData->reluMode0 != 0;
348#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)348#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
349- intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->convApiTiling.reluMode0);349+ intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->reluMode0);
350- if (self_->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {350+ if (self_->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
351 intriParams.reluScalar = self_->ctx.preReluScalar0;351 intriParams.reluScalar = self_->ctx.preReluScalar0;
352- } else if (self_->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {352+ } else if (self_->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
353 intriParams.vectorRelu =353 intriParams.vectorRelu =
354 self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr()].GetPhyAddr();354 self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr()].GetPhyAddr();
355 }355 }
356#endif356#endif
357 intriParams.deqScalar = self_->ctx.deqScalar0;357 intriParams.deqScalar = self_->ctx.deqScalar0;
358 } else {358 } else {
359- intriParams.reluEn = self_->ctx.convTilingData->convApiTiling.reluMode1 != 0;359+ intriParams.reluEn = self_->ctx.convTilingData->reluMode1 != 0;
360#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)360#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
361- intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->convApiTiling.reluMode1);361+ intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->reluMode1);
362- if (self_->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {362+ if (self_->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
363 intriParams.reluScalar = self_->ctx.preReluScalar1;363 intriParams.reluScalar = self_->ctx.preReluScalar1;
364- } else if(self_->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {364+ } else if(self_->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
365 intriParams.vectorRelu =365 intriParams.vectorRelu =
366 self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr() + self_->ctx.reluWeight1L1offset].GetPhyAddr();366 self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr() + self_->ctx.reluWeight1L1offset].GetPhyAddr();
367 }367 }
@@ -375,7 +375,7 @@ private:
375 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;375 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
376 } else {376 } else {
377 if constexpr (Intf::isExtendConv2d) {377 if constexpr (Intf::isExtendConv2d) {
378- if (self_->ctx.convTilingData->convApiTiling.dualOutput) {378+ if (self_->ctx.convTilingData->dualOutput) {
379 intriParams.unitFlag = UNIT_FLAG_ENABLE_ONLY;379 intriParams.unitFlag = UNIT_FLAG_ENABLE_ONLY;
380 } else {380 } else {
381 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;381 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
@@ -445,29 +445,29 @@ private:
445 offset = self_->ctx.batchIter * self_->ctx.outputOneBatchSize;445 offset = self_->ctx.batchIter * self_->ctx.outputOneBatchSize;
446 }446 }
447 if constexpr (Intf::isInnerBatchFlag) {447 if constexpr (Intf::isInnerBatchFlag) {
448- offset *= self_->ctx.convTilingData->convApiTiling.innerBatch;448+ offset *= self_->ctx.convTilingData->innerBatch;
449 }449 }
450 450 
451 uint64_t offsetCout = 0;451 uint64_t offsetCout = 0;
452 if constexpr (Intf::groupOptPreloadFlag) {452 if constexpr (Intf::groupOptPreloadFlag) {
453- offsetCout += self_->ctx.groupOptIter * self_->ctx.convTilingData->convApiTiling.orgCo / self_->ctx.convTilingData->convApiTiling.groups *453+ offsetCout += self_->ctx.groupOptIter * self_->ctx.convTilingData->orgCo / self_->ctx.convTilingData->groups *
454- self_->ctx.convTilingData->convApiTiling.enlarge;454+ self_->ctx.convTilingData->enlarge;
455 }455 }
456 if constexpr (!Intf::isKL1NL0FullLoad) {456 if constexpr (!Intf::isKL1NL0FullLoad) {
457- offsetCout += self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +457+ offsetCout += self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
458- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;458+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0;
459 }459 }
460 uint64_t offsetMAL1 = self_->ctx.mAL1Iter * self_->ctx.mAL1 +460 uint64_t offsetMAL1 = self_->ctx.mAL1Iter * self_->ctx.mAL1 +
461 self_->ctx.mL0Iter * self_->ctx.mL0;461 self_->ctx.mL0Iter * self_->ctx.mL0;
462 462
463 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {463 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {
464- offset += offsetCout * self_->ctx.convTilingData->convApiTiling.orgDo * valueHoWo_ + self_->ctx.dOutIter * valueHoWo_ + offsetMAL1;464+ offset += offsetCout * self_->ctx.convTilingData->orgDo * valueHoWo_ + self_->ctx.dOutIter * valueHoWo_ + offsetMAL1;
465 } else if constexpr (Intf::formatOutput == ConvFormat::NDHWC) {465 } else if constexpr (Intf::formatOutput == ConvFormat::NDHWC) {
466- offset += self_->ctx.dOutIter * valueHoWo_ * self_->ctx.convTilingData->convApiTiling.orgCo + offsetMAL1 * self_->ctx.convTilingData->convApiTiling.orgCo + offsetCout;466+ offset += self_->ctx.dOutIter * valueHoWo_ * self_->ctx.convTilingData->orgCo + offsetMAL1 * self_->ctx.convTilingData->orgCo + offsetCout;
467 } else if constexpr (Intf::formatOutput == ConvFormat::NCHW) {467 } else if constexpr (Intf::formatOutput == ConvFormat::NCHW) {
468 offset += offsetCout * valueHoWo_ + offsetMAL1;468 offset += offsetCout * valueHoWo_ + offsetMAL1;
469 } else {469 } else {
470- offset += offsetMAL1 * self_->ctx.convTilingData->convApiTiling.orgCo + offsetCout;470+ offset += offsetMAL1 * self_->ctx.convTilingData->orgCo + offsetCout;
471 }471 }
472 472 
473 return offset;473 return offset;
@@ -497,9 +497,9 @@ private:
497 return 0;497 return 0;
498 }498 }
499 if constexpr (Intf::isQuantScene) {499 if constexpr (Intf::isQuantScene) {
500- if (self_->ctx.convTilingData->convApiTiling.hasScale != 0 && self_->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {500+ if (self_->ctx.convTilingData->hasScale != 0 && self_->ctx.convTilingData->fixpParamsFullLoadFlag) {
501- return self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +501+ return self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
502- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;502+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0;
503 }503 }
504 }504 }
505 505 
@@ -511,9 +511,9 @@ private:
511 if constexpr (Intf::isKL1NL0FullLoad) {511 if constexpr (Intf::isKL1NL0FullLoad) {
512 return 0;512 return 0;
513 }513 }
514- if (self_->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {514+ if (self_->ctx.convTilingData->fixpParamsFullLoadFlag) {
515- return self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +515+ return self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
516- self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;516+ self_->ctx.nL0Iter * self_->ctx.convTilingData->nL0;
517 }517 }
518 return 0;518 return 0;
519 }519 }
Mconv/common/op_kernel/arch35/conv_instr_opt_group_impl.h+31-31
@@ -63,14 +63,14 @@ public:
63 uint64_t weightOneGroupSize = 0;63 uint64_t weightOneGroupSize = 0;
64 if constexpr (Intf::formatWeight == ConvFormat::NCHW) {64 if constexpr (Intf::formatWeight == ConvFormat::NCHW) {
65 weightOneGroupSize = self_->ctx.coPerGroup * self_->ctx.ciPerGroup * self_->ctx.enlarge *65 weightOneGroupSize = self_->ctx.coPerGroup * self_->ctx.ciPerGroup * self_->ctx.enlarge *
66- self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;66+ self_->ctx.convTilingData->kernelHxkernelWxkernelD;
67 } else if constexpr (Intf::formatWeight == ConvFormat::HWCN) {67 } else if constexpr (Intf::formatWeight == ConvFormat::HWCN) {
68 weightOneGroupSize = self_->ctx.coPerGroup * self_->ctx.enlarge;68 weightOneGroupSize = self_->ctx.coPerGroup * self_->ctx.enlarge;
69 }69 }
70 gmOffset = weightOneGroupSize * self_->ctx.groupOptIter;70 gmOffset = weightOneGroupSize * self_->ctx.groupOptIter;
71 }71 }
72 if constexpr (Intf::formatOutput == ConvFormat::NDHWC || Intf::formatOutput == ConvFormat::NHWC) {72 if constexpr (Intf::formatOutput == ConvFormat::NDHWC || Intf::formatOutput == ConvFormat::NHWC) {
73- uint64_t curSrcCoOpt = self_->ctx.convTilingData->convApiTiling.orgCo;73+ uint64_t curSrcCoOpt = self_->ctx.convTilingData->orgCo;
74 copyParamsHWC.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = curSrcCoOpt;74 copyParamsHWC.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = curSrcCoOpt;
75 copyParamsHWC.loopInfo.loopSrcStride[NDDMA_LOOP3_INDEX] = curSrcCoOpt * self_->ctx.ciPerGroup;75 copyParamsHWC.loopInfo.loopSrcStride[NDDMA_LOOP3_INDEX] = curSrcCoOpt * self_->ctx.ciPerGroup;
76 copyParamsHWC.loopInfo.loopSize[NDDMA_LOOP2_INDEX] = self_->ctx.singleGroups;76 copyParamsHWC.loopInfo.loopSize[NDDMA_LOOP2_INDEX] = self_->ctx.singleGroups;
@@ -86,7 +86,7 @@ public:
86private:86private:
87__aicore__ inline void NDDMAFirstSetCopyParamsCHW()87__aicore__ inline void NDDMAFirstSetCopyParamsCHW()
88 {88 {
89- uint64_t srcKSize = self_->ctx.ciPerGroup * self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;89+ uint64_t srcKSize = self_->ctx.ciPerGroup * self_->ctx.convTilingData->kernelHxkernelWxkernelD;
90 // NDDMA Loop0 params90 // NDDMA Loop0 params
91 copyParams.loopInfo.loopSize[NDDMA_LOOP0_INDEX] = srcKSize;91 copyParams.loopInfo.loopSize[NDDMA_LOOP0_INDEX] = srcKSize;
92 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 1;92 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 1;
@@ -115,7 +115,7 @@ __aicore__ inline void NDDMAFirstSetCopyParamsCHW()
115 copyParamsHWC.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = self_->ctx.coOptAlign * self_->ctx.ciPerGroup +115 copyParamsHWC.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = self_->ctx.coOptAlign * self_->ctx.ciPerGroup +
116 self_->ctx.coPerGroup;116 self_->ctx.coPerGroup;
117 // NDDMA Loop3 params117 // NDDMA Loop3 params
118- copyParamsHWC.loopInfo.loopSize[NDDMA_LOOP3_INDEX] = self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;118+ copyParamsHWC.loopInfo.loopSize[NDDMA_LOOP3_INDEX] = self_->ctx.convTilingData->kernelHxkernelWxkernelD;
119 copyParamsHWC.loopInfo.loopDstStride[NDDMA_LOOP3_INDEX] = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;119 copyParamsHWC.loopInfo.loopDstStride[NDDMA_LOOP3_INDEX] = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;
120 }120 }
121 121 
@@ -173,7 +173,7 @@ private:
173 for (uint8_t idx = 0; idx < Intf::k0; ++idx) {173 for (uint8_t idx = 0; idx < Intf::k0; ++idx) {
174 indexTensor.SetValue(idx, curValue);174 indexTensor.SetValue(idx, curValue);
175 if constexpr (Intf::formatOutput == ConvFormat::NCDHW || Intf::formatOutput == ConvFormat::NCHW) {175 if constexpr (Intf::formatOutput == ConvFormat::NCDHW || Intf::formatOutput == ConvFormat::NCHW) {
176- curValue += self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;176+ curValue += self_->ctx.convTilingData->kernelHxkernelWxkernelD;
177 } else {177 } else {
178 curValue += self_->ctx.coOptAlign;178 curValue += self_->ctx.coOptAlign;
179 }179 }
@@ -212,10 +212,10 @@ private:
212 {212 {
213 uint16_t ciLoopTimes = self_->ctx.ci1Opt;213 uint16_t ciLoopTimes = self_->ctx.ci1Opt;
214 uint16_t coLoopTimes = coOptLoopTimes;214 uint16_t coLoopTimes = coOptLoopTimes;
215- uint16_t khkwLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;215+ uint16_t khkwLoopTimes = self_->ctx.convTilingData->kernelHxkernelW;
216- uint32_t srcCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0;216+ uint32_t srcCiStride = self_->ctx.convTilingData->kernelHxkernelW * Intf::k0;
217 uint32_t srcCoStride = coPerReg * self_->ctx.kUbSize;217 uint32_t srcCoStride = coPerReg * self_->ctx.kUbSize;
218- uint32_t dstCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;218+ uint32_t dstCiStride = self_->ctx.convTilingData->kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;
219 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;219 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;
220 uint32_t dstCoStride = coPerReg * Intf::k0;220 uint32_t dstCoStride = coPerReg * Intf::k0;
221 221 
@@ -265,16 +265,16 @@ private:
265 265 
266 __aicore__ inline void TransNCDHW2NZ()266 __aicore__ inline void TransNCDHW2NZ()
267 {267 {
268- uint16_t kdLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelD;268+ uint16_t kdLoopTimes = self_->ctx.convTilingData->kernelD;
269 uint16_t ciLoopTimes = self_->ctx.ci1Opt;269 uint16_t ciLoopTimes = self_->ctx.ci1Opt;
270 uint16_t coLoopTimes = coOptLoopTimes;270 uint16_t coLoopTimes = coOptLoopTimes;
271- uint16_t khkwLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;271+ uint16_t khkwLoopTimes = self_->ctx.convTilingData->kernelHxkernelW;
272- uint32_t srcCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD * Intf::k0;272+ uint32_t srcCiStride = self_->ctx.convTilingData->kernelHxkernelWxkernelD * Intf::k0;
273 uint32_t srcCoStride = coPerReg * self_->ctx.kUbSize;273 uint32_t srcCoStride = coPerReg * self_->ctx.kUbSize;
274- uint32_t dstCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;274+ uint32_t dstCiStride = self_->ctx.convTilingData->kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;
275 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;275 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;
276 uint32_t dstCoStride = coPerReg * Intf::k0;276 uint32_t dstCoStride = coPerReg * Intf::k0;
277- uint32_t srcKdStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;277+ uint32_t srcKdStride = self_->ctx.convTilingData->kernelHxkernelW;
278 uint32_t dstKdStride = self_->ctx.ci1Opt * dstCiStride;278 uint32_t dstKdStride = self_->ctx.ci1Opt * dstCiStride;
279 279 
280 __VEC_SCOPE__280 __VEC_SCOPE__
@@ -324,19 +324,19 @@ private:
324 324 
325 __aicore__ inline void TransNDHWC2NZ()325 __aicore__ inline void TransNDHWC2NZ()
326 {326 {
327- uint16_t kdLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelD;327+ uint16_t kdLoopTimes = self_->ctx.convTilingData->kernelD;
328 uint16_t ciLoopTimes = self_->ctx.ci1Opt;328 uint16_t ciLoopTimes = self_->ctx.ci1Opt;
329 uint16_t coLoopTimes = coOptLoopTimes;329 uint16_t coLoopTimes = coOptLoopTimes;
330- uint16_t khkwLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;330+ uint16_t khkwLoopTimes = self_->ctx.convTilingData->kernelHxkernelW;
331 uint32_t srcGroupOptSize = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;331 uint32_t srcGroupOptSize = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;
332 uint32_t srcCiStride = self_->ctx.coOptAlign * Intf::k0;332 uint32_t srcCiStride = self_->ctx.coOptAlign * Intf::k0;
333 uint32_t srcCoStride = coPerReg;333 uint32_t srcCoStride = coPerReg;
334- uint32_t dstCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;334+ uint32_t dstCiStride = self_->ctx.convTilingData->kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;
335 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;335 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;
336 uint32_t srcKhKwStride = srcGroupOptSize;336 uint32_t srcKhKwStride = srcGroupOptSize;
337 uint32_t dstCoStride = coPerReg * Intf::k0;337 uint32_t dstCoStride = coPerReg * Intf::k0;
338- uint32_t srcKdStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * srcGroupOptSize;338+ uint32_t srcKdStride = self_->ctx.convTilingData->kernelHxkernelW * srcGroupOptSize;
339- uint32_t dstKdStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * self_->ctx.coOptAlign *339+ uint32_t dstKdStride = self_->ctx.convTilingData->kernelHxkernelW * self_->ctx.coOptAlign *
340 self_->ctx.ci1Opt * Intf::k0; // ci1Opt has updated in groupOptTail340 self_->ctx.ci1Opt * Intf::k0; // ci1Opt has updated in groupOptTail
341 __VEC_SCOPE__341 __VEC_SCOPE__
342 {342 {
@@ -386,12 +386,12 @@ private:
386 __aicore__ inline void TransNHWC2NZ()386 __aicore__ inline void TransNHWC2NZ()
387 {387 {
388 uint16_t ciLoopTimes = self_->ctx.ci1Opt;388 uint16_t ciLoopTimes = self_->ctx.ci1Opt;
389- uint16_t khkwLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;389+ uint16_t khkwLoopTimes = self_->ctx.convTilingData->kernelHxkernelW;
390 uint16_t coLoopTimes = coOptLoopTimes;390 uint16_t coLoopTimes = coOptLoopTimes;
391 uint32_t srcCiStride = self_->ctx.coOptAlign * Intf::k0;391 uint32_t srcCiStride = self_->ctx.coOptAlign * Intf::k0;
392 uint32_t srcKhKwStride = self_->ctx.coOptAlign * self_->ctx.ciOptAlign; 392 uint32_t srcKhKwStride = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;
393 uint32_t srcCoStride = coPerReg;393 uint32_t srcCoStride = coPerReg;
394- uint32_t dstCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;394+ uint32_t dstCiStride = self_->ctx.convTilingData->kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;
395 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;395 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;
396 uint32_t dstCoStride = coPerReg * Intf::k0;396 uint32_t dstCoStride = coPerReg * Intf::k0;
397 397
@@ -485,37 +485,37 @@ private:
485 __aicore__ inline void SetCopyParams2D()485 __aicore__ inline void SetCopyParams2D()
486 {486 {
487 if (unlikely(self_->ctx.groupOptIter == self_->ctx.vecId)) {487 if (unlikely(self_->ctx.groupOptIter == self_->ctx.vecId)) {
488- copyParams.blockCount = self_->ctx.convTilingData->convApiTiling.kBL1 / Intf::k0;488+ copyParams.blockCount = self_->ctx.convTilingData->kBL1 / Intf::k0;
489- copyParams.blockLen = self_->ctx.convTilingData->convApiTiling.nBL1;489+ copyParams.blockLen = self_->ctx.convTilingData->nBL1;
490- copyParams.srcStride = self_->ctx.co1Opt * BLOCK_L0_N - self_->ctx.convTilingData->convApiTiling.nBL1;490+ copyParams.srcStride = self_->ctx.co1Opt * BLOCK_L0_N - self_->ctx.convTilingData->nBL1;
491 }491 }
492 492 
493 if constexpr (Intf::isKL1NL0FullLoad) {493 if constexpr (Intf::isKL1NL0FullLoad) {
494 srcOffset = self_->ctx.coStartPos * Intf::k0 +494 srcOffset = self_->ctx.coStartPos * Intf::k0 +
495- self_->ctx.kBL1Iter * self_->ctx.convTilingData->convApiTiling.kBL1 * self_->ctx.coOptAlign;495+ self_->ctx.kBL1Iter * self_->ctx.convTilingData->kBL1 * self_->ctx.coOptAlign;
496 } else {496 } else {
497- srcOffset = (self_->ctx.coStartPos + self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1) * Intf::k0 +497+ srcOffset = (self_->ctx.coStartPos + self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1) * Intf::k0 +
498- self_->ctx.kBL1Iter * self_->ctx.convTilingData->convApiTiling.kBL1 * self_->ctx.coOptAlign;498+ self_->ctx.kBL1Iter * self_->ctx.convTilingData->kBL1 * self_->ctx.coOptAlign;
499 }499 }
500 }500 }
501 501 
502 __aicore__ inline void SetCopyParams3D()502 __aicore__ inline void SetCopyParams3D()
503 {503 {
504 if (unlikely(self_->ctx.loadUB2L1Iter == 0)) {504 if (unlikely(self_->ctx.loadUB2L1Iter == 0)) {
505- copyParams.blockLen = self_->ctx.convTilingData->convApiTiling.nBL1;505+ copyParams.blockLen = self_->ctx.convTilingData->nBL1;
506- copyParams.srcStride = self_->ctx.co1Opt * BLOCK_L0_N - self_->ctx.convTilingData->convApiTiling.nBL1;506+ copyParams.srcStride = self_->ctx.co1Opt * BLOCK_L0_N - self_->ctx.convTilingData->nBL1;
507- kOffset = self_->ctx.bL1Dk * self_->ctx.bL1Cin * self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;507+ kOffset = self_->ctx.bL1Dk * self_->ctx.bL1Cin * self_->ctx.convTilingData->kernelHxkernelW;
508 }508 }
509 509 
510 uint64_t currentBL1Dk = IsKBL1Tail() ? self_->ctx.bL1DkTail : self_->ctx.bL1Dk;510 uint64_t currentBL1Dk = IsKBL1Tail() ? self_->ctx.bL1DkTail : self_->ctx.bL1Dk;
511 uint64_t currentBL1Cin1 = IsKBL1Tail() ? self_->ctx.bL1CinTail : self_->ctx.bL1Cin;511 uint64_t currentBL1Cin1 = IsKBL1Tail() ? self_->ctx.bL1CinTail : self_->ctx.bL1Cin;
512 copyParams.blockCount = currentBL1Dk * CeilDiv(currentBL1Cin1, Intf::k0) *512 copyParams.blockCount = currentBL1Dk * CeilDiv(currentBL1Cin1, Intf::k0) *
513- self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;513+ self_->ctx.convTilingData->kernelHxkernelW;
514 514 
515 if constexpr (Intf::isKL1NL0FullLoad) {515 if constexpr (Intf::isKL1NL0FullLoad) {
516 srcOffset = self_->ctx.coStartPos * Intf::k0 + self_->ctx.kBL1Iter * kOffset * self_->ctx.coOptAlign;516 srcOffset = self_->ctx.coStartPos * Intf::k0 + self_->ctx.kBL1Iter * kOffset * self_->ctx.coOptAlign;
517 } else {517 } else {
518- srcOffset = (self_->ctx.coStartPos + self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1) * Intf::k0 +518+ srcOffset = (self_->ctx.coStartPos + self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1) * Intf::k0 +
519 self_->ctx.kBL1Iter * kOffset * self_->ctx.coOptAlign;519 self_->ctx.kBL1Iter * kOffset * self_->ctx.coOptAlign;
520 }520 }
521 }521 }
Mconv/common/op_kernel/arch35/conv_iterate_base_impl.h+33-33
@@ -61,10 +61,10 @@ template <class Intf>
61static __aicore__ inline void InitBatchDirectionValue(Intf *self)61static __aicore__ inline void InitBatchDirectionValue(Intf *self)
62{62{
63 if constexpr (Intf::isInnerBatchFlag) {63 if constexpr (Intf::isInnerBatchFlag) {
64- self->ctx.ddr2l1LoopBatch = CeilDiv(self->ctx.singleCoreBatch, self->ctx.convTilingData->convApiTiling.innerBatch);64+ self->ctx.ddr2l1LoopBatch = CeilDiv(self->ctx.singleCoreBatch, self->ctx.convTilingData->innerBatch);
65- self->ctx.innerBatchTail = self->ctx.singleCoreBatch % self->ctx.convTilingData->convApiTiling.innerBatch;65+ self->ctx.innerBatchTail = self->ctx.singleCoreBatch % self->ctx.convTilingData->innerBatch;
66 self->ctx.innerBatchTail = self->ctx.innerBatchTail == 0 ?66 self->ctx.innerBatchTail = self->ctx.innerBatchTail == 0 ?
67- self->ctx.convTilingData->convApiTiling.innerBatch : self->ctx.innerBatchTail;67+ self->ctx.convTilingData->innerBatch : self->ctx.innerBatchTail;
68 } else {68 } else {
69 self->ctx.ddr2l1LoopBatch = self->ctx.singleCoreBatch;69 self->ctx.ddr2l1LoopBatch = self->ctx.singleCoreBatch;
70 }70 }
@@ -73,16 +73,16 @@ static __aicore__ inline void InitBatchDirectionValue(Intf *self)
73template <class Intf>73template <class Intf>
74static __aicore__ inline void InitCoDirectionValue(Intf *self)74static __aicore__ inline void InitCoDirectionValue(Intf *self)
75{75{
76- self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->convApiTiling.nBL1;76+ self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->nBL1;
77- self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.nBL1 : self->ctx.nBL1Tail;77+ self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->nBL1 : self->ctx.nBL1Tail;
78- self->ctx.nL0Tail = self->ctx.nBL1Tail % self->ctx.convTilingData->convApiTiling.nL0;78+ self->ctx.nL0Tail = self->ctx.nBL1Tail % self->ctx.convTilingData->nL0;
79- self->ctx.nL0Tail = self->ctx.nL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.nL0 : self->ctx.nL0Tail;79+ self->ctx.nL0Tail = self->ctx.nL0Tail == 0 ? self->ctx.convTilingData->nL0 : self->ctx.nL0Tail;
80 80 
81 if constexpr (!Intf::hasNL1IterFlag) {81 if constexpr (!Intf::hasNL1IterFlag) {
82 self->ctx.maxNBL1Iter = 0;82 self->ctx.maxNBL1Iter = 0;
83 self->ctx.ddr2l1LoopN = 1;83 self->ctx.ddr2l1LoopN = 1;
84 } else {84 } else {
85- self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->convApiTiling.nBL1);85+ self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->nBL1);
86 self->ctx.maxNBL1Iter = self->ctx.ddr2l1LoopN - 1;86 self->ctx.maxNBL1Iter = self->ctx.ddr2l1LoopN - 1;
87 }87 }
88 88 
@@ -91,7 +91,7 @@ static __aicore__ inline void InitCoDirectionValue(Intf *self)
91 self->ctx.maxNL0Iter = 0;91 self->ctx.maxNL0Iter = 0;
92 } else if constexpr (Intf::groupOptFlag || Intf::weightUbTrans) {92 } else if constexpr (Intf::groupOptFlag || Intf::weightUbTrans) {
93 self->ctx.l12l0LoopN = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?93 self->ctx.l12l0LoopN = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?
94- CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->convApiTiling.nL0) : self->ctx.convTilingData->convApiTiling.multiNBL1;94+ CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->nL0) : self->ctx.convTilingData->multiNBL1;
95 }95 }
96 96 
97 if constexpr (Intf::WEIGHT_NZ_FLAG) {97 if constexpr (Intf::WEIGHT_NZ_FLAG) {
@@ -105,7 +105,7 @@ __aicore__ inline void CalcCoDirectionVar(Intf *self)
105{105{
106 if constexpr (Intf::hasNL0IterFlag) {106 if constexpr (Intf::hasNL0IterFlag) {
107 self->ctx.l12l0LoopN = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?107 self->ctx.l12l0LoopN = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?
108- CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->convApiTiling.nL0) : self->ctx.convTilingData->convApiTiling.multiNBL1;108+ CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->nL0) : self->ctx.convTilingData->multiNBL1;
109 self->ctx.maxNL0Iter = self->ctx.l12l0LoopN - 1;109 self->ctx.maxNL0Iter = self->ctx.l12l0LoopN - 1;
110 }110 }
111}111}
@@ -118,7 +118,7 @@ __aicore__ inline uint64_t CalcCurrentNL0(Intf *self)
118 return self->ctx.nL0Tail;118 return self->ctx.nL0Tail;
119 } else {119 } else {
120 bool isNL0Tail = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter && self->ctx.nL0Iter == self->ctx.maxNL0Iter;120 bool isNL0Tail = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter && self->ctx.nL0Iter == self->ctx.maxNL0Iter;
121- uint64_t currentNL0 = isNL0Tail ? self->ctx.nL0Tail : self->ctx.convTilingData->convApiTiling.nL0;121+ uint64_t currentNL0 = isNL0Tail ? self->ctx.nL0Tail : self->ctx.convTilingData->nL0;
122 self->ctx.currentNL0Align = isNL0Tail ? AlignB(currentNL0, BLOCK_L0_N) : currentNL0;122 self->ctx.currentNL0Align = isNL0Tail ? AlignB(currentNL0, BLOCK_L0_N) : currentNL0;
123 return currentNL0;123 return currentNL0;
124 }124 }
@@ -144,9 +144,9 @@ __aicore__ inline uint64_t CalcCurrentML0HWMode(Intf *self)
144 return currentML0;144 return currentML0;
145 } else {145 } else {
146 uint64_t currentML0 = self->ctx.currentHoL0 * self->ctx.currentWoL0;146 uint64_t currentML0 = self->ctx.currentHoL0 * self->ctx.currentWoL0;
147- if (self->ctx.currentWoL0 == self->ctx.convTilingData->convApiTiling.woL0 &&147+ if (self->ctx.currentWoL0 == self->ctx.convTilingData->woL0 &&
148- self->ctx.currentHoL0 == self->ctx.convTilingData->convApiTiling.hoL0) {148+ self->ctx.currentHoL0 == self->ctx.convTilingData->hoL0) {
149- self->ctx.currentML0Align = self->ctx.convTilingData->convApiTiling.mStep;149+ self->ctx.currentML0Align = self->ctx.convTilingData->mStep;
150 } else {150 } else {
151 self->ctx.currentML0Align = AlignB(currentML0, BLOCK_L0_N);151 self->ctx.currentML0Align = AlignB(currentML0, BLOCK_L0_N);
152 }152 }
@@ -159,7 +159,7 @@ template <class Intf>
159__aicore__ inline void LoadAL1BaseModule(Intf *self)159__aicore__ inline void LoadAL1BaseModule(Intf *self)
160{160{
161 if constexpr (Intf::preFusionFlag) {161 if constexpr (Intf::preFusionFlag) {
162- if ((self->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {162+ if ((self->ctx.convTilingData->pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {
163 self->ctx.al1PingPongFlag ^= 1;163 self->ctx.al1PingPongFlag ^= 1;
164 }164 }
165 self->ctx.al1 = LocalTensor<typename Intf::FmapT>(TPosition::A1, 0, 0);165 self->ctx.al1 = LocalTensor<typename Intf::FmapT>(TPosition::A1, 0, 0);
@@ -216,7 +216,7 @@ __aicore__ inline void LoadBL1BaseModule(Intf *self)
216 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(self->ctx.pingPongFlag * VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);216 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(self->ctx.pingPongFlag * VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);
217 } else {217 } else {
218 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE3_MTE1);218 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE3_MTE1);
219- if (self->ctx.convTilingData->convApiTiling.nBL1 > BLOCK_L0_N) {219+ if (self->ctx.convTilingData->nBL1 > BLOCK_L0_N) {
220 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);220 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);
221 }221 }
222 }222 }
@@ -241,7 +241,7 @@ __aicore__ inline void LoadBL1BaseModule(Intf *self, TempIters& tempIters)
241 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(self->ctx.pingPongFlag * VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);241 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(self->ctx.pingPongFlag * VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);
242 } else {242 } else {
243 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE3_MTE1);243 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE3_MTE1);
244- if (self->ctx.convTilingData->convApiTiling.nBL1 > BLOCK_L0_N) {244+ if (self->ctx.convTilingData->nBL1 > BLOCK_L0_N) {
245 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);245 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);
246 }246 }
247 }247 }
@@ -278,7 +278,7 @@ const uint64_t &KStartPosition, const uint64_t &kStep, const MmadParams &mmadPar
278 event_t eventID = static_cast<event_t>(L0_SYBC_DB_CLOSE);278 event_t eventID = static_cast<event_t>(L0_SYBC_DB_CLOSE);
279 AscendC::WaitFlag<AscendC::HardEvent::M_MTE1>(eventID);279 AscendC::WaitFlag<AscendC::HardEvent::M_MTE1>(eventID);
280 const LocalTensor<typename Intf::FmapT> &al0 = self->ctx.wholeAl0Tensor;280 const LocalTensor<typename Intf::FmapT> &al0 = self->ctx.wholeAl0Tensor;
281- if (self->ctx.convTilingData->convApiTiling.kAL1 != self->ctx.convTilingData->convApiTiling.kL0 || self->ctx.loadAL0Flag) {281+ if (self->ctx.convTilingData->kAL1 != self->ctx.convTilingData->kL0 || self->ctx.loadAL0Flag) {
282 self->ctx.loadAL0Ins.LoadAL0(currentKL0, posK, kIter, al0);282 self->ctx.loadAL0Ins.LoadAL0(currentKL0, posK, kIter, al0);
283 }283 }
284 284 
@@ -289,11 +289,11 @@ const uint64_t &KStartPosition, const uint64_t &kStep, const MmadParams &mmadPar
289 LocalTensor<typename Intf::WeightT> bl0;289 LocalTensor<typename Intf::WeightT> bl0;
290 if constexpr (Intf::isL0BFullLoadable) {290 if constexpr (Intf::isL0BFullLoadable) {
291 bl0 = 291 bl0 =
292- self->ctx.wholeBl0Tensor[kIter * self->ctx.convTilingData->convApiTiling.nL0 * self->ctx.convTilingData->convApiTiling.kL0];292+ self->ctx.wholeBl0Tensor[kIter * self->ctx.convTilingData->nL0 * self->ctx.convTilingData->kL0];
293 } else {293 } else {
294 bl0 = self->ctx.wholeBl0Tensor;294 bl0 = self->ctx.wholeBl0Tensor;
295 }295 }
296- if (self->ctx.convTilingData->convApiTiling.kBL1 != self->ctx.convTilingData->convApiTiling.kL0 || self->ctx.loadBL0Flag) {296+ if (self->ctx.convTilingData->kBL1 != self->ctx.convTilingData->kL0 || self->ctx.loadBL0Flag) {
297 if constexpr (!Intf::isL0BFullLoadable) {297 if constexpr (!Intf::isL0BFullLoadable) {
298 self->ctx.loadBL0Ins.LoadBL0(KStartPosition, kStep, bl0);298 self->ctx.loadBL0Ins.LoadBL0(KStartPosition, kStep, bl0);
299 }299 }
@@ -331,11 +331,11 @@ const uint64_t &posK, const uint64_t &KStartPosition, const uint64_t &kStep, con
331 LocalTensor<typename Intf::WeightT> bl0;331 LocalTensor<typename Intf::WeightT> bl0;
332 if constexpr (Intf::isL0BFullLoadable) {332 if constexpr (Intf::isL0BFullLoadable) {
333 bl0 = 333 bl0 =
334- self->ctx.wholeBl0Tensor[kIter * self->ctx.convTilingData->convApiTiling.nL0 * self->ctx.convTilingData->convApiTiling.kL0];334+ self->ctx.wholeBl0Tensor[kIter * self->ctx.convTilingData->nL0 * self->ctx.convTilingData->kL0];
335 } else {335 } else {
336 bl0 = self->ctx.wholeBl0Tensor;336 bl0 = self->ctx.wholeBl0Tensor;
337 }337 }
338- if (self->ctx.convTilingData->convApiTiling.kBL1 != self->ctx.convTilingData->convApiTiling.kL0 || self->ctx.loadBL0Flag) {338+ if (self->ctx.convTilingData->kBL1 != self->ctx.convTilingData->kL0 || self->ctx.loadBL0Flag) {
339 // BL0 wait MMAD339 // BL0 wait MMAD
340 event_t e = static_cast<event_t>(al0PingPongFlag ^ 1);340 event_t e = static_cast<event_t>(al0PingPongFlag ^ 1);
341 AscendC::WaitFlag<AscendC::HardEvent::M_MTE1>(e);341 AscendC::WaitFlag<AscendC::HardEvent::M_MTE1>(e);
@@ -374,7 +374,7 @@ const uint64_t &currentKL0, const uint64_t &posK, const uint64_t &KStartPosition
374 374
375 if constexpr (Intf::isL0BFullLoadable) {375 if constexpr (Intf::isL0BFullLoadable) {
376 bl0 = 376 bl0 =
377- self->ctx.wholeBl0Tensor[kIter * self->ctx.convTilingData->convApiTiling.nL0 * self->ctx.convTilingData->convApiTiling.kL0];377+ self->ctx.wholeBl0Tensor[kIter * self->ctx.convTilingData->nL0 * self->ctx.convTilingData->kL0];
378 } else {378 } else {
379 bl0 = self->ctx.wholeBl0Tensor[(bl0PingPongFlag) * L0B_HALF_SIZE / Intf::sizeOfWeight];379 bl0 = self->ctx.wholeBl0Tensor[(bl0PingPongFlag) * L0B_HALF_SIZE / Intf::sizeOfWeight];
380 self->ctx.loadBL0Ins.LoadBL0(KStartPosition, kStep, bl0);380 self->ctx.loadBL0Ins.LoadBL0(KStartPosition, kStep, bl0);
@@ -390,7 +390,7 @@ const uint64_t &currentKL0, const uint64_t &posK, const uint64_t &KStartPosition
390 self->ctx.weightUbProcessTools.WeightUbTransSyncSet(self, kIter);390 self->ctx.weightUbProcessTools.WeightUbTransSyncSet(self, kIter);
391 }391 }
392 392 
393- if (self->ctx.convTilingData->convApiTiling.kAL1 != self->ctx.convTilingData->convApiTiling.kL0 || self->ctx.loadAL0Flag) {393+ if (self->ctx.convTilingData->kAL1 != self->ctx.convTilingData->kL0 || self->ctx.loadAL0Flag) {
394 // AL0 wait MMAD394 // AL0 wait MMAD
395 event_t e = static_cast<event_t>((bl0PingPongFlag ^ 1));395 event_t e = static_cast<event_t>((bl0PingPongFlag ^ 1));
396 AscendC::WaitFlag<AscendC::HardEvent::M_MTE1>(e);396 AscendC::WaitFlag<AscendC::HardEvent::M_MTE1>(e);
@@ -423,7 +423,7 @@ const uint64_t &currentKL0, const uint64_t &posK, const uint64_t &KStartPosition
423 423 
424 LocalTensor<typename Intf::WeightT> bl0;424 LocalTensor<typename Intf::WeightT> bl0;
425 if constexpr (Intf::isL0BFullLoadable) {425 if constexpr (Intf::isL0BFullLoadable) {
426- bl0 = self->ctx.wholeBl0Tensor[kIter * self->ctx.convTilingData->convApiTiling.nL0 * self->ctx.convTilingData->convApiTiling.kL0];426+ bl0 = self->ctx.wholeBl0Tensor[kIter * self->ctx.convTilingData->nL0 * self->ctx.convTilingData->kL0];
427 } else {427 } else {
428 bl0 = self->ctx.wholeBl0Tensor[(al0PingPongFlag) * L0B_HALF_SIZE / Intf::sizeOfWeight];428 bl0 = self->ctx.wholeBl0Tensor[(al0PingPongFlag) * L0B_HALF_SIZE / Intf::sizeOfWeight];
429 self->ctx.loadBL0Ins.LoadBL0(KStartPosition, kStep, bl0);429 self->ctx.loadBL0Ins.LoadBL0(KStartPosition, kStep, bl0);
@@ -463,9 +463,9 @@ template <class Intf>
463__aicore__ inline void SetMNBeforeIterateK(Intf *self, MmadParams &mmadParams)463__aicore__ inline void SetMNBeforeIterateK(Intf *self, MmadParams &mmadParams)
464{464{
465 if constexpr (Intf::isInnerBatchFlag) {465 if constexpr (Intf::isInnerBatchFlag) {
466- uint16_t maxBatchIter = CeilDiv(self->ctx.singleCoreBatch, self->ctx.convTilingData->convApiTiling.innerBatch) - 1;466+ uint16_t maxBatchIter = CeilDiv(self->ctx.singleCoreBatch, self->ctx.convTilingData->innerBatch) - 1;
467 self->ctx.innerBatch = self->ctx.batchIter == maxBatchIter ?467 self->ctx.innerBatch = self->ctx.batchIter == maxBatchIter ?
468- self->ctx.innerBatchTail : self->ctx.convTilingData->convApiTiling.innerBatch;468+ self->ctx.innerBatchTail : self->ctx.convTilingData->innerBatch;
469 }469 }
470 470 
471 uint64_t currentML0 = 0;471 uint64_t currentML0 = 0;
@@ -515,8 +515,8 @@ __aicore__ inline void FreeL1Tensor(Intf *self)
515 if constexpr (Intf::groupOptPreloadFlag) {515 if constexpr (Intf::groupOptPreloadFlag) {
516 self->ctx.queueAL1.FreeTensor(self->ctx.al1);516 self->ctx.queueAL1.FreeTensor(self->ctx.al1);
517 } else if constexpr (Intf::isMPreLoad) {517 } else if constexpr (Intf::isMPreLoad) {
518- if ((self->ctx.mL0Iter + 1) % CeilDiv(self->ctx.convTilingData->convApiTiling.hoL1,518+ if ((self->ctx.mL0Iter + 1) % CeilDiv(self->ctx.convTilingData->hoL1,
519- self->ctx.convTilingData->convApiTiling.hoL0) == 0) {519+ self->ctx.convTilingData->hoL0) == 0) {
520 self->ctx.queueAL1.FreeTensor(self->ctx.al1);520 self->ctx.queueAL1.FreeTensor(self->ctx.al1);
521 }521 }
522 } else {522 } else {
@@ -533,7 +533,7 @@ __aicore__ inline void FreeL1Tensor(Intf *self)
533 }533 }
534 534 
535 if (self->ctx.enableBias) {535 if (self->ctx.enableBias) {
536- if (!self->ctx.convTilingData->convApiTiling.biasFullLoadFlag) {536+ if (!self->ctx.convTilingData->biasFullLoadFlag) {
537 self->ctx.queueBiasL1.FreeTensor(self->ctx.biasL1);537 self->ctx.queueBiasL1.FreeTensor(self->ctx.biasL1);
538 }538 }
539 539 
@@ -546,14 +546,14 @@ __aicore__ inline bool CheckReduceOneKNotSupportDBCase(Intf *self)
546{546{
547 // when only L0A pingpong and L0B not pingpong, L0B tensor need full load for template ReduceKOpenL0APingPong547 // when only L0A pingpong and L0B not pingpong, L0B tensor need full load for template ReduceKOpenL0APingPong
548 if constexpr (Intf::l0PingPong == 1) {548 if constexpr (Intf::l0PingPong == 1) {
549- if (self->ctx.convTilingData->convApiTiling.nL0 < self->ctx.singleCoreCo) {549+ if (self->ctx.convTilingData->nL0 < self->ctx.singleCoreCo) {
550 return true;550 return true;
551 }551 }
552 }552 }
553 // when only L0B pingpong and L0A not pingpong, L0A tensor need full load for template ReduceKOpenL0BPingPong553 // when only L0B pingpong and L0A not pingpong, L0A tensor need full load for template ReduceKOpenL0BPingPong
554 if constexpr (Intf::l0PingPong == 2) {554 if constexpr (Intf::l0PingPong == 2) {
555- if (self->ctx.convTilingData->convApiTiling.hoL0 < self->ctx.convTilingData->convApiTiling.singleCoreHo ||555+ if (self->ctx.convTilingData->hoL0 < self->ctx.convTilingData->singleCoreHo ||
556- self->ctx.convTilingData->convApiTiling.woL0 < self->ctx.convTilingData->convApiTiling.singleCoreWo) {556+ self->ctx.convTilingData->woL0 < self->ctx.convTilingData->singleCoreWo) {
557 return true;557 return true;
558 }558 }
559 }559 }
Mconv/common/op_kernel/arch35/conv_iterate_hw_mode_impl.h+42-42
@@ -28,11 +28,11 @@ template <class Intf>
28static __aicore__ inline void InitHoDirectionValue(Intf *self)28static __aicore__ inline void InitHoDirectionValue(Intf *self)
29{29{
30 // Ho方向L1及L0均全载时相关变量赋值30 // Ho方向L1及L0均全载时相关变量赋值
31- self->ctx.hoAL1Tail = self->ctx.singleCoreHo % self->ctx.convTilingData->convApiTiling.hoL1;31+ self->ctx.hoAL1Tail = self->ctx.singleCoreHo % self->ctx.convTilingData->hoL1;
32- self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.hoL1 : self->ctx.hoAL1Tail;32+ self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTilingData->hoL1 : self->ctx.hoAL1Tail;
33 self->ctx.currentHoL1 = self->ctx.hoAL1Tail;33 self->ctx.currentHoL1 = self->ctx.hoAL1Tail;
34- self->ctx.hoL0Tail = self->ctx.currentHoL1 % self->ctx.convTilingData->convApiTiling.hoL0;34+ self->ctx.hoL0Tail = self->ctx.currentHoL1 % self->ctx.convTilingData->hoL0;
35- self->ctx.hoL0Tail = self->ctx.hoL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.hoL0 : self->ctx.hoL0Tail;35+ self->ctx.hoL0Tail = self->ctx.hoL0Tail == 0 ? self->ctx.convTilingData->hoL0 : self->ctx.hoL0Tail;
36 self->ctx.currentHoL0 = self->ctx.hoL0Tail;36 self->ctx.currentHoL0 = self->ctx.hoL0Tail;
37 37 
38 // Ho方向变量计算38 // Ho方向变量计算
@@ -40,7 +40,7 @@ static __aicore__ inline void InitHoDirectionValue(Intf *self)
40 self->ctx.ddr2l1LoopH = 1;40 self->ctx.ddr2l1LoopH = 1;
41 self->ctx.maxHoL1Iter = 0;41 self->ctx.maxHoL1Iter = 0;
42 } else {42 } else {
43- self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.hoL1);43+ self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->hoL1);
44 self->ctx.maxHoL1Iter = self->ctx.ddr2l1LoopH - 1;44 self->ctx.maxHoL1Iter = self->ctx.ddr2l1LoopH - 1;
45 }45 }
46}46}
@@ -49,26 +49,26 @@ template <class Intf>
49static __aicore__ inline void InitWoDirectionValue(Intf *self)49static __aicore__ inline void InitWoDirectionValue(Intf *self)
50{50{
51 // Wo方向L1及L0均全载时相关变量赋值51 // Wo方向L1及L0均全载时相关变量赋值
52- self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->convApiTiling.woL1;52+ self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->woL1;
53- self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL1 : self->ctx.woAL1Tail;53+ self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->woL1 : self->ctx.woAL1Tail;
54 self->ctx.currentWoL1 = self->ctx.woAL1Tail;54 self->ctx.currentWoL1 = self->ctx.woAL1Tail;
55 self->ctx.maxWoL1Iter = 0;55 self->ctx.maxWoL1Iter = 0;
56- self->ctx.woL0Tail = self->ctx.currentWoL1 % self->ctx.convTilingData->convApiTiling.woL0;56+ self->ctx.woL0Tail = self->ctx.currentWoL1 % self->ctx.convTilingData->woL0;
57- self->ctx.woL0Tail = self->ctx.woL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL0 : self->ctx.woL0Tail;57+ self->ctx.woL0Tail = self->ctx.woL0Tail == 0 ? self->ctx.convTilingData->woL0 : self->ctx.woL0Tail;
58 self->ctx.currentWoL0 = self->ctx.woL0Tail;58 self->ctx.currentWoL0 = self->ctx.woL0Tail;
59 self->ctx.maxWoL0Iter = 0;59 self->ctx.maxWoL0Iter = 0;
60 self->ctx.l12l0LoopW = 1;60 self->ctx.l12l0LoopW = 1;
61 61 
62 // Wo方向变量计算62 // Wo方向变量计算
63 if constexpr (Intf::hasWL1IterFlag) {63 if constexpr (Intf::hasWL1IterFlag) {
64- self->ctx.maxWoL1Iter = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL1) - 1;64+ self->ctx.maxWoL1Iter = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL1) - 1;
65 }65 }
66 66 
67 if constexpr (!Intf::isDmaFlag) {67 if constexpr (!Intf::isDmaFlag) {
68- if (Intf::hasWL0IterFlag && self->ctx.convTilingData->convApiTiling.hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&68+ if (Intf::hasWL0IterFlag && self->ctx.convTilingData->hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&
69- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0) > 1) {69+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->woL0) > 1) {
70- self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->convApiTiling.woL0;70+ self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->woL0;
71- self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->convApiTiling.woL0) * self->ctx.convTilingData->convApiTiling.woL0;71+ self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->woL0) * self->ctx.convTilingData->woL0;
72 }72 }
73 73
74 if (self->ctx.woL1SmallTail > 0) {74 if (self->ctx.woL1SmallTail > 0) {
@@ -89,17 +89,17 @@ __aicore__ inline void CalcWoDirectionVar(Intf *self)
89 } else if (self->ctx.woAL1Iter == self->ctx.maxWoL1Iter - 1) {89 } else if (self->ctx.woAL1Iter == self->ctx.maxWoL1Iter - 1) {
90 self->ctx.currentWoL1 = self->ctx.woAL1Tail;90 self->ctx.currentWoL1 = self->ctx.woAL1Tail;
91 } else {91 } else {
92- self->ctx.currentWoL1 = self->ctx.convTilingData->convApiTiling.woL1;92+ self->ctx.currentWoL1 = self->ctx.convTilingData->woL1;
93 }93 }
94 } else {94 } else {
95 self->ctx.currentWoL1 = self->ctx.woAL1Iter == self->ctx.maxWoL1Iter ?95 self->ctx.currentWoL1 = self->ctx.woAL1Iter == self->ctx.maxWoL1Iter ?
96- self->ctx.woAL1Tail : self->ctx.convTilingData->convApiTiling.woL1;96+ self->ctx.woAL1Tail : self->ctx.convTilingData->woL1;
97 }97 }
98 }98 }
99 if constexpr (Intf::hasWL0IterFlag) {99 if constexpr (Intf::hasWL0IterFlag) {
100- self->ctx.woL0Tail = self->ctx.currentWoL1 % self->ctx.convTilingData->convApiTiling.woL0;100+ self->ctx.woL0Tail = self->ctx.currentWoL1 % self->ctx.convTilingData->woL0;
101- self->ctx.woL0Tail = self->ctx.woL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL0 : self->ctx.woL0Tail;101+ self->ctx.woL0Tail = self->ctx.woL0Tail == 0 ? self->ctx.convTilingData->woL0 : self->ctx.woL0Tail;
102- self->ctx.maxWoL0Iter = CeilDiv(self->ctx.currentWoL1, self->ctx.convTilingData->convApiTiling.woL0) - 1;102+ self->ctx.maxWoL0Iter = CeilDiv(self->ctx.currentWoL1, self->ctx.convTilingData->woL0) - 1;
103 self->ctx.l12l0LoopW = self->ctx.maxWoL0Iter + 1;103 self->ctx.l12l0LoopW = self->ctx.maxWoL0Iter + 1;
104 }104 }
105}105}
@@ -109,13 +109,13 @@ __aicore__ inline void CalcHoDirectionVar(Intf *self)
109{109{
110 if constexpr (Intf::hasHL1IterFlag) {110 if constexpr (Intf::hasHL1IterFlag) {
111 self->ctx.currentHoL1 = self->ctx.hoAL1Iter == self->ctx.maxHoL1Iter ?111 self->ctx.currentHoL1 = self->ctx.hoAL1Iter == self->ctx.maxHoL1Iter ?
112- self->ctx.hoAL1Tail : self->ctx.convTilingData->convApiTiling.hoL1;112+ self->ctx.hoAL1Tail : self->ctx.convTilingData->hoL1;
113 }113 }
114 114
115 if constexpr (Intf::hasHL0IterFlag) {115 if constexpr (Intf::hasHL0IterFlag) {
116- self->ctx.hoL0Tail = self->ctx.currentHoL1 % self->ctx.convTilingData->convApiTiling.hoL0;116+ self->ctx.hoL0Tail = self->ctx.currentHoL1 % self->ctx.convTilingData->hoL0;
117- self->ctx.hoL0Tail = self->ctx.hoL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.hoL0 : self->ctx.hoL0Tail;117+ self->ctx.hoL0Tail = self->ctx.hoL0Tail == 0 ? self->ctx.convTilingData->hoL0 : self->ctx.hoL0Tail;
118- self->ctx.maxHoL0Iter = CeilDiv(self->ctx.currentHoL1, self->ctx.convTilingData->convApiTiling.hoL0) - 1;118+ self->ctx.maxHoL0Iter = CeilDiv(self->ctx.currentHoL1, self->ctx.convTilingData->hoL0) - 1;
119 self->ctx.l12l0LoopH = self->ctx.maxHoL0Iter + 1;119 self->ctx.l12l0LoopH = self->ctx.maxHoL0Iter + 1;
120 }120 }
121}121}
@@ -124,36 +124,36 @@ template <class Intf>
124__aicore__ inline void UpdateHoL0WoL0(Intf *self)124__aicore__ inline void UpdateHoL0WoL0(Intf *self)
125{125{
126 self->ctx.currentHoL0 = self->ctx.hoL0Iter == self->ctx.maxHoL0Iter ?126 self->ctx.currentHoL0 = self->ctx.hoL0Iter == self->ctx.maxHoL0Iter ?
127- self->ctx.hoL0Tail : self->ctx.convTilingData->convApiTiling.hoL0;127+ self->ctx.hoL0Tail : self->ctx.convTilingData->hoL0;
128 self->ctx.currentWoL0 = self->ctx.woL0Iter == self->ctx.maxWoL0Iter ?128 self->ctx.currentWoL0 = self->ctx.woL0Iter == self->ctx.maxWoL0Iter ?
129- self->ctx.woL0Tail : self->ctx.convTilingData->convApiTiling.woL0;129+ self->ctx.woL0Tail : self->ctx.convTilingData->woL0;
130}130}
131 131 
132template <class Intf>132template <class Intf>
133__aicore__ inline void CalcGroupOptParamForHWMode(Intf *self)133__aicore__ inline void CalcGroupOptParamForHWMode(Intf *self)
134{134{
135 if (((self->ctx.groupOptIter + 1 == self->ctx.singleGroupOpt - 1 && self->ctx.groupOptIter != 0) ||135 if (((self->ctx.groupOptIter + 1 == self->ctx.singleGroupOpt - 1 && self->ctx.groupOptIter != 0) ||
136- self->ctx.singleGroupOpt == 1) && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {136+ self->ctx.singleGroupOpt == 1) && self->ctx.updateEnlarge != self->ctx.convTilingData->enlarge) {
137 self->ctx.singleGroups = self->ctx.updateEnlarge;137 self->ctx.singleGroups = self->ctx.updateEnlarge;
138- self->ctx.singleGroups = self->ctx.singleGroups == 0 ? self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;138+ self->ctx.singleGroups = self->ctx.singleGroups == 0 ? self->ctx.convTilingData->enlarge : self->ctx.singleGroups;
139 }139 }
140- uint64_t enlargeTail = self->ctx.singleGroups % self->ctx.convTilingData->convApiTiling.enlarge;140+ uint64_t enlargeTail = self->ctx.singleGroups % self->ctx.convTilingData->enlarge;
141- enlargeTail = enlargeTail == 0 ? self->ctx.convTilingData->convApiTiling.enlarge : enlargeTail;141+ enlargeTail = enlargeTail == 0 ? self->ctx.convTilingData->enlarge : enlargeTail;
142- if (enlargeTail != self->ctx.convTilingData->convApiTiling.enlarge) {142+ if (enlargeTail != self->ctx.convTilingData->enlarge) {
143- self->ctx.singleCoreCi = enlargeTail * (self->ctx.convTilingData->convApiTiling.orgCi / self->ctx.convTilingData->convApiTiling.groups);143+ self->ctx.singleCoreCi = enlargeTail * (self->ctx.convTilingData->orgCi / self->ctx.convTilingData->groups);
144 if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {144 if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {
145 self->ctx.singleCoreCo = self->ctx.updateSingleCoOpt;145 self->ctx.singleCoreCo = self->ctx.updateSingleCoOpt;
146 146 
147- uint64_t totalKAlignK0 = AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelW;147+ uint64_t totalKAlignK0 = AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->kernelHxkernelW;
148- self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTilingData->convApiTiling.kL0);148+ self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTilingData->kL0);
149 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;149 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;
150- self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTilingData->convApiTiling.kL0;150+ self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTilingData->kL0;
151 if constexpr (Intf::k0 != Intf::k0FmapTail) {151 if constexpr (Intf::k0 != Intf::k0FmapTail) {
152 self->ctx.kAL0Tail = AlignB(self->ctx.singleCoreCi, Intf::k0FmapTail) *152 self->ctx.kAL0Tail = AlignB(self->ctx.singleCoreCi, Intf::k0FmapTail) *
153- self->ctx.convTilingData->convApiTiling.kernelHxkernelW % self->ctx.convTilingData->convApiTiling.kL0;153+ self->ctx.convTilingData->kernelHxkernelW % self->ctx.convTilingData->kL0;
154- self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.kL0 : self->ctx.kAL0Tail;154+ self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTilingData->kL0 : self->ctx.kAL0Tail;
155 }155 }
156- self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.kL0 : self->ctx.kL0Tail;156+ self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTilingData->kL0 : self->ctx.kL0Tail;
157 157
158 InitCoDirectionValue<Intf>(self);158 InitCoDirectionValue<Intf>(self);
159 }159 }
@@ -199,18 +199,18 @@ __aicore__ inline void FirstIterateImplHWMode(Intf *self)
199 CalcCoDirectionVar<Intf>(self);199 CalcCoDirectionVar<Intf>(self);
200 200 
201 if constexpr (Intf::groupOptPreloadFlag) {201 if constexpr (Intf::groupOptPreloadFlag) {
202- if (self->ctx.singleGroupOpt == 1 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {202+ if (self->ctx.singleGroupOpt == 1 && self->ctx.updateEnlarge != self->ctx.convTilingData->enlarge) {
203 self->ctx.singleGroups = self->ctx.updateEnlarge;203 self->ctx.singleGroups = self->ctx.updateEnlarge;
204 self->ctx.singleGroups = self->ctx.singleGroups == 0 ?204 self->ctx.singleGroups = self->ctx.singleGroups == 0 ?
205- self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;205+ self->ctx.convTilingData->enlarge : self->ctx.singleGroups;
206 CalcGroupOptParamForHWMode<Intf>(self);206 CalcGroupOptParamForHWMode<Intf>(self);
207 }207 }
208 LoadAL1BaseModule<Intf>(self);208 LoadAL1BaseModule<Intf>(self);
209 self->ctx.loadAL1Flag = true;209 self->ctx.loadAL1Flag = true;
210- if (self->ctx.singleGroupOpt == 2 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {210+ if (self->ctx.singleGroupOpt == 2 && self->ctx.updateEnlarge != self->ctx.convTilingData->enlarge) {
211 self->ctx.singleGroups = self->ctx.updateEnlarge;211 self->ctx.singleGroups = self->ctx.updateEnlarge;
212 self->ctx.singleGroups = self->ctx.singleGroups == 0 ?212 self->ctx.singleGroups = self->ctx.singleGroups == 0 ?
213- self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;213+ self->ctx.convTilingData->enlarge : self->ctx.singleGroups;
214 CalcGroupOptParamForHWMode<Intf>(self);214 CalcGroupOptParamForHWMode<Intf>(self);
215 }215 }
216 }216 }
@@ -367,7 +367,7 @@ __aicore__ inline bool IterateMFirstHWMode(Intf *self)
367 if (self->ctx.groupOptIter != self->ctx.singleGroupOpt) {367 if (self->ctx.groupOptIter != self->ctx.singleGroupOpt) {
368 return true;368 return true;
369 } else if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {369 } else if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {
370- if (self->ctx.updateSingleCoOpt == 0 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {370+ if (self->ctx.updateSingleCoOpt == 0 && self->ctx.updateEnlarge != self->ctx.convTilingData->enlarge) {
371 return false;371 return false;
372 }372 }
373 return true;373 return true;
Mconv/common/op_kernel/arch35/conv_iterate_impl.h+27-27
@@ -114,7 +114,7 @@ struct PreFusionProcess {
114 if constexpr (!Intf::preFusionFlag) {114 if constexpr (!Intf::preFusionFlag) {
115 return;115 return;
116 }116 }
117- if ((self->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {117+ if ((self->ctx.convTilingData->pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {
118 if (!self->ctx.al1PingPongFlag) {118 if (!self->ctx.al1PingPongFlag) {
119 eventIdMte3ToMte1 = self->ctx.eventIdMte3ToMte1Ping;119 eventIdMte3ToMte1 = self->ctx.eventIdMte3ToMte1Ping;
120 eventIdMte1ToMte3 = self->ctx.eventIdMte1ToMte3Ping;120 eventIdMte1ToMte3 = self->ctx.eventIdMte1ToMte3Ping;
@@ -122,7 +122,7 @@ struct PreFusionProcess {
122 eventIdMte3ToMte1 = self->ctx.eventIdMte3ToMte1Pong;122 eventIdMte3ToMte1 = self->ctx.eventIdMte3ToMte1Pong;
123 eventIdMte1ToMte3 = self->ctx.eventIdMte1ToMte3Pong;123 eventIdMte1ToMte3 = self->ctx.eventIdMte1ToMte3Pong;
124 }124 }
125- aL1Offset = self->ctx.al1PingPongFlag * self->ctx.convTilingData->convApiTiling.aL1SpaceSize;125+ aL1Offset = self->ctx.al1PingPongFlag * self->ctx.convTilingData->aL1SpaceSize;
126 } else {126 } else {
127 eventIdMte3ToMte1 = self->ctx.eventIdMte3ToMte1Ping;127 eventIdMte3ToMte1 = self->ctx.eventIdMte3ToMte1Ping;
128 eventIdMte1ToMte3 = self->ctx.eventIdMte1ToMte3Ping;128 eventIdMte1ToMte3 = self->ctx.eventIdMte1ToMte3Ping;
@@ -379,7 +379,7 @@ template <class Intf, uint32_t ImplType>
379__aicore__ void Iterate<Intf, ImplType>::ReduceKPreload(Intf *self, MmadParams &mmadParams)379__aicore__ void Iterate<Intf, ImplType>::ReduceKPreload(Intf *self, MmadParams &mmadParams)
380{380{
381 // updateIterByFmapTag is true means fm update; false means weight update381 // updateIterByFmapTag is true means fm update; false means weight update
382- bool updateIterByFmapTag = self->ctx.convTilingData->convApiTiling.kAL1 > self->ctx.convTilingData->convApiTiling.kBL1;382+ bool updateIterByFmapTag = self->ctx.convTilingData->kAL1 > self->ctx.convTilingData->kBL1;
383 if (self->ctx.kAL1fullload && !self->ctx.kBL1fullload) {383 if (self->ctx.kAL1fullload && !self->ctx.kBL1fullload) {
384 updateIterByFmapTag = false;384 updateIterByFmapTag = false;
385 } else if (!self->ctx.kAL1fullload && self->ctx.kBL1fullload) {385 } else if (!self->ctx.kAL1fullload && self->ctx.kBL1fullload) {
@@ -408,10 +408,10 @@ __aicore__ void Iterate<Intf, ImplType>::ReduceKPreload(Intf *self, MmadParams &
408 // state408 // state
409 TempIters tempIters;409 TempIters tempIters;
410 410 
411- uint64_t currentKL0 = self->ctx.convTilingData->convApiTiling.kL0;411+ uint64_t currentKL0 = self->ctx.convTilingData->kL0;
412 uint64_t posK = 0;412 uint64_t posK = 0;
413 uint64_t KStartPosition = 0;413 uint64_t KStartPosition = 0;
414- uint64_t kStep = self->ctx.convTilingData->convApiTiling.kStep;414+ uint64_t kStep = self->ctx.convTilingData->kStep;
415 uint64_t multiKAL1 = self->ctx.multiKAL1;415 uint64_t multiKAL1 = self->ctx.multiKAL1;
416 uint64_t multiKBL1 = self->ctx.multiKBL1;416 uint64_t multiKBL1 = self->ctx.multiKBL1;
417 417 
@@ -470,7 +470,7 @@ template <class Intf, uint32_t ImplType>
470__aicore__ void Iterate<Intf, ImplType>::ReduceKFmapPreload(Intf *self, MmadParams &mmadParams)470__aicore__ void Iterate<Intf, ImplType>::ReduceKFmapPreload(Intf *self, MmadParams &mmadParams)
471{471{
472 // updateIterByFmapTag is true means fm update; false means weight update472 // updateIterByFmapTag is true means fm update; false means weight update
473- bool updateIterByFmapTag = self->ctx.convTilingData->convApiTiling.kAL1 > self->ctx.convTilingData->convApiTiling.kBL1;473+ bool updateIterByFmapTag = self->ctx.convTilingData->kAL1 > self->ctx.convTilingData->kBL1;
474 if (self->ctx.kAL1fullload && !self->ctx.kBL1fullload) {474 if (self->ctx.kAL1fullload && !self->ctx.kBL1fullload) {
475 updateIterByFmapTag = false;475 updateIterByFmapTag = false;
476 } else if (!self->ctx.kAL1fullload && self->ctx.kBL1fullload) {476 } else if (!self->ctx.kAL1fullload && self->ctx.kBL1fullload) {
@@ -488,10 +488,10 @@ __aicore__ void Iterate<Intf, ImplType>::ReduceKFmapPreload(Intf *self, MmadPara
488 // state488 // state
489 TempIters tempIters;489 TempIters tempIters;
490 490 
491- uint64_t currentKL0 = self->ctx.convTilingData->convApiTiling.kL0;491+ uint64_t currentKL0 = self->ctx.convTilingData->kL0;
492 uint64_t posK = 0;492 uint64_t posK = 0;
493 uint64_t KStartPosition = 0;493 uint64_t KStartPosition = 0;
494- uint64_t kStep = self->ctx.convTilingData->convApiTiling.kStep;494+ uint64_t kStep = self->ctx.convTilingData->kStep;
495 uint64_t multiKAL1 = self->ctx.multiKAL1;495 uint64_t multiKAL1 = self->ctx.multiKAL1;
496 uint64_t multiKBL1 = self->ctx.multiKBL1;496 uint64_t multiKBL1 = self->ctx.multiKBL1;
497 497 
@@ -559,11 +559,11 @@ __aicore__ void Iterate<Intf, ImplType>::ReduceKPreloadWithWeightFullloadL0(Intf
559 if (self->ctx.loadAL1Flag) {559 if (self->ctx.loadAL1Flag) {
560 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {560 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {
561 if constexpr (Intf::isNoPad) {561 if constexpr (Intf::isNoPad) {
562- self->ctx.loadAl1Ins.SetLoad3dFMatrixNoPad(self->ctx.convTilingData->convApiTiling.orgWi);562+ self->ctx.loadAl1Ins.SetLoad3dFMatrixNoPad(self->ctx.convTilingData->orgWi);
563 } else {563 } else {
564- self->ctx.loadAl1Ins.SetLoad3dFMatrix(self->ctx.convTilingData->convApiTiling.padLeft,564+ self->ctx.loadAl1Ins.SetLoad3dFMatrix(self->ctx.convTilingData->padLeft,
565- self->ctx.convTilingData->convApiTiling.padRight,565+ self->ctx.convTilingData->padRight,
566- self->ctx.convTilingData->convApiTiling.orgWi);566+ self->ctx.convTilingData->orgWi);
567 }567 }
568 }568 }
569 // If the current m direction iterator is not the last one, load the next fmap block to be processed.569 // If the current m direction iterator is not the last one, load the next fmap block to be processed.
@@ -585,10 +585,10 @@ __aicore__ void Iterate<Intf, ImplType>::ReduceKPreloadWithWeightFullloadL0(Intf
585 }585 }
586 self->ctx.loadAL0Flag = true;586 self->ctx.loadAL0Flag = true;
587 587 
588- uint64_t currentKL0 = self->ctx.convTilingData->convApiTiling.kL0;588+ uint64_t currentKL0 = self->ctx.convTilingData->kL0;
589 uint64_t posK = 0;589 uint64_t posK = 0;
590 uint64_t KStartPosition = 0;590 uint64_t KStartPosition = 0;
591- uint64_t kStep = self->ctx.convTilingData->convApiTiling.kStep;591+ uint64_t kStep = self->ctx.convTilingData->kStep;
592 uint64_t multiKAL1 = self->ctx.multiKAL1;592 uint64_t multiKAL1 = self->ctx.multiKAL1;
593 uint64_t multiKBL1 = self->ctx.multiKBL1;593 uint64_t multiKBL1 = self->ctx.multiKBL1;
594 594 
@@ -636,10 +636,10 @@ template <class Intf, uint32_t ImplType>
636__aicore__ void Iterate<Intf, ImplType>::ReduceK(Intf *self, MmadParams &mmadParams)636__aicore__ void Iterate<Intf, ImplType>::ReduceK(Intf *self, MmadParams &mmadParams)
637{637{
638 // state638 // state
639- uint64_t currentKL0 = self->ctx.convTilingData->convApiTiling.kL0;639+ uint64_t currentKL0 = self->ctx.convTilingData->kL0;
640 uint64_t posK = 0;640 uint64_t posK = 0;
641 uint64_t KStartPosition = 0;641 uint64_t KStartPosition = 0;
642- uint64_t kStep = self->ctx.convTilingData->convApiTiling.kStep;642+ uint64_t kStep = self->ctx.convTilingData->kStep;
643 uint64_t multiKAL1 = self->ctx.multiKAL1;643 uint64_t multiKAL1 = self->ctx.multiKAL1;
644 uint64_t multiKBL1 = self->ctx.multiKBL1;644 uint64_t multiKBL1 = self->ctx.multiKBL1;
645 645 
@@ -832,10 +832,10 @@ __aicore__ void Iterate<Intf, ImplType>::ReduceGroupOptFmapPreload(Intf *self, M
832 }832 }
833 833 
834 // state834 // state
835- uint64_t currentKL0 = self->ctx.convTilingData->convApiTiling.kL0;835+ uint64_t currentKL0 = self->ctx.convTilingData->kL0;
836 uint64_t posK = 0;836 uint64_t posK = 0;
837 uint64_t KStartPosition = 0;837 uint64_t KStartPosition = 0;
838- uint64_t kStep = self->ctx.convTilingData->convApiTiling.kStep;838+ uint64_t kStep = self->ctx.convTilingData->kStep;
839 uint64_t multiKAL1 = self->ctx.multiKAL1;839 uint64_t multiKAL1 = self->ctx.multiKAL1;
840 uint64_t multiKBL1 = self->ctx.multiKBL1;840 uint64_t multiKBL1 = self->ctx.multiKBL1;
841 841
@@ -898,7 +898,7 @@ __aicore__ void Iterate<Intf, ImplType>::IterateK(Intf *self)
898 if constexpr (Intf::isInnerBatchFlag || Intf::isDeQuantFlag) {898 if constexpr (Intf::isInnerBatchFlag || Intf::isDeQuantFlag) {
899 self->ctx.cl0 = self->ctx.queueCL0.template AllocTensor<typename Intf::L0cT>();899 self->ctx.cl0 = self->ctx.queueCL0.template AllocTensor<typename Intf::L0cT>();
900 } else {900 } else {
901- if ((self->ctx.convTilingData->convApiTiling.pBufferFlag & 0x04) >> 2) { // cl0 db901+ if ((self->ctx.convTilingData->pBufferFlag & 0x04) >> 2) { // cl0 db
902 self->ctx.cl0 =902 self->ctx.cl0 =
903 self->ctx.wholeCl0Tensor[(self->ctx.cl0PingPongFlag & 0x1) * L0C_HALF_SIZE / Intf::sizeOfL0c];903 self->ctx.wholeCl0Tensor[(self->ctx.cl0PingPongFlag & 0x1) * L0C_HALF_SIZE / Intf::sizeOfL0c];
904 } else {904 } else {
@@ -937,7 +937,7 @@ template <class Intf, uint32_t ImplType>
937__aicore__ void Iterate<Intf, ImplType>::IterateBiasScale(Intf *self)937__aicore__ void Iterate<Intf, ImplType>::IterateBiasScale(Intf *self)
938{938{
939 if (self->ctx.enableBias) {939 if (self->ctx.enableBias) {
940- if (!self->ctx.convTilingData->convApiTiling.biasFullLoadFlag) {940+ if (!self->ctx.convTilingData->biasFullLoadFlag) {
941 self->ctx.biasL1 = self->ctx.queueBiasL1.template AllocTensor<typename Intf::BiasT>();941 self->ctx.biasL1 = self->ctx.queueBiasL1.template AllocTensor<typename Intf::BiasT>();
942 self->ctx.loadBiasL1Ins.LoadChannelWiseL1(self->ctx.biasL1, self->ctx.biasgm);942 self->ctx.loadBiasL1Ins.LoadChannelWiseL1(self->ctx.biasL1, self->ctx.biasgm);
943 self->ctx.queueBiasL1.EnQue(self->ctx.biasL1);943 self->ctx.queueBiasL1.EnQue(self->ctx.biasL1);
@@ -949,7 +949,7 @@ __aicore__ void Iterate<Intf, ImplType>::IterateBiasScale(Intf *self)
949 self->ctx.biasBT = self->ctx.queueBiasBT.template DeQue<typename Intf::L0cT>();949 self->ctx.biasBT = self->ctx.queueBiasBT.template DeQue<typename Intf::L0cT>();
950 }950 }
951 951 
952- if (!self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {952+ if (!self->ctx.convTilingData->fixpParamsFullLoadFlag) {
953 if (self->ctx.enableVectorQuant || self->ctx.enableVectorRelu) {953 if (self->ctx.enableVectorQuant || self->ctx.enableVectorRelu) {
954 event_t eventId = static_cast<event_t>(self->ctx.pipe.FetchEventID(HardEvent::FIX_MTE2));954 event_t eventId = static_cast<event_t>(self->ctx.pipe.FetchEventID(HardEvent::FIX_MTE2));
955 SetFlag<HardEvent::FIX_MTE2>(eventId);955 SetFlag<HardEvent::FIX_MTE2>(eventId);
@@ -958,11 +958,11 @@ __aicore__ void Iterate<Intf, ImplType>::IterateBiasScale(Intf *self)
958 if (self->ctx.enableVectorQuant) {958 if (self->ctx.enableVectorQuant) {
959 self->ctx.scaleL1 = self->ctx.queueScaleL1.template AllocTensor<typename Intf::ScaleT>();959 self->ctx.scaleL1 = self->ctx.queueScaleL1.template AllocTensor<typename Intf::ScaleT>();
960 if constexpr (Intf::isExtendConv2d) {960 if constexpr (Intf::isExtendConv2d) {
961- if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {961+ if (self->ctx.convTilingData->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
962 self->ctx.loadScaleL1Ins.LoadChannelWiseL1(self->ctx.scaleL1, self->ctx.scalegm);962 self->ctx.loadScaleL1Ins.LoadChannelWiseL1(self->ctx.scaleL1, self->ctx.scalegm);
963 }963 }
964- if (self->ctx.convTilingData->convApiTiling.dualOutput &&964+ if (self->ctx.convTilingData->dualOutput &&
965- self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {965+ self->ctx.convTilingData->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
966 self->ctx.loadScaleL1Ins.LoadChannelWiseL1(self->ctx.scaleL1[self->ctx.scale1L1offset],966 self->ctx.loadScaleL1Ins.LoadChannelWiseL1(self->ctx.scaleL1[self->ctx.scale1L1offset],
967 self->ctx.scale1gm);967 self->ctx.scale1gm);
968 }968 }
@@ -975,11 +975,11 @@ __aicore__ void Iterate<Intf, ImplType>::IterateBiasScale(Intf *self)
975 if constexpr (Intf::isExtendConv2d) {975 if constexpr (Intf::isExtendConv2d) {
976 if (self->ctx.enableVectorRelu) {976 if (self->ctx.enableVectorRelu) {
977 self->ctx.reluWeightL1 = self->ctx.queueReluWeightL1.template AllocTensor<typename Intf::ReluWeightT>();977 self->ctx.reluWeightL1 = self->ctx.queueReluWeightL1.template AllocTensor<typename Intf::ReluWeightT>();
978- if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {978+ if (self->ctx.convTilingData->reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
979 self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1(self->ctx.reluWeightL1, self->ctx.reluWeightGM);979 self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1(self->ctx.reluWeightL1, self->ctx.reluWeightGM);
980 }980 }
981- if (self->ctx.convTilingData->convApiTiling.dualOutput &&981+ if (self->ctx.convTilingData->dualOutput &&
982- self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {982+ self->ctx.convTilingData->reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
983 self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1(self->ctx.reluWeightL1[self->ctx.reluWeight1L1offset],983 self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1(self->ctx.reluWeightL1[self->ctx.reluWeight1L1offset],
984 self->ctx.reluWeight1GM);984 self->ctx.reluWeight1GM);
985 }985 }
Mconv/common/op_kernel/arch35/conv_iterate_m_mode_impl.h+17-17
@@ -51,28 +51,28 @@ template <class Intf>
51__aicore__ inline void CalcGroupOptParamForMMode(Intf *self)51__aicore__ inline void CalcGroupOptParamForMMode(Intf *self)
52{52{
53 if (((self->ctx.groupOptIter + 1 == self->ctx.singleGroupOpt - 1 && self->ctx.groupOptIter != 0) ||53 if (((self->ctx.groupOptIter + 1 == self->ctx.singleGroupOpt - 1 && self->ctx.groupOptIter != 0) ||
54- self->ctx.singleGroupOpt == 1) && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {54+ self->ctx.singleGroupOpt == 1) && self->ctx.updateEnlarge != self->ctx.convTilingData->enlarge) {
55 self->ctx.singleGroups = self->ctx.updateEnlarge;55 self->ctx.singleGroups = self->ctx.updateEnlarge;
56- self->ctx.singleGroups = self->ctx.singleGroups == 0 ? self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;56+ self->ctx.singleGroups = self->ctx.singleGroups == 0 ? self->ctx.convTilingData->enlarge : self->ctx.singleGroups;
57 }57 }
58 58 
59- uint64_t enlargeTail = self->ctx.singleGroups % self->ctx.convTilingData->convApiTiling.enlarge;59+ uint64_t enlargeTail = self->ctx.singleGroups % self->ctx.convTilingData->enlarge;
60- enlargeTail = enlargeTail == 0 ? self->ctx.convTilingData->convApiTiling.enlarge : enlargeTail;60+ enlargeTail = enlargeTail == 0 ? self->ctx.convTilingData->enlarge : enlargeTail;
61- if (enlargeTail != self->ctx.convTilingData->convApiTiling.enlarge) {61+ if (enlargeTail != self->ctx.convTilingData->enlarge) {
62- self->ctx.singleCoreCi = enlargeTail * (self->ctx.convTilingData->convApiTiling.orgCi / self->ctx.convTilingData->convApiTiling.groups);62+ self->ctx.singleCoreCi = enlargeTail * (self->ctx.convTilingData->orgCi / self->ctx.convTilingData->groups);
63 if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {63 if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {
64 self->ctx.singleCoreCo = self->ctx.updateSingleCoOpt;64 self->ctx.singleCoreCo = self->ctx.updateSingleCoOpt;
65 65 
66- uint64_t totalKAlignK0 = AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelW;66+ uint64_t totalKAlignK0 = AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->kernelHxkernelW;
67- self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTilingData->convApiTiling.kL0);67+ self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTilingData->kL0);
68 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;68 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;
69- self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTilingData->convApiTiling.kL0;69+ self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTilingData->kL0;
70 if constexpr (Intf::k0 != Intf::k0FmapTail) {70 if constexpr (Intf::k0 != Intf::k0FmapTail) {
71 self->ctx.kAL0Tail = AlignB(self->ctx.singleCoreCi, Intf::k0FmapTail) *71 self->ctx.kAL0Tail = AlignB(self->ctx.singleCoreCi, Intf::k0FmapTail) *
72- self->ctx.convTilingData->convApiTiling.kernelHxkernelW % self->ctx.convTilingData->convApiTiling.kL0;72+ self->ctx.convTilingData->kernelHxkernelW % self->ctx.convTilingData->kL0;
73- self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.kL0 : self->ctx.kAL0Tail;73+ self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTilingData->kL0 : self->ctx.kAL0Tail;
74 }74 }
75- self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.kL0 : self->ctx.kL0Tail;75+ self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTilingData->kL0 : self->ctx.kL0Tail;
76 76
77 InitCoDirectionValue<Intf>(self);77 InitCoDirectionValue<Intf>(self);
78 }78 }
@@ -114,18 +114,18 @@ __aicore__ inline void FirstIterateImplMMode(Intf *self)
114 CalcCoDirectionVar<Intf>(self);114 CalcCoDirectionVar<Intf>(self);
115 115 
116 if constexpr (Intf::groupOptPreloadFlag) {116 if constexpr (Intf::groupOptPreloadFlag) {
117- if (self->ctx.singleGroupOpt == 1 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {117+ if (self->ctx.singleGroupOpt == 1 && self->ctx.updateEnlarge != self->ctx.convTilingData->enlarge) {
118 self->ctx.singleGroups = self->ctx.updateEnlarge;118 self->ctx.singleGroups = self->ctx.updateEnlarge;
119 self->ctx.singleGroups = self->ctx.singleGroups == 0 ?119 self->ctx.singleGroups = self->ctx.singleGroups == 0 ?
120- self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;120+ self->ctx.convTilingData->enlarge : self->ctx.singleGroups;
121 CalcGroupOptParamForMMode<Intf>(self);121 CalcGroupOptParamForMMode<Intf>(self);
122 }122 }
123 LoadAL1BaseModule<Intf>(self);123 LoadAL1BaseModule<Intf>(self);
124 self->ctx.loadAL1Flag = true;124 self->ctx.loadAL1Flag = true;
125- if (self->ctx.singleGroupOpt == 2 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {125+ if (self->ctx.singleGroupOpt == 2 && self->ctx.updateEnlarge != self->ctx.convTilingData->enlarge) {
126 self->ctx.singleGroups = self->ctx.updateEnlarge;126 self->ctx.singleGroups = self->ctx.updateEnlarge;
127 self->ctx.singleGroups = self->ctx.singleGroups == 0 ?127 self->ctx.singleGroups = self->ctx.singleGroups == 0 ?
128- self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;128+ self->ctx.convTilingData->enlarge : self->ctx.singleGroups;
129 CalcGroupOptParamForMMode<Intf>(self);129 CalcGroupOptParamForMMode<Intf>(self);
130 }130 }
131 } else if constexpr (Intf::isMPreLoad) {131 } else if constexpr (Intf::isMPreLoad) {
@@ -246,7 +246,7 @@ __aicore__ inline bool IterateMFirstMMode(Intf *self)
246 if (self->ctx.groupOptIter < self->ctx.singleGroupOpt - 1) {246 if (self->ctx.groupOptIter < self->ctx.singleGroupOpt - 1) {
247 return true;247 return true;
248 } else if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {248 } else if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {
249- if (self->ctx.updateSingleCoOpt == 0 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {249+ if (self->ctx.updateSingleCoOpt == 0 && self->ctx.updateEnlarge != self->ctx.convTilingData->enlarge) {
250 return false;250 return false;
251 }251 }
252 return true;252 return true;
Mconv/common/op_kernel/arch35/conv_opt_group_impl.h+6-6
@@ -139,7 +139,7 @@ template <class Intf>
139__aicore__ inline void OptGroupUpdateLoopN(Intf *self) {139__aicore__ inline void OptGroupUpdateLoopN(Intf *self) {
140 if constexpr (Intf::hasNL0IterFlag) {140 if constexpr (Intf::hasNL0IterFlag) {
141 self->ctx.l12l0LoopN = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?141 self->ctx.l12l0LoopN = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?
142- CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->convApiTiling.nL0) : self->ctx.convTilingData->convApiTiling.multiNBL1;142+ CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->nL0) : self->ctx.convTilingData->multiNBL1;
143 self->ctx.ddr2l1LoopInner = self->ctx.ddr2l1LoopTmp * self->ctx.l12l0LoopN;143 self->ctx.ddr2l1LoopInner = self->ctx.ddr2l1LoopTmp * self->ctx.l12l0LoopN;
144 }144 }
145}145}
@@ -159,24 +159,24 @@ __aicore__ inline void OptGroupUpdateLoopInner(Intf *self) {
159 } else if (self->ctx.woAL1Iter == self->ctx.maxWoL1Iter) {159 } else if (self->ctx.woAL1Iter == self->ctx.maxWoL1Iter) {
160 self->ctx.currentWoL1 = self->ctx.woL1SmallTail;160 self->ctx.currentWoL1 = self->ctx.woL1SmallTail;
161 } else {161 } else {
162- self->ctx.currentWoL1 = self->ctx.convTilingData->convApiTiling.woL1;162+ self->ctx.currentWoL1 = self->ctx.convTilingData->woL1;
163 }163 }
164 } else {164 } else {
165 self->ctx.currentWoL1 = self->ctx.woAL1Iter == self->ctx.maxWoL1Iter ?165 self->ctx.currentWoL1 = self->ctx.woAL1Iter == self->ctx.maxWoL1Iter ?
166- self->ctx.woAL1Tail : self->ctx.convTilingData->convApiTiling.woL1;166+ self->ctx.woAL1Tail : self->ctx.convTilingData->woL1;
167 }167 }
168 }168 }
169 if constexpr (Intf::hasWL0IterFlag) {169 if constexpr (Intf::hasWL0IterFlag) {
170- self->ctx.l12l0LoopW = CeilDiv(self->ctx.currentWoL1, self->ctx.convTilingData->convApiTiling.woL0);170+ self->ctx.l12l0LoopW = CeilDiv(self->ctx.currentWoL1, self->ctx.convTilingData->woL0);
171 }171 }
172 172 
173 if constexpr (Intf::hasHL1IterFlag) {173 if constexpr (Intf::hasHL1IterFlag) {
174 self->ctx.hoAL1Iter = (self->ctx.outerIter / self->ctx.ddr2l1LoopW) % self->ctx.ddr2l1LoopH;174 self->ctx.hoAL1Iter = (self->ctx.outerIter / self->ctx.ddr2l1LoopW) % self->ctx.ddr2l1LoopH;
175 self->ctx.currentHoL1 = self->ctx.hoAL1Iter == self->ctx.maxHoL1Iter ?175 self->ctx.currentHoL1 = self->ctx.hoAL1Iter == self->ctx.maxHoL1Iter ?
176- self->ctx.hoAL1Tail : self->ctx.convTilingData->convApiTiling.hoL1;176+ self->ctx.hoAL1Tail : self->ctx.convTilingData->hoL1;
177 }177 }
178 if constexpr (Intf::hasHL0IterFlag) {178 if constexpr (Intf::hasHL0IterFlag) {
179- self->ctx.l12l0LoopH = CeilDiv(self->ctx.currentHoL1, self->ctx.convTilingData->convApiTiling.hoL0);179+ self->ctx.l12l0LoopH = CeilDiv(self->ctx.currentHoL1, self->ctx.convTilingData->hoL0);
180 }180 }
181 181 
182 self->ctx.ddr2l1LoopTmp = self->ctx.l12l0LoopW * self->ctx.l12l0LoopH;182 self->ctx.ddr2l1LoopTmp = self->ctx.l12l0LoopW * self->ctx.l12l0LoopH;
Mconv/common/op_kernel/arch35/conv_opt_group_init_impl.h+58-58
@@ -29,19 +29,19 @@ __aicore__ inline void OptGroupCalcBL1LoadTimesHWMode(Intf *self)
29{29{
30 if (!self->ctx.kBL1fullload) {30 if (!self->ctx.kBL1fullload) {
31 self->ctx.ddr2l1LoopKB = self->ctx.maxKBL1Iter + 1;31 self->ctx.ddr2l1LoopKB = self->ctx.maxKBL1Iter + 1;
32- uint64_t ddr2l0LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.hoL0);32+ uint64_t ddr2l0LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->hoL0);
33 uint64_t ddr2l0LoopW = 0;33 uint64_t ddr2l0LoopW = 0;
34 if constexpr (Intf::hasWL0IterFlag) {34 if constexpr (Intf::hasWL0IterFlag) {
35 if (self->ctx.woL1SmallTail > 0) {35 if (self->ctx.woL1SmallTail > 0) {
36 ddr2l0LoopW = (self->ctx.ddr2l1LoopW - W_TAIL_NUM) *36 ddr2l0LoopW = (self->ctx.ddr2l1LoopW - W_TAIL_NUM) *
37- CeilDiv(self->ctx.convTilingData->convApiTiling.woL1, self->ctx.convTilingData->convApiTiling.woL0) +37+ CeilDiv(self->ctx.convTilingData->woL1, self->ctx.convTilingData->woL0) +
38- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0) +38+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->woL0) +
39- CeilDiv(self->ctx.woL1SmallTail, self->ctx.convTilingData->convApiTiling.woL0);39+ CeilDiv(self->ctx.woL1SmallTail, self->ctx.convTilingData->woL0);
40 } else {40 } else {
41- ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL0);41+ ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL0);
42 }42 }
43 } else {43 } else {
44- ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL0);44+ ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL0);
45 }45 }
46 self->ctx.bL1LoadTimes = ddr2l0LoopH * ddr2l0LoopW * self->ctx.ddr2l1LoopN * self->ctx.l12l0LoopN *46 self->ctx.bL1LoadTimes = ddr2l0LoopH * ddr2l0LoopW * self->ctx.ddr2l1LoopN * self->ctx.l12l0LoopN *
47 self->ctx.ddr2l1LoopBatch * self->ctx.ddr2l1LoopKB;47 self->ctx.ddr2l1LoopBatch * self->ctx.ddr2l1LoopKB;
@@ -98,27 +98,27 @@ template <class Intf>
98__aicore__ inline void OptGroupInitIterValueMfirstHWMode(Intf *self)98__aicore__ inline void OptGroupInitIterValueMfirstHWMode(Intf *self)
99{99{
100 if (!self->ctx.kBL1fullload) {100 if (!self->ctx.kBL1fullload) {
101- uint64_t ddr2l0LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.hoL0);101+ uint64_t ddr2l0LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->hoL0);
102 uint64_t ddr2l0LoopW = 0;102 uint64_t ddr2l0LoopW = 0;
103 if constexpr (Intf::hasWL0IterFlag) {103 if constexpr (Intf::hasWL0IterFlag) {
104- self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->convApiTiling.woL1;104+ self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->woL1;
105- self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL1 : self->ctx.woAL1Tail;105+ self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->woL1 : self->ctx.woAL1Tail;
106- if (self->ctx.convTilingData->convApiTiling.hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&106+ if (self->ctx.convTilingData->hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&
107- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0) > 1) {107+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->woL0) > 1) {
108- self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->convApiTiling.woL0;108+ self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->woL0;
109- self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->convApiTiling.woL0) * self->ctx.convTilingData->convApiTiling.woL0;109+ self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->woL0) * self->ctx.convTilingData->woL0;
110 }110 }
111 111 
112- self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL1);112+ self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL1);
113 ddr2l0LoopW = 113 ddr2l0LoopW =
114- (self->ctx.ddr2l1LoopW - 1) * CeilDiv(self->ctx.convTilingData->convApiTiling.woL1, self->ctx.convTilingData->convApiTiling.woL0) +114+ (self->ctx.ddr2l1LoopW - 1) * CeilDiv(self->ctx.convTilingData->woL1, self->ctx.convTilingData->woL0) +
115- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0);115+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->woL0);
116 116 
117 if (self->ctx.woL1SmallTail > 0) {117 if (self->ctx.woL1SmallTail > 0) {
118- ddr2l0LoopW += CeilDiv(self->ctx.woL1SmallTail, self->ctx.convTilingData->convApiTiling.woL0);118+ ddr2l0LoopW += CeilDiv(self->ctx.woL1SmallTail, self->ctx.convTilingData->woL0);
119 }119 }
120 } else {120 } else {
121- ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL0);121+ ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL0);
122 }122 }
123 123 
124 self->ctx.ddr2l1LoopTmp = self->ctx.ddr2l1LoopBatch * ddr2l0LoopH * ddr2l0LoopW;124 self->ctx.ddr2l1LoopTmp = self->ctx.ddr2l1LoopBatch * ddr2l0LoopH * ddr2l0LoopW;
@@ -135,23 +135,23 @@ __aicore__ inline void OptGroupInitIterValueNfirstHWMode(Intf *self)
135 self->ctx.ddr2l1LoopH = 1;135 self->ctx.ddr2l1LoopH = 1;
136 self->ctx.maxHoL1Iter = 0;136 self->ctx.maxHoL1Iter = 0;
137 } else {137 } else {
138- self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.hoL1);138+ self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->hoL1);
139 self->ctx.maxHoL1Iter = self->ctx.ddr2l1LoopH - 1;139 self->ctx.maxHoL1Iter = self->ctx.ddr2l1LoopH - 1;
140 }140 }
141 141 
142 if constexpr (Intf::hasWL1IterFlag) {142 if constexpr (Intf::hasWL1IterFlag) {
143- self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL1);143+ self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL1);
144 } else {144 } else {
145 self->ctx.ddr2l1LoopW = 1;145 self->ctx.ddr2l1LoopW = 1;
146 }146 }
147 147 
148- self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->convApiTiling.woL1;148+ self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->woL1;
149- self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL1 : self->ctx.woAL1Tail;149+ self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->woL1 : self->ctx.woAL1Tail;
150 if constexpr (Intf::hasWL0IterFlag) {150 if constexpr (Intf::hasWL0IterFlag) {
151- if (self->ctx.convTilingData->convApiTiling.hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&151+ if (self->ctx.convTilingData->hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&
152- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0) > 1) {152+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->woL0) > 1) {
153- self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->convApiTiling.woL0;153+ self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->woL0;
154- self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->convApiTiling.woL0) * self->ctx.convTilingData->convApiTiling.woL0;154+ self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->woL0) * self->ctx.convTilingData->woL0;
155 }155 }
156 156 
157 if (self->ctx.woL1SmallTail > 0) {157 if (self->ctx.woL1SmallTail > 0) {
@@ -160,8 +160,8 @@ __aicore__ inline void OptGroupInitIterValueNfirstHWMode(Intf *self)
160 }160 }
161 161 
162 if (!self->ctx.kBL1fullload) {162 if (!self->ctx.kBL1fullload) {
163- self->ctx.hoAL1Tail = self->ctx.singleCoreHo % self->ctx.convTilingData->convApiTiling.hoL1;163+ self->ctx.hoAL1Tail = self->ctx.singleCoreHo % self->ctx.convTilingData->hoL1;
164- self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.hoL1 : self->ctx.hoAL1Tail;164+ self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTilingData->hoL1 : self->ctx.hoAL1Tail;
165 self->ctx.currentHoL1 = self->ctx.hoAL1Tail;165 self->ctx.currentHoL1 = self->ctx.hoAL1Tail;
166 self->ctx.currentWoL1 = self->ctx.woAL1Tail;166 self->ctx.currentWoL1 = self->ctx.woAL1Tail;
167 self->ctx.maxWoL1Iter = self->ctx.ddr2l1LoopW - 1;167 self->ctx.maxWoL1Iter = self->ctx.ddr2l1LoopW - 1;
@@ -225,15 +225,15 @@ template <class Intf>
225__aicore__ inline void OptGroupInitKValue(Intf *self)225__aicore__ inline void OptGroupInitKValue(Intf *self)
226{226{
227 self->ctx.ddr2l1LoopKB =227 self->ctx.ddr2l1LoopKB =
228- CeilDiv(AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD,228+ CeilDiv(AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->kernelHxkernelWxkernelD,
229- self->ctx.convTilingData->convApiTiling.kBL1);229+ self->ctx.convTilingData->kBL1);
230 self->ctx.maxKBL1Iter = self->ctx.ddr2l1LoopKB - 1;230 self->ctx.maxKBL1Iter = self->ctx.ddr2l1LoopKB - 1;
231 self->ctx.kBL1fullload = self->ctx.ddr2l1LoopKB == 1;231 self->ctx.kBL1fullload = self->ctx.ddr2l1LoopKB == 1;
232 232 
233 self->ctx.ci1Opt = CeilDiv(self->ctx.singleCoreCi, Intf::k0);233 self->ctx.ci1Opt = CeilDiv(self->ctx.singleCoreCi, Intf::k0);
234 234 
235 if constexpr (Intf::isConv3D) {235 if constexpr (Intf::isConv3D) {
236- self->ctx.bL1Cin = self->ctx.convTilingData->convApiTiling.cinBInCore / self->ctx.bL1Dk;236+ self->ctx.bL1Cin = self->ctx.convTilingData->cinBInCore / self->ctx.bL1Dk;
237 self->ctx.bL1CinTail = self->ctx.bL1Dk > 1 ? self->ctx.bL1Cin : self->ctx.singleCoreCi % self->ctx.bL1Cin;237 self->ctx.bL1CinTail = self->ctx.bL1Dk > 1 ? self->ctx.bL1Cin : self->ctx.singleCoreCi % self->ctx.bL1Cin;
238 self->ctx.bL1CinTail = self->ctx.bL1CinTail == 0 ? self->ctx.bL1Cin : self->ctx.bL1CinTail;238 self->ctx.bL1CinTail = self->ctx.bL1CinTail == 0 ? self->ctx.bL1Cin : self->ctx.bL1CinTail;
239 self->ctx.bL1CinLoadNum = CeilDiv(self->ctx.singleCoreCi, self->ctx.bL1Cin);239 self->ctx.bL1CinLoadNum = CeilDiv(self->ctx.singleCoreCi, self->ctx.bL1Cin);
@@ -243,28 +243,28 @@ __aicore__ inline void OptGroupInitKValue(Intf *self)
243template <class Intf>243template <class Intf>
244__aicore__ inline void OptGroupInitNValue(Intf *self)244__aicore__ inline void OptGroupInitNValue(Intf *self)
245{245{
246- self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->convApiTiling.nBL1;246+ self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->nBL1;
247- self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.nBL1 : self->ctx.nBL1Tail;247+ self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->nBL1 : self->ctx.nBL1Tail;
248 248 
249 if constexpr (!Intf::hasNL1IterFlag) {249 if constexpr (!Intf::hasNL1IterFlag) {
250 self->ctx.ddr2l1LoopN = 1;250 self->ctx.ddr2l1LoopN = 1;
251 self->ctx.maxNBL1Iter = 0;251 self->ctx.maxNBL1Iter = 0;
252 } else {252 } else {
253- self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->convApiTiling.nBL1);253+ self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->nBL1);
254 self->ctx.maxNBL1Iter = self->ctx.ddr2l1LoopN - 1;254 self->ctx.maxNBL1Iter = self->ctx.ddr2l1LoopN - 1;
255 }255 }
256 256 
257 if constexpr (!Intf::hasNL0IterFlag) {257 if constexpr (!Intf::hasNL0IterFlag) {
258 self->ctx.l12l0LoopN = 1;258 self->ctx.l12l0LoopN = 1;
259 } else {259 } else {
260- self->ctx.l12l0LoopN = self->ctx.convTilingData->convApiTiling.multiNBL1;260+ self->ctx.l12l0LoopN = self->ctx.convTilingData->multiNBL1;
261 }261 }
262}262}
263 263 
264template <class Intf>264template <class Intf>
265__aicore__ inline void OptGroupInitBuf(Intf *self)265__aicore__ inline void OptGroupInitBuf(Intf *self)
266{266{
267- self->ctx.ubBufSize = self->ctx.ci1Opt * self->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD * self->ctx.co1Opt *267+ self->ctx.ubBufSize = self->ctx.ci1Opt * self->ctx.convTilingData->kernelHxkernelWxkernelD * self->ctx.co1Opt *
268 Intf::k0 * BLOCK_L0_N;268 Intf::k0 * BLOCK_L0_N;
269 self->ctx.pipe.InitBuffer(self->ctx.ndUbBuf, self->ctx.ubBufSize * Intf::sizeOfWeight);269 self->ctx.pipe.InitBuffer(self->ctx.ndUbBuf, self->ctx.ubBufSize * Intf::sizeOfWeight);
270 self->ctx.pipe.InitBuffer(self->ctx.nzUbBuf, self->ctx.ubBufSize * Intf::sizeOfWeight);270 self->ctx.pipe.InitBuffer(self->ctx.nzUbBuf, self->ctx.ubBufSize * Intf::sizeOfWeight);
@@ -272,11 +272,11 @@ __aicore__ inline void OptGroupInitBuf(Intf *self)
272 272 
273 self->ctx.nzTensor = self->ctx.nzUbBuf.template Get<typename Intf::WeightT>();273 self->ctx.nzTensor = self->ctx.nzUbBuf.template Get<typename Intf::WeightT>();
274 274 
275- uint32_t aL1SpaceSize = self->ctx.convTilingData->convApiTiling.aL1SpaceSize;275+ uint32_t aL1SpaceSize = self->ctx.convTilingData->aL1SpaceSize;
276- if ((self->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {276+ if ((self->ctx.convTilingData->pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {
277 aL1SpaceSize *= DOUBLE_BUF;277 aL1SpaceSize *= DOUBLE_BUF;
278 }278 }
279- self->ctx.bL1SpaceSize = self->ctx.convTilingData->convApiTiling.nBL1 * self->ctx.convTilingData->convApiTiling.kBL1;279+ self->ctx.bL1SpaceSize = self->ctx.convTilingData->nBL1 * self->ctx.convTilingData->kBL1;
280 280 
281 if constexpr (Intf::bL1DBFlag) {281 if constexpr (Intf::bL1DBFlag) {
282 self->ctx.pipe.InitBuffer(self->ctx.bL1TBuf,282 self->ctx.pipe.InitBuffer(self->ctx.bL1TBuf,
@@ -294,37 +294,37 @@ __aicore__ inline void OptGroupVecInit(Intf *self)
294{294{
295 self->ctx.vecId = GetSubBlockIdx();295 self->ctx.vecId = GetSubBlockIdx();
296 296 
297- self->ctx.singleCoreCi = self->ctx.convTilingData->convApiTiling.singleCoreCi;297+ self->ctx.singleCoreCi = self->ctx.convTilingData->singleCoreCi;
298- self->ctx.singleCoreCo = self->ctx.convTilingData->convApiTiling.singleCoreCo;298+ self->ctx.singleCoreCo = self->ctx.convTilingData->singleCoreCo;
299- self->ctx.singleGroups = self->ctx.convTilingData->convApiTiling.singleCoreGroups;299+ self->ctx.singleGroups = self->ctx.convTilingData->singleCoreGroups;
300- self->ctx.singleGroupOpt = self->ctx.convTilingData->convApiTiling.singleCoreGroupOpt;300+ self->ctx.singleGroupOpt = self->ctx.convTilingData->singleCoreGroupOpt;
301 301 
302- self->ctx.enlarge = self->ctx.convTilingData->convApiTiling.enlarge;302+ self->ctx.enlarge = self->ctx.convTilingData->enlarge;
303- self->ctx.ciPerGroup = self->ctx.convTilingData->convApiTiling.orgCi / self->ctx.convTilingData->convApiTiling.groups;303+ self->ctx.ciPerGroup = self->ctx.convTilingData->orgCi / self->ctx.convTilingData->groups;
304- self->ctx.coPerGroup = self->ctx.convTilingData->convApiTiling.orgCo / self->ctx.convTilingData->convApiTiling.groups;304+ self->ctx.coPerGroup = self->ctx.convTilingData->orgCo / self->ctx.convTilingData->groups;
305- self->ctx.ciOpt = self->ctx.ciPerGroup * self->ctx.convTilingData->convApiTiling.enlarge;305+ self->ctx.ciOpt = self->ctx.ciPerGroup * self->ctx.convTilingData->enlarge;
306 self->ctx.ci1Opt = CeilDiv(self->ctx.ciOpt, Intf::k0);306 self->ctx.ci1Opt = CeilDiv(self->ctx.ciOpt, Intf::k0);
307 self->ctx.ciOptAlign = self->ctx.ci1Opt * Intf::k0;307 self->ctx.ciOptAlign = self->ctx.ci1Opt * Intf::k0;
308- self->ctx.kUbSize = self->ctx.ciOptAlign * self->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;308+ self->ctx.kUbSize = self->ctx.ciOptAlign * self->ctx.convTilingData->kernelHxkernelWxkernelD;
309- self->ctx.coOpt = self->ctx.coPerGroup * self->ctx.convTilingData->convApiTiling.enlarge;309+ self->ctx.coOpt = self->ctx.coPerGroup * self->ctx.convTilingData->enlarge;
310 self->ctx.co1Opt = CeilDiv(self->ctx.coOpt, BLOCK_L0_N);310 self->ctx.co1Opt = CeilDiv(self->ctx.coOpt, BLOCK_L0_N);
311 self->ctx.coOptAlign = self->ctx.co1Opt * BLOCK_L0_N;311 self->ctx.coOptAlign = self->ctx.co1Opt * BLOCK_L0_N;
312 312 
313 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {313 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {
314- self->ctx.singleCoreM = self->ctx.convTilingData->convApiTiling.singleCoreHo;314+ self->ctx.singleCoreM = self->ctx.convTilingData->singleCoreHo;
315- self->ctx.mAL1 = self->ctx.convTilingData->convApiTiling.hoL1;315+ self->ctx.mAL1 = self->ctx.convTilingData->hoL1;
316- self->ctx.mL0 = self->ctx.convTilingData->convApiTiling.hoL0;316+ self->ctx.mL0 = self->ctx.convTilingData->hoL0;
317 } else {317 } else {
318- self->ctx.singleCoreHo = self->ctx.convTilingData->convApiTiling.singleCoreHo;318+ self->ctx.singleCoreHo = self->ctx.convTilingData->singleCoreHo;
319- self->ctx.singleCoreWo = self->ctx.convTilingData->convApiTiling.singleCoreWo;319+ self->ctx.singleCoreWo = self->ctx.convTilingData->singleCoreWo;
320 }320 }
321 321 
322 if constexpr (Intf::isConv3D) {322 if constexpr (Intf::isConv3D) {
323- self->ctx.singleCoreDo = self->ctx.convTilingData->convApiTiling.singleCoreDo;323+ self->ctx.singleCoreDo = self->ctx.convTilingData->singleCoreDo;
324 self->ctx.cin1xcin0 = AlignB(self->ctx.singleCoreCi, Intf::k0);324 self->ctx.cin1xcin0 = AlignB(self->ctx.singleCoreCi, Intf::k0);
325- self->ctx.bL1Dk = self->ctx.convTilingData->convApiTiling.cinBInCore <= self->ctx.cin1xcin0 ?325+ self->ctx.bL1Dk = self->ctx.convTilingData->cinBInCore <= self->ctx.cin1xcin0 ?
326- 1 : self->ctx.convTilingData->convApiTiling.cinBInCore / self->ctx.cin1xcin0;326+ 1 : self->ctx.convTilingData->cinBInCore / self->ctx.cin1xcin0;
327- self->ctx.bL1DkTail = self->ctx.convTilingData->convApiTiling.kernelD % self->ctx.bL1Dk;327+ self->ctx.bL1DkTail = self->ctx.convTilingData->kernelD % self->ctx.bL1Dk;
328 self->ctx.bL1DkTail = self->ctx.bL1DkTail == 0 ? self->ctx.bL1Dk : self->ctx.bL1DkTail;328 self->ctx.bL1DkTail = self->ctx.bL1DkTail == 0 ? self->ctx.bL1Dk : self->ctx.bL1DkTail;
329 }329 }
330 330 
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_api_tiling.cpp+9-8
@@ -14,6 +14,7 @@
14 */14 */
15 15 
16#include "conv2d_v2_api_tiling.h"16#include "conv2d_v2_api_tiling.h"
17+#include "../conv2d_v2_tiling.h"
17#include <cstdint>18#include <cstdint>
18#include <algorithm>19#include <algorithm>
19#include <set>20#include <set>
@@ -31,7 +32,7 @@ using namespace std;
31#define UNUSED __attribute__((unused))32#define UNUSED __attribute__((unused))
32 33 
33namespace conv_tiling {34namespace conv_tiling {
34-int64_t Conv2dTiling::GetTiling(optiling::TConv2DTiling &tiling)35+int64_t Conv2dTiling::GetTiling(optiling::Conv2DTilingData &tiling)
35{36{
36 if (!CheckParams()) {37 if (!CheckParams()) {
37 OP_LOGE(nodeType, "conv2d api tiling check params failed.");38 OP_LOGE(nodeType, "conv2d api tiling check params failed.");
@@ -57,7 +58,7 @@ int64_t Conv2dTiling::GetTiling(optiling::TConv2DTiling &tiling)
57 return 0;58 return 0;
58}59}
59 60 
60-bool Conv2dTiling::GetTiling(Conv2DBasicBlockInfo& conv2DBasicBlockInfo, optiling::TConv2DTiling &tiling)61+bool Conv2dTiling::GetTiling(Conv2DBasicBlockInfo& conv2DBasicBlockInfo, optiling::Conv2DTilingData &tiling)
61{62{
62 if (!CheckTilingAlgorithmType(conv2DBasicBlockInfo, BB_PHASE_2)) {63 if (!CheckTilingAlgorithmType(conv2DBasicBlockInfo, BB_PHASE_2)) {
63 return false;64 return false;
@@ -260,7 +261,7 @@ void Conv2dTiling::GetDmaUbTiling(ConvDmaParams& params)
260 params.khUb = static_cast<uint32_t>(khUbRange[khUbIdx]);261 params.khUb = static_cast<uint32_t>(khUbRange[khUbIdx]);
261}262}
262 263 
263-void Conv2dTiling::SetScalarParams(optiling::TConv2DTiling& tiling)264+void Conv2dTiling::SetScalarParams(optiling::Conv2DTilingData& tiling)
264{265{
265 // calculate the follow params in tiling process, for scalar optimization in kernel266 // calculate the follow params in tiling process, for scalar optimization in kernel
266 uint32_t kernelHxkernelW = tiling.get_kernelH() * tiling.get_kernelW();267 uint32_t kernelHxkernelW = tiling.get_kernelH() * tiling.get_kernelW();
@@ -309,7 +310,7 @@ void Conv2dTiling::SetScalarParams(optiling::TConv2DTiling& tiling)
309 tiling.set_aL1SpaceSize(CalcAL1SpaceSize(tiling));310 tiling.set_aL1SpaceSize(CalcAL1SpaceSize(tiling));
310}311}
311 312 
312-void Conv2dTiling::SetUbTiling(optiling::TConv2DTiling& tiling)313+void Conv2dTiling::SetUbTiling(optiling::Conv2DTilingData& tiling)
313{314{
314 if (isDmaFlag) {315 if (isDmaFlag) {
315 ConvDmaParams params = {l1TilingInfo.hoAL1, l1TilingInfo.woAL1, l1TilingInfo.khL1,316 ConvDmaParams params = {l1TilingInfo.hoAL1, l1TilingInfo.woAL1, l1TilingInfo.khL1,
@@ -350,7 +351,7 @@ void Conv2dTiling::SetUbTiling(optiling::TConv2DTiling& tiling)
350 tiling.set_bUbKStep(0);351 tiling.set_bUbKStep(0);
351}352}
352 353 
353-void Conv2dTiling::SetExtendConv2DParams(optiling::TConv2DTiling& tiling)354+void Conv2dTiling::SetExtendConv2DParams(optiling::Conv2DTilingData& tiling)
354{355{
355 // set extendConv2d fixpipe mode to tilingdata356 // set extendConv2d fixpipe mode to tilingdata
356 tiling.set_quantMode0(shapeInfo.quantMode0);357 tiling.set_quantMode0(shapeInfo.quantMode0);
@@ -362,7 +363,7 @@ void Conv2dTiling::SetExtendConv2DParams(optiling::TConv2DTiling& tiling)
362 tiling.set_dualOutput(shapeInfo.dualOutput);363 tiling.set_dualOutput(shapeInfo.dualOutput);
363}364}
364 365 
365-void Conv2dTiling::SetTilingData(optiling::TConv2DTiling& tiling)366+void Conv2dTiling::SetTilingData(optiling::Conv2DTilingData& tiling)
366{367{
367 if (outputOrder == static_cast<int8_t>(OutputOrder::M)) {368 if (outputOrder == static_cast<int8_t>(OutputOrder::M)) {
368 tiling.set_singleCoreHo(static_cast<uint64_t>(shapeInfo.singleM));369 tiling.set_singleCoreHo(static_cast<uint64_t>(shapeInfo.singleM));
@@ -417,7 +418,7 @@ void Conv2dTiling::SetTilingData(optiling::TConv2DTiling& tiling)
417 SetExtendConv2DParams(tiling);418 SetExtendConv2DParams(tiling);
418}419}
419 420 
420-void Conv2dTiling::SetAttrsTilingData(optiling::TConv2DTiling& tiling)421+void Conv2dTiling::SetAttrsTilingData(optiling::Conv2DTilingData& tiling)
421{422{
422 tiling.set_strideH(static_cast<uint32_t>(attrInfo.strideH));423 tiling.set_strideH(static_cast<uint32_t>(attrInfo.strideH));
423 tiling.set_strideW(static_cast<uint32_t>(attrInfo.strideW));424 tiling.set_strideW(static_cast<uint32_t>(attrInfo.strideW));
@@ -440,7 +441,7 @@ void Conv2dTiling::SetAttrsTilingData(optiling::TConv2DTiling& tiling)
440 tiling.set_roundMode(attrInfo.roundMode);441 tiling.set_roundMode(attrInfo.roundMode);
441}442}
442 443 
443-uint32_t Conv2dTiling::CalcAL1SpaceSize(optiling::TConv2DTiling& tiling)444+uint32_t Conv2dTiling::CalcAL1SpaceSize(optiling::Conv2DTilingData& tiling)
444{445{
445 uint64_t aL1SpaceSize = 0;446 uint64_t aL1SpaceSize = 0;
446 uint64_t fmapSize = DTYPE_SIZE_TAB.at(descInfo.fMapType.dtype);447 uint64_t fmapSize = DTYPE_SIZE_TAB.at(descInfo.fMapType.dtype);
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_api_tiling.h+12-8
@@ -22,6 +22,10 @@
22#include "../../../../common/op_host/op_tiling/arch35/conv_api_tiling_base.h"22#include "../../../../common/op_host/op_tiling/arch35/conv_api_tiling_base.h"
23#include "../../../../common/op_host/op_tiling/arch35/conv_api_tiling_algorithm_base.h"23#include "../../../../common/op_host/op_tiling/arch35/conv_api_tiling_algorithm_base.h"
24 24 
25+namespace optiling {
26+ class Conv2DTilingData;
27+}
28+ 
25namespace conv_tiling {29namespace conv_tiling {
26using optiling::conv_ops_tiling::FixpipeInfo;30using optiling::conv_ops_tiling::FixpipeInfo;
27using optiling::conv_ops_tiling::ConvOriGroupInfo;31using optiling::conv_ops_tiling::ConvOriGroupInfo;
@@ -76,8 +80,8 @@ public:
76 Conv2dTiling() {};80 Conv2dTiling() {};
77 explicit Conv2dTiling(const PlatformInfo& platform) : ConvTilingBase(platform) {};81 explicit Conv2dTiling(const PlatformInfo& platform) : ConvTilingBase(platform) {};
78 ~Conv2dTiling() override {};82 ~Conv2dTiling() override {};
79- int64_t GetTiling(optiling::TConv2DTiling &tiling);83+ int64_t GetTiling(optiling::Conv2DTilingData &tiling);
80- bool GetTiling(Conv2DBasicBlockInfo& conv2DBasicBlockInfo, optiling::TConv2DTiling &tiling);84+ bool GetTiling(Conv2DBasicBlockInfo& conv2DBasicBlockInfo, optiling::Conv2DTilingData &tiling);
81 int64_t Compute() override;85 int64_t Compute() override;
82 86 
83 void SetOrgWeightShape(int64_t orgCo, int64_t orgkH, int64_t orgkW);87 void SetOrgWeightShape(int64_t orgCo, int64_t orgkH, int64_t orgkW);
@@ -113,14 +117,14 @@ public:
113 void GetDmaUbTiling(ConvDmaParams& params);117 void GetDmaUbTiling(ConvDmaParams& params);
114private:118private:
115 std::shared_ptr<ConvTilingAlgorithmBase> algoPtr;119 std::shared_ptr<ConvTilingAlgorithmBase> algoPtr;
116- void SetTilingData(optiling::TConv2DTiling& tiling);120+ void SetTilingData(optiling::Conv2DTilingData& tiling);
117- void SetAttrsTilingData(optiling::TConv2DTiling& tiling);121+ void SetAttrsTilingData(optiling::Conv2DTilingData& tiling);
118- void SetScalarParams(optiling::TConv2DTiling& tiling);122+ void SetScalarParams(optiling::Conv2DTilingData& tiling);
119- void SetUbTiling(optiling::TConv2DTiling& tiling);123+ void SetUbTiling(optiling::Conv2DTilingData& tiling);
120- void SetExtendConv2DParams(optiling::TConv2DTiling& tiling);124+ void SetExtendConv2DParams(optiling::Conv2DTilingData& tiling);
121 uint64_t CalcWeightUBSize(ConvWeightUbTransParams& params, uint64_t ci1Ub, uint64_t co1Ub) const;125 uint64_t CalcWeightUBSize(ConvWeightUbTransParams& params, uint64_t ci1Ub, uint64_t co1Ub) const;
122 uint64_t CalcDmaUBSize(ConvDmaParams& params, uint64_t khUb, uint64_t kwUb) const;126 uint64_t CalcDmaUBSize(ConvDmaParams& params, uint64_t khUb, uint64_t kwUb) const;
123- uint32_t CalcAL1SpaceSize(optiling::TConv2DTiling& tiling);127+ uint32_t CalcAL1SpaceSize(optiling::Conv2DTilingData& tiling);
124 void SetDefaultDdim();128 void SetDefaultDdim();
125 void Infer5hdShape();129 void Infer5hdShape();
126 bool CheckParams();130 bool CheckParams();
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling.cpp+5-5
@@ -440,12 +440,12 @@ ge::graphStatus Conv2dBaseTiling::GetFeatureFlag()
440 440 
441void Conv2dBaseTiling::SetNumBlocksRes()441void Conv2dBaseTiling::SetNumBlocksRes()
442{442{
443- numBlocksRes.batchDim = tilingData_.convRunInfo.get_batchDim();443+ numBlocksRes.batchDim = tilingData_.get_batchDim();
444- numBlocksRes.nDim = tilingData_.convRunInfo.get_nDim();444+ numBlocksRes.nDim = tilingData_.get_nDim();
445- numBlocksRes.hoDim = tilingData_.convRunInfo.get_hoDim();445+ numBlocksRes.hoDim = tilingData_.get_hoDim();
446- numBlocksRes.woDim = tilingData_.convRunInfo.get_woDim();446+ numBlocksRes.woDim = tilingData_.get_woDim();
447 numBlocksRes.mDim = numBlocksRes.hoDim;447 numBlocksRes.mDim = numBlocksRes.hoDim;
448- numBlocksRes.groupDim = tilingData_.convRunInfo.get_groupDim();448+ numBlocksRes.groupDim = tilingData_.get_groupDim();
449}449}
450 450 
451// reset conv2d API's tilingdata451// reset conv2d API's tilingdata
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_cache.cpp+150-158
@@ -63,68 +63,68 @@ bool Conv2dBaseTiling::AddTilingToCache()
63 63 
64void Conv2dBaseTiling::GetCachedTilingDataAux()64void Conv2dBaseTiling::GetCachedTilingDataAux()
65{65{
66- cachedTilingData_.fmapKStride = tilingData_.convApiTiling.get_fmapKStride();66+ cachedTilingData_.fmapKStride = tilingData_.get_fmapKStride();
67- cachedTilingData_.weightKStride = tilingData_.convApiTiling.get_weightKStride();67+ cachedTilingData_.weightKStride = tilingData_.get_weightKStride();
68- cachedTilingData_.cinOffsetBlockInGM = tilingData_.convApiTiling.get_cinOffsetBlockInGM();68+ cachedTilingData_.cinOffsetBlockInGM = tilingData_.get_cinOffsetBlockInGM();
69- cachedTilingData_.coutOffsetBlock = tilingData_.convApiTiling.get_coutOffsetBlock();69+ cachedTilingData_.coutOffsetBlock = tilingData_.get_coutOffsetBlock();
70- cachedTilingData_.nL1DivBlockSize = tilingData_.convApiTiling.get_nL1DivBlockSize();70+ cachedTilingData_.nL1DivBlockSize = tilingData_.get_nL1DivBlockSize();
71- cachedTilingData_.iterateMNOrder = tilingData_.convApiTiling.get_iterateMNOrder();71+ cachedTilingData_.iterateMNOrder = tilingData_.get_iterateMNOrder();
72- cachedTilingData_.biasFullLoadFlag = tilingData_.convApiTiling.get_biasFullLoadFlag();72+ cachedTilingData_.biasFullLoadFlag = tilingData_.get_biasFullLoadFlag();
73- cachedTilingData_.fixpParamsFullLoadFlag = tilingData_.convApiTiling.get_fixpParamsFullLoadFlag();73+ cachedTilingData_.fixpParamsFullLoadFlag = tilingData_.get_fixpParamsFullLoadFlag();
74- cachedTilingData_.hf32Enable = tilingData_.convApiTiling.get_hf32Enable();74+ cachedTilingData_.hf32Enable = tilingData_.get_hf32Enable();
75- cachedTilingData_.hf32TransMode = tilingData_.convApiTiling.get_hf32TransMode();75+ cachedTilingData_.hf32TransMode = tilingData_.get_hf32TransMode();
76- cachedTilingData_.batchDim = tilingData_.convRunInfo.get_batchDim();76+ cachedTilingData_.batchDim = tilingData_.get_batchDim();
77- cachedTilingData_.groupDim = tilingData_.convRunInfo.get_groupDim();77+ cachedTilingData_.groupDim = tilingData_.get_groupDim();
78- cachedTilingData_.nDim = tilingData_.convRunInfo.get_nDim();78+ cachedTilingData_.nDim = tilingData_.get_nDim();
79- cachedTilingData_.hoDim = tilingData_.convRunInfo.get_hoDim();79+ cachedTilingData_.hoDim = tilingData_.get_hoDim();
80- cachedTilingData_.woDim = tilingData_.convRunInfo.get_woDim();80+ cachedTilingData_.woDim = tilingData_.get_woDim();
81- cachedTilingData_.cinOpt = tilingData_.convRunInfo.get_cinOpt();81+ cachedTilingData_.cinOpt = tilingData_.get_cinOpt();
82- cachedTilingData_.coutOpt = tilingData_.convRunInfo.get_coutOpt();82+ cachedTilingData_.coutOpt = tilingData_.get_coutOpt();
83- cachedTilingData_.groupOpt = tilingData_.convRunInfo.get_groupOpt();83+ cachedTilingData_.groupOpt = tilingData_.get_groupOpt();
84 cachedTilingData_.enableC04Flag = flagInfo_.enableC04Flag;84 cachedTilingData_.enableC04Flag = flagInfo_.enableC04Flag;
85 cachedTilingData_.mSplitModeFlag = flagInfo_.mSplitModeFlag;85 cachedTilingData_.mSplitModeFlag = flagInfo_.mSplitModeFlag;
86- cachedTilingData_.unionDataXt = tilingData_.convApiTiling.get_unionDataXt();86+ cachedTilingData_.unionDataXt = tilingData_.get_unionDataXt();
87}87}
88 88 
89void Conv2dBaseTiling::GetCachedTilingData()89void Conv2dBaseTiling::GetCachedTilingData()
90{90{
91- cachedTilingData_.singleCoreBatch = tilingData_.convApiTiling.get_singleCoreBatch();91+ cachedTilingData_.singleCoreBatch = tilingData_.get_singleCoreBatch();
92- cachedTilingData_.singleCoreHo = tilingData_.convApiTiling.get_singleCoreHo();92+ cachedTilingData_.singleCoreHo = tilingData_.get_singleCoreHo();
93- cachedTilingData_.singleCoreWo = tilingData_.convApiTiling.get_singleCoreWo();93+ cachedTilingData_.singleCoreWo = tilingData_.get_singleCoreWo();
94- cachedTilingData_.singleCoreCi = tilingData_.convApiTiling.get_singleCoreCi();94+ cachedTilingData_.singleCoreCi = tilingData_.get_singleCoreCi();
95- cachedTilingData_.singleCoreCo = tilingData_.convApiTiling.get_singleCoreCo();95+ cachedTilingData_.singleCoreCo = tilingData_.get_singleCoreCo();
96- cachedTilingData_.hoL1 = tilingData_.convApiTiling.get_hoL1();96+ cachedTilingData_.hoL1 = tilingData_.get_hoL1();
97- cachedTilingData_.woL1 = tilingData_.convApiTiling.get_woL1();97+ cachedTilingData_.woL1 = tilingData_.get_woL1();
98- cachedTilingData_.kAL1 = tilingData_.convApiTiling.get_kAL1();98+ cachedTilingData_.kAL1 = tilingData_.get_kAL1();
99- cachedTilingData_.kBL1 = tilingData_.convApiTiling.get_kBL1();99+ cachedTilingData_.kBL1 = tilingData_.get_kBL1();
100- cachedTilingData_.khL1 = tilingData_.convApiTiling.get_khL1();100+ cachedTilingData_.khL1 = tilingData_.get_khL1();
101- cachedTilingData_.kwL1 = tilingData_.convApiTiling.get_kwL1();101+ cachedTilingData_.kwL1 = tilingData_.get_kwL1();
102- cachedTilingData_.nBL1 = tilingData_.convApiTiling.get_nBL1();102+ cachedTilingData_.nBL1 = tilingData_.get_nBL1();
103- cachedTilingData_.hoL0 = tilingData_.convApiTiling.get_hoL0();103+ cachedTilingData_.hoL0 = tilingData_.get_hoL0();
104- cachedTilingData_.woL0 = tilingData_.convApiTiling.get_woL0();104+ cachedTilingData_.woL0 = tilingData_.get_woL0();
105- cachedTilingData_.kL0 = tilingData_.convApiTiling.get_kL0();105+ cachedTilingData_.kL0 = tilingData_.get_kL0();
106- cachedTilingData_.nL0 = tilingData_.convApiTiling.get_nL0();106+ cachedTilingData_.nL0 = tilingData_.get_nL0();
107- cachedTilingData_.pBufferFlag = tilingData_.convApiTiling.get_pBufferFlag();107+ cachedTilingData_.pBufferFlag = tilingData_.get_pBufferFlag();
108- cachedTilingData_.enlarge = tilingData_.convApiTiling.get_enlarge();108+ cachedTilingData_.enlarge = tilingData_.get_enlarge();
109- cachedTilingData_.singleCoreGroups = tilingData_.convApiTiling.get_singleCoreGroups();109+ cachedTilingData_.singleCoreGroups = tilingData_.get_singleCoreGroups();
110- cachedTilingData_.singleCoreGroupOpt = tilingData_.convApiTiling.get_singleCoreGroupOpt();110+ cachedTilingData_.singleCoreGroupOpt = tilingData_.get_singleCoreGroupOpt();
111- cachedTilingData_.bUbNStep = tilingData_.convApiTiling.get_bUbNStep();111+ cachedTilingData_.bUbNStep = tilingData_.get_bUbNStep();
112- cachedTilingData_.bUbKStep = tilingData_.convApiTiling.get_bUbKStep();112+ cachedTilingData_.bUbKStep = tilingData_.get_bUbKStep();
113- cachedTilingData_.khUb = tilingData_.convApiTiling.get_khUb();113+ cachedTilingData_.khUb = tilingData_.get_khUb();
114- cachedTilingData_.kwUb = tilingData_.convApiTiling.get_kwUb();114+ cachedTilingData_.kwUb = tilingData_.get_kwUb();
115- cachedTilingData_.orgHixWi = tilingData_.convApiTiling.get_orgHixWi();115+ cachedTilingData_.orgHixWi = tilingData_.get_orgHixWi();
116- cachedTilingData_.kernelHxkernelW = tilingData_.convApiTiling.get_kernelHxkernelW();116+ cachedTilingData_.kernelHxkernelW = tilingData_.get_kernelHxkernelW();
117- cachedTilingData_.kernelHxkernelWxkernelD = tilingData_.convApiTiling.get_kernelHxkernelWxkernelD();117+ cachedTilingData_.kernelHxkernelWxkernelD = tilingData_.get_kernelHxkernelWxkernelD();
118- cachedTilingData_.aL1SpaceSize = tilingData_.convApiTiling.get_aL1SpaceSize();118+ cachedTilingData_.aL1SpaceSize = tilingData_.get_aL1SpaceSize();
119- cachedTilingData_.multiNBL1 = tilingData_.convApiTiling.get_multiNBL1();119+ cachedTilingData_.multiNBL1 = tilingData_.get_multiNBL1();
120- cachedTilingData_.cinAInCore = tilingData_.convApiTiling.get_cinAInCore();120+ cachedTilingData_.cinAInCore = tilingData_.get_cinAInCore();
121- cachedTilingData_.cinATailInCore = tilingData_.convApiTiling.get_cinATailInCore();121+ cachedTilingData_.cinATailInCore = tilingData_.get_cinATailInCore();
122- cachedTilingData_.cinBInCore = tilingData_.convApiTiling.get_cinBInCore();122+ cachedTilingData_.cinBInCore = tilingData_.get_cinBInCore();
123- cachedTilingData_.cinBTailInCore = tilingData_.convApiTiling.get_cinBTailInCore();123+ cachedTilingData_.cinBTailInCore = tilingData_.get_cinBTailInCore();
124- cachedTilingData_.mStep = tilingData_.convApiTiling.get_mStep();124+ cachedTilingData_.mStep = tilingData_.get_mStep();
125- cachedTilingData_.kStep = tilingData_.convApiTiling.get_kStep();125+ cachedTilingData_.kStep = tilingData_.get_kStep();
126- cachedTilingData_.nStep = tilingData_.convApiTiling.get_nStep();126+ cachedTilingData_.nStep = tilingData_.get_nStep();
127- cachedTilingData_.innerBatch = tilingData_.convApiTiling.get_innerBatch();127+ cachedTilingData_.innerBatch = tilingData_.get_innerBatch();
128 GetCachedTilingDataAux();128 GetCachedTilingDataAux();
129}129}
130 130 
@@ -188,116 +188,108 @@ void Conv2dBaseTiling::GetCacheTilingInputArgs()
188 188 
189void Conv2dBaseTiling::TranslateCachedRunInfo()189void Conv2dBaseTiling::TranslateCachedRunInfo()
190{190{
191- tilingData_.convRunInfo.set_batch(cacheInputArgs_.inputShapeN);191+ tilingData_.set_batch(cacheInputArgs_.inputShapeN);
192- tilingData_.convRunInfo.set_hin(cacheInputArgs_.inputShapeH);192+ tilingData_.set_hin(cacheInputArgs_.inputShapeH);
193- tilingData_.convRunInfo.set_win(cacheInputArgs_.inputShapeW);193+ tilingData_.set_win(cacheInputArgs_.inputShapeW);
194- tilingData_.convRunInfo.set_batchDim(cachedTilingData_.batchDim);194+ tilingData_.set_batchDim(cachedTilingData_.batchDim);
195- tilingData_.convRunInfo.set_hoDim(cachedTilingData_.hoDim);195+ tilingData_.set_hoDim(cachedTilingData_.hoDim);
196- tilingData_.convRunInfo.set_woDim(cachedTilingData_.woDim);196+ tilingData_.set_woDim(cachedTilingData_.woDim);
197- tilingData_.convRunInfo.set_nDim(cachedTilingData_.nDim);197+ tilingData_.set_nDim(cachedTilingData_.nDim);
198- tilingData_.convRunInfo.set_cin(shapeInfo_.ci);198+ tilingData_.set_cin(shapeInfo_.ci);
199- tilingData_.convRunInfo.set_cout(cacheInputArgs_.weightShapeN);199+ tilingData_.set_cout(cacheInputArgs_.weightShapeN);
200- tilingData_.convRunInfo.set_kh(cacheInputArgs_.weightShapeH);200+ tilingData_.set_kh(cacheInputArgs_.weightShapeH);
201- tilingData_.convRunInfo.set_kw(cacheInputArgs_.weightShapeW);201+ tilingData_.set_kw(cacheInputArgs_.weightShapeW);
202- tilingData_.convRunInfo.set_hout(cacheInputArgs_.outputShapeH);202+ tilingData_.set_hout(cacheInputArgs_.outputShapeH);
203- tilingData_.convRunInfo.set_wout(cacheInputArgs_.outputShapeW);203+ tilingData_.set_wout(cacheInputArgs_.outputShapeW);
204- tilingData_.convRunInfo.set_strideH(cacheInputArgs_.strideH);204+ tilingData_.set_cinOpt(cachedTilingData_.cinOpt);
205- tilingData_.convRunInfo.set_strideW(cacheInputArgs_.strideW);205+ tilingData_.set_coutOpt(cachedTilingData_.coutOpt);
206- tilingData_.convRunInfo.set_dilationH(cacheInputArgs_.dilationH);206+ tilingData_.set_groupOpt(cachedTilingData_.groupOpt);
207- tilingData_.convRunInfo.set_dilationW(cacheInputArgs_.dilationW);207+ tilingData_.set_enlarge(cachedTilingData_.enlarge);
208- tilingData_.convRunInfo.set_padTop(cacheInputArgs_.padTop);208+ tilingData_.set_groupDim(cachedTilingData_.groupDim);
209- tilingData_.convRunInfo.set_padLeft(cacheInputArgs_.padLeft);
210- tilingData_.convRunInfo.set_hasBias(cacheInputArgs_.biasFlag);
211- tilingData_.convRunInfo.set_groups(cacheInputArgs_.groups);
212- tilingData_.convRunInfo.set_cinOpt(cachedTilingData_.cinOpt);
213- tilingData_.convRunInfo.set_coutOpt(cachedTilingData_.coutOpt);
214- tilingData_.convRunInfo.set_groupOpt(cachedTilingData_.groupOpt);
215- tilingData_.convRunInfo.set_enlarge(cachedTilingData_.enlarge);
216- tilingData_.convRunInfo.set_groupDim(cachedTilingData_.groupDim);
217}209}
218 210 
219void Conv2dBaseTiling::TranslateCachedApiTilingPartTwo()211void Conv2dBaseTiling::TranslateCachedApiTilingPartTwo()
220{212{
221- tilingData_.convApiTiling.set_strideH(cacheInputArgs_.strideH);213+ tilingData_.set_strideH(cacheInputArgs_.strideH);
222- tilingData_.convApiTiling.set_strideW(cacheInputArgs_.strideW);214+ tilingData_.set_strideW(cacheInputArgs_.strideW);
223- tilingData_.convApiTiling.set_dilationH(cacheInputArgs_.dilationH);215+ tilingData_.set_dilationH(cacheInputArgs_.dilationH);
224- tilingData_.convApiTiling.set_dilationW(cacheInputArgs_.dilationW);216+ tilingData_.set_dilationW(cacheInputArgs_.dilationW);
225- tilingData_.convApiTiling.set_padTop(cacheInputArgs_.padTop);217+ tilingData_.set_padTop(cacheInputArgs_.padTop);
226- tilingData_.convApiTiling.set_padBottom(cacheInputArgs_.padBottom);218+ tilingData_.set_padBottom(cacheInputArgs_.padBottom);
227- tilingData_.convApiTiling.set_padLeft(cacheInputArgs_.padLeft);219+ tilingData_.set_padLeft(cacheInputArgs_.padLeft);
228- tilingData_.convApiTiling.set_padRight(cacheInputArgs_.padRight);220+ tilingData_.set_padRight(cacheInputArgs_.padRight);
229- tilingData_.convApiTiling.set_iterateMNOrder(cachedTilingData_.iterateMNOrder);221+ tilingData_.set_iterateMNOrder(cachedTilingData_.iterateMNOrder);
230- tilingData_.convApiTiling.set_biasFullLoadFlag(cachedTilingData_.biasFullLoadFlag);222+ tilingData_.set_biasFullLoadFlag(cachedTilingData_.biasFullLoadFlag);
231- tilingData_.convApiTiling.set_fixpParamsFullLoadFlag(cachedTilingData_.fixpParamsFullLoadFlag);223+ tilingData_.set_fixpParamsFullLoadFlag(cachedTilingData_.fixpParamsFullLoadFlag);
232- tilingData_.convApiTiling.set_hf32Enable(cachedTilingData_.hf32Enable);224+ tilingData_.set_hf32Enable(cachedTilingData_.hf32Enable);
233- tilingData_.convApiTiling.set_hf32TransMode(cachedTilingData_.hf32TransMode);225+ tilingData_.set_hf32TransMode(cachedTilingData_.hf32TransMode);
234- tilingData_.convApiTiling.set_hasBias(cacheInputArgs_.biasFlag);226+ tilingData_.set_hasBias(cacheInputArgs_.biasFlag);
235- tilingData_.convApiTiling.set_hasScale(static_cast<uint8_t>(flagInfo_.quantFlag || flagInfo_.extendConvFlag));227+ tilingData_.set_hasScale(static_cast<uint8_t>(flagInfo_.quantFlag || flagInfo_.extendConvFlag));
236- tilingData_.convApiTiling.set_offsetx(attrInfo_.offsetx);228+ tilingData_.set_offsetx(attrInfo_.offsetx);
237- tilingData_.convApiTiling.set_roundMode(attrInfo_.roundMode);229+ tilingData_.set_roundMode(attrInfo_.roundMode);
238- tilingData_.convApiTiling.set_dualOutput(fixpipeInfo_.dualOutput);230+ tilingData_.set_dualOutput(fixpipeInfo_.dualOutput);
239- tilingData_.convApiTiling.set_quantMode0(fixpipeInfo_.quantMode0);231+ tilingData_.set_quantMode0(fixpipeInfo_.quantMode0);
240- tilingData_.convApiTiling.set_reluMode0(fixpipeInfo_.reluMode0);232+ tilingData_.set_reluMode0(fixpipeInfo_.reluMode0);
241- tilingData_.convApiTiling.set_clipMode0(fixpipeInfo_.clipMode0);233+ tilingData_.set_clipMode0(fixpipeInfo_.clipMode0);
242- tilingData_.convApiTiling.set_quantMode1(fixpipeInfo_.quantMode1);234+ tilingData_.set_quantMode1(fixpipeInfo_.quantMode1);
243- tilingData_.convApiTiling.set_reluMode1(fixpipeInfo_.reluMode1);235+ tilingData_.set_reluMode1(fixpipeInfo_.reluMode1);
244- tilingData_.convApiTiling.set_clipMode1(fixpipeInfo_.clipMode1);236+ tilingData_.set_clipMode1(fixpipeInfo_.clipMode1);
245- tilingData_.convApiTiling.set_innerBatch(cachedTilingData_.innerBatch);237+ tilingData_.set_innerBatch(cachedTilingData_.innerBatch);
246- tilingData_.convApiTiling.set_unionDataXt(cachedTilingData_.unionDataXt);238+ tilingData_.set_unionDataXt(cachedTilingData_.unionDataXt);
247}239}
248 240 
249void Conv2dBaseTiling::TranslateCachedApiTilingPartOne()241void Conv2dBaseTiling::TranslateCachedApiTilingPartOne()
250{242{
251- tilingData_.convApiTiling.set_orgHi(cacheInputArgs_.inputShapeH);243+ tilingData_.set_orgHi(cacheInputArgs_.inputShapeH);
252- tilingData_.convApiTiling.set_orgWi(cacheInputArgs_.inputShapeW);244+ tilingData_.set_orgWi(cacheInputArgs_.inputShapeW);
253- tilingData_.convApiTiling.set_orgHo(cacheInputArgs_.outputShapeH);245+ tilingData_.set_orgHo(cacheInputArgs_.outputShapeH);
254- tilingData_.convApiTiling.set_orgWo(cacheInputArgs_.outputShapeW);246+ tilingData_.set_orgWo(cacheInputArgs_.outputShapeW);
255- tilingData_.convApiTiling.set_singleCoreBatch(cachedTilingData_.singleCoreBatch);247+ tilingData_.set_singleCoreBatch(cachedTilingData_.singleCoreBatch);
256- tilingData_.convApiTiling.set_singleCoreHo(cachedTilingData_.singleCoreHo);248+ tilingData_.set_singleCoreHo(cachedTilingData_.singleCoreHo);
257- tilingData_.convApiTiling.set_singleCoreWo(cachedTilingData_.singleCoreWo);249+ tilingData_.set_singleCoreWo(cachedTilingData_.singleCoreWo);
258- tilingData_.convApiTiling.set_orgCi(shapeInfo_.ci);250+ tilingData_.set_orgCi(shapeInfo_.ci);
259- tilingData_.convApiTiling.set_orgCo(cacheInputArgs_.weightShapeN);251+ tilingData_.set_orgCo(cacheInputArgs_.weightShapeN);
260- tilingData_.convApiTiling.set_singleCoreCi(cachedTilingData_.singleCoreCi);252+ tilingData_.set_singleCoreCi(cachedTilingData_.singleCoreCi);
261- tilingData_.convApiTiling.set_singleCoreCo(cachedTilingData_.singleCoreCo);253+ tilingData_.set_singleCoreCo(cachedTilingData_.singleCoreCo);
262- tilingData_.convApiTiling.set_hoL1(cachedTilingData_.hoL1);254+ tilingData_.set_hoL1(cachedTilingData_.hoL1);
263- tilingData_.convApiTiling.set_woL1(cachedTilingData_.woL1);255+ tilingData_.set_woL1(cachedTilingData_.woL1);
264- tilingData_.convApiTiling.set_kAL1(cachedTilingData_.kAL1);256+ tilingData_.set_kAL1(cachedTilingData_.kAL1);
265- tilingData_.convApiTiling.set_kBL1(cachedTilingData_.kBL1);257+ tilingData_.set_kBL1(cachedTilingData_.kBL1);
266- tilingData_.convApiTiling.set_khL1(cachedTilingData_.khL1);258+ tilingData_.set_khL1(cachedTilingData_.khL1);
267- tilingData_.convApiTiling.set_kwL1(cachedTilingData_.kwL1);259+ tilingData_.set_kwL1(cachedTilingData_.kwL1);
268- tilingData_.convApiTiling.set_nBL1(cachedTilingData_.nBL1);260+ tilingData_.set_nBL1(cachedTilingData_.nBL1);
269- tilingData_.convApiTiling.set_hoL0(cachedTilingData_.hoL0);261+ tilingData_.set_hoL0(cachedTilingData_.hoL0);
270- tilingData_.convApiTiling.set_woL0(cachedTilingData_.woL0);262+ tilingData_.set_woL0(cachedTilingData_.woL0);
271- tilingData_.convApiTiling.set_kL0(cachedTilingData_.kL0);263+ tilingData_.set_kL0(cachedTilingData_.kL0);
272- tilingData_.convApiTiling.set_nL0(cachedTilingData_.nL0);264+ tilingData_.set_nL0(cachedTilingData_.nL0);
273- tilingData_.convApiTiling.set_pBufferFlag(cachedTilingData_.pBufferFlag);265+ tilingData_.set_pBufferFlag(cachedTilingData_.pBufferFlag);
274- tilingData_.convApiTiling.set_groups(cacheInputArgs_.groups);266+ tilingData_.set_groups(cacheInputArgs_.groups);
275- tilingData_.convApiTiling.set_enlarge(cachedTilingData_.enlarge);267+ tilingData_.set_enlarge(cachedTilingData_.enlarge);
276- tilingData_.convApiTiling.set_singleCoreGroups(cachedTilingData_.singleCoreGroups);268+ tilingData_.set_singleCoreGroups(cachedTilingData_.singleCoreGroups);
277- tilingData_.convApiTiling.set_singleCoreGroupOpt(cachedTilingData_.singleCoreGroupOpt);269+ tilingData_.set_singleCoreGroupOpt(cachedTilingData_.singleCoreGroupOpt);
278- tilingData_.convApiTiling.set_bUbNStep(cachedTilingData_.bUbNStep);270+ tilingData_.set_bUbNStep(cachedTilingData_.bUbNStep);
279- tilingData_.convApiTiling.set_bUbKStep(cachedTilingData_.bUbKStep);271+ tilingData_.set_bUbKStep(cachedTilingData_.bUbKStep);
280- tilingData_.convApiTiling.set_khUb(cachedTilingData_.khUb);272+ tilingData_.set_khUb(cachedTilingData_.khUb);
281- tilingData_.convApiTiling.set_kwUb(cachedTilingData_.kwUb);273+ tilingData_.set_kwUb(cachedTilingData_.kwUb);
282- tilingData_.convApiTiling.set_orgHixWi(cachedTilingData_.orgHixWi);274+ tilingData_.set_orgHixWi(cachedTilingData_.orgHixWi);
283- tilingData_.convApiTiling.set_kernelHxkernelW(cachedTilingData_.kernelHxkernelW);275+ tilingData_.set_kernelHxkernelW(cachedTilingData_.kernelHxkernelW);
284- tilingData_.convApiTiling.set_kernelHxkernelWxkernelD(cachedTilingData_.kernelHxkernelWxkernelD);276+ tilingData_.set_kernelHxkernelWxkernelD(cachedTilingData_.kernelHxkernelWxkernelD);
285- tilingData_.convApiTiling.set_aL1SpaceSize(cachedTilingData_.aL1SpaceSize);277+ tilingData_.set_aL1SpaceSize(cachedTilingData_.aL1SpaceSize);
286- tilingData_.convApiTiling.set_multiNBL1(cachedTilingData_.multiNBL1);278+ tilingData_.set_multiNBL1(cachedTilingData_.multiNBL1);
287- tilingData_.convApiTiling.set_cinAInCore(cachedTilingData_.cinAInCore);279+ tilingData_.set_cinAInCore(cachedTilingData_.cinAInCore);
288- tilingData_.convApiTiling.set_cinATailInCore(cachedTilingData_.cinATailInCore);280+ tilingData_.set_cinATailInCore(cachedTilingData_.cinATailInCore);
289- tilingData_.convApiTiling.set_cinBInCore(cachedTilingData_.cinBInCore);281+ tilingData_.set_cinBInCore(cachedTilingData_.cinBInCore);
290- tilingData_.convApiTiling.set_cinBTailInCore(cachedTilingData_.cinBTailInCore);282+ tilingData_.set_cinBTailInCore(cachedTilingData_.cinBTailInCore);
291- tilingData_.convApiTiling.set_mStep(cachedTilingData_.mStep);283+ tilingData_.set_mStep(cachedTilingData_.mStep);
292- tilingData_.convApiTiling.set_kStep(cachedTilingData_.kStep);284+ tilingData_.set_kStep(cachedTilingData_.kStep);
293- tilingData_.convApiTiling.set_nStep(cachedTilingData_.nStep);285+ tilingData_.set_nStep(cachedTilingData_.nStep);
294- tilingData_.convApiTiling.set_fmapKStride(cachedTilingData_.fmapKStride);286+ tilingData_.set_fmapKStride(cachedTilingData_.fmapKStride);
295- tilingData_.convApiTiling.set_weightKStride(cachedTilingData_.weightKStride);287+ tilingData_.set_weightKStride(cachedTilingData_.weightKStride);
296- tilingData_.convApiTiling.set_cinOffsetBlockInGM(cachedTilingData_.cinOffsetBlockInGM);288+ tilingData_.set_cinOffsetBlockInGM(cachedTilingData_.cinOffsetBlockInGM);
297- tilingData_.convApiTiling.set_coutOffsetBlock(cachedTilingData_.coutOffsetBlock);289+ tilingData_.set_coutOffsetBlock(cachedTilingData_.coutOffsetBlock);
298- tilingData_.convApiTiling.set_nL1DivBlockSize(cachedTilingData_.nL1DivBlockSize);290+ tilingData_.set_nL1DivBlockSize(cachedTilingData_.nL1DivBlockSize);
299- tilingData_.convApiTiling.set_kernelH(cacheInputArgs_.weightShapeH);291+ tilingData_.set_kernelH(cacheInputArgs_.weightShapeH);
300- tilingData_.convApiTiling.set_kernelW(cacheInputArgs_.weightShapeW);292+ tilingData_.set_kernelW(cacheInputArgs_.weightShapeW);
301}293}
302 294 
303void Conv2dBaseTiling::TranslateCachedTilingData()295void Conv2dBaseTiling::TranslateCachedTilingData()
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_fast_tiling.cpp+30-30
@@ -251,12 +251,12 @@ ge::graphStatus Conv2dBaseTiling::GetConv2dApiTiling()
251 Conv2dApiTilingSetShape();251 Conv2dApiTilingSetShape();
252 252 
253 if (flagInfo_.mBasicBlockFlag) {253 if (flagInfo_.mBasicBlockFlag) {
254- if (!conv2dApiTiling_.GetTiling(conv2dBasicBlockInfo_, tilingData_.convApiTiling)) {254+ if (!conv2dApiTiling_.GetTiling(conv2dBasicBlockInfo_, tilingData_)) {
255 OP_LOGE(context_->GetNodeName(), "%s AscendC: get api tiling wrong", paramInfo_.nodeType.c_str());255 OP_LOGE(context_->GetNodeName(), "%s AscendC: get api tiling wrong", paramInfo_.nodeType.c_str());
256 return ge::GRAPH_FAILED;256 return ge::GRAPH_FAILED;
257 }257 }
258 } else {258 } else {
259- if (conv2dApiTiling_.GetTiling(tilingData_.convApiTiling) == -1) {259+ if (conv2dApiTiling_.GetTiling(tilingData_) == -1) {
260 OP_LOGE(context_->GetNodeName(), "%s AscendC: get api tiling wrong", paramInfo_.nodeType.c_str());260 OP_LOGE(context_->GetNodeName(), "%s AscendC: get api tiling wrong", paramInfo_.nodeType.c_str());
261 return ge::GRAPH_FAILED;261 return ge::GRAPH_FAILED;
262 }262 }
@@ -334,39 +334,39 @@ void Conv2dBaseTiling::Conv2dOpTilingSetAttr()
334 334 
335ge::graphStatus Conv2dBaseTiling::GetConv2dOpsTiling()335ge::graphStatus Conv2dBaseTiling::GetConv2dOpsTiling()
336{336{
337- tilingData_.convRunInfo.set_hin(static_cast<uint64_t>(shapeInfo_.hi));337+ tilingData_.set_hin(static_cast<uint64_t>(shapeInfo_.hi));
338- tilingData_.convRunInfo.set_win(static_cast<uint64_t>(shapeInfo_.wi));338+ tilingData_.set_win(static_cast<uint64_t>(shapeInfo_.wi));
339- tilingData_.convRunInfo.set_hout(static_cast<uint64_t>(shapeInfo_.ho));339+ tilingData_.set_hout(static_cast<uint64_t>(shapeInfo_.ho));
340- tilingData_.convRunInfo.set_wout(static_cast<uint64_t>(shapeInfo_.wo));340+ tilingData_.set_wout(static_cast<uint64_t>(shapeInfo_.wo));
341- tilingData_.convRunInfo.set_batch(static_cast<uint32_t>(shapeInfo_.batch));341+ tilingData_.set_batch(static_cast<uint32_t>(shapeInfo_.batch));
342- tilingData_.convRunInfo.set_cin(static_cast<uint32_t>(shapeInfo_.ci));342+ tilingData_.set_cin(static_cast<uint32_t>(shapeInfo_.ci));
343- tilingData_.convRunInfo.set_cout(static_cast<uint32_t>(shapeInfo_.co));343+ tilingData_.set_cout(static_cast<uint32_t>(shapeInfo_.co));
344- tilingData_.convRunInfo.set_kh(static_cast<uint32_t>(shapeInfo_.kh));344+ tilingData_.set_kh(static_cast<uint32_t>(shapeInfo_.kh));
345- tilingData_.convRunInfo.set_kw(static_cast<uint32_t>(shapeInfo_.kw));345+ tilingData_.set_kw(static_cast<uint32_t>(shapeInfo_.kw));
346- tilingData_.convRunInfo.set_strideH(static_cast<uint32_t>(attrInfo_.strideH));346+ tilingData_.set_strideH(static_cast<uint32_t>(attrInfo_.strideH));
347- tilingData_.convRunInfo.set_strideW(static_cast<uint32_t>(attrInfo_.strideW));347+ tilingData_.set_strideW(static_cast<uint32_t>(attrInfo_.strideW));
348- tilingData_.convRunInfo.set_dilationH(static_cast<uint32_t>(attrInfo_.dilationH));348+ tilingData_.set_dilationH(static_cast<uint32_t>(attrInfo_.dilationH));
349- tilingData_.convRunInfo.set_dilationW(static_cast<uint32_t>(attrInfo_.dilationW));349+ tilingData_.set_dilationW(static_cast<uint32_t>(attrInfo_.dilationW));
350- tilingData_.convRunInfo.set_padTop(static_cast<uint32_t>(attrInfo_.padTop));350+ tilingData_.set_padTop(static_cast<uint32_t>(attrInfo_.padTop));
351- tilingData_.convRunInfo.set_padLeft(static_cast<uint32_t>(attrInfo_.padLeft));351+ tilingData_.set_padLeft(static_cast<uint32_t>(attrInfo_.padLeft));
352- tilingData_.convRunInfo.set_hasBias(static_cast<uint8_t>(flagInfo_.hasBias));352+ tilingData_.set_hasBias(static_cast<uint8_t>(flagInfo_.hasBias));
353- tilingData_.convRunInfo.set_batchDim(static_cast<uint32_t>(numBlocksRes.batchDim));353+ tilingData_.set_batchDim(static_cast<uint32_t>(numBlocksRes.batchDim));
354- tilingData_.convRunInfo.set_nDim(static_cast<uint32_t>(numBlocksRes.nDim));354+ tilingData_.set_nDim(static_cast<uint32_t>(numBlocksRes.nDim));
355- tilingData_.convRunInfo.set_groupDim(static_cast<uint32_t>(numBlocksRes.groupDim));355+ tilingData_.set_groupDim(static_cast<uint32_t>(numBlocksRes.groupDim));
356- tilingData_.convRunInfo.set_groups(static_cast<uint32_t>(attrInfo_.groups));356+ tilingData_.set_groups(static_cast<uint32_t>(attrInfo_.groups));
357 if (flagInfo_.convGroupType == ConvGroupType::OPT_GROUP_CONV) {357 if (flagInfo_.convGroupType == ConvGroupType::OPT_GROUP_CONV) {
358- tilingData_.convRunInfo.set_cinOpt(static_cast<uint32_t>(optGroupInfo_.cinOpt));358+ tilingData_.set_cinOpt(static_cast<uint32_t>(optGroupInfo_.cinOpt));
359- tilingData_.convRunInfo.set_coutOpt(static_cast<uint32_t>(optGroupInfo_.coutOpt));359+ tilingData_.set_coutOpt(static_cast<uint32_t>(optGroupInfo_.coutOpt));
360- tilingData_.convRunInfo.set_groupOpt(static_cast<uint32_t>(optGroupInfo_.groupOpt));360+ tilingData_.set_groupOpt(static_cast<uint32_t>(optGroupInfo_.groupOpt));
361- tilingData_.convRunInfo.set_enlarge(static_cast<uint32_t>(optGroupInfo_.enlarge));361+ tilingData_.set_enlarge(static_cast<uint32_t>(optGroupInfo_.enlarge));
362 }362 }
363 363 
364 if (flagInfo_.mSplitModeFlag) {364 if (flagInfo_.mSplitModeFlag) {
365- tilingData_.convRunInfo.set_hoDim(static_cast<uint32_t>(numBlocksRes.mDim));365+ tilingData_.set_hoDim(static_cast<uint32_t>(numBlocksRes.mDim));
366- tilingData_.convRunInfo.set_woDim(static_cast<uint32_t>(1));366+ tilingData_.set_woDim(static_cast<uint32_t>(1));
367 } else {367 } else {
368- tilingData_.convRunInfo.set_hoDim(static_cast<uint32_t>(numBlocksRes.hoDim));368+ tilingData_.set_hoDim(static_cast<uint32_t>(numBlocksRes.hoDim));
369- tilingData_.convRunInfo.set_woDim(static_cast<uint32_t>(numBlocksRes.woDim));369+ tilingData_.set_woDim(static_cast<uint32_t>(numBlocksRes.woDim));
370 }370 }
371 return ge::GRAPH_SUCCESS;371 return ge::GRAPH_SUCCESS;
372}372}
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_print.cpp+96-96
@@ -23,26 +23,26 @@ void Conv2dBaseTiling::PrintOpTilingData()
23 do not modify the param name if not necessarily23 do not modify the param name if not necessarily
24 */24 */
25 std::stringstream ss;25 std::stringstream ss;
26- ss << "hin: " << tilingData_.convRunInfo.get_hin() << ", win: " << tilingData_.convRunInfo.get_win()26+ ss << "hin: " << tilingData_.get_hin() << ", win: " << tilingData_.get_win()
27- << ", hout: " << tilingData_.convRunInfo.get_hout() << ", wout: " << tilingData_.convRunInfo.get_wout()27+ << ", hout: " << tilingData_.get_hout() << ", wout: " << tilingData_.get_wout()
28- << ", batch: " << tilingData_.convRunInfo.get_batch() << ", cin: " << tilingData_.convRunInfo.get_cin()28+ << ", batch: " << tilingData_.get_batch() << ", cin: " << tilingData_.get_cin()
29- << ", cout: " << tilingData_.convRunInfo.get_cout() << ", kh: " << tilingData_.convRunInfo.get_kh()29+ << ", cout: " << tilingData_.get_cout() << ", kh: " << tilingData_.get_kh()
30- << ", kw: " << tilingData_.convRunInfo.get_kw() << ", batchDim: " << tilingData_.convRunInfo.get_batchDim()30+ << ", kw: " << tilingData_.get_kw() << ", batchDim: " << tilingData_.get_batchDim()
31- << ", hoDim: " << tilingData_.convRunInfo.get_hoDim() << ", woDim: " << tilingData_.convRunInfo.get_woDim()31+ << ", hoDim: " << tilingData_.get_hoDim() << ", woDim: " << tilingData_.get_woDim()
32- << ", nDim: " << tilingData_.convRunInfo.get_nDim()32+ << ", nDim: " << tilingData_.get_nDim()
33- << ", strideH: " << tilingData_.convRunInfo.get_strideH()33+ << ", strideH: " << tilingData_.get_strideH()
34- << ", strideW: " << tilingData_.convRunInfo.get_strideW()34+ << ", strideW: " << tilingData_.get_strideW()
35- << ", dilationH: " << tilingData_.convRunInfo.get_dilationH()35+ << ", dilationH: " << tilingData_.get_dilationH()
36- << ", dilationW: " << tilingData_.convRunInfo.get_dilationW()36+ << ", dilationW: " << tilingData_.get_dilationW()
37- << ", padTop: " << tilingData_.convRunInfo.get_padTop()37+ << ", padTop: " << tilingData_.get_padTop()
38- << ", padLeft: " << tilingData_.convRunInfo.get_padLeft()38+ << ", padLeft: " << tilingData_.get_padLeft()
39- << ", groups: " << tilingData_.convRunInfo.get_groups()39+ << ", groups: " << tilingData_.get_groups()
40- << ", cinOpt: " << tilingData_.convRunInfo.get_cinOpt()40+ << ", cinOpt: " << tilingData_.get_cinOpt()
41- << ", coutOpt: " << tilingData_.convRunInfo.get_coutOpt()41+ << ", coutOpt: " << tilingData_.get_coutOpt()
42- << ", groupOpt: " << tilingData_.convRunInfo.get_groupOpt()42+ << ", groupOpt: " << tilingData_.get_groupOpt()
43- << ", enlarge: " << tilingData_.convRunInfo.get_enlarge()43+ << ", enlarge: " << tilingData_.get_enlarge()
44- << ", groupDim: " << tilingData_.convRunInfo.get_groupDim()44+ << ", groupDim: " << tilingData_.get_groupDim()
45- << ", hasBias: " << static_cast<uint32_t>(tilingData_.convRunInfo.get_hasBias());45+ << ", hasBias: " << static_cast<uint32_t>(tilingData_.get_hasBias());
46 OP_LOGD(context_->GetNodeName(), "%s AscendC: ops tilingdata: %s", paramInfo_.nodeType.c_str(), ss.str().c_str());46 OP_LOGD(context_->GetNodeName(), "%s AscendC: ops tilingdata: %s", paramInfo_.nodeType.c_str(), ss.str().c_str());
47}47}
48 48 
@@ -73,54 +73,54 @@ void Conv2dBaseTiling::PrintTilingInfo() const
73 73 
74void Conv2dBaseTiling::PrintLibApiTilingDataPartOne(std::stringstream &ss)74void Conv2dBaseTiling::PrintLibApiTilingDataPartOne(std::stringstream &ss)
75{75{
76- ss << "singleCoreHo: " << tilingData_.convApiTiling.get_singleCoreHo()76+ ss << "singleCoreHo: " << tilingData_.get_singleCoreHo()
77- << ", singleCoreWo: " << tilingData_.convApiTiling.get_singleCoreWo()77+ << ", singleCoreWo: " << tilingData_.get_singleCoreWo()
78- << ", singleCoreBatch: " << tilingData_.convApiTiling.get_singleCoreBatch()78+ << ", singleCoreBatch: " << tilingData_.get_singleCoreBatch()
79- << ", orgHi: " << tilingData_.convApiTiling.get_orgHi()79+ << ", orgHi: " << tilingData_.get_orgHi()
80- << ", orgWi: " << tilingData_.convApiTiling.get_orgWi()80+ << ", orgWi: " << tilingData_.get_orgWi()
81- << ", orgHo: " << tilingData_.convApiTiling.get_orgHo()81+ << ", orgHo: " << tilingData_.get_orgHo()
82- << ", orgWo: " << tilingData_.convApiTiling.get_orgWo()82+ << ", orgWo: " << tilingData_.get_orgWo()
83- << ", groups: " << tilingData_.convApiTiling.get_groups()83+ << ", groups: " << tilingData_.get_groups()
84- << ", orgCi: " << tilingData_.convApiTiling.get_orgCi()84+ << ", orgCi: " << tilingData_.get_orgCi()
85- << ", orgCo: " << tilingData_.convApiTiling.get_orgCo()85+ << ", orgCo: " << tilingData_.get_orgCo()
86- << ", kernelH: " << tilingData_.convApiTiling.get_kernelH()86+ << ", kernelH: " << tilingData_.get_kernelH()
87- << ", kernelW: " << tilingData_.convApiTiling.get_kernelW()87+ << ", kernelW: " << tilingData_.get_kernelW()
88- << ", singleCoreCi: " << tilingData_.convApiTiling.get_singleCoreCi()88+ << ", singleCoreCi: " << tilingData_.get_singleCoreCi()
89- << ", singleCoreCo: " << tilingData_.convApiTiling.get_singleCoreCo()89+ << ", singleCoreCo: " << tilingData_.get_singleCoreCo()
90- << ", hoL1: " << tilingData_.convApiTiling.get_hoL1()90+ << ", hoL1: " << tilingData_.get_hoL1()
91- << ", woL1: " << tilingData_.convApiTiling.get_woL1()91+ << ", woL1: " << tilingData_.get_woL1()
92- << ", kAL1: " << tilingData_.convApiTiling.get_kAL1()92+ << ", kAL1: " << tilingData_.get_kAL1()
93- << ", kBL1: " << tilingData_.convApiTiling.get_kBL1()93+ << ", kBL1: " << tilingData_.get_kBL1()
94- << ", nBL1: " << tilingData_.convApiTiling.get_nBL1()94+ << ", nBL1: " << tilingData_.get_nBL1()
95- << ", hoL0: " << tilingData_.convApiTiling.get_hoL0()95+ << ", hoL0: " << tilingData_.get_hoL0()
96- << ", woL0: " << tilingData_.convApiTiling.get_woL0()96+ << ", woL0: " << tilingData_.get_woL0()
97- << ", kL0: " << tilingData_.convApiTiling.get_kL0()97+ << ", kL0: " << tilingData_.get_kL0()
98- << ", nL0: " << tilingData_.convApiTiling.get_nL0()98+ << ", nL0: " << tilingData_.get_nL0()
99- << ", pBufferFlag: " << tilingData_.convApiTiling.get_pBufferFlag()99+ << ", pBufferFlag: " << tilingData_.get_pBufferFlag()
100- << ", multiNBL1: " << tilingData_.convApiTiling.get_multiNBL1()100+ << ", multiNBL1: " << tilingData_.get_multiNBL1()
101- << ", strideH: " << tilingData_.convApiTiling.get_strideH()101+ << ", strideH: " << tilingData_.get_strideH()
102- << ", strideW: " << tilingData_.convApiTiling.get_strideW()102+ << ", strideW: " << tilingData_.get_strideW()
103- << ", dilationH: " << tilingData_.convApiTiling.get_dilationH()103+ << ", dilationH: " << tilingData_.get_dilationH()
104- << ", dilationW: " << tilingData_.convApiTiling.get_dilationW()104+ << ", dilationW: " << tilingData_.get_dilationW()
105- << ", padTop: " << tilingData_.convApiTiling.get_padTop()105+ << ", padTop: " << tilingData_.get_padTop()
106- << ", padBottom: " << tilingData_.convApiTiling.get_padBottom()106+ << ", padBottom: " << tilingData_.get_padBottom()
107- << ", padLeft: " << tilingData_.convApiTiling.get_padLeft()107+ << ", padLeft: " << tilingData_.get_padLeft()
108- << ", padRight: " << tilingData_.convApiTiling.get_padRight()108+ << ", padRight: " << tilingData_.get_padRight()
109- << ", aL1SpaceSize: " << tilingData_.convApiTiling.get_aL1SpaceSize()109+ << ", aL1SpaceSize: " << tilingData_.get_aL1SpaceSize()
110- << ", singleCoreGroups: " << tilingData_.convApiTiling.get_singleCoreGroups()110+ << ", singleCoreGroups: " << tilingData_.get_singleCoreGroups()
111- << ", singleCoreGroupOpt: " << tilingData_.convApiTiling.get_singleCoreGroupOpt()111+ << ", singleCoreGroupOpt: " << tilingData_.get_singleCoreGroupOpt()
112- << ", enlarge: " << tilingData_.convApiTiling.get_enlarge()112+ << ", enlarge: " << tilingData_.get_enlarge()
113- << ", bUbNStep: " << tilingData_.convApiTiling.get_bUbNStep()113+ << ", bUbNStep: " << tilingData_.get_bUbNStep()
114- << ", iterateMNOrder: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_iterateMNOrder())114+ << ", iterateMNOrder: " << static_cast<uint32_t>(tilingData_.get_iterateMNOrder())
115- << ", biasFullLoadFlag: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_biasFullLoadFlag())115+ << ", biasFullLoadFlag: " << static_cast<uint32_t>(tilingData_.get_biasFullLoadFlag())
116- << ", fixpParamsFullLoadFlag: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_fixpParamsFullLoadFlag())116+ << ", fixpParamsFullLoadFlag: " << static_cast<uint32_t>(tilingData_.get_fixpParamsFullLoadFlag())
117- << ", hf32Enable: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_hf32Enable())117+ << ", hf32Enable: " << static_cast<uint32_t>(tilingData_.get_hf32Enable())
118- << ", hf32TransMode: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_hf32TransMode())118+ << ", hf32TransMode: " << static_cast<uint32_t>(tilingData_.get_hf32TransMode())
119- << ", hasBias: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_hasBias())119+ << ", hasBias: " << static_cast<uint32_t>(tilingData_.get_hasBias())
120- << ", hasScale: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_hasScale())120+ << ", hasScale: " << static_cast<uint32_t>(tilingData_.get_hasScale())
121- << ", offsetx: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_offsetx())121+ << ", offsetx: " << static_cast<uint32_t>(tilingData_.get_offsetx())
122- << ", roundMode: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_roundMode())122+ << ", roundMode: " << static_cast<uint32_t>(tilingData_.get_roundMode())
123- << ", innerBatch: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_innerBatch());123+ << ", innerBatch: " << static_cast<uint32_t>(tilingData_.get_innerBatch());
124}124}
125 125 
126 126 
@@ -133,34 +133,34 @@ void Conv2dBaseTiling::PrintLibApiTilingData()
133 std::stringstream ss;133 std::stringstream ss;
134 PrintLibApiTilingDataPartOne(ss);134 PrintLibApiTilingDataPartOne(ss);
135 std::stringstream ssPartTwo;135 std::stringstream ssPartTwo;
136- ssPartTwo << "bUbKStep: " << tilingData_.convApiTiling.get_bUbKStep()136+ ssPartTwo << "bUbKStep: " << tilingData_.get_bUbKStep()
137- << ", orgHixWi: " << tilingData_.convApiTiling.get_orgHixWi()137+ << ", orgHixWi: " << tilingData_.get_orgHixWi()
138- << ", kernelHxkernelW: " << tilingData_.convApiTiling.get_kernelHxkernelW()138+ << ", kernelHxkernelW: " << tilingData_.get_kernelHxkernelW()
139- << ", kernelHxkernelWxkernelD: " << tilingData_.convApiTiling.get_kernelHxkernelWxkernelD()139+ << ", kernelHxkernelWxkernelD: " << tilingData_.get_kernelHxkernelWxkernelD()
140- << ", cinAInCore: " << tilingData_.convApiTiling.get_cinAInCore()140+ << ", cinAInCore: " << tilingData_.get_cinAInCore()
141- << ", cinATailInCore: " << tilingData_.convApiTiling.get_cinATailInCore()141+ << ", cinATailInCore: " << tilingData_.get_cinATailInCore()
142- << ", cinBInCore: " << tilingData_.convApiTiling.get_cinBInCore()142+ << ", cinBInCore: " << tilingData_.get_cinBInCore()
143- << ", cinBTailInCore: " << tilingData_.convApiTiling.get_cinBTailInCore()143+ << ", cinBTailInCore: " << tilingData_.get_cinBTailInCore()
144- << ", mStep: " << tilingData_.convApiTiling.get_mStep()144+ << ", mStep: " << tilingData_.get_mStep()
145- << ", kStep: " << tilingData_.convApiTiling.get_kStep()145+ << ", kStep: " << tilingData_.get_kStep()
146- << ", nStep: " << tilingData_.convApiTiling.get_nStep()146+ << ", nStep: " << tilingData_.get_nStep()
147- << ", fmapKStride: " << tilingData_.convApiTiling.get_fmapKStride()147+ << ", fmapKStride: " << tilingData_.get_fmapKStride()
148- << ", weightKStride: " << tilingData_.convApiTiling.get_weightKStride()148+ << ", weightKStride: " << tilingData_.get_weightKStride()
149- << ", cinOffsetBlockInGM: " << tilingData_.convApiTiling.get_cinOffsetBlockInGM()149+ << ", cinOffsetBlockInGM: " << tilingData_.get_cinOffsetBlockInGM()
150- << ", coutOffsetBlock: " << tilingData_.convApiTiling.get_coutOffsetBlock()150+ << ", coutOffsetBlock: " << tilingData_.get_coutOffsetBlock()
151- << ", nL1DivBlockSize: " << tilingData_.convApiTiling.get_nL1DivBlockSize()151+ << ", nL1DivBlockSize: " << tilingData_.get_nL1DivBlockSize()
152- << ", dualOutput: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_dualOutput())152+ << ", dualOutput: " << static_cast<uint32_t>(tilingData_.get_dualOutput())
153- << ", quantMode0: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_quantMode0())153+ << ", quantMode0: " << static_cast<uint32_t>(tilingData_.get_quantMode0())
154- << ", reluMode0: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_reluMode0())154+ << ", reluMode0: " << static_cast<uint32_t>(tilingData_.get_reluMode0())
155- << ", clipMode0: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_clipMode0())155+ << ", clipMode0: " << static_cast<uint32_t>(tilingData_.get_clipMode0())
156- << ", quantMode1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_quantMode1())156+ << ", quantMode1: " << static_cast<uint32_t>(tilingData_.get_quantMode1())
157- << ", reluMode1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_reluMode1())157+ << ", reluMode1: " << static_cast<uint32_t>(tilingData_.get_reluMode1())
158- << ", clipMode1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_clipMode1())158+ << ", clipMode1: " << static_cast<uint32_t>(tilingData_.get_clipMode1())
159- << ", khL1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_khL1())159+ << ", khL1: " << static_cast<uint32_t>(tilingData_.get_khL1())
160- << ", kwL1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_kwL1())160+ << ", kwL1: " << static_cast<uint32_t>(tilingData_.get_kwL1())
161- << ", khUb: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_khUb())161+ << ", khUb: " << static_cast<uint32_t>(tilingData_.get_khUb())
162- << ", kwUb: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_kwUb())162+ << ", kwUb: " << static_cast<uint32_t>(tilingData_.get_kwUb())
163- << ", unionDataXt: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_unionDataXt());163+ << ", unionDataXt: " << static_cast<uint32_t>(tilingData_.get_unionDataXt());
164 OP_LOGD(context_->GetNodeName(), "%s AscendC: api tilingdata: %s", paramInfo_.nodeType.c_str(), ss.str().c_str());164 OP_LOGD(context_->GetNodeName(), "%s AscendC: api tilingdata: %s", paramInfo_.nodeType.c_str(), ss.str().c_str());
165 OP_LOGD(context_->GetNodeName(), "%s", ssPartTwo.str().c_str());165 OP_LOGD(context_->GetNodeName(), "%s", ssPartTwo.str().c_str());
166}166}
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_repo.cpp+107-117
@@ -118,51 +118,51 @@ bool Conv2dBaseTiling::TranslateRepoTiling(tuningtiling::TuningTilingDefPtr &tun
118 118 
119void Conv2dBaseTiling::TranslateApiTiling(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling)119void Conv2dBaseTiling::TranslateApiTiling(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling)
120{120{
121- tilingData_.convApiTiling.set_groups(convRepoTiling->groups);121+ tilingData_.set_groups(convRepoTiling->groups);
122- tilingData_.convApiTiling.set_orgCi(convRepoTiling->orgCi);122+ tilingData_.set_orgCi(convRepoTiling->orgCi);
123- tilingData_.convApiTiling.set_orgHi(convRepoTiling->orgHi);123+ tilingData_.set_orgHi(convRepoTiling->orgHi);
124- tilingData_.convApiTiling.set_orgWi(convRepoTiling->orgWi);124+ tilingData_.set_orgWi(convRepoTiling->orgWi);
125- tilingData_.convApiTiling.set_orgCo(convRepoTiling->orgCo);125+ tilingData_.set_orgCo(convRepoTiling->orgCo);
126- tilingData_.convApiTiling.set_orgHo(convRepoTiling->orgHo);126+ tilingData_.set_orgHo(convRepoTiling->orgHo);
127- tilingData_.convApiTiling.set_orgWo(convRepoTiling->orgWo);127+ tilingData_.set_orgWo(convRepoTiling->orgWo);
128- tilingData_.convApiTiling.set_kernelH(convRepoTiling->kernelH);128+ tilingData_.set_kernelH(convRepoTiling->kernelH);
129- tilingData_.convApiTiling.set_kernelW(convRepoTiling->kernelW);129+ tilingData_.set_kernelW(convRepoTiling->kernelW);
130- tilingData_.convApiTiling.set_singleCoreCi(convRepoTiling->singleCoreCi);130+ tilingData_.set_singleCoreCi(convRepoTiling->singleCoreCi);
131- tilingData_.convApiTiling.set_singleCoreCo(convRepoTiling->singleCoreCo);131+ tilingData_.set_singleCoreCo(convRepoTiling->singleCoreCo);
132- tilingData_.convApiTiling.set_singleCoreHo(convRepoTiling->singleCoreHo);132+ tilingData_.set_singleCoreHo(convRepoTiling->singleCoreHo);
133- tilingData_.convApiTiling.set_singleCoreWo(convRepoTiling->singleCoreWo);133+ tilingData_.set_singleCoreWo(convRepoTiling->singleCoreWo);
134- tilingData_.convApiTiling.set_singleCoreBatch(ConvCeilDiv(shapeInfo_.batch, convRepoTiling->batchDim));134+ tilingData_.set_singleCoreBatch(ConvCeilDiv(shapeInfo_.batch, convRepoTiling->batchDim));
135- tilingData_.convApiTiling.set_hoL1(convRepoTiling->hoL1);135+ tilingData_.set_hoL1(convRepoTiling->hoL1);
136- tilingData_.convApiTiling.set_woL1(convRepoTiling->woL1);136+ tilingData_.set_woL1(convRepoTiling->woL1);
137- tilingData_.convApiTiling.set_kAL1(convRepoTiling->kAL1);137+ tilingData_.set_kAL1(convRepoTiling->kAL1);
138- tilingData_.convApiTiling.set_kBL1(convRepoTiling->kBL1);138+ tilingData_.set_kBL1(convRepoTiling->kBL1);
139- tilingData_.convApiTiling.set_nBL1(convRepoTiling->nBL1);139+ tilingData_.set_nBL1(convRepoTiling->nBL1);
140- tilingData_.convApiTiling.set_hoL0(convRepoTiling->hoL0);140+ tilingData_.set_hoL0(convRepoTiling->hoL0);
141- tilingData_.convApiTiling.set_woL0(convRepoTiling->woL0);141+ tilingData_.set_woL0(convRepoTiling->woL0);
142- tilingData_.convApiTiling.set_kL0(convRepoTiling->kL0);142+ tilingData_.set_kL0(convRepoTiling->kL0);
143- tilingData_.convApiTiling.set_nL0(convRepoTiling->nL0);143+ tilingData_.set_nL0(convRepoTiling->nL0);
144- tilingData_.convApiTiling.set_pBufferFlag(convRepoTiling->pBufferFlag);144+ tilingData_.set_pBufferFlag(convRepoTiling->pBufferFlag);
145- tilingData_.convApiTiling.set_strideH(convRepoTiling->strideH);145+ tilingData_.set_strideH(convRepoTiling->strideH);
146- tilingData_.convApiTiling.set_strideW(convRepoTiling->strideW);146+ tilingData_.set_strideW(convRepoTiling->strideW);
147- tilingData_.convApiTiling.set_dilationH(convRepoTiling->dilationH);147+ tilingData_.set_dilationH(convRepoTiling->dilationH);
148- tilingData_.convApiTiling.set_dilationW(convRepoTiling->dilationW);148+ tilingData_.set_dilationW(convRepoTiling->dilationW);
149- tilingData_.convApiTiling.set_padTop(convRepoTiling->padTop);149+ tilingData_.set_padTop(convRepoTiling->padTop);
150- tilingData_.convApiTiling.set_padBottom(convRepoTiling->padBottom);150+ tilingData_.set_padBottom(convRepoTiling->padBottom);
151- tilingData_.convApiTiling.set_padLeft(convRepoTiling->padLeft);151+ tilingData_.set_padLeft(convRepoTiling->padLeft);
152- tilingData_.convApiTiling.set_padRight(convRepoTiling->padRight);152+ tilingData_.set_padRight(convRepoTiling->padRight);
153- tilingData_.convApiTiling.set_iterateMNOrder(convRepoTiling->iterateMNOrder);153+ tilingData_.set_iterateMNOrder(convRepoTiling->iterateMNOrder);
154- tilingData_.convApiTiling.set_biasFullLoadFlag(convRepoTiling->biasFullLoadFlag);154+ tilingData_.set_biasFullLoadFlag(convRepoTiling->biasFullLoadFlag);
155- tilingData_.convApiTiling.set_fixpParamsFullLoadFlag(convRepoTiling->fixpParamsFullLoadFlag);155+ tilingData_.set_fixpParamsFullLoadFlag(convRepoTiling->fixpParamsFullLoadFlag);
156- tilingData_.convApiTiling.set_offsetx(attrInfo_.offsetx);156+ tilingData_.set_offsetx(attrInfo_.offsetx);
157- tilingData_.convApiTiling.set_hf32Enable(convRepoTiling->hf32Enable);157+ tilingData_.set_hf32Enable(convRepoTiling->hf32Enable);
158- tilingData_.convApiTiling.set_hf32TransMode(convRepoTiling->hf32TransMode);158+ tilingData_.set_hf32TransMode(convRepoTiling->hf32TransMode);
159- tilingData_.convApiTiling.set_roundMode(attrInfo_.roundMode);159+ tilingData_.set_roundMode(attrInfo_.roundMode);
160 if (convRepoTiling->enlarge > 0) {160 if (convRepoTiling->enlarge > 0) {
161- tilingData_.convApiTiling.set_enlarge(convRepoTiling->enlarge);161+ tilingData_.set_enlarge(convRepoTiling->enlarge);
162 } else {162 } else {
163- tilingData_.convApiTiling.set_enlarge(optGroupInfo_.enlarge);163+ tilingData_.set_enlarge(optGroupInfo_.enlarge);
164 }164 }
165- tilingData_.convApiTiling.set_innerBatch(convRepoTiling->innerBatch);165+ tilingData_.set_innerBatch(convRepoTiling->innerBatch);
166}166}
167 167 
168void Conv2dBaseTiling::TranslateApiTilingAux(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling)168void Conv2dBaseTiling::TranslateApiTilingAux(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling)
@@ -188,31 +188,31 @@ void Conv2dBaseTiling::TranslateApiTilingAux(shared_ptr<tuningtiling::Conv2DV2Tu
188 kBTail = kBTail == 0 ? convRepoTiling->kBL1 : kBTail;188 kBTail = kBTail == 0 ? convRepoTiling->kBL1 : kBTail;
189 189 
190 if (featureFlagInfo_ == ConvAscendcFeatureFlag::IS_DMA_FLAG) {190 if (featureFlagInfo_ == ConvAscendcFeatureFlag::IS_DMA_FLAG) {
191- tilingData_.convApiTiling.set_khL1(convRepoTiling->khL1);191+ tilingData_.set_khL1(convRepoTiling->khL1);
192- tilingData_.convApiTiling.set_kwL1(convRepoTiling->kwL1);192+ tilingData_.set_kwL1(convRepoTiling->kwL1);
193 } else {193 } else {
194- tilingData_.convApiTiling.set_khL1(0);194+ tilingData_.set_khL1(0);
195- tilingData_.convApiTiling.set_kwL1(0);195+ tilingData_.set_kwL1(0);
196 }196 }
197- tilingData_.convApiTiling.set_kernelHxkernelW(kernelHW);197+ tilingData_.set_kernelHxkernelW(kernelHW);
198- tilingData_.convApiTiling.set_kernelHxkernelWxkernelD(kernelHW);198+ tilingData_.set_kernelHxkernelWxkernelD(kernelHW);
199- tilingData_.convApiTiling.set_multiNBL1(ConvCeilDiv(convRepoTiling->nBL1, convRepoTiling->nL0));199+ tilingData_.set_multiNBL1(ConvCeilDiv(convRepoTiling->nBL1, convRepoTiling->nL0));
200- tilingData_.convApiTiling.set_cinAInCore(cinAInCore);200+ tilingData_.set_cinAInCore(cinAInCore);
201- tilingData_.convApiTiling.set_cinATailInCore(kATail / kernelValueInKSize);201+ tilingData_.set_cinATailInCore(kATail / kernelValueInKSize);
202- tilingData_.convApiTiling.set_orgHixWi(orgHiWi);202+ tilingData_.set_orgHixWi(orgHiWi);
203- tilingData_.convApiTiling.set_cinOffsetBlockInGM(orgHiWi * cinAInCore);203+ tilingData_.set_cinOffsetBlockInGM(orgHiWi * cinAInCore);
204- tilingData_.convApiTiling.set_mStep(mStep);204+ tilingData_.set_mStep(mStep);
205- tilingData_.convApiTiling.set_fmapKStride(fmapKStride);205+ tilingData_.set_fmapKStride(fmapKStride);
206- tilingData_.convApiTiling.set_nStep(ConvCeilDiv(convRepoTiling->nL0, convOpsConstParams_.n0));206+ tilingData_.set_nStep(ConvCeilDiv(convRepoTiling->nL0, convOpsConstParams_.n0));
207- tilingData_.convApiTiling.set_kStep(convRepoTiling->kL0 / convOpsConstParams_.k0);207+ tilingData_.set_kStep(convRepoTiling->kL0 / convOpsConstParams_.k0);
208- tilingData_.convApiTiling.set_weightKStride(weightKStride);208+ tilingData_.set_weightKStride(weightKStride);
209- tilingData_.convApiTiling.set_coutOffsetBlock((convRepoTiling->orgCi / convRepoTiling->groups) * kernelHW);209+ tilingData_.set_coutOffsetBlock((convRepoTiling->orgCi / convRepoTiling->groups) * kernelHW);
210- tilingData_.convApiTiling.set_cinBInCore(cinBInCore);210+ tilingData_.set_cinBInCore(cinBInCore);
211- tilingData_.convApiTiling.set_cinBTailInCore(kBTail / kernelValueInKSize);211+ tilingData_.set_cinBTailInCore(kBTail / kernelValueInKSize);
212- tilingData_.convApiTiling.set_nL1DivBlockSize(convRepoTiling->nBL1 / convOpsConstParams_.n0);212+ tilingData_.set_nL1DivBlockSize(convRepoTiling->nBL1 / convOpsConstParams_.n0);
213- tilingData_.convApiTiling.set_aL1SpaceSize(CalcAL1SpaceSize(convRepoTiling));213+ tilingData_.set_aL1SpaceSize(CalcAL1SpaceSize(convRepoTiling));
214- tilingData_.convApiTiling.set_hasBias(static_cast<uint8_t>(flagInfo_.hasBias));214+ tilingData_.set_hasBias(static_cast<uint8_t>(flagInfo_.hasBias));
215- tilingData_.convApiTiling.set_hasScale(static_cast<uint8_t>(flagInfo_.quantFlag || flagInfo_.extendConvFlag));215+ tilingData_.set_hasScale(static_cast<uint8_t>(flagInfo_.quantFlag || flagInfo_.extendConvFlag));
216 uint64_t singleGroups = 0;216 uint64_t singleGroups = 0;
217 uint64_t singleGroupOpt = 0;217 uint64_t singleGroupOpt = 0;
218 if (flagInfo_.convGroupType == ConvGroupType::OPT_GROUP_CONV) {218 if (flagInfo_.convGroupType == ConvGroupType::OPT_GROUP_CONV) {
@@ -222,8 +222,8 @@ void Conv2dBaseTiling::TranslateApiTilingAux(shared_ptr<tuningtiling::Conv2DV2Tu
222 ConvCeilDiv(optGroupInfo_.groupOpt, convRepoTiling->groupDim);222 ConvCeilDiv(optGroupInfo_.groupOpt, convRepoTiling->groupDim);
223 }223 }
224 224 
225- tilingData_.convApiTiling.set_singleCoreGroups(singleGroups);225+ tilingData_.set_singleCoreGroups(singleGroups);
226- tilingData_.convApiTiling.set_singleCoreGroupOpt(singleGroupOpt);226+ tilingData_.set_singleCoreGroupOpt(singleGroupOpt);
227 227 
228 SetUbTiling(convRepoTiling);228 SetUbTiling(convRepoTiling);
229 SetFixpipeTiling();229 SetFixpipeTiling();
@@ -240,18 +240,18 @@ void Conv2dBaseTiling::SetUbTiling(shared_ptr<tuningtiling::Conv2DV2TunnerTiling
240 c04Info_.n0 = convOpsConstParams_.n0;240 c04Info_.n0 = convOpsConstParams_.n0;
241 c04Info_.k0 = convOpsConstParams_.k0;241 c04Info_.k0 = convOpsConstParams_.k0;
242 c04Info_.weightDtype = dtypeMap.at(descInfo_.weightDtype);242 c04Info_.weightDtype = dtypeMap.at(descInfo_.weightDtype);
243- tilingData_.convApiTiling.set_bUbNStep(static_cast<uint32_t>(conv2dApiTiling_.CalcC04UbLoadNsize(c04Info_)));243+ tilingData_.set_bUbNStep(static_cast<uint32_t>(conv2dApiTiling_.CalcC04UbLoadNsize(c04Info_)));
244- tilingData_.convApiTiling.set_bUbKStep(0);244+ tilingData_.set_bUbKStep(0);
245 } else if (convRepoTiling->isWeightUbTransFlag != 0) {245 } else if (convRepoTiling->isWeightUbTransFlag != 0) {
246 conv_tiling::ConvWeightUbTransParams params = {convRepoTiling->nBL1, convRepoTiling->kBL1,246 conv_tiling::ConvWeightUbTransParams params = {convRepoTiling->nBL1, convRepoTiling->kBL1,
247 static_cast<uint64_t>(convRepoTiling->kernelH), static_cast<uint64_t>(convRepoTiling->kernelW),247 static_cast<uint64_t>(convRepoTiling->kernelH), static_cast<uint64_t>(convRepoTiling->kernelW),
248 convOpsConstParams_.k0, convOpsConstParams_.n0, dtypeMap.at(descInfo_.weightDtype)};248 convOpsConstParams_.k0, convOpsConstParams_.n0, dtypeMap.at(descInfo_.weightDtype)};
249 conv2dApiTiling_.GetWeightUBTiling(params);249 conv2dApiTiling_.GetWeightUBTiling(params);
250- tilingData_.convApiTiling.set_bUbNStep(params.bUbNStep);250+ tilingData_.set_bUbNStep(params.bUbNStep);
251- tilingData_.convApiTiling.set_bUbKStep(params.bUbKStep);251+ tilingData_.set_bUbKStep(params.bUbKStep);
252 } else {252 } else {
253- tilingData_.convApiTiling.set_bUbNStep(0);253+ tilingData_.set_bUbNStep(0);
254- tilingData_.convApiTiling.set_bUbKStep(0);254+ tilingData_.set_bUbKStep(0);
255 }255 }
256 256 
257 if (featureFlagInfo_ == ConvAscendcFeatureFlag::IS_DMA_FLAG) {257 if (featureFlagInfo_ == ConvAscendcFeatureFlag::IS_DMA_FLAG) {
@@ -259,61 +259,51 @@ void Conv2dBaseTiling::SetUbTiling(shared_ptr<tuningtiling::Conv2DV2TunnerTiling
259 static_cast<uint64_t>(convRepoTiling->khL1), static_cast<uint64_t>(convRepoTiling->kwL1),259 static_cast<uint64_t>(convRepoTiling->khL1), static_cast<uint64_t>(convRepoTiling->kwL1),
260 convOpsConstParams_.k0, dtypeMap.at(descInfo_.fMapDtype)};260 convOpsConstParams_.k0, dtypeMap.at(descInfo_.fMapDtype)};
261 conv2dApiTiling_.GetDmaUbTiling(params);261 conv2dApiTiling_.GetDmaUbTiling(params);
262- tilingData_.convApiTiling.set_khUb(params.khUb);262+ tilingData_.set_khUb(params.khUb);
263- tilingData_.convApiTiling.set_kwUb(params.kwUb);263+ tilingData_.set_kwUb(params.kwUb);
264 } else {264 } else {
265- tilingData_.convApiTiling.set_khUb(0);265+ tilingData_.set_khUb(0);
266- tilingData_.convApiTiling.set_kwUb(0);266+ tilingData_.set_kwUb(0);
267 }267 }
268}268}
269 269 
270void Conv2dBaseTiling::SetFixpipeTiling()270void Conv2dBaseTiling::SetFixpipeTiling()
271{271{
272- tilingData_.convApiTiling.set_dualOutput(fixpipeInfo_.dualOutput);272+ tilingData_.set_dualOutput(fixpipeInfo_.dualOutput);
273- tilingData_.convApiTiling.set_quantMode0(fixpipeInfo_.quantMode0);273+ tilingData_.set_quantMode0(fixpipeInfo_.quantMode0);
274- tilingData_.convApiTiling.set_reluMode0(fixpipeInfo_.reluMode0);274+ tilingData_.set_reluMode0(fixpipeInfo_.reluMode0);
275- tilingData_.convApiTiling.set_clipMode0(fixpipeInfo_.clipMode0);275+ tilingData_.set_clipMode0(fixpipeInfo_.clipMode0);
276- tilingData_.convApiTiling.set_quantMode1(fixpipeInfo_.quantMode1);276+ tilingData_.set_quantMode1(fixpipeInfo_.quantMode1);
277- tilingData_.convApiTiling.set_reluMode1(fixpipeInfo_.reluMode1);277+ tilingData_.set_reluMode1(fixpipeInfo_.reluMode1);
278- tilingData_.convApiTiling.set_clipMode1(fixpipeInfo_.clipMode1);278+ tilingData_.set_clipMode1(fixpipeInfo_.clipMode1);
279}279}
280 280 
281void Conv2dBaseTiling::TranslateRunInfo(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling)281void Conv2dBaseTiling::TranslateRunInfo(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling)
282{282{
283- tilingData_.convRunInfo.set_batch(shapeInfo_.batch);283+ tilingData_.set_batch(shapeInfo_.batch);
284- tilingData_.convRunInfo.set_hin(convRepoTiling->orgHi);284+ tilingData_.set_hin(convRepoTiling->orgHi);
285- tilingData_.convRunInfo.set_win(convRepoTiling->orgWi);285+ tilingData_.set_win(convRepoTiling->orgWi);
286- tilingData_.convRunInfo.set_batchDim(convRepoTiling->batchDim);286+ tilingData_.set_batchDim(convRepoTiling->batchDim);
287- tilingData_.convRunInfo.set_hoDim(convRepoTiling->hoDim);287+ tilingData_.set_hoDim(convRepoTiling->hoDim);
288- tilingData_.convRunInfo.set_woDim(featureFlagInfo_ == ConvAscendcFeatureFlag::IS_CONV1D_FLAG ?288+ tilingData_.set_woDim(featureFlagInfo_ == ConvAscendcFeatureFlag::IS_CONV1D_FLAG ?
289 convRepoTiling->woDim : 1);289 convRepoTiling->woDim : 1);
290- tilingData_.convRunInfo.set_nDim(convRepoTiling->nDim);290+ tilingData_.set_nDim(convRepoTiling->nDim);
291- tilingData_.convRunInfo.set_cin(convRepoTiling->orgCi);291+ tilingData_.set_cin(convRepoTiling->orgCi);
292- tilingData_.convRunInfo.set_cout(convRepoTiling->orgCo);292+ tilingData_.set_cout(convRepoTiling->orgCo);
293- tilingData_.convRunInfo.set_kh(convRepoTiling->kernelH);293+ tilingData_.set_kh(convRepoTiling->kernelH);
294- tilingData_.convRunInfo.set_kw(convRepoTiling->kernelW);294+ tilingData_.set_kw(convRepoTiling->kernelW);
295- tilingData_.convRunInfo.set_hout(convRepoTiling->orgHo);295+ tilingData_.set_hout(convRepoTiling->orgHo);
296- tilingData_.convRunInfo.set_wout(convRepoTiling->orgWo);296+ tilingData_.set_wout(convRepoTiling->orgWo);
297- tilingData_.convRunInfo.set_strideH(convRepoTiling->strideH);
298- tilingData_.convRunInfo.set_strideW(convRepoTiling->strideW);
299- tilingData_.convRunInfo.set_dilationH(convRepoTiling->dilationH);
300- tilingData_.convRunInfo.set_dilationW(convRepoTiling->dilationW);
301- tilingData_.convRunInfo.set_padTop(convRepoTiling->padTop);
302- tilingData_.convRunInfo.set_padLeft(convRepoTiling->padLeft);
303- tilingData_.convRunInfo.set_hasBias(flagInfo_.hasBias);
304- tilingData_.convRunInfo.set_groups(convRepoTiling->groups);
305 if (convRepoTiling->enlarge > 0) {297 if (convRepoTiling->enlarge > 0) {
306- tilingData_.convRunInfo.set_cinOpt(oriGroupInfo_.ciPerGroup * convRepoTiling->enlarge);298+ tilingData_.set_cinOpt(oriGroupInfo_.ciPerGroup * convRepoTiling->enlarge);
307- tilingData_.convRunInfo.set_coutOpt(oriGroupInfo_.coPerGroup * convRepoTiling->enlarge);299+ tilingData_.set_coutOpt(oriGroupInfo_.coPerGroup * convRepoTiling->enlarge);
308- tilingData_.convRunInfo.set_groupOpt(ConvCeilDiv(convRepoTiling->groups, convRepoTiling->enlarge));300+ tilingData_.set_groupOpt(ConvCeilDiv(convRepoTiling->groups, convRepoTiling->enlarge));
309- tilingData_.convRunInfo.set_enlarge(convRepoTiling->enlarge);
310 } else {301 } else {
311- tilingData_.convRunInfo.set_cinOpt(optGroupInfo_.cinOpt);302+ tilingData_.set_cinOpt(optGroupInfo_.cinOpt);
312- tilingData_.convRunInfo.set_coutOpt(optGroupInfo_.coutOpt);303+ tilingData_.set_coutOpt(optGroupInfo_.coutOpt);
313- tilingData_.convRunInfo.set_groupOpt(optGroupInfo_.groupOpt);304+ tilingData_.set_groupOpt(optGroupInfo_.groupOpt);
314- tilingData_.convRunInfo.set_enlarge(optGroupInfo_.enlarge);
315 }305 }
316- tilingData_.convRunInfo.set_groupDim(convRepoTiling->groupDim);306+ tilingData_.set_groupDim(convRepoTiling->groupDim);
317}307}
318 308 
319uint32_t Conv2dBaseTiling::CalcAL1SpaceSize(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling)309uint32_t Conv2dBaseTiling::CalcAL1SpaceSize(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling)
@@ -335,7 +325,7 @@ uint32_t Conv2dBaseTiling::CalcAL1SpaceSize(shared_ptr<tuningtiling::Conv2DV2Tun
335 uint64_t hoL1Max = std::min(mL1Max / convRepoTiling->orgWo + 2, convRepoTiling->orgHo);325 uint64_t hoL1Max = std::min(mL1Max / convRepoTiling->orgWo + 2, convRepoTiling->orgHo);
336 uint64_t hiAL1Max = ConvInferHiL1(hoL1Max, convRepoTiling->orgHi, convRepoTiling->kernelH, convRepoTiling->dilationH,326 uint64_t hiAL1Max = ConvInferHiL1(hoL1Max, convRepoTiling->orgHi, convRepoTiling->kernelH, convRepoTiling->dilationH,
337 convRepoTiling->strideH);327 convRepoTiling->strideH);
338- aL1SpaceSize = tilingData_.convApiTiling.get_cinAInCore() * hiAL1Max * convRepoTiling->orgWi;328+ aL1SpaceSize = tilingData_.get_cinAInCore() * hiAL1Max * convRepoTiling->orgWi;
339 }329 }
340 } else {330 } else {
341 uint64_t hiAL1Max = ConvInferHiL1(convRepoTiling->hoL1, convRepoTiling->orgHi, convRepoTiling->kernelH, convRepoTiling->dilationH,331 uint64_t hiAL1Max = ConvInferHiL1(convRepoTiling->hoL1, convRepoTiling->orgHi, convRepoTiling->kernelH, convRepoTiling->dilationH,
@@ -347,7 +337,7 @@ uint32_t Conv2dBaseTiling::CalcAL1SpaceSize(shared_ptr<tuningtiling::Conv2DV2Tun
347 } else {337 } else {
348 wiAL1Max = ConvInferWiL1(convRepoTiling->woL1, convRepoTiling->orgWi, convRepoTiling->kernelW, convRepoTiling->dilationW,338 wiAL1Max = ConvInferWiL1(convRepoTiling->woL1, convRepoTiling->orgWi, convRepoTiling->kernelW, convRepoTiling->dilationW,
349 convRepoTiling->strideW);339 convRepoTiling->strideW);
350- aL1SpaceSize = tilingData_.convApiTiling.get_cinAInCore() * hiAL1Max * wiAL1Max;340+ aL1SpaceSize = tilingData_.get_cinAInCore() * hiAL1Max * wiAL1Max;
351 }341 }
352 }342 }
353 aL1SpaceSize = ConvAlignB(aL1SpaceSize * fmapSize, C0_SIZE) * convRepoTiling->innerBatch;343 aL1SpaceSize = ConvAlignB(aL1SpaceSize * fmapSize, C0_SIZE) * convRepoTiling->innerBatch;
@@ -366,7 +356,7 @@ void Conv2dBaseTiling::SetUnionDataXt(shared_ptr<tuningtiling::Conv2DV2TunnerTil
366 unionDataXt.bf.dilationW = static_cast<uint64_t>(convRepoTiling->dilationW);356 unionDataXt.bf.dilationW = static_cast<uint64_t>(convRepoTiling->dilationW);
367 unionDataXt.bf.strideH = static_cast<uint64_t>(convRepoTiling->strideH) & 0x3f;357 unionDataXt.bf.strideH = static_cast<uint64_t>(convRepoTiling->strideH) & 0x3f;
368 unionDataXt.bf.strideW = static_cast<uint64_t>(convRepoTiling->strideW) & 0x3f;358 unionDataXt.bf.strideW = static_cast<uint64_t>(convRepoTiling->strideW) & 0x3f;
369- tilingData_.convApiTiling.set_unionDataXt(unionDataXt.n);359+ tilingData_.set_unionDataXt(unionDataXt.n);
370}360}
371 361 
372void Conv2dBaseTiling::PrintInputArgs(shared_ptr<tuningtiling::Conv2DV2InputArgs> conv2DInput)362void Conv2dBaseTiling::PrintInputArgs(shared_ptr<tuningtiling::Conv2DV2InputArgs> conv2DInput)
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_tilingkey.cpp+29-29
@@ -27,7 +27,7 @@ uint64_t Conv2dBaseTiling::GetSmallWeightVal()
27 }27 }
28 28 
29 // Not useful to weight ub29 // Not useful to weight ub
30- if (tilingData_.convApiTiling.get_bUbNStep() > 0 && tilingData_.convApiTiling.get_bUbKStep() > 0) {30+ if (tilingData_.get_bUbNStep() > 0 && tilingData_.get_bUbKStep() > 0) {
31 return CONV_NOT_SMALL_WEIGHT;31 return CONV_NOT_SMALL_WEIGHT;
32 }32 }
33 33 
@@ -37,9 +37,9 @@ uint64_t Conv2dBaseTiling::GetSmallWeightVal()
37 ci1 * shapeInfo_.kh * shapeInfo_.kw * convOpsConstParams_.k0;37 ci1 * shapeInfo_.kh * shapeInfo_.kw * convOpsConstParams_.k0;
38 uint64_t singleCoreNSize = ConvAlignB(ConvCeilDiv(shapeInfo_.co, numBlocksRes.nDim), convOpsConstParams_.n0);38 uint64_t singleCoreNSize = ConvAlignB(ConvCeilDiv(shapeInfo_.co, numBlocksRes.nDim), convOpsConstParams_.n0);
39 int64_t weightDtypeSize = dtypeSizeTab.at(descInfo_.weightDtype);39 int64_t weightDtypeSize = dtypeSizeTab.at(descInfo_.weightDtype);
40- if (weightKSize == tilingData_.convApiTiling.get_kBL1() &&40+ if (weightKSize == tilingData_.get_kBL1() &&
41- weightKSize == tilingData_.convApiTiling.get_kAL1() &&41+ weightKSize == tilingData_.get_kAL1() &&
42- tilingData_.convApiTiling.get_nL0() == singleCoreNSize) {42+ tilingData_.get_nL0() == singleCoreNSize) {
43 if (weightKSize * singleCoreNSize * weightDtypeSize <= apiInputPlatformInfo.l0BSize) {43 if (weightKSize * singleCoreNSize * weightDtypeSize <= apiInputPlatformInfo.l0BSize) {
44 return CONV_WEIGHT_SMALLER_THAN_BL0;44 return CONV_WEIGHT_SMALLER_THAN_BL0;
45 }45 }
@@ -58,7 +58,7 @@ uint64_t Conv2dBaseTiling::GetFmpTilingVal()
58 uint64_t fmpKSize = flagInfo_.enableC04Flag ? ConvAlignB(C04_CIN_SIZE *58 uint64_t fmpKSize = flagInfo_.enableC04Flag ? ConvAlignB(C04_CIN_SIZE *
59 shapeInfo_.kh * shapeInfo_.kw, convOpsConstParams_.k0) :59 shapeInfo_.kh * shapeInfo_.kw, convOpsConstParams_.k0) :
60 ci1 * shapeInfo_.kh * shapeInfo_.kw * convOpsConstParams_.k0;60 ci1 * shapeInfo_.kh * shapeInfo_.kw * convOpsConstParams_.k0;
61- bool kAL1FullloadFlag = tilingData_.convApiTiling.get_kAL1() == fmpKSize;61+ bool kAL1FullloadFlag = tilingData_.get_kAL1() == fmpKSize;
62 if (flagInfo_.mSplitModeFlag) {62 if (flagInfo_.mSplitModeFlag) {
63 return GetFmpTilingValForMSplit(kAL1FullloadFlag);63 return GetFmpTilingValForMSplit(kAL1FullloadFlag);
64 }64 }
@@ -67,10 +67,10 @@ uint64_t Conv2dBaseTiling::GetFmpTilingVal()
67 67
68uint64_t Conv2dBaseTiling::GetFmpTilingValForMSplit(bool kAL1FullloadFlag)68uint64_t Conv2dBaseTiling::GetFmpTilingValForMSplit(bool kAL1FullloadFlag)
69{69{
70- bool mL1FullloadFlag = tilingData_.convApiTiling.get_innerBatch() == 1 ?70+ bool mL1FullloadFlag = tilingData_.get_innerBatch() == 1 ?
71- tilingData_.convApiTiling.get_singleCoreHo() <= tilingData_.convApiTiling.get_hoL1() :71+ tilingData_.get_singleCoreHo() <= tilingData_.get_hoL1() :
72- tilingData_.convApiTiling.get_innerBatch() == tilingData_.convApiTiling.get_singleCoreBatch();72+ tilingData_.get_innerBatch() == tilingData_.get_singleCoreBatch();
73- bool mL0FullloadFlag = tilingData_.convApiTiling.get_hoL1() == tilingData_.convApiTiling.get_hoL0();73+ bool mL0FullloadFlag = tilingData_.get_hoL1() == tilingData_.get_hoL0();
74 if (kAL1FullloadFlag && mL1FullloadFlag) {74 if (kAL1FullloadFlag && mL1FullloadFlag) {
75 return FULLLOAD_AL1;75 return FULLLOAD_AL1;
76 } else if (!kAL1FullloadFlag && mL1FullloadFlag && mL0FullloadFlag) {76 } else if (!kAL1FullloadFlag && mL1FullloadFlag && mL0FullloadFlag) {
@@ -81,7 +81,7 @@ uint64_t Conv2dBaseTiling::GetFmpTilingValForMSplit(bool kAL1FullloadFlag)
81 81 
82uint64_t Conv2dBaseTiling::GetL0PingPongVal()82uint64_t Conv2dBaseTiling::GetL0PingPongVal()
83{83{
84- return static_cast<uint64_t>(tilingData_.convApiTiling.get_pBufferFlag() & L0A_L0B_PB_FLAG_MASK);84+ return static_cast<uint64_t>(tilingData_.get_pBufferFlag() & L0A_L0B_PB_FLAG_MASK);
85}85}
86 86 
87uint64_t Conv2dBaseTiling::GetWeightTilingVal()87uint64_t Conv2dBaseTiling::GetWeightTilingVal()
@@ -96,11 +96,11 @@ uint64_t Conv2dBaseTiling::GetWeightTilingVal()
96 shapeInfo_.kh * shapeInfo_.kw, convOpsConstParams_.k0) :96 shapeInfo_.kh * shapeInfo_.kw, convOpsConstParams_.k0) :
97 ci1 * shapeInfo_.kh * shapeInfo_.kw * convOpsConstParams_.k0;97 ci1 * shapeInfo_.kh * shapeInfo_.kw * convOpsConstParams_.k0;
98 uint64_t singleCoreNSize = ConvAlignB(ConvCeilDiv(shapeInfo_.co, numBlocksRes.nDim), convOpsConstParams_.n0);98 uint64_t singleCoreNSize = ConvAlignB(ConvCeilDiv(shapeInfo_.co, numBlocksRes.nDim), convOpsConstParams_.n0);
99- if (tilingData_.convApiTiling.get_kBL1() == weightKSize) {99+ if (tilingData_.get_kBL1() == weightKSize) {
100 kBL1FullloadFlag = true;100 kBL1FullloadFlag = true;
101 }101 }
102 102 
103- if (tilingData_.convApiTiling.get_nBL1() == singleCoreNSize) {103+ if (tilingData_.get_nBL1() == singleCoreNSize) {
104 nBL1FullloadFlag = true;104 nBL1FullloadFlag = true;
105 }105 }
106 106 
@@ -108,7 +108,7 @@ uint64_t Conv2dBaseTiling::GetWeightTilingVal()
108 return FULLLOAD_BL1;108 return FULLLOAD_BL1;
109 }109 }
110 110 
111- if (!kBL1FullloadFlag && tilingData_.convApiTiling.get_nL0() == singleCoreNSize) {111+ if (!kBL1FullloadFlag && tilingData_.get_nL0() == singleCoreNSize) {
112 return ONLY_N_FULLLOAD_BL1_BL0;112 return ONLY_N_FULLLOAD_BL1_BL0;
113 }113 }
114 return WEIGHT_OTHER;114 return WEIGHT_OTHER;
@@ -124,16 +124,16 @@ uint64_t Conv2dBaseTiling::GetOutputOrderVal()
124 124
125uint64_t Conv2dBaseTiling::GetFmpTilingValForHWSplit(bool kAL1FullloadFlag)125uint64_t Conv2dBaseTiling::GetFmpTilingValForHWSplit(bool kAL1FullloadFlag)
126{126{
127- bool hoL1FullloadFlag = tilingData_.convApiTiling.get_singleCoreHo() <= tilingData_.convApiTiling.get_hoL1();127+ bool hoL1FullloadFlag = tilingData_.get_singleCoreHo() <= tilingData_.get_hoL1();
128 bool woL1FullloadFlag = false;128 bool woL1FullloadFlag = false;
129- if (tilingData_.convApiTiling.get_singleCoreWo() <= tilingData_.convApiTiling.get_woL1() &&129+ if (tilingData_.get_singleCoreWo() <= tilingData_.get_woL1() &&
130- !(ConvCeilDiv(tilingData_.convApiTiling.get_singleCoreWo(), tilingData_.convApiTiling.get_woL0()) > 1 &&130+ !(ConvCeilDiv(tilingData_.get_singleCoreWo(), tilingData_.get_woL0()) > 1 &&
131- tilingData_.convApiTiling.get_singleCoreWo() % convOpsConstParams_.m0 > 0 &&131+ tilingData_.get_singleCoreWo() % convOpsConstParams_.m0 > 0 &&
132- tilingData_.convApiTiling.get_hoL0() > 1)) {132+ tilingData_.get_hoL0() > 1)) {
133 woL1FullloadFlag = true;133 woL1FullloadFlag = true;
134 }134 }
135- bool hoL0FullloadFlag = tilingData_.convApiTiling.get_hoL1() == tilingData_.convApiTiling.get_hoL0();135+ bool hoL0FullloadFlag = tilingData_.get_hoL1() == tilingData_.get_hoL0();
136- bool woL0FullloadFlag = tilingData_.convApiTiling.get_woL1() == tilingData_.convApiTiling.get_woL0();136+ bool woL0FullloadFlag = tilingData_.get_woL1() == tilingData_.get_woL0();
137 if (kAL1FullloadFlag && hoL1FullloadFlag && woL1FullloadFlag) {137 if (kAL1FullloadFlag && hoL1FullloadFlag && woL1FullloadFlag) {
138 return FULLLOAD_AL1;138 return FULLLOAD_AL1;
139 } else if (!kAL1FullloadFlag && hoL1FullloadFlag && hoL0FullloadFlag && woL1FullloadFlag && woL0FullloadFlag) {139 } else if (!kAL1FullloadFlag && hoL1FullloadFlag && hoL0FullloadFlag && woL1FullloadFlag && woL0FullloadFlag) {
@@ -144,7 +144,7 @@ uint64_t Conv2dBaseTiling::GetFmpTilingValForHWSplit(bool kAL1FullloadFlag)
144 144 
145uint64_t Conv2dBaseTiling::GetL1PingPongVal()145uint64_t Conv2dBaseTiling::GetL1PingPongVal()
146{146{
147- uint64_t l1PingPong = static_cast<uint64_t>(tilingData_.convApiTiling.get_pBufferFlag() &147+ uint64_t l1PingPong = static_cast<uint64_t>(tilingData_.get_pBufferFlag() &
148 L1A_L1B_PB_FLAG_MASK) >> L1_PB_OFFSET;148 L1A_L1B_PB_FLAG_MASK) >> L1_PB_OFFSET;
149 // in group conv: only care about bl1 pingpong149 // in group conv: only care about bl1 pingpong
150 if (flagInfo_.convGroupType != ConvGroupType::NORMAL_CONV) {150 if (flagInfo_.convGroupType != ConvGroupType::NORMAL_CONV) {
@@ -160,7 +160,7 @@ uint64_t Conv2dBaseTiling::GetL1PingPongVal()
160uint64_t Conv2dBaseTiling::GetWeightUbTrans()160uint64_t Conv2dBaseTiling::GetWeightUbTrans()
161{161{
162 // bUbKStep is always 0 except weight ub trans mode.162 // bUbKStep is always 0 except weight ub trans mode.
163- if (tilingData_.convApiTiling.get_bUbNStep() > 0 && tilingData_.convApiTiling.get_bUbKStep() > 0) {163+ if (tilingData_.get_bUbNStep() > 0 && tilingData_.get_bUbKStep() > 0) {
164 return WEIGHT_UB_TRANS_OPEN;164 return WEIGHT_UB_TRANS_OPEN;
165 }165 }
166 166 
@@ -169,7 +169,7 @@ uint64_t Conv2dBaseTiling::GetWeightUbTrans()
169 169 
170uint64_t Conv2dBaseTiling::GetEnableInnerBatch()170uint64_t Conv2dBaseTiling::GetEnableInnerBatch()
171{171{
172- if (tilingData_.convApiTiling.get_innerBatch() > 1) {172+ if (tilingData_.get_innerBatch() > 1) {
173 return shapeInfo_.kh == 1 && shapeInfo_.kw == 1 && attrInfo_.padTop == 0 && attrInfo_.padBottom == 0 &&173 return shapeInfo_.kh == 1 && shapeInfo_.kw == 1 && attrInfo_.padTop == 0 && attrInfo_.padBottom == 0 &&
174 attrInfo_.padLeft == 0 && attrInfo_.padRight == 0 && attrInfo_.strideH == 1 && attrInfo_.strideW == 1 &&174 attrInfo_.padLeft == 0 && attrInfo_.padRight == 0 && attrInfo_.strideH == 1 && attrInfo_.strideW == 1 &&
175 attrInfo_.dilationH == 1 && attrInfo_.dilationW == 1 ?175 attrInfo_.dilationH == 1 && attrInfo_.dilationW == 1 ?
@@ -181,7 +181,7 @@ uint64_t Conv2dBaseTiling::GetEnableInnerBatch()
181uint64_t Conv2dBaseTiling::GetFmapCopyMode()181uint64_t Conv2dBaseTiling::GetFmapCopyMode()
182{182{
183 // bUbKStep is always 0 except weight ub trans mode.183 // bUbKStep is always 0 except weight ub trans mode.
184- if (tilingData_.convApiTiling.get_khUb() > 0 && tilingData_.convApiTiling.get_kwUb() > 0) {184+ if (tilingData_.get_khUb() > 0 && tilingData_.get_kwUb() > 0) {
185 return FMAP_DMA_MODE;185 return FMAP_DMA_MODE;
186 }186 }
187 187 
@@ -205,10 +205,10 @@ void Conv2dBaseTiling::ReSetTilingKeyPara()
205 tilingKeyPara_.fmpTiling = FMP_OTHER;205 tilingKeyPara_.fmpTiling = FMP_OTHER;
206 }206 }
207 bool sceneFlag = flagInfo_.convGroupType == ConvGroupType::OPT_GROUP_CONV &&207 bool sceneFlag = flagInfo_.convGroupType == ConvGroupType::OPT_GROUP_CONV &&
208- tilingKeyPara_.fmapCppyMode == FMAP_LOAD3D_MODE && tilingData_.convApiTiling.get_innerBatch() == 1 &&208+ tilingKeyPara_.fmapCppyMode == FMAP_LOAD3D_MODE && tilingData_.get_innerBatch() == 1 &&
209- tilingData_.convApiTiling.get_hoL1() == tilingData_.convApiTiling.get_hoL0();209+ tilingData_.get_hoL1() == tilingData_.get_hoL0();
210 if (!flagInfo_.mSplitModeFlag) {210 if (!flagInfo_.mSplitModeFlag) {
211- sceneFlag = sceneFlag && tilingData_.convApiTiling.get_woL1() == tilingData_.convApiTiling.get_woL0();211+ sceneFlag = sceneFlag && tilingData_.get_woL1() == tilingData_.get_woL0();
212 }212 }
213 bool otherFlag = tilingKeyPara_.iterOrder == 0 && tilingKeyPara_.l1PingPong == L1_PB_ALL_OPEN;213 bool otherFlag = tilingKeyPara_.iterOrder == 0 && tilingKeyPara_.l1PingPong == L1_PB_ALL_OPEN;
214 if (sceneFlag && otherFlag) {214 if (sceneFlag && otherFlag) {
@@ -238,7 +238,7 @@ ge::graphStatus Conv2dBaseTiling::SetTilingKey()
238 tilingKeyPara_.l1PingPong = GetL1PingPongVal();238 tilingKeyPara_.l1PingPong = GetL1PingPongVal();
239 tilingKeyPara_.l0PingPong = GetL0PingPongVal();239 tilingKeyPara_.l0PingPong = GetL0PingPongVal();
240 tilingKeyPara_.outputOrder = GetOutputOrderVal();240 tilingKeyPara_.outputOrder = GetOutputOrderVal();
241- tilingKeyPara_.iterOrder = static_cast<uint64_t>(tilingData_.convApiTiling.get_iterateMNOrder());241+ tilingKeyPara_.iterOrder = static_cast<uint64_t>(tilingData_.get_iterateMNOrder());
242 tilingKeyPara_.groupType = static_cast<uint64_t>(flagInfo_.convGroupType);242 tilingKeyPara_.groupType = static_cast<uint64_t>(flagInfo_.convGroupType);
243 tilingKeyPara_.enableSmallChannel = static_cast<uint64_t>(flagInfo_.enableC04Flag);243 tilingKeyPara_.enableSmallChannel = static_cast<uint64_t>(flagInfo_.enableC04Flag);
244 tilingKeyPara_.weightUbTrans = GetWeightUbTrans();244 tilingKeyPara_.weightUbTrans = GetWeightUbTrans();
@@ -246,7 +246,7 @@ ge::graphStatus Conv2dBaseTiling::SetTilingKey()
246 tilingKeyPara_.innerBatch = GetEnableInnerBatch();246 tilingKeyPara_.innerBatch = GetEnableInnerBatch();
247 tilingKeyPara_.disContinuous = static_cast<uint64_t>(flagInfo_.disContinuousFlag);247 tilingKeyPara_.disContinuous = static_cast<uint64_t>(flagInfo_.disContinuousFlag);
248 if (IsWeightNZFormat(descInfo_.weightFormat)) {248 if (IsWeightNZFormat(descInfo_.weightFormat)) {
249- tilingKeyPara_.batchOne = static_cast<uint64_t>(tilingData_.convApiTiling.get_singleCoreBatch() == 1);249+ tilingKeyPara_.batchOne = static_cast<uint64_t>(tilingData_.get_singleCoreBatch() == 1);
250 tilingKeyPara_.noPad = GetNoPad();250 tilingKeyPara_.noPad = GetNoPad();
251 tilingKeyPara_.smallWeight = GetSmallWeightVal();251 tilingKeyPara_.smallWeight = GetSmallWeightVal();
252 }252 }
Mconv/conv2d_v2/op_host/op_tiling/conv2d_v2_tiling.h+77-17
@@ -22,16 +22,76 @@
22#include "tiling/tiling_api.h"22#include "tiling/tiling_api.h"
23 23 
24/**24/**
25- * Conv2DRunInfo define here
26 * BEGIN_TILING_DATA_DEF and TILING_DATA_FIELD_DEF should same with CANN25 * BEGIN_TILING_DATA_DEF and TILING_DATA_FIELD_DEF should same with CANN
27*/26*/
28 27 
29namespace optiling {28namespace optiling {
30- BEGIN_TILING_DATA_DEF(Conv2DRunInfo)29+ 
30+ BEGIN_TILING_DATA_DEF(Conv2DTilingData)
31+ TILING_DATA_FIELD_DEF(uint64_t, orgHi);
32+ TILING_DATA_FIELD_DEF(uint64_t, orgWi);
33+ TILING_DATA_FIELD_DEF(uint64_t, orgHo);
34+ TILING_DATA_FIELD_DEF(uint64_t, orgWo);
35+ TILING_DATA_FIELD_DEF(uint64_t, orgHixWi);
36+ TILING_DATA_FIELD_DEF(uint64_t, singleCoreBatch);
37+ TILING_DATA_FIELD_DEF(uint64_t, singleCoreHo);
38+ TILING_DATA_FIELD_DEF(uint64_t, singleCoreWo);
39+ TILING_DATA_FIELD_DEF(uint64_t, unionDataXt);
31 TILING_DATA_FIELD_DEF(uint64_t, hin);40 TILING_DATA_FIELD_DEF(uint64_t, hin);
32 TILING_DATA_FIELD_DEF(uint64_t, win);41 TILING_DATA_FIELD_DEF(uint64_t, win);
33 TILING_DATA_FIELD_DEF(uint64_t, hout);42 TILING_DATA_FIELD_DEF(uint64_t, hout);
34 TILING_DATA_FIELD_DEF(uint64_t, wout);43 TILING_DATA_FIELD_DEF(uint64_t, wout);
44+ TILING_DATA_FIELD_DEF(uint32_t, orgCi);
45+ TILING_DATA_FIELD_DEF(uint32_t, orgCo);
46+ TILING_DATA_FIELD_DEF(uint32_t, singleCoreCi);
47+ TILING_DATA_FIELD_DEF(uint32_t, singleCoreCo);
48+ TILING_DATA_FIELD_DEF(uint32_t, hoL1);
49+ TILING_DATA_FIELD_DEF(uint32_t, woL1);
50+ TILING_DATA_FIELD_DEF(uint32_t, kAL1);
51+ TILING_DATA_FIELD_DEF(uint32_t, kBL1);
52+ TILING_DATA_FIELD_DEF(uint32_t, khL1);
53+ TILING_DATA_FIELD_DEF(uint32_t, kwL1);
54+ TILING_DATA_FIELD_DEF(uint32_t, nBL1);
55+ TILING_DATA_FIELD_DEF(uint32_t, hoL0);
56+ TILING_DATA_FIELD_DEF(uint32_t, woL0);
57+ TILING_DATA_FIELD_DEF(uint32_t, kL0);
58+ TILING_DATA_FIELD_DEF(uint32_t, nL0);
59+ TILING_DATA_FIELD_DEF(uint32_t, pBufferFlag);
60+ TILING_DATA_FIELD_DEF(uint32_t, groups);
61+ TILING_DATA_FIELD_DEF(uint32_t, enlarge);
62+ TILING_DATA_FIELD_DEF(uint32_t, singleCoreGroups);
63+ TILING_DATA_FIELD_DEF(uint32_t, singleCoreGroupOpt);
64+ TILING_DATA_FIELD_DEF(uint32_t, bUbNStep);
65+ TILING_DATA_FIELD_DEF(uint32_t, bUbKStep);
66+ TILING_DATA_FIELD_DEF(uint32_t, khUb);
67+ TILING_DATA_FIELD_DEF(uint32_t, kwUb);
68+ TILING_DATA_FIELD_DEF(uint32_t, kernelHxkernelW);
69+ TILING_DATA_FIELD_DEF(uint32_t, kernelHxkernelWxkernelD);
70+ TILING_DATA_FIELD_DEF(uint32_t, aL1SpaceSize);
71+ TILING_DATA_FIELD_DEF(uint32_t, multiNBL1);
72+ TILING_DATA_FIELD_DEF(uint32_t, cinAInCore);
73+ TILING_DATA_FIELD_DEF(uint32_t, cinATailInCore);
74+ TILING_DATA_FIELD_DEF(uint32_t, cinBInCore);
75+ TILING_DATA_FIELD_DEF(uint32_t, cinBTailInCore);
76+ TILING_DATA_FIELD_DEF(uint32_t, mStep);
77+ TILING_DATA_FIELD_DEF(uint32_t, kStep);
78+ TILING_DATA_FIELD_DEF(uint32_t, nStep);
79+ TILING_DATA_FIELD_DEF(uint32_t, fmapKStride);
80+ TILING_DATA_FIELD_DEF(uint32_t, weightKStride);
81+ TILING_DATA_FIELD_DEF(uint32_t, cinOffsetBlockInGM);
82+ TILING_DATA_FIELD_DEF(uint32_t, coutOffsetBlock);
83+ TILING_DATA_FIELD_DEF(uint32_t, nL1DivBlockSize);
84+ TILING_DATA_FIELD_DEF(uint32_t, kernelH);
85+ TILING_DATA_FIELD_DEF(uint32_t, kernelW);
86+ TILING_DATA_FIELD_DEF(uint32_t, strideH);
87+ TILING_DATA_FIELD_DEF(uint32_t, strideW);
88+ TILING_DATA_FIELD_DEF(uint32_t, dilationH);
89+ TILING_DATA_FIELD_DEF(uint32_t, dilationW);
90+ TILING_DATA_FIELD_DEF(uint32_t, padTop);
91+ TILING_DATA_FIELD_DEF(uint32_t, padBottom);
92+ TILING_DATA_FIELD_DEF(uint32_t, padLeft);
93+ TILING_DATA_FIELD_DEF(uint32_t, padRight);
94+ TILING_DATA_FIELD_DEF(uint32_t, innerBatch);
35 TILING_DATA_FIELD_DEF(uint32_t, batch);95 TILING_DATA_FIELD_DEF(uint32_t, batch);
36 TILING_DATA_FIELD_DEF(uint32_t, cin);96 TILING_DATA_FIELD_DEF(uint32_t, cin);
37 TILING_DATA_FIELD_DEF(uint32_t, cout);97 TILING_DATA_FIELD_DEF(uint32_t, cout);
@@ -42,25 +102,25 @@ namespace optiling {
42 TILING_DATA_FIELD_DEF(uint32_t, nDim);102 TILING_DATA_FIELD_DEF(uint32_t, nDim);
43 TILING_DATA_FIELD_DEF(uint32_t, hoDim);103 TILING_DATA_FIELD_DEF(uint32_t, hoDim);
44 TILING_DATA_FIELD_DEF(uint32_t, woDim);104 TILING_DATA_FIELD_DEF(uint32_t, woDim);
45- TILING_DATA_FIELD_DEF(uint32_t, strideH);
46- TILING_DATA_FIELD_DEF(uint32_t, strideW);
47- TILING_DATA_FIELD_DEF(uint32_t, dilationH);
48- TILING_DATA_FIELD_DEF(uint32_t, dilationW);
49- TILING_DATA_FIELD_DEF(uint32_t, padTop);
50- TILING_DATA_FIELD_DEF(uint32_t, padLeft);
51- TILING_DATA_FIELD_DEF(uint32_t, groups);
52- TILING_DATA_FIELD_DEF(uint32_t, enlarge);
53 TILING_DATA_FIELD_DEF(uint32_t, cinOpt);105 TILING_DATA_FIELD_DEF(uint32_t, cinOpt);
54 TILING_DATA_FIELD_DEF(uint32_t, coutOpt);106 TILING_DATA_FIELD_DEF(uint32_t, coutOpt);
55 TILING_DATA_FIELD_DEF(uint32_t, groupOpt);107 TILING_DATA_FIELD_DEF(uint32_t, groupOpt);
108+ TILING_DATA_FIELD_DEF(uint8_t, iterateMNOrder);
109+ TILING_DATA_FIELD_DEF(uint8_t, biasFullLoadFlag);
110+ TILING_DATA_FIELD_DEF(uint8_t, fixpParamsFullLoadFlag);
111+ TILING_DATA_FIELD_DEF(uint8_t, hf32Enable);
112+ TILING_DATA_FIELD_DEF(uint8_t, hf32TransMode);
56 TILING_DATA_FIELD_DEF(uint8_t, hasBias);113 TILING_DATA_FIELD_DEF(uint8_t, hasBias);
57- END_TILING_DATA_DEF;114+ TILING_DATA_FIELD_DEF(uint8_t, hasScale);
58- 115+ TILING_DATA_FIELD_DEF(uint8_t, dualOutput);
59- REGISTER_TILING_DATA_CLASS(Conv2DRunInfoOp, Conv2DRunInfo)116+ TILING_DATA_FIELD_DEF(uint8_t, quantMode0);
60- 117+ TILING_DATA_FIELD_DEF(uint8_t, reluMode0);
61- BEGIN_TILING_DATA_DEF(Conv2DTilingData)118+ TILING_DATA_FIELD_DEF(uint8_t, clipMode0);
62- TILING_DATA_FIELD_DEF_STRUCT(TConv2DTiling, convApiTiling);119+ TILING_DATA_FIELD_DEF(uint8_t, quantMode1);
63- TILING_DATA_FIELD_DEF_STRUCT(Conv2DRunInfo, convRunInfo);120+ TILING_DATA_FIELD_DEF(uint8_t, reluMode1);
121+ TILING_DATA_FIELD_DEF(uint8_t, clipMode1);
122+ TILING_DATA_FIELD_DEF(int8_t, offsetx);
123+ TILING_DATA_FIELD_DEF(int8_t, roundMode);
64 END_TILING_DATA_DEF;124 END_TILING_DATA_DEF;
65 125 
66 REGISTER_TILING_DATA_CLASS(Conv2DV2, Conv2DTilingData)126 REGISTER_TILING_DATA_CLASS(Conv2DV2, Conv2DTilingData)
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2.h+19-19
@@ -184,23 +184,23 @@ __aicore__ inline bool Conv2dBase<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYPE
184 184 
185 if constexpr (A_FORMAT == ConvFormat::NCHW) {185 if constexpr (A_FORMAT == ConvFormat::NCHW) {
186 if constexpr (isMMode) {186 if constexpr (isMMode) {
187- if (!InitSingleCoreData(convTilingData->convRunInfo.mDim, 1, 0, 0)) {187+ if (!InitSingleCoreData(convTilingData->mDim, 1, 0, 0)) {
188 return false;188 return false;
189 }189 }
190 } else {190 } else {
191- if (!InitSingleCoreData(convTilingData->convRunInfo.hoDim * convTilingData->convRunInfo.woDim, 0,191+ if (!InitSingleCoreData(convTilingData->hoDim * convTilingData->woDim, 0,
192- convTilingData->convRunInfo.woDim, 1)) {192+ convTilingData->woDim, 1)) {
193 return false;193 return false;
194 }194 }
195 }195 }
196 } else {196 } else {
197 if constexpr (isMMode) {197 if constexpr (isMMode) {
198- if (!InitSingleCoreData(1, convTilingData->convRunInfo.nDim, 0, 0)) {198+ if (!InitSingleCoreData(1, convTilingData->nDim, 0, 0)) {
199 return false;199 return false;
200 }200 }
201 } else {201 } else {
202- if (!InitSingleCoreData(1, 0, convTilingData->convRunInfo.woDim * convTilingData->convRunInfo.nDim,202+ if (!InitSingleCoreData(1, 0, convTilingData->woDim * convTilingData->nDim,
203- convTilingData->convRunInfo.nDim)) {203+ convTilingData->nDim)) {
204 return false;204 return false;
205 }205 }
206 }206 }
@@ -217,38 +217,38 @@ __aicore__ inline bool Conv2dBase<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYPE
217{217{
218 DimDataToFill batchToFill(singleCoreBatch, batchIdxStart, isBatchDimTail);218 DimDataToFill batchToFill(singleCoreBatch, batchIdxStart, isBatchDimTail);
219 bool isRealDim =219 bool isRealDim =
220- convCommon.CalcDimData(convTilingData->convRunInfo.hoDim * convTilingData->convRunInfo.nDim * convTilingData->convRunInfo.woDim,220+ convCommon.CalcDimData(convTilingData->hoDim * convTilingData->nDim * convTilingData->woDim,
221- convTilingData->convRunInfo.batchDim, convTilingData->convRunInfo.batch, convTilingData->convRunInfo.batch, batchToFill);221+ convTilingData->batchDim, convTilingData->batch, convTilingData->batch, batchToFill);
222 if (unlikely(!isRealDim)) {222 if (unlikely(!isRealDim)) {
223 return false;223 return false;
224 }224 }
225 225 
226 DimDataToFill nToFill(singleCoreN, nIdxStart, isNDimTail);226 DimDataToFill nToFill(singleCoreN, nIdxStart, isNDimTail);
227- isRealDim = convCommon.CalcNDimDataAlign(blockPerNDim, convTilingData->convRunInfo.nDim,227+ isRealDim = convCommon.CalcNDimDataAlign(blockPerNDim, convTilingData->nDim,
228- convTilingData->convRunInfo.cout, nToFill);228+ convTilingData->cout, nToFill);
229 if (unlikely(!isRealDim)) {229 if (unlikely(!isRealDim)) {
230 return false;230 return false;
231 }231 }
232 232 
233 if constexpr (isMMode) {233 if constexpr (isMMode) {
234 DimDataToFill mToFill(singleCoreM, mIdxStart, isMDimTail);234 DimDataToFill mToFill(singleCoreM, mIdxStart, isMDimTail);
235- uint64_t totalM = convTilingData->convRunInfo.hout * convTilingData->convRunInfo.wout;235+ uint64_t totalM = convTilingData->hout * convTilingData->wout;
236- isRealDim = convCommon.CalcDimData(blockPerMDim, convTilingData->convRunInfo.mDim, convCommon.AlignB(totalM, M0),236+ isRealDim = convCommon.CalcDimData(blockPerMDim, convTilingData->mDim, convCommon.AlignB(totalM, M0),
237 totalM, mToFill);237 totalM, mToFill);
238 if (unlikely(!isRealDim)) {238 if (unlikely(!isRealDim)) {
239 return false;239 return false;
240 }240 }
241 } else {241 } else {
242 DimDataToFill hoToFill(singleCoreHo, hoIdxStart, isHoDimTail);242 DimDataToFill hoToFill(singleCoreHo, hoIdxStart, isHoDimTail);
243- isRealDim = convCommon.CalcDimData(blockPerHoDim, convTilingData->convRunInfo.hoDim, convTilingData->convRunInfo.hout,243+ isRealDim = convCommon.CalcDimData(blockPerHoDim, convTilingData->hoDim, convTilingData->hout,
244- convTilingData->convRunInfo.hout, hoToFill);244+ convTilingData->hout, hoToFill);
245 if (unlikely(!isRealDim)) {245 if (unlikely(!isRealDim)) {
246 return false;246 return false;
247 }247 }
248 248 
249 DimDataToFill woToFill(singleCoreWo, woIdxStart, isWoDimTail);249 DimDataToFill woToFill(singleCoreWo, woIdxStart, isWoDimTail);
250- isRealDim = convCommon.CalcDimData(blockPerWoDim, convTilingData->convRunInfo.woDim, convTilingData->convRunInfo.wout,250+ isRealDim = convCommon.CalcDimData(blockPerWoDim, convTilingData->woDim, convTilingData->wout,
251- convTilingData->convRunInfo.wout, woToFill);251+ convTilingData->wout, woToFill);
252 if (unlikely(!isRealDim)) {252 if (unlikely(!isRealDim)) {
253 return false;253 return false;
254 }254 }
@@ -297,7 +297,7 @@ __aicore__ inline void Conv2dBase<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYPE
297 }297 }
298 298 
299 conv.SetWeight(filterGm);299 conv.SetWeight(filterGm);
300- if (convTilingData->convRunInfo.hasBias) {300+ if (convTilingData->hasBias) {
301 conv.SetBias(biasGm);301 conv.SetBias(biasGm);
302 }302 }
303 303 
@@ -309,10 +309,10 @@ __aicore__ inline void Conv2dBase<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYPE
309 309 
310 conv.SetFmap(fmapGm);310 conv.SetFmap(fmapGm);
311 if constexpr (CONV_CFG::disContinuous) {311 if constexpr (CONV_CFG::disContinuous) {
312- conv.SetOrgBatch(convTilingData->convRunInfo.batch);312+ conv.SetOrgBatch(convTilingData->batch);
313 }313 }
314 if constexpr (CONV_CFG::isExtendConv2d) {314 if constexpr (CONV_CFG::isExtendConv2d) {
315- if (convTilingData->convApiTiling.dualOutput) {315+ if (convTilingData->dualOutput) {
316 conv.IterateAll(outputGm, output1Gm);316 conv.IterateAll(outputGm, output1Gm);
317 } else {317 } else {
318 conv.IterateAll(outputGm);318 conv.IterateAll(outputGm);
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2_c04_impl.h+24-24
@@ -78,7 +78,7 @@ public:
78 } else {78 } else {
79 CrossCoreSetFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE1_MTE3);79 CrossCoreSetFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE1_MTE3);
80 // Each core deal with half of nBL1 at least n0, when nBL1 <= n0, vec1 doesn't have data to deal.80 // Each core deal with half of nBL1 at least n0, when nBL1 <= n0, vec1 doesn't have data to deal.
81- if (self->ctx.convTilingData->convApiTiling.nBL1 > BLOCK_L0_N) {81+ if (self->ctx.convTilingData->nBL1 > BLOCK_L0_N) {
82 CrossCoreSetFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE1_MTE3);82 CrossCoreSetFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE1_MTE3);
83 }83 }
84 }84 }
@@ -99,12 +99,12 @@ public:
99 self->ctx.ddr2l1LoopH = 1;99 self->ctx.ddr2l1LoopH = 1;
100 self->ctx.maxHoL1Iter = 0;100 self->ctx.maxHoL1Iter = 0;
101 } else {101 } else {
102- self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.hoL1);102+ self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->hoL1);
103 self->ctx.maxHoL1Iter = self->ctx.ddr2l1LoopH - 1;103 self->ctx.maxHoL1Iter = self->ctx.ddr2l1LoopH - 1;
104 }104 }
105 105 
106 if constexpr (Intf::hasWL1IterFlag) {106 if constexpr (Intf::hasWL1IterFlag) {
107- self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL1);107+ self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL1);
108 } else {108 } else {
109 self->ctx.ddr2l1LoopW = 1;109 self->ctx.ddr2l1LoopW = 1;
110 }110 }
@@ -127,21 +127,21 @@ public:
127 127 
128 __aicore__ inline void C04InitNValue(Intf *self)128 __aicore__ inline void C04InitNValue(Intf *self)
129 {129 {
130- self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->convApiTiling.nBL1;130+ self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->nBL1;
131- self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.nBL1 : self->ctx.nBL1Tail;131+ self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->nBL1 : self->ctx.nBL1Tail;
132 132 
133 if constexpr (!Intf::hasNL1IterFlag) {133 if constexpr (!Intf::hasNL1IterFlag) {
134 self->ctx.ddr2l1LoopN = 1;134 self->ctx.ddr2l1LoopN = 1;
135 self->ctx.maxNBL1Iter = 0;135 self->ctx.maxNBL1Iter = 0;
136 } else {136 } else {
137- self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->convApiTiling.nBL1);137+ self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->nBL1);
138 self->ctx.maxNBL1Iter = self->ctx.ddr2l1LoopN - 1;138 self->ctx.maxNBL1Iter = self->ctx.ddr2l1LoopN - 1;
139 }139 }
140 }140 }
141 141 
142 __aicore__ inline void C04InitBuf(Intf *self)142 __aicore__ inline void C04InitBuf(Intf *self)
143 {143 {
144- self->ctx.ubBufSize = self->ctx.convTilingData->convApiTiling.bUbNStep * self->ctx.convTilingData->convApiTiling.kBL1;144+ self->ctx.ubBufSize = self->ctx.convTilingData->bUbNStep * self->ctx.convTilingData->kBL1;
145 self->ctx.pipe.InitBuffer(self->ctx.ndUbBuf, self->ctx.ubBufSize * Intf::sizeOfWeight);145 self->ctx.pipe.InitBuffer(self->ctx.ndUbBuf, self->ctx.ubBufSize * Intf::sizeOfWeight);
146 self->ctx.pipe.InitBuffer(self->ctx.nzUbBuf, self->ctx.ubBufSize * Intf::sizeOfWeight);146 self->ctx.pipe.InitBuffer(self->ctx.nzUbBuf, self->ctx.ubBufSize * Intf::sizeOfWeight);
147 self->ctx.pipe.InitBuffer(self->ctx.indexUbBuf, REG_SIZE);147 self->ctx.pipe.InitBuffer(self->ctx.indexUbBuf, REG_SIZE);
@@ -149,11 +149,11 @@ public:
149 self->ctx.ndTensor = self->ctx.ndUbBuf.template Get<typename Intf::WeightT>();149 self->ctx.ndTensor = self->ctx.ndUbBuf.template Get<typename Intf::WeightT>();
150 self->ctx.nzTensor = self->ctx.nzUbBuf.template Get<typename Intf::WeightT>();150 self->ctx.nzTensor = self->ctx.nzUbBuf.template Get<typename Intf::WeightT>();
151 151 
152- uint32_t aL1SpaceSize = self->ctx.convTilingData->convApiTiling.aL1SpaceSize;152+ uint32_t aL1SpaceSize = self->ctx.convTilingData->aL1SpaceSize;
153- if ((self->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {153+ if ((self->ctx.convTilingData->pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {
154 aL1SpaceSize *= DOUBLE_BUF;154 aL1SpaceSize *= DOUBLE_BUF;
155 }155 }
156- self->ctx.bL1SpaceSize = self->ctx.convTilingData->convApiTiling.nBL1 * self->ctx.convTilingData->convApiTiling.kBL1;156+ self->ctx.bL1SpaceSize = self->ctx.convTilingData->nBL1 * self->ctx.convTilingData->kBL1;
157 157 
158 if constexpr (Intf::bL1DBFlag) {158 if constexpr (Intf::bL1DBFlag) {
159 self->ctx.pipe.InitBuffer(self->ctx.bL1TBuf,159 self->ctx.pipe.InitBuffer(self->ctx.bL1TBuf,
@@ -170,17 +170,17 @@ public:
170 {170 {
171 self->ctx.vecId = GetSubBlockIdx();171 self->ctx.vecId = GetSubBlockIdx();
172 172 
173- self->ctx.singleCoreCo = self->ctx.convTilingData->convApiTiling.singleCoreCo;173+ self->ctx.singleCoreCo = self->ctx.convTilingData->singleCoreCo;
174- self->ctx.kSizeC04 = self->ctx.convTilingData->convApiTiling.kernelHxkernelW * C04_CIN_SIZE;174+ self->ctx.kSizeC04 = self->ctx.convTilingData->kernelHxkernelW * C04_CIN_SIZE;
175- self->ctx.ddr2l1LoopBatch = self->ctx.convTilingData->convApiTiling.singleCoreBatch;175+ self->ctx.ddr2l1LoopBatch = self->ctx.convTilingData->singleCoreBatch;
176 176 
177 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {177 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {
178- self->ctx.singleCoreM = self->ctx.convTilingData->convApiTiling.singleCoreHo;178+ self->ctx.singleCoreM = self->ctx.convTilingData->singleCoreHo;
179- self->ctx.mAL1 = self->ctx.convTilingData->convApiTiling.hoL1;179+ self->ctx.mAL1 = self->ctx.convTilingData->hoL1;
180- self->ctx.mL0 = self->ctx.convTilingData->convApiTiling.hoL0;180+ self->ctx.mL0 = self->ctx.convTilingData->hoL0;
181 } else {181 } else {
182- self->ctx.singleCoreHo = self->ctx.convTilingData->convApiTiling.singleCoreHo;182+ self->ctx.singleCoreHo = self->ctx.convTilingData->singleCoreHo;
183- self->ctx.singleCoreWo = self->ctx.convTilingData->convApiTiling.singleCoreWo;183+ self->ctx.singleCoreWo = self->ctx.convTilingData->singleCoreWo;
184 }184 }
185 185 
186 C04InitBuf(self);186 C04InitBuf(self);
@@ -196,14 +196,14 @@ public:
196 __aicore__ inline void C04UpdateNUbValue(Intf *self)196 __aicore__ inline void C04UpdateNUbValue(Intf *self)
197 {197 {
198 self->ctx.currentUbNStep = self->ctx.vecNIter == self->ctx.maxVecNIter ?198 self->ctx.currentUbNStep = self->ctx.vecNIter == self->ctx.maxVecNIter ?
199- self->ctx.bUbNTailStep : self->ctx.convTilingData->convApiTiling.bUbNStep;199+ self->ctx.bUbNTailStep : self->ctx.convTilingData->bUbNStep;
200 self->ctx.currentUbNStepAilgn = AlignB(self->ctx.currentUbNStep, BLOCK_L0_N);200 self->ctx.currentUbNStepAilgn = AlignB(self->ctx.currentUbNStep, BLOCK_L0_N);
201 }201 }
202 202 
203 __aicore__ inline void C04UpdateNL1Value(Intf *self)203 __aicore__ inline void C04UpdateNL1Value(Intf *self)
204 {204 {
205 self->ctx.currentNBL1 = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?205 self->ctx.currentNBL1 = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?
206- self->ctx.nBL1Tail : self->ctx.convTilingData->convApiTiling.nBL1;206+ self->ctx.nBL1Tail : self->ctx.convTilingData->nBL1;
207 self->ctx.currentNBL1Align = AlignB(self->ctx.currentNBL1, BLOCK_L0_N);207 self->ctx.currentNBL1Align = AlignB(self->ctx.currentNBL1, BLOCK_L0_N);
208 208 
209 if constexpr (!Intf::bL1DBFlag) {209 if constexpr (!Intf::bL1DBFlag) {
@@ -219,10 +219,10 @@ public:
219 }219 }
220 }220 }
221 221 
222- self->ctx.vecNLoopTimes = CeilDiv(self->ctx.currentNBL1, self->ctx.convTilingData->convApiTiling.bUbNStep);222+ self->ctx.vecNLoopTimes = CeilDiv(self->ctx.currentNBL1, self->ctx.convTilingData->bUbNStep);
223 self->ctx.maxVecNIter = self->ctx.vecNLoopTimes - 1;223 self->ctx.maxVecNIter = self->ctx.vecNLoopTimes - 1;
224- self->ctx.bUbNTailStep = self->ctx.currentNBL1 % self->ctx.convTilingData->convApiTiling.bUbNStep;224+ self->ctx.bUbNTailStep = self->ctx.currentNBL1 % self->ctx.convTilingData->bUbNStep;
225- self->ctx.bUbNTailStep = self->ctx.bUbNTailStep == 0 ? self->ctx.convTilingData->convApiTiling.bUbNStep : self->ctx.bUbNTailStep;225+ self->ctx.bUbNTailStep = self->ctx.bUbNTailStep == 0 ? self->ctx.convTilingData->bUbNStep : self->ctx.bUbNTailStep;
226 226 
227 self->ctx.currentNLoopRpSize =227 self->ctx.currentNLoopRpSize =
228 self->ctx.nBL1Iter == self->ctx.maxNBL1Iter && self->ctx.vecNIter == self->ctx.maxVecNIter ?228 self->ctx.nBL1Iter == self->ctx.maxNBL1Iter && self->ctx.vecNIter == self->ctx.maxVecNIter ?
@@ -330,7 +330,7 @@ public:
330 __aicore__ inline bool C04VecImpl(Intf *self)330 __aicore__ inline bool C04VecImpl(Intf *self)
331 {331 {
332 if constexpr (!Intf::bL1DBFlag) {332 if constexpr (!Intf::bL1DBFlag) {
333- if (self->ctx.convTilingData->convApiTiling.nBL1 <= BLOCK_L0_N && self->ctx.vecId == 1) {333+ if (self->ctx.convTilingData->nBL1 <= BLOCK_L0_N && self->ctx.vecId == 1) {
334 return false;334 return false;
335 }335 }
336 }336 }
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2_common_func.h+48-48
@@ -47,14 +47,14 @@ using TypeFalse = struct {
47template <class Intf>47template <class Intf>
48static __aicore__ inline void InitKDirectionValueC04(Intf *self)48static __aicore__ inline void InitKDirectionValueC04(Intf *self)
49{49{
50- size_t multiKL1 = self->ctx.convTilingData->convApiTiling .kAL1 / self->ctx.convTilingData->convApiTiling .kL0;50+ size_t multiKL1 = self->ctx.convTilingData->kAL1 / self->ctx.convTilingData->kL0;
51 self->ctx.maxKAL1Iter = 0;51 self->ctx.maxKAL1Iter = 0;
52 self->ctx.maxKBL1Iter = 0;52 self->ctx.maxKBL1Iter = 0;
53 self->ctx.maxKL0Iter = multiKL1 - 1;53 self->ctx.maxKL0Iter = multiKL1 - 1;
54 self->ctx.ddr2l0LoopK = multiKL1;54 self->ctx.ddr2l0LoopK = multiKL1;
55- self->ctx.kAL1Tail = self->ctx.convTilingData->convApiTiling .kAL1;55+ self->ctx.kAL1Tail = self->ctx.convTilingData->kAL1;
56- self->ctx.kBL1Tail = self->ctx.convTilingData->convApiTiling .kBL1;56+ self->ctx.kBL1Tail = self->ctx.convTilingData->kBL1;
57- self->ctx.kL0Tail = self->ctx.convTilingData->convApiTiling .kL0;57+ self->ctx.kL0Tail = self->ctx.convTilingData->kL0;
58 self->ctx.multiKAL1 = multiKL1;58 self->ctx.multiKAL1 = multiKL1;
59 self->ctx.multiKBL1 = multiKL1;59 self->ctx.multiKBL1 = multiKL1;
60 self->ctx.kAL1fullload = true;60 self->ctx.kAL1fullload = true;
@@ -69,43 +69,43 @@ static __aicore__ inline void InitKDirectionValue(Intf *self)
69 InitKDirectionValueC04(self);69 InitKDirectionValueC04(self);
70 return;70 return;
71 }71 }
72- uint64_t totalKAlignK0 = AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling .kernelHxkernelW;72+ uint64_t totalKAlignK0 = AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->kernelHxkernelW;
73- self->ctx.maxKAL1Iter = CeilDiv(totalKAlignK0, self->ctx.convTilingData->convApiTiling .kAL1) - 1;73+ self->ctx.maxKAL1Iter = CeilDiv(totalKAlignK0, self->ctx.convTilingData->kAL1) - 1;
74- self->ctx.maxKBL1Iter = CeilDiv(totalKAlignK0, self->ctx.convTilingData->convApiTiling .kBL1) - 1;74+ self->ctx.maxKBL1Iter = CeilDiv(totalKAlignK0, self->ctx.convTilingData->kBL1) - 1;
75- self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTilingData->convApiTiling .kL0);75+ self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTilingData->kL0);
76 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;76 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;
77- self->ctx.kAL1Tail = (self->ctx.singleCoreCi * self->ctx.convTilingData->convApiTiling .kernelHxkernelW) % self->ctx.convTilingData->convApiTiling .kAL1;77+ self->ctx.kAL1Tail = (self->ctx.singleCoreCi * self->ctx.convTilingData->kernelHxkernelW) % self->ctx.convTilingData->kAL1;
78- self->ctx.kAL1Tail = self->ctx.kAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling .kAL1 : self->ctx.kAL1Tail;78+ self->ctx.kAL1Tail = self->ctx.kAL1Tail == 0 ? self->ctx.convTilingData->kAL1 : self->ctx.kAL1Tail;
79- self->ctx.kBL1Tail = (self->ctx.singleCoreCi * self->ctx.convTilingData->convApiTiling .kernelHxkernelW) % self->ctx.convTilingData->convApiTiling .kBL1;79+ self->ctx.kBL1Tail = (self->ctx.singleCoreCi * self->ctx.convTilingData->kernelHxkernelW) % self->ctx.convTilingData->kBL1;
80- self->ctx.kBL1Tail = self->ctx.kBL1Tail == 0 ? self->ctx.convTilingData->convApiTiling .kBL1 : self->ctx.kBL1Tail;80+ self->ctx.kBL1Tail = self->ctx.kBL1Tail == 0 ? self->ctx.convTilingData->kBL1 : self->ctx.kBL1Tail;
81- self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTilingData->convApiTiling .kL0;81+ self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTilingData->kL0;
82 if constexpr (Intf::k0 != Intf::k0FmapTail) {82 if constexpr (Intf::k0 != Intf::k0FmapTail) {
83- self->ctx.kAL0Tail = AlignB(self->ctx.singleCoreCi, Intf::k0FmapTail) * self->ctx.convTilingData->convApiTiling .kernelHxkernelW %83+ self->ctx.kAL0Tail = AlignB(self->ctx.singleCoreCi, Intf::k0FmapTail) * self->ctx.convTilingData->kernelHxkernelW %
84- self->ctx.convTilingData->convApiTiling .kL0;84+ self->ctx.convTilingData->kL0;
85- self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTilingData->convApiTiling .kL0 : self->ctx.kAL0Tail;85+ self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTilingData->kL0 : self->ctx.kAL0Tail;
86 }86 }
87- self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTilingData->convApiTiling .kL0 : self->ctx.kL0Tail;87+ self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTilingData->kL0 : self->ctx.kL0Tail;
88- self->ctx.multiKAL1 = CeilDiv(self->ctx.convTilingData->convApiTiling .kAL1, self->ctx.convTilingData->convApiTiling .kL0);88+ self->ctx.multiKAL1 = CeilDiv(self->ctx.convTilingData->kAL1, self->ctx.convTilingData->kL0);
89- self->ctx.multiKBL1 = CeilDiv(self->ctx.convTilingData->convApiTiling .kBL1, self->ctx.convTilingData->convApiTiling .kL0);89+ self->ctx.multiKBL1 = CeilDiv(self->ctx.convTilingData->kBL1, self->ctx.convTilingData->kL0);
90 90 
91 if constexpr (Intf::kPreLoadFlag || Intf::WEIGHT_NZ_FLAG) {91 if constexpr (Intf::kPreLoadFlag || Intf::WEIGHT_NZ_FLAG) {
92- self->ctx.kBL1AlignK0Tail = totalKAlignK0 % self->ctx.convTilingData->convApiTiling .kBL1;92+ self->ctx.kBL1AlignK0Tail = totalKAlignK0 % self->ctx.convTilingData->kBL1;
93 self->ctx.kBL1AlignK0Tail =93 self->ctx.kBL1AlignK0Tail =
94- self->ctx.kBL1AlignK0Tail == 0 ? self->ctx.convTilingData->convApiTiling .kBL1 : self->ctx.kBL1AlignK0Tail;94+ self->ctx.kBL1AlignK0Tail == 0 ? self->ctx.convTilingData->kBL1 : self->ctx.kBL1AlignK0Tail;
95 }95 }
96 if constexpr (Intf::kPreLoadFlag) {96 if constexpr (Intf::kPreLoadFlag) {
97- self->ctx.kAL1AlignK0Tail = totalKAlignK0 % self->ctx.convTilingData->convApiTiling .kAL1;97+ self->ctx.kAL1AlignK0Tail = totalKAlignK0 % self->ctx.convTilingData->kAL1;
98 self->ctx.kAL1AlignK0Tail =98 self->ctx.kAL1AlignK0Tail =
99- self->ctx.kAL1AlignK0Tail == 0 ? self->ctx.convTilingData->convApiTiling .kAL1 : self->ctx.kAL1AlignK0Tail;99+ self->ctx.kAL1AlignK0Tail == 0 ? self->ctx.convTilingData->kAL1 : self->ctx.kAL1AlignK0Tail;
100 self->ctx.lastLoopKAL1StartPos = self->ctx.ddr2l0LoopK - self->ctx.multiKAL1;100 self->ctx.lastLoopKAL1StartPos = self->ctx.ddr2l0LoopK - self->ctx.multiKAL1;
101 self->ctx.lastLoopKBL1StartPos = self->ctx.ddr2l0LoopK - self->ctx.multiKBL1;101 self->ctx.lastLoopKBL1StartPos = self->ctx.ddr2l0LoopK - self->ctx.multiKBL1;
102 self->ctx.lastLoopKAL1StartPosTail =102 self->ctx.lastLoopKAL1StartPosTail =
103- self->ctx.ddr2l0LoopK - CeilDiv(self->ctx.kAL1AlignK0Tail, self->ctx.convTilingData->convApiTiling .kL0);103+ self->ctx.ddr2l0LoopK - CeilDiv(self->ctx.kAL1AlignK0Tail, self->ctx.convTilingData->kL0);
104 self->ctx.lastLoopKBL1StartPosTail =104 self->ctx.lastLoopKBL1StartPosTail =
105- self->ctx.ddr2l0LoopK - CeilDiv(self->ctx.kBL1AlignK0Tail, self->ctx.convTilingData->convApiTiling .kL0);105+ self->ctx.ddr2l0LoopK - CeilDiv(self->ctx.kBL1AlignK0Tail, self->ctx.convTilingData->kL0);
106 }106 }
107 if constexpr (Intf::isDmaFlag) {107 if constexpr (Intf::isDmaFlag) {
108- self->ctx.cinBL1 = self->ctx.convTilingData->convApiTiling .kBL1 / (self->ctx.convTilingData->convApiTiling .khL1 * self->ctx.convTilingData->convApiTiling .kwL1);108+ self->ctx.cinBL1 = self->ctx.convTilingData->kBL1 / (self->ctx.convTilingData->khL1 * self->ctx.convTilingData->kwL1);
109 self->ctx.cinBL1LoopTimes = CeilDiv(self->ctx.singleCoreCi, self->ctx.cinBL1);109 self->ctx.cinBL1LoopTimes = CeilDiv(self->ctx.singleCoreCi, self->ctx.cinBL1);
110 self->ctx.maxCinBL1Iter = self->ctx.cinBL1LoopTimes - 1;110 self->ctx.maxCinBL1Iter = self->ctx.cinBL1LoopTimes - 1;
111 }111 }
@@ -116,7 +116,7 @@ struct Init {
116 static __aicore__ inline void call(Intf *self, const void *__restrict convTiling)116 static __aicore__ inline void call(Intf *self, const void *__restrict convTiling)
117 {117 {
118 self->ctx.convTilingData = (Conv2DTilingData *)convTiling;118 self->ctx.convTilingData = (Conv2DTilingData *)convTiling;
119- self->ctx.singleCoreBatch = self->ctx.convTilingData->convApiTiling .singleCoreBatch;119+ self->ctx.singleCoreBatch = self->ctx.convTilingData->singleCoreBatch;
120 InitBatchDirectionValue<Intf>(self);120 InitBatchDirectionValue<Intf>(self);
121 if ASCEND_IS_AIC_CONV {121 if ASCEND_IS_AIC_CONV {
122 InitBaseValue(self);122 InitBaseValue(self);
@@ -134,16 +134,16 @@ struct Init {
134 InitSubApiParams<Intf>(self);134 InitSubApiParams<Intf>(self);
135 135 
136 if constexpr (Intf::groupOptPreloadFlag) {136 if constexpr (Intf::groupOptPreloadFlag) {
137- self->ctx.singleGroups = self->ctx.convTilingData->convApiTiling.singleCoreGroups;137+ self->ctx.singleGroups = self->ctx.convTilingData->singleCoreGroups;
138- self->ctx.singleGroupOpt = self->ctx.convTilingData->convApiTiling.singleCoreGroupOpt;138+ self->ctx.singleGroupOpt = self->ctx.convTilingData->singleCoreGroupOpt;
139- self->ctx.ciPerGroup = self->ctx.convTilingData->convApiTiling.orgCi / self->ctx.convTilingData->convApiTiling.groups;139+ self->ctx.ciPerGroup = self->ctx.convTilingData->orgCi / self->ctx.convTilingData->groups;
140 if ASCEND_IS_AIC_CONV {140 if ASCEND_IS_AIC_CONV {
141 InitKDirectionValue<Intf>(self);141 InitKDirectionValue<Intf>(self);
142 OptGroupCalcBL1LoadTimes<Intf>(self);142 OptGroupCalcBL1LoadTimes<Intf>(self);
143 }143 }
144 } else if constexpr (Intf::groupOptFlag) {144 } else if constexpr (Intf::groupOptFlag) {
145- self->ctx.ciPerGroup = self->ctx.convTilingData->convApiTiling.orgCi / self->ctx.convTilingData->convApiTiling.groups;145+ self->ctx.ciPerGroup = self->ctx.convTilingData->orgCi / self->ctx.convTilingData->groups;
146- self->ctx.singleGroupOpt = self->ctx.convTilingData->convApiTiling.singleCoreGroupOpt;146+ self->ctx.singleGroupOpt = self->ctx.convTilingData->singleCoreGroupOpt;
147 } else if constexpr (Intf::c04NDFlag) {147 } else if constexpr (Intf::c04NDFlag) {
148 self->ctx.c04ProcessTools.C04InitIterValue(self);148 self->ctx.c04ProcessTools.C04InitIterValue(self);
149 } else if constexpr (Intf::weightUbTrans) {149 } else if constexpr (Intf::weightUbTrans) {
@@ -174,32 +174,32 @@ struct Init {
174 174 
175 static __aicore__ inline void InitBaseValue(Intf *self)175 static __aicore__ inline void InitBaseValue(Intf *self)
176 {176 {
177- self->ctx.dilatedKernelH = 1 + (self->ctx.convTilingData->convApiTiling.kernelH - 1) * self->ctx.convTilingData->convApiTiling.dilationH;177+ self->ctx.dilatedKernelH = 1 + (self->ctx.convTilingData->kernelH - 1) * self->ctx.convTilingData->dilationH;
178- self->ctx.dilatedKernelW = 1 + (self->ctx.convTilingData->convApiTiling.kernelW - 1) * self->ctx.convTilingData->convApiTiling.dilationW;178+ self->ctx.dilatedKernelW = 1 + (self->ctx.convTilingData->kernelW - 1) * self->ctx.convTilingData->dilationW;
179- self->ctx.singleCoreCi = self->ctx.convTilingData->convApiTiling.singleCoreCi; // Cin Size in Single Core179+ self->ctx.singleCoreCi = self->ctx.convTilingData->singleCoreCi; // Cin Size in Single Core
180 uint64_t alignCinKhKw =180 uint64_t alignCinKhKw =
181- AlignB(self->ctx.convTilingData->convApiTiling.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelW;181+ AlignB(self->ctx.convTilingData->singleCoreCi, Intf::k0) * self->ctx.convTilingData->kernelHxkernelW;
182- self->ctx.kAL1fullload = alignCinKhKw == self->ctx.convTilingData->convApiTiling.kAL1;182+ self->ctx.kAL1fullload = alignCinKhKw == self->ctx.convTilingData->kAL1;
183- self->ctx.kBL1fullload = alignCinKhKw == self->ctx.convTilingData->convApiTiling.kBL1;183+ self->ctx.kBL1fullload = alignCinKhKw == self->ctx.convTilingData->kBL1;
184- self->ctx.singleCoreCo = self->ctx.convTilingData->convApiTiling.singleCoreCo; // Cout Size in Single Core184+ self->ctx.singleCoreCo = self->ctx.convTilingData->singleCoreCo; // Cout Size in Single Core
185- self->ctx.fmapOneBatchSize = self->ctx.convTilingData->convApiTiling.orgCi * self->ctx.convTilingData->convApiTiling.orgHi *185+ self->ctx.fmapOneBatchSize = self->ctx.convTilingData->orgCi * self->ctx.convTilingData->orgHi *
186- self->ctx.convTilingData->convApiTiling.orgWi;186+ self->ctx.convTilingData->orgWi;
187- self->ctx.outputOneBatchSize = self->ctx.convTilingData->convApiTiling.orgCo * self->ctx.convTilingData->convApiTiling.orgHo *187+ self->ctx.outputOneBatchSize = self->ctx.convTilingData->orgCo * self->ctx.convTilingData->orgHo *
188- self->ctx.convTilingData->convApiTiling.orgWo;188+ self->ctx.convTilingData->orgWo;
189 }189 }
190 190 
191 static __aicore__ inline void InitValueMMode(Intf *self)191 static __aicore__ inline void InitValueMMode(Intf *self)
192 {192 {
193- self->ctx.mAL1 = self->ctx.convTilingData->convApiTiling.hoL1;193+ self->ctx.mAL1 = self->ctx.convTilingData->hoL1;
194- self->ctx.mL0 = self->ctx.convTilingData->convApiTiling.hoL0;194+ self->ctx.mL0 = self->ctx.convTilingData->hoL0;
195- self->ctx.singleCoreM = self->ctx.convTilingData->convApiTiling.singleCoreHo;195+ self->ctx.singleCoreM = self->ctx.convTilingData->singleCoreHo;
196 InitMDirectionValue<Intf>(self);196 InitMDirectionValue<Intf>(self);
197 }197 }
198 198 
199 static __aicore__ inline void InitValueHWMode(Intf *self)199 static __aicore__ inline void InitValueHWMode(Intf *self)
200 {200 {
201- self->ctx.singleCoreHo = self->ctx.convTilingData->convApiTiling.singleCoreHo; // Ho Size in Single Core201+ self->ctx.singleCoreHo = self->ctx.convTilingData->singleCoreHo; // Ho Size in Single Core
202- self->ctx.singleCoreWo = self->ctx.convTilingData->convApiTiling.singleCoreWo; // Wo Size in Single Core202+ self->ctx.singleCoreWo = self->ctx.convTilingData->singleCoreWo; // Wo Size in Single Core
203 InitHoDirectionValue<Intf>(self);203 InitHoDirectionValue<Intf>(self);
204 InitWoDirectionValue<Intf>(self);204 InitWoDirectionValue<Intf>(self);
205 }205 }
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2_dma_impl.h+46-46
@@ -66,7 +66,7 @@ public:
66 }66 }
67 67
68 CrossCoreSetFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE1_MTE3);68 CrossCoreSetFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE1_MTE3);
69- if (self->ctx.convTilingData->convApiTiling.cinAInCore > Intf::k0) {69+ if (self->ctx.convTilingData->cinAInCore > Intf::k0) {
70 CrossCoreSetFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE1_MTE3);70 CrossCoreSetFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE1_MTE3);
71 }71 }
72 }72 }
@@ -74,7 +74,7 @@ public:
74 __aicore__ inline void DmaSyncWait(Intf *self)74 __aicore__ inline void DmaSyncWait(Intf *self)
75 {75 {
76 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE3_MTE1);76 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(CV_SYNC_ID_MTE3_MTE1);
77- if (self->ctx.convTilingData->convApiTiling.cinAInCore > Intf::k0) {77+ if (self->ctx.convTilingData->cinAInCore > Intf::k0) {
78 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);78 CrossCoreWaitFlag<CV_ENHANCE_MODE, PIPE_MTE1>(VEC_ID_MAX + CV_SYNC_ID_MTE3_MTE1);
79 }79 }
80 self->ctx.loadUB2L1Iter++;80 self->ctx.loadUB2L1Iter++;
@@ -83,9 +83,9 @@ public:
83 __aicore__ inline void DmaCalcAL1LoadTimes(Intf *self)83 __aicore__ inline void DmaCalcAL1LoadTimes(Intf *self)
84 {84 {
85 if (!self->ctx.kAL1fullload) {85 if (!self->ctx.kAL1fullload) {
86- uint64_t ddr2L0LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->convApiTiling.nL0);86+ uint64_t ddr2L0LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->nL0);
87- uint64_t ddr2L0LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.hoL0);87+ uint64_t ddr2L0LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->hoL0);
88- uint64_t ddr2L0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL0);88+ uint64_t ddr2L0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL0);
89 self->ctx.ddr2l1LoopKA = self->ctx.maxKAL1Iter + 1;89 self->ctx.ddr2l1LoopKA = self->ctx.maxKAL1Iter + 1;
90 90
91 self->ctx.aL1LoadTimes = ddr2L0LoopN * ddr2L0LoopH * ddr2L0LoopW * self->ctx.ddr2l1LoopKA *91 self->ctx.aL1LoadTimes = ddr2L0LoopN * ddr2L0LoopH * ddr2L0LoopW * self->ctx.ddr2l1LoopKA *
@@ -103,59 +103,59 @@ public:
103 __aicore__ inline void DmaUpdateHWValue(Intf *self)103 __aicore__ inline void DmaUpdateHWValue(Intf *self)
104 {104 {
105 self->ctx.currentHoL1 = self->ctx.hoAL1Iter == self->ctx.maxHoL1Iter ?105 self->ctx.currentHoL1 = self->ctx.hoAL1Iter == self->ctx.maxHoL1Iter ?
106- self->ctx.hoAL1Tail : self->ctx.convTilingData->convApiTiling.hoL1;106+ self->ctx.hoAL1Tail : self->ctx.convTilingData->hoL1;
107 107
108 self->ctx.currentWoL1 = self->ctx.woAL1Iter == self->ctx.maxWoL1Iter ?108 self->ctx.currentWoL1 = self->ctx.woAL1Iter == self->ctx.maxWoL1Iter ?
109- self->ctx.woAL1Tail : self->ctx.convTilingData->convApiTiling.woL1;109+ self->ctx.woAL1Tail : self->ctx.convTilingData->woL1;
110 110
111 self->ctx.currentHoL1xWoL1Align = self->ctx.currentHoL1 * AlignB(self->ctx.currentWoL1, BLOCK_L0_M);111 self->ctx.currentHoL1xWoL1Align = self->ctx.currentHoL1 * AlignB(self->ctx.currentWoL1, BLOCK_L0_M);
112 112
113 if (!self->ctx.kAL1fullload) {113 if (!self->ctx.kAL1fullload) {
114- self->ctx.l12l0LoopH = CeilDiv(self->ctx.currentHoL1, self->ctx.convTilingData->convApiTiling.hoL0);114+ self->ctx.l12l0LoopH = CeilDiv(self->ctx.currentHoL1, self->ctx.convTilingData->hoL0);
115- self->ctx.l12l0LoopW = CeilDiv(self->ctx.currentWoL1, self->ctx.convTilingData->convApiTiling.woL0);115+ self->ctx.l12l0LoopW = CeilDiv(self->ctx.currentWoL1, self->ctx.convTilingData->woL0);
116 self->ctx.ddr2l1LoopInner = self->ctx.ddr2l1LoopTmp * self->ctx.l12l0LoopH * self->ctx.l12l0LoopW;116 self->ctx.ddr2l1LoopInner = self->ctx.ddr2l1LoopTmp * self->ctx.l12l0LoopH * self->ctx.l12l0LoopW;
117 }117 }
118 }118 }
119 119
120 __aicore__ inline void DmaInitKValue(Intf *self)120 __aicore__ inline void DmaInitKValue(Intf *self)
121 {121 {
122- self->ctx.kAL1Tail = (AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelW) % self->ctx.convTilingData->convApiTiling.kAL1;122+ self->ctx.kAL1Tail = (AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->kernelHxkernelW) % self->ctx.convTilingData->kAL1;
123- self->ctx.kAL1Tail = self->ctx.kAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.kAL1 : self->ctx.kAL1Tail;123+ self->ctx.kAL1Tail = self->ctx.kAL1Tail == 0 ? self->ctx.convTilingData->kAL1 : self->ctx.kAL1Tail;
124- if (self->ctx.vecId == 0 && CeilDiv(self->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0) > 1) {124+ if (self->ctx.vecId == 0 && CeilDiv(self->ctx.convTilingData->cinATailInCore, Intf::k0) > 1) {
125 self->ctx.ciTail = Intf::k0;125 self->ctx.ciTail = Intf::k0;
126 } else {126 } else {
127 self->ctx.ciTail = self->ctx.singleCoreCi % Intf::k0;127 self->ctx.ciTail = self->ctx.singleCoreCi % Intf::k0;
128 self->ctx.ciTail = self->ctx.ciTail == 0 ? Intf::k0 : self->ctx.ciTail;128 self->ctx.ciTail = self->ctx.ciTail == 0 ? Intf::k0 : self->ctx.ciTail;
129 }129 }
130 130
131- self->ctx.ddr2l1LoopKA = CeilDiv(AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelW,131+ self->ctx.ddr2l1LoopKA = CeilDiv(AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->kernelHxkernelW,
132- self->ctx.convTilingData->convApiTiling.kAL1);132+ self->ctx.convTilingData->kAL1);
133 self->ctx.maxKAL1Iter = self->ctx.ddr2l1LoopKA - 1;133 self->ctx.maxKAL1Iter = self->ctx.ddr2l1LoopKA - 1;
134 self->ctx.kAL1fullload = self->ctx.ddr2l1LoopKA == 1;134 self->ctx.kAL1fullload = self->ctx.ddr2l1LoopKA == 1;
135 135
136- uint32_t currentCi1 = self->ctx.convTilingData->convApiTiling.cinAInCore / Intf::k0;136+ uint32_t currentCi1 = self->ctx.convTilingData->cinAInCore / Intf::k0;
137 self->ctx.currentCi1Ub = CeilDiv(currentCi1, VEC_NUM);137 self->ctx.currentCi1Ub = CeilDiv(currentCi1, VEC_NUM);
138 138
139 if (self->ctx.vecId == 1) {139 if (self->ctx.vecId == 1) {
140 self->ctx.currentVec0Ci = self->ctx.currentCi1Ub * Intf::k0;140 self->ctx.currentVec0Ci = self->ctx.currentCi1Ub * Intf::k0;
141 self->ctx.currentCi1Ub = currentCi1 - self->ctx.currentCi1Ub;141 self->ctx.currentCi1Ub = currentCi1 - self->ctx.currentCi1Ub;
142- self->ctx.vec0TotalSize = self->ctx.convTilingData->convApiTiling.hoL1 * self->ctx.convTilingData->convApiTiling.woL1 * self->ctx.currentVec0Ci *142+ self->ctx.vec0TotalSize = self->ctx.convTilingData->hoL1 * self->ctx.convTilingData->woL1 * self->ctx.currentVec0Ci *
143- self->ctx.convTilingData->convApiTiling.khL1 * self->ctx.convTilingData->convApiTiling.kwL1;143+ self->ctx.convTilingData->khL1 * self->ctx.convTilingData->kwL1;
144 }144 }
145 self->ctx.maxVecCi1Iter = self->ctx.currentCi1Ub - 1;145 self->ctx.maxVecCi1Iter = self->ctx.currentCi1Ub - 1;
146 }146 }
147 147
148 __aicore__ inline void DmaInitHWValue(Intf *self)148 __aicore__ inline void DmaInitHWValue(Intf *self)
149 {149 {
150- self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->convApiTiling.woL1;150+ self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->woL1;
151- self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL1 : self->ctx.woAL1Tail;151+ self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->woL1 : self->ctx.woAL1Tail;
152 152
153- self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL1);153+ self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->woL1);
154 self->ctx.maxWoL1Iter = self->ctx.ddr2l1LoopW - 1;154 self->ctx.maxWoL1Iter = self->ctx.ddr2l1LoopW - 1;
155 155
156- self->ctx.hoAL1Tail = self->ctx.singleCoreHo % self->ctx.convTilingData->convApiTiling.hoL1;156+ self->ctx.hoAL1Tail = self->ctx.singleCoreHo % self->ctx.convTilingData->hoL1;
157- self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.hoL1 : self->ctx.hoAL1Tail;157+ self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTilingData->hoL1 : self->ctx.hoAL1Tail;
158- self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.hoL1);158+ self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->hoL1);
159 self->ctx.maxHoL1Iter = self->ctx.ddr2l1LoopH - 1;159 self->ctx.maxHoL1Iter = self->ctx.ddr2l1LoopH - 1;
160 160
161 DmaUpdateHWValue(self);161 DmaUpdateHWValue(self);
@@ -164,31 +164,31 @@ public:
164 __aicore__ inline void DmaInitIterValue(Intf *self)164 __aicore__ inline void DmaInitIterValue(Intf *self)
165 {165 {
166 if constexpr (Intf::iterateMFirstFlag) {166 if constexpr (Intf::iterateMFirstFlag) {
167- self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->convApiTiling.nBL1;167+ self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->nBL1;
168- self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.nBL1 : self->ctx.nBL1Tail;168+ self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->nBL1 : self->ctx.nBL1Tail;
169 169
170- self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->convApiTiling.nBL1);170+ self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->nBL1);
171 self->ctx.maxNBL1Iter = self->ctx.ddr2l1LoopN - 1;171 self->ctx.maxNBL1Iter = self->ctx.ddr2l1LoopN - 1;
172 172
173 if (!self->ctx.kAL1fullload) {173 if (!self->ctx.kAL1fullload) {
174- self->ctx.ddr2l1LoopTmp = self->ctx.convTilingData->convApiTiling.multiNBL1;174+ self->ctx.ddr2l1LoopTmp = self->ctx.convTilingData->multiNBL1;
175 }175 }
176 } else {176 } else {
177 if (!self->ctx.kAL1fullload) {177 if (!self->ctx.kAL1fullload) {
178- self->ctx.ddr2l1LoopTmp = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->convApiTiling.nL0);178+ self->ctx.ddr2l1LoopTmp = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->nL0);
179 }179 }
180 }180 }
181 }181 }
182 182
183 __aicore__ inline void DmaInitBuf(Intf *self)183 __aicore__ inline void DmaInitBuf(Intf *self)
184 {184 {
185- self->ctx.ubBufSize = self->ctx.convTilingData->convApiTiling.hoL1 * AlignB(self->ctx.convTilingData->convApiTiling.woL1, BLOCK_L0_M) *185+ self->ctx.ubBufSize = self->ctx.convTilingData->hoL1 * AlignB(self->ctx.convTilingData->woL1, BLOCK_L0_M) *
186- self->ctx.convTilingData->convApiTiling.khUb * self->ctx.convTilingData->convApiTiling.kwUb * Intf::k0;186+ self->ctx.convTilingData->khUb * self->ctx.convTilingData->kwUb * Intf::k0;
187 187
188 self->ctx.pipe.InitBuffer(self->ctx.ubBuf, self->ctx.ubBufSize * Intf::sizeOfFmap);188 self->ctx.pipe.InitBuffer(self->ctx.ubBuf, self->ctx.ubBufSize * Intf::sizeOfFmap);
189 self->ctx.img2ColTensor = self->ctx.ubBuf.template Get<typename Intf::FmapT>();189 self->ctx.img2ColTensor = self->ctx.ubBuf.template Get<typename Intf::FmapT>();
190 190
191- self->ctx.pipe.InitBuffer(self->ctx.aL1TBuf, self->ctx.convTilingData->convApiTiling.aL1SpaceSize);191+ self->ctx.pipe.InitBuffer(self->ctx.aL1TBuf, self->ctx.convTilingData->aL1SpaceSize);
192 self->ctx.al1 = self->ctx.aL1TBuf.template Get<typename Intf::FmapT>();192 self->ctx.al1 = self->ctx.aL1TBuf.template Get<typename Intf::FmapT>();
193 }193 }
194 194
@@ -196,23 +196,23 @@ public:
196 {196 {
197 self->ctx.vecId = GetSubBlockIdx();197 self->ctx.vecId = GetSubBlockIdx();
198 198 
199- self->ctx.singleCoreCo = self->ctx.convTilingData->convApiTiling.singleCoreCo;199+ self->ctx.singleCoreCo = self->ctx.convTilingData->singleCoreCo;
200- self->ctx.singleCoreHo = self->ctx.convTilingData->convApiTiling.singleCoreHo;200+ self->ctx.singleCoreHo = self->ctx.convTilingData->singleCoreHo;
201- self->ctx.singleCoreWo = self->ctx.convTilingData->convApiTiling.singleCoreWo;201+ self->ctx.singleCoreWo = self->ctx.convTilingData->singleCoreWo;
202- self->ctx.singleCoreCi = self->ctx.convTilingData->convApiTiling.singleCoreCi;202+ self->ctx.singleCoreCi = self->ctx.convTilingData->singleCoreCi;
203- self->ctx.vecKhLoopTimes = CeilDiv(self->ctx.convTilingData->convApiTiling.khL1, self->ctx.convTilingData->convApiTiling.khUb);203+ self->ctx.vecKhLoopTimes = CeilDiv(self->ctx.convTilingData->khL1, self->ctx.convTilingData->khUb);
204 self->ctx.maxVecKhIter = self->ctx.vecKhLoopTimes - 1;204 self->ctx.maxVecKhIter = self->ctx.vecKhLoopTimes - 1;
205- self->ctx.vecKwLoopTimes = CeilDiv(self->ctx.convTilingData->convApiTiling.kwL1, self->ctx.convTilingData->convApiTiling.kwUb);205+ self->ctx.vecKwLoopTimes = CeilDiv(self->ctx.convTilingData->kwL1, self->ctx.convTilingData->kwUb);
206 self->ctx.maxVecKwIter = self->ctx.vecKwLoopTimes - 1;206 self->ctx.maxVecKwIter = self->ctx.vecKwLoopTimes - 1;
207- self->ctx.ddr2L1LoopKh = CeilDiv(self->ctx.convTilingData->convApiTiling.kernelH, self->ctx.convTilingData->convApiTiling.khL1);207+ self->ctx.ddr2L1LoopKh = CeilDiv(self->ctx.convTilingData->kernelH, self->ctx.convTilingData->khL1);
208 self->ctx.maxKhAL1Iter = self->ctx.ddr2L1LoopKh - 1;208 self->ctx.maxKhAL1Iter = self->ctx.ddr2L1LoopKh - 1;
209- self->ctx.ddr2L1LoopKw = CeilDiv(self->ctx.convTilingData->convApiTiling.kernelW, self->ctx.convTilingData->convApiTiling.kwL1);209+ self->ctx.ddr2L1LoopKw = CeilDiv(self->ctx.convTilingData->kernelW, self->ctx.convTilingData->kwL1);
210 self->ctx.maxKwAL1Iter = self->ctx.ddr2L1LoopKw - 1;210 self->ctx.maxKwAL1Iter = self->ctx.ddr2L1LoopKw - 1;
211- self->ctx.cinAL1 = self->ctx.convTilingData->convApiTiling.kAL1 / (self->ctx.convTilingData->convApiTiling.khL1 * self->ctx.convTilingData->convApiTiling.kwL1);211+ self->ctx.cinAL1 = self->ctx.convTilingData->kAL1 / (self->ctx.convTilingData->khL1 * self->ctx.convTilingData->kwL1);
212 self->ctx.cinAL1LoopTimes = CeilDiv(self->ctx.singleCoreCi, self->ctx.cinAL1);212 self->ctx.cinAL1LoopTimes = CeilDiv(self->ctx.singleCoreCi, self->ctx.cinAL1);
213 self->ctx.maxCinAL1Iter = self->ctx.cinAL1LoopTimes - 1;213 self->ctx.maxCinAL1Iter = self->ctx.cinAL1LoopTimes - 1;
214 214
215- self->ctx.fmapOneBatchSize = self->ctx.convTilingData->convApiTiling.orgCi * self->ctx.convTilingData->convApiTiling.orgHixWi;215+ self->ctx.fmapOneBatchSize = self->ctx.convTilingData->orgCi * self->ctx.convTilingData->orgHixWi;
216 216
217 DmaInitBuf(self);217 DmaInitBuf(self);
218 DmaInitKValue(self);218 DmaInitKValue(self);
@@ -225,12 +225,12 @@ public:
225 225 
226 __aicore__ inline void DmaCubeInit(Intf *self)226 __aicore__ inline void DmaCubeInit(Intf *self)
227 {227 {
228- self->ctx.ddr2L1LoopKh = CeilDiv(self->ctx.convTilingData->convApiTiling.kernelH, self->ctx.convTilingData->convApiTiling.khL1);228+ self->ctx.ddr2L1LoopKh = CeilDiv(self->ctx.convTilingData->kernelH, self->ctx.convTilingData->khL1);
229 self->ctx.maxKhAL1Iter = self->ctx.ddr2L1LoopKh - 1;229 self->ctx.maxKhAL1Iter = self->ctx.ddr2L1LoopKh - 1;
230- self->ctx.ddr2L1LoopKw = CeilDiv(self->ctx.convTilingData->convApiTiling.kernelW, self->ctx.convTilingData->convApiTiling.kwL1);230+ self->ctx.ddr2L1LoopKw = CeilDiv(self->ctx.convTilingData->kernelW, self->ctx.convTilingData->kwL1);
231 self->ctx.maxKwAL1Iter = self->ctx.ddr2L1LoopKw - 1;231 self->ctx.maxKwAL1Iter = self->ctx.ddr2L1LoopKw - 1;
232- self->ctx.cinBL1 = CeilDiv(self->ctx.convTilingData->convApiTiling.kBL1, self->ctx.convTilingData->convApiTiling.khL1 * self->ctx.convTilingData->convApiTiling.kwL1);232+ self->ctx.cinBL1 = CeilDiv(self->ctx.convTilingData->kBL1, self->ctx.convTilingData->khL1 * self->ctx.convTilingData->kwL1);
233- self->ctx.cinBL1LoopTimes = CeilDiv(self->ctx.convTilingData->convApiTiling.singleCoreCi, self->ctx.cinBL1);233+ self->ctx.cinBL1LoopTimes = CeilDiv(self->ctx.convTilingData->singleCoreCi, self->ctx.cinBL1);
234 }234 }
235 235
236 __aicore__ inline void DmaUpdateLoopInner(Intf *self)236 __aicore__ inline void DmaUpdateLoopInner(Intf *self)
@@ -241,7 +241,7 @@ public:
241 241
242 if constexpr (Intf::iterateMFirstFlag) {242 if constexpr (Intf::iterateMFirstFlag) {
243 self->ctx.ddr2l1LoopTmp = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ? 243 self->ctx.ddr2l1LoopTmp = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?
244- CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->convApiTiling.nL0) : self->ctx.convTilingData->convApiTiling.multiNBL1;244+ CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->nL0) : self->ctx.convTilingData->multiNBL1;
245 }245 }
246 246 
247 self->ctx.ddr2l1LoopInner = self->ctx.ddr2l1LoopTmp * self->ctx.l12l0LoopH * self->ctx.l12l0LoopW;247 self->ctx.ddr2l1LoopInner = self->ctx.ddr2l1LoopTmp * self->ctx.l12l0LoopH * self->ctx.l12l0LoopW;
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2_group.h+31-31
@@ -99,14 +99,14 @@ __aicore__ inline bool GroupConv2d<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYP
99 hasScale = (extendParams != nullptr);99 hasScale = (extendParams != nullptr);
100 convTilingData = &tilingData;100 convTilingData = &tilingData;
101 if constexpr (CONV_CFG::isExtendConv2d) {101 if constexpr (CONV_CFG::isExtendConv2d) {
102- dualOutput = convTilingData->convApiTiling.dualOutput;102+ dualOutput = convTilingData->dualOutput;
103 }103 }
104 convCommon.Init(this, convTilingData, hasScale);104 convCommon.Init(this, convTilingData, hasScale);
105 105 
106 conv.Init(convTilingData);106 conv.Init(convTilingData);
107 107 
108- ciPerGroup = convTilingData->convRunInfo.cin / convTilingData->convRunInfo.groups;108+ ciPerGroup = convTilingData->cin / convTilingData->groups;
109- coPerGroup = convTilingData->convRunInfo.cout / convTilingData->convRunInfo.groups;109+ coPerGroup = convTilingData->cout / convTilingData->groups;
110 110 
111 if (!InitSingleCoreData()) {111 if (!InitSingleCoreData()) {
112 return false;112 return false;
@@ -124,30 +124,30 @@ __aicore__ inline bool GroupConv2d<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYP
124 InitBlockNums();124 InitBlockNums();
125 125 
126 DimDataToFill batchToFill(this->singleCoreBatch, this->batchIdxStart, this->isBatchDimTail);126 DimDataToFill batchToFill(this->singleCoreBatch, this->batchIdxStart, this->isBatchDimTail);
127- bool isRealDim = convCommon.CalcDimData(this->batchBlockNums, convTilingData->convRunInfo.batchDim, convTilingData->convRunInfo.batch,127+ bool isRealDim = convCommon.CalcDimData(this->batchBlockNums, convTilingData->batchDim, convTilingData->batch,
128- convTilingData->convRunInfo.batch, batchToFill);128+ convTilingData->batch, batchToFill);
129 if (unlikely(!isRealDim)) {129 if (unlikely(!isRealDim)) {
130 return false;130 return false;
131 }131 }
132 132 
133 if constexpr (isMMode) {133 if constexpr (isMMode) {
134 DimDataToFill mToFill(this->singleCoreM, this->mIdxStart, this->isMDimTail);134 DimDataToFill mToFill(this->singleCoreM, this->mIdxStart, this->isMDimTail);
135- uint64_t totalM = convTilingData->convRunInfo.hout * convTilingData->convRunInfo.wout;135+ uint64_t totalM = convTilingData->hout * convTilingData->wout;
136- isRealDim = convCommon.CalcDimData(this->hoBlockNums, convTilingData->convRunInfo.hoDim, convCommon.AlignB(totalM, M0),136+ isRealDim = convCommon.CalcDimData(this->hoBlockNums, convTilingData->hoDim, convCommon.AlignB(totalM, M0),
137 totalM, mToFill);137 totalM, mToFill);
138 if (unlikely(!isRealDim)) {138 if (unlikely(!isRealDim)) {
139 return false;139 return false;
140 }140 }
141 } else {141 } else {
142 DimDataToFill hoToFill(this->singleCoreHo, this->hoIdxStart, this->isHoDimTail);142 DimDataToFill hoToFill(this->singleCoreHo, this->hoIdxStart, this->isHoDimTail);
143- isRealDim = convCommon.CalcDimData(this->hoBlockNums, convTilingData->convRunInfo.hoDim, convTilingData->convRunInfo.hout,143+ isRealDim = convCommon.CalcDimData(this->hoBlockNums, convTilingData->hoDim, convTilingData->hout,
144- convTilingData->convRunInfo.hout, hoToFill);144+ convTilingData->hout, hoToFill);
145 if (unlikely(!isRealDim)) {145 if (unlikely(!isRealDim)) {
146 return false;146 return false;
147 }147 }
148 DimDataToFill woToFill(this->singleCoreWo, this->woIdxStart, this->isWoDimTail);148 DimDataToFill woToFill(this->singleCoreWo, this->woIdxStart, this->isWoDimTail);
149- isRealDim = convCommon.CalcDimData(this->woBlockNums, convTilingData->convRunInfo.woDim, convTilingData->convRunInfo.wout,149+ isRealDim = convCommon.CalcDimData(this->woBlockNums, convTilingData->woDim, convTilingData->wout,
150- convTilingData->convRunInfo.wout, woToFill);150+ convTilingData->wout, woToFill);
151 if (unlikely(!isRealDim)) {151 if (unlikely(!isRealDim)) {
152 return false;152 return false;
153 }153 }
@@ -170,28 +170,28 @@ __aicore__ inline void GroupConv2d<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYP
170 // NCHW: batchDim -> 'groupDim -> nDim' -> hoDim -> woDim / batchDim -> 'groupDim -> nDim' -> mDim170 // NCHW: batchDim -> 'groupDim -> nDim' -> hoDim -> woDim / batchDim -> 'groupDim -> nDim' -> mDim
171 // NHWC: batchDim -> hoDim -> woDim -> 'groupDim -> nDim' / batchDim -> mDim -> 'groupDim -> nDim'171 // NHWC: batchDim -> hoDim -> woDim -> 'groupDim -> nDim' / batchDim -> mDim -> 'groupDim -> nDim'
172 if constexpr (isMMode) {172 if constexpr (isMMode) {
173- this->batchBlockNums = convTilingData->convRunInfo.groupDim * convTilingData->convRunInfo.nDim * convTilingData->convRunInfo.hoDim;173+ this->batchBlockNums = convTilingData->groupDim * convTilingData->nDim * convTilingData->hoDim;
174 if constexpr (A_FORMAT == ConvFormat::NCHW) {174 if constexpr (A_FORMAT == ConvFormat::NCHW) {
175- this->groupBlockNums = convTilingData->convRunInfo.nDim * convTilingData->convRunInfo.hoDim;175+ this->groupBlockNums = convTilingData->nDim * convTilingData->hoDim;
176- this->nBlockNums = convTilingData->convRunInfo.hoDim;176+ this->nBlockNums = convTilingData->hoDim;
177 this->hoBlockNums = 1;177 this->hoBlockNums = 1;
178 } else {178 } else {
179- this->hoBlockNums = convTilingData->convRunInfo.groupDim * convTilingData->convRunInfo.nDim;179+ this->hoBlockNums = convTilingData->groupDim * convTilingData->nDim;
180- this->groupBlockNums = convTilingData->convRunInfo.nDim;180+ this->groupBlockNums = convTilingData->nDim;
181 this->nBlockNums = 1;181 this->nBlockNums = 1;
182 }182 }
183 } else {183 } else {
184- this->batchBlockNums = convTilingData->convRunInfo.groupDim * convTilingData->convRunInfo.nDim * convTilingData->convRunInfo.hoDim *184+ this->batchBlockNums = convTilingData->groupDim * convTilingData->nDim * convTilingData->hoDim *
185- convTilingData->convRunInfo.woDim;185+ convTilingData->woDim;
186 if constexpr (A_FORMAT == ConvFormat::NCHW) {186 if constexpr (A_FORMAT == ConvFormat::NCHW) {
187- this->groupBlockNums = convTilingData->convRunInfo.nDim * convTilingData->convRunInfo.hoDim * convTilingData->convRunInfo.woDim;187+ this->groupBlockNums = convTilingData->nDim * convTilingData->hoDim * convTilingData->woDim;
188- this->nBlockNums = convTilingData->convRunInfo.hoDim * convTilingData->convRunInfo.woDim;188+ this->nBlockNums = convTilingData->hoDim * convTilingData->woDim;
189- this->hoBlockNums = convTilingData->convRunInfo.woDim;189+ this->hoBlockNums = convTilingData->woDim;
190 this->woBlockNums = 1;190 this->woBlockNums = 1;
191 } else {191 } else {
192- this->hoBlockNums = convTilingData->convRunInfo.woDim * convTilingData->convRunInfo.groupDim * convTilingData->convRunInfo.nDim;192+ this->hoBlockNums = convTilingData->woDim * convTilingData->groupDim * convTilingData->nDim;
193- this->woBlockNums = convTilingData->convRunInfo.groupDim * convTilingData->convRunInfo.nDim;193+ this->woBlockNums = convTilingData->groupDim * convTilingData->nDim;
194- this->groupBlockNums = convTilingData->convRunInfo.nDim;194+ this->groupBlockNums = convTilingData->nDim;
195 this->nBlockNums = 1;195 this->nBlockNums = 1;
196 }196 }
197 }197 }
@@ -202,14 +202,14 @@ __aicore__ inline bool GroupConv2d<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYP
202 InitSingleCoreDataOriGroup()202 InitSingleCoreDataOriGroup()
203{203{
204 DimDataToFill groupToFill(singleGroups, groupIdxStart, isGroupDimTail);204 DimDataToFill groupToFill(singleGroups, groupIdxStart, isGroupDimTail);
205- bool isRealDim = convCommon.CalcDimData(this->groupBlockNums, convTilingData->convRunInfo.groupDim,205+ bool isRealDim = convCommon.CalcDimData(this->groupBlockNums, convTilingData->groupDim,
206- convTilingData->convRunInfo.groups, convTilingData->convRunInfo.groups, groupToFill);206+ convTilingData->groups, convTilingData->groups, groupToFill);
207 if (unlikely(!isRealDim)) {207 if (unlikely(!isRealDim)) {
208 return false;208 return false;
209 }209 }
210 210 
211 DimDataToFill nToFill(this->singleCoreN, this->nIdxStart, this->isNDimTail);211 DimDataToFill nToFill(this->singleCoreN, this->nIdxStart, this->isNDimTail);
212- isRealDim = convCommon.CalcNDimDataAlign(this->nBlockNums, convTilingData->convRunInfo.nDim, coPerGroup, nToFill);212+ isRealDim = convCommon.CalcNDimDataAlign(this->nBlockNums, convTilingData->nDim, coPerGroup, nToFill);
213 if (unlikely(!isRealDim)) {213 if (unlikely(!isRealDim)) {
214 return false;214 return false;
215 }215 }
@@ -222,15 +222,15 @@ __aicore__ inline bool GroupConv2d<FMAP_TYPE, WEIGHT_TYPE, OUTPUT_TYPE, BIAS_TYP
222 InitSingleCoreDataOptGroup()222 InitSingleCoreDataOptGroup()
223{223{
224 DimDataToFill groupToFill(singleGroupOpt, groupIdxStart, isGroupDimTail);224 DimDataToFill groupToFill(singleGroupOpt, groupIdxStart, isGroupDimTail);
225- bool isRealDim = convCommon.CalcDimData(this->groupBlockNums, convTilingData->convRunInfo.groupDim,225+ bool isRealDim = convCommon.CalcDimData(this->groupBlockNums, convTilingData->groupDim,
226- convTilingData->convRunInfo.groupOpt, convTilingData->convRunInfo.groupOpt, groupToFill);226+ convTilingData->groupOpt, convTilingData->groupOpt, groupToFill);
227 if (unlikely(!isRealDim)) {227 if (unlikely(!isRealDim)) {
228 return false;228 return false;
229 }229 }
230 230 
231 DimDataToFill nToFill(singleCoOpt, this->nIdxStart, this->isNDimTail);231 DimDataToFill nToFill(singleCoOpt, this->nIdxStart, this->isNDimTail);
232- isRealDim = convCommon.CalcNDimDataAlign(this->nBlockNums, convTilingData->convRunInfo.nDim,232+ isRealDim = convCommon.CalcNDimDataAlign(this->nBlockNums, convTilingData->nDim,
233- convTilingData->convRunInfo.coutOpt, nToFill);233+ convTilingData->coutOpt, nToFill);
234 if (unlikely(!isRealDim)) {234 if (unlikely(!isRealDim)) {
235 return false;235 return false;
236 }236 }
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2_instr_base_impl.h+135-135
@@ -31,7 +31,7 @@ public:
31 __aicore__ inline void SetIntf(Intf *self)31 __aicore__ inline void SetIntf(Intf *self)
32 {32 {
33 self_ = self;33 self_ = self;
34- buffAddr.dataLen = self_->ctx.convTilingData->convApiTiling.aL1SpaceSize;34+ buffAddr.dataLen = self_->ctx.convTilingData->aL1SpaceSize;
35 buffAddr.logicPos = static_cast<uint8_t>(QuePosition::A1);35 buffAddr.logicPos = static_cast<uint8_t>(QuePosition::A1);
36 }36 }
37 37 
@@ -39,16 +39,16 @@ public:
39 {39 {
40 if constexpr (Intf::isQuantScene) {40 if constexpr (Intf::isQuantScene) {
41 uint16_t padValue = 41 uint16_t padValue =
42- (static_cast<uint16_t>(self_->ctx.convTilingData->convApiTiling.offsetx)) << BIT_OFFSET_8 |42+ (static_cast<uint16_t>(self_->ctx.convTilingData->offsetx)) << BIT_OFFSET_8 |
43- (static_cast<uint16_t>(self_->ctx.convTilingData->convApiTiling.offsetx));43+ (static_cast<uint16_t>(self_->ctx.convTilingData->offsetx));
44 InitConstValueParams<uint16_t> params(44 InitConstValueParams<uint16_t> params(
45- 1, static_cast<uint16_t>(self_->ctx.convTilingData->convApiTiling.aL1SpaceSize / C0_SIZE), 0, padValue);45+ 1, static_cast<uint16_t>(self_->ctx.convTilingData->aL1SpaceSize / C0_SIZE), 0, padValue);
46 buffAddr.bufferAddr = self_->ctx.al1.GetPhyAddr();46 buffAddr.bufferAddr = self_->ctx.al1.GetPhyAddr();
47 al1tmp.SetAddr(buffAddr);47 al1tmp.SetAddr(buffAddr);
48 InitConstValue<uint16_t>(al1tmp, params);48 InitConstValue<uint16_t>(al1tmp, params);
49 } else {49 } else {
50 InitConstValueParams<typename Intf::FmapT> params(50 InitConstValueParams<typename Intf::FmapT> params(
51- 1, static_cast<uint16_t>(self_->ctx.convTilingData->convApiTiling.aL1SpaceSize / C0_SIZE), 0, 0);51+ 1, static_cast<uint16_t>(self_->ctx.convTilingData->aL1SpaceSize / C0_SIZE), 0, 0);
52 InitConstValue<typename Intf::FmapT>(self_->ctx.al1, params);52 InitConstValue<typename Intf::FmapT>(self_->ctx.al1, params);
53 }53 }
54 }54 }
@@ -64,22 +64,22 @@ public:
64 padList[PAD_IDX_T] = padTopL1;64 padList[PAD_IDX_T] = padTopL1;
65 Load3DSetFMatrixCal(hiLoadL1, wiLoadL1, padList);65 Load3DSetFMatrixCal(hiLoadL1, wiLoadL1, padList);
66 }66 }
67- Load3DSetPaddingCal(self_->ctx.convTilingData->convApiTiling.offsetx);67+ Load3DSetPaddingCal(self_->ctx.convTilingData->offsetx);
68 }68 }
69 69 
70 __aicore__ inline void SetLoad3dFMatrixNoPad(uint64_t wiLoadL1)70 __aicore__ inline void SetLoad3dFMatrixNoPad(uint64_t wiLoadL1)
71 {71 {
72 uint8_t padList[4] = {0};72 uint8_t padList[4] = {0};
73 Load3DSetFMatrixCal(hiLoadL1, wiLoadL1, padList);73 Load3DSetFMatrixCal(hiLoadL1, wiLoadL1, padList);
74- Load3DSetPaddingCal(self_->ctx.convTilingData->convApiTiling.offsetx);74+ Load3DSetPaddingCal(self_->ctx.convTilingData->offsetx);
75 }75 }
76 76 
77 __aicore__ inline void SetLoad3dFMatrixForOptPreload()77 __aicore__ inline void SetLoad3dFMatrixForOptPreload()
78 {78 {
79 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {79 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {
80- SetLoad3dFMatrix(self_->ctx.convTilingData->convApiTiling.padLeft,80+ SetLoad3dFMatrix(self_->ctx.convTilingData->padLeft,
81- self_->ctx.convTilingData->convApiTiling.padRight,81+ self_->ctx.convTilingData->padRight,
82- self_->ctx.convTilingData->convApiTiling.orgWi);82+ self_->ctx.convTilingData->orgWi);
83 } else {83 } else {
84 SetLoad3dFMatrix(padLeftL1, padRightL1, wiLoadL1);84 SetLoad3dFMatrix(padLeftL1, padRightL1, wiLoadL1);
85 }85 }
@@ -106,18 +106,18 @@ public:
106 {106 {
107 self_ = self;107 self_ = self;
108 108 
109- hiLoadL1 = (self_->ctx.convTilingData->convApiTiling.orgHo - 1) *109+ hiLoadL1 = (self_->ctx.convTilingData->orgHo - 1) *
110- self_->ctx.convTilingData->convApiTiling.strideH + self_->ctx.dilatedKernelH;110+ self_->ctx.convTilingData->strideH + self_->ctx.dilatedKernelH;
111- hiLoadL1 = hiLoadL1 > self_->ctx.convTilingData->convApiTiling.orgHi ?111+ hiLoadL1 = hiLoadL1 > self_->ctx.convTilingData->orgHi ?
112- self_->ctx.convTilingData->convApiTiling.orgHi : hiLoadL1;112+ self_->ctx.convTilingData->orgHi : hiLoadL1;
113- realHixWi = hiLoadL1 * self_->ctx.convTilingData->convApiTiling.orgWi;113+ realHixWi = hiLoadL1 * self_->ctx.convTilingData->orgWi;
114 114 
115- padList[PAD_IDX_L] = static_cast<uint8_t>(self_->ctx.convTilingData->convApiTiling.padLeft);115+ padList[PAD_IDX_L] = static_cast<uint8_t>(self_->ctx.convTilingData->padLeft);
116- padList[PAD_IDX_R] = static_cast<uint8_t>(self_->ctx.convTilingData->convApiTiling.padRight);116+ padList[PAD_IDX_R] = static_cast<uint8_t>(self_->ctx.convTilingData->padRight);
117- padList[PAD_IDX_T] = static_cast<uint8_t>(self_->ctx.convTilingData->convApiTiling.padTop);117+ padList[PAD_IDX_T] = static_cast<uint8_t>(self_->ctx.convTilingData->padTop);
118- Load3DSetPaddingCal(self_->ctx.convTilingData->convApiTiling.offsetx);118+ Load3DSetPaddingCal(self_->ctx.convTilingData->offsetx);
119 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {119 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {
120- Load3DSetFMatrixCal(hiLoadL1, self_->ctx.convTilingData->convApiTiling.orgWi, padList);120+ Load3DSetFMatrixCal(hiLoadL1, self_->ctx.convTilingData->orgWi, padList);
121 }121 }
122 }122 }
123 123 
@@ -131,11 +131,11 @@ public:
131 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {131 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {
132 if constexpr (Intf::formatFmap == ConvFormat::NCHW && Intf::c04Flag) {132 if constexpr (Intf::formatFmap == ConvFormat::NCHW && Intf::c04Flag) {
133 Load3DSetFMatrixCal(self_->ctx.innerBatch,133 Load3DSetFMatrixCal(self_->ctx.innerBatch,
134- AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->convApiTiling.orgHixWi, Intf::k0) / C04_CIN_SIZE,134+ AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->orgHixWi, Intf::k0) / C04_CIN_SIZE,
135 padList);135 padList);
136 } else {136 } else {
137 Load3DSetFMatrixCal(self_->ctx.innerBatch * hiLoadL1,137 Load3DSetFMatrixCal(self_->ctx.innerBatch * hiLoadL1,
138- self_->ctx.convTilingData->convApiTiling.orgWi, padList);138+ self_->ctx.convTilingData->orgWi, padList);
139 }139 }
140 }140 }
141 if constexpr (Intf::disContinuousFlag) {141 if constexpr (Intf::disContinuousFlag) {
@@ -143,10 +143,10 @@ public:
143 } else if constexpr (Intf::formatFmap == ConvFormat::NCHW) {143 } else if constexpr (Intf::formatFmap == ConvFormat::NCHW) {
144 uint64_t aL1GmOffset = 0;144 uint64_t aL1GmOffset = 0;
145 if constexpr (!Intf::isKL1NL0FullLoad) {145 if constexpr (!Intf::isKL1NL0FullLoad) {
146- aL1GmOffset += kAL1Iter * self_->ctx.convTilingData->convApiTiling.cinOffsetBlockInGM;146+ aL1GmOffset += kAL1Iter * self_->ctx.convTilingData->cinOffsetBlockInGM;
147 }147 }
148 if constexpr (!Intf::isOneBatch) {148 if constexpr (!Intf::isOneBatch) {
149- aL1GmOffset += batchIter * self_->ctx.convTilingData->convApiTiling.innerBatch *149+ aL1GmOffset += batchIter * self_->ctx.convTilingData->innerBatch *
150 self_->ctx.fmapOneBatchSize;150 self_->ctx.fmapOneBatchSize;
151 }151 }
152 152 
@@ -161,10 +161,10 @@ public:
161 } else if constexpr (Intf::formatOutput == ConvFormat::NHWC) {161 } else if constexpr (Intf::formatOutput == ConvFormat::NHWC) {
162 uint64_t aL1GmOffset = 0;162 uint64_t aL1GmOffset = 0;
163 if constexpr (!Intf::isKL1NL0FullLoad) {163 if constexpr (!Intf::isKL1NL0FullLoad) {
164- aL1GmOffset += kAL1Iter * self_->ctx.convTilingData->convApiTiling.cinAInCore;164+ aL1GmOffset += kAL1Iter * self_->ctx.convTilingData->cinAInCore;
165 }165 }
166 if constexpr (!Intf::isOneBatch) {166 if constexpr (!Intf::isOneBatch) {
167- aL1GmOffset += batchIter * self_->ctx.convTilingData->convApiTiling.innerBatch *167+ aL1GmOffset += batchIter * self_->ctx.convTilingData->innerBatch *
168 self_->ctx.fmapOneBatchSize;168 self_->ctx.fmapOneBatchSize;
169 }169 }
170 170 
@@ -181,9 +181,9 @@ public:
181 181 
182 __aicore__ inline void LoadAL1InputHWNC(uint64_t kAL1Iter, uint64_t mAL1Iter, uint64_t batchIter)182 __aicore__ inline void LoadAL1InputHWNC(uint64_t kAL1Iter, uint64_t mAL1Iter, uint64_t batchIter)
183 {183 {
184- uint64_t aL1GmOffset = batchIter * self_->ctx.convTilingData->convApiTiling.innerBatch *184+ uint64_t aL1GmOffset = batchIter * self_->ctx.convTilingData->innerBatch *
185- self_->ctx.convTilingData->convApiTiling.orgCi +185+ self_->ctx.convTilingData->orgCi +
186- kAL1Iter * self_->ctx.convTilingData->convApiTiling.cinAInCore;186+ kAL1Iter * self_->ctx.convTilingData->cinAInCore;
187 187 
188 Nd2NzParams intriParams;188 Nd2NzParams intriParams;
189 SetNd2NzIntriParamsInputHWNC(intriParams, kAL1Iter);189 SetNd2NzIntriParamsInputHWNC(intriParams, kAL1Iter);
@@ -194,33 +194,33 @@ private:
194 __aicore__ inline void SetDn2NzIntriParamsC04(Dn2NzParams &intriParams)194 __aicore__ inline void SetDn2NzIntriParamsC04(Dn2NzParams &intriParams)
195 {195 {
196 intriParams.dnNum = self_->ctx.innerBatch;196 intriParams.dnNum = self_->ctx.innerBatch;
197- intriParams.nValue = self_->ctx.convTilingData->convApiTiling.orgHixWi;197+ intriParams.nValue = self_->ctx.convTilingData->orgHixWi;
198- intriParams.dValue = self_->ctx.convTilingData->convApiTiling.orgCi;198+ intriParams.dValue = self_->ctx.convTilingData->orgCi;
199- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.orgHixWi;199+ intriParams.srcDValue = self_->ctx.convTilingData->orgHixWi;
200- intriParams.srcDnMatrixStride = self_->ctx.convTilingData->convApiTiling.orgCi *200+ intriParams.srcDnMatrixStride = self_->ctx.convTilingData->orgCi *
201- self_->ctx.convTilingData->convApiTiling.orgHixWi;201+ self_->ctx.convTilingData->orgHixWi;
202 intriParams.dstNzNStride = 1;202 intriParams.dstNzNStride = 1;
203 intriParams.dstNzMatrixStride = 203 intriParams.dstNzMatrixStride =
204- AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->convApiTiling.orgHixWi, Intf::k0);204+ AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->orgHixWi, Intf::k0);
205 }205 }
206 206
207 __aicore__ inline void SetDn2NzIntriParams(Dn2NzParams &intriParams, uint64_t kAL1Iter)207 __aicore__ inline void SetDn2NzIntriParams(Dn2NzParams &intriParams, uint64_t kAL1Iter)
208 {208 {
209 uint32_t al1Ci;209 uint32_t al1Ci;
210 if constexpr (Intf::isKL1NL0FullLoad) {210 if constexpr (Intf::isKL1NL0FullLoad) {
211- al1Ci = self_->ctx.convTilingData->convApiTiling.cinATailInCore;211+ al1Ci = self_->ctx.convTilingData->cinATailInCore;
212 } else {212 } else {
213 al1Ci = IsKAL1Tail(kAL1Iter) ?213 al1Ci = IsKAL1Tail(kAL1Iter) ?
214- self_->ctx.convTilingData->convApiTiling.cinATailInCore :214+ self_->ctx.convTilingData->cinATailInCore :
215- self_->ctx.convTilingData->convApiTiling.cinAInCore;215+ self_->ctx.convTilingData->cinAInCore;
216 }216 }
217 intriParams.dnNum = self_->ctx.innerBatch;217 intriParams.dnNum = self_->ctx.innerBatch;
218 intriParams.nValue = realHixWi;218 intriParams.nValue = realHixWi;
219 intriParams.dValue = al1Ci;219 intriParams.dValue = al1Ci;
220- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.orgHixWi;220+ intriParams.srcDValue = self_->ctx.convTilingData->orgHixWi;
221 intriParams.dstNzNStride = 1;221 intriParams.dstNzNStride = 1;
222- intriParams.srcDnMatrixStride = self_->ctx.convTilingData->convApiTiling.orgCi *222+ intriParams.srcDnMatrixStride = self_->ctx.convTilingData->orgCi *
223- self_->ctx.convTilingData->convApiTiling.orgHixWi;223+ self_->ctx.convTilingData->orgHixWi;
224 224 
225 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {225 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {
226 intriParams.dstNzC0Stride = self_->ctx.innerBatch * realHixWi;226 intriParams.dstNzC0Stride = self_->ctx.innerBatch * realHixWi;
@@ -235,36 +235,36 @@ private:
235 {235 {
236 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {236 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {
237 intriParams.ndNum = 1;237 intriParams.ndNum = 1;
238- intriParams.nValue = self_->ctx.innerBatch * self_->ctx.convTilingData->convApiTiling.orgHixWi;238+ intriParams.nValue = self_->ctx.innerBatch * self_->ctx.convTilingData->orgHixWi;
239 } else {239 } else {
240 intriParams.ndNum = self_->ctx.innerBatch;240 intriParams.ndNum = self_->ctx.innerBatch;
241 intriParams.nValue = realHixWi;241 intriParams.nValue = realHixWi;
242- intriParams.srcNdMatrixStride = self_->ctx.convTilingData->convApiTiling.orgCi *242+ intriParams.srcNdMatrixStride = self_->ctx.convTilingData->orgCi *
243- self_->ctx.convTilingData->convApiTiling.orgHixWi;243+ self_->ctx.convTilingData->orgHixWi;
244 intriParams.dstNzNStride = 1;244 intriParams.dstNzNStride = 1;
245 intriParams.dstNzMatrixStride =245 intriParams.dstNzMatrixStride =
246- AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->convApiTiling.orgHixWi, Intf::k0);246+ AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->orgHixWi, Intf::k0);
247 }247 }
248- intriParams.dValue = self_->ctx.convTilingData->convApiTiling.orgCi;248+ intriParams.dValue = self_->ctx.convTilingData->orgCi;
249- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.orgCi;249+ intriParams.srcDValue = self_->ctx.convTilingData->orgCi;
250 }250 }
251 251 
252 __aicore__ inline void SetNd2NzIntriParams(Nd2NzParams &intriParams, uint64_t kAL1Iter)252 __aicore__ inline void SetNd2NzIntriParams(Nd2NzParams &intriParams, uint64_t kAL1Iter)
253 {253 {
254 uint32_t al1Ci;254 uint32_t al1Ci;
255 if constexpr (Intf::isKL1NL0FullLoad) {255 if constexpr (Intf::isKL1NL0FullLoad) {
256- al1Ci = self_->ctx.convTilingData->convApiTiling.cinATailInCore;256+ al1Ci = self_->ctx.convTilingData->cinATailInCore;
257 } else {257 } else {
258 al1Ci = IsKAL1Tail(kAL1Iter) ?258 al1Ci = IsKAL1Tail(kAL1Iter) ?
259- self_->ctx.convTilingData->convApiTiling.cinATailInCore :259+ self_->ctx.convTilingData->cinATailInCore :
260- self_->ctx.convTilingData->convApiTiling.cinAInCore;260+ self_->ctx.convTilingData->cinAInCore;
261 }261 }
262 intriParams.dValue = al1Ci;262 intriParams.dValue = al1Ci;
263 intriParams.ndNum = self_->ctx.innerBatch;263 intriParams.ndNum = self_->ctx.innerBatch;
264 intriParams.nValue = realHixWi;264 intriParams.nValue = realHixWi;
265- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.orgCi;265+ intriParams.srcDValue = self_->ctx.convTilingData->orgCi;
266- intriParams.srcNdMatrixStride = self_->ctx.convTilingData->convApiTiling.orgCi *266+ intriParams.srcNdMatrixStride = self_->ctx.convTilingData->orgCi *
267- self_->ctx.convTilingData->convApiTiling.orgHixWi;267+ self_->ctx.convTilingData->orgHixWi;
268 intriParams.dstNzNStride = 1;268 intriParams.dstNzNStride = 1;
269 269 
270 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {270 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {
@@ -279,14 +279,14 @@ private:
279 __aicore__ inline void SetNd2NzIntriParamsInputHWNC(Nd2NzParams &intriParams, uint64_t kAL1Iter)279 __aicore__ inline void SetNd2NzIntriParamsInputHWNC(Nd2NzParams &intriParams, uint64_t kAL1Iter)
280 {280 {
281 uint32_t al1Ci = IsKAL1Tail(kAL1Iter) ?281 uint32_t al1Ci = IsKAL1Tail(kAL1Iter) ?
282- self_->ctx.convTilingData->convApiTiling.cinATailInCore :282+ self_->ctx.convTilingData->cinATailInCore :
283- self_->ctx.convTilingData->convApiTiling.cinAInCore;283+ self_->ctx.convTilingData->cinAInCore;
284 intriParams.ndNum = self_->ctx.innerBatch;284 intriParams.ndNum = self_->ctx.innerBatch;
285 intriParams.nValue = realHixWi;285 intriParams.nValue = realHixWi;
286 intriParams.dValue = al1Ci;286 intriParams.dValue = al1Ci;
287- intriParams.srcNdMatrixStride = self_->ctx.convTilingData->convApiTiling.orgCi;287+ intriParams.srcNdMatrixStride = self_->ctx.convTilingData->orgCi;
288- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.orgCi *288+ intriParams.srcDValue = self_->ctx.convTilingData->orgCi *
289- self_->ctx.convTilingData->convRunInfo.batch;289+ self_->ctx.convTilingData->batch;
290 intriParams.dstNzNStride = 1;290 intriParams.dstNzNStride = 1;
291 291 
292 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {292 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {
@@ -318,7 +318,7 @@ public:
318 {318 {
319 self_ = self;319 self_ = self;
320 320 
321- isOneByOneKernelScenario = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW == 1 &&321+ isOneByOneKernelScenario = self_->ctx.convTilingData->kernelHxkernelW == 1 &&
322 !(AscendC::IsSameType<typename Intf::WeightT, hifloat8_t>::value) &&322 !(AscendC::IsSameType<typename Intf::WeightT, hifloat8_t>::value) &&
323 !(AscendC::IsSameType<typename Intf::WeightT, fp8_e4m3fn_t>::value);323 !(AscendC::IsSameType<typename Intf::WeightT, fp8_e4m3fn_t>::value);
324 }324 }
@@ -350,14 +350,14 @@ private:
350 {350 {
351 uint64_t bL1GmOffset;351 uint64_t bL1GmOffset;
352 if constexpr (!Intf::hasNL1IterFlag) {352 if constexpr (!Intf::hasNL1IterFlag) {
353- bL1GmOffset = kBL1Iter * self_->ctx.convTilingData->convApiTiling.kBL1;353+ bL1GmOffset = kBL1Iter * self_->ctx.convTilingData->kBL1;
354 } else {354 } else {
355 bL1GmOffset = 355 bL1GmOffset =
356- nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 *356+ nBL1Iter * self_->ctx.convTilingData->nBL1 *
357- self_->ctx.convTilingData->convApiTiling.coutOffsetBlock +357+ self_->ctx.convTilingData->coutOffsetBlock +
358- kBL1Iter * self_->ctx.convTilingData->convApiTiling.kBL1;358+ kBL1Iter * self_->ctx.convTilingData->kBL1;
359 self_->ctx.currentNBL1 = nBL1Iter == self_->ctx.maxNBL1Iter ?359 self_->ctx.currentNBL1 = nBL1Iter == self_->ctx.maxNBL1Iter ?
360- self_->ctx.nBL1Tail : self_->ctx.convTilingData->convApiTiling.nBL1;360+ self_->ctx.nBL1Tail : self_->ctx.convTilingData->nBL1;
361 }361 }
362 362 
363 if (isOneByOneKernelScenario) {363 if (isOneByOneKernelScenario) {
@@ -365,11 +365,11 @@ private:
365 intriParams.ndNum = 1;365 intriParams.ndNum = 1;
366 intriParams.nValue = self_->ctx.currentNBL1;366 intriParams.nValue = self_->ctx.currentNBL1;
367 intriParams.dValue = kBL1Iter == self_->ctx.maxKBL1Iter ?367 intriParams.dValue = kBL1Iter == self_->ctx.maxKBL1Iter ?
368- self_->ctx.kBL1Tail : self_->ctx.convTilingData->convApiTiling.kBL1;368+ self_->ctx.kBL1Tail : self_->ctx.convTilingData->kBL1;
369 intriParams.srcNdMatrixStride = 0;369 intriParams.srcNdMatrixStride = 0;
370 intriParams.srcDValue =370 intriParams.srcDValue =
371- self_->ctx.convTilingData->convApiTiling.orgCi / self_->ctx.convTilingData->convApiTiling.groups;371+ self_->ctx.convTilingData->orgCi / self_->ctx.convTilingData->groups;
372- intriParams.dstNzC0Stride = self_->ctx.convTilingData->convApiTiling.nBL1;372+ intriParams.dstNzC0Stride = self_->ctx.convTilingData->nBL1;
373 intriParams.dstNzNStride = 1;373 intriParams.dstNzNStride = 1;
374 intriParams.dstNzMatrixStride = 0;374 intriParams.dstNzMatrixStride = 0;
375 DataCopy<typename Intf::WeightT>(self_->ctx.bl1, self_->ctx.bgm[bL1GmOffset], intriParams);375 DataCopy<typename Intf::WeightT>(self_->ctx.bl1, self_->ctx.bgm[bL1GmOffset], intriParams);
@@ -377,43 +377,43 @@ private:
377 }377 }
378 Dn2NzParams intriParams;378 Dn2NzParams intriParams;
379 intriParams.dnNum = self_->ctx.currentNBL1;379 intriParams.dnNum = self_->ctx.currentNBL1;
380- intriParams.nValue = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;380+ intriParams.nValue = self_->ctx.convTilingData->kernelHxkernelW;
381 intriParams.dValue = kBL1Iter == self_->ctx.maxKBL1Iter ?381 intriParams.dValue = kBL1Iter == self_->ctx.maxKBL1Iter ?
382- self_->ctx.convTilingData->convApiTiling.cinBTailInCore :382+ self_->ctx.convTilingData->cinBTailInCore :
383- self_->ctx.convTilingData->convApiTiling.cinBInCore;383+ self_->ctx.convTilingData->cinBInCore;
384- intriParams.srcDnMatrixStride = self_->ctx.convTilingData->convApiTiling.coutOffsetBlock;384+ intriParams.srcDnMatrixStride = self_->ctx.convTilingData->coutOffsetBlock;
385- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;385+ intriParams.srcDValue = self_->ctx.convTilingData->kernelHxkernelW;
386- intriParams.dstNzC0Stride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW *386+ intriParams.dstNzC0Stride = self_->ctx.convTilingData->kernelHxkernelW *
387- self_->ctx.convTilingData->convApiTiling.nBL1;387+ self_->ctx.convTilingData->nBL1;
388- intriParams.dstNzNStride = self_->ctx.convTilingData->convApiTiling.nBL1;388+ intriParams.dstNzNStride = self_->ctx.convTilingData->nBL1;
389 intriParams.dstNzMatrixStride = Intf::k0;389 intriParams.dstNzMatrixStride = Intf::k0;
390 DataCopy<typename Intf::WeightT>(self_->ctx.bl1, self_->ctx.bgm[bL1GmOffset], intriParams);390 DataCopy<typename Intf::WeightT>(self_->ctx.bl1, self_->ctx.bgm[bL1GmOffset], intriParams);
391 }391 }
392 392 
393 __aicore__ inline void LoadBL1DataHWC(uint64_t kBL1Iter, uint64_t nBL1Iter)393 __aicore__ inline void LoadBL1DataHWC(uint64_t kBL1Iter, uint64_t nBL1Iter)
394 {394 {
395- uint64_t bL1GmOffset = kBL1Iter * self_->ctx.convTilingData->convApiTiling.cinBInCore *395+ uint64_t bL1GmOffset = kBL1Iter * self_->ctx.convTilingData->cinBInCore *
396- self_->ctx.convTilingData->convApiTiling.orgCo;396+ self_->ctx.convTilingData->orgCo;
397 if constexpr (Intf::hasNL1IterFlag) {397 if constexpr (Intf::hasNL1IterFlag) {
398- bL1GmOffset += nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1;398+ bL1GmOffset += nBL1Iter * self_->ctx.convTilingData->nBL1;
399 self_->ctx.currentNBL1 = nBL1Iter == self_->ctx.maxNBL1Iter ?399 self_->ctx.currentNBL1 = nBL1Iter == self_->ctx.maxNBL1Iter ?
400- self_->ctx.nBL1Tail : self_->ctx.convTilingData->convApiTiling.nBL1;400+ self_->ctx.nBL1Tail : self_->ctx.convTilingData->nBL1;
401 }401 }
402 402 
403 Dn2NzParams intriParams;403 Dn2NzParams intriParams;
404- intriParams.dnNum = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;404+ intriParams.dnNum = self_->ctx.convTilingData->kernelHxkernelW;
405 intriParams.nValue = self_->ctx.currentNBL1;405 intriParams.nValue = self_->ctx.currentNBL1;
406 intriParams.dValue = kBL1Iter == self_->ctx.maxKBL1Iter ?406 intriParams.dValue = kBL1Iter == self_->ctx.maxKBL1Iter ?
407- self_->ctx.convTilingData->convApiTiling.cinBTailInCore :407+ self_->ctx.convTilingData->cinBTailInCore :
408- self_->ctx.convTilingData->convApiTiling.cinBInCore;408+ self_->ctx.convTilingData->cinBInCore;
409 intriParams.srcDnMatrixStride = 409 intriParams.srcDnMatrixStride =
410- self_->ctx.convTilingData->convApiTiling.orgCi / self_->ctx.convTilingData->convApiTiling.groups *410+ self_->ctx.convTilingData->orgCi / self_->ctx.convTilingData->groups *
411- self_->ctx.convTilingData->convApiTiling.orgCo;411+ self_->ctx.convTilingData->orgCo;
412- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.orgCo;412+ intriParams.srcDValue = self_->ctx.convTilingData->orgCo;
413- intriParams.dstNzC0Stride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW *413+ intriParams.dstNzC0Stride = self_->ctx.convTilingData->kernelHxkernelW *
414- self_->ctx.convTilingData->convApiTiling.nBL1;414+ self_->ctx.convTilingData->nBL1;
415 intriParams.dstNzNStride = 1;415 intriParams.dstNzNStride = 1;
416- intriParams.dstNzMatrixStride = self_->ctx.convTilingData->convApiTiling.nBL1 * Intf::k0;416+ intriParams.dstNzMatrixStride = self_->ctx.convTilingData->nBL1 * Intf::k0;
417 DataCopy<typename Intf::WeightT>(self_->ctx.bl1, self_->ctx.bgm[bL1GmOffset], intriParams);417 DataCopy<typename Intf::WeightT>(self_->ctx.bl1, self_->ctx.bgm[bL1GmOffset], intriParams);
418 }418 }
419 419 
@@ -422,80 +422,80 @@ private:
422 uint64_t bL1GmOffset;422 uint64_t bL1GmOffset;
423 if constexpr (!Intf::hasNL1IterFlag) {423 if constexpr (!Intf::hasNL1IterFlag) {
424 bL1GmOffset = 424 bL1GmOffset =
425- self_->ctx.cinBL1Iter * self_->ctx.cinBL1 * self_->ctx.convTilingData->convApiTiling.kernelHxkernelW +425+ self_->ctx.cinBL1Iter * self_->ctx.cinBL1 * self_->ctx.convTilingData->kernelHxkernelW +
426- self_->ctx.khBL1Iter * self_->ctx.convTilingData->convApiTiling.khL1 *426+ self_->ctx.khBL1Iter * self_->ctx.convTilingData->khL1 *
427- self_->ctx.convTilingData->convApiTiling.kernelW +427+ self_->ctx.convTilingData->kernelW +
428- self_->ctx.kwBL1Iter * self_->ctx.convTilingData->convApiTiling.kwL1;428+ self_->ctx.kwBL1Iter * self_->ctx.convTilingData->kwL1;
429 } else {429 } else {
430 bL1GmOffset = 430 bL1GmOffset =
431- nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 *431+ nBL1Iter * self_->ctx.convTilingData->nBL1 *
432- self_->ctx.convTilingData->convApiTiling.coutOffsetBlock +432+ self_->ctx.convTilingData->coutOffsetBlock +
433- self_->ctx.cinBL1Iter * self_->ctx.cinBL1 * self_->ctx.convTilingData->convApiTiling.kernelHxkernelW +433+ self_->ctx.cinBL1Iter * self_->ctx.cinBL1 * self_->ctx.convTilingData->kernelHxkernelW +
434- self_->ctx.khBL1Iter * self_->ctx.convTilingData->convApiTiling.khL1 *434+ self_->ctx.khBL1Iter * self_->ctx.convTilingData->khL1 *
435- self_->ctx.convTilingData->convApiTiling.kernelW +435+ self_->ctx.convTilingData->kernelW +
436- self_->ctx.kwBL1Iter * self_->ctx.convTilingData->convApiTiling.kwL1;436+ self_->ctx.kwBL1Iter * self_->ctx.convTilingData->kwL1;
437 self_->ctx.currentNBL1 = nBL1Iter == self_->ctx.maxNBL1Iter ?437 self_->ctx.currentNBL1 = nBL1Iter == self_->ctx.maxNBL1Iter ?
438- self_->ctx.nBL1Tail : self_->ctx.convTilingData->convApiTiling.nBL1;438+ self_->ctx.nBL1Tail : self_->ctx.convTilingData->nBL1;
439 }439 }
440 Dn2NzParams intriParams;440 Dn2NzParams intriParams;
441 intriParams.dnNum = self_->ctx.currentNBL1;441 intriParams.dnNum = self_->ctx.currentNBL1;
442- intriParams.nValue = self_->ctx.convTilingData->convApiTiling.kwL1;442+ intriParams.nValue = self_->ctx.convTilingData->kwL1;
443 intriParams.dValue = self_->ctx.cinBL1Iter == self_->ctx.maxCinBL1Iter ?443 intriParams.dValue = self_->ctx.cinBL1Iter == self_->ctx.maxCinBL1Iter ?
444- self_->ctx.convTilingData->convApiTiling.cinBTailInCore :444+ self_->ctx.convTilingData->cinBTailInCore :
445- self_->ctx.convTilingData->convApiTiling.cinBInCore;445+ self_->ctx.convTilingData->cinBInCore;
446- intriParams.srcDnMatrixStride = self_->ctx.convTilingData->convApiTiling.coutOffsetBlock;446+ intriParams.srcDnMatrixStride = self_->ctx.convTilingData->coutOffsetBlock;
447- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;447+ intriParams.srcDValue = self_->ctx.convTilingData->kernelHxkernelW;
448- intriParams.dstNzC0Stride = self_->ctx.convTilingData->convApiTiling.kwL1 *448+ intriParams.dstNzC0Stride = self_->ctx.convTilingData->kwL1 *
449- self_->ctx.convTilingData->convApiTiling.khL1 *449+ self_->ctx.convTilingData->khL1 *
450- self_->ctx.convTilingData->convApiTiling.nBL1;450+ self_->ctx.convTilingData->nBL1;
451- intriParams.dstNzNStride = self_->ctx.convTilingData->convApiTiling.nBL1;451+ intriParams.dstNzNStride = self_->ctx.convTilingData->nBL1;
452 intriParams.dstNzMatrixStride = Intf::k0;452 intriParams.dstNzMatrixStride = Intf::k0;
453 uint64_t bl1DstOffset = 0;453 uint64_t bl1DstOffset = 0;
454- for (uint16_t khIterIdx = 0; khIterIdx < self_->ctx.convTilingData->convApiTiling.khL1; khIterIdx++) {454+ for (uint16_t khIterIdx = 0; khIterIdx < self_->ctx.convTilingData->khL1; khIterIdx++) {
455- bl1DstOffset = khIterIdx * self_->ctx.convTilingData->convApiTiling.kwL1 *455+ bl1DstOffset = khIterIdx * self_->ctx.convTilingData->kwL1 *
456- self_->ctx.convTilingData->convApiTiling.nBL1 * Intf::k0;456+ self_->ctx.convTilingData->nBL1 * Intf::k0;
457 DataCopy<typename Intf::WeightT>(self_->ctx.bl1[bl1DstOffset], self_->ctx.bgm[bL1GmOffset], intriParams);457 DataCopy<typename Intf::WeightT>(self_->ctx.bl1[bl1DstOffset], self_->ctx.bgm[bL1GmOffset], intriParams);
458- bL1GmOffset += self_->ctx.convTilingData->convApiTiling.kernelW;458+ bL1GmOffset += self_->ctx.convTilingData->kernelW;
459 }459 }
460 }460 }
461 461 
462 __aicore__ inline void LoadBL1DataKernelSplitHWC(uint64_t kBL1Iter, uint64_t nBL1Iter)462 __aicore__ inline void LoadBL1DataKernelSplitHWC(uint64_t kBL1Iter, uint64_t nBL1Iter)
463 {463 {
464 uint64_t bL1GmOffset = 464 uint64_t bL1GmOffset =
465- self_->ctx.cinBL1Iter * self_->ctx.cinBL1 * self_->ctx.convTilingData->convApiTiling.orgCo +465+ self_->ctx.cinBL1Iter * self_->ctx.cinBL1 * self_->ctx.convTilingData->orgCo +
466- self_->ctx.khBL1Iter * self_->ctx.convTilingData->convApiTiling.khL1 *466+ self_->ctx.khBL1Iter * self_->ctx.convTilingData->khL1 *
467- self_->ctx.convTilingData->convApiTiling.kernelW * self_->ctx.convTilingData->convApiTiling.orgCi /467+ self_->ctx.convTilingData->kernelW * self_->ctx.convTilingData->orgCi /
468- self_->ctx.convTilingData->convApiTiling.groups * self_->ctx.convTilingData->convApiTiling.orgCo +468+ self_->ctx.convTilingData->groups * self_->ctx.convTilingData->orgCo +
469- self_->ctx.kwBL1Iter * self_->ctx.convTilingData->convApiTiling.kwL1 *469+ self_->ctx.kwBL1Iter * self_->ctx.convTilingData->kwL1 *
470- self_->ctx.convTilingData->convApiTiling.orgCi / self_->ctx.convTilingData->convApiTiling.groups *470+ self_->ctx.convTilingData->orgCi / self_->ctx.convTilingData->groups *
471- self_->ctx.convTilingData->convApiTiling.orgCo;471+ self_->ctx.convTilingData->orgCo;
472 if constexpr (Intf::hasNL1IterFlag) {472 if constexpr (Intf::hasNL1IterFlag) {
473- bL1GmOffset += nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1;473+ bL1GmOffset += nBL1Iter * self_->ctx.convTilingData->nBL1;
474 self_->ctx.currentNBL1 = nBL1Iter == self_->ctx.maxNBL1Iter ?474 self_->ctx.currentNBL1 = nBL1Iter == self_->ctx.maxNBL1Iter ?
475- self_->ctx.nBL1Tail : self_->ctx.convTilingData->convApiTiling.nBL1;475+ self_->ctx.nBL1Tail : self_->ctx.convTilingData->nBL1;
476 }476 }
477 477 
478 Dn2NzParams intriParams;478 Dn2NzParams intriParams;
479- intriParams.dnNum = self_->ctx.convTilingData->convApiTiling.kwL1;479+ intriParams.dnNum = self_->ctx.convTilingData->kwL1;
480 intriParams.nValue = self_->ctx.currentNBL1;480 intriParams.nValue = self_->ctx.currentNBL1;
481 intriParams.dValue = self_->ctx.cinBL1Iter == self_->ctx.maxCinBL1Iter ?481 intriParams.dValue = self_->ctx.cinBL1Iter == self_->ctx.maxCinBL1Iter ?
482- self_->ctx.convTilingData->convApiTiling.cinBTailInCore :482+ self_->ctx.convTilingData->cinBTailInCore :
483- self_->ctx.convTilingData->convApiTiling.cinBInCore;483+ self_->ctx.convTilingData->cinBInCore;
484- intriParams.srcDnMatrixStride = self_->ctx.convTilingData->convApiTiling.orgCi /484+ intriParams.srcDnMatrixStride = self_->ctx.convTilingData->orgCi /
485- self_->ctx.convTilingData->convApiTiling.groups * self_->ctx.convTilingData->convApiTiling.orgCo;485+ self_->ctx.convTilingData->groups * self_->ctx.convTilingData->orgCo;
486- intriParams.srcDValue = self_->ctx.convTilingData->convApiTiling.orgCo;486+ intriParams.srcDValue = self_->ctx.convTilingData->orgCo;
487- intriParams.dstNzC0Stride = self_->ctx.convTilingData->convApiTiling.khL1 *487+ intriParams.dstNzC0Stride = self_->ctx.convTilingData->khL1 *
488- self_->ctx.convTilingData->convApiTiling.kwL1 * self_->ctx.convTilingData->convApiTiling.nBL1;488+ self_->ctx.convTilingData->kwL1 * self_->ctx.convTilingData->nBL1;
489 intriParams.dstNzNStride = 1;489 intriParams.dstNzNStride = 1;
490- intriParams.dstNzMatrixStride = self_->ctx.convTilingData->convApiTiling.nBL1 * Intf::k0;490+ intriParams.dstNzMatrixStride = self_->ctx.convTilingData->nBL1 * Intf::k0;
491 uint64_t bl1DstOffset = 0;491 uint64_t bl1DstOffset = 0;
492- for (uint16_t khIterIdx = 0; khIterIdx < self_->ctx.convTilingData->convApiTiling.khL1; khIterIdx++) {492+ for (uint16_t khIterIdx = 0; khIterIdx < self_->ctx.convTilingData->khL1; khIterIdx++) {
493- bl1DstOffset = khIterIdx * self_->ctx.convTilingData->convApiTiling.kwL1 *493+ bl1DstOffset = khIterIdx * self_->ctx.convTilingData->kwL1 *
494- self_->ctx.convTilingData->convApiTiling.nBL1 * Intf::k0;494+ self_->ctx.convTilingData->nBL1 * Intf::k0;
495 DataCopy<typename Intf::WeightT>(self_->ctx.bl1[bl1DstOffset], self_->ctx.bgm[bL1GmOffset], intriParams);495 DataCopy<typename Intf::WeightT>(self_->ctx.bl1[bl1DstOffset], self_->ctx.bgm[bL1GmOffset], intriParams);
496 bL1GmOffset +=496 bL1GmOffset +=
497- self_->ctx.convTilingData->convApiTiling.kernelW * self_->ctx.convTilingData->convApiTiling.orgCi /497+ self_->ctx.convTilingData->kernelW * self_->ctx.convTilingData->orgCi /
498- self_->ctx.convTilingData->convApiTiling.groups * self_->ctx.convTilingData->convApiTiling.orgCo;498+ self_->ctx.convTilingData->groups * self_->ctx.convTilingData->orgCo;
499 }499 }
500 }500 }
501 501 
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2_instr_c04_impl.h+25-25
@@ -73,32 +73,32 @@ private:
73 {73 {
74 if (unlikely(self_->ctx.isFirstIterate)) {74 if (unlikely(self_->ctx.isFirstIterate)) {
75 // NDDMA Loop0 params75 // NDDMA Loop0 params
76- copyParams.loopInfo.loopSize[NDDMA_LOOP0_INDEX] = self_->ctx.convTilingData->convApiTiling.singleCoreCi;76+ copyParams.loopInfo.loopSize[NDDMA_LOOP0_INDEX] = self_->ctx.convTilingData->singleCoreCi;
77 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] =77 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] =
78- self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;78+ self_->ctx.convTilingData->kernelHxkernelW;
79 copyParams.loopInfo.loopDstStride[NDDMA_LOOP0_INDEX] = 1;79 copyParams.loopInfo.loopDstStride[NDDMA_LOOP0_INDEX] = 1;
80 copyParams.loopInfo.loopRpSize[NDDMA_LOOP0_INDEX] =80 copyParams.loopInfo.loopRpSize[NDDMA_LOOP0_INDEX] =
81- C04_CIN_SIZE - self_->ctx.convTilingData->convApiTiling.singleCoreCi;81+ C04_CIN_SIZE - self_->ctx.convTilingData->singleCoreCi;
82 // NDDMA Loop1 params82 // NDDMA Loop1 params
83- copyParams.loopInfo.loopSize[NDDMA_LOOP1_INDEX] = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;83+ copyParams.loopInfo.loopSize[NDDMA_LOOP1_INDEX] = self_->ctx.convTilingData->kernelHxkernelW;
84 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = 1;84 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = 1;
85 copyParams.loopInfo.loopDstStride[NDDMA_LOOP1_INDEX] = C04_CIN_SIZE;85 copyParams.loopInfo.loopDstStride[NDDMA_LOOP1_INDEX] = C04_CIN_SIZE;
86 // NDDMA Loop2 params86 // NDDMA Loop2 params
87 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP2_INDEX] =87 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP2_INDEX] =
88- self_->ctx.convTilingData->convApiTiling.kernelHxkernelW *88+ self_->ctx.convTilingData->kernelHxkernelW *
89- self_->ctx.convTilingData->convApiTiling.singleCoreCi;89+ self_->ctx.convTilingData->singleCoreCi;
90- copyParams.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = self_->ctx.convTilingData->convApiTiling.kBL1;90+ copyParams.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = self_->ctx.convTilingData->kBL1;
91 copyParams.constantValue = 0;91 copyParams.constantValue = 0;
92 }92 }
93 copyParams.loopInfo.loopSize[NDDMA_LOOP2_INDEX] = self_->ctx.currentUbNStep;93 copyParams.loopInfo.loopSize[NDDMA_LOOP2_INDEX] = self_->ctx.currentUbNStep;
94 copyParams.loopInfo.loopRpSize[NDDMA_LOOP2_INDEX] = self_->ctx.currentNLoopRpSize;94 copyParams.loopInfo.loopRpSize[NDDMA_LOOP2_INDEX] = self_->ctx.currentNLoopRpSize;
95 95 
96- srcOffset = (self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +96+ srcOffset = (self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
97- self_->ctx.vecNIter * self_->ctx.convTilingData->convApiTiling.bUbNStep) *97+ self_->ctx.vecNIter * self_->ctx.convTilingData->bUbNStep) *
98- self_->ctx.convTilingData->convApiTiling.coutOffsetBlock;98+ self_->ctx.convTilingData->coutOffsetBlock;
99 if constexpr (!Intf::bL1DBFlag) {99 if constexpr (!Intf::bL1DBFlag) {
100 if (self_->ctx.vecId == 1) {100 if (self_->ctx.vecId == 1) {
101- srcOffset += self_->ctx.nBL1Vec0 * self_->ctx.convTilingData->convApiTiling.coutOffsetBlock;101+ srcOffset += self_->ctx.nBL1Vec0 * self_->ctx.convTilingData->coutOffsetBlock;
102 }102 }
103 }103 }
104 }104 }
@@ -108,25 +108,25 @@ private:
108 if (unlikely(self_->ctx.isFirstIterate)) {108 if (unlikely(self_->ctx.isFirstIterate)) {
109 // NDDMA Loop0 params109 // NDDMA Loop0 params
110 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 1;110 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 1;
111- copyParams.loopInfo.loopDstStride[NDDMA_LOOP0_INDEX] = self_->ctx.convTilingData->convApiTiling.kBL1;111+ copyParams.loopInfo.loopDstStride[NDDMA_LOOP0_INDEX] = self_->ctx.convTilingData->kBL1;
112 // NDDMA Loop1 params112 // NDDMA Loop1 params
113- copyParams.loopInfo.loopSize[NDDMA_LOOP1_INDEX] = self_->ctx.convTilingData->convApiTiling.singleCoreCi;113+ copyParams.loopInfo.loopSize[NDDMA_LOOP1_INDEX] = self_->ctx.convTilingData->singleCoreCi;
114- copyParams.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = self_->ctx.convTilingData->convApiTiling.orgCo;114+ copyParams.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = self_->ctx.convTilingData->orgCo;
115 copyParams.loopInfo.loopDstStride[NDDMA_LOOP1_INDEX] = 1;115 copyParams.loopInfo.loopDstStride[NDDMA_LOOP1_INDEX] = 1;
116 copyParams.loopInfo.loopRpSize[NDDMA_LOOP1_INDEX] =116 copyParams.loopInfo.loopRpSize[NDDMA_LOOP1_INDEX] =
117- C04_CIN_SIZE - self_->ctx.convTilingData->convApiTiling.singleCoreCi;117+ C04_CIN_SIZE - self_->ctx.convTilingData->singleCoreCi;
118 // NDDMA Loop2 params118 // NDDMA Loop2 params
119- copyParams.loopInfo.loopSize[NDDMA_LOOP2_INDEX] = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;119+ copyParams.loopInfo.loopSize[NDDMA_LOOP2_INDEX] = self_->ctx.convTilingData->kernelHxkernelW;
120 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP2_INDEX] =120 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP2_INDEX] =
121- self_->ctx.convTilingData->convApiTiling.orgCo * self_->ctx.convTilingData->convApiTiling.singleCoreCi;121+ self_->ctx.convTilingData->orgCo * self_->ctx.convTilingData->singleCoreCi;
122 copyParams.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = C04_CIN_SIZE;122 copyParams.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = C04_CIN_SIZE;
123 copyParams.constantValue = 0;123 copyParams.constantValue = 0;
124 }124 }
125 copyParams.loopInfo.loopSize[NDDMA_LOOP0_INDEX] = self_->ctx.currentUbNStep;125 copyParams.loopInfo.loopSize[NDDMA_LOOP0_INDEX] = self_->ctx.currentUbNStep;
126 copyParams.loopInfo.loopRpSize[NDDMA_LOOP0_INDEX] = self_->ctx.currentNLoopRpSize;126 copyParams.loopInfo.loopRpSize[NDDMA_LOOP0_INDEX] = self_->ctx.currentNLoopRpSize;
127 127
128- srcOffset = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +128+ srcOffset = self_->ctx.nBL1Iter * self_->ctx.convTilingData->nBL1 +
129- self_->ctx.vecNIter * self_->ctx.convTilingData->convApiTiling.bUbNStep;129+ self_->ctx.vecNIter * self_->ctx.convTilingData->bUbNStep;
130 if constexpr (!Intf::bL1DBFlag) {130 if constexpr (!Intf::bL1DBFlag) {
131 if (self_->ctx.vecId == 1) {131 if (self_->ctx.vecId == 1) {
132 srcOffset += self_->ctx.nBL1Vec0;132 srcOffset += self_->ctx.nBL1Vec0;
@@ -159,11 +159,11 @@ public:
159 SetIndex();159 SetIndex();
160 }160 }
161 161 
162- uint16_t kLoopTimes = self_->ctx.convTilingData->convApiTiling.kBL1 / Intf::k0;162+ uint16_t kLoopTimes = self_->ctx.convTilingData->kBL1 / Intf::k0;
163 uint16_t nLoopTimes = self_->ctx.currentUbNStepAilgn / BLOCK_L0_N * CO0_LOOP_TIMES;163 uint16_t nLoopTimes = self_->ctx.currentUbNStepAilgn / BLOCK_L0_N * CO0_LOOP_TIMES;
164 164 
165 uint32_t srcKStride = Intf::k0;165 uint32_t srcKStride = Intf::k0;
166- uint32_t srcNStride = coPerReg * self_->ctx.convTilingData->convApiTiling.kBL1;166+ uint32_t srcNStride = coPerReg * self_->ctx.convTilingData->kBL1;
167 uint32_t dstKStride = Intf::k0 * self_->ctx.currentUbNStepAilgn;167 uint32_t dstKStride = Intf::k0 * self_->ctx.currentUbNStepAilgn;
168 uint32_t dstNStride = Intf::k0 * coPerReg;168 uint32_t dstNStride = Intf::k0 * coPerReg;
169 169 
@@ -213,7 +213,7 @@ private:
213 uint8_t dstOffset = Intf::k0;213 uint8_t dstOffset = Intf::k0;
214 uint8_t elesPerRepeat = Intf::k0;214 uint8_t elesPerRepeat = Intf::k0;
215 uint32_t maskL = Intf::k0;215 uint32_t maskL = Intf::k0;
216- IndexT nStride = static_cast<IndexT>(self_->ctx.convTilingData->convApiTiling.kBL1);216+ IndexT nStride = static_cast<IndexT>(self_->ctx.convTilingData->kBL1);
217 217 
218 __VEC_SCOPE__218 __VEC_SCOPE__
219 {219 {
@@ -254,13 +254,13 @@ public:
254 __aicore__ inline void LoadUB2L1()254 __aicore__ inline void LoadUB2L1()
255 {255 {
256 if (unlikely(self_->ctx.isFirstIterate)) {256 if (unlikely(self_->ctx.isFirstIterate)) {
257- copyParams.blockCount = self_->ctx.convTilingData->convApiTiling.kBL1 / Intf::k0;257+ copyParams.blockCount = self_->ctx.convTilingData->kBL1 / Intf::k0;
258 copyParams.srcStride = 0;258 copyParams.srcStride = 0;
259 }259 }
260 copyParams.blockLen = self_->ctx.currentUbNStepAilgn;260 copyParams.blockLen = self_->ctx.currentUbNStepAilgn;
261- copyParams.dstStride = self_->ctx.convTilingData->convApiTiling.nBL1 - self_->ctx.currentUbNStepAilgn;261+ copyParams.dstStride = self_->ctx.convTilingData->nBL1 - self_->ctx.currentUbNStepAilgn;
262 262 
263- uint64_t dstOffset = self_->ctx.vecNIter * self_->ctx.convTilingData->convApiTiling.bUbNStep * Intf::k0;263+ uint64_t dstOffset = self_->ctx.vecNIter * self_->ctx.convTilingData->bUbNStep * Intf::k0;
264 if (self_->ctx.vecId == 1) {264 if (self_->ctx.vecId == 1) {
265 if constexpr (Intf::bL1DBFlag) {265 if constexpr (Intf::bL1DBFlag) {
266 dstOffset += self_->ctx.bL1SpaceSize;266 dstOffset += self_->ctx.bL1SpaceSize;
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2_instr_dma_impl.h+71-71
@@ -33,14 +33,14 @@ public:
33 {33 {
34 self_ = self;34 self_ = self;
35 if constexpr (Intf::formatFmap == ConvFormat::NCHW) {35 if constexpr (Intf::formatFmap == ConvFormat::NCHW) {
36- srcCiStride = self_->ctx.convTilingData->convApiTiling.orgHixWi * Intf::k0;36+ srcCiStride = self_->ctx.convTilingData->orgHixWi * Intf::k0;
37 } else if constexpr (Intf::formatFmap == ConvFormat::NHWC) {37 } else if constexpr (Intf::formatFmap == ConvFormat::NHWC) {
38 srcCiStride = Intf::k0;38 srcCiStride = Intf::k0;
39 }39 }
40 srcKhStride =40 srcKhStride =
41- self_->ctx.convTilingData->convApiTiling.khUb * self_->ctx.convTilingData->convApiTiling.dilationH;41+ self_->ctx.convTilingData->khUb * self_->ctx.convTilingData->dilationH;
42 srcKwStride =42 srcKwStride =
43- self_->ctx.convTilingData->convApiTiling.kwUb * self_->ctx.convTilingData->convApiTiling.dilationW;43+ self_->ctx.convTilingData->kwUb * self_->ctx.convTilingData->dilationW;
44 }44 }
45 45 
46 __aicore__ inline void LoadGM2UB()46 __aicore__ inline void LoadGM2UB()
@@ -49,28 +49,28 @@ public:
49 self_->ctx.khAL1Iter = self_->ctx.kAL1Iter / self_->ctx.ddr2L1LoopKw % self_->ctx.ddr2L1LoopKh;49 self_->ctx.khAL1Iter = self_->ctx.kAL1Iter / self_->ctx.ddr2L1LoopKw % self_->ctx.ddr2L1LoopKh;
50 self_->ctx.cinAL1Iter = (self_->ctx.kAL1Iter / (self_->ctx.ddr2L1LoopKw * self_->ctx.ddr2L1LoopKh)) %50 self_->ctx.cinAL1Iter = (self_->ctx.kAL1Iter / (self_->ctx.ddr2L1LoopKw * self_->ctx.ddr2L1LoopKh)) %
51 self_->ctx.cinAL1LoopTimes;51 self_->ctx.cinAL1LoopTimes;
52- srckhAL1IterOffset = self_->ctx.khAL1Iter * self_->ctx.convTilingData->convApiTiling.dilationH *52+ srckhAL1IterOffset = self_->ctx.khAL1Iter * self_->ctx.convTilingData->dilationH *
53- self_->ctx.convTilingData->convApiTiling.khL1 + self_->ctx.vecKhIter * srcKhStride;53+ self_->ctx.convTilingData->khL1 + self_->ctx.vecKhIter * srcKhStride;
54- srckwAL1IterOffset = self_->ctx.kwAL1Iter * self_->ctx.convTilingData->convApiTiling.dilationW *54+ srckwAL1IterOffset = self_->ctx.kwAL1Iter * self_->ctx.convTilingData->dilationW *
55- self_->ctx.convTilingData->convApiTiling.kwL1 + self_->ctx.vecKwIter * srcKwStride;55+ self_->ctx.convTilingData->kwL1 + self_->ctx.vecKwIter * srcKwStride;
56 if (unlikely(self_->ctx.isFirstIterate)) {56 if (unlikely(self_->ctx.isFirstIterate)) {
57 if constexpr (Intf::formatFmap == ConvFormat::NCHW) {57 if constexpr (Intf::formatFmap == ConvFormat::NCHW) {
58 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 58 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] =
59- self_->ctx.convTilingData->convApiTiling.orgHixWi;59+ self_->ctx.convTilingData->orgHixWi;
60 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = 60 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] =
61- self_->ctx.convTilingData->convApiTiling.dilationW;61+ self_->ctx.convTilingData->dilationW;
62 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP2_INDEX] = 62 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP2_INDEX] =
63- self_->ctx.convTilingData->convApiTiling.dilationH *63+ self_->ctx.convTilingData->dilationH *
64- self_->ctx.convTilingData->convApiTiling.orgWi;64+ self_->ctx.convTilingData->orgWi;
65 } else {65 } else {
66 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 1;66 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 1;
67 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] =67 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] =
68- self_->ctx.convTilingData->convApiTiling.dilationW *68+ self_->ctx.convTilingData->dilationW *
69- self_->ctx.convTilingData->convApiTiling.orgCi;69+ self_->ctx.convTilingData->orgCi;
70 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP2_INDEX] =70 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP2_INDEX] =
71- self_->ctx.convTilingData->convApiTiling.dilationH *71+ self_->ctx.convTilingData->dilationH *
72- self_->ctx.convTilingData->convApiTiling.orgWi *72+ self_->ctx.convTilingData->orgWi *
73- self_->ctx.convTilingData->convApiTiling.orgCi;73+ self_->ctx.convTilingData->orgCi;
74 }74 }
75 copyParams.loopInfo.loopDstStride[NDDMA_LOOP0_INDEX] = 1;75 copyParams.loopInfo.loopDstStride[NDDMA_LOOP0_INDEX] = 1;
76 }76 }
@@ -90,9 +90,9 @@ public:
90 uint32_t baseOffset = self_->ctx.batchIter * self_->ctx.fmapOneBatchSize +90 uint32_t baseOffset = self_->ctx.batchIter * self_->ctx.fmapOneBatchSize +
91 self_->ctx.vecCi1Iter * srcCiStride;91 self_->ctx.vecCi1Iter * srcCiStride;
92 if constexpr (Intf::formatFmap == ConvFormat::NCHW) {92 if constexpr (Intf::formatFmap == ConvFormat::NCHW) {
93- baseOffset += self_->ctx.cinAL1Iter * self_->ctx.convTilingData->convApiTiling.cinOffsetBlockInGM;93+ baseOffset += self_->ctx.cinAL1Iter * self_->ctx.convTilingData->cinOffsetBlockInGM;
94 if (self_->ctx.vecId == 1) {94 if (self_->ctx.vecId == 1) {
95- baseOffset += self_->ctx.currentVec0Ci * self_->ctx.convTilingData->convApiTiling.orgHixWi;95+ baseOffset += self_->ctx.currentVec0Ci * self_->ctx.convTilingData->orgHixWi;
96 }96 }
97 } else {97 } else {
98 baseOffset += self_->ctx.cinAL1Iter * self_->ctx.cinAL1;98 baseOffset += self_->ctx.cinAL1Iter * self_->ctx.cinAL1;
@@ -101,13 +101,13 @@ public:
101 }101 }
102 }102 }
103 103
104- hoL1Idx = self_->ctx.hoAL1Iter * self_->ctx.convTilingData->convApiTiling.hoL1 *104+ hoL1Idx = self_->ctx.hoAL1Iter * self_->ctx.convTilingData->hoL1 *
105- self_->ctx.convTilingData->convApiTiling.strideH;105+ self_->ctx.convTilingData->strideH;
106- woL1Idx = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 *106+ woL1Idx = self_->ctx.woAL1Iter * self_->ctx.convTilingData->woL1 *
107- self_->ctx.convTilingData->convApiTiling.strideW;107+ self_->ctx.convTilingData->strideW;
108 uint32_t dstHoStride = AlignB(self_->ctx.currentWoL1, BLOCK_L0_M) * Intf::k0;108 uint32_t dstHoStride = AlignB(self_->ctx.currentWoL1, BLOCK_L0_M) * Intf::k0;
109 uint32_t dstKwStride = self_->ctx.currentHoL1xWoL1Align * Intf::k0;109 uint32_t dstKwStride = self_->ctx.currentHoL1xWoL1Align * Intf::k0;
110- uint32_t dstKhStride = dstKwStride * self_->ctx.convTilingData->convApiTiling.kwUb;110+ uint32_t dstKhStride = dstKwStride * self_->ctx.convTilingData->kwUb;
111 copyParams.loopInfo.loopDstStride[NDDMA_LOOP1_INDEX] = dstKwStride;111 copyParams.loopInfo.loopDstStride[NDDMA_LOOP1_INDEX] = dstKwStride;
112 copyParams.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = dstKhStride;112 copyParams.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = dstKhStride;
113 uint32_t srcOffset = 0;113 uint32_t srcOffset = 0;
@@ -118,17 +118,17 @@ public:
118 continue;118 continue;
119 }119 }
120 copyParams.loopInfo.loopSize[NDDMA_LOOP1_INDEX] =120 copyParams.loopInfo.loopSize[NDDMA_LOOP1_INDEX] =
121- self_->ctx.convTilingData->convApiTiling.kwUb - kwUbOnPadLeft - kwUbOnPadRight;121+ self_->ctx.convTilingData->kwUb - kwUbOnPadLeft - kwUbOnPadRight;
122 copyParams.loopInfo.loopSize[NDDMA_LOOP2_INDEX] =122 copyParams.loopInfo.loopSize[NDDMA_LOOP2_INDEX] =
123- self_->ctx.convTilingData->convApiTiling.khUb - khUbOnPadTop - khUbOnPadBottom;123+ self_->ctx.convTilingData->khUb - khUbOnPadTop - khUbOnPadBottom;
124 if constexpr (Intf::formatFmap == ConvFormat::NCHW) {124 if constexpr (Intf::formatFmap == ConvFormat::NCHW) {
125 srcOffset = baseOffset +125 srcOffset = baseOffset +
126- hIdxWithPadStart * self_->ctx.convTilingData->convApiTiling.orgWi + wIdxWithPadStart;126+ hIdxWithPadStart * self_->ctx.convTilingData->orgWi + wIdxWithPadStart;
127 } else {127 } else {
128 srcOffset = baseOffset +128 srcOffset = baseOffset +
129- hIdxWithPadStart * self_->ctx.convTilingData->convApiTiling.orgWi *129+ hIdxWithPadStart * self_->ctx.convTilingData->orgWi *
130- self_->ctx.convTilingData->convApiTiling.orgCi +130+ self_->ctx.convTilingData->orgCi +
131- wIdxWithPadStart * self_->ctx.convTilingData->convApiTiling.orgCi;131+ wIdxWithPadStart * self_->ctx.convTilingData->orgCi;
132 }132 }
133 uint32_t dstOffset = hoIdx * dstHoStride + woIdx * Intf::k0;133 uint32_t dstOffset = hoIdx * dstHoStride + woIdx * Intf::k0;
134 if (unlikely(khUbOnPadTop != 0)) {134 if (unlikely(khUbOnPadTop != 0)) {
@@ -151,32 +151,32 @@ private:
151 151 
152 __aicore__ inline bool IsOnPadAndUpdateKernelSize(uint16_t hoIdx, uint16_t woIdx)152 __aicore__ inline bool IsOnPadAndUpdateKernelSize(uint16_t hoIdx, uint16_t woIdx)
153 {153 {
154- hIdxWithPadStart = hoL1Idx + hoIdx * self_->ctx.convTilingData->convApiTiling.strideH + srckhAL1IterOffset;154+ hIdxWithPadStart = hoL1Idx + hoIdx * self_->ctx.convTilingData->strideH + srckhAL1IterOffset;
155- wIdxWithPadStart = woL1Idx + woIdx * self_->ctx.convTilingData->convApiTiling.strideW + srckwAL1IterOffset;155+ wIdxWithPadStart = woL1Idx + woIdx * self_->ctx.convTilingData->strideW + srckwAL1IterOffset;
156 156 
157 hIdxWithPadStartOrg = hIdxWithPadStart + self_->ctx.hiStartPos;157 hIdxWithPadStartOrg = hIdxWithPadStart + self_->ctx.hiStartPos;
158 wIdxWithPadStartOrg = wIdxWithPadStart + self_->ctx.wiStartPos;158 wIdxWithPadStartOrg = wIdxWithPadStart + self_->ctx.wiStartPos;
159 hIdxWithPadStart = self_->ctx.hiStartPos < 0 ? hIdxWithPadStart + self_->ctx.hiStartPos : hIdxWithPadStart;159 hIdxWithPadStart = self_->ctx.hiStartPos < 0 ? hIdxWithPadStart + self_->ctx.hiStartPos : hIdxWithPadStart;
160 wIdxWithPadStart = self_->ctx.wiStartPos < 0 ? wIdxWithPadStart + self_->ctx.wiStartPos : wIdxWithPadStart;160 wIdxWithPadStart = self_->ctx.wiStartPos < 0 ? wIdxWithPadStart + self_->ctx.wiStartPos : wIdxWithPadStart;
161- hIdxWithPadEndOrg = hIdxWithPadStartOrg + (self_->ctx.convTilingData->convApiTiling.khUb - 1) *161+ hIdxWithPadEndOrg = hIdxWithPadStartOrg + (self_->ctx.convTilingData->khUb - 1) *
162- self_->ctx.convTilingData->convApiTiling.dilationH;162+ self_->ctx.convTilingData->dilationH;
163- wIdxWithPadEndOrg = wIdxWithPadStartOrg + (self_->ctx.convTilingData->convApiTiling.kwUb - 1) *163+ wIdxWithPadEndOrg = wIdxWithPadStartOrg + (self_->ctx.convTilingData->kwUb - 1) *
164- self_->ctx.convTilingData->convApiTiling.dilationW;164+ self_->ctx.convTilingData->dilationW;
165 165 
166 kwUbOnPadLeft = 0;166 kwUbOnPadLeft = 0;
167 kwUbOnPadRight = 0;167 kwUbOnPadRight = 0;
168 khUbOnPadTop = 0;168 khUbOnPadTop = 0;
169 khUbOnPadBottom = 0;169 khUbOnPadBottom = 0;
170 if (unlikely(wIdxWithPadEndOrg < 0 ||170 if (unlikely(wIdxWithPadEndOrg < 0 ||
171- wIdxWithPadStartOrg >= static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgWi) ||171+ wIdxWithPadStartOrg >= static_cast<int64_t>(self_->ctx.convTilingData->orgWi) ||
172 hIdxWithPadEndOrg < 0 ||172 hIdxWithPadEndOrg < 0 ||
173- hIdxWithPadStartOrg >= static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgHi))) {173+ hIdxWithPadStartOrg >= static_cast<int64_t>(self_->ctx.convTilingData->orgHi))) {
174 return true;174 return true;
175 }175 }
176 if (likely(hIdxWithPadStartOrg >= 0 &&176 if (likely(hIdxWithPadStartOrg >= 0 &&
177- hIdxWithPadEndOrg < static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgHi) &&177+ hIdxWithPadEndOrg < static_cast<int64_t>(self_->ctx.convTilingData->orgHi) &&
178 wIdxWithPadStartOrg >= 0 &&178 wIdxWithPadStartOrg >= 0 &&
179- wIdxWithPadEndOrg < static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgWi))) {179+ wIdxWithPadEndOrg < static_cast<int64_t>(self_->ctx.convTilingData->orgWi))) {
180 return false;180 return false;
181 }181 }
182 UpdateRealKernelSize();182 UpdateRealKernelSize();
@@ -188,44 +188,44 @@ private:
188 {188 {
189 // update real khUbSize in fmap.189 // update real khUbSize in fmap.
190 if (hIdxWithPadStartOrg < 0 &&190 if (hIdxWithPadStartOrg < 0 &&
191- hIdxWithPadEndOrg < static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgHi)) {191+ hIdxWithPadEndOrg < static_cast<int64_t>(self_->ctx.convTilingData->orgHi)) {
192- khUbOnPadTop = (-hIdxWithPadStartOrg - 1) / self_->ctx.convTilingData->convApiTiling.dilationH + 1;192+ khUbOnPadTop = (-hIdxWithPadStartOrg - 1) / self_->ctx.convTilingData->dilationH + 1;
193- hIdxWithPadStart += khUbOnPadTop * self_->ctx.convTilingData->convApiTiling.dilationH;193+ hIdxWithPadStart += khUbOnPadTop * self_->ctx.convTilingData->dilationH;
194 } else if (hIdxWithPadStartOrg >= 0 &&194 } else if (hIdxWithPadStartOrg >= 0 &&
195- hIdxWithPadEndOrg >= static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgHi)) {195+ hIdxWithPadEndOrg >= static_cast<int64_t>(self_->ctx.convTilingData->orgHi)) {
196 uint32_t khUbOnFmapCountTemp = 196 uint32_t khUbOnFmapCountTemp =
197- (static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgHi) - 1 - hIdxWithPadStartOrg) /197+ (static_cast<int64_t>(self_->ctx.convTilingData->orgHi) - 1 - hIdxWithPadStartOrg) /
198- self_->ctx.convTilingData->convApiTiling.dilationH + 1;198+ self_->ctx.convTilingData->dilationH + 1;
199- khUbOnPadBottom = self_->ctx.convTilingData->convApiTiling.khUb - khUbOnFmapCountTemp;199+ khUbOnPadBottom = self_->ctx.convTilingData->khUb - khUbOnFmapCountTemp;
200 } else if (hIdxWithPadStartOrg < 0 &&200 } else if (hIdxWithPadStartOrg < 0 &&
201- hIdxWithPadEndOrg >= static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgHi)) {201+ hIdxWithPadEndOrg >= static_cast<int64_t>(self_->ctx.convTilingData->orgHi)) {
202- khUbOnPadTop = (-hIdxWithPadStartOrg - 1) / self_->ctx.convTilingData->convApiTiling.dilationH + 1;202+ khUbOnPadTop = (-hIdxWithPadStartOrg - 1) / self_->ctx.convTilingData->dilationH + 1;
203- hIdxWithPadStart += khUbOnPadTop * self_->ctx.convTilingData->convApiTiling.dilationH;203+ hIdxWithPadStart += khUbOnPadTop * self_->ctx.convTilingData->dilationH;
204 uint32_t khUbOnPadTopAndFmapCountTemp =204 uint32_t khUbOnPadTopAndFmapCountTemp =
205- (static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgHi) - 1 - hIdxWithPadStartOrg) /205+ (static_cast<int64_t>(self_->ctx.convTilingData->orgHi) - 1 - hIdxWithPadStartOrg) /
206- self_->ctx.convTilingData->convApiTiling.dilationH + 1;206+ self_->ctx.convTilingData->dilationH + 1;
207- khUbOnPadBottom = self_->ctx.convTilingData->convApiTiling.khUb - khUbOnPadTopAndFmapCountTemp;207+ khUbOnPadBottom = self_->ctx.convTilingData->khUb - khUbOnPadTopAndFmapCountTemp;
208 }208 }
209 209 
210 // update real kwUbSize in fmap.210 // update real kwUbSize in fmap.
211 if (wIdxWithPadStartOrg < 0 &&211 if (wIdxWithPadStartOrg < 0 &&
212- wIdxWithPadEndOrg < static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgWi)) {212+ wIdxWithPadEndOrg < static_cast<int64_t>(self_->ctx.convTilingData->orgWi)) {
213- kwUbOnPadLeft = (-wIdxWithPadStartOrg - 1) / self_->ctx.convTilingData->convApiTiling.dilationW + 1;213+ kwUbOnPadLeft = (-wIdxWithPadStartOrg - 1) / self_->ctx.convTilingData->dilationW + 1;
214- wIdxWithPadStart += kwUbOnPadLeft * self_->ctx.convTilingData->convApiTiling.dilationW;214+ wIdxWithPadStart += kwUbOnPadLeft * self_->ctx.convTilingData->dilationW;
215 } else if (wIdxWithPadStartOrg >= 0 &&215 } else if (wIdxWithPadStartOrg >= 0 &&
216- wIdxWithPadEndOrg >= static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgWi)) {216+ wIdxWithPadEndOrg >= static_cast<int64_t>(self_->ctx.convTilingData->orgWi)) {
217 uint32_t kwUbOnFmapCountTemp =217 uint32_t kwUbOnFmapCountTemp =
218- (static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgWi) - 1 - wIdxWithPadStartOrg) /218+ (static_cast<int64_t>(self_->ctx.convTilingData->orgWi) - 1 - wIdxWithPadStartOrg) /
219- self_->ctx.convTilingData->convApiTiling.dilationW + 1;219+ self_->ctx.convTilingData->dilationW + 1;
220- kwUbOnPadRight = self_->ctx.convTilingData->convApiTiling.kwUb - kwUbOnFmapCountTemp;220+ kwUbOnPadRight = self_->ctx.convTilingData->kwUb - kwUbOnFmapCountTemp;
221 } else if (wIdxWithPadStartOrg < 0 &&221 } else if (wIdxWithPadStartOrg < 0 &&
222- wIdxWithPadEndOrg >= static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgWi)) {222+ wIdxWithPadEndOrg >= static_cast<int64_t>(self_->ctx.convTilingData->orgWi)) {
223- kwUbOnPadLeft = (-wIdxWithPadStartOrg - 1) / self_->ctx.convTilingData->convApiTiling.dilationW + 1;223+ kwUbOnPadLeft = (-wIdxWithPadStartOrg - 1) / self_->ctx.convTilingData->dilationW + 1;
224- wIdxWithPadStart += kwUbOnPadLeft * self_->ctx.convTilingData->convApiTiling.dilationW;224+ wIdxWithPadStart += kwUbOnPadLeft * self_->ctx.convTilingData->dilationW;
225 uint32_t kwUbOnPadLeftAndFmapCountTemp =225 uint32_t kwUbOnPadLeftAndFmapCountTemp =
226- (static_cast<int64_t>(self_->ctx.convTilingData->convApiTiling.orgWi) - 1 - wIdxWithPadStartOrg) /226+ (static_cast<int64_t>(self_->ctx.convTilingData->orgWi) - 1 - wIdxWithPadStartOrg) /
227- self_->ctx.convTilingData->convApiTiling.dilationW + 1;227+ self_->ctx.convTilingData->dilationW + 1;
228- kwUbOnPadRight = self_->ctx.convTilingData->convApiTiling.kwUb - kwUbOnPadLeftAndFmapCountTemp;228+ kwUbOnPadRight = self_->ctx.convTilingData->kwUb - kwUbOnPadLeftAndFmapCountTemp;
229 } 229 }
230 }230 }
231 231 
@@ -269,21 +269,21 @@ public:
269 self_ = self;269 self_ = self;
270 270 
271 dstStrideBase = 271 dstStrideBase =
272- (self_->ctx.convTilingData->convApiTiling.kwL1 - self_->ctx.convTilingData->convApiTiling.kwUb) * Intf::k0;272+ (self_->ctx.convTilingData->kwL1 - self_->ctx.convTilingData->kwUb) * Intf::k0;
273 ci1StrideBase = 273 ci1StrideBase =
274- self_->ctx.convTilingData->convApiTiling.khL1 * self_->ctx.convTilingData->convApiTiling.kwL1 * Intf::k0;274+ self_->ctx.convTilingData->khL1 * self_->ctx.convTilingData->kwL1 * Intf::k0;
275 khStrideBase = 275 khStrideBase =
276- self_->ctx.convTilingData->convApiTiling.khUb * self_->ctx.convTilingData->convApiTiling.kwL1 * Intf::k0;276+ self_->ctx.convTilingData->khUb * self_->ctx.convTilingData->kwL1 * Intf::k0;
277- kwStrideBase = self_->ctx.convTilingData->convApiTiling.kwUb * Intf::k0;277+ kwStrideBase = self_->ctx.convTilingData->kwUb * Intf::k0;
278 }278 }
279 279 
280 __aicore__ inline void LoadUB2L1()280 __aicore__ inline void LoadUB2L1()
281 {281 {
282 if (unlikely(self_->ctx.isFirstIterate)) {282 if (unlikely(self_->ctx.isFirstIterate)) {
283- copyParams.blockCount = self_->ctx.convTilingData->convApiTiling.khUb;283+ copyParams.blockCount = self_->ctx.convTilingData->khUb;
284 copyParams.srcStride = 0;284 copyParams.srcStride = 0;
285 }285 }
286- copyParams.blockLen = self_->ctx.currentHoL1xWoL1Align * self_->ctx.convTilingData->convApiTiling.kwUb;286+ copyParams.blockLen = self_->ctx.currentHoL1xWoL1Align * self_->ctx.convTilingData->kwUb;
287 copyParams.dstStride = self_->ctx.currentHoL1xWoL1Align * dstStrideBase;287 copyParams.dstStride = self_->ctx.currentHoL1xWoL1Align * dstStrideBase;
288 288 
289 uint64_t dstOffset = self_->ctx.vecCi1Iter * self_->ctx.currentHoL1xWoL1Align * ci1StrideBase +289 uint64_t dstOffset = self_->ctx.vecCi1Iter * self_->ctx.currentHoL1xWoL1Align * ci1StrideBase +
Mconv/conv2d_v2/op_kernel/arch35/conv2d_v2_instr_fz_impl.h+11-11
@@ -33,14 +33,14 @@ public:
33 self_ = self;33 self_ = self;
34 maxNBL1Iter_ = self->ctx.maxNBL1Iter;34 maxNBL1Iter_ = self->ctx.maxNBL1Iter;
35 maxKBL1Iter_ = self->ctx.maxKBL1Iter;35 maxKBL1Iter_ = self->ctx.maxKBL1Iter;
36- orgCoAlignN0 = AlignB(self_->ctx.convTilingData->convApiTiling.orgCo, BLOCK_L0_N);36+ orgCoAlignN0 = AlignB(self_->ctx.convTilingData->orgCo, BLOCK_L0_N);
37 if constexpr (Intf::groupOptNZFlag) {37 if constexpr (Intf::groupOptNZFlag) {
38 self->ctx.coPerGroup =38 self->ctx.coPerGroup =
39- self->ctx.convTilingData->convApiTiling.orgCo / self->ctx.convTilingData->convApiTiling.groups;39+ self->ctx.convTilingData->orgCo / self->ctx.convTilingData->groups;
40- self->ctx.coOpt = self->ctx.coPerGroup * self->ctx.convTilingData->convApiTiling.enlarge;40+ self->ctx.coOpt = self->ctx.coPerGroup * self->ctx.convTilingData->enlarge;
41 orgCoAlignN0 = AlignB(self->ctx.coOpt, BLOCK_L0_N);41 orgCoAlignN0 = AlignB(self->ctx.coOpt, BLOCK_L0_N);
42 } else {42 } else {
43- orgCoAlignN0 = AlignB(self_->ctx.convTilingData->convApiTiling.orgCo, BLOCK_L0_N);43+ orgCoAlignN0 = AlignB(self_->ctx.convTilingData->orgCo, BLOCK_L0_N);
44 }44 }
45 }45 }
46 46 
@@ -53,18 +53,18 @@ public:
53 {53 {
54 uint64_t bL1GmOffset;54 uint64_t bL1GmOffset;
55 if constexpr (!Intf::hasNL1IterFlag) {55 if constexpr (!Intf::hasNL1IterFlag) {
56- bL1GmOffset = kBL1Iter * self_->ctx.convTilingData->convApiTiling.kBL1 * orgCoAlignN0;56+ bL1GmOffset = kBL1Iter * self_->ctx.convTilingData->kBL1 * orgCoAlignN0;
57 } else {57 } else {
58- bL1GmOffset = kBL1Iter * self_->ctx.convTilingData->convApiTiling.kBL1 * orgCoAlignN0 +58+ bL1GmOffset = kBL1Iter * self_->ctx.convTilingData->kBL1 * orgCoAlignN0 +
59- nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 * Intf::k0;59+ nBL1Iter * self_->ctx.convTilingData->nBL1 * Intf::k0;
60 self_->ctx.currentNBL1 = nBL1Iter == maxNBL1Iter_ ?60 self_->ctx.currentNBL1 = nBL1Iter == maxNBL1Iter_ ?
61- self_->ctx.nBL1Tail : self_->ctx.convTilingData->convApiTiling.nBL1;61+ self_->ctx.nBL1Tail : self_->ctx.convTilingData->nBL1;
62 }62 }
63 63 
64- uint64_t currentKBL1 = self_->ctx.convTilingData->convApiTiling.kBL1;64+ uint64_t currentKBL1 = self_->ctx.convTilingData->kBL1;
65 if constexpr (!Intf::c04Flag) {65 if constexpr (!Intf::c04Flag) {
66 currentKBL1 = kBL1Iter == maxKBL1Iter_ ?66 currentKBL1 = kBL1Iter == maxKBL1Iter_ ?
67- self_->ctx.kBL1AlignK0Tail : self_->ctx.convTilingData->convApiTiling.kBL1;67+ self_->ctx.kBL1AlignK0Tail : self_->ctx.convTilingData->kBL1;
68 }68 }
69 DataCopyPadExtParams<typename Intf::WeightT> padParams;69 DataCopyPadExtParams<typename Intf::WeightT> padParams;
70 DataCopyExtParams dataCopyParams;70 DataCopyExtParams dataCopyParams;
@@ -76,7 +76,7 @@ public:
76 dataCopyParams.blockCount = CeilDiv(currentKBL1, Intf::k0);76 dataCopyParams.blockCount = CeilDiv(currentKBL1, Intf::k0);
77 dataCopyParams.blockLen = self_->ctx.currentNBL1 * C0_SIZE;77 dataCopyParams.blockLen = self_->ctx.currentNBL1 * C0_SIZE;
78 dataCopyParams.srcStride = (orgCoAlignN0 - self_->ctx.currentNBL1) * C0_SIZE;78 dataCopyParams.srcStride = (orgCoAlignN0 - self_->ctx.currentNBL1) * C0_SIZE;
79- dataCopyParams.dstStride = self_->ctx.convTilingData->convApiTiling.nBL1 - self_->ctx.currentNBL1;79+ dataCopyParams.dstStride = self_->ctx.convTilingData->nBL1 - self_->ctx.currentNBL1;
80 }80 }
81 DataCopyPad<typename Intf::WeightT>(self_->ctx.bl1, self_->ctx.bgm[bL1GmOffset], dataCopyParams, padParams);81 DataCopyPad<typename Intf::WeightT>(self_->ctx.bl1, self_->ctx.bgm[bL1GmOffset], dataCopyParams, padParams);
82 }82 }