Star1324
2076
代码
代码
Issues746
Pull Requests625
Actions
项目讨论
Wiki
分析
Star1324
2076
已合并
【性能优化】conv正向卷积算子性能优化 #3793
ray-shaw创建于 4月14日
【性能优化】conv正向卷积算子性能优化 #3793
已合并
ray-shaw创建于 4月14日
从ray-shaw/ops-nnmaster合入到cann/ops-nnmaster
共 88 个文件变更+4424-3574
Mconv/common/op_host/op_tiling/arch35/conv_api_tiling_base.h+23-0
@@ -159,6 +159,26 @@ struct ShapeInfo {
159 uint8_t dualOutput = 0;159 uint8_t dualOutput = 0;
160};160};
161 161 
162+#pragma pack(push, 1) // 确保无填充，严格按位排列
163+typedef struct {
164+ uint64_t strideW : 6; // bits 0-5
165+ uint64_t strideH : 6; // bits 6-11
166+ uint64_t kernelW : 8; // bits 12-19
167+ uint64_t kernelH : 8; // bits 20-27
168+ uint64_t dilationW : 8; // bits 28-35
169+ uint64_t dilationH : 8; // bits 36-43
170+ uint64_t kernelW_highest_bit : 1; // bit 44
171+ uint64_t kernelH_highest_bit : 1; // bit 45
172+ uint64_t reserve1 : 2; // bits 46-47
173+ uint64_t channelSize : 16; // bits 48-63
174+} BitFieldData;
175+#pragma pack(pop)
176+
177+union UnionDataXt {
178+ uint64_t n;
179+ BitFieldData bf; // n与bf共享同一块内存空间
180+};
181+ 
162struct AttrInfo {182struct AttrInfo {
163 int32_t padHead = 0;183 int32_t padHead = 0;
164 int32_t padTail = 0;184 int32_t padTail = 0;
@@ -178,6 +198,9 @@ struct AttrInfo {
178 int32_t groups = 1;198 int32_t groups = 1;
179 int8_t offsetx = 0;199 int8_t offsetx = 0;
180 int8_t roundMode = 0;200 int8_t roundMode = 0;
201+ 
202+ UnionDataXt unionDataXt;
203+ AttrInfo() : unionDataXt{} {}
181};204};
182 205 
183struct DescInfo {206struct DescInfo {
Mconv/common/op_host/op_tiling/arch35/conv_base_utils.h+10-2
@@ -132,7 +132,11 @@ const std::vector<std::vector<ge::Format>> SUPPORT_CONV2D_FORMAT_LIST = {
132const std::vector<std::vector<ge::Format>> SUPPORT_CONV2D_FORMAT_LIST_MDC = {132const std::vector<std::vector<ge::Format>> SUPPORT_CONV2D_FORMAT_LIST_MDC = {
133 {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW},133 {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW},
134 {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NCHW},134 {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NCHW},
135- {ge::Format::FORMAT_NHWC, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NHWC}135+ {ge::Format::FORMAT_NHWC, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NHWC},
136+ {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NHWC},
137+ {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NHWC},
138+ {ge::Format::FORMAT_NHWC, ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW},
139+ {ge::Format::FORMAT_NHWC, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NCHW}
136};140};
137 141 
138const std::vector<std::vector<ge::Format>> SUPPORT_QUANT_CONV2D_FORMAT_LIST = {142const std::vector<std::vector<ge::Format>> SUPPORT_QUANT_CONV2D_FORMAT_LIST = {
@@ -167,7 +171,11 @@ const std::vector<std::vector<ge::Format>> EXTENDCONV2D_SUPPORT_FORMAT_LIST = {
167const std::vector<std::vector<ge::Format>> EXTENDCONV2D_SUPPORT_FORMAT_LIST_MDC = {171const std::vector<std::vector<ge::Format>> EXTENDCONV2D_SUPPORT_FORMAT_LIST_MDC = {
168 {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW},172 {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW},
169 {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NCHW},173 {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NCHW},
170- {ge::Format::FORMAT_NHWC, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NHWC}174+ {ge::Format::FORMAT_NHWC, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NHWC},
175+ {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NHWC},
176+ {ge::Format::FORMAT_NCHW, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NHWC},
177+ {ge::Format::FORMAT_NHWC, ge::Format::FORMAT_NCHW, ge::Format::FORMAT_NCHW},
178+ {ge::Format::FORMAT_NHWC, ge::Format::FORMAT_HWCN, ge::Format::FORMAT_NCHW}
171};179};
172 180 
173struct ConvParamInfo {181struct ConvParamInfo {
Mconv/common/op_host/op_tiling/arch35/conv_cache_tiling.h+2-0
@@ -134,6 +134,7 @@ struct Conv2dCacheTilingData {
134 uint32_t singleCoreCi = 0;134 uint32_t singleCoreCi = 0;
135 uint32_t singleCoreCo = 0;135 uint32_t singleCoreCo = 0;
136 uint64_t orgHixWi = 0;136 uint64_t orgHixWi = 0;
137+ uint64_t unionDataXt = 0;
137 uint32_t hoL1 = 0;138 uint32_t hoL1 = 0;
138 uint32_t woL1 = 0;139 uint32_t woL1 = 0;
139 uint32_t kAL1 = 0;140 uint32_t kAL1 = 0;
@@ -207,6 +208,7 @@ struct Conv3dCacheTilingData {
207 uint64_t oriWixk0 = 0;208 uint64_t oriWixk0 = 0;
208 uint64_t orgHixWi = 0;209 uint64_t orgHixWi = 0;
209 uint64_t orgHoxWo = 0;210 uint64_t orgHoxWo = 0;
211+ uint64_t unionDataXt = 0;
210 212 
211 // uint32_t fields213 // uint32_t fields
212 uint32_t orgCi = 0;214 uint32_t orgCi = 0;
Mconv/common/op_host/op_tiling/arch35/conv_template_utils.h+8-0
@@ -222,8 +222,13 @@ constexpr uint64_t CONV_INNER_BATCH_SINGLE = 0;
222constexpr uint64_t CONV_INNER_BATCH_KERNEL_1X1_MULTI = 1;222constexpr uint64_t CONV_INNER_BATCH_KERNEL_1X1_MULTI = 1;
223constexpr uint64_t CONV_INNER_BATCH_MULTI = 2;223constexpr uint64_t CONV_INNER_BATCH_MULTI = 2;
224 224 
225+constexpr uint64_t CONV_NOT_SMALL_WEIGHT = 0;
226+constexpr uint64_t CONV_FULLLOAD_KL1_NL0 = 1;
227+constexpr uint64_t CONV_WEIGHT_SMALLER_THAN_BL0 = 2;
228+ 
225constexpr uint32_t FP16_DTYPE_SIZE = 2;229constexpr uint32_t FP16_DTYPE_SIZE = 2;
226constexpr uint32_t INT64_DTYPE_SIZE_COMPARE_FP16 = 4;230constexpr uint32_t INT64_DTYPE_SIZE_COMPARE_FP16 = 4;
231+constexpr uint32_t FLOAT_DTYPE_SIZE_COMPARE_FP16 = 2;
227 232 
228// idxList233// idxList
229constexpr size_t IDX_LIST_N_IDX = 0;234constexpr size_t IDX_LIST_N_IDX = 0;
@@ -425,6 +430,9 @@ struct ConvTilingKeyPara {
425 uint64_t fmapCppyMode = 0;430 uint64_t fmapCppyMode = 0;
426 uint64_t innerBatch = 0;431 uint64_t innerBatch = 0;
427 uint64_t disContinuous = 0;432 uint64_t disContinuous = 0;
433+ uint64_t batchOne = 0;
434+ uint64_t noPad = 0;
435+ uint64_t smallWeight = 0;
428};436};
429 437 
430static std::map<ge::DataType, uint32_t> dtypeSizeTab = {438static std::map<ge::DataType, uint32_t> dtypeSizeTab = {
Mconv/common/op_kernel/arch35/conv_common.h+100-89
@@ -53,10 +53,10 @@ struct ExtendParams {
53 GM_ADDR y1 = nullptr;53 GM_ADDR y1 = nullptr;
54};54};
55 55 
56-template <class CONV, class RUN_INFO>56+template <class CONV, class CONV_TILING>
57class ConvCommon {57class ConvCommon {
58public:58public:
59- __aicore__ __forceinline__ void Init(CONV *convPtr, const RUN_INFO* tilingPtr, bool hasScaleFlag);59+ __aicore__ __forceinline__ void Init(CONV *convPtr, const CONV_TILING* tilingPtr, bool hasScaleFlag);
60 60 
61 // Calc singleCore dim data61 // Calc singleCore dim data
62 __aicore__ __forceinline__ bool CalcDimData(const uint32_t& blockPerDim, const uint32_t& dim,62 __aicore__ __forceinline__ bool CalcDimData(const uint32_t& blockPerDim, const uint32_t& dim,
@@ -100,21 +100,22 @@ public:
100 100 
101protected:101protected:
102 CONV *convOps;102 CONV *convOps;
103- const RUN_INFO *convRunInfo;103+ const CONV_TILING *convTilingData;
104 bool hasScale = true;104 bool hasScale = true;
105 uint64_t fmStartAddr = 0;105 uint64_t fmStartAddr = 0;
106 uint64_t weightStartAddr = 0;106 uint64_t weightStartAddr = 0;
107 uint64_t biasStartAddr = 0;107 uint64_t biasStartAddr = 0;
108 uint64_t scaleStartAddr = 0;108 uint64_t scaleStartAddr = 0;
109+ uint64_t reluWeightStartAddr = 0;
109 uint64_t outputStartAddr = 0;110 uint64_t outputStartAddr = 0;
110 111 
111 uint64_t hwIn = 0;112 uint64_t hwIn = 0;
112 uint64_t hwOut = 0;113 uint64_t hwOut = 0;
113};114};
114 115 
115-template <class CONV, class RUN_INFO>116+template <class CONV, class CONV_TILING>
116-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::117+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
117- Init(CONV *convPtr, const RUN_INFO* tilingPtr, bool hasScaleFlag)118+ Init(CONV *convPtr, const CONV_TILING* tilingPtr, bool hasScaleFlag)
118{119{
119 if ASCEND_IS_AIV {120 if ASCEND_IS_AIV {
120 blockIdx = GetBlockIdx();121 blockIdx = GetBlockIdx();
@@ -124,12 +125,12 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
124 blockIdx = GetBlockIdx();125 blockIdx = GetBlockIdx();
125 }126 }
126 convOps = convPtr;127 convOps = convPtr;
127- convRunInfo = tilingPtr;128+ convTilingData = tilingPtr;
128 hasScale = hasScaleFlag;129 hasScale = hasScaleFlag;
129}130}
130 131 
131-template <class CONV, class RUN_INFO>132+template <class CONV, class CONV_TILING>
132-__aicore__ __forceinline__ bool ConvCommon<CONV, RUN_INFO>::133+__aicore__ __forceinline__ bool ConvCommon<CONV, CONV_TILING>::
133 CalcDimData(const uint32_t& blockPerDim, const uint32_t& dim, const uint64_t& wholeDim,134 CalcDimData(const uint32_t& blockPerDim, const uint32_t& dim, const uint64_t& wholeDim,
134 const uint64_t &realWholeDim, DimDataToFill& curStruct)135 const uint64_t &realWholeDim, DimDataToFill& curStruct)
135{136{
@@ -147,8 +148,8 @@ __aicore__ __forceinline__ bool ConvCommon<CONV, RUN_INFO>::
147 return true;148 return true;
148}149}
149 150 
150-template <class CONV, class RUN_INFO>151+template <class CONV, class CONV_TILING>
151-__aicore__ __forceinline__ bool ConvCommon<CONV, RUN_INFO>::CalcNDimDataAlign(const uint64_t& blockPerDim,152+__aicore__ __forceinline__ bool ConvCommon<CONV, CONV_TILING>::CalcNDimDataAlign(const uint64_t& blockPerDim,
152 const uint64_t& dim,153 const uint64_t& dim,
153 const uint64_t& wholeDim,154 const uint64_t& wholeDim,
154 DimDataToFill& curStruct)155 DimDataToFill& curStruct)
@@ -167,17 +168,17 @@ __aicore__ __forceinline__ bool ConvCommon<CONV, RUN_INFO>::CalcNDimDataAlign(co
167 return true;168 return true;
168}169}
169 170 
170-template <class CONV, class RUN_INFO>171+template <class CONV, class CONV_TILING>
171-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::172+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
172 CalcStartAddrCommon(const uint32_t din, const uint32_t dout)173 CalcStartAddrCommon(const uint32_t din, const uint32_t dout)
173{174{
174- hwIn = convRunInfo->hin * convRunInfo->win;175+ hwIn = convTilingData->convRunInfo.hin * convTilingData->convRunInfo.win;
175- hwOut = convRunInfo->hout * convRunInfo->wout;176+ hwOut = convTilingData->convRunInfo.hout * convTilingData->convRunInfo.wout;
176 177 
177- convOps->fmapOneBatchSize = convRunInfo->cin * din * hwIn;178+ convOps->fmapOneBatchSize = convTilingData->convRunInfo.cin * din * hwIn;
178- convOps->outputOneBatchSize = convRunInfo->cout * dout * hwOut;179+ convOps->outputOneBatchSize = convTilingData->convRunInfo.cout * dout * hwOut;
179 180 
180- if (convRunInfo->hasBias) {181+ if (convTilingData->convRunInfo.hasBias) {
181 biasStartAddr = convOps->nIdxStart;182 biasStartAddr = convOps->nIdxStart;
182 }183 }
183 184 
@@ -188,22 +189,22 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
188 }189 }
189}190}
190 191 
191-template <class CONV, class RUN_INFO>192+template <class CONV, class CONV_TILING>
192-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::193+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
193 CalcStartAddrMMode(const uint32_t din, const uint32_t dout, const uint32_t kd, const uint64_t doIdxStart,194 CalcStartAddrMMode(const uint32_t din, const uint32_t dout, const uint32_t kd, const uint64_t doIdxStart,
194 const int64_t diIdxStart)195 const int64_t diIdxStart)
195{196{
196 CalcStartAddrCommon(din, dout);197 CalcStartAddrCommon(din, dout);
197 198 
198 if constexpr (CONV::DIS_CONTINUOUS) {199 if constexpr (CONV::DIS_CONTINUOUS) {
199- fmStartAddr = convOps->batchIdxStart * convRunInfo->cin;200+ fmStartAddr = convOps->batchIdxStart * convTilingData->convRunInfo.cin;
200 } else {201 } else {
201 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize;202 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize;
202 }203 }
203 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {204 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {
204 weightStartAddr = convOps->nIdxStart * convOps->k0;205 weightStartAddr = convOps->nIdxStart * convOps->k0;
205 } else {206 } else {
206- weightStartAddr = convOps->nIdxStart * convRunInfo->cin * kd * convRunInfo->kh * convRunInfo->kw;207+ weightStartAddr = convOps->nIdxStart * convTilingData->convRunInfo.cin * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;
207 }208 }
208 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +209 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
209 convOps->nIdxStart * dout * hwOut +210 convOps->nIdxStart * dout * hwOut +
@@ -215,37 +216,42 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
215 }216 }
216 217 
217 if constexpr (CONV::IS_EXTEND_CONV2D) {218 if constexpr (CONV::IS_EXTEND_CONV2D) {
218- if (convOps->conv2dApiTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||219+ if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
219- convOps->conv2dApiTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {220+ convOps->convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
220 scaleStartAddr = convOps->nIdxStart;221 scaleStartAddr = convOps->nIdxStart;
221 }222 }
223+ if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
224+ convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
225+ reluWeightStartAddr = convOps->nIdxStart;
226+ }
222 }227 }
223}228}
224 229 
225-template <class CONV, class RUN_INFO>230+template <class CONV, class CONV_TILING>
226-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::231+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
227 CalcStartAddrHWode(const uint32_t din, const uint32_t dout, const uint32_t kd, const uint64_t doIdxStart,232 CalcStartAddrHWode(const uint32_t din, const uint32_t dout, const uint32_t kd, const uint64_t doIdxStart,
228 const int64_t diIdxStart)233 const int64_t diIdxStart)
229{234{
230 CalcStartAddrCommon(din, dout);235 CalcStartAddrCommon(din, dout);
231 236 
232- int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convRunInfo->strideH) -237+ int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->convRunInfo.strideH) -
233- static_cast<int64_t>(convRunInfo->padTop);238+ static_cast<int64_t>(convTilingData->convRunInfo.padTop);
234 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;239 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;
235 if constexpr (CONV::DIS_CONTINUOUS) {240 if constexpr (CONV::DIS_CONTINUOUS) {
236- fmStartAddr = convOps->singleCoreHiStartPos * convRunInfo->win * convRunInfo->batch * convRunInfo->cin +241+ fmStartAddr = convOps->singleCoreHiStartPos * convTilingData->convRunInfo.win *
237- convOps->batchIdxStart * convRunInfo->cin;242+ convTilingData->convRunInfo.batch * convTilingData->convRunInfo.cin +
243+ convOps->batchIdxStart * convTilingData->convRunInfo.cin;
238 } else {244 } else {
239 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *245 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *
240- convRunInfo->win;246+ convTilingData->convRunInfo.win;
241 }247 }
242 convOps->singleCoreHiStartPos = hiStartPosTmp;248 convOps->singleCoreHiStartPos = hiStartPosTmp;
243 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {249 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {
244- int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convRunInfo->strideW) -250+ int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->convRunInfo.strideW) -
245- static_cast<int64_t>(convRunInfo->padLeft);251+ static_cast<int64_t>(convTilingData->convRunInfo.padLeft);
246 convOps->singleCoreWiStartPos = wiStartPosTmp < 0 ? 0 : wiStartPosTmp;252 convOps->singleCoreWiStartPos = wiStartPosTmp < 0 ? 0 : wiStartPosTmp;
247 if constexpr (CONV::DIS_CONTINUOUS) {253 if constexpr (CONV::DIS_CONTINUOUS) {
248- fmStartAddr += convOps->singleCoreWiStartPos * convRunInfo->batch * convRunInfo->cin;254+ fmStartAddr += convOps->singleCoreWiStartPos * convTilingData->convRunInfo.batch * convTilingData->convRunInfo.cin;
249 } else {255 } else {
250 fmStartAddr += convOps->singleCoreWiStartPos;256 fmStartAddr += convOps->singleCoreWiStartPos;
251 }257 }
@@ -255,12 +261,12 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
255 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {261 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z || CONV::B_FORMAT == ConvFormat::FRACTAL_Z_C04) {
256 weightStartAddr = convOps->nIdxStart * convOps->k0;262 weightStartAddr = convOps->nIdxStart * convOps->k0;
257 } else {263 } else {
258- weightStartAddr = convOps->nIdxStart * convRunInfo->cin * kd * convRunInfo->kh * convRunInfo->kw;264+ weightStartAddr = convOps->nIdxStart * convTilingData->convRunInfo.cin * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;
259 }265 }
260 266 
261 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +267 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
262 convOps->nIdxStart * dout * hwOut +268 convOps->nIdxStart * dout * hwOut +
263- convOps->hoIdxStart * convRunInfo->wout;269+ convOps->hoIdxStart * convTilingData->convRunInfo.wout;
264 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {270 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {
265 outputStartAddr += convOps->woIdxStart;271 outputStartAddr += convOps->woIdxStart;
266 }272 }
@@ -271,8 +277,8 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
271 }277 }
272}278}
273 279 
274-template <class CONV, class RUN_INFO>280+template <class CONV, class CONV_TILING>
275-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::281+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
276 CalcStartAddrMModeHWC(const uint32_t din, const uint32_t dout, const uint64_t doIdxStart, const int64_t diIdxStart)282 CalcStartAddrMModeHWC(const uint32_t din, const uint32_t dout, const uint64_t doIdxStart, const int64_t diIdxStart)
277{283{
278 CalcStartAddrCommon(din, dout);284 CalcStartAddrCommon(din, dout);
@@ -280,50 +286,50 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
280 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize;286 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize;
281 weightStartAddr = convOps->nIdxStart;287 weightStartAddr = convOps->nIdxStart;
282 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +288 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
283- convOps->mIdxStart * convRunInfo->cout +289+ convOps->mIdxStart * convTilingData->convRunInfo.cout +
284 convOps->nIdxStart;290 convOps->nIdxStart;
285 291 
286 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {292 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
287- fmStartAddr += diIdxStart * hwIn * convRunInfo->cin;293+ fmStartAddr += diIdxStart * hwIn * convTilingData->convRunInfo.cin;
288- outputStartAddr += doIdxStart * hwOut * convRunInfo->cout;294+ outputStartAddr += doIdxStart * hwOut * convTilingData->convRunInfo.cout;
289 }295 }
290}296}
291 297 
292-template <class CONV, class RUN_INFO>298+template <class CONV, class CONV_TILING>
293-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::299+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
294 CalcStartAddrHWodeHWC(const uint32_t din, const uint32_t dout, const uint64_t doIdxStart, const int64_t diIdxStart)300 CalcStartAddrHWodeHWC(const uint32_t din, const uint32_t dout, const uint64_t doIdxStart, const int64_t diIdxStart)
295{301{
296 CalcStartAddrCommon(din, dout);302 CalcStartAddrCommon(din, dout);
297 303 
298- int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convRunInfo->strideH) -304+ int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->convRunInfo.strideH) -
299- static_cast<int64_t>(convRunInfo->padTop);305+ static_cast<int64_t>(convTilingData->convRunInfo.padTop);
300 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;306 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;
301 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *307 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *
302- convRunInfo->win * convRunInfo->cin;308+ convTilingData->convRunInfo.win * convTilingData->convRunInfo.cin;
303 convOps->singleCoreHiStartPos = hiStartPosTmp;309 convOps->singleCoreHiStartPos = hiStartPosTmp;
304 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {310 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {
305- int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convRunInfo->strideW) -311+ int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->convRunInfo.strideW) -
306- static_cast<int64_t>(convRunInfo->padLeft);312+ static_cast<int64_t>(convTilingData->convRunInfo.padLeft);
307 convOps->singleCoreWiStartPos = wiStartPosTmp;313 convOps->singleCoreWiStartPos = wiStartPosTmp;
308- fmStartAddr += Max(convOps->singleCoreWiStartPos, 0) * convRunInfo->cin;314+ fmStartAddr += Max(convOps->singleCoreWiStartPos, 0) * convTilingData->convRunInfo.cin;
309 }315 }
310 316 
311 weightStartAddr = convOps->nIdxStart;317 weightStartAddr = convOps->nIdxStart;
312 318 
313 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize + convOps->hoIdxStart *319 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize + convOps->hoIdxStart *
314- convRunInfo->wout * convRunInfo->cout + convOps->nIdxStart;320+ convTilingData->convRunInfo.wout * convTilingData->convRunInfo.cout + convOps->nIdxStart;
315 321 
316 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {322 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {
317- outputStartAddr += convOps->woIdxStart * convRunInfo->cout;323+ outputStartAddr += convOps->woIdxStart * convTilingData->convRunInfo.cout;
318 }324 }
319 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {325 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
320- fmStartAddr += diIdxStart * hwIn * convRunInfo->cin;326+ fmStartAddr += diIdxStart * hwIn * convTilingData->convRunInfo.cin;
321- outputStartAddr += doIdxStart * hwOut * convRunInfo->cout;327+ outputStartAddr += doIdxStart * hwOut * convTilingData->convRunInfo.cout;
322 }328 }
323}329}
324 330 
325-template <class CONV, class RUN_INFO>331+template <class CONV, class CONV_TILING>
326-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::332+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
327 CalcStartAddrDeQuant(const uint32_t din, const uint32_t dout, const uint32_t kd,333 CalcStartAddrDeQuant(const uint32_t din, const uint32_t dout, const uint32_t kd,
328 const uint64_t doIdxStart, const int64_t diIdxStart)334 const uint64_t doIdxStart, const int64_t diIdxStart)
329{335{
@@ -333,60 +339,65 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
333 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + diIdxStart * hwIn;339 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + diIdxStart * hwIn;
334 340 
335 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +341 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
336- convOps->mIdxStart * convRunInfo->cout + convOps->nIdxStart +342+ convOps->mIdxStart * convTilingData->convRunInfo.cout + convOps->nIdxStart +
337- doIdxStart * hwOut * convRunInfo->cout;343+ doIdxStart * hwOut * convTilingData->convRunInfo.cout;
338 } else {344 } else {
339- int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convRunInfo->strideH) -345+ int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->convRunInfo.strideH) -
340- static_cast<int64_t>(convRunInfo->padTop);346+ static_cast<int64_t>(convTilingData->convRunInfo.padTop);
341 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;347 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;
342 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *348 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize + convOps->singleCoreHiStartPos *
343- convRunInfo->win + diIdxStart * hwIn;349+ convTilingData->convRunInfo.win + diIdxStart * hwIn;
344 convOps->singleCoreHiStartPos = hiStartPosTmp;350 convOps->singleCoreHiStartPos = hiStartPosTmp;
345 351 
346 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize + convOps->hoIdxStart *352 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize + convOps->hoIdxStart *
347- convRunInfo->wout * convRunInfo->cout + convOps->nIdxStart +353+ convTilingData->convRunInfo.wout * convTilingData->convRunInfo.cout + convOps->nIdxStart +
348- doIdxStart * hwOut * convRunInfo->cout;354+ doIdxStart * hwOut * convTilingData->convRunInfo.cout;
349 }355 }
350- weightStartAddr = convOps->nIdxStart * convRunInfo->cin * kd * convRunInfo->kh * convRunInfo->kw;356+ weightStartAddr = convOps->nIdxStart * convTilingData->convRunInfo.cin * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;
351}357}
352 358 
353-template <class CONV, class RUN_INFO>359+template <class CONV, class CONV_TILING>
354-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::360+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
355 InitFixpipeBuffer(const ExtendParams* extendParams)361 InitFixpipeBuffer(const ExtendParams* extendParams)
356{362{
357 if constexpr (CONV::isQuant || CONV::IS_EXTEND_CONV2D) {363 if constexpr (CONV::isQuant || CONV::IS_EXTEND_CONV2D) {
358 if (hasScale) {364 if (hasScale) {
359 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {365 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {
360- if (convOps->conv2dApiTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {366+ if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
361 convOps->fixpipeParams.scale0.SetGlobalBuffer(367 convOps->fixpipeParams.scale0.SetGlobalBuffer(
362 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale0 + scaleStartAddr *368 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale0 + scaleStartAddr *
363 sizeof(typename CONV::SCALE_T)));369 sizeof(typename CONV::SCALE_T)));
364- } else if (convOps->conv2dApiTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {370+ } else if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {
365 convOps->fixpipeParams.scale0.SetGlobalBuffer(371 convOps->fixpipeParams.scale0.SetGlobalBuffer(
366 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale0));372 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale0));
367 }373 }
368-#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
369- if (convOps->conv2dApiTiling->reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
370- convOps->fixpipeParams.reluWeight0.SetGlobalBuffer(
371- reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight0));
372- }
373-#endif
374 if constexpr (CONV::IS_EXTEND_CONV2D) {374 if constexpr (CONV::IS_EXTEND_CONV2D) {
375- if (convOps->dualOutput) {375+ if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
376- if (convOps->conv2dApiTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {376+ convOps->fixpipeParams.reluWeight0.SetGlobalBuffer(
377+ reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight0));
378+ } else if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
379+ // copy relu_weight from om to gm
380+ convOps->fixpipeParams.reluWeight0.SetGlobalBuffer(
381+ reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight0 +
382+ reluWeightStartAddr * sizeof(typename CONV::RELU_WEIGHT_T)));
383+ }
384+ if (convTilingData->convApiTiling.dualOutput) {
385+ if (convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
377 convOps->fixpipeParams.scale1.SetGlobalBuffer(386 convOps->fixpipeParams.scale1.SetGlobalBuffer(
378 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale1 + scaleStartAddr *387 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale1 + scaleStartAddr *
379 sizeof(typename CONV::SCALE_T)));388 sizeof(typename CONV::SCALE_T)));
380- } else if (convOps->conv2dApiTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {389+ } else if (convOps->convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {
381 convOps->fixpipeParams.scale1.SetGlobalBuffer(390 convOps->fixpipeParams.scale1.SetGlobalBuffer(
382 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale1));391 reinterpret_cast<__gm__ typename CONV::SCALE_T*>(extendParams->scale1));
383 }392 }
384-#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)393+ if (convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
385- if (convOps->conv2dApiTiling->reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
386 convOps->fixpipeParams.reluWeight1.SetGlobalBuffer(394 convOps->fixpipeParams.reluWeight1.SetGlobalBuffer(
387 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight1));395 reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight1));
396+ } else if (convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
397+ convOps->fixpipeParams.reluWeight1.SetGlobalBuffer(
398+ reinterpret_cast<__gm__ typename CONV::RELU_WEIGHT_T*>(extendParams->reluWeight1 +
399+ reluWeightStartAddr * sizeof(typename CONV::RELU_WEIGHT_T)));
388 }400 }
389-#endif
390 }401 }
391 }402 }
392 } else {403 } else {
@@ -398,8 +409,8 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
398 }409 }
399}410}
400 411 
401-template <class CONV, class RUN_INFO>412+template <class CONV, class CONV_TILING>
402-__aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::413+__aicore__ __forceinline__ void ConvCommon<CONV, CONV_TILING>::
403 InitBufferCommon(GM_ADDR x, GM_ADDR filter, GM_ADDR bias, GM_ADDR y, const ExtendParams* extendParams)414 InitBufferCommon(GM_ADDR x, GM_ADDR filter, GM_ADDR bias, GM_ADDR y, const ExtendParams* extendParams)
404{415{
405 convOps->fmapGm.SetGlobalBuffer(416 convOps->fmapGm.SetGlobalBuffer(
@@ -409,13 +420,13 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
409 convOps->outputGm.SetGlobalBuffer(420 convOps->outputGm.SetGlobalBuffer(
410 reinterpret_cast<__gm__ typename CONV::OUTPUT_T*>(y + outputStartAddr * sizeof(typename CONV::OUTPUT_T)));421 reinterpret_cast<__gm__ typename CONV::OUTPUT_T*>(y + outputStartAddr * sizeof(typename CONV::OUTPUT_T)));
411 if constexpr (CONV::IS_EXTEND_CONV2D) {422 if constexpr (CONV::IS_EXTEND_CONV2D) {
412- if (convOps->dualOutput) {423+ if (convTilingData->convApiTiling.dualOutput) {
413 convOps->output1Gm.SetGlobalBuffer(424 convOps->output1Gm.SetGlobalBuffer(
414 reinterpret_cast<__gm__ typename CONV::OUTPUT1_T*>(425 reinterpret_cast<__gm__ typename CONV::OUTPUT1_T*>(
415 extendParams->y1 + outputStartAddr * sizeof(typename CONV::OUTPUT1_T)));426 extendParams->y1 + outputStartAddr * sizeof(typename CONV::OUTPUT1_T)));
416 }427 }
417 }428 }
418- if (convRunInfo->hasBias) {429+ if (convTilingData->convRunInfo.hasBias) {
419 convOps->biasGm.SetGlobalBuffer(430 convOps->biasGm.SetGlobalBuffer(
420 reinterpret_cast<__gm__ typename CONV::BIAS_T*>(bias + biasStartAddr * sizeof(typename CONV::BIAS_T)));431 reinterpret_cast<__gm__ typename CONV::BIAS_T*>(bias + biasStartAddr * sizeof(typename CONV::BIAS_T)));
421 }432 }
@@ -423,22 +434,22 @@ __aicore__ __forceinline__ void ConvCommon<CONV, RUN_INFO>::
423 InitFixpipeBuffer(extendParams);434 InitFixpipeBuffer(extendParams);
424}435}
425 436 
426-template <class CONV, class RUN_INFO>437+template <class CONV, class CONV_TILING>
427-__aicore__ __forceinline__ uint64_t ConvCommon<CONV, RUN_INFO>::438+__aicore__ __forceinline__ uint64_t ConvCommon<CONV, CONV_TILING>::
428 AlignB(const uint64_t numA, const uint64_t numB)439 AlignB(const uint64_t numA, const uint64_t numB)
429{440{
430 return ((numA + numB - 1) / numB) * numB;441 return ((numA + numB - 1) / numB) * numB;
431}442}
432 443 
433-template <class CONV, class RUN_INFO>444+template <class CONV, class CONV_TILING>
434-__aicore__ __forceinline__ uint64_t ConvCommon<CONV, RUN_INFO>::445+__aicore__ __forceinline__ uint64_t ConvCommon<CONV, CONV_TILING>::
435 CeilDiv(const uint64_t numA, const uint64_t numB)446 CeilDiv(const uint64_t numA, const uint64_t numB)
436{447{
437 return (numA + numB - 1) / numB;448 return (numA + numB - 1) / numB;
438}449}
439 450 
440-template <class CONV, class RUN_INFO>451+template <class CONV, class CONV_TILING>
441-__aicore__ __forceinline__ int64_t ConvCommon<CONV, RUN_INFO>::452+__aicore__ __forceinline__ int64_t ConvCommon<CONV, CONV_TILING>::
442 Max(const int64_t numA, const int64_t numB)453 Max(const int64_t numA, const int64_t numB)
443{454{
444 return numA > numB ? numA : numB;455 return numA > numB ? numA : numB;
Mconv/common/op_kernel/arch35/conv_common_func.h+147-88
@@ -104,45 +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.convTiling->quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {107+ if (self->ctx.convTilingData->convApiTiling.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.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {109+ if (self->ctx.convTilingData->convApiTiling.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 defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)113+ if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
114- if (self->ctx.convTiling->reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
115 float m2 = fixpipeParams.reluWeight0.GetValue(0);114 float m2 = fixpipeParams.reluWeight0.GetValue(0);
116 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)) {
117+ // copy from fixpipe paramIn to global tensor
118+ self->ctx.reluWeightGM.SetGlobalBuffer(fixpipeParams.reluWeight0.GetPhyAddr(0),
119+ fixpipeParams.reluWeight0.GetSize());
117 }120 }
118-#endif
119 if constexpr (Intf::isExtendConv2d) {121 if constexpr (Intf::isExtendConv2d) {
120- if (self->ctx.convTiling->dualOutput &&122+ if (self->ctx.convTilingData->convApiTiling.dualOutput &&
121- self->ctx.convTiling->quantMode1 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {123+ self->ctx.convTilingData->convApiTiling.quantMode1 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {
122 self->ctx.scale1gm.SetGlobalBuffer(fixpipeParams.scale1.GetPhyAddr(0), fixpipeParams.scale1.GetSize());124 self->ctx.scale1gm.SetGlobalBuffer(fixpipeParams.scale1.GetPhyAddr(0), fixpipeParams.scale1.GetSize());
123- if (self->ctx.convTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {125+ if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::SCALAR_QUANT)) {
124 self->ctx.deqScalar1 = fixpipeParams.scale1.GetValue(0);126 self->ctx.deqScalar1 = fixpipeParams.scale1.GetValue(0);
125 }127 }
126-#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)128+ if (self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
127- if (self->ctx.convTiling->reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
128 float m2 = fixpipeParams.reluWeight1.GetValue(0);129 float m2 = fixpipeParams.reluWeight1.GetValue(0);
129 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)) {
132+ self->ctx.reluWeight1GM.SetGlobalBuffer(fixpipeParams.reluWeight1.GetPhyAddr(0),
133+ fixpipeParams.reluWeight1.GetSize());
130 }134 }
131-#endif
132 }135 }
133- if (self->ctx.convTiling->quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT) ||136+ if (self->ctx.convTilingData->convApiTiling.quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT) ||
134- self->ctx.convTiling->quantMode1 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {137+ self->ctx.convTilingData->convApiTiling.quantMode1 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {
135 self->ctx.isFirstIterate = true;138 self->ctx.isFirstIterate = true;
136 }139 }
137- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||140+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
138- self->ctx.convTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {141+ self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
139 self->ctx.enableVectorQuant = true;142 self->ctx.enableVectorQuant = true;
140 }143 }
144+ if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
145+ self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
146+ self->ctx.enableVectorRelu = true;
147+ }
141 } else {148 } else {
142- if (self->ctx.convTiling->quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {149+ if (self->ctx.convTilingData->convApiTiling.quantMode0 != static_cast<uint8_t>(QuantModeType::NO_QUANT)) {
143 self->ctx.isFirstIterate = true;150 self->ctx.isFirstIterate = true;
144 }151 }
145- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {152+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
146 self->ctx.enableVectorQuant = true;153 self->ctx.enableVectorQuant = true;
147 }154 }
148 }155 }
@@ -168,7 +175,7 @@ struct GetTensorC {
168 if ASCEND_IS_AIC_CONV {175 if ASCEND_IS_AIC_CONV {
169 self->ctx.copyOutIns.template CopyOut<TensorTypeT, config>(output0, ubInfo);176 self->ctx.copyOutIns.template CopyOut<TensorTypeT, config>(output0, ubInfo);
170 if constexpr (Intf::isExtendConv2d) {177 if constexpr (Intf::isExtendConv2d) {
171- if (self->ctx.convTiling->dualOutput) {178+ if (self->ctx.convTilingData->convApiTiling.dualOutput) {
172 self->ctx.copyOutIns1.template CopyOut<GlobalTensor, config>(output1, ubInfo);179 self->ctx.copyOutIns1.template CopyOut<GlobalTensor, config>(output1, ubInfo);
173 }180 }
174 }181 }
@@ -177,9 +184,13 @@ struct GetTensorC {
177 self->ctx.queueCL0.FreeTensor(self->ctx.cl0);184 self->ctx.queueCL0.FreeTensor(self->ctx.cl0);
178 }185 }
179 186 
180- if (self->ctx.enableVectorQuant && !self->ctx.convTiling->fixpParamsFullLoadFlag) {187+ if (self->ctx.enableVectorQuant && !self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
181 self->ctx.queueScaleL1.FreeTensor(self->ctx.scaleL1);188 self->ctx.queueScaleL1.FreeTensor(self->ctx.scaleL1);
182 }189 }
190+ 
191+ if (self->ctx.enableVectorRelu && !self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
192+ self->ctx.queueReluWeightL1.FreeTensor(self->ctx.reluWeightL1);
193+ }
183 }194 }
184 }195 }
185};196};
@@ -190,44 +201,62 @@ struct ConvPreProcess {
190 {201 {
191 if ASCEND_IS_AIC_CONV {202 if ASCEND_IS_AIC_CONV {
192 if constexpr (!Intf::isDeQuantFlag) {203 if constexpr (!Intf::isDeQuantFlag) {
193- if (self->ctx.convTiling->biasFullLoadFlag && self->ctx.enableBias) {204+ if (self->ctx.convTilingData->convApiTiling.biasFullLoadFlag && self->ctx.enableBias) {
194 self->ctx.biasL1 = self->ctx.queueBiasL1.template AllocTensor<typename Intf::BiasT>();205 self->ctx.biasL1 = self->ctx.queueBiasL1.template AllocTensor<typename Intf::BiasT>();
195 uint64_t biasLoadNum = self->ctx.singleCoreCo;206 uint64_t biasLoadNum = self->ctx.singleCoreCo;
196 if constexpr (Intf::groupOptPreloadFlag) {207 if constexpr (Intf::groupOptPreloadFlag) {
197- biasLoadNum = self->ctx.orgCo;208+ biasLoadNum = self->ctx.convTilingData->convApiTiling.orgCo;
198 }209 }
199 self->ctx.loadBiasL1Ins.LoadChannelWiseL1FullLoad(self->ctx.biasL1, self->ctx.biasgm,210 self->ctx.loadBiasL1Ins.LoadChannelWiseL1FullLoad(self->ctx.biasL1, self->ctx.biasgm,
200 biasLoadNum, 0);211 biasLoadNum, 0);
201 self->ctx.queueBiasL1.EnQue(self->ctx.biasL1);212 self->ctx.queueBiasL1.EnQue(self->ctx.biasL1);
202 self->ctx.biasL1 = self->ctx.queueBiasL1.template DeQue<typename Intf::BiasT>();213 self->ctx.biasL1 = self->ctx.queueBiasL1.template DeQue<typename Intf::BiasT>();
203 }214 }
204- if (self->ctx.convTiling->fixpParamsFullLoadFlag && self->ctx.enableVectorQuant) {215+ if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
205- if constexpr (Intf::groupType != static_cast<int8_t>(ConvGroupType::NORMAL_CONV)) {216+ if (self->ctx.enableVectorQuant || self->ctx.enableVectorRelu) {
206- event_t eventId = static_cast<event_t>(self->ctx.pipe.FetchEventID(HardEvent::FIX_MTE2));217+ if constexpr (Intf::groupType != static_cast<int8_t>(ConvGroupType::NORMAL_CONV)) {
207- SetFlag<HardEvent::FIX_MTE2>(eventId);218+ event_t eventId = static_cast<event_t>(self->ctx.pipe.FetchEventID(HardEvent::FIX_MTE2));
208- WaitFlag<HardEvent::FIX_MTE2>(eventId);219+ SetFlag<HardEvent::FIX_MTE2>(eventId);
220+ WaitFlag<HardEvent::FIX_MTE2>(eventId);
221+ }
209 }222 }
210- self->ctx.scaleL1 = self->ctx.queueScaleL1.template AllocTensor<typename Intf::ScaleT>();
211 uint64_t scaleLoadNum = self->ctx.singleCoreCo;223 uint64_t scaleLoadNum = self->ctx.singleCoreCo;
212 if constexpr (Intf::groupOptPreloadFlag) {224 if constexpr (Intf::groupOptPreloadFlag) {
213- scaleLoadNum = self->ctx.orgCo;225+ scaleLoadNum = self->ctx.convTilingData->convApiTiling.orgCo;
214 }226 }
215- if constexpr (Intf::isExtendConv2d) {227+ if (self->ctx.enableVectorQuant) {
216- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {228+ self->ctx.scaleL1 = self->ctx.queueScaleL1.template AllocTensor<typename Intf::ScaleT>();
229+ if constexpr (Intf::isExtendConv2d) {
230+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
231+ self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1, self->ctx.scalegm,
232+ scaleLoadNum, 0);
233+ }
234+ if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
235+ self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1[self->ctx.scale1L1offset],
236+ self->ctx.scale1gm, scaleLoadNum, 0);
237+ }
238+ } else {
217 self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1, self->ctx.scalegm,239 self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1, self->ctx.scalegm,
218 scaleLoadNum, 0);240 scaleLoadNum, 0);
219 }241 }
220- if (self->ctx.convTiling->dualOutput &&242+ self->ctx.queueScaleL1.EnQue(self->ctx.scaleL1);
221- self->ctx.convTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {243+ self->ctx.scaleL1 = self->ctx.queueScaleL1.template DeQue<typename Intf::ScaleT>();
222- self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1[self->ctx.scale1L1offset],244+ }
223- self->ctx.scale1gm, scaleLoadNum, 0);245+ if constexpr (Intf::isExtendConv2d) {
224- }246+ if (self->ctx.enableVectorRelu) {
225- } else {247+ self->ctx.reluWeightL1 =
226- self->ctx.loadScaleL1Ins.LoadChannelWiseL1FullLoad(self->ctx.scaleL1, self->ctx.scalegm,248+ self->ctx.queueReluWeightL1.template AllocTensor<typename Intf::ReluWeightT>();
227- scaleLoadNum, 0);249+ if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
250+ self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1FullLoad(self->ctx.reluWeightL1,
251+ self->ctx.reluWeightGM, scaleLoadNum, 0);
252+ }
253+ if (self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
254+ self->ctx.loadReluWeightL1Ins.LoadChannelWiseL1FullLoad(
255+ self->ctx.reluWeightL1[self->ctx.reluWeight1L1offset],
256+ self->ctx.reluWeight1GM, scaleLoadNum, 0);
257+ }
258+ }
228 }259 }
229- self->ctx.queueScaleL1.EnQue(self->ctx.scaleL1);
230- self->ctx.scaleL1 = self->ctx.queueScaleL1.template DeQue<typename Intf::ScaleT>();
231 }260 }
232 }261 }
233 }262 }
@@ -242,13 +271,20 @@ struct ConvPostProcess {
242 {271 {
243 if ASCEND_IS_AIC_CONV {272 if ASCEND_IS_AIC_CONV {
244 if constexpr (!Intf::isDeQuantFlag) {273 if constexpr (!Intf::isDeQuantFlag) {
245- if (self->ctx.convTiling->biasFullLoadFlag && self->ctx.enableBias) {274+ if (self->ctx.convTilingData->convApiTiling.biasFullLoadFlag && self->ctx.enableBias) {
246 self->ctx.queueBiasL1.FreeTensor(self->ctx.biasL1);275 self->ctx.queueBiasL1.FreeTensor(self->ctx.biasL1);
247 }276 }
248- if (self->ctx.convTiling->fixpParamsFullLoadFlag && self->ctx.enableVectorQuant) {277+ if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag && self->ctx.enableVectorQuant) {
249 self->ctx.queueScaleL1.FreeTensor(self->ctx.scaleL1);278 self->ctx.queueScaleL1.FreeTensor(self->ctx.scaleL1);
250 }279 }
251 }280 }
281+ if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag && self->ctx.enableVectorRelu) {
282+ self->ctx.queueReluWeightL1.FreeTensor(self->ctx.reluWeightL1);
283+ }
284+ 
285+ if constexpr (Intf::isMPreLoad) {
286+ self->ctx.queueBL1.FreeTensor(self->ctx.bl1);
287+ }
252 288 
253 self->ctx.isFirstIterate = true;289 self->ctx.isFirstIterate = true;
254 }290 }
@@ -315,7 +351,7 @@ struct IterateAll {
315 const GlobalTensor<typename Intf::Output1T> &output1) {351 const GlobalTensor<typename Intf::Output1T> &output1) {
316 if constexpr (Intf::formatOutput == ConvFormat::NHWC) {352 if constexpr (Intf::formatOutput == ConvFormat::NHWC) {
317 if constexpr (Intf::isFixedPoint) {353 if constexpr (Intf::isFixedPoint) {
318- if (self->ctx.convTiling->dualOutput) {354+ if (self->ctx.convTilingData->convApiTiling.dualOutput) {
319 GetTensorC<Intf, ImplType>::template355 GetTensorC<Intf, ImplType>::template
320 call<GlobalTensor, CFG_ROW_MAJOR_FIXED_POINT, sync>(self, output0, output1);356 call<GlobalTensor, CFG_ROW_MAJOR_FIXED_POINT, sync>(self, output0, output1);
321 } else {357 } else {
@@ -323,7 +359,7 @@ struct IterateAll {
323 call<GlobalTensor, CFG_ROW_MAJOR_FIXED_POINT, sync>(self, output0);359 call<GlobalTensor, CFG_ROW_MAJOR_FIXED_POINT, sync>(self, output0);
324 }360 }
325 } else {361 } else {
326- if (self->ctx.convTiling->dualOutput) {362+ if (self->ctx.convTilingData->convApiTiling.dualOutput) {
327 GetTensorC<Intf, ImplType>::template363 GetTensorC<Intf, ImplType>::template
328 call<GlobalTensor, CFG_ROW_MAJOR, sync>(self, output0, output1);364 call<GlobalTensor, CFG_ROW_MAJOR, sync>(self, output0, output1);
329 } else {365 } else {
@@ -333,7 +369,7 @@ struct IterateAll {
333 }369 }
334 } else {370 } else {
335 if constexpr (Intf::isFixedPoint) {371 if constexpr (Intf::isFixedPoint) {
336- if (self->ctx.convTiling->dualOutput) {372+ if (self->ctx.convTilingData->convApiTiling.dualOutput) {
337 GetTensorC<Intf, ImplType>::template373 GetTensorC<Intf, ImplType>::template
338 call<GlobalTensor, CFG_COLUMN_MAJOR_FIXED_POINT, sync>(self, output0, output1);374 call<GlobalTensor, CFG_COLUMN_MAJOR_FIXED_POINT, sync>(self, output0, output1);
339 } else {375 } else {
@@ -341,7 +377,7 @@ struct IterateAll {
341 call<GlobalTensor, CFG_COLUMN_MAJOR_FIXED_POINT, sync>(self, output0);377 call<GlobalTensor, CFG_COLUMN_MAJOR_FIXED_POINT, sync>(self, output0);
342 }378 }
343 } else {379 } else {
344- if (self->ctx.convTiling->dualOutput) {380+ if (self->ctx.convTilingData->convApiTiling.dualOutput) {
345 GetTensorC<Intf, ImplType>::template381 GetTensorC<Intf, ImplType>::template
346 call<GlobalTensor, CFG_COLUMN_MAJOR, sync>(self, output0, output1);382 call<GlobalTensor, CFG_COLUMN_MAJOR, sync>(self, output0, output1);
347 } else {383 } else {
@@ -363,6 +399,7 @@ struct End {
363 if constexpr (!Intf::isDeQuantFlag) {399 if constexpr (!Intf::isDeQuantFlag) {
364 self->ctx.queueBiasL1.FreeAllEvent();400 self->ctx.queueBiasL1.FreeAllEvent();
365 self->ctx.queueScaleL1.FreeAllEvent();401 self->ctx.queueScaleL1.FreeAllEvent();
402+ self->ctx.queueReluWeightL1.FreeAllEvent();
366 self->ctx.queueBiasBT.FreeAllEvent();403 self->ctx.queueBiasBT.FreeAllEvent();
367 }404 }
368 405 
@@ -370,6 +407,15 @@ struct End {
370 self->ctx.queueCL0.FreeAllEvent();407 self->ctx.queueCL0.FreeAllEvent();
371 }408 }
372 }409 }
410+ if constexpr (Intf::preFusionFlag) {
411+ GetTPipePtr()->ReleaseEventID<AscendC::HardEvent::MTE3_MTE1>(self->ctx.eventIdMte3ToMte1Ping);
412+ GetTPipePtr()->ReleaseEventID<AscendC::HardEvent::MTE3_MTE1>(self->ctx.eventIdMte3ToMte1Pong);
413+ GetTPipePtr()->ReleaseEventID<AscendC::HardEvent::MTE1_MTE3>(self->ctx.eventIdMte1ToMte3Ping);
414+ GetTPipePtr()->ReleaseEventID<AscendC::HardEvent::MTE1_MTE3>(self->ctx.eventIdMte1ToMte3Pong);
415+ if constexpr ((Intf::preFusionFlag && ConvParam::weightTiling == 0)) {
416+ self->ctx.queueBL1.FreeTensor(self->ctx.bl1);
417+ }
418+ }
373 }419 }
374};420};
375 421 
@@ -381,14 +427,11 @@ __aicore__ inline void InitBufferWithDoubleBuf(Intf *self)
381 self->ctx.wholeAl0Tensor = self->ctx.al0Buf.template Get<typename Intf::FmapT>();427 self->ctx.wholeAl0Tensor = self->ctx.al0Buf.template Get<typename Intf::FmapT>();
382 self->ctx.wholeBl0Tensor = self->ctx.bl0Buf.template Get<typename Intf::WeightT>();428 self->ctx.wholeBl0Tensor = self->ctx.bl0Buf.template Get<typename Intf::WeightT>();
383 429 
384- int8_t cl0db = (self->ctx.convTiling->pBufferFlag & 0x04) >> 2;430+ int8_t cl0db = (self->ctx.convTilingData->convApiTiling.pBufferFlag & 0x04) >> 2;
385- int8_t al1db = (self->ctx.convTiling->pBufferFlag & 0x08) >> 3;
386- int8_t bl1db = (self->ctx.convTiling->pBufferFlag & 0x10) >> 4;
387- 
388 if constexpr (Intf::isInnerBatchFlag || Intf::isDeQuantFlag) {431 if constexpr (Intf::isInnerBatchFlag || Intf::isDeQuantFlag) {
389- uint64_t cl0Spacesize = self->ctx.convTiling->mStep * self->ctx.convTiling->nL0;432+ uint64_t cl0Spacesize = self->ctx.convTilingData->convApiTiling.mStep * self->ctx.convTilingData->convApiTiling.nL0;
390 if constexpr (Intf::isInnerBatchFlag) {433 if constexpr (Intf::isInnerBatchFlag) {
391- cl0Spacesize *= self->ctx.convTiling->innerBatch;434+ cl0Spacesize *= self->ctx.convTilingData->convApiTiling.innerBatch;
392 }435 }
393 if (!cl0db) {436 if (!cl0db) {
394 self->ctx.pipe.InitBuffer(self->ctx.queueCL0, 1, cl0Spacesize * Intf::sizeOfL0c);437 self->ctx.pipe.InitBuffer(self->ctx.queueCL0, 1, cl0Spacesize * Intf::sizeOfL0c);
@@ -400,10 +443,10 @@ __aicore__ inline void InitBufferWithDoubleBuf(Intf *self)
400 self->ctx.wholeCl0Tensor = self->ctx.l0cBuf.template Get<typename Intf::L0cT>();443 self->ctx.wholeCl0Tensor = self->ctx.l0cBuf.template Get<typename Intf::L0cT>();
401 }444 }
402 445 
403- if (!al1db) {446+ if (!((self->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET)) {
404- self->ctx.pipe.InitBuffer(self->ctx.queueAL1, 1, self->ctx.convTiling->aL1SpaceSize);447+ self->ctx.pipe.InitBuffer(self->ctx.queueAL1, 1, self->ctx.convTilingData->convApiTiling.aL1SpaceSize);
405 } else {448 } else {
406- self->ctx.pipe.InitBuffer(self->ctx.queueAL1, DOUBLE_BUF, self->ctx.convTiling->aL1SpaceSize);449+ self->ctx.pipe.InitBuffer(self->ctx.queueAL1, DOUBLE_BUF, self->ctx.convTilingData->convApiTiling.aL1SpaceSize);
407 }450 }
408 451 
409 if constexpr (!Intf::bL1DBFlag) {452 if constexpr (!Intf::bL1DBFlag) {
@@ -420,7 +463,7 @@ __aicore__ inline void InitBufferWithDoubleBuf(Intf *self)
420 463 
421 if constexpr (Intf::isDeQuantFlag) {464 if constexpr (Intf::isDeQuantFlag) {
422 self->ctx.pipe.InitBuffer(self->ctx.mmadResUbBuf,465 self->ctx.pipe.InitBuffer(self->ctx.mmadResUbBuf,
423- self->ctx.convTiling->mUB * self->ctx.convTiling->nUB * Intf::sizeOfL0c);466+ self->ctx.convTilingData->convApiTiling.mUB * self->ctx.convTilingData->convApiTiling.nUB * Intf::sizeOfL0c);
424 self->ctx.mmadResUbTensor = self->ctx.mmadResUbBuf.template Get<typename Intf::L0cT>();467 self->ctx.mmadResUbTensor = self->ctx.mmadResUbBuf.template Get<typename Intf::L0cT>();
425 }468 }
426}469}
@@ -428,21 +471,21 @@ __aicore__ inline void InitBufferWithDoubleBuf(Intf *self)
428template <class Intf>471template <class Intf>
429__aicore__ inline void InitBuffer(Intf *self)472__aicore__ inline void InitBuffer(Intf *self)
430{473{
431- self->ctx.bL1SpaceSize = self->ctx.convTiling->nBL1 * self->ctx.convTiling->kBL1;474+ self->ctx.bL1SpaceSize = self->ctx.convTilingData->convApiTiling.nBL1 * self->ctx.convTilingData->convApiTiling.kBL1;
432 475 
433 InitBufferWithDoubleBuf<Intf>(self);476 InitBufferWithDoubleBuf<Intf>(self);
434 477
435 if constexpr (!Intf::isDeQuantFlag) {478 if constexpr (!Intf::isDeQuantFlag) {
436- if (self->ctx.convTiling->hasBias) {479+ if (self->ctx.convTilingData->convApiTiling.hasBias) {
437- uint64_t biasl1Spacesize = self->ctx.convTiling->biasFullLoadFlag ? AlignB(480+ uint64_t biasl1Spacesize = self->ctx.convTilingData->convApiTiling.biasFullLoadFlag ? AlignB(
438 self->ctx.singleCoreCo * Intf::sizeOfBias, BLOCK_L0_N * Intf::sizeOfBias) :481 self->ctx.singleCoreCo * Intf::sizeOfBias, BLOCK_L0_N * Intf::sizeOfBias) :
439- self->ctx.convTiling->nL0 * Intf::sizeOfBias;482+ self->ctx.convTilingData->convApiTiling.nL0 * Intf::sizeOfBias;
440 if constexpr (Intf::groupOptPreloadFlag) {483 if constexpr (Intf::groupOptPreloadFlag) {
441- if (self->ctx.convTiling->biasFullLoadFlag) {484+ if (self->ctx.convTilingData->convApiTiling.biasFullLoadFlag) {
442- biasl1Spacesize = AlignB(self->ctx.orgCo * Intf::sizeOfBias, BLOCK_L0_N * Intf::sizeOfBias);485+ biasl1Spacesize = AlignB(self->ctx.convTilingData->convApiTiling.orgCo * Intf::sizeOfBias, BLOCK_L0_N * Intf::sizeOfBias);
443 }486 }
444 }487 }
445- uint64_t biasBTSpacesize = self->ctx.convTiling->nL0;488+ uint64_t biasBTSpacesize = self->ctx.convTilingData->convApiTiling.nL0;
446 self->ctx.pipe.InitBuffer(self->ctx.queueBiasL1, 1, AlignB(biasl1Spacesize, C0_SIZE));489 self->ctx.pipe.InitBuffer(self->ctx.queueBiasL1, 1, AlignB(biasl1Spacesize, C0_SIZE));
447 self->ctx.pipe.InitBuffer(490 self->ctx.pipe.InitBuffer(
448 self->ctx.queueBiasBT, 1, AlignB(biasBTSpacesize * Intf::sizeOfL0c, BT_SIZE));491 self->ctx.queueBiasBT, 1, AlignB(biasBTSpacesize * Intf::sizeOfL0c, BT_SIZE));
@@ -450,34 +493,50 @@ __aicore__ inline void InitBuffer(Intf *self)
450 }493 }
451 494 
452 if constexpr (Intf::isExtendConv2d) {495 if constexpr (Intf::isExtendConv2d) {
453- uint64_t scaleL1SpaceSize = 0;496+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
454- uint64_t scale0L1Size = self->ctx.convTiling->fixpParamsFullLoadFlag ? AlignB(497+ self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
455- self->ctx.singleCoreCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale) :498+ uint64_t scaleL1SpaceSize = 0;
456- self->ctx.convTiling->nL0 * Intf::sizeOfScale;499+ uint64_t scale0L1Size = self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag ? AlignB(
457- if constexpr (Intf::groupOptPreloadFlag) {500+ self->ctx.singleCoreCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale) :
458- if (self->ctx.convTiling->fixpParamsFullLoadFlag) {501+ self->ctx.convTilingData->convApiTiling.nL0 * Intf::sizeOfScale;
459- scale0L1Size = AlignB(self->ctx.orgCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale);502+ if constexpr (Intf::groupOptPreloadFlag) {
503+ if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
504+ scale0L1Size = AlignB(self->ctx.convTilingData->convApiTiling.orgCo * Intf::sizeOfScale,
505+ BLOCK_L0_N * Intf::sizeOfScale);
506+ }
507+ }
508+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
509+ scaleL1SpaceSize += scale0L1Size;
510+ }
511+ if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
512+ self->ctx.scale1L1offset = scaleL1SpaceSize / Intf::sizeOfScale;
513+ scaleL1SpaceSize += scale0L1Size;
460 }514 }
461- }
462- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
463- scaleL1SpaceSize += scale0L1Size;
464- }
465- if (self->ctx.convTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
466- self->ctx.scale1L1offset = scaleL1SpaceSize / Intf::sizeOfScale;
467- scaleL1SpaceSize += scale0L1Size;
468- }
469- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
470- self->ctx.convTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
471 self->ctx.pipe.InitBuffer(self->ctx.queueScaleL1, 1, AlignB(scaleL1SpaceSize, C0_SIZE));515 self->ctx.pipe.InitBuffer(self->ctx.queueScaleL1, 1, AlignB(scaleL1SpaceSize, C0_SIZE));
472 }516 }
517+ if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
518+ self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
519+ uint32_t reluWeightL1SpaceSize = 0;
520+ uint32_t reluWeight0L1Size = self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag ? AlignB(
521+ self->ctx.singleCoreCo * Intf::sizeOfReluWeight, BLOCK_L0_N * Intf::sizeOfReluWeight) :
522+ self->ctx.convTilingData->convApiTiling.nL0 * Intf::sizeOfReluWeight;
523+ if (self->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
524+ reluWeightL1SpaceSize += reluWeight0L1Size;
525+ }
526+ if (self->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
527+ self->ctx.reluWeight1L1offset = reluWeightL1SpaceSize / Intf::sizeOfReluWeight;
528+ reluWeightL1SpaceSize += reluWeight0L1Size;
529+ }
530+ self->ctx.pipe.InitBuffer(self->ctx.queueReluWeightL1, 1, AlignB(reluWeightL1SpaceSize, C0_SIZE));
531+ }
473 } else if constexpr (Intf::isQuantScene) {532 } else if constexpr (Intf::isQuantScene) {
474- if (self->ctx.convTiling->hasScale != 0) {533+ if (self->ctx.convTilingData->convApiTiling.hasScale != 0) {
475- uint64_t scaleL1SpaceSize = self->ctx.convTiling->fixpParamsFullLoadFlag ? AlignB(534+ uint64_t scaleL1SpaceSize = self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag ? AlignB(
476 self->ctx.singleCoreCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale) :535 self->ctx.singleCoreCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale) :
477- self->ctx.convTiling->nL0 * Intf::sizeOfScale;536+ self->ctx.convTilingData->convApiTiling.nL0 * Intf::sizeOfScale;
478 if constexpr (Intf::groupOptPreloadFlag) {537 if constexpr (Intf::groupOptPreloadFlag) {
479- if (self->ctx.convTiling->fixpParamsFullLoadFlag) {538+ if (self->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
480- scaleL1SpaceSize = AlignB(self->ctx.orgCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale);539+ scaleL1SpaceSize = AlignB(self->ctx.convTilingData->convApiTiling.orgCo * Intf::sizeOfScale, BLOCK_L0_N * Intf::sizeOfScale);
481 }540 }
482 }541 }
483 self->ctx.pipe.InitBuffer(self->ctx.queueScaleL1, 1, AlignB(scaleL1SpaceSize, C0_SIZE));542 self->ctx.pipe.InitBuffer(self->ctx.queueScaleL1, 1, AlignB(scaleL1SpaceSize, C0_SIZE));
@@ -488,9 +547,9 @@ __aicore__ inline void InitBuffer(Intf *self)
488template <class Intf>547template <class Intf>
489__aicore__ inline void InitHf32Mode(Intf *self)548__aicore__ inline void InitHf32Mode(Intf *self)
490{549{
491- if (self->ctx.convTiling->hf32Enable) {550+ if (self->ctx.convTilingData->convApiTiling.hf32Enable) {
492- SetHF32Mode(self->ctx.convTiling->hf32Enable);551+ SetHF32Mode(self->ctx.convTilingData->convApiTiling.hf32Enable);
493- SetHF32TransMode(self->ctx.convTiling->hf32TransMode);552+ SetHF32TransMode(self->ctx.convTilingData->convApiTiling.hf32TransMode);
494 } else {553 } else {
495 SetHF32Mode(0);554 SetHF32Mode(0);
496 SetHF32TransMode(0);555 SetHF32TransMode(0);
Mconv/common/op_kernel/arch35/conv_config.h+22-18
@@ -59,14 +59,14 @@ struct Extendconv2dFixpipeParams {
59 59 
60struct CopyUbInfo {60struct CopyUbInfo {
61 // only support batchUb = 1 in tiling hw mode61 // only support batchUb = 1 in tiling hw mode
62- // when batchUb > 1, mUb should equal to convTiling->hoL062+ // when batchUb > 1, mUb should equal to hoL0
63 uint64_t batchUb = 0;63 uint64_t batchUb = 0;
64 // nUb should be aligned to BLOCK_L0_N64 // nUb should be aligned to BLOCK_L0_N
65 uint64_t nUb = 0;65 uint64_t nUb = 0;
66 // mUb should be aligned to BLOCK_L0_M66 // mUb should be aligned to BLOCK_L0_M
67- // In tiling m mode, mUb is part of convTiling->hoL067+ // In tiling m mode, mUb is part of hoL0
68- // In tiling hw mode, mUb is part of convTiling->hoL0*convTiling->woL0,68+ // In tiling hw mode, mUb is part of hoL0*woL0,
69- // additionally, when convTiling->hoL0 != 1, mUb must be >= convTiling->woL0.69+ // additionally, when hoL0 != 1, mUb must be >= woL0.
70 uint64_t mUb = 0;70 uint64_t mUb = 0;
71 uint64_t batchLoopIdx = 0;71 uint64_t batchLoopIdx = 0;
72 uint64_t nLoopIdx = 0;72 uint64_t nLoopIdx = 0;
@@ -164,6 +164,9 @@ struct ConvParam {
164 constexpr static int8_t iterOrder = 0;164 constexpr static int8_t iterOrder = 0;
165 constexpr static int8_t groupType = 0;165 constexpr static int8_t groupType = 0;
166 constexpr static int8_t innerBatch = static_cast<int8_t>(ConvInnerBatch::SINGLE_BATCH);166 constexpr static int8_t innerBatch = static_cast<int8_t>(ConvInnerBatch::SINGLE_BATCH);
167+ constexpr static int8_t batchOne = 0;
168+ constexpr static int8_t noPad = 0;
169+ constexpr static int8_t smallWeight = 0;
167};170};
168 171 
169template <typename T>172template <typename T>
@@ -263,8 +266,11 @@ public:
263 GlobalTensor<typename ConvDataType::BiasT> biasgm;266 GlobalTensor<typename ConvDataType::BiasT> biasgm;
264 GlobalTensor<typename ConvDataType::ScaleT> scalegm;267 GlobalTensor<typename ConvDataType::ScaleT> scalegm;
265 GlobalTensor<typename ConvDataType::ScaleT> scale1gm;268 GlobalTensor<typename ConvDataType::ScaleT> scale1gm;
269+ GlobalTensor<typename ConvDataType::ReluWeightT> reluWeightGM;
270+ GlobalTensor<typename ConvDataType::ReluWeightT> reluWeight1GM;
266 TQue<QuePosition::A1, 1> queueBiasL1; // BiasL1271 TQue<QuePosition::A1, 1> queueBiasL1; // BiasL1
267 TQue<QuePosition::A1, 1> queueScaleL1; // ScaleL1272 TQue<QuePosition::A1, 1> queueScaleL1; // ScaleL1
273+ TQue<QuePosition::A1, 1> queueReluWeightL1; // ReluWeightL1
268 TQue<TPosition::C2, 1> queueBiasBT; // BT274 TQue<TPosition::C2, 1> queueBiasBT; // BT
269 TQue<QuePosition::CO1, 1> queueCL0;275 TQue<QuePosition::CO1, 1> queueCL0;
270 TBuf<TPosition::A2> al0Buf;276 TBuf<TPosition::A2> al0Buf;
@@ -275,10 +281,9 @@ public:
275 LocalTensor<typename ConvDataType::WeightT> bl1;281 LocalTensor<typename ConvDataType::WeightT> bl1;
276 LocalTensor<typename ConvDataType::BiasT> biasL1;282 LocalTensor<typename ConvDataType::BiasT> biasL1;
277 LocalTensor<typename ConvDataType::ScaleT> scaleL1;283 LocalTensor<typename ConvDataType::ScaleT> scaleL1;
284+ LocalTensor<typename ConvDataType::ReluWeightT> reluWeightL1;
278 LocalTensor<typename ConvDataType::FmapT> wholeAl0Tensor;285 LocalTensor<typename ConvDataType::FmapT> wholeAl0Tensor;
279- LocalTensor<typename ConvDataType::FmapT> al0;
280 LocalTensor<typename ConvDataType::WeightT> wholeBl0Tensor;286 LocalTensor<typename ConvDataType::WeightT> wholeBl0Tensor;
281- LocalTensor<typename ConvDataType::WeightT> bl0;
282 LocalTensor<typename ConvDataType::L0cT> biasBT;287 LocalTensor<typename ConvDataType::L0cT> biasBT;
283 LocalTensor<typename ConvDataType::L0cT> wholeCl0Tensor =288 LocalTensor<typename ConvDataType::L0cT> wholeCl0Tensor =
284 LocalTensor<typename ConvDataType::L0cT>(TPosition::CO1, 0, 0);289 LocalTensor<typename ConvDataType::L0cT>(TPosition::CO1, 0, 0);
@@ -287,7 +292,9 @@ public:
287 292 
288 uint8_t enableBias = false; // 是否有bias293 uint8_t enableBias = false; // 是否有bias
289 uint8_t enableVectorQuant = false; // 是否有vector类型scale，双输出场景下任一scale为vector类型，即为true294 uint8_t enableVectorQuant = false; // 是否有vector类型scale，双输出场景下任一scale为vector类型，即为true
295+ uint8_t enableVectorRelu = false; // 是否有vector类型relu_weight，双输出场景下任一relu_weight为vector类型，即为true
290 uint64_t scale1L1offset = 0; // 双输出时scale1在L1上的偏移296 uint64_t scale1L1offset = 0; // 双输出时scale1在L1上的偏移
297+ uint64_t reluWeight1L1offset = 0; // 双输出时reluWeight1在L1上的偏移
291 uint64_t deqScalar0 = DEQ_SCALAR_ONE; // m1 = 1.0, other bits are 0.298 uint64_t deqScalar0 = DEQ_SCALAR_ONE; // m1 = 1.0, other bits are 0.
292 uint64_t deqScalar1 = DEQ_SCALAR_ONE; // m1 = 1.0, other bits are 0.299 uint64_t deqScalar1 = DEQ_SCALAR_ONE; // m1 = 1.0, other bits are 0.
293 uint64_t preReluScalar0 = 0;300 uint64_t preReluScalar0 = 0;
@@ -396,6 +403,14 @@ public:
396 uint8_t loadBL1Flag = true; // 是否载入BL1的标志403 uint8_t loadBL1Flag = true; // 是否载入BL1的标志
397 uint8_t loadAL0Flag = true; // 是否载入AL0的标志404 uint8_t loadAL0Flag = true; // 是否载入AL0的标志
398 uint8_t loadBL0Flag = true; // 是否载入BL0的标志405 uint8_t loadBL0Flag = true; // 是否载入BL0的标志
406+ uint8_t updateL1Flag = false; // 是否更新AL1数据的标志
407+ // 用于管理前融合场景下MTE1/MTE3之前的同步管理，在后续处理中赋值后使用
408+ AscendC::TEventID eventIdMte3ToMte1;
409+ AscendC::TEventID eventIdMte1ToMte3;
410+ AscendC::TEventID eventIdMte3ToMte1Ping;
411+ AscendC::TEventID eventIdMte3ToMte1Pong;
412+ AscendC::TEventID eventIdMte1ToMte3Ping;
413+ AscendC::TEventID eventIdMte1ToMte3Pong;
399 414 
400 uint64_t kIter = 0; // k方向迭代器，从DDR到L0415 uint64_t kIter = 0; // k方向迭代器，从DDR到L0
401 uint64_t batchIter = 0; // batch方向迭代器416 uint64_t batchIter = 0; // batch方向迭代器
@@ -406,7 +421,6 @@ public:
406 uint64_t kBL0Iter = 0; // L1B 到L0方向的迭代器421 uint64_t kBL0Iter = 0; // L1B 到L0方向的迭代器
407 uint64_t nL0Iter = 0; // BL0上n方向迭代器422 uint64_t nL0Iter = 0; // BL0上n方向迭代器
408 423 
409- uint64_t maxBatchIter = 0;
410 uint64_t multiKAL1 = 1;424 uint64_t multiKAL1 = 1;
411 uint64_t multiKBL1 = 1;425 uint64_t multiKBL1 = 1;
412 uint64_t maxKAL1Iter = 0;426 uint64_t maxKAL1Iter = 0;
@@ -444,17 +458,7 @@ public:
444 uint16_t kL0FullLoadAl0PingPongFlag = 0;458 uint16_t kL0FullLoadAl0PingPongFlag = 0;
445 uint16_t kL0FullLoadBl0PingPongFlag = 0;459 uint16_t kL0FullLoadBl0PingPongFlag = 0;
446 uint16_t cl0PingPongFlag = 0;460 uint16_t cl0PingPongFlag = 0;
447- 461+ uint16_t al1PingPongFlag = 0;
448- // Input shape info
449- uint64_t orgBatch = 0; // fmap上batch大小
450- uint64_t orgCi = 0; // fmap上cin大小
451- uint64_t orgCo = 0; // weight上cout大小
452- uint64_t orgHi = 0; // fmap上h大小
453- uint64_t orgWi = 0; // fmap上w大小
454- uint64_t orgHo = 0; // output上h大小
455- uint64_t orgWo = 0; // output上w大小
456- uint64_t kernelH = 0; // weight上h大小
457- uint64_t kernelW = 0; // weight上w大小
458 };462 };
459};463};
460 464 
Mconv/common/op_kernel/arch35/conv_framework_util.h+3-0
@@ -62,6 +62,9 @@ using namespace AscendC;
62#define CONV_CHECK_FUN_TEMPLATE(T, NAMESPACE, ARG, ...) \62#define CONV_CHECK_FUN_TEMPLATE(T, NAMESPACE, ARG, ...) \
63 (!IsSameType<decltype(T::template call<ARG>(__VA_ARGS__)), NAMESPACE::TypeFalse>::value)63 (!IsSameType<decltype(T::template call<ARG>(__VA_ARGS__)), NAMESPACE::TypeFalse>::value)
64 64 
65+#define CONV_CHECK_FUN_THREE_TEMPLATE(T, NAMESPACE, ARG0, ARG1, ARG2, ...) \
66+ (!IsSameType<decltype(T::template call<ARG0, ARG1, ARG2>(__VA_ARGS__)), NAMESPACE::TypeFalse>::value)
67+ 
65// Define struct _has_member_MEMBER for check whether MEMBER exists.68// Define struct _has_member_MEMBER for check whether MEMBER exists.
66#define CONV_DECLARE_CHECK_MEMBER(MEMBER) namespace __ConvFramework { \69#define CONV_DECLARE_CHECK_MEMBER(MEMBER) namespace __ConvFramework { \
67 template <typename T> \70 template <typename T> \
Mconv/common/op_kernel/arch35/conv_group_common.h+116-105
@@ -18,8 +18,8 @@
18 18 
19#include "conv_common.h"19#include "conv_common.h"
20 20 
21-template <class CONV, class RUN_INFO>21+template <class CONV, class CONV_TILING>
22-class ConvGroupCommon : public ConvCommon<CONV, RUN_INFO> {22+class ConvGroupCommon : public ConvCommon<CONV, CONV_TILING> {
23public:23public:
24 __aicore__ __forceinline__ void CalcStartAddrFmapHW();24 __aicore__ __forceinline__ void CalcStartAddrFmapHW();
25 25 
@@ -41,14 +41,16 @@ public:
41 __aicore__ __forceinline__ void SetOptGroupTail();41 __aicore__ __forceinline__ void SetOptGroupTail();
42 42 
43private:43private:
44- using ConvCommon<CONV, RUN_INFO>::convOps;44+ using ConvCommon<CONV, CONV_TILING>::convOps;
45- using ConvCommon<CONV, RUN_INFO>::convRunInfo;45+ // 2d and 3d is conflict, need to fix in future
46- using ConvCommon<CONV, RUN_INFO>::hasScale;46+ using ConvCommon<CONV, CONV_TILING>::convTilingData;
47- using ConvCommon<CONV, RUN_INFO>::fmStartAddr;47+ using ConvCommon<CONV, CONV_TILING>::hasScale;
48- using ConvCommon<CONV, RUN_INFO>::weightStartAddr;48+ using ConvCommon<CONV, CONV_TILING>::fmStartAddr;
49- using ConvCommon<CONV, RUN_INFO>::biasStartAddr;49+ using ConvCommon<CONV, CONV_TILING>::weightStartAddr;
50- using ConvCommon<CONV, RUN_INFO>::scaleStartAddr;50+ using ConvCommon<CONV, CONV_TILING>::biasStartAddr;
51- using ConvCommon<CONV, RUN_INFO>::outputStartAddr;51+ using ConvCommon<CONV, CONV_TILING>::scaleStartAddr;
52+ using ConvCommon<CONV, CONV_TILING>::reluWeightStartAddr;
53+ using ConvCommon<CONV, CONV_TILING>::outputStartAddr;
52 54 
53 uint64_t fmapOneGroupSize = 0;55 uint64_t fmapOneGroupSize = 0;
54 uint64_t weightOneCoSize = 0;56 uint64_t weightOneCoSize = 0;
@@ -62,28 +64,28 @@ private:
62 uint64_t dhwIn = 0;64 uint64_t dhwIn = 0;
63};65};
64 66 
65-template <class CONV, class RUN_INFO>67+template <class CONV, class CONV_TILING>
66-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::68+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
67 CalcStartAddrFmapHW()69 CalcStartAddrFmapHW()
68{70{
69 if constexpr (!CONV::isMMode) {71 if constexpr (!CONV::isMMode) {
70- int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convRunInfo->strideH) -72+ int64_t hiStartPosTmp = static_cast<int64_t>(convOps->hoIdxStart * convTilingData->convRunInfo.strideH) -
71- static_cast<int64_t>(convRunInfo->padTop);73+ static_cast<int64_t>(convTilingData->convRunInfo.padTop);
72 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;74 convOps->singleCoreHiStartPos = hiStartPosTmp < 0 ? 0 : hiStartPosTmp;
73 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {75 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {
74- fmStartAddr += convOps->singleCoreHiStartPos * convRunInfo->win * convRunInfo->cin;76+ fmStartAddr += convOps->singleCoreHiStartPos * convTilingData->convRunInfo.win * convTilingData->convRunInfo.cin;
75 } else {77 } else {
76- fmStartAddr += convOps->singleCoreHiStartPos * convRunInfo->win;78+ fmStartAddr += convOps->singleCoreHiStartPos * convTilingData->convRunInfo.win;
77 }79 }
78 80
79 // Conv2D Split W81 // Conv2D Split W
80 convOps->singleCoreHiStartPos = hiStartPosTmp;82 convOps->singleCoreHiStartPos = hiStartPosTmp;
81 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {83 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {
82- int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convRunInfo->strideW) -84+ int64_t wiStartPosTmp = static_cast<int64_t>(convOps->woIdxStart * convTilingData->convRunInfo.strideW) -
83- static_cast<int64_t>(convRunInfo->padLeft);85+ static_cast<int64_t>(convTilingData->convRunInfo.padLeft);
84 convOps->singleCoreWiStartPos = wiStartPosTmp < 0 ? 0 : wiStartPosTmp;86 convOps->singleCoreWiStartPos = wiStartPosTmp < 0 ? 0 : wiStartPosTmp;
85 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {87 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {
86- fmStartAddr += convOps->singleCoreWiStartPos * convRunInfo->cin;88+ fmStartAddr += convOps->singleCoreWiStartPos * convTilingData->convRunInfo.cin;
87 } else {89 } else {
88 fmStartAddr += convOps->singleCoreWiStartPos;90 fmStartAddr += convOps->singleCoreWiStartPos;
89 }91 }
@@ -92,8 +94,8 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
92 }94 }
93}95}
94 96 
95-template <class CONV, class RUN_INFO>97+template <class CONV, class CONV_TILING>
96-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::98+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
97 CalcStartAddrOriGroupHWC(const uint64_t doIdxStart, const int64_t diIdxStart)99 CalcStartAddrOriGroupHWC(const uint64_t doIdxStart, const int64_t diIdxStart)
98{100{
99 // fmap: Batch -> Di -> HiWi -> Group -> Ci_perg101 // fmap: Batch -> Di -> HiWi -> Group -> Ci_perg
@@ -107,21 +109,21 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
107 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +109 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
108 convOps->groupIdxStart * convOps->coPerGroup + convOps->nIdxStart;110 convOps->groupIdxStart * convOps->coPerGroup + convOps->nIdxStart;
109 if constexpr (CONV::isMMode) {111 if constexpr (CONV::isMMode) {
110- outputStartAddr += convOps->mIdxStart * convRunInfo->cout;112+ outputStartAddr += convOps->mIdxStart * convTilingData->convRunInfo.cout;
111 } else {113 } else {
112- outputStartAddr += convOps->hoIdxStart * convRunInfo->wout * convRunInfo->cout;114+ outputStartAddr += convOps->hoIdxStart * convTilingData->convRunInfo.wout * convTilingData->convRunInfo.cout;
113 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {115 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {
114- outputStartAddr += convOps->woIdxStart * convRunInfo->cout;116+ outputStartAddr += convOps->woIdxStart * convTilingData->convRunInfo.cout;
115 }117 }
116 }118 }
117 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {119 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
118- fmStartAddr += diIdxStart * hwIn * convRunInfo->cin;120+ fmStartAddr += diIdxStart * hwIn * convTilingData->convRunInfo.cin;
119- outputStartAddr += doIdxStart * hwOut * convRunInfo->cout;121+ outputStartAddr += doIdxStart * hwOut * convTilingData->convRunInfo.cout;
120 }122 }
121}123}
122 124 
123-template <class CONV, class RUN_INFO>125+template <class CONV, class CONV_TILING>
124-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::126+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
125 CalcStartAddrOriGroupCHW(const uint64_t doIdxStart, const int64_t diIdxStart)127 CalcStartAddrOriGroupCHW(const uint64_t doIdxStart, const int64_t diIdxStart)
126{128{
127 // fmap: Batch -> Group -> Ci_perg -> Di -> HiWi129 // fmap: Batch -> Group -> Ci_perg -> Di -> HiWi
@@ -138,7 +140,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
138 if constexpr (CONV::isMMode) {140 if constexpr (CONV::isMMode) {
139 outputStartAddr += convOps->mIdxStart;141 outputStartAddr += convOps->mIdxStart;
140 } else {142 } else {
141- outputStartAddr += convOps->hoIdxStart * convRunInfo->wout;143+ outputStartAddr += convOps->hoIdxStart * convTilingData->convRunInfo.wout;
142 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {144 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {
143 outputStartAddr += convOps->woIdxStart;145 outputStartAddr += convOps->woIdxStart;
144 }146 }
@@ -150,17 +152,17 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
150 }152 }
151}153}
152 154 
153-template <class CONV, class RUN_INFO>155+template <class CONV, class CONV_TILING>
154-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::156+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
155 CalcStartAddrOriGroup(const uint64_t din, const uint64_t dout, const uint32_t kd, const uint64_t doIdxStart,157 CalcStartAddrOriGroup(const uint64_t din, const uint64_t dout, const uint32_t kd, const uint64_t doIdxStart,
156 const int64_t diIdxStart)158 const int64_t diIdxStart)
157{159{
158- hwIn = convRunInfo->hin * convRunInfo->win;160+ hwIn = convTilingData->convRunInfo.hin * convTilingData->convRunInfo.win;
159- hwOut = convRunInfo->hout * convRunInfo->wout;161+ hwOut = convTilingData->convRunInfo.hout * convTilingData->convRunInfo.wout;
160 dhwOut = dout * hwOut;162 dhwOut = dout * hwOut;
161 dhwIn = din * hwIn;163 dhwIn = din * hwIn;
162- convOps->fmapOneBatchSize = dhwIn * convRunInfo->cin;164+ convOps->fmapOneBatchSize = dhwIn * convTilingData->convRunInfo.cin;
163- convOps->outputOneBatchSize = dhwOut * convRunInfo->cout;165+ convOps->outputOneBatchSize = dhwOut * convTilingData->convRunInfo.cout;
164 166 
165 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {167 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {
166 fmapOneGroupSize = convOps->ciPerGroup;168 fmapOneGroupSize = convOps->ciPerGroup;
@@ -169,13 +171,13 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
169 CalcStartAddrOriGroupHWC(doIdxStart, diIdxStart);171 CalcStartAddrOriGroupHWC(doIdxStart, diIdxStart);
170 } else {172 } else {
171 fmapOneGroupSize = convOps->ciPerGroup * dhwIn;173 fmapOneGroupSize = convOps->ciPerGroup * dhwIn;
172- weightOneCoSize = convOps->ciPerGroup * kd * convRunInfo->kh * convRunInfo->kw;174+ weightOneCoSize = convOps->ciPerGroup * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;
173 weightOneGroupSize = convOps->coPerGroup * weightOneCoSize;175 weightOneGroupSize = convOps->coPerGroup * weightOneCoSize;
174 outputOneGroupSize = convOps->coPerGroup * dhwOut;176 outputOneGroupSize = convOps->coPerGroup * dhwOut;
175 CalcStartAddrOriGroupCHW(doIdxStart, diIdxStart);177 CalcStartAddrOriGroupCHW(doIdxStart, diIdxStart);
176 }178 }
177 179 
178- if (convRunInfo->hasBias) {180+ if (convTilingData->convRunInfo.hasBias) {
179 biasStartAddr = convOps->groupIdxStart * convOps->coPerGroup + convOps->nIdxStart;181 biasStartAddr = convOps->groupIdxStart * convOps->coPerGroup + convOps->nIdxStart;
180 }182 }
181 183 
@@ -186,34 +188,38 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
186 }188 }
187}189}
188 190 
189-template <class CONV, class RUN_INFO>191+template <class CONV, class CONV_TILING>
190-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::192+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
191 DealFixpiepParams(uint64_t groupIter, uint64_t coPerGroup)193 DealFixpiepParams(uint64_t groupIter, uint64_t coPerGroup)
192{194{
193 if constexpr (CONV::isQuant || CONV::IS_EXTEND_CONV2D) {195 if constexpr (CONV::isQuant || CONV::IS_EXTEND_CONV2D) {
194 if (hasScale) {196 if (hasScale) {
195 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {197 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW || CONV::A_FORMAT == ConvFormat::NHWC) {
196 Extendconv2dFixpipeParams fixpipeParamsCopy(convOps->fixpipeParams);198 Extendconv2dFixpipeParams fixpipeParamsCopy(convOps->fixpipeParams);
197- if (convOps->conv2dApiTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {199+ if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
198 fixpipeParamsCopy.scale0 = convOps->fixpipeParams.scale0[groupIter * coPerGroup];200 fixpipeParamsCopy.scale0 = convOps->fixpipeParams.scale0[groupIter * coPerGroup];
199 }201 }
200 if constexpr (CONV::IS_EXTEND_CONV2D) {202 if constexpr (CONV::IS_EXTEND_CONV2D) {
201- if (convOps->dualOutput) {203+ if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
202- if (convOps->conv2dApiTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {204+ fixpipeParamsCopy.reluWeight0 = convOps->fixpipeParams.reluWeight0[groupIter * coPerGroup];
205+ }
206+ if (convTilingData->convApiTiling.dualOutput) {
207+ if (convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
203 fixpipeParamsCopy.scale1 = convOps->fixpipeParams.scale1[groupIter * coPerGroup];208 fixpipeParamsCopy.scale1 = convOps->fixpipeParams.scale1[groupIter * coPerGroup];
204 }209 }
205- }210+ if (convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
206- }211+ fixpipeParamsCopy.reluWeight1 = convOps->fixpipeParams.reluWeight1[groupIter * coPerGroup];
207- if constexpr (CONV::IS_EXTEND_CONV2D) {212+ }
208- if (convOps->dualOutput) {213+ if (convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
209- if (convOps->conv2dApiTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||214+ convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
210- convOps->conv2dApiTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {215+ convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
216+ convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
211 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);217 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);
212 }218 }
213- } else if (convOps->conv2dApiTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {219+ } else if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
214 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);220 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);
215 }221 }
216- } else if (convOps->conv2dApiTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {222+ } else if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
217 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);223 convOps->conv.SetFixpipeParams(fixpipeParamsCopy);
218 }224 }
219 } else {225 } else {
@@ -223,21 +229,21 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
223 }229 }
224}230}
225 231 
226-template <class CONV, class RUN_INFO>232+template <class CONV, class CONV_TILING>
227-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::233+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
228 ConvKernelImplOriGroup()234 ConvKernelImplOriGroup()
229{235{
230 for (uint64_t groupIter = 0; groupIter < convOps->singleGroups; ++groupIter) {236 for (uint64_t groupIter = 0; groupIter < convOps->singleGroups; ++groupIter) {
231 convOps->conv.SetWeight(convOps->filterGm[groupIter * weightOneGroupSize]);237 convOps->conv.SetWeight(convOps->filterGm[groupIter * weightOneGroupSize]);
232 convOps->conv.SetFmap(convOps->fmapGm[groupIter * fmapOneGroupSize]);238 convOps->conv.SetFmap(convOps->fmapGm[groupIter * fmapOneGroupSize]);
233- if (convRunInfo->hasBias) {239+ if (convTilingData->convRunInfo.hasBias) {
234 convOps->conv.SetBias(convOps->biasGm[groupIter * convOps->coPerGroup]);240 convOps->conv.SetBias(convOps->biasGm[groupIter * convOps->coPerGroup]);
235 }241 }
236 242 
237 DealFixpiepParams(groupIter, convOps->coPerGroup);243 DealFixpiepParams(groupIter, convOps->coPerGroup);
238 244 
239 if constexpr (CONV::IS_EXTEND_CONV2D) {245 if constexpr (CONV::IS_EXTEND_CONV2D) {
240- if (convOps->dualOutput) {246+ if (this->convTilingData->convApiTiling.dualOutput) {
241 convOps->conv.IterateAll(convOps->outputGm[groupIter * outputOneGroupSize],247 convOps->conv.IterateAll(convOps->outputGm[groupIter * outputOneGroupSize],
242 convOps->output1Gm[groupIter * outputOneGroupSize]);248 convOps->output1Gm[groupIter * outputOneGroupSize]);
243 } else {249 } else {
@@ -250,13 +256,14 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
250 }256 }
251}257}
252 258 
253-template <class CONV, class RUN_INFO>259+template <class CONV, class CONV_TILING>
254-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::260+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
255 UpdateRealCoutOptGroup()261 UpdateRealCoutOptGroup()
256{262{
257 convOps->singleCoreN = convOps->singleCoOpt;263 convOps->singleCoreN = convOps->singleCoOpt;
258 264 
259- enlargeTail = convRunInfo->groups % convRunInfo->enlarge;265+ // update singleGroups : real groups nums in singlecore
266+ enlargeTail = convTilingData->convRunInfo.groups % convTilingData->convRunInfo.enlarge;
260 267 
261 if (unlikely(convOps->isGroupDimTail && enlargeTail != 0)) {268 if (unlikely(convOps->isGroupDimTail && enlargeTail != 0)) {
262 uint64_t realCout = enlargeTail * convOps->coPerGroup;269 uint64_t realCout = enlargeTail * convOps->coPerGroup;
@@ -270,12 +277,12 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
270 }277 }
271 278 
272 if (unlikely(convOps->isGroupDimTail)) {279 if (unlikely(convOps->isGroupDimTail)) {
273- convOps->singleGroups = enlargeTail == 0 ? convRunInfo->enlarge : enlargeTail;280+ convOps->singleGroups = enlargeTail == 0 ? convTilingData->convRunInfo.enlarge : enlargeTail;
274 }281 }
275}282}
276 283 
277-template <class CONV, class RUN_INFO>284+template <class CONV, class CONV_TILING>
278-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::285+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
279 CalcStartAddrOptGroupCHW(const uint64_t doIdxStart, const int64_t diIdxStart, const uint32_t kd)286 CalcStartAddrOptGroupCHW(const uint64_t doIdxStart, const int64_t diIdxStart, const uint32_t kd)
280{287{
281 // fmap: Batch -> Group(enlarge) -> Ci_opt -> Di -> HiWi288 // fmap: Batch -> Group(enlarge) -> Ci_opt -> Di -> HiWi
@@ -285,10 +292,10 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
285 convOps->groupIdxStart * fmapOneGroupSize;292 convOps->groupIdxStart * fmapOneGroupSize;
286 CalcStartAddrFmapHW();293 CalcStartAddrFmapHW();
287 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z) {294 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z) {
288- weightOneGroupSize = convRunInfo->coutOpt * convRunInfo->cinOpt * kd * convRunInfo->kh * convRunInfo->kw;295+ weightOneGroupSize = convTilingData->convRunInfo.coutOpt * convTilingData->convRunInfo.cinOpt * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;
289 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize + convOps->nIdxStart * convOps->k0;296 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize + convOps->nIdxStart * convOps->k0;
290 } else {297 } else {
291- weightOneGroupSize = convRunInfo->coutOpt * convOps->ciPerGroup * kd * convRunInfo->kh * convRunInfo->kw;298+ weightOneGroupSize = convTilingData->convRunInfo.coutOpt * convOps->ciPerGroup * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;
292 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize;299 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize;
293 }300 }
294 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +301 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
@@ -303,90 +310,94 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
303 if constexpr (CONV::isMMode) {310 if constexpr (CONV::isMMode) {
304 outputStartAddr += convOps->mIdxStart;311 outputStartAddr += convOps->mIdxStart;
305 } else {312 } else {
306- outputStartAddr += convOps->hoIdxStart * convRunInfo->wout;313+ outputStartAddr += convOps->hoIdxStart * convTilingData->convRunInfo.wout;
307 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {314 if constexpr (CONV::A_FORMAT == ConvFormat::NCHW) {
308 outputStartAddr += convOps->woIdxStart;315 outputStartAddr += convOps->woIdxStart;
309 }316 }
310 }317 }
311}318}
312 319 
313-template <class CONV, class RUN_INFO>320+template <class CONV, class CONV_TILING>
314-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::321+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
315 CalcStartAddrOptGroupHWC(const uint64_t doIdxStart, const int64_t diIdxStart, const uint32_t kd)322 CalcStartAddrOptGroupHWC(const uint64_t doIdxStart, const int64_t diIdxStart, const uint32_t kd)
316{323{
317 // fmap: Batch -> Di -> HiWi -> Group(enlarge) -> Ci_opt324 // fmap: Batch -> Di -> HiWi -> Group(enlarge) -> Ci_opt
318 // weight: Kd -> KhKw -> Ci_opt -> Group(enlarge) -> Co_opt325 // weight: Kd -> KhKw -> Ci_opt -> Group(enlarge) -> Co_opt
319 // output: Batch -> Do -> HoWo -> Group(enlarge) -> Co_opt326 // output: Batch -> Do -> HoWo -> Group(enlarge) -> Co_opt
320 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize +327 fmStartAddr = convOps->batchIdxStart * convOps->fmapOneBatchSize +
321- convOps->groupIdxStart * convRunInfo->cinOpt;328+ convOps->groupIdxStart * convTilingData->convRunInfo.cinOpt;
322 CalcStartAddrFmapHW();329 CalcStartAddrFmapHW();
323 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z) {330 if constexpr (CONV::B_FORMAT == ConvFormat::FRACTAL_Z) {
324- weightOneGroupSize = convRunInfo->coutOpt * convRunInfo->cinOpt * kd * convRunInfo->kh * convRunInfo->kw;331+ weightOneGroupSize = convTilingData->convRunInfo.coutOpt * convTilingData->convRunInfo.cinOpt * kd * convTilingData->convRunInfo.kh * convTilingData->convRunInfo.kw;
325 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize + convOps->nIdxStart * convOps->k0;332 weightStartAddr = convOps->groupIdxStart * weightOneGroupSize + convOps->nIdxStart * convOps->k0;
326 } else {333 } else {
327- weightOneGroupSize = convRunInfo->coutOpt;334+ weightOneGroupSize = convTilingData->convRunInfo.coutOpt;
328- weightStartAddr = convOps->groupIdxStart * convRunInfo->coutOpt;335+ weightStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt;
329 }336 }
330 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +337 outputStartAddr = convOps->batchIdxStart * convOps->outputOneBatchSize +
331- convOps->groupIdxStart * convRunInfo->coutOpt +338+ convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt +
332 convOps->nIdxStart;339 convOps->nIdxStart;
333 340 
334 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {341 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC) {
335- fmStartAddr += diIdxStart * hwIn * convRunInfo->cin;342+ fmStartAddr += diIdxStart * hwIn * convTilingData->convRunInfo.cin;
336- outputStartAddr += doIdxStart * hwOut * convRunInfo->cout;343+ outputStartAddr += doIdxStart * hwOut * convTilingData->convRunInfo.cout;
337 }344 }
338 345 
339 if constexpr (CONV::isMMode) {346 if constexpr (CONV::isMMode) {
340- outputStartAddr += convOps->mIdxStart * convRunInfo->cout;347+ outputStartAddr += convOps->mIdxStart * convTilingData->convRunInfo.cout;
341 } else {348 } else {
342- outputStartAddr += convOps->hoIdxStart * convRunInfo->wout * convRunInfo->cout;349+ outputStartAddr += convOps->hoIdxStart * convTilingData->convRunInfo.wout * convTilingData->convRunInfo.cout;
343 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {350 if constexpr (CONV::A_FORMAT == ConvFormat::NHWC) {
344- outputStartAddr += convOps->woIdxStart * convRunInfo->cout;351+ outputStartAddr += convOps->woIdxStart * convTilingData->convRunInfo.cout;
345 }352 }
346 }353 }
347}354}
348 355 
349-template <class CONV, class RUN_INFO>356+template <class CONV, class CONV_TILING>
350-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::357+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
351 CalcStartAddrOptGroup(const uint64_t din, const uint64_t dout, const uint32_t kd, const uint64_t doIdxStart,358 CalcStartAddrOptGroup(const uint64_t din, const uint64_t dout, const uint32_t kd, const uint64_t doIdxStart,
352 const int64_t diIdxStart)359 const int64_t diIdxStart)
353{360{
354- hwIn = convRunInfo->hin * convRunInfo->win;361+ hwIn = convTilingData->convRunInfo.hin * convTilingData->convRunInfo.win;
355- hwOut = convRunInfo->hout * convRunInfo->wout;362+ hwOut = convTilingData->convRunInfo.hout * convTilingData->convRunInfo.wout;
356 dhwOut = dout * hwOut;363 dhwOut = dout * hwOut;
357 dhwIn = din * hwIn;364 dhwIn = din * hwIn;
358- convOps->fmapOneBatchSize = convRunInfo->cin * dhwIn;365+ convOps->fmapOneBatchSize = convTilingData->convRunInfo.cin * dhwIn;
359- convOps->outputOneBatchSize = convRunInfo->cout * dhwOut;366+ convOps->outputOneBatchSize = convTilingData->convRunInfo.cout * dhwOut;
360 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {367 if constexpr (CONV::A_FORMAT == ConvFormat::NDHWC || CONV::A_FORMAT == ConvFormat::NHWC) {
361- fmapOneGroupSize = convRunInfo->cinOpt;368+ fmapOneGroupSize = convTilingData->convRunInfo.cinOpt;
362- outputOneGroupSize = convRunInfo->coutOpt;369+ outputOneGroupSize = convTilingData->convRunInfo.coutOpt;
363 CalcStartAddrOptGroupHWC(doIdxStart, diIdxStart, kd);370 CalcStartAddrOptGroupHWC(doIdxStart, diIdxStart, kd);
364 } else {371 } else {
365- fmapOneGroupSize = convRunInfo->cinOpt * dhwIn;372+ fmapOneGroupSize = convTilingData->convRunInfo.cinOpt * dhwIn;
366- outputOneGroupSize = convRunInfo->coutOpt * dhwOut;373+ outputOneGroupSize = convTilingData->convRunInfo.coutOpt * dhwOut;
367 CalcStartAddrOptGroupCHW(doIdxStart, diIdxStart, kd);374 CalcStartAddrOptGroupCHW(doIdxStart, diIdxStart, kd);
368 }375 }
369 376 
370- if (convRunInfo->hasBias) {377+ if (convTilingData->convRunInfo.hasBias) {
371- biasStartAddr = convOps->groupIdxStart * convRunInfo->coutOpt + convOps->nIdxStart;378+ biasStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt + convOps->nIdxStart;
372 }379 }
373 380 
374 if constexpr (CONV::isQuant) {381 if constexpr (CONV::isQuant) {
375 if (hasScale) {382 if (hasScale) {
376- scaleStartAddr = convOps->groupIdxStart * convRunInfo->coutOpt + convOps->nIdxStart;383+ scaleStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt + convOps->nIdxStart;
377 }384 }
378 }385 }
379 386 
380 if constexpr (CONV::IS_EXTEND_CONV2D) {387 if constexpr (CONV::IS_EXTEND_CONV2D) {
381- if (convOps->conv2dApiTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||388+ if (convOps->convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT) ||
382- convOps->conv2dApiTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {389+ convOps->convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
383- scaleStartAddr = convOps->groupIdxStart * convRunInfo->coutOpt + convOps->nIdxStart;390+ scaleStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt + convOps->nIdxStart;
391+ }
392+ if (convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU) ||
393+ convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
394+ reluWeightStartAddr = convOps->groupIdxStart * convTilingData->convRunInfo.coutOpt + convOps->nIdxStart;
384 }395 }
385 }396 }
386}397}
387 398 
388-template <class CONV, class RUN_INFO>399+template <class CONV, class CONV_TILING>
389-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::400+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
390 ConvKernelImplOptGroup()401 ConvKernelImplOptGroup()
391{402{
392 convOps->conv.SetWeightStartPosition(convOps->nIdxStart);403 convOps->conv.SetWeightStartPosition(convOps->nIdxStart);
@@ -408,14 +419,14 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
408 convOps->conv.SetFmap(convOps->fmapGm[groupOptIter * fmapOneGroupSize]);419 convOps->conv.SetFmap(convOps->fmapGm[groupOptIter * fmapOneGroupSize]);
409 convOps->conv.SetWeight(convOps->filterGm[groupOptIter * weightOneGroupSize]);420 convOps->conv.SetWeight(convOps->filterGm[groupOptIter * weightOneGroupSize]);
410 421 
411- if (convRunInfo->hasBias) {422+ if (convTilingData->convRunInfo.hasBias) {
412- convOps->conv.SetBias(convOps->biasGm[groupOptIter * convRunInfo->coutOpt]);423+ convOps->conv.SetBias(convOps->biasGm[groupOptIter * convTilingData->convRunInfo.coutOpt]);
413 }424 }
414 425 
415- DealFixpiepParams(groupOptIter, convRunInfo->coutOpt);426+ DealFixpiepParams(groupOptIter, convTilingData->convRunInfo.coutOpt);
416 427 
417 if constexpr (CONV::IS_EXTEND_CONV2D) {428 if constexpr (CONV::IS_EXTEND_CONV2D) {
418- if (convOps->dualOutput) {429+ if (this->convTilingData->convApiTiling.dualOutput) {
419 convOps->conv.IterateAll(convOps->outputGm[groupOptIter * outputOneGroupSize],430 convOps->conv.IterateAll(convOps->outputGm[groupOptIter * outputOneGroupSize],
420 convOps->output1Gm[groupOptIter * outputOneGroupSize]);431 convOps->output1Gm[groupOptIter * outputOneGroupSize]);
421 } else {432 } else {
@@ -428,8 +439,8 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
428 }439 }
429}440}
430 441 
431-template <class CONV, class RUN_INFO>442+template <class CONV, class CONV_TILING>
432-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::443+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
433 ConvKernelImplOptGroupPreload()444 ConvKernelImplOptGroupPreload()
434{445{
435 uint64_t groupOptIter = 0;446 uint64_t groupOptIter = 0;
@@ -440,7 +451,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
440 }451 }
441 convOps->conv.SetOptGroupParams(enlargeTail, convOps->singleGroupOpt, convOps->singleCoOpt);452 convOps->conv.SetOptGroupParams(enlargeTail, convOps->singleGroupOpt, convOps->singleCoOpt);
442 } else {453 } else {
443- convOps->conv.SetOptGroupParams(convRunInfo->enlarge, convOps->singleGroupOpt, convOps->singleCoOpt);454+ convOps->conv.SetOptGroupParams(convTilingData->convRunInfo.enlarge, convOps->singleGroupOpt, convOps->singleCoOpt);
444 }455 }
445 456 
446 convOps->conv.SetIterIndex(groupOptIter);457 convOps->conv.SetIterIndex(groupOptIter);
@@ -448,12 +459,12 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
448 convOps->conv.SetFmap(convOps->fmapGm);459 convOps->conv.SetFmap(convOps->fmapGm);
449 convOps->conv.SetWeight(convOps->filterGm);460 convOps->conv.SetWeight(convOps->filterGm);
450 461 
451- if (convRunInfo->hasBias) {462+ if (convTilingData->convRunInfo.hasBias) {
452 convOps->conv.SetBias(convOps->biasGm);463 convOps->conv.SetBias(convOps->biasGm);
453 }464 }
454 465
455 // quant need, current not care466 // quant need, current not care
456- DealFixpiepParams(groupOptIter, convRunInfo->coutOpt);467+ DealFixpiepParams(groupOptIter, convTilingData->convRunInfo.coutOpt);
457 468 
458 if constexpr (CONV::IS_EXTEND_CONV2D) {469 if constexpr (CONV::IS_EXTEND_CONV2D) {
459 if (convOps->dualOutput) {470 if (convOps->dualOutput) {
@@ -467,8 +478,8 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
467 convOps->conv.End();478 convOps->conv.End();
468}479}
469 480 
470-template <class CONV, class RUN_INFO>481+template <class CONV, class CONV_TILING>
471-__aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::482+__aicore__ __forceinline__ void ConvGroupCommon<CONV, CONV_TILING>::
472 SetOptGroupTail()483 SetOptGroupTail()
473{484{
474 if (convOps->singleCoOpt != convOps->singleCoreN) {485 if (convOps->singleCoOpt != convOps->singleCoreN) {
@@ -478,7 +489,7 @@ __aicore__ __forceinline__ void ConvGroupCommon<CONV, RUN_INFO>::
478 convOps->singleCoreBatch);489 convOps->singleCoreBatch);
479 } else {490 } else {
480 convOps->conv.SetSingleOutputShape(convOps->singleCoOpt, convOps->singleCoreDout, convOps->singleCoreHo,491 convOps->conv.SetSingleOutputShape(convOps->singleCoOpt, convOps->singleCoreDout, convOps->singleCoreHo,
481- convRunInfo->wout, convOps->singleCoreBatch);492+ convTilingData->convRunInfo.wout, convOps->singleCoreBatch);
482 }493 }
483 } else {494 } else {
484 if constexpr (CONV::isMMode) {495 if constexpr (CONV::isMMode) {
Mconv/common/op_kernel/arch35/conv_instr_hw_mode_impl.h+126-106
@@ -32,24 +32,32 @@ 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.convTiling->cinAInCore :35+ channelSize_ = self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ? self_->ctx.convTilingData->convApiTiling.cinAInCore :
36- self_->ctx.convTiling->cinATailInCore;36+ self_->ctx.convTilingData->convApiTiling.cinATailInCore;
37 } else {37 } else {
38- channelSize_ = self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ? self_->ctx.convTiling->cinAInCore :38+ channelSize_ = self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ? self_->ctx.convTilingData->convApiTiling.cinAInCore :
39- self_->ctx.convTiling->cinATailInCore;39+ self_->ctx.convTilingData->convApiTiling.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.convTiling->kernelHxkernelW) % self_->ctx.convTiling->kL0;44+ c04KStepTail = (channelSize_ * self_->ctx.convTilingData->convApiTiling.kernelHxkernelW) % self_->ctx.convTilingData->convApiTiling.kL0;
45- c04KStepTail = c04KStepTail == 0 ? self_->ctx.convTiling->kL0 : c04KStepTail;45+ c04KStepTail = c04KStepTail == 0 ? self_->ctx.convTilingData->convApiTiling.kL0 : c04KStepTail;
46 } else {46 } else {
47 if constexpr (Intf::kPreLoadFlag) {47 if constexpr (Intf::kPreLoadFlag) {
48- channelSize_ = self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ?48+ if constexpr (Intf::isKL1NL0FullLoad) {
49- self_->ctx.convTiling->cinAInCore : AlignB(self_->ctx.convTiling->cinATailInCore, Intf::k0);49+ channelSize_ = AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0);
50+ } else {
51+ channelSize_ = self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ?
52+ self_->ctx.convTilingData->convApiTiling.cinAInCore : AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0);
53+ }
50 } else {54 } else {
51- channelSize_ = self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ?55+ if constexpr (Intf::isKL1NL0FullLoad) {
52- self_->ctx.convTiling->cinAInCore : AlignB(self_->ctx.convTiling->cinATailInCore, Intf::k0);56+ channelSize_ = AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0);
57+ } else {
58+ channelSize_ = self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ?
59+ self_->ctx.convTilingData->convApiTiling.cinAInCore : AlignB(self_->ctx.convTilingData->convApiTiling.cinATailInCore, Intf::k0);
60+ }
53 }61 }
54 }62 }
55 }63 }
@@ -60,8 +68,8 @@ public:
60 currentML0_ = m;68 currentML0_ = m;
61 if constexpr (!Intf::isDmaFlag) {69 if constexpr (!Intf::isDmaFlag) {
62 currentML0Align_ = mNotAlign;70 currentML0Align_ = mNotAlign;
63- mStartPt_ = self_->ctx.hoL0Iter * self_->ctx.currentWoL1 * self_->ctx.convTiling->hoL0 +71+ mStartPt_ = self_->ctx.hoL0Iter * self_->ctx.currentWoL1 * self_->ctx.convTilingData->convApiTiling.hoL0 +
64- self_->ctx.woL0Iter * self_->ctx.convTiling->woL0;72+ self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;
65 SetLoad3dRepeatParams();73 SetLoad3dRepeatParams();
66 }74 }
67 }75 }
@@ -74,14 +82,14 @@ public:
74 uint8_t repeatMode = 0;82 uint8_t repeatMode = 0;
75 mExtension_ = currentML0Align_;83 mExtension_ = currentML0Align_;
76 84 
77- if (self_->ctx.currentWoL0 == self_->ctx.convTiling->woL0 &&85+ if (self_->ctx.currentWoL0 == self_->ctx.convTilingData->convApiTiling.woL0 &&
78- self_->ctx.currentHoL0 == self_->ctx.convTiling->hoL0) {86+ self_->ctx.currentHoL0 == self_->ctx.convTilingData->convApiTiling.hoL0) {
79- dstStride = self_->ctx.convTiling->fmapKStride;87+ dstStride = self_->ctx.convTilingData->convApiTiling.fmapKStride;
80 } else {88 } else {
81 dstStride = currentML0_ / BLOCK_L0_M;89 dstStride = currentML0_ / BLOCK_L0_M;
82 }90 }
83 91 
84- if (unlikely(CeilDiv(self_->ctx.currentWoL1, self_->ctx.convTiling->woL0) > 1)) {92+ if (unlikely(CeilDiv(self_->ctx.currentWoL1, self_->ctx.convTilingData->convApiTiling.woL0) > 1)) {
85 mExtension_ = self_->ctx.currentWoL0;93 mExtension_ = self_->ctx.currentWoL0;
86 repeatStride = self_->ctx.currentWoL1 / BLOCK_L0_M;94 repeatStride = self_->ctx.currentWoL1 / BLOCK_L0_M;
87 repeatTime = self_->ctx.currentHoL0;95 repeatTime = self_->ctx.currentHoL0;
@@ -91,70 +99,63 @@ public:
91 SetLoadDataRepeatWithStride(repeatParams);99 SetLoadDataRepeatWithStride(repeatParams);
92 }100 }
93 101 
94- __aicore__ inline void LoadAL0(bool isFirst = true)102+ __aicore__ inline void SetFirst()
103+ {
104+ xm_.bf.mExtension_ = mExtension_ & MASK_16;
105+ xm_.bf.mStartPt_ = mStartPt_ & MASK_16;
106+ 
107+ xt_.n = static_cast<uint64_t>(self_->ctx.convTilingData->convApiTiling.unionDataXt);
108+ xt_.bf.channelSize = channelSize_;
109+ param_.SetConfig1(xt_.n);
110+ }
111+ 
112+ __aicore__ inline uint64_t GetC04KStepTail()
113+ {
114+ return c04KStepTail;
115+ }
116+ 
117+ __aicore__ inline void LoadAL0(const uint64_t &currentKL0, const uint64_t &posK, const uint64_t &kIter, const LocalTensor<typename Intf::FmapT> &al0)
95 {118 {
96 if ASCEND_IS_AIV {119 if ASCEND_IS_AIV {
97 return;120 return;
98 }121 }
99 if constexpr (Intf::isDmaFlag) {122 if constexpr (Intf::isDmaFlag) {
100- DmaLoad2DImpl();123+ DmaLoad2DImpl(kIter, al0);
101 return;124 return;
102 } else {125 } else {
103- uint64_t currentKL0 = 0;126+ xm_.bf.currentKL0 = currentKL0 & MASK_16;
104- if constexpr (Intf::c04Flag) {127+ xm_.bf.posK = posK & MASK_16;
105- currentKL0 = IsKL0Tail() ? c04KStepTail : self_->ctx.convTiling->kL0;128+ param_.SetConfig0(xm_.n);
106- } else {129+ 
107- if constexpr (Intf::k0 == Intf::k0FmapTail) {130+ LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(al0, self_->ctx.al1, param_);
108- currentKL0 = IsKL0Tail() ? self_->ctx.kL0Tail : self_->ctx.convTiling->kL0;
109- } else {
110- currentKL0 = IsKL0Tail() ? self_->ctx.kAL0Tail : self_->ctx.convTiling->kL0;
111- }
112- }
113- uint64_t posK = self_->ctx.kAL0Iter * self_->ctx.convTiling->kL0;
114- if (isFirst) {
115- xmtmp_ = ((mExtension_ & MASK_16) << MSTEP_OFFSET) | ((mStartPt_ & MASK_16) << POSM_OFFSET);
116- xt_ = ((static_cast<uint64_t>(self_->ctx.convTiling->strideW) & MASK_6) << 0) |
117- ((static_cast<uint64_t>(self_->ctx.convTiling->strideH) & MASK_6) << STRIDEH_OFFSET) |
118- ((static_cast<uint64_t>(self_->ctx.kernelW) & MASK_8) << KERNELW_OFFSET) |
119- ((static_cast<uint64_t>(self_->ctx.kernelW) & NINTH_BIT_MASK) << KERNELW_HIGHEST_BIT_OFFSET) |
120- ((static_cast<uint64_t>(self_->ctx.kernelH) & MASK_8) << KERNELH_OFFSET) |
121- ((static_cast<uint64_t>(self_->ctx.kernelH) & NINTH_BIT_MASK) << KERNELH_HIGHEST_BIT_OFFSET) |
122- ((static_cast<uint64_t>(self_->ctx.convTiling->dilationW) & MASK_8) << DILATIONW_OFFSET) |
123- ((static_cast<uint64_t>(self_->ctx.convTiling->dilationH) & MASK_8) << DILATIONH_OFFSET) |
124- ((static_cast<uint64_t>(channelSize_) & MASK_16) << CIN_OFFSET);
125- param_.SetConfig1(xt_);
126- }
127- xm_ = ((currentKL0 & MASK_16) << 0) | ((posK & MASK_16) << POSK_OFFSET) | xmtmp_;
128- param_.SetConfig0(xm_);
129- LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(self_->ctx.al0, self_->ctx.al1, param_);
130 }131 }
131 };132 };
132 133 
133private:134private:
134- __aicore__ inline bool IsKL0Tail()135+ __aicore__ inline bool IsKL0Tail(const uint64_t &kIter)
135 {136 {
136- return self_->ctx.kIter == self_->ctx.maxKL0Iter;137+ return kIter == self_->ctx.maxKL0Iter;
137 }138 }
138 139 
139- __aicore__ inline void DmaLoad2DImpl()140+ __aicore__ inline void DmaLoad2DImpl(const uint64_t &kIter, const LocalTensor<typename Intf::FmapT> &al0)
140 {141 {
141 uint64_t currentWoL1Align = AlignB(self_->ctx.currentWoL1, BLOCK_L0_M);142 uint64_t currentWoL1Align = AlignB(self_->ctx.currentWoL1, BLOCK_L0_M);
142- uint64_t currentKL0 = IsKL0Tail() ? self_->ctx.kL0Tail : self_->ctx.convTiling->kL0;143+ uint64_t currentKL0 = IsKL0Tail(kIter) ? self_->ctx.kL0Tail : self_->ctx.convTilingData->convApiTiling.kL0;
143 144
144 LoadData2DParamsV2 loadParams;145 LoadData2DParamsV2 loadParams;
145- loadParams.kStartPosition = self_->ctx.kAL0Iter * self_->ctx.convTiling->kL0 / Intf::k0;146+ loadParams.kStartPosition = (kIter % self_->ctx.multiKAL1) * self_->ctx.convTilingData->convApiTiling.kL0 / Intf::k0;
146 loadParams.mStep = CeilDiv(self_->ctx.currentWoL0, BLOCK_L0_M);147 loadParams.mStep = CeilDiv(self_->ctx.currentWoL0, BLOCK_L0_M);
147 loadParams.kStep = currentKL0 / Intf::k0;148 loadParams.kStep = currentKL0 / Intf::k0;
148 loadParams.srcStride = self_->ctx.currentHoL1 * currentWoL1Align / BLOCK_L0_M;149 loadParams.srcStride = self_->ctx.currentHoL1 * currentWoL1Align / BLOCK_L0_M;
149 loadParams.dstStride = currentML0_ / BLOCK_L0_M;150 loadParams.dstStride = currentML0_ / BLOCK_L0_M;
150 151
151- uint32_t mStartPosition = self_->ctx.hoL0Iter * self_->ctx.convTiling->hoL0 * currentWoL1Align +152+ uint32_t mStartPosition = self_->ctx.hoL0Iter * self_->ctx.convTilingData->convApiTiling.hoL0 * currentWoL1Align +
152- self_->ctx.woL0Iter * self_->ctx.convTiling->woL0;153+ self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;
153 uint32_t dstOffset = 0;154 uint32_t dstOffset = 0;
154 uint32_t dstOffsetStride = AlignB(self_->ctx.currentWoL0, BLOCK_L0_M) * Intf::k0;155 uint32_t dstOffsetStride = AlignB(self_->ctx.currentWoL0, BLOCK_L0_M) * Intf::k0;
155 for (uint16_t hoL0Idx = 0; hoL0Idx < self_->ctx.currentHoL0; ++ hoL0Idx) {156 for (uint16_t hoL0Idx = 0; hoL0Idx < self_->ctx.currentHoL0; ++ hoL0Idx) {
156 loadParams.mStartPosition = mStartPosition / BLOCK_L0_M;157 loadParams.mStartPosition = mStartPosition / BLOCK_L0_M;
157- LoadData<typename Intf::FmapT>(self_->ctx.al0[dstOffset], self_->ctx.al1, loadParams);158+ LoadData<typename Intf::FmapT>(al0[dstOffset], self_->ctx.al1, loadParams);
158 mStartPosition += currentWoL1Align;159 mStartPosition += currentWoL1Align;
159 dstOffset += dstOffsetStride;160 dstOffset += dstOffsetStride;
160 }161 }
@@ -168,10 +169,9 @@ private:
168 uint16_t mStartPt_ = 0;169 uint16_t mStartPt_ = 0;
169 uint16_t channelSize_ = 0;170 uint16_t channelSize_ = 0;
170 uint64_t c04KStepTail = 0;171 uint64_t c04KStepTail = 0;
171- uint64_t xm_ = 0;
172- uint64_t xt_ = 0;
173- uint64_t xmtmp_ = 0;
174 Load3DBitModeParam param_;172 Load3DBitModeParam param_;
173+ UnionDataXt xt_;
174+ UnionDataXm xm_;
175};175};
176 176 
177template <class Intf, typename OutputT, uint64_t FixpipeIdx = 0>177template <class Intf, typename OutputT, uint64_t FixpipeIdx = 0>
@@ -181,7 +181,7 @@ public:
181 __aicore__ inline void SetParams(Intf *self)181 __aicore__ inline void SetParams(Intf *self)
182 {182 {
183 self_ = self;183 self_ = self;
184- valueHoWo_ = self_->ctx.orgHo * self_->ctx.orgWo;184+ valueHoWo_ = self_->ctx.convTilingData->convApiTiling.orgHo * self_->ctx.convTilingData->convApiTiling.orgWo;
185 }185 }
186 186 
187 __aicore__ inline void SetMN(uint64_t m, uint64_t n)187 __aicore__ inline void SetMN(uint64_t m, uint64_t n)
@@ -198,7 +198,7 @@ public:
198 intriParams.params.dnNum = self_->ctx.currentHoL0;198 intriParams.params.dnNum = self_->ctx.currentHoL0;
199 intriParams.params.dstDnMatrixStride = self_->ctx.singleCoreWo;199 intriParams.params.dstDnMatrixStride = self_->ctx.singleCoreWo;
200 } else {200 } else {
201- if (likely(self_->ctx.convTiling->woL0 >= self_->ctx.singleCoreWo)) {201+ if (likely(self_->ctx.convTilingData->convApiTiling.woL0 >= self_->ctx.singleCoreWo)) {
202 intriParams.mSize = currentML0_;202 intriParams.mSize = currentML0_;
203 intriParams.params.srcNzMatrixStride = 0;203 intriParams.params.srcNzMatrixStride = 0;
204 intriParams.params.dnNum = 1;204 intriParams.params.dnNum = 1;
@@ -214,7 +214,7 @@ public:
214 intriParams.nSize = currentNL0_;214 intriParams.nSize = currentNL0_;
215 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);215 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);
216 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {216 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {
217- intriParams.dstStride = self_->ctx.orgDo * valueHoWo_;217+ intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgDo * valueHoWo_;
218 } else {218 } else {
219 intriParams.dstStride = valueHoWo_;219 intriParams.dstStride = valueHoWo_;
220 }220 }
@@ -228,8 +228,8 @@ public:
228 intriParams.mSize = self_->ctx.currentWoL0;228 intriParams.mSize = self_->ctx.currentWoL0;
229 intriParams.params.srcNdStride = AlignB(self_->ctx.currentWoL0, BLOCK_L0_M);229 intriParams.params.srcNdStride = AlignB(self_->ctx.currentWoL0, BLOCK_L0_M);
230 intriParams.params.ndNum = self_->ctx.currentHoL0;230 intriParams.params.ndNum = self_->ctx.currentHoL0;
231- intriParams.params.dstNdStride = self_->ctx.singleCoreWo * self_->ctx.orgCo;231+ intriParams.params.dstNdStride = self_->ctx.singleCoreWo * self_->ctx.convTilingData->convApiTiling.orgCo;
232- } else if (likely(self_->ctx.convTiling->woL0 >= self_->ctx.singleCoreWo)) {232+ } else if (likely(self_->ctx.convTilingData->convApiTiling.woL0 >= self_->ctx.singleCoreWo)) {
233 intriParams.mSize = currentML0_;233 intriParams.mSize = currentML0_;
234 intriParams.params.srcNdStride = 0;234 intriParams.params.srcNdStride = 0;
235 intriParams.params.ndNum = 1;235 intriParams.params.ndNum = 1;
@@ -238,10 +238,10 @@ public:
238 intriParams.mSize = self_->ctx.currentWoL0;238 intriParams.mSize = self_->ctx.currentWoL0;
239 intriParams.params.srcNdStride = self_->ctx.currentWoL0;239 intriParams.params.srcNdStride = self_->ctx.currentWoL0;
240 intriParams.params.ndNum = self_->ctx.currentHoL0;240 intriParams.params.ndNum = self_->ctx.currentHoL0;
241- intriParams.params.dstNdStride = self_->ctx.singleCoreWo * self_->ctx.orgCo;241+ intriParams.params.dstNdStride = self_->ctx.singleCoreWo * self_->ctx.convTilingData->convApiTiling.orgCo;
242 }242 }
243 243 
244- intriParams.dstStride = self_->ctx.orgCo;244+ intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgCo;
245 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);245 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);
246 intriParams.nSize = currentNL0_;246 intriParams.nSize = currentNL0_;
247 SetBaseParams<CO2Layout::ROW_MAJOR>(intriParams);247 SetBaseParams<CO2Layout::ROW_MAJOR>(intriParams);
@@ -260,20 +260,24 @@ public:
260 ubInfo->realWUb = unUsedML0 < ubInfo->mUb ? unUsedML0 : ubInfo->mUb;260 ubInfo->realWUb = unUsedML0 < ubInfo->mUb ? unUsedML0 : ubInfo->mUb;
261 ubInfo->realBatchUb = 1;261 ubInfo->realBatchUb = 1;
262 ubInfo->outBatchIdx = self_->ctx.batchIter + ubInfo->batchLoopIdx * ubInfo->batchUb;262 ubInfo->outBatchIdx = self_->ctx.batchIter + ubInfo->batchLoopIdx * ubInfo->batchUb;
263- ubInfo->outCIdx = self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +263+ if constexpr (Intf::isKL1NL0FullLoad) {
264- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0 + ubInfo->nLoopIdx * ubInfo->nUb;264+ ubInfo->outCIdx = ubInfo->nLoopIdx * ubInfo->nUb;
265- ubInfo->outHIdx = self_->ctx.hoAL1Iter * self_->ctx.convTiling->hoL1 +265+ } else {
266- self_->ctx.hoL0Iter * self_->ctx.convTiling->hoL0 + ubInfo->mLoopIdx * ubInfo->mUb / self_->ctx.convTiling->woL0;266+ ubInfo->outCIdx = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
267+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0 + ubInfo->nLoopIdx * ubInfo->nUb;
268+ }
269+ ubInfo->outHIdx = self_->ctx.hoAL1Iter * self_->ctx.convTilingData->convApiTiling.hoL1 +
270+ self_->ctx.hoL0Iter * self_->ctx.convTilingData->convApiTiling.hoL0 + ubInfo->mLoopIdx * ubInfo->mUb / self_->ctx.convTilingData->convApiTiling.woL0;
267 if (self_->ctx.woL1SmallTail == 0) {271 if (self_->ctx.woL1SmallTail == 0) {
268- ubInfo->outWIdx = self_->ctx.woAL1Iter * self_->ctx.convTiling->woL1 +272+ ubInfo->outWIdx = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 +
269- self_->ctx.woL0Iter * self_->ctx.convTiling->woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTiling->woL0;273+ self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->convApiTiling.woL0;
270 } else {274 } else {
271 if (self_->ctx.woAL1Iter == self_->ctx.maxWoL1Iter) {275 if (self_->ctx.woAL1Iter == self_->ctx.maxWoL1Iter) {
272- ubInfo->outWIdx = ((self_->ctx.woAL1Iter - 1) * self_->ctx.convTiling->woL1 + self_->ctx.woAL1Tail) +276+ ubInfo->outWIdx = ((self_->ctx.woAL1Iter - 1) * self_->ctx.convTilingData->convApiTiling.woL1 + self_->ctx.woAL1Tail) +
273- self_->ctx.woL0Iter * self_->ctx.convTiling->woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTiling->woL0;277+ self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->convApiTiling.woL0;
274 } else {278 } else {
275- ubInfo->outWIdx = self_->ctx.woAL1Iter * self_->ctx.convTiling->woL1 +279+ ubInfo->outWIdx = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 +
276- self_->ctx.woL0Iter * self_->ctx.convTiling->woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTiling->woL0;280+ self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0 + ubInfo->mLoopIdx * ubInfo->mUb % self_->ctx.convTilingData->convApiTiling.woL0;
277 }281 }
278 }282 }
279 283 
@@ -299,25 +303,31 @@ public:
299 __aicore__ inline void SetBaseParams(FixpipeParamsC310<format> &intriParams)303 __aicore__ inline void SetBaseParams(FixpipeParamsC310<format> &intriParams)
300 {304 {
301 intriParams.quantPre = GetQuantPre<Intf, OutputT, FixpipeIdx>(self_);305 intriParams.quantPre = GetQuantPre<Intf, OutputT, FixpipeIdx>(self_);
302- if (self_->ctx.convTiling->hasScale == 0) {306+ if (self_->ctx.convTilingData->convApiTiling.hasScale == 0) {
303 intriParams.deqScalar = DEQ_SCALAR_ONE;307 intriParams.deqScalar = DEQ_SCALAR_ONE;
304 }308 }
305 if constexpr (Intf::isExtendConv2d) {309 if constexpr (Intf::isExtendConv2d) {
306 if constexpr (FixpipeIdx == 0) {310 if constexpr (FixpipeIdx == 0) {
307- intriParams.reluEn = self_->ctx.convTiling->reluMode0 != 0;311+ intriParams.reluEn = self_->ctx.convTilingData->convApiTiling.reluMode0 != 0;
308#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)312#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
309- intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTiling->reluMode0);313+ intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->convApiTiling.reluMode0);
310- if (self_->ctx.convTiling->reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {314+ if (self_->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
311 intriParams.reluScalar = self_->ctx.preReluScalar0;315 intriParams.reluScalar = self_->ctx.preReluScalar0;
316+ } else if (self_->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
317+ intriParams.vectorRelu =
318+ self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr()].GetPhyAddr();
312 }319 }
313#endif320#endif
314 intriParams.deqScalar = self_->ctx.deqScalar0;321 intriParams.deqScalar = self_->ctx.deqScalar0;
315 } else {322 } else {
316- intriParams.reluEn = self_->ctx.convTiling->reluMode1 != 0;323+ intriParams.reluEn = self_->ctx.convTilingData->convApiTiling.reluMode1 != 0;
317#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)324#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
318- intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTiling->reluMode1);325+ intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->convApiTiling.reluMode1);
319- if (self_->ctx.convTiling->reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {326+ if (self_->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
320 intriParams.reluScalar = self_->ctx.preReluScalar1;327 intriParams.reluScalar = self_->ctx.preReluScalar1;
328+ } else if(self_->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
329+ intriParams.vectorRelu =
330+ self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr() + self_->ctx.reluWeight1L1offset].GetPhyAddr();
321 }331 }
322#endif332#endif
323 intriParams.deqScalar = self_->ctx.deqScalar1;333 intriParams.deqScalar = self_->ctx.deqScalar1;
@@ -327,7 +337,7 @@ public:
327 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;337 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
328 } else {338 } else {
329 if constexpr (Intf::isExtendConv2d) {339 if constexpr (Intf::isExtendConv2d) {
330- if (self_->ctx.convTiling->dualOutput) {340+ if (self_->ctx.convTilingData->convApiTiling.dualOutput) {
331 intriParams.unitFlag = UNIT_FLAG_ENABLE_ONLY;341 intriParams.unitFlag = UNIT_FLAG_ENABLE_ONLY;
332 } else {342 } else {
333 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;343 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
@@ -341,38 +351,42 @@ public:
341 __aicore__ inline uint64_t CalcFixpipeOffset()351 __aicore__ inline uint64_t CalcFixpipeOffset()
342 {352 {
343 uint64_t offset = self_->ctx.batchIter * self_->ctx.outputOneBatchSize;353 uint64_t offset = self_->ctx.batchIter * self_->ctx.outputOneBatchSize;
344- uint64_t offsetH = self_->ctx.hoAL1Iter * self_->ctx.convTiling->hoL1 +354+ uint64_t offsetH = self_->ctx.hoAL1Iter * self_->ctx.convTilingData->convApiTiling.hoL1 +
345- self_->ctx.hoL0Iter * self_->ctx.convTiling->hoL0;355+ self_->ctx.hoL0Iter * self_->ctx.convTilingData->convApiTiling.hoL0;
346 uint64_t offsetW;356 uint64_t offsetW;
347 if (self_->ctx.woL1SmallTail == 0) {357 if (self_->ctx.woL1SmallTail == 0) {
348- offsetW = self_->ctx.woAL1Iter * self_->ctx.convTiling->woL1 +358+ offsetW = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 +
349- self_->ctx.woL0Iter * self_->ctx.convTiling->woL0;359+ self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;
350 } else {360 } else {
351 if (self_->ctx.woAL1Iter == self_->ctx.maxWoL1Iter) {361 if (self_->ctx.woAL1Iter == self_->ctx.maxWoL1Iter) {
352- offsetW = ((self_->ctx.woAL1Iter - 1) * self_->ctx.convTiling->woL1 + self_->ctx.woAL1Tail) +362+ offsetW = ((self_->ctx.woAL1Iter - 1) * self_->ctx.convTilingData->convApiTiling.woL1 + self_->ctx.woAL1Tail) +
353- self_->ctx.woL0Iter * self_->ctx.convTiling->woL0;363+ self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;
354 } else {364 } else {
355- offsetW = self_->ctx.woAL1Iter * self_->ctx.convTiling->woL1 +365+ offsetW = self_->ctx.woAL1Iter * self_->ctx.convTilingData->convApiTiling.woL1 +
356- self_->ctx.woL0Iter * self_->ctx.convTiling->woL0;366+ self_->ctx.woL0Iter * self_->ctx.convTilingData->convApiTiling.woL0;
357 }367 }
358 }368 }
359 369 
360- uint64_t offsetCout = self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +370+ uint64_t offsetCout = 0U;
361- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0;371+ if constexpr (!Intf::isKL1NL0FullLoad) {
372+ offsetCout = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
373+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;
374+ }
362 if constexpr (Intf::groupOptPreloadFlag) {375 if constexpr (Intf::groupOptPreloadFlag) {
363- offsetCout += self_->ctx.groupOptIter * self_->ctx.convTiling->orgCo / self_->ctx.convTiling->groups *376+ offsetCout += self_->ctx.groupOptIter * self_->ctx.convTilingData->convApiTiling.orgCo /
364- self_->ctx.convTiling->enlarge;377+ self_->ctx.convTilingData->convApiTiling.groups *
378+ self_->ctx.convTilingData->convApiTiling.enlarge;
365 }379 }
366 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {380 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {
367- offset += offsetCout * self_->ctx.orgDo * valueHoWo_ + self_->ctx.dOutIter * valueHoWo_ +381+ offset += offsetCout * self_->ctx.convTilingData->convApiTiling.orgDo * valueHoWo_ + self_->ctx.dOutIter * valueHoWo_ +
368- offsetH * self_->ctx.orgWo + offsetW;382+ offsetH * self_->ctx.convTilingData->convApiTiling.orgWo + offsetW;
369 } else if constexpr (Intf::formatOutput == ConvFormat::NDHWC) {383 } else if constexpr (Intf::formatOutput == ConvFormat::NDHWC) {
370- offset += self_->ctx.dOutIter * valueHoWo_ * self_->ctx.orgCo + offsetH * self_->ctx.orgWo *384+ offset += self_->ctx.dOutIter * valueHoWo_ * self_->ctx.convTilingData->convApiTiling.orgCo + offsetH * self_->ctx.convTilingData->convApiTiling.orgWo *
371- self_->ctx.orgCo + offsetW * self_->ctx.orgCo + offsetCout;385+ self_->ctx.convTilingData->convApiTiling.orgCo + offsetW * self_->ctx.convTilingData->convApiTiling.orgCo + offsetCout;
372 } else if constexpr (Intf::formatOutput == ConvFormat::NCHW) {386 } else if constexpr (Intf::formatOutput == ConvFormat::NCHW) {
373- offset += offsetCout * valueHoWo_ + offsetH * self_->ctx.orgWo + offsetW;387+ offset += offsetCout * valueHoWo_ + offsetH * self_->ctx.convTilingData->convApiTiling.orgWo + offsetW;
374 } else {388 } else {
375- offset += offsetH * self_->ctx.orgWo * self_->ctx.orgCo + offsetW * self_->ctx.orgCo + offsetCout;389+ offset += offsetH * self_->ctx.convTilingData->convApiTiling.orgWo * self_->ctx.convTilingData->convApiTiling.orgCo + offsetW * self_->ctx.convTilingData->convApiTiling.orgCo + offsetCout;
376 }390 }
377 391 
378 return offset;392 return offset;
@@ -398,9 +412,12 @@ public:
398 412 
399 __aicore__ inline uint64_t GetExtendConv2dScaleL1Addr()413 __aicore__ inline uint64_t GetExtendConv2dScaleL1Addr()
400 {414 {
401- if (self_->ctx.convTiling->fixpParamsFullLoadFlag) {415+ if constexpr (Intf::isKL1NL0FullLoad) {
402- return self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +416+ return 0;
403- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0;417+ }
418+ if (self_->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
419+ return self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
420+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;
404 }421 }
405 return 0;422 return 0;
406 }423 }
@@ -428,7 +445,7 @@ public:
428 if constexpr (Intf::isExtendConv2d) {445 if constexpr (Intf::isExtendConv2d) {
429 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);446 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);
430 } else if constexpr (Intf::isQuantScene) {447 } else if constexpr (Intf::isQuantScene) {
431- if (self_->ctx.convTiling->hasScale != 0) {448+ if (self_->ctx.convTilingData->convApiTiling.hasScale != 0) {
432 Fixpipe<OutputT, typename Intf::L0cT, config>(449 Fixpipe<OutputT, typename Intf::L0cT, config>(
433 output[offset], self_->ctx.cl0, self_->ctx.scaleL1[GetScaleL1Addr()], intriParams);450 output[offset], self_->ctx.cl0, self_->ctx.scaleL1[GetScaleL1Addr()], intriParams);
434 } else {451 } else {
@@ -447,10 +464,13 @@ private:
447 464 
448 __aicore__ inline uint64_t GetScaleL1Addr()465 __aicore__ inline uint64_t GetScaleL1Addr()
449 {466 {
467+ if constexpr (Intf::isKL1NL0FullLoad) {
468+ return 0;
469+ }
450 if constexpr (Intf::isQuantScene) {470 if constexpr (Intf::isQuantScene) {
451- if (self_->ctx.convTiling->hasScale != 0 && self_->ctx.convTiling->fixpParamsFullLoadFlag) {471+ if (self_->ctx.convTilingData->convApiTiling.hasScale != 0 && self_->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
452- return self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +472+ return self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
453- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0;473+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;
454 }474 }
455 }475 }
456 return 0;476 return 0;
Mconv/common/op_kernel/arch35/conv_instr_impl.h+77-99
@@ -52,10 +52,13 @@ public:
52 __aicore__ inline void LoadChannelWiseL1(const LocalTensor<ChannelWiseT> &tensorL1,52 __aicore__ inline void LoadChannelWiseL1(const LocalTensor<ChannelWiseT> &tensorL1,
53 const GlobalTensor<ChannelWiseT> &tensorGm)53 const GlobalTensor<ChannelWiseT> &tensorGm)
54 {54 {
55- uint64_t tensorGmOffset = self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +55+ if constexpr (Intf::isKL1NL0FullLoad) {
56- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0;56+ LoadChannelWiseL1FullLoad(tensorL1, tensorGm, currentNL0_, 0);
57- 57+ } else {
58- LoadChannelWiseL1FullLoad(tensorL1, tensorGm, currentNL0_, tensorGmOffset);58+ uint64_t tensorGmOffset = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
59+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;
60+ LoadChannelWiseL1FullLoad(tensorL1, tensorGm, currentNL0_, tensorGmOffset);
61+ }
59 }62 }
60 63 
61private:64private:
@@ -86,12 +89,16 @@ public:
86 return;89 return;
87 }90 }
88 uint32_t offset = 0;91 uint32_t offset = 0;
89- if (self_->ctx.convTiling->biasFullLoadFlag) {92+
90- offset = self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +93+ if (self_->ctx.convTilingData->convApiTiling.biasFullLoadFlag) {
91- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0;94+ if constexpr (!Intf::isKL1NL0FullLoad) {
95+ offset += self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
96+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;
97+ }
92 if constexpr (Intf::groupOptPreloadFlag) {98 if constexpr (Intf::groupOptPreloadFlag) {
93- offset += self_->ctx.groupOptIter * self_->ctx.convTiling->orgCo / self_->ctx.convTiling->groups *99+ offset += self_->ctx.groupOptIter * self_->ctx.convTilingData->convApiTiling.orgCo /
94- self_->ctx.convTiling->enlarge;100+ self_->ctx.convTilingData->convApiTiling.groups *
101+ self_->ctx.convTilingData->convApiTiling.enlarge;
95 }102 }
96 }103 }
97 104 
@@ -114,41 +121,59 @@ public:
114 __aicore__ inline void SetParams(Intf *self)121 __aicore__ inline void SetParams(Intf *self)
115 {122 {
116 self_ = self;123 self_ = self;
117- nStep_ = self_->ctx.convTiling->nL0 / BLOCK_L0_N;124+ nStep_ = self_->ctx.convTilingData->convApiTiling.nL0 / BLOCK_L0_N;
118 }125 }
119 126 
120 __aicore__ inline void SetN(uint64_t n)127 __aicore__ inline void SetN(uint64_t n)
121 {128 {
122- ratioOfNToN0 = (self_->ctx.nBL1Iter == self_->ctx.maxNBL1Iter && self_->ctx.nL0Iter == self_->ctx.maxNL0Iter) ?129+ if constexpr (Intf::isKL1NL0FullLoad) {
123- n / BLOCK_L0_N : self_->ctx.convTiling->nStep;130+ ratioOfNToN0 = n / BLOCK_L0_N;
124- }
125- 
126- __aicore__ inline void LoadBL0(bool isFirst)
127- {
128- uint64_t kStep = (self_->ctx.kIter != self_->ctx.maxKL0Iter) ?
129- self_->ctx.convTiling->kStep : (self_->ctx.kL0Tail / Intf::k0);
130- if (unlikely(isFirst)) {
131- uint64_t mStartPosition = self_->ctx.nL0Iter * nStep_;
132- param_.SetMStartPosition(static_cast<uint32_t>(mStartPosition));
133- param_.SetKStartPosition(static_cast<uint32_t>(self_->ctx.kBL0Iter * self_->ctx.convTiling->kStep));
134- param_.SetMStep(static_cast<uint16_t>(ratioOfNToN0));
135- param_.SetKStep(static_cast<uint16_t>(kStep));
136- param_.SetSrcStride(static_cast<int32_t>(self_->ctx.convTiling->nL1DivBlockSize));
137- param_.SetDstStride(static_cast<uint16_t>(ratioOfNToN0));
138- param_.SetIfTranspose(false);
139 } else {131 } else {
140- param_.SetKStartPosition(static_cast<uint32_t>(self_->ctx.kBL0Iter * self_->ctx.convTiling->kStep));132+ ratioOfNToN0 = (self_->ctx.nBL1Iter == self_->ctx.maxNBL1Iter && self_->ctx.nL0Iter == self_->ctx.maxNL0Iter) ?
141- param_.SetKStep(static_cast<uint16_t>(kStep));133+ n / BLOCK_L0_N : self_->ctx.convTilingData->convApiTiling.nStep;
142 }134 }
143- LoadData<TPosition::B2, TPosition::B1, typename Intf::WeightT>(self_->ctx.bl0, self_->ctx.bl1, param_);
144 }135 }
145 136 
137+ __aicore__ inline void SetFirst()
138+ {
139+ if constexpr (Intf::isKL1NL0FullLoad) {
140+ param_.SetMStartPosition(static_cast<uint32_t>(0));
141+ } else {
142+ param_.SetMStartPosition(static_cast<uint32_t>(self_->ctx.nL0Iter * nStep_));
143+ }
144+ param_.SetMStep(static_cast<uint16_t>(ratioOfNToN0));
145+ param_.SetSrcStride(static_cast<int32_t>(self_->ctx.convTilingData->convApiTiling.nL1DivBlockSize));
146+ param_.SetDstStride(static_cast<uint16_t>(ratioOfNToN0));
147+ param_.SetIfTranspose(false);
148+ }
149+ 
150+ __aicore__ inline void LoadBL0(const uint64_t &KStartPosition, const uint64_t &kStep, const LocalTensor<typename Intf::WeightT> &bl0)
151+ {
152+ param_.SetKStartPosition(static_cast<uint32_t>(KStartPosition));
153+ param_.SetKStep(static_cast<uint16_t>(kStep));
154+ LoadData<TPosition::B2, TPosition::B1, typename Intf::WeightT>(bl0, self_->ctx.bl1, param_);
155+ }
156+ 
157+ __aicore__ inline void FullLoadBL0(const LocalTensor<typename Intf::WeightT> &bl0)
158+ {
159+ static uint32_t isLoaded = false;
160+ if (isLoaded) {
161+ return;
162+ }
163+ isLoaded = true;
164+ param_.SetMStartPosition(0);
165+ param_.SetKStartPosition(0);
166+ param_.SetMStep(static_cast<uint16_t>(self_->ctx.convTilingData->convApiTiling.nL0 / BLOCK_L0_N));
167+ param_.SetKStep(static_cast<uint16_t>(AlignB(self_->ctx.convTilingData->convApiTiling.kBL1, Intf::k0)/Intf::k0));
168+ uint16_t stride = static_cast<uint16_t>(self_->ctx.convTilingData->convApiTiling.nBL1 / BLOCK_L0_N);
169+ param_.SetSrcStride(stride);
170+ param_.SetDstStride(stride);
171+ param_.SetIfTranspose(false);
172+ LoadData<TPosition::B2, TPosition::B1, typename Intf::WeightT>(bl0, self_->ctx.bl1, param_);
173+ }
146private:174private:
147 Intf *self_ = nullptr;175 Intf *self_ = nullptr;
148 uint64_t ratioOfNToN0 = 0;176 uint64_t ratioOfNToN0 = 0;
149- uint64_t xm_ = 0;
150- uint64_t xt_ = 0;
151- uint64_t xmtmp_ = 0;
152 uint64_t nStep_ = 0;177 uint64_t nStep_ = 0;
153 Load2DBitModeParam param_;178 Load2DBitModeParam param_;
154};179};
@@ -164,80 +189,33 @@ public:
164 self_ = self;189 self_ = self;
165 }190 }
166 191 
167- __aicore__ inline void SetMN(uint64_t m, uint64_t n)192+ __aicore__ inline void Mad(const uint64_t &kIter, const LocalTensor<typename Intf::FmapT> &al0, const LocalTensor<typename Intf::WeightT> &bl0, const MmadParams &mmadParams)
168 {193 {
169- currentML0_ = m;
170- currentNL0_ = n;
171- }
172- 
173- __aicore__ inline void Mad()
174- {
175- MmadParams mmadParams;
176- SetMmadParams(mmadParams);
177- 
178 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {194 if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {
179 uint32_t srcOffset = 0;195 uint32_t srcOffset = 0;
180 uint32_t dstOffset = 0;196 uint32_t dstOffset = 0;
181- uint32_t srcBatchStride = self_->ctx.currentML0Align * self_->ctx.convTiling->kL0;197+ uint32_t srcBatchStride = self_->ctx.currentML0Align * self_->ctx.convTilingData->convApiTiling.kL0;
182 uint32_t dstBatchStride = self_->ctx.currentML0Align * self_->ctx.currentNL0Align;198 uint32_t dstBatchStride = self_->ctx.currentML0Align * self_->ctx.currentNL0Align;
183 for (uint32_t batchIdx = 0; batchIdx < self_->ctx.innerBatch; batchIdx++) {199 for (uint32_t batchIdx = 0; batchIdx < self_->ctx.innerBatch; batchIdx++) {
184 Mmad<typename Intf::L0cT, typename Intf::FmapT, typename Intf::WeightT>(200 Mmad<typename Intf::L0cT, typename Intf::FmapT, typename Intf::WeightT>(
185- self_->ctx.cl0[dstOffset], self_->ctx.al0[srcOffset], self_->ctx.bl0, mmadParams);201+ self_->ctx.cl0[dstOffset], al0[srcOffset], bl0, mmadParams);
186 srcOffset += srcBatchStride;202 srcOffset += srcBatchStride;
187 dstOffset += dstBatchStride;203 dstOffset += dstBatchStride;
188 }204 }
189 } else {205 } else {
190 Mmad<typename Intf::L0cT, typename Intf::FmapT, typename Intf::WeightT>(206 Mmad<typename Intf::L0cT, typename Intf::FmapT, typename Intf::WeightT>(
191- self_->ctx.cl0, self_->ctx.al0, self_->ctx.bl0, mmadParams);207+ self_->ctx.cl0, al0, bl0, mmadParams);
192 }208 }
193 }209 }
194 210 
195private:211private:
196- __aicore__ inline void SetMmadParams(MmadParams &mmadParams)212+ __aicore__ inline bool IsKL0Tail(const uint64_t &kIter)
197 {213 {
198- if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::KERNEL_1X1_MULTI_BATCH)) {214+ return kIter == self_->ctx.maxKL0Iter;
199- if constexpr (Intf::formatOutput == ConvFormat::NCHW && Intf::c04Flag) {
200- mmadParams.m = self_->ctx.innerBatch * AlignB(C04_CIN_SIZE * currentML0_, Intf::k0) / C04_CIN_SIZE;
201- } else {
202- mmadParams.m = self_->ctx.innerBatch * currentML0_;
203- }
204- } else {
205- mmadParams.m = currentML0_;
206- }
207- mmadParams.n = currentNL0_;
208- if constexpr (Intf::k0 == Intf::k0FmapTail) {
209- mmadParams.k = IsKL0Tail() ? self_->ctx.kL0Tail : self_->ctx.convTiling->kL0;
210- } else {
211- mmadParams.k = IsKL0Tail() ? self_->ctx.kAL0Tail : self_->ctx.convTiling->kL0;
212- }
213- 
214- if constexpr (!Intf::isDeQuantFlag) {
215- if (!self_->ctx.enableBias) {
216- mmadParams.cmatrixInitVal = self_->ctx.kIter == 0;
217- mmadParams.cmatrixSource = false;
218- } else {
219- mmadParams.cmatrixInitVal = false;
220- mmadParams.cmatrixSource = self_->ctx.kIter == 0;
221- }
222- } else {
223- mmadParams.cmatrixInitVal = self_->ctx.kIter == 0;
224- mmadParams.cmatrixSource = false;
225- }
226- 
227- if constexpr (!Intf::isInnerBatchFlag && !Intf::isDeQuantFlag) {
228- mmadParams.unitFlag = IsKL0Tail() ? UNIT_FLAG_ENABLE_WITH_FLIP : UNIT_FLAG_ENABLE_ONLY;
229- }
230- }
231- 
232- __aicore__ inline bool IsKL0Tail()
233- {
234- return self_->ctx.kIter == self_->ctx.maxKL0Iter;
235 }215 }
236 216 
237private:217private:
238 Intf *self_ = nullptr;218 Intf *self_ = nullptr;
239- uint64_t currentML0_ = 0;
240- uint64_t currentNL0_ = 0;
241};219};
242 220 
243template <class Intf, typename OutputT, uint64_t FixpipeIdx = 0>221template <class Intf, typename OutputT, uint64_t FixpipeIdx = 0>
@@ -247,7 +225,7 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
247 // extend conv2d: may be scalar or vector quant225 // extend conv2d: may be scalar or vector quant
248 if constexpr (AscendC::IsSameType<OutputT, float>::value) {226 if constexpr (AscendC::IsSameType<OutputT, float>::value) {
249 if constexpr (Intf::isExtendConv2d) {227 if constexpr (Intf::isExtendConv2d) {
250- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {228+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
251 return QuantMode_t::VQF322F32_PRE;229 return QuantMode_t::VQF322F32_PRE;
252 } else {230 } else {
253 return QuantMode_t::QF322F32_PRE;231 return QuantMode_t::QF322F32_PRE;
@@ -259,7 +237,7 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
259 237 
260 if constexpr (AscendC::IsSameType<OutputT, half>::value) {238 if constexpr (AscendC::IsSameType<OutputT, half>::value) {
261 if constexpr (Intf::isExtendConv2d) {239 if constexpr (Intf::isExtendConv2d) {
262- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {240+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
263 return QuantMode_t::VQF322F16_PRE;241 return QuantMode_t::VQF322F16_PRE;
264 } else {242 } else {
265 return QuantMode_t::QF322F16_PRE;243 return QuantMode_t::QF322F16_PRE;
@@ -271,7 +249,7 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
271 249 
272 if constexpr (AscendC::IsSameType<OutputT, bfloat16_t>::value) {250 if constexpr (AscendC::IsSameType<OutputT, bfloat16_t>::value) {
273 if constexpr (Intf::isExtendConv2d) {251 if constexpr (Intf::isExtendConv2d) {
274- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {252+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
275 return QuantMode_t::VQF322BF16_PRE;253 return QuantMode_t::VQF322BF16_PRE;
276 } else {254 } else {
277 return QuantMode_t::QF322BF16_PRE;255 return QuantMode_t::QF322BF16_PRE;
@@ -283,16 +261,16 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
283 261 
284 if constexpr (AscendC::IsSameType<OutputT, hifloat8_t>::value) {262 if constexpr (AscendC::IsSameType<OutputT, hifloat8_t>::value) {
285 if constexpr (Intf::isExtendConv2d) {263 if constexpr (Intf::isExtendConv2d) {
286- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {264+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
287 return QuantMode_t::VQF322HIF8_PRE;265 return QuantMode_t::VQF322HIF8_PRE;
288 } else {266 } else {
289 return QuantMode_t::QF322HIF8_PRE;267 return QuantMode_t::QF322HIF8_PRE;
290 }268 }
291 } else {269 } else {
292- if (self->ctx.convTiling->hasScale == 0) {270+ if (self->ctx.convTilingData->convApiTiling.hasScale == 0) {
293 // conv2d support hif8 in hif8 out271 // conv2d support hif8 in hif8 out
294 return QuantMode_t::QF322HIF8_PRE;272 return QuantMode_t::QF322HIF8_PRE;
295- } else if (self->ctx.convTiling->roundMode == ROUND_MODE_ROUND) {273+ } else if (self->ctx.convTilingData->convApiTiling.roundMode == ROUND_MODE_ROUND) {
296 // quantconv2d/quantconv3d274 // quantconv2d/quantconv3d
297 return QuantMode_t::VQF322HIF8_PRE;275 return QuantMode_t::VQF322HIF8_PRE;
298 }276 }
@@ -301,7 +279,7 @@ __aicore__ inline QuantMode_t GetQuantPreHif8Fp8(Intf *self)
301 279 
302 if constexpr (AscendC::IsSameType<OutputT, fp8_e4m3fn_t>::value) {280 if constexpr (AscendC::IsSameType<OutputT, fp8_e4m3fn_t>::value) {
303 if constexpr (Intf::isExtendConv2d) {281 if constexpr (Intf::isExtendConv2d) {
304- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {282+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
305 return QuantMode_t::VQF322FP8_PRE;283 return QuantMode_t::VQF322FP8_PRE;
306 } else {284 } else {
307 return QuantMode_t::QF322FP8_PRE;285 return QuantMode_t::QF322FP8_PRE;
@@ -321,13 +299,13 @@ __aicore__ inline QuantMode_t GetQuantPreInt32(Intf *self)
321 if constexpr (AscendC::IsSameType<OutputT, half>::value) {299 if constexpr (AscendC::IsSameType<OutputT, half>::value) {
322 if constexpr (Intf::isExtendConv2d) {300 if constexpr (Intf::isExtendConv2d) {
323 if constexpr (FixpipeIdx == 0) {301 if constexpr (FixpipeIdx == 0) {
324- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {302+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
325 return QuantMode_t::VDEQF16;303 return QuantMode_t::VDEQF16;
326 } else {304 } else {
327 return QuantMode_t::DEQF16;305 return QuantMode_t::DEQF16;
328 }306 }
329 } else {307 } else {
330- if (self->ctx.convTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {308+ if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
331 return QuantMode_t::VDEQF16;309 return QuantMode_t::VDEQF16;
332 } else {310 } else {
333 return QuantMode_t::DEQF16;311 return QuantMode_t::DEQF16;
@@ -342,13 +320,13 @@ __aicore__ inline QuantMode_t GetQuantPreInt32(Intf *self)
342 } else if constexpr (AscendC::IsSameType<OutputT, int8_t>::value) {320 } else if constexpr (AscendC::IsSameType<OutputT, int8_t>::value) {
343 if constexpr (Intf::isExtendConv2d) {321 if constexpr (Intf::isExtendConv2d) {
344 if constexpr (FixpipeIdx == 0) {322 if constexpr (FixpipeIdx == 0) {
345- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {323+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
346 return QuantMode_t::VREQ8;324 return QuantMode_t::VREQ8;
347 } else {325 } else {
348 return QuantMode_t::REQ8;326 return QuantMode_t::REQ8;
349 }327 }
350 } else {328 } else {
351- if (self->ctx.convTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {329+ if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
352 return QuantMode_t::VREQ8;330 return QuantMode_t::VREQ8;
353 } else {331 } else {
354 return QuantMode_t::REQ8;332 return QuantMode_t::REQ8;
@@ -378,13 +356,13 @@ __aicore__ inline QuantMode_t GetQuantPreFp32(Intf *self)
378 } else if constexpr (AscendC::IsSameType<OutputT, int8_t>::value) {356 } else if constexpr (AscendC::IsSameType<OutputT, int8_t>::value) {
379 if constexpr (Intf::isExtendConv2d) {357 if constexpr (Intf::isExtendConv2d) {
380 if constexpr (FixpipeIdx == 0) {358 if constexpr (FixpipeIdx == 0) {
381- if (self->ctx.convTiling->quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {359+ if (self->ctx.convTilingData->convApiTiling.quantMode0 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
382 return QuantMode_t::VQF322B8_PRE;360 return QuantMode_t::VQF322B8_PRE;
383 } else {361 } else {
384 return QuantMode_t::QF322B8_PRE;362 return QuantMode_t::QF322B8_PRE;
385 }363 }
386 } else {364 } else {
387- if (self->ctx.convTiling->quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {365+ if (self->ctx.convTilingData->convApiTiling.quantMode1 == static_cast<uint8_t>(QuantModeType::VECTOR_QUANT)) {
388 return QuantMode_t::VQF322B8_PRE;366 return QuantMode_t::VQF322B8_PRE;
389 } else {367 } else {
390 return QuantMode_t::QF322B8_PRE;368 return QuantMode_t::QF322B8_PRE;
Mconv/common/op_kernel/arch35/conv_instr_m_mode_impl.h+122-97
@@ -32,27 +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.convTiling->cinATailInCore, Intf::k0FmapTail);35+ alignCinATailInCore_ = AlignB(self_->ctx.convTilingData->convApiTiling.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 = (channelSize_ * self_->ctx.convTiling->kernelHxkernelW) % self_->ctx.convTiling->kL0;39+ c04KStepTail = (conv::C04_CIN_SIZE * self_->ctx.convTilingData->convApiTiling.kernelHxkernelW) % self_->ctx.convTilingData->convApiTiling.kL0;
40- c04KStepTail = c04KStepTail == 0 ? self_->ctx.convTiling->kL0 : c04KStepTail;40+ c04KStepTail = c04KStepTail == 0 ? self_->ctx.convTilingData->convApiTiling.kL0 : c04KStepTail;
41 } else {41 } else {
42 if constexpr (Intf::kPreLoadFlag) {42 if constexpr (Intf::kPreLoadFlag) {
43- channelSize_ = (self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ?43+ if constexpr (Intf::isKL1NL0FullLoad) {
44- self_->ctx.convTiling->cinAInCore : alignCinATailInCore_);44+ channelSize_ = alignCinATailInCore_;
45+ } else {
46+ channelSize_ = (self_->ctx.kAL1Iter - 1 != self_->ctx.maxKAL1Iter ?
47+ self_->ctx.convTilingData->convApiTiling.cinAInCore : alignCinATailInCore_);
48+ }
45 } else {49 } else {
46- channelSize_ = (self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ?50+ if constexpr (Intf::isKL1NL0FullLoad) {
47- self_->ctx.convTiling->cinAInCore : alignCinATailInCore_);51+ channelSize_ = alignCinATailInCore_;
52+ } else {
53+ channelSize_ = (self_->ctx.kAL1Iter != self_->ctx.maxKAL1Iter ?
54+ self_->ctx.convTilingData->convApiTiling.cinAInCore : alignCinATailInCore_);
55+ }
48 }56 }
49 }57 }
50 58 
51 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)) {
52- uint64_t hiLoadL1 = (self_->ctx.convTiling->orgHo - 1) * self_->ctx.convTiling->strideH +60+ uint64_t hiLoadL1 = (self_->ctx.convTilingData->convApiTiling.orgHo - 1) * self_->ctx.convTilingData->convApiTiling.strideH +
53 self_->ctx.dilatedKernelH;61 self_->ctx.dilatedKernelH;
54- hiLoadL1 = hiLoadL1 > self_->ctx.convTiling->orgHi ? self_->ctx.convTiling->orgHi : hiLoadL1;62+ hiLoadL1 = hiLoadL1 > self_->ctx.convTilingData->convApiTiling.orgHi ? self_->ctx.convTilingData->convApiTiling.orgHi : hiLoadL1;
55- realHixWi = hiLoadL1 * self_->ctx.convTiling->orgWi;63+ realHixWi = hiLoadL1 * self_->ctx.convTilingData->convApiTiling.orgWi;
56 }64 }
57 }65 }
58 66 
@@ -74,79 +82,73 @@ public:
74 SetLoadDataRepeatWithStride(repeatParams);82 SetLoadDataRepeatWithStride(repeatParams);
75 }83 }
76 84 
77- __aicore__ inline void LoadAL0(bool isFirst = true)85+ __aicore__ inline void SetFirst()
86+ {
87+ uint64_t posM = self_->ctx.mL0Iter * self_->ctx.mL0 +
88+ (self_->ctx.mStartPos + self_->ctx.mAL1Iter * self_->ctx.mAL1) % self_->ctx.convTilingData->convApiTiling.orgWo;
89+ xm_.bf.mExtension_ = currentML0_ & MASK_16;
90+ xm_.bf.mStartPt_ = posM & MASK_16;
91+ 
92+ xt_.n = static_cast<uint64_t>(self_->ctx.convTilingData->convApiTiling.unionDataXt);
93+ xt_.bf.channelSize = channelSize_;
94+ param_.SetConfig1(xt_.n);
95+ }
96+ 
97+ __aicore__ inline uint64_t GetC04KStepTail()
98+ {
99+ return c04KStepTail;
100+ }
101+ 
102+ __aicore__ inline void LoadAL0(const uint64_t &currentKL0, const uint64_t &posK, const uint64_t &kIter, const LocalTensor<typename Intf::FmapT> &al0)
78 {103 {
79 if ASCEND_IS_AIV {104 if ASCEND_IS_AIV {
80 return;105 return;
81 }106 }
82- if constexpr (Intf::c04Flag) {107+ xm_.bf.currentKL0 = currentKL0 & MASK_16;
83- currentKL0 = IsKL0Tail() ? c04KStepTail : self_->ctx.convTiling->kL0;108+ xm_.bf.posK = posK & MASK_16;
84- } else if constexpr (Intf::k0 == Intf::k0FmapTail) {109+ param_.SetConfig0(xm_.n);
85- currentKL0 = IsKL0Tail() ? self_->ctx.kL0Tail : self_->ctx.convTiling->kL0;
86- } else {
87- currentKL0 = IsKL0Tail() ? self_->ctx.kAL0Tail : self_->ctx.convTiling->kL0;
88- }
89 110 
90- uint64_t posK = self_->ctx.kAL0Iter * self_->ctx.convTiling->kL0;111+ if constexpr (Intf::preFusionFlag) {
91- if (isFirst) {112+ uint64_t aL1Offset = 0;
92- uint64_t posM = self_->ctx.mL0Iter * self_->ctx.mL0 +113+ if ((self_->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {
93- (self_->ctx.mStartPos + self_->ctx.mAL1Iter * self_->ctx.mAL1) % self_->ctx.orgWo;114+ aL1Offset = (self_->ctx.al1PingPongFlag^1) * self_->ctx.convTilingData->convApiTiling.aL1SpaceSize / Intf::sizeOfFmap;
94- xmtmp_ = ((currentML0_ & MASK_16) << MSTEP_OFFSET) | ((posM & MASK_16) << POSM_OFFSET);115+ }
95- xt_ = ((static_cast<uint64_t>(self_->ctx.convTiling->strideW) & MASK_6) << 0) |116+ LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(al0,
96- ((static_cast<uint64_t>(self_->ctx.kernelW) & MASK_8) << KERNELW_OFFSET) |117+ self_->ctx.al1[aL1Offset], param_);
97- ((static_cast<uint64_t>(self_->ctx.kernelH) & MASK_8) << KERNELH_OFFSET) |118+ } else if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {
98- ((static_cast<uint64_t>(self_->ctx.convTiling->strideH) & MASK_6) << STRIDEH_OFFSET) |
99- ((static_cast<uint64_t>(self_->ctx.kernelW) & NINTH_BIT_MASK) << KERNELW_HIGHEST_BIT_OFFSET) |
100- ((static_cast<uint64_t>(self_->ctx.kernelH) & NINTH_BIT_MASK) << KERNELH_HIGHEST_BIT_OFFSET) |
101- ((static_cast<uint64_t>(self_->ctx.convTiling->dilationW) & MASK_8) << DILATIONW_OFFSET) |
102- ((static_cast<uint64_t>(self_->ctx.convTiling->dilationH) & MASK_8) << DILATIONH_OFFSET) |
103- ((static_cast<uint64_t>(channelSize_) & MASK_16) << CIN_OFFSET);
104- param_.SetConfig1(xt_);
105- }
106- xm_ = currentKL0 | (posK << POSK_OFFSET) | xmtmp_;
107- param_.SetConfig0(xm_);
108- 
109- if constexpr (Intf::ConvParam::innerBatch == static_cast<int8_t>(ConvInnerBatch::MULTI_BATCH)) {
110 uint32_t srcOffset = 0;119 uint32_t srcOffset = 0;
111 uint32_t dstOffset = 0;120 uint32_t dstOffset = 0;
112 uint32_t srcBatchStride = 0;121 uint32_t srcBatchStride = 0;
113 if constexpr (Intf::c04Flag) {122 if constexpr (Intf::c04Flag) {
114- srcBatchStride = AlignB(C04_CIN_SIZE * self_->ctx.convTiling->orgHixWi, Intf::k0);123+ srcBatchStride = AlignB(C04_CIN_SIZE * self_->ctx.convTilingData->convApiTiling.orgHixWi, Intf::k0);
115 } else {124 } else {
116- srcBatchStride = ((self_->ctx.kIter / self_->ctx.multiKAL1) != self_->ctx.maxKAL1Iter ?125+ srcBatchStride = ((kIter / self_->ctx.multiKAL1) != self_->ctx.maxKAL1Iter ?
117- self_->ctx.convTiling->cinAInCore : alignCinATailInCore_) * realHixWi;126+ self_->ctx.convTilingData->convApiTiling.cinAInCore : alignCinATailInCore_) * realHixWi;
118- } 127+ }
119- uint32_t dstBatchStride = currentML0Align_ * self_->ctx.convTiling->kL0;128+ uint32_t dstBatchStride = currentML0Align_ * self_->ctx.convTilingData->convApiTiling.kL0;
120 for (uint16_t batchIter = 0; batchIter < self_->ctx.innerBatch; batchIter++) {129 for (uint16_t batchIter = 0; batchIter < self_->ctx.innerBatch; batchIter++) {
121- LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(self_->ctx.al0[dstOffset],130+ LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(al0[dstOffset],
122 self_->ctx.al1[srcOffset], param_);131 self_->ctx.al1[srcOffset], param_);
123 srcOffset += srcBatchStride;132 srcOffset += srcBatchStride;
124 dstOffset += dstBatchStride;133 dstOffset += dstBatchStride;
125 }134 }
126 } else {135 } else {
127- LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(self_->ctx.al0, self_->ctx.al1, param_);136+ LoadData<TPosition::A2, TPosition::A1, typename Intf::FmapT>(al0, self_->ctx.al1, param_);
128 }137 }
129 };138 };
130 139 
131-private:
132- __aicore__ inline bool IsKL0Tail()
133- {
134- return self_->ctx.kIter == self_->ctx.maxKL0Iter;
135- }
136 140 
137private:141private:
138 Intf *self_ = nullptr;142 Intf *self_ = nullptr;
139 uint64_t currentML0_ = 0;143 uint64_t currentML0_ = 0;
140 uint64_t currentML0Align_ = 0;144 uint64_t currentML0Align_ = 0;
141- uint64_t xm_ = 0;
142- uint64_t xt_ = 0;
143- uint64_t xmtmp_ = 0;
144 uint64_t alignCinATailInCore_ = 0;145 uint64_t alignCinATailInCore_ = 0;
145 uint16_t channelSize_ = 0;146 uint16_t channelSize_ = 0;
146 uint64_t c04KStepTail = 0;147 uint64_t c04KStepTail = 0;
147 uint64_t realHixWi = 0;148 uint64_t realHixWi = 0;
148- uint64_t currentKL0 = 0;
149 Load3DBitModeParam param_;149 Load3DBitModeParam param_;
150+ UnionDataXt xt_;
151+ UnionDataXm xm_;
150};152};
151 153 
152template <class Intf, typename OutputT, uint64_t FixpipeIdx = 0>154template <class Intf, typename OutputT, uint64_t FixpipeIdx = 0>
@@ -155,18 +157,18 @@ public:
155 __aicore__ inline CopyOutToolsMMode()157 __aicore__ inline CopyOutToolsMMode()
156 {}158 {}
157 159 
160+ __aicore__ inline void SetParams(Intf *self)
161+ {
162+ self_ = self;
163+ valueHoWo_ = self_->ctx.convTilingData->convApiTiling.orgHo * self_->ctx.convTilingData->convApiTiling.orgWo;
164+ }
165+ 
158 __aicore__ inline void SetMN(uint64_t m, uint64_t n)166 __aicore__ inline void SetMN(uint64_t m, uint64_t n)
159 {167 {
160 currentML0_ = m;168 currentML0_ = m;
161 currentNL0_ = n;169 currentNL0_ = n;
162 }170 }
163 171 
164- __aicore__ inline void SetParams(Intf *self)
165- {
166- self_ = self;
167- valueHoWo_ = self_->ctx.orgHo * self_->ctx.orgWo;
168- }
169- 
170 template <template <typename> class TensorTypeT, const FixpipeConfig &config>172 template <template <typename> class TensorTypeT, const FixpipeConfig &config>
171 __aicore__ inline void CopyOut(const TensorTypeT<OutputT> &output, CopyUbInfo* ubInfo = nullptr)173 __aicore__ inline void CopyOut(const TensorTypeT<OutputT> &output, CopyUbInfo* ubInfo = nullptr)
172 {174 {
@@ -193,7 +195,7 @@ public:
193 if constexpr (Intf::isExtendConv2d) {195 if constexpr (Intf::isExtendConv2d) {
194 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);196 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);
195 } else if constexpr (Intf::isQuantScene) {197 } else if constexpr (Intf::isQuantScene) {
196- if (self_->ctx.convTiling->hasScale == 0) {198+ if (self_->ctx.convTilingData->convApiTiling.hasScale == 0) {
197 Fixpipe<OutputT, typename Intf::L0cT, config>(output[offset], self_->ctx.cl0, intriParams);199 Fixpipe<OutputT, typename Intf::L0cT, config>(output[offset], self_->ctx.cl0, intriParams);
198 } else {200 } else {
199 Fixpipe<OutputT, typename Intf::L0cT, config>(201 Fixpipe<OutputT, typename Intf::L0cT, config>(
@@ -225,7 +227,7 @@ private:
225 if constexpr (Intf::isExtendConv2d) {227 if constexpr (Intf::isExtendConv2d) {
226 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);228 ExtendConv2DFixpipe<TensorTypeT, config>(output, intriParams, offset);
227 } else if constexpr (Intf::isQuantScene) {229 } else if constexpr (Intf::isQuantScene) {
228- if (self_->ctx.convTiling->hasScale != 0) {230+ if (self_->ctx.convTilingData->convApiTiling.hasScale != 0) {
229 Fixpipe<OutputT, typename Intf::L0cT, config>(231 Fixpipe<OutputT, typename Intf::L0cT, config>(
230 output[offset], self_->ctx.cl0, self_->ctx.scaleL1[GetScaleL1Addr()], intriParams);232 output[offset], self_->ctx.cl0, self_->ctx.scaleL1[GetScaleL1Addr()], intriParams);
231 } else {233 } else {
@@ -255,7 +257,7 @@ private:
255 intriParams.params.srcNzMatrixStride = self_->ctx.currentML0Align * CeilDiv(currentNL0_ , BLOCK_L0_N);257 intriParams.params.srcNzMatrixStride = self_->ctx.currentML0Align * CeilDiv(currentNL0_ , BLOCK_L0_N);
256 }258 }
257 intriParams.params.srcNzC0Stride = 1;259 intriParams.params.srcNzC0Stride = 1;
258- intriParams.params.dstDnMatrixStride = self_->ctx.orgCo * valueHoWo_;260+ intriParams.params.dstDnMatrixStride = self_->ctx.convTilingData->convApiTiling.orgCo * valueHoWo_;
259 intriParams.dstStride = valueHoWo_;261 intriParams.dstStride = valueHoWo_;
260 }262 }
261 263 
@@ -271,8 +273,8 @@ private:
271 intriParams.srcStride = self_->ctx.currentML0Align;273 intriParams.srcStride = self_->ctx.currentML0Align;
272 intriParams.params.srcNdStride = self_->ctx.currentML0Align * CeilDiv(currentNL0_ , BLOCK_L0_N);274 intriParams.params.srcNdStride = self_->ctx.currentML0Align * CeilDiv(currentNL0_ , BLOCK_L0_N);
273 }275 }
274- intriParams.params.dstNdStride = self_->ctx.orgCo * valueHoWo_;276+ intriParams.params.dstNdStride = self_->ctx.convTilingData->convApiTiling.orgCo * valueHoWo_;
275- intriParams.dstStride = self_->ctx.orgCo;277+ intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgCo;
276 }278 }
277 279 
278 __aicore__ inline void SetFixpipeIntriParamsHWC(FixpipeParamsC310<CO2Layout::ROW_MAJOR> &intriParams)280 __aicore__ inline void SetFixpipeIntriParamsHWC(FixpipeParamsC310<CO2Layout::ROW_MAJOR> &intriParams)
@@ -280,7 +282,7 @@ private:
280 intriParams.nSize = currentNL0_;282 intriParams.nSize = currentNL0_;
281 intriParams.mSize = currentML0_;283 intriParams.mSize = currentML0_;
282 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);284 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);
283- intriParams.dstStride = self_->ctx.orgCo;285+ intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgCo;
284 intriParams.params.ndNum = 1;286 intriParams.params.ndNum = 1;
285 intriParams.params.dstNdStride = 0;287 intriParams.params.dstNdStride = 0;
286 intriParams.params.srcNdStride = 0;288 intriParams.params.srcNdStride = 0;
@@ -293,7 +295,7 @@ private:
293 intriParams.mSize = currentML0_;295 intriParams.mSize = currentML0_;
294 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);296 intriParams.srcStride = AlignB(currentML0_, BLOCK_L0_M);
295 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {297 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {
296- intriParams.dstStride = self_->ctx.orgDo * valueHoWo_;298+ intriParams.dstStride = self_->ctx.convTilingData->convApiTiling.orgDo * valueHoWo_;
297 } else {299 } else {
298 intriParams.dstStride = valueHoWo_;300 intriParams.dstStride = valueHoWo_;
299 }301 }
@@ -318,8 +320,12 @@ private:
318 ubInfo->realBatchUb = 1;320 ubInfo->realBatchUb = 1;
319 }321 }
320 ubInfo->outBatchIdx = self_->ctx.batchIter + ubInfo->batchLoopIdx * ubInfo->batchUb;322 ubInfo->outBatchIdx = self_->ctx.batchIter + ubInfo->batchLoopIdx * ubInfo->batchUb;
321- ubInfo->outCIdx = self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +323+ if constexpr (Intf::isKL1NL0FullLoad) {
322- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0 + ubInfo->nLoopIdx * ubInfo->nUb;324+ ubInfo->outCIdx = ubInfo->nLoopIdx * ubInfo->nUb;
325+ } else {
326+ ubInfo->outCIdx = self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
327+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0 + ubInfo->nLoopIdx * ubInfo->nUb;
328+ }
323 ubInfo->outHIdx = 0;329 ubInfo->outHIdx = 0;
324 ubInfo->outWIdx = self_->ctx.mAL1Iter * self_->ctx.mAL1 +330 ubInfo->outWIdx = self_->ctx.mAL1Iter * self_->ctx.mAL1 +
325 self_->ctx.mL0Iter * self_->ctx.mL0 + ubInfo->mLoopIdx * ubInfo->mUb;331 self_->ctx.mL0Iter * self_->ctx.mL0 + ubInfo->mLoopIdx * ubInfo->mUb;
@@ -329,28 +335,34 @@ private:
329 __aicore__ inline void SetBaseParams(FixpipeParamsC310<format> &intriParams)335 __aicore__ inline void SetBaseParams(FixpipeParamsC310<format> &intriParams)
330 {336 {
331 intriParams.quantPre = GetQuantPre<Intf, OutputT, FixpipeIdx>(self_);337 intriParams.quantPre = GetQuantPre<Intf, OutputT, FixpipeIdx>(self_);
332- if (self_->ctx.convTiling->hasScale == 0) {338+ if (self_->ctx.convTilingData->convApiTiling.hasScale == 0) {
333 intriParams.deqScalar = DEQ_SCALAR_ONE;339 intriParams.deqScalar = DEQ_SCALAR_ONE;
334 }340 }
335 if constexpr (Intf::isExtendConv2d) {341 if constexpr (Intf::isExtendConv2d) {
336 if constexpr (FixpipeIdx == 0) {342 if constexpr (FixpipeIdx == 0) {
337- intriParams.deqScalar = self_->ctx.deqScalar0;343+ intriParams.reluEn = self_->ctx.convTilingData->convApiTiling.reluMode0 != 0;
338- intriParams.reluEn = self_->ctx.convTiling->reluMode0 != 0;
339#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)344#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
340- intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTiling->reluMode0);345+ intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->convApiTiling.reluMode0);
341- if (self_->ctx.convTiling->reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {346+ if (self_->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
342 intriParams.reluScalar = self_->ctx.preReluScalar0;347 intriParams.reluScalar = self_->ctx.preReluScalar0;
348+ } else if (self_->ctx.convTilingData->convApiTiling.reluMode0 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
349+ intriParams.vectorRelu =
350+ self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr()].GetPhyAddr();
343 }351 }
344#endif352#endif
353+ intriParams.deqScalar = self_->ctx.deqScalar0;
345 } else {354 } else {
346- intriParams.deqScalar = self_->ctx.deqScalar1;355+ intriParams.reluEn = self_->ctx.convTilingData->convApiTiling.reluMode1 != 0;
347- intriParams.reluEn = self_->ctx.convTiling->reluMode1 != 0;
348#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)356#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 5102)
349- intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTiling->reluMode1);357+ intriParams.preReluMode = static_cast<ReluMode>(self_->ctx.convTilingData->convApiTiling.reluMode1);
350- if (self_->ctx.convTiling->reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {358+ if (self_->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::SCALAR_RELU)) {
351 intriParams.reluScalar = self_->ctx.preReluScalar1;359 intriParams.reluScalar = self_->ctx.preReluScalar1;
360+ } else if(self_->ctx.convTilingData->convApiTiling.reluMode1 == static_cast<uint8_t>(ReluMode::VECTOR_RELU)) {
361+ intriParams.vectorRelu =
362+ self_->ctx.reluWeightL1[GetExtendConv2dScaleL1Addr() + self_->ctx.reluWeight1L1offset].GetPhyAddr();
352 }363 }
353#endif364#endif
365+ intriParams.deqScalar = self_->ctx.deqScalar1;
354 }366 }
355 }367 }
356 368 
@@ -359,10 +371,10 @@ private:
359 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;371 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
360 } else {372 } else {
361 if constexpr (Intf::isExtendConv2d) {373 if constexpr (Intf::isExtendConv2d) {
362- if (!self_->ctx.convTiling->dualOutput) {374+ if (self_->ctx.convTilingData->convApiTiling.dualOutput) {
363- intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
364- } else {
365 intriParams.unitFlag = UNIT_FLAG_ENABLE_ONLY;375 intriParams.unitFlag = UNIT_FLAG_ENABLE_ONLY;
376+ } else {
377+ intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
366 }378 }
367 } else {379 } else {
368 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;380 intriParams.unitFlag = UNIT_FLAG_ENABLE_WITH_FLIP;
@@ -424,27 +436,34 @@ private:
424 436 
425 __aicore__ inline uint64_t CalcFixpipeOffset()437 __aicore__ inline uint64_t CalcFixpipeOffset()
426 {438 {
427- uint64_t offset = self_->ctx.batchIter * self_->ctx.outputOneBatchSize;439+ uint64_t offset = 0;
428- if constexpr (Intf::isInnerBatchFlag) {440+ if constexpr (!Intf::isOneBatch) {
429- offset *= self_->ctx.convTiling->innerBatch;441+ offset = self_->ctx.batchIter * self_->ctx.outputOneBatchSize;
430 }442 }
431- uint64_t offsetCout = self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +443+ if constexpr (Intf::isInnerBatchFlag) {
432- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0;444+ offset *= self_->ctx.convTilingData->convApiTiling.innerBatch;
445+ }
446+ 
447+ uint64_t offsetCout = 0;
433 if constexpr (Intf::groupOptPreloadFlag) {448 if constexpr (Intf::groupOptPreloadFlag) {
434- offsetCout += self_->ctx.groupOptIter * self_->ctx.convTiling->orgCo / self_->ctx.convTiling->groups *449+ offsetCout += self_->ctx.groupOptIter * self_->ctx.convTilingData->convApiTiling.orgCo / self_->ctx.convTilingData->convApiTiling.groups *
435- self_->ctx.convTiling->enlarge;450+ self_->ctx.convTilingData->convApiTiling.enlarge;
451+ }
452+ if constexpr (!Intf::isKL1NL0FullLoad) {
453+ offsetCout += self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
454+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;
436 }455 }
437 uint64_t offsetMAL1 = self_->ctx.mAL1Iter * self_->ctx.mAL1 +456 uint64_t offsetMAL1 = self_->ctx.mAL1Iter * self_->ctx.mAL1 +
438 self_->ctx.mL0Iter * self_->ctx.mL0;457 self_->ctx.mL0Iter * self_->ctx.mL0;
439 458
440 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {459 if constexpr (Intf::formatOutput == ConvFormat::NCDHW) {
441- offset += offsetCout * self_->ctx.orgDo * valueHoWo_ + self_->ctx.dOutIter * valueHoWo_ + offsetMAL1;460+ offset += offsetCout * self_->ctx.convTilingData->convApiTiling.orgDo * valueHoWo_ + self_->ctx.dOutIter * valueHoWo_ + offsetMAL1;
442 } else if constexpr (Intf::formatOutput == ConvFormat::NDHWC) {461 } else if constexpr (Intf::formatOutput == ConvFormat::NDHWC) {
443- offset += self_->ctx.dOutIter * valueHoWo_ * self_->ctx.orgCo + offsetMAL1 * self_->ctx.orgCo + offsetCout;462+ offset += self_->ctx.dOutIter * valueHoWo_ * self_->ctx.convTilingData->convApiTiling.orgCo + offsetMAL1 * self_->ctx.convTilingData->convApiTiling.orgCo + offsetCout;
444 } else if constexpr (Intf::formatOutput == ConvFormat::NCHW) {463 } else if constexpr (Intf::formatOutput == ConvFormat::NCHW) {
445 offset += offsetCout * valueHoWo_ + offsetMAL1;464 offset += offsetCout * valueHoWo_ + offsetMAL1;
446 } else {465 } else {
447- offset += offsetMAL1 * self_->ctx.orgCo + offsetCout;466+ offset += offsetMAL1 * self_->ctx.convTilingData->convApiTiling.orgCo + offsetCout;
448 }467 }
449 468 
450 return offset;469 return offset;
@@ -470,10 +489,13 @@ private:
470 489 
471 __aicore__ inline uint64_t GetScaleL1Addr()490 __aicore__ inline uint64_t GetScaleL1Addr()
472 {491 {
492+ if constexpr (Intf::isKL1NL0FullLoad) {
493+ return 0;
494+ }
473 if constexpr (Intf::isQuantScene) {495 if constexpr (Intf::isQuantScene) {
474- if (self_->ctx.convTiling->fixpParamsFullLoadFlag && self_->ctx.convTiling->hasScale != 0) {496+ if (self_->ctx.convTilingData->convApiTiling.hasScale != 0 && self_->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
475- return self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +497+ return self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
476- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0;498+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;
477 }499 }
478 }500 }
479 501 
@@ -482,9 +504,12 @@ private:
482 504 
483 __aicore__ inline uint64_t GetExtendConv2dScaleL1Addr()505 __aicore__ inline uint64_t GetExtendConv2dScaleL1Addr()
484 {506 {
485- if (self_->ctx.convTiling->fixpParamsFullLoadFlag) {507+ if constexpr (Intf::isKL1NL0FullLoad) {
486- return self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1 +508+ return 0;
487- self_->ctx.nL0Iter * self_->ctx.convTiling->nL0;509+ }
510+ if (self_->ctx.convTilingData->convApiTiling.fixpParamsFullLoadFlag) {
511+ return self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1 +
512+ self_->ctx.nL0Iter * self_->ctx.convTilingData->convApiTiling.nL0;
488 }513 }
489 return 0;514 return 0;
490 }515 }
Mconv/common/op_kernel/arch35/conv_instr_opt_group_impl.h+40-31
@@ -62,12 +62,12 @@ public:
62 if constexpr (Intf::groupOptPreloadFlag) {62 if constexpr (Intf::groupOptPreloadFlag) {
63 uint64_t weightOneGroupSize = self_->ctx.coPerGroup * self_->ctx.ciPerGroup * self_->ctx.enlarge;63 uint64_t weightOneGroupSize = self_->ctx.coPerGroup * self_->ctx.ciPerGroup * self_->ctx.enlarge;
64 if constexpr (Intf::formatOutput == ConvFormat::NCHW) {64 if constexpr (Intf::formatOutput == ConvFormat::NCHW) {
65- weightOneGroupSize *= self_->ctx.convTiling->kernelHxkernelWxkernelD;65+ weightOneGroupSize *= self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;
66 }66 }
67 gmOffset = weightOneGroupSize * self_->ctx.groupOptIter;67 gmOffset = weightOneGroupSize * self_->ctx.groupOptIter;
68 }68 }
69 if constexpr (Intf::formatOutput == ConvFormat::NDHWC || Intf::formatOutput == ConvFormat::NHWC) {69 if constexpr (Intf::formatOutput == ConvFormat::NDHWC || Intf::formatOutput == ConvFormat::NHWC) {
70- uint64_t curSrcCoOpt = self_->ctx.convTiling->orgCo;70+ uint64_t curSrcCoOpt = self_->ctx.convTilingData->convApiTiling.orgCo;
71 copyParamsHWC.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = curSrcCoOpt;71 copyParamsHWC.loopInfo.loopSrcStride[NDDMA_LOOP1_INDEX] = curSrcCoOpt;
72 copyParamsHWC.loopInfo.loopSrcStride[NDDMA_LOOP3_INDEX] = curSrcCoOpt * self_->ctx.ciPerGroup;72 copyParamsHWC.loopInfo.loopSrcStride[NDDMA_LOOP3_INDEX] = curSrcCoOpt * self_->ctx.ciPerGroup;
73 copyParamsHWC.loopInfo.loopSize[NDDMA_LOOP2_INDEX] = self_->ctx.singleGroups;73 copyParamsHWC.loopInfo.loopSize[NDDMA_LOOP2_INDEX] = self_->ctx.singleGroups;
@@ -83,7 +83,7 @@ public:
83private:83private:
84__aicore__ inline void NDDMAFirstSetCopyParamsCHW()84__aicore__ inline void NDDMAFirstSetCopyParamsCHW()
85 {85 {
86- uint64_t srcKSize = self_->ctx.ciPerGroup * self_->ctx.convTiling->kernelHxkernelWxkernelD;86+ uint64_t srcKSize = self_->ctx.ciPerGroup * self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;
87 // NDDMA Loop0 params87 // NDDMA Loop0 params
88 copyParams.loopInfo.loopSize[NDDMA_LOOP0_INDEX] = srcKSize;88 copyParams.loopInfo.loopSize[NDDMA_LOOP0_INDEX] = srcKSize;
89 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 1;89 copyParams.loopInfo.loopSrcStride[NDDMA_LOOP0_INDEX] = 1;
@@ -112,7 +112,7 @@ __aicore__ inline void NDDMAFirstSetCopyParamsCHW()
112 copyParamsHWC.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = self_->ctx.coOptAlign * self_->ctx.ciPerGroup +112 copyParamsHWC.loopInfo.loopDstStride[NDDMA_LOOP2_INDEX] = self_->ctx.coOptAlign * self_->ctx.ciPerGroup +
113 self_->ctx.coPerGroup;113 self_->ctx.coPerGroup;
114 // NDDMA Loop3 params114 // NDDMA Loop3 params
115- copyParamsHWC.loopInfo.loopSize[NDDMA_LOOP3_INDEX] = self_->ctx.convTiling->kernelHxkernelWxkernelD;115+ copyParamsHWC.loopInfo.loopSize[NDDMA_LOOP3_INDEX] = self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;
116 copyParamsHWC.loopInfo.loopDstStride[NDDMA_LOOP3_INDEX] = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;116 copyParamsHWC.loopInfo.loopDstStride[NDDMA_LOOP3_INDEX] = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;
117 }117 }
118 118 
@@ -170,7 +170,7 @@ private:
170 for (uint8_t idx = 0; idx < Intf::k0; ++idx) {170 for (uint8_t idx = 0; idx < Intf::k0; ++idx) {
171 indexTensor.SetValue(idx, curValue);171 indexTensor.SetValue(idx, curValue);
172 if constexpr (Intf::formatOutput == ConvFormat::NCDHW || Intf::formatOutput == ConvFormat::NCHW) {172 if constexpr (Intf::formatOutput == ConvFormat::NCDHW || Intf::formatOutput == ConvFormat::NCHW) {
173- curValue += self_->ctx.convTiling->kernelHxkernelWxkernelD;173+ curValue += self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;
174 } else {174 } else {
175 curValue += self_->ctx.coOptAlign;175 curValue += self_->ctx.coOptAlign;
176 }176 }
@@ -209,10 +209,10 @@ private:
209 {209 {
210 uint16_t ciLoopTimes = self_->ctx.ci1Opt;210 uint16_t ciLoopTimes = self_->ctx.ci1Opt;
211 uint16_t coLoopTimes = coOptLoopTimes;211 uint16_t coLoopTimes = coOptLoopTimes;
212- uint16_t khkwLoopTimes = self_->ctx.convTiling->kernelHxkernelW;212+ uint16_t khkwLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;
213- uint32_t srcCiStride = self_->ctx.convTiling->kernelHxkernelW * Intf::k0;213+ uint32_t srcCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0;
214 uint32_t srcCoStride = coPerReg * self_->ctx.kUbSize;214 uint32_t srcCoStride = coPerReg * self_->ctx.kUbSize;
215- uint32_t dstCiStride = self_->ctx.convTiling->kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;215+ uint32_t dstCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;
216 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;216 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;
217 uint32_t dstCoStride = coPerReg * Intf::k0;217 uint32_t dstCoStride = coPerReg * Intf::k0;
218 218 
@@ -262,16 +262,16 @@ private:
262 262 
263 __aicore__ inline void TransNCDHW2NZ()263 __aicore__ inline void TransNCDHW2NZ()
264 {264 {
265- uint16_t kdLoopTimes = self_->ctx.convTiling->kernelD;265+ uint16_t kdLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelD;
266 uint16_t ciLoopTimes = self_->ctx.ci1Opt;266 uint16_t ciLoopTimes = self_->ctx.ci1Opt;
267 uint16_t coLoopTimes = coOptLoopTimes;267 uint16_t coLoopTimes = coOptLoopTimes;
268- uint16_t khkwLoopTimes = self_->ctx.convTiling->kernelHxkernelW;268+ uint16_t khkwLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;
269- uint32_t srcCiStride = self_->ctx.convTiling->kernelHxkernelWxkernelD * Intf::k0;269+ uint32_t srcCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD * Intf::k0;
270 uint32_t srcCoStride = coPerReg * self_->ctx.kUbSize;270 uint32_t srcCoStride = coPerReg * self_->ctx.kUbSize;
271- uint32_t dstCiStride = self_->ctx.convTiling->kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;271+ uint32_t dstCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;
272 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;272 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;
273 uint32_t dstCoStride = coPerReg * Intf::k0;273 uint32_t dstCoStride = coPerReg * Intf::k0;
274- uint32_t srcKdStride = self_->ctx.convTiling->kernelHxkernelW;274+ uint32_t srcKdStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;
275 uint32_t dstKdStride = self_->ctx.ci1Opt * dstCiStride;275 uint32_t dstKdStride = self_->ctx.ci1Opt * dstCiStride;
276 276 
277 __VEC_SCOPE__277 __VEC_SCOPE__
@@ -321,19 +321,19 @@ private:
321 321 
322 __aicore__ inline void TransNDHWC2NZ()322 __aicore__ inline void TransNDHWC2NZ()
323 {323 {
324- uint16_t kdLoopTimes = self_->ctx.convTiling->kernelD;324+ uint16_t kdLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelD;
325 uint16_t ciLoopTimes = self_->ctx.ci1Opt;325 uint16_t ciLoopTimes = self_->ctx.ci1Opt;
326 uint16_t coLoopTimes = coOptLoopTimes;326 uint16_t coLoopTimes = coOptLoopTimes;
327- uint16_t khkwLoopTimes = self_->ctx.convTiling->kernelHxkernelW;327+ uint16_t khkwLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;
328 uint32_t srcGroupOptSize = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;328 uint32_t srcGroupOptSize = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;
329 uint32_t srcCiStride = self_->ctx.coOptAlign * Intf::k0;329 uint32_t srcCiStride = self_->ctx.coOptAlign * Intf::k0;
330 uint32_t srcCoStride = coPerReg;330 uint32_t srcCoStride = coPerReg;
331- uint32_t dstCiStride = self_->ctx.convTiling->kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;331+ uint32_t dstCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;
332 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;332 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;
333 uint32_t srcKhKwStride = srcGroupOptSize;333 uint32_t srcKhKwStride = srcGroupOptSize;
334 uint32_t dstCoStride = coPerReg * Intf::k0;334 uint32_t dstCoStride = coPerReg * Intf::k0;
335- uint32_t srcKdStride = self_->ctx.convTiling->kernelHxkernelW * srcGroupOptSize;335+ uint32_t srcKdStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * srcGroupOptSize;
336- uint32_t dstKdStride = self_->ctx.convTiling->kernelHxkernelW * self_->ctx.coOptAlign *336+ uint32_t dstKdStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * self_->ctx.coOptAlign *
337 self_->ctx.ci1Opt * Intf::k0; // ci1Opt has updated in groupOptTail337 self_->ctx.ci1Opt * Intf::k0; // ci1Opt has updated in groupOptTail
338 __VEC_SCOPE__338 __VEC_SCOPE__
339 {339 {
@@ -383,12 +383,12 @@ private:
383 __aicore__ inline void TransNHWC2NZ()383 __aicore__ inline void TransNHWC2NZ()
384 {384 {
385 uint16_t ciLoopTimes = self_->ctx.ci1Opt;385 uint16_t ciLoopTimes = self_->ctx.ci1Opt;
386- uint16_t khkwLoopTimes = self_->ctx.convTiling->kernelHxkernelW;386+ uint16_t khkwLoopTimes = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;
387 uint16_t coLoopTimes = coOptLoopTimes;387 uint16_t coLoopTimes = coOptLoopTimes;
388 uint32_t srcCiStride = self_->ctx.coOptAlign * Intf::k0;388 uint32_t srcCiStride = self_->ctx.coOptAlign * Intf::k0;
389 uint32_t srcKhKwStride = self_->ctx.coOptAlign * self_->ctx.ciOptAlign; 389 uint32_t srcKhKwStride = self_->ctx.coOptAlign * self_->ctx.ciOptAlign;
390 uint32_t srcCoStride = coPerReg;390 uint32_t srcCoStride = coPerReg;
391- uint32_t dstCiStride = self_->ctx.convTiling->kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;391+ uint32_t dstCiStride = self_->ctx.convTilingData->convApiTiling.kernelHxkernelW * Intf::k0 * self_->ctx.coOptAlign;
392 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;392 uint32_t dstKhKwStride = Intf::k0 * self_->ctx.coOptAlign;
393 uint32_t dstCoStride = coPerReg * Intf::k0;393 uint32_t dstCoStride = coPerReg * Intf::k0;
394 394
@@ -482,30 +482,39 @@ private:
482 __aicore__ inline void SetCopyParams2D()482 __aicore__ inline void SetCopyParams2D()
483 {483 {
484 if (unlikely(self_->ctx.groupOptIter == self_->ctx.vecId)) {484 if (unlikely(self_->ctx.groupOptIter == self_->ctx.vecId)) {
485- copyParams.blockCount = self_->ctx.convTiling->kBL1 / Intf::k0;485+ copyParams.blockCount = self_->ctx.convTilingData->convApiTiling.kBL1 / Intf::k0;
486- copyParams.blockLen = self_->ctx.convTiling->nBL1;486+ copyParams.blockLen = self_->ctx.convTilingData->convApiTiling.nBL1;
487- copyParams.srcStride = self_->ctx.co1Opt * BLOCK_L0_N - self_->ctx.convTiling->nBL1;487+ copyParams.srcStride = self_->ctx.co1Opt * BLOCK_L0_N - self_->ctx.convTilingData->convApiTiling.nBL1;
488 }488 }
489 489 
490- srcOffset = (self_->ctx.coStartPos + self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1) * Intf::k0 +490+ if constexpr (Intf::isKL1NL0FullLoad) {
491- self_->ctx.kBL1Iter * self_->ctx.convTiling->kBL1 * self_->ctx.coOptAlign;491+ srcOffset = self_->ctx.coStartPos * Intf::k0 +
492+ self_->ctx.kBL1Iter * self_->ctx.convTilingData->convApiTiling.kBL1 * self_->ctx.coOptAlign;
493+ } else {
494+ srcOffset = (self_->ctx.coStartPos + self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1) * Intf::k0 +
495+ self_->ctx.kBL1Iter * self_->ctx.convTilingData->convApiTiling.kBL1 * self_->ctx.coOptAlign;
496+ }
492 }497 }
493 498 
494 __aicore__ inline void SetCopyParams3D()499 __aicore__ inline void SetCopyParams3D()
495 {500 {
496 if (unlikely(self_->ctx.loadUB2L1Iter == 0)) {501 if (unlikely(self_->ctx.loadUB2L1Iter == 0)) {
497- copyParams.blockLen = self_->ctx.convTiling->nBL1;502+ copyParams.blockLen = self_->ctx.convTilingData->convApiTiling.nBL1;
498- copyParams.srcStride = self_->ctx.co1Opt * BLOCK_L0_N - self_->ctx.convTiling->nBL1;503+ copyParams.srcStride = self_->ctx.co1Opt * BLOCK_L0_N - self_->ctx.convTilingData->convApiTiling.nBL1;
499- kOffset = self_->ctx.bL1Dk * self_->ctx.bL1Cin * self_->ctx.convTiling->kernelHxkernelW;504+ kOffset = self_->ctx.bL1Dk * self_->ctx.bL1Cin * self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;
500 }505 }
501 506 
502 uint64_t currentBL1Dk = IsKBL1Tail() ? self_->ctx.bL1DkTail : self_->ctx.bL1Dk;507 uint64_t currentBL1Dk = IsKBL1Tail() ? self_->ctx.bL1DkTail : self_->ctx.bL1Dk;
503 uint64_t currentBL1Cin1 = IsKBL1Tail() ? self_->ctx.bL1CinTail : self_->ctx.bL1Cin;508 uint64_t currentBL1Cin1 = IsKBL1Tail() ? self_->ctx.bL1CinTail : self_->ctx.bL1Cin;
504 copyParams.blockCount = currentBL1Dk * CeilDiv(currentBL1Cin1, Intf::k0) *509 copyParams.blockCount = currentBL1Dk * CeilDiv(currentBL1Cin1, Intf::k0) *
505- self_->ctx.convTiling->kernelHxkernelW;510+ self_->ctx.convTilingData->convApiTiling.kernelHxkernelW;
506 511 
507- srcOffset = (self_->ctx.coStartPos + self_->ctx.nBL1Iter * self_->ctx.convTiling->nBL1) * Intf::k0 +512+ if constexpr (Intf::isKL1NL0FullLoad) {
508- self_->ctx.kBL1Iter * kOffset * self_->ctx.coOptAlign;513+ srcOffset = self_->ctx.coStartPos * Intf::k0 + self_->ctx.kBL1Iter * kOffset * self_->ctx.coOptAlign;
514+ } else {
515+ srcOffset = (self_->ctx.coStartPos + self_->ctx.nBL1Iter * self_->ctx.convTilingData->convApiTiling.nBL1) * Intf::k0 +
516+ self_->ctx.kBL1Iter * kOffset * self_->ctx.coOptAlign;
517+ }
509 }518 }
510 519 
511 __aicore__ inline bool IsKBL1Tail()520 __aicore__ inline bool IsKBL1Tail()
Mconv/common/op_kernel/arch35/conv_iterate_base_impl.h+171-119
@@ -61,11 +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.convTiling->innerBatch);64+ self->ctx.ddr2l1LoopBatch = CeilDiv(self->ctx.singleCoreBatch, self->ctx.convTilingData->convApiTiling.innerBatch);
65- self->ctx.maxBatchIter = self->ctx.ddr2l1LoopBatch - 1;65+ self->ctx.innerBatchTail = self->ctx.singleCoreBatch % self->ctx.convTilingData->convApiTiling.innerBatch;
66- self->ctx.innerBatchTail = self->ctx.singleCoreBatch % self->ctx.convTiling->innerBatch;
67 self->ctx.innerBatchTail = self->ctx.innerBatchTail == 0 ?66 self->ctx.innerBatchTail = self->ctx.innerBatchTail == 0 ?
68- self->ctx.convTiling->innerBatch : self->ctx.innerBatchTail;67+ self->ctx.convTilingData->convApiTiling.innerBatch : self->ctx.innerBatchTail;
69 } else {68 } else {
70 self->ctx.ddr2l1LoopBatch = self->ctx.singleCoreBatch;69 self->ctx.ddr2l1LoopBatch = self->ctx.singleCoreBatch;
71 }70 }
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.convTiling->hoL1;31+ self->ctx.hoAL1Tail = self->ctx.singleCoreHo % self->ctx.convTilingData->convApiTiling.hoL1;
32- self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTiling->hoL1 : self->ctx.hoAL1Tail;32+ self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.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.convTiling->hoL0;34+ self->ctx.hoL0Tail = self->ctx.currentHoL1 % self->ctx.convTilingData->convApiTiling.hoL0;
35- self->ctx.hoL0Tail = self->ctx.hoL0Tail == 0 ? self->ctx.convTiling->hoL0 : self->ctx.hoL0Tail;35+ self->ctx.hoL0Tail = self->ctx.hoL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.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.convTiling->hoL1);43+ self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.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.convTiling->woL1;52+ self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->convApiTiling.woL1;
53- self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTiling->woL1 : self->ctx.woAL1Tail;53+ self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.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.convTiling->woL0;56+ self->ctx.woL0Tail = self->ctx.currentWoL1 % self->ctx.convTilingData->convApiTiling.woL0;
57- self->ctx.woL0Tail = self->ctx.woL0Tail == 0 ? self->ctx.convTiling->woL0 : self->ctx.woL0Tail;57+ self->ctx.woL0Tail = self->ctx.woL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.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.convTiling->woL1) - 1;64+ self->ctx.maxWoL1Iter = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL1) - 1;
65 }65 }
66 66 
67 if constexpr (!Intf::isDmaFlag) {67 if constexpr (!Intf::isDmaFlag) {
68- if (Intf::hasWL0IterFlag && self->ctx.convTiling->hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&68+ if (Intf::hasWL0IterFlag && self->ctx.convTilingData->convApiTiling.hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&
69- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTiling->woL0) > 1) {69+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0) > 1) {
70- self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTiling->woL0;70+ self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->convApiTiling.woL0;
71- self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTiling->woL0) * self->ctx.convTiling->woL0;71+ self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->convApiTiling.woL0) * self->ctx.convTilingData->convApiTiling.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.convTiling->woL1;92+ self->ctx.currentWoL1 = self->ctx.convTilingData->convApiTiling.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.convTiling->woL1;96+ self->ctx.woAL1Tail : self->ctx.convTilingData->convApiTiling.woL1;
97 }97 }
98 }98 }
99 if constexpr (Intf::hasWL0IterFlag) {99 if constexpr (Intf::hasWL0IterFlag) {
100- self->ctx.woL0Tail = self->ctx.currentWoL1 % self->ctx.convTiling->woL0;100+ self->ctx.woL0Tail = self->ctx.currentWoL1 % self->ctx.convTilingData->convApiTiling.woL0;
101- self->ctx.woL0Tail = self->ctx.woL0Tail == 0 ? self->ctx.convTiling->woL0 : self->ctx.woL0Tail;101+ self->ctx.woL0Tail = self->ctx.woL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL0 : self->ctx.woL0Tail;
102- self->ctx.maxWoL0Iter = CeilDiv(self->ctx.currentWoL1, self->ctx.convTiling->woL0) - 1;102+ self->ctx.maxWoL0Iter = CeilDiv(self->ctx.currentWoL1, self->ctx.convTilingData->convApiTiling.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.convTiling->hoL1;112+ self->ctx.hoAL1Tail : self->ctx.convTilingData->convApiTiling.hoL1;
113 }113 }
114 114
115 if constexpr (Intf::hasHL0IterFlag) {115 if constexpr (Intf::hasHL0IterFlag) {
116- self->ctx.hoL0Tail = self->ctx.currentHoL1 % self->ctx.convTiling->hoL0;116+ self->ctx.hoL0Tail = self->ctx.currentHoL1 % self->ctx.convTilingData->convApiTiling.hoL0;
117- self->ctx.hoL0Tail = self->ctx.hoL0Tail == 0 ? self->ctx.convTiling->hoL0 : self->ctx.hoL0Tail;117+ self->ctx.hoL0Tail = self->ctx.hoL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.hoL0 : self->ctx.hoL0Tail;
118- self->ctx.maxHoL0Iter = CeilDiv(self->ctx.currentHoL1, self->ctx.convTiling->hoL0) - 1;118+ self->ctx.maxHoL0Iter = CeilDiv(self->ctx.currentHoL1, self->ctx.convTilingData->convApiTiling.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.convTiling->hoL0;127+ self->ctx.hoL0Tail : self->ctx.convTilingData->convApiTiling.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.convTiling->woL0;129+ self->ctx.woL0Tail : self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge) {136+ self->ctx.singleGroupOpt == 1) && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge : self->ctx.singleGroups;138+ self->ctx.singleGroups = self->ctx.singleGroups == 0 ? self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;
139 }139 }
140- uint64_t enlargeTail = self->ctx.singleGroups % self->ctx.convTiling->enlarge;140+ uint64_t enlargeTail = self->ctx.singleGroups % self->ctx.convTilingData->convApiTiling.enlarge;
141- enlargeTail = enlargeTail == 0 ? self->ctx.convTiling->enlarge : enlargeTail;141+ enlargeTail = enlargeTail == 0 ? self->ctx.convTilingData->convApiTiling.enlarge : enlargeTail;
142- if (enlargeTail != self->ctx.convTiling->enlarge) {142+ if (enlargeTail != self->ctx.convTilingData->convApiTiling.enlarge) {
143- self->ctx.singleCoreCi = enlargeTail * (self->ctx.convTiling->orgCi / self->ctx.convTiling->groups);143+ self->ctx.singleCoreCi = enlargeTail * (self->ctx.convTilingData->convApiTiling.orgCi / self->ctx.convTilingData->convApiTiling.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.convTiling->kernelHxkernelW;147+ uint64_t totalKAlignK0 = AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelW;
148- self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTiling->kL0);148+ self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTilingData->convApiTiling.kL0);
149 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;149 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;
150- self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTiling->kL0;150+ self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTilingData->convApiTiling.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.convTiling->kernelHxkernelW % self->ctx.convTiling->kL0;153+ self->ctx.convTilingData->convApiTiling.kernelHxkernelW % self->ctx.convTilingData->convApiTiling.kL0;
154- self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTiling->kL0 : self->ctx.kAL0Tail;154+ self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.kL0 : self->ctx.kAL0Tail;
155 }155 }
156- self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTiling->kL0 : self->ctx.kL0Tail;156+ self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge) {202+ if (self->ctx.singleGroupOpt == 1 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge : self->ctx.singleGroups;205+ self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge) {210+ if (self->ctx.singleGroupOpt == 2 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge : self->ctx.singleGroups;213+ self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge) {370+ if (self->ctx.updateSingleCoOpt == 0 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {
371 return false;371 return false;
372 }372 }
373 return true;373 return true;
Mconv/common/op_kernel/arch35/conv_iterate_m_mode_impl.h+42-25
@@ -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.convTiling->enlarge) {54+ self->ctx.singleGroupOpt == 1) && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge : self->ctx.singleGroups;56+ self->ctx.singleGroups = self->ctx.singleGroups == 0 ? self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;
57 }57 }
58 58 
59- uint64_t enlargeTail = self->ctx.singleGroups % self->ctx.convTiling->enlarge;59+ uint64_t enlargeTail = self->ctx.singleGroups % self->ctx.convTilingData->convApiTiling.enlarge;
60- enlargeTail = enlargeTail == 0 ? self->ctx.convTiling->enlarge : enlargeTail;60+ enlargeTail = enlargeTail == 0 ? self->ctx.convTilingData->convApiTiling.enlarge : enlargeTail;
61- if (enlargeTail != self->ctx.convTiling->enlarge) {61+ if (enlargeTail != self->ctx.convTilingData->convApiTiling.enlarge) {
62- self->ctx.singleCoreCi = enlargeTail * (self->ctx.convTiling->orgCi / self->ctx.convTiling->groups);62+ self->ctx.singleCoreCi = enlargeTail * (self->ctx.convTilingData->convApiTiling.orgCi / self->ctx.convTilingData->convApiTiling.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.convTiling->kernelHxkernelW;66+ uint64_t totalKAlignK0 = AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelW;
67- self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTiling->kL0);67+ self->ctx.ddr2l0LoopK = CeilDiv(totalKAlignK0, self->ctx.convTilingData->convApiTiling.kL0);
68 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;68 self->ctx.maxKL0Iter = self->ctx.ddr2l0LoopK - 1;
69- self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTiling->kL0;69+ self->ctx.kL0Tail = totalKAlignK0 % self->ctx.convTilingData->convApiTiling.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.convTiling->kernelHxkernelW % self->ctx.convTiling->kL0;72+ self->ctx.convTilingData->convApiTiling.kernelHxkernelW % self->ctx.convTilingData->convApiTiling.kL0;
73- self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTiling->kL0 : self->ctx.kAL0Tail;73+ self->ctx.kAL0Tail = self->ctx.kAL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.kL0 : self->ctx.kAL0Tail;
74 }74 }
75- self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTiling->kL0 : self->ctx.kL0Tail;75+ self->ctx.kL0Tail = self->ctx.kL0Tail == 0 ? self->ctx.convTilingData->convApiTiling.kL0 : self->ctx.kL0Tail;
76 76
77 InitCoDirectionValue<Intf>(self);77 InitCoDirectionValue<Intf>(self);
78 }78 }
@@ -113,21 +113,32 @@ __aicore__ inline void FirstIterateImplMMode(Intf *self)
113 CalcMDirectionVar<Intf>(self);113 CalcMDirectionVar<Intf>(self);
114 CalcCoDirectionVar<Intf>(self);114 CalcCoDirectionVar<Intf>(self);
115 115 
116- if (Intf::groupOptPreloadFlag) {116+ if constexpr (Intf::groupOptPreloadFlag) {
117- if (self->ctx.singleGroupOpt == 1 && self->ctx.updateEnlarge != self->ctx.convTiling->enlarge) {117+ if (self->ctx.singleGroupOpt == 1 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge : self->ctx.singleGroups;120+ self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge) {125+ if (self->ctx.singleGroupOpt == 2 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.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.convTiling->enlarge : self->ctx.singleGroups;128+ self->ctx.convTilingData->convApiTiling.enlarge : self->ctx.singleGroups;
129 CalcGroupOptParamForMMode<Intf>(self);129 CalcGroupOptParamForMMode<Intf>(self);
130 }130 }
131+ } else if constexpr (Intf::isMPreLoad) {
132+ self->ctx.al1 = self->ctx.queueAL1.template AllocTensor<typename Intf::FmapT>();
133+ self->ctx.loadAl1Ins.LoadAL1();
134+ self->ctx.queueAL1.EnQue(self->ctx.al1);
135+ self->ctx.mAL1UpdateFlag = true;
136+ LoadBL1BaseModule<Intf>(self);
137+ self->ctx.bl1 = self->ctx.queueBL1.template DeQue<typename Intf::WeightT>();
138+ self->ctx.loadBL0Flag = true;
139+ if constexpr (Intf::isL0BFullLoadable) {
140+ self->ctx.loadBL0Ins.FullLoadBL0(self->ctx.wholeBl0Tensor[0]);
141+ }
131 }142 }
132}143}
133 144 
@@ -175,7 +186,7 @@ __aicore__ inline bool IterateMFirstMMode(Intf *self)
175 }186 }
176 187 
177 if ASCEND_IS_AIC_CONV {188 if ASCEND_IS_AIC_CONV {
178- if constexpr (!Intf::groupOptPreloadFlag) {189+ if constexpr (!Intf::isMPreLoad && !Intf::groupOptPreloadFlag) {
179 if (self->ctx.kAL1fullload) {190 if (self->ctx.kAL1fullload) {
180 self->ctx.queueAL1.FreeTensor(self->ctx.al1);191 self->ctx.queueAL1.FreeTensor(self->ctx.al1);
181 }192 }
@@ -211,7 +222,7 @@ __aicore__ inline bool IterateMFirstMMode(Intf *self)
211 }222 }
212 223 
213 if ASCEND_IS_AIC_CONV {224 if ASCEND_IS_AIC_CONV {
214- if constexpr (!Intf::groupOptPreloadFlag) {225+ if constexpr (!Intf::groupOptPreloadFlag && !Intf::isMPreLoad) {
215 if (self->ctx.kBL1fullload) {226 if (self->ctx.kBL1fullload) {
216 self->ctx.queueBL1.FreeTensor(self->ctx.bl1);227 self->ctx.queueBL1.FreeTensor(self->ctx.bl1);
217 }228 }
@@ -235,7 +246,7 @@ __aicore__ inline bool IterateMFirstMMode(Intf *self)
235 if (self->ctx.groupOptIter < self->ctx.singleGroupOpt - 1) {246 if (self->ctx.groupOptIter < self->ctx.singleGroupOpt - 1) {
236 return true;247 return true;
237 } else if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {248 } else if (self->ctx.groupOptIter == self->ctx.singleGroupOpt - 1) {
238- if (self->ctx.updateSingleCoOpt == 0 && self->ctx.updateEnlarge != self->ctx.convTiling->enlarge) {249+ if (self->ctx.updateSingleCoOpt == 0 && self->ctx.updateEnlarge != self->ctx.convTilingData->convApiTiling.enlarge) {
239 return false;250 return false;
240 }251 }
241 return true;252 return true;
@@ -341,8 +352,10 @@ __aicore__ inline bool IterateNFirstMMode(Intf *self)
341 }352 }
342 353 
343 if ASCEND_IS_AIC_CONV {354 if ASCEND_IS_AIC_CONV {
344- if (self->ctx.kBL1fullload) {355+ if constexpr (!(Intf::preFusionFlag && ConvParam::weightTiling == 0)) {
345- self->ctx.queueBL1.FreeTensor(self->ctx.bl1);356+ if (self->ctx.kBL1fullload) {
357+ self->ctx.queueBL1.FreeTensor(self->ctx.bl1);
358+ }
346 }359 }
347 }360 }
348 361 
@@ -359,8 +372,10 @@ __aicore__ inline bool IterateNFirstMMode(Intf *self)
359 }372 }
360 373 
361 if ASCEND_IS_AIC_CONV {374 if ASCEND_IS_AIC_CONV {
362- if (self->ctx.kAL1fullload) {375+ if constexpr (!Intf::preFusionFlag) {
363- self->ctx.queueAL1.FreeTensor(self->ctx.al1);376+ if (self->ctx.kAL1fullload) {
377+ self->ctx.queueAL1.FreeTensor(self->ctx.al1);
378+ }
364 }379 }
365 }380 }
366 381 
@@ -386,7 +401,9 @@ __aicore__ inline bool IterateNFirstMMode(Intf *self)
386 self->ctx.batchIter++;401 self->ctx.batchIter++;
387 if (likely(self->ctx.batchIter != self->ctx.ddr2l1LoopBatch)) {402 if (likely(self->ctx.batchIter != self->ctx.ddr2l1LoopBatch)) {
388 self->ctx.loadAL1Flag = true;403 self->ctx.loadAL1Flag = true;
389- self->ctx.loadBL1Flag = true;404+ if constexpr (!(Intf::preFusionFlag && ConvParam::weightTiling == 0)) {
405+ self->ctx.loadBL1Flag = true;
406+ }
390 return true;407 return true;
391 }408 }
392 409 
Mconv/common/op_kernel/arch35/conv_opt_group_impl.h+17-17
@@ -49,10 +49,10 @@ struct SetIterIndex {
49};49};
50 50 
51template <class Intf>51template <class Intf>
52-__aicore__ inline bool OptGroupUpdateBL1(Intf *self)52+__aicore__ inline bool OptGroupUpdateBL1(Intf *self, const uint64_t &kIter)
53{53{
54- if (!((self->ctx.kIter == self->ctx.ddr2l0LoopK - 1) ||54+ if (!((kIter == self->ctx.ddr2l0LoopK - 1) ||
55- ((self->ctx.kIter + 1) % self->ctx.multiKBL1 == 0))) {55+ ((kIter + 1) % self->ctx.multiKBL1 == 0))) {
56 return false;56 return false;
57 }57 }
58 58 
@@ -78,9 +78,9 @@ __aicore__ inline bool OptGroupUpdateBL1(Intf *self)
78 }78 }
79 } else if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {79 } else if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {
80 if constexpr (Intf::iterateMFirstFlag) {80 if constexpr (Intf::iterateMFirstFlag) {
81- updateBL1Flag = self->ctx.mAL1Iter == self->ctx.maxMAL1Iter &&81+ updateBL1Flag = self->ctx.nL0Iter == self->ctx.maxNL0Iter &&
82- self->ctx.nL0Iter == self->ctx.maxNL0Iter &&
83 self->ctx.mL0Iter == self->ctx.maxML0Iter &&82 self->ctx.mL0Iter == self->ctx.maxML0Iter &&
83+ self->ctx.mAL1Iter == self->ctx.maxMAL1Iter &&
84 self->ctx.batchIter == self->ctx.ddr2l1LoopBatch - 1;84 self->ctx.batchIter == self->ctx.ddr2l1LoopBatch - 1;
85 } else {85 } else {
86 updateBL1Flag = self->ctx.nL0Iter == self->ctx.maxNL0Iter &&86 updateBL1Flag = self->ctx.nL0Iter == self->ctx.maxNL0Iter &&
@@ -96,9 +96,9 @@ __aicore__ inline bool OptGroupUpdateBL1(Intf *self)
96}96}
97 97 
98template <class Intf>98template <class Intf>
99-__aicore__ inline void OptGroupSyncSet(Intf *self)99+__aicore__ inline void OptGroupSyncSet(Intf *self, const uint64_t &kIter)
100{100{
101- if (!OptGroupUpdateBL1<Intf>(self)) {101+ if (!OptGroupUpdateBL1<Intf>(self, kIter)) {
102 return;102 return;
103 }103 }
104 104 
@@ -124,9 +124,9 @@ __aicore__ inline void OptGroupSyncSet(Intf *self)
124}124}
125 125 
126template <class Intf>126template <class Intf>
127-__aicore__ inline void OptGroupPreloadSyncSet(Intf *self)127+__aicore__ inline void OptGroupPreloadSyncSet(Intf *self, const uint64_t &kIter)
128{128{
129- if (!OptGroupUpdateBL1<Intf>(self)) {129+ if (!OptGroupUpdateBL1<Intf>(self, kIter)) {
130 return;130 return;
131 }131 }
132 132 
@@ -138,8 +138,8 @@ __aicore__ inline void OptGroupPreloadSyncSet(Intf *self)
138template <class Intf>138template <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.maxNBL1Iter == self->ctx.nBL1Iter ?141+ self->ctx.l12l0LoopN = self->ctx.nBL1Iter == self->ctx.maxNBL1Iter ?
142- CeilDiv(self->ctx.nBL1Tail, self->ctx.convTiling->nL0) : self->ctx.convTiling->multiNBL1;142+ CeilDiv(self->ctx.nBL1Tail, self->ctx.convTilingData->convApiTiling.nL0) : self->ctx.convTilingData->convApiTiling.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.convTiling->woL1;162+ self->ctx.currentWoL1 = self->ctx.convTilingData->convApiTiling.woL1;
163 }163 }
164 } else {164 } else {
165- self->ctx.currentWoL1 = self->ctx.maxWoL1Iter == self->ctx.woAL1Iter ?165+ self->ctx.currentWoL1 = self->ctx.woAL1Iter == self->ctx.maxWoL1Iter ?
166- self->ctx.woAL1Tail : self->ctx.convTiling->woL1;166+ self->ctx.woAL1Tail : self->ctx.convTilingData->convApiTiling.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.convTiling->woL0);170+ self->ctx.l12l0LoopW = CeilDiv(self->ctx.currentWoL1, self->ctx.convTilingData->convApiTiling.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.convTiling->hoL1;176+ self->ctx.hoAL1Tail : self->ctx.convTilingData->convApiTiling.hoL1;
177 }177 }
178 if constexpr (Intf::hasHL0IterFlag) {178 if constexpr (Intf::hasHL0IterFlag) {
179- self->ctx.l12l0LoopH = CeilDiv(self->ctx.currentHoL1, self->ctx.convTiling->hoL0);179+ self->ctx.l12l0LoopH = CeilDiv(self->ctx.currentHoL1, self->ctx.convTilingData->convApiTiling.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+58-59
@@ -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.convTiling->hoL0);32+ uint64_t ddr2l0LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.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.convTiling->woL1, self->ctx.convTiling->woL0) +37+ CeilDiv(self->ctx.convTilingData->convApiTiling.woL1, self->ctx.convTilingData->convApiTiling.woL0) +
38- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTiling->woL0) +38+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0) +
39- CeilDiv(self->ctx.woL1SmallTail, self->ctx.convTiling->woL0);39+ CeilDiv(self->ctx.woL1SmallTail, self->ctx.convTilingData->convApiTiling.woL0);
40 } else {40 } else {
41- ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTiling->woL0);41+ ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL0);
42 }42 }
43 } else {43 } else {
44- ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTiling->woL0);44+ ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.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 ddr2l0LoopW = 0;102 uint64_t ddr2l0LoopW = 0;
102- uint64_t ddr2l0LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTiling->hoL0);
103 if constexpr (Intf::hasWL0IterFlag) {103 if constexpr (Intf::hasWL0IterFlag) {
104- self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTiling->woL1;104+ self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->convApiTiling.woL1;
105- self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTiling->woL1 : self->ctx.woAL1Tail;105+ self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL1 : self->ctx.woAL1Tail;
106- if (self->ctx.convTiling->hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&106+ if (self->ctx.convTilingData->convApiTiling.hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&
107- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTiling->woL0) > 1) {107+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0) > 1) {
108- self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTiling->woL0) * self->ctx.convTiling->woL0;108+ self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->convApiTiling.woL0;
109- self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTiling->woL0;109+ self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->convApiTiling.woL0) * self->ctx.convTilingData->convApiTiling.woL0;
110 }110 }
111 111 
112- self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTiling->woL1);112+ self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.woL1);
113 ddr2l0LoopW = 113 ddr2l0LoopW =
114- (self->ctx.ddr2l1LoopW - 1) * CeilDiv(self->ctx.woAL1Tail, self->ctx.convTiling->woL0) +114+ (self->ctx.ddr2l1LoopW - 1) * CeilDiv(self->ctx.convTilingData->convApiTiling.woL1, self->ctx.convTilingData->convApiTiling.woL0) +
115- CeilDiv(self->ctx.convTiling->woL1, self->ctx.convTiling->woL0);115+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0);
116 116 
117 if (self->ctx.woL1SmallTail > 0) {117 if (self->ctx.woL1SmallTail > 0) {
118- ddr2l0LoopW += CeilDiv(self->ctx.woL1SmallTail, self->ctx.convTiling->woL0);118+ ddr2l0LoopW += CeilDiv(self->ctx.woL1SmallTail, self->ctx.convTilingData->convApiTiling.woL0);
119 }119 }
120 } else {120 } else {
121- ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTiling->woL0);121+ ddr2l0LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.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.convTiling->hoL1);138+ self->ctx.ddr2l1LoopH = CeilDiv(self->ctx.singleCoreHo, self->ctx.convTilingData->convApiTiling.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.convTiling->woL1);143+ self->ctx.ddr2l1LoopW = CeilDiv(self->ctx.singleCoreWo, self->ctx.convTilingData->convApiTiling.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.convTiling->woL1;148+ self->ctx.woAL1Tail = self->ctx.singleCoreWo % self->ctx.convTilingData->convApiTiling.woL1;
149- self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTiling->woL1 : self->ctx.woAL1Tail;149+ self->ctx.woAL1Tail = self->ctx.woAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.woL1 : self->ctx.woAL1Tail;
150 if constexpr (Intf::hasWL0IterFlag) {150 if constexpr (Intf::hasWL0IterFlag) {
151- if (self->ctx.convTiling->hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&151+ if (self->ctx.convTilingData->convApiTiling.hoL0 > 1 && self->ctx.woAL1Tail % BLOCK_L0_N > 0 &&
152- CeilDiv(self->ctx.woAL1Tail, self->ctx.convTiling->woL0) > 1) {152+ CeilDiv(self->ctx.woAL1Tail, self->ctx.convTilingData->convApiTiling.woL0) > 1) {
153- self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTiling->woL0;153+ self->ctx.woL1SmallTail = self->ctx.woAL1Tail % self->ctx.convTilingData->convApiTiling.woL0;
154- self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTiling->woL0) * self->ctx.convTiling->woL0;154+ self->ctx.woAL1Tail = (self->ctx.woAL1Tail / self->ctx.convTilingData->convApiTiling.woL0) * self->ctx.convTilingData->convApiTiling.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.convTiling->hoL1;163+ self->ctx.hoAL1Tail = self->ctx.singleCoreHo % self->ctx.convTilingData->convApiTiling.hoL1;
164- self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTiling->hoL1 : self->ctx.hoAL1Tail;164+ self->ctx.hoAL1Tail = self->ctx.hoAL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.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.convTiling->kernelHxkernelWxkernelD,228+ CeilDiv(AlignB(self->ctx.singleCoreCi, Intf::k0) * self->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD,
229- self->ctx.convTiling->kBL1);229+ self->ctx.convTilingData->convApiTiling.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.convTiling->cinBInCore / self->ctx.bL1Dk;236+ self->ctx.bL1Cin = self->ctx.convTilingData->convApiTiling.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.convTiling->nBL1;246+ self->ctx.nBL1Tail = self->ctx.singleCoreCo % self->ctx.convTilingData->convApiTiling.nBL1;
247- self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTiling->nBL1 : self->ctx.nBL1Tail;247+ self->ctx.nBL1Tail = self->ctx.nBL1Tail == 0 ? self->ctx.convTilingData->convApiTiling.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.convTiling->nBL1);253+ self->ctx.ddr2l1LoopN = CeilDiv(self->ctx.singleCoreCo, self->ctx.convTilingData->convApiTiling.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.convTiling->multiNBL1;260+ self->ctx.l12l0LoopN = self->ctx.convTilingData->convApiTiling.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.convTiling->kernelHxkernelWxkernelD * self->ctx.co1Opt *267+ self->ctx.ubBufSize = self->ctx.ci1Opt * self->ctx.convTilingData->convApiTiling.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);
@@ -272,12 +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- int8_t al1db = (self->ctx.convTiling->pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET;275+ uint32_t aL1SpaceSize = self->ctx.convTilingData->convApiTiling.aL1SpaceSize;
276- uint32_t aL1SpaceSize = self->ctx.convTiling->aL1SpaceSize;276+ if ((self->ctx.convTilingData->convApiTiling.pBufferFlag & AL1_DB_IDX) >> AL1_DB_OFFSET) {
277- if (al1db) {
278 aL1SpaceSize *= DOUBLE_BUF;277 aL1SpaceSize *= DOUBLE_BUF;
279 }278 }
280- self->ctx.bL1SpaceSize = self->ctx.convTiling->nBL1 * self->ctx.convTiling->kBL1;279+ self->ctx.bL1SpaceSize = self->ctx.convTilingData->convApiTiling.nBL1 * self->ctx.convTilingData->convApiTiling.kBL1;
281 280 
282 if constexpr (Intf::bL1DBFlag) {281 if constexpr (Intf::bL1DBFlag) {
283 self->ctx.pipe.InitBuffer(self->ctx.bL1TBuf,282 self->ctx.pipe.InitBuffer(self->ctx.bL1TBuf,
@@ -295,37 +294,37 @@ __aicore__ inline void OptGroupVecInit(Intf *self)
295{294{
296 self->ctx.vecId = GetSubBlockIdx();295 self->ctx.vecId = GetSubBlockIdx();
297 296 
298- self->ctx.singleCoreCi = self->ctx.convTiling->singleCoreCi;297+ self->ctx.singleCoreCi = self->ctx.convTilingData->convApiTiling.singleCoreCi;
299- self->ctx.singleCoreCo = self->ctx.convTiling->singleCoreCo;298+ self->ctx.singleCoreCo = self->ctx.convTilingData->convApiTiling.singleCoreCo;
300- self->ctx.singleGroups = self->ctx.convTiling->singleCoreGroups;299+ self->ctx.singleGroups = self->ctx.convTilingData->convApiTiling.singleCoreGroups;
301- self->ctx.singleGroupOpt = self->ctx.convTiling->singleCoreGroupOpt;300+ self->ctx.singleGroupOpt = self->ctx.convTilingData->convApiTiling.singleCoreGroupOpt;
302 301 
303- self->ctx.enlarge = self->ctx.convTiling->enlarge;302+ self->ctx.enlarge = self->ctx.convTilingData->convApiTiling.enlarge;
304- self->ctx.ciPerGroup = self->ctx.convTiling->orgCi / self->ctx.convTiling->groups;303+ self->ctx.ciPerGroup = self->ctx.convTilingData->convApiTiling.orgCi / self->ctx.convTilingData->convApiTiling.groups;
305- self->ctx.coPerGroup = self->ctx.convTiling->orgCo / self->ctx.convTiling->groups;304+ self->ctx.coPerGroup = self->ctx.convTilingData->convApiTiling.orgCo / self->ctx.convTilingData->convApiTiling.groups;
306- self->ctx.ciOpt = self->ctx.ciPerGroup * self->ctx.convTiling->enlarge;305+ self->ctx.ciOpt = self->ctx.ciPerGroup * self->ctx.convTilingData->convApiTiling.enlarge;
307 self->ctx.ci1Opt = CeilDiv(self->ctx.ciOpt, Intf::k0);306 self->ctx.ci1Opt = CeilDiv(self->ctx.ciOpt, Intf::k0);
308 self->ctx.ciOptAlign = self->ctx.ci1Opt * Intf::k0;307 self->ctx.ciOptAlign = self->ctx.ci1Opt * Intf::k0;
309- self->ctx.kUbSize = self->ctx.ciOptAlign * self->ctx.convTiling->kernelHxkernelWxkernelD;308+ self->ctx.kUbSize = self->ctx.ciOptAlign * self->ctx.convTilingData->convApiTiling.kernelHxkernelWxkernelD;
310- self->ctx.coOpt = self->ctx.coPerGroup * self->ctx.convTiling->enlarge;309+ self->ctx.coOpt = self->ctx.coPerGroup * self->ctx.convTilingData->convApiTiling.enlarge;
311 self->ctx.co1Opt = CeilDiv(self->ctx.coOpt, BLOCK_L0_N);310 self->ctx.co1Opt = CeilDiv(self->ctx.coOpt, BLOCK_L0_N);
312 self->ctx.coOptAlign = self->ctx.co1Opt * BLOCK_L0_N;311 self->ctx.coOptAlign = self->ctx.co1Opt * BLOCK_L0_N;
313 312 
314 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {313 if constexpr (Intf::outputOrder == static_cast<int8_t>(ConvOutputOrder::M_MODE)) {
315- self->ctx.singleCoreM = self->ctx.convTiling->singleCoreHo;314+ self->ctx.singleCoreM = self->ctx.convTilingData->convApiTiling.singleCoreHo;
316- self->ctx.mAL1 = self->ctx.convTiling->hoL1;315+ self->ctx.mAL1 = self->ctx.convTilingData->convApiTiling.hoL1;
317- self->ctx.mL0 = self->ctx.convTiling->hoL0;316+ self->ctx.mL0 = self->ctx.convTilingData->convApiTiling.hoL0;
318 } else {317 } else {
319- self->ctx.singleCoreHo = self->ctx.convTiling->singleCoreHo;318+ self->ctx.singleCoreHo = self->ctx.convTilingData->convApiTiling.singleCoreHo;
320- self->ctx.singleCoreWo = self->ctx.convTiling->singleCoreWo;319+ self->ctx.singleCoreWo = self->ctx.convTilingData->convApiTiling.singleCoreWo;
321 }320 }
322 321 
323 if constexpr (Intf::isConv3D) {322 if constexpr (Intf::isConv3D) {
324- self->ctx.singleCoreDo = self->ctx.convTiling->singleCoreDo;323+ self->ctx.singleCoreDo = self->ctx.convTilingData->convApiTiling.singleCoreDo;
325 self->ctx.cin1xcin0 = AlignB(self->ctx.singleCoreCi, Intf::k0);324 self->ctx.cin1xcin0 = AlignB(self->ctx.singleCoreCi, Intf::k0);
326- self->ctx.bL1Dk = self->ctx.convTiling->cinBInCore <= self->ctx.cin1xcin0 ?325+ self->ctx.bL1Dk = self->ctx.convTilingData->convApiTiling.cinBInCore <= self->ctx.cin1xcin0 ?
327- 1 : self->ctx.convTiling->cinBInCore / self->ctx.cin1xcin0;326+ 1 : self->ctx.convTilingData->convApiTiling.cinBInCore / self->ctx.cin1xcin0;
328- self->ctx.bL1DkTail = self->ctx.kernelD % self->ctx.bL1Dk;327+ self->ctx.bL1DkTail = self->ctx.convTilingData->convApiTiling.kernelD % self->ctx.bL1Dk;
329 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;
330 }329 }
331 330 
Mconv/common/op_kernel/arch35/conv_tilingkey.h+10-0
@@ -62,6 +62,16 @@ namespace ConvKey {
62#define CONV_DIS_CONTINUOUS_CLOSE 062#define CONV_DIS_CONTINUOUS_CLOSE 0
63#define CONV_DIS_CONTINUOUS_INPUT_HWNC 163#define CONV_DIS_CONTINUOUS_INPUT_HWNC 1
64 64 
65+#define CONV_HAS_PAD 0
66+#define CONV_NO_PAD 1
67+ 
68+#define CONV_MULTI_BATCH 0
69+#define CONV_ONE_BATCH 1
70+ 
71+#define CONV_NOT_SMALL_WEIGHT 0
72+#define CONV_FULLLOAD_KL1_NL0 1
73+#define CONV_WEIGHT_SMALLER_THAN_BL0 2
74+ 
65#define CONV_COMMON_ONLY_MN_FULLLOAD_SEL() \75#define CONV_COMMON_ONLY_MN_FULLLOAD_SEL() \
66ASCENDC_TPL_UINT_SEL(FmapTiling, ASCENDC_TPL_UI_LIST, \76ASCENDC_TPL_UINT_SEL(FmapTiling, ASCENDC_TPL_UI_LIST, \
67 CONV_FMAP_TILING_ONLY_M_FULLLOAD_AL1_AL0), \77 CONV_FMAP_TILING_ONLY_M_FULLLOAD_AL1_AL0), \
Mconv/common/op_kernel/arch35/conv_util.h+37-0
@@ -21,6 +21,43 @@
21namespace conv {21namespace conv {
22using namespace AscendC;22using namespace AscendC;
23 23 
24+// 新加bitmode
25+#pragma pack(push, 1) // 确保无填充，严格按位排列
26+typedef struct {
27+ uint64_t strideW : 6; // bits 0-5
28+ uint64_t strideH : 6; // bits 6-11
29+ uint64_t kernelW : 8; // bits 12-19
30+ uint64_t kernelH : 8; // bits 20-27
31+ uint64_t dilationW : 8; // bits 28-35
32+ uint64_t dilationH : 8; // bits 36-43
33+ uint64_t kernelW_highest_bit : 1; // bit 44
34+ uint64_t kernelH_highest_bit : 1; // bit 45
35+ uint64_t reserve1 : 2; // bits 46-47
36+ uint64_t channelSize : 16; // bits 48-64
37+} BitFieldXtData;
38+#pragma pack(pop)
39+
40+#pragma pack(push, 1) // 确保无填充，严格按位排列
41+typedef struct {
42+ uint64_t currentKL0 : 16; // bits 0-15
43+ uint64_t mExtension_ : 16; // bits 16-31
44+ uint64_t posK : 16; // bits 32-47
45+ uint64_t mStartPt_ : 16; // bits 48-63
46+} BitFieldXmData;
47+#pragma pack(pop)
48+
49+union UnionDataXt {
50+ uint64_t n;
51+ BitFieldXtData bf;
52+ __aicore__ inline UnionDataXt() : n(0) {}
53+};
54+
55+union UnionDataXm {
56+ uint64_t n;
57+ BitFieldXmData bf;
58+ __aicore__ inline UnionDataXm() : n(0) {}
59+};
60+ 
24const static uint64_t L0A_SIZE = 65536;61const static uint64_t L0A_SIZE = 65536;
25const static uint64_t L0B_SIZE = 65536;62const static uint64_t L0B_SIZE = 65536;
26const static uint64_t L0C_SIZE = 262144;63const static uint64_t L0C_SIZE = 262144;
Dconv/conv2d_v2/op_host/conv2d_v2_mdc_def.cpp+0-92
@@ -1,92 +0,0 @@
1-/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3- * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4- * CANN Open Software License Agreement Version 2.0 (the "License").
5- * Please refer to the License for details. You may not use this file except in compliance with the License.
6- * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7- * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8- * See LICENSE in the root of the software repository for the full text of the License.
9- */
10- 
11-/*!
12- * \file conv2d_mdc_def.cpp
13- * \brief
14- */
15- 
16-#include <cstdint>
17-#include <cstring>
18-#include "register/op_config_registry.h"
19- 
20-namespace ops {
21-std::vector<ge::DataType> conv2dv2FmapDataTypeList = {
22- ge::DT_FLOAT16, ge::DT_FLOAT16, ge::DT_FLOAT16, ge::DT_FLOAT16
23-};
24-std::vector<ge::DataType> conv2dv2WeightDataTypeList = {
25- ge::DT_FLOAT16, ge::DT_FLOAT16, ge::DT_FLOAT16, ge::DT_FLOAT16
26-};
27-std::vector<ge::DataType> conv2dv2BiasDataTypeList = {
28- ge::DT_FLOAT16, ge::DT_FLOAT16, ge::DT_FLOAT16, ge::DT_FLOAT16
29-};
30-std::vector<ge::DataType> conv2dv2OffsetWDataTypeList = {
31- ge::DT_INT8, ge::DT_INT8, ge::DT_INT8, ge::DT_INT8
32-};
33-std::vector<ge::DataType> conv2dv2OutputDataTypeList = {
34- ge::DT_FLOAT16, ge::DT_FLOAT16, ge::DT_FLOAT16, ge::DT_FLOAT16
35-};
36-std::vector<ge::Format> conv2dV2FmapFormatList = {
37- ge::FORMAT_NCHW, ge::FORMAT_NCHW, ge::FORMAT_NHWC, ge::FORMAT_NHWC
38-};
39-std::vector<ge::Format> conv2dV2WeightFormatList = {
40- ge::FORMAT_FRACTAL_Z, ge::FORMAT_FRACTAL_Z_C04, ge::FORMAT_FRACTAL_Z, ge::FORMAT_FRACTAL_Z_C04
41-};
42-std::vector<ge::Format> conv2dV2BiasFormatList = {
43- ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND
44-};
45-std::vector<ge::Format> conv2dV2OffsetWFormatList = {
46- ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND
47-};
48-std::vector<ge::Format> conv2dV2OutputFormatList = {
49- ge::FORMAT_NCHW, ge::FORMAT_NCHW, ge::FORMAT_NHWC, ge::FORMAT_NHWC
50-};
51- 
52-REGISTER_OP_AICORE_CONFIG(Conv2DV2, mc62cm12a, []() {
53- OpAICoreConfig aicoreConfig("mc62cm12a");
54- aicoreConfig.Input("x")
55- .ParamType(REQUIRED)
56- .DataType(conv2dv2FmapDataTypeList)
57- .Format(conv2dV2FmapFormatList)
58- .UnknownShapeFormat(conv2dV2FmapFormatList);
59- aicoreConfig.Input("filter")
60- .ParamType(REQUIRED)
61- .DataType(conv2dv2WeightDataTypeList)
62- .Format(conv2dV2WeightFormatList)
63- .UnknownShapeFormat(conv2dV2WeightFormatList);
64- aicoreConfig.Input("bias")
65- .ParamType(OPTIONAL)
66- .DataType(conv2dv2BiasDataTypeList)
67- .Format(conv2dV2BiasFormatList)
68- .UnknownShapeFormat(conv2dV2BiasFormatList);
69- aicoreConfig.Input("offset_w")
70- .ParamType(OPTIONAL)
71- .DataType(conv2dv2OffsetWDataTypeList)
72- .Format(conv2dV2OffsetWFormatList)
73- .UnknownShapeFormat(conv2dV2OffsetWFormatList);
74- aicoreConfig.Output("y")
75- .ParamType(REQUIRED)
76- .DataType(conv2dv2OutputDataTypeList)
77- .Format(conv2dV2OutputFormatList)
78- .UnknownShapeFormat(conv2dV2OutputFormatList);
79- aicoreConfig.DynamicCompileStaticFlag(true)
80- .DynamicFormatFlag(true)
81- .DynamicRankSupportFlag(true)
82- .DynamicShapeSupportFlag(true)
83- .NeedCheckSupportFlag(false)
84- .PrecisionReduceFlag(true)
85- .ExtendCfgInfo("opFile.value", "conv2d_v2")
86- .ExtendCfgInfo("opInterface.value", "conv2dv2")
87- .ExtendCfgInfo("aclnnSupport.value", "support_aclnn")
88- .ExtendCfgInfo("jitCompile.flag", "false");
89- 
90- return aicoreConfig;
91-});
92-}
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_api_tiling.cpp+9-1
@@ -425,7 +425,7 @@ void Conv2dTiling::SetAttrsTilingData(optiling::TConv2DTiling& tiling)
425 tiling.set_padBottom(static_cast<uint32_t>(attrInfo.padBottom));425 tiling.set_padBottom(static_cast<uint32_t>(attrInfo.padBottom));
426 tiling.set_padLeft(static_cast<uint32_t>(attrInfo.padLeft));426 tiling.set_padLeft(static_cast<uint32_t>(attrInfo.padLeft));
427 tiling.set_padRight(static_cast<uint32_t>(attrInfo.padRight));427 tiling.set_padRight(static_cast<uint32_t>(attrInfo.padRight));
428- 428+ tiling.set_unionDataXt(static_cast<uint64_t>(attrInfo.unionDataXt.n));
429 tiling.set_groups(static_cast<uint32_t>(attrInfo.groups));429 tiling.set_groups(static_cast<uint32_t>(attrInfo.groups));
430 430 
431 if (this->optGroupFlag) {431 if (this->optGroupFlag) {
@@ -492,6 +492,10 @@ void Conv2dTiling::SetOrgWeightShape(int64_t orgCo, int64_t orgkH, int64_t orgkW
492 shapeInfo.orgkH = orgkH;492 shapeInfo.orgkH = orgkH;
493 shapeInfo.orgkW = orgkW;493 shapeInfo.orgkW = orgkW;
494 shapeInfo.orgCo = orgCo;494 shapeInfo.orgCo = orgCo;
495+ attrInfo.unionDataXt.bf.kernelW = static_cast<uint64_t>(orgkW);
496+ attrInfo.unionDataXt.bf.kernelW_highest_bit = (static_cast<uint64_t>(orgkW) & 0x100) >> 8; // 8 kernelW lower 8 bit
497+ attrInfo.unionDataXt.bf.kernelH = static_cast<uint64_t>(orgkH);
498+ attrInfo.unionDataXt.bf.kernelH_highest_bit = (static_cast<uint64_t>(orgkH) & 0x100) >> 8; // 8 kernelH lower 8 bit
495}499}
496 500 
497void Conv2dTiling::SetSingleWeightShape(int64_t singleCi, int64_t singlekH, int64_t singlekW)501void Conv2dTiling::SetSingleWeightShape(int64_t singleCi, int64_t singlekH, int64_t singlekW)
@@ -549,12 +553,16 @@ void Conv2dTiling::SetDilation(int32_t dilationH, int32_t dilationW)
549{553{
550 attrInfo.dilationH = dilationH;554 attrInfo.dilationH = dilationH;
551 attrInfo.dilationW = dilationW;555 attrInfo.dilationW = dilationW;
556+ attrInfo.unionDataXt.bf.dilationH = static_cast<uint64_t>(dilationH);
557+ attrInfo.unionDataXt.bf.dilationW = static_cast<uint64_t>(dilationW);
552}558}
553 559 
554void Conv2dTiling::SetStride(int32_t strideH, int32_t strideW)560void Conv2dTiling::SetStride(int32_t strideH, int32_t strideW)
555{561{
556 attrInfo.strideH = strideH;562 attrInfo.strideH = strideH;
557 attrInfo.strideW = strideW;563 attrInfo.strideW = strideW;
564+ attrInfo.unionDataXt.bf.strideH = static_cast<uint64_t>(strideH) & 0x3f;
565+ attrInfo.unionDataXt.bf.strideW = static_cast<uint64_t>(strideW) & 0x3f;
558}566}
559 567 
560void Conv2dTiling::SetBiasType(TPosition pos, ConvFormat format, ConvDtype dtype)568void Conv2dTiling::SetBiasType(TPosition pos, ConvFormat format, ConvDtype dtype)
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_api_tilingdata.h+1-0
@@ -29,6 +29,7 @@ TILING_DATA_FIELD_DEF(uint64_t, orgHixWi);
29TILING_DATA_FIELD_DEF(uint64_t, singleCoreBatch);29TILING_DATA_FIELD_DEF(uint64_t, singleCoreBatch);
30TILING_DATA_FIELD_DEF(uint64_t, singleCoreHo);30TILING_DATA_FIELD_DEF(uint64_t, singleCoreHo);
31TILING_DATA_FIELD_DEF(uint64_t, singleCoreWo);31TILING_DATA_FIELD_DEF(uint64_t, singleCoreWo);
32+TILING_DATA_FIELD_DEF(uint64_t, unionDataXt);
32TILING_DATA_FIELD_DEF(uint32_t, orgCi);33TILING_DATA_FIELD_DEF(uint32_t, orgCi);
33TILING_DATA_FIELD_DEF(uint32_t, orgCo);34TILING_DATA_FIELD_DEF(uint32_t, orgCo);
34TILING_DATA_FIELD_DEF(uint32_t, singleCoreCi);35TILING_DATA_FIELD_DEF(uint32_t, singleCoreCi);
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling.cpp+5-5
@@ -446,12 +446,12 @@ ge::graphStatus Conv2dBaseTiling::GetFeatureFlag()
446 446 
447void Conv2dBaseTiling::SetNumBlocksRes()447void Conv2dBaseTiling::SetNumBlocksRes()
448{448{
449- numBlocksRes.batchDim = tilingData_.conv2dRunInfo.get_batchDim();449+ numBlocksRes.batchDim = tilingData_.convRunInfo.get_batchDim();
450- numBlocksRes.nDim = tilingData_.conv2dRunInfo.get_nDim();450+ numBlocksRes.nDim = tilingData_.convRunInfo.get_nDim();
451- numBlocksRes.hoDim = tilingData_.conv2dRunInfo.get_hoDim();451+ numBlocksRes.hoDim = tilingData_.convRunInfo.get_hoDim();
452- numBlocksRes.woDim = tilingData_.conv2dRunInfo.get_woDim();452+ numBlocksRes.woDim = tilingData_.convRunInfo.get_woDim();
453 numBlocksRes.mDim = numBlocksRes.hoDim;453 numBlocksRes.mDim = numBlocksRes.hoDim;
454- numBlocksRes.groupDim = tilingData_.conv2dRunInfo.get_groupDim();454+ numBlocksRes.groupDim = tilingData_.convRunInfo.get_groupDim();
455}455}
456 456 
457// reset conv2d API's tilingdata457// reset conv2d API's tilingdata
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling.h+4-0
@@ -207,6 +207,8 @@ private:
207 uint64_t GetFmapCopyMode();207 uint64_t GetFmapCopyMode();
208 uint64_t GetEnableInnerBatch();208 uint64_t GetEnableInnerBatch();
209 void ReSetTilingKeyPara();209 void ReSetTilingKeyPara();
210+ uint64_t GetNoPad() const;
211+ uint64_t GetSmallWeightVal();
210 212 
211 int32_t BasicBlockSortFWDimScores(vector<tuple<uint64_t, float, uint32_t, double>>& scores);213 int32_t BasicBlockSortFWDimScores(vector<tuple<uint64_t, float, uint32_t, double>>& scores);
212 ge::graphStatus GetConv2DAxisPosInfo();214 ge::graphStatus GetConv2DAxisPosInfo();
@@ -240,6 +242,8 @@ private:
240 bool CheckScaleLegal(uint32_t scaleIndex, uint8_t& quantMode, const std::string& scaleType);242 bool CheckScaleLegal(uint32_t scaleIndex, uint8_t& quantMode, const std::string& scaleType);
241 ge::Format GetWeightFormat() const;243 ge::Format GetWeightFormat() const;
242 gert::Shape GetWeightShape(const gert::StorageShape* weightShapePtr) const;244 gert::Shape GetWeightShape(const gert::StorageShape* weightShapePtr) const;
245+ void SetUnionDataXt(shared_ptr<tuningtiling::Conv2DV2TunnerTiling> convRepoTiling);
246+ void PrintInputArgs(shared_ptr<tuningtiling::Conv2DV2InputArgs> conv2DInput);
243};247};
244}248}
245}249}
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_basic_block.cpp+7-2
@@ -402,8 +402,13 @@ void Conv2dBaseTiling::GetInitBasicBlockMN(uint64_t& basicBlockM, uint64_t& basi
402{402{
403 uint64_t howo = shapeInfo_.ho * shapeInfo_.wo;403 uint64_t howo = shapeInfo_.ho * shapeInfo_.wo;
404 if (shapeInfo_.co <= BASICBLOCK_BOUNDARY_VALUE_64) {404 if (shapeInfo_.co <= BASICBLOCK_BOUNDARY_VALUE_64) {
405- basicBlockM = BASICBLOCK_INIT_VALUE_1024;405+ if (descInfo_.fMapDtype == ge::DT_FLOAT16 && descInfo_.weightDtype == ge::DT_INT8) {
406- basicBlockN = BASICBLOCK_INIT_VALUE_64;406+ basicBlockM = BASICBLOCK_INIT_VALUE_512;
407+ basicBlockN = BASICBLOCK_INIT_VALUE_64;
408+ } else {
409+ basicBlockM = BASICBLOCK_INIT_VALUE_1024;
410+ basicBlockN = BASICBLOCK_INIT_VALUE_64;
411+ }
407 } else if (shapeInfo_.co > BASICBLOCK_BOUNDARY_VALUE_64412 } else if (shapeInfo_.co > BASICBLOCK_BOUNDARY_VALUE_64
408 && shapeInfo_.co <= BASICBLOCK_BOUNDARY_VALUE_128) {413 && shapeInfo_.co <= BASICBLOCK_BOUNDARY_VALUE_128) {
409 basicBlockM = BASICBLOCK_INIT_VALUE_512;414 basicBlockM = BASICBLOCK_INIT_VALUE_512;
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_cache.cpp+158-156
@@ -63,67 +63,68 @@ bool Conv2dBaseTiling::AddTilingToCache()
63 63 
64void Conv2dBaseTiling::GetCachedTilingDataAux()64void Conv2dBaseTiling::GetCachedTilingDataAux()
65{65{
66- cachedTilingData_.fmapKStride = tilingData_.conv2dApiTiling.get_fmapKStride();66+ cachedTilingData_.fmapKStride = tilingData_.convApiTiling.get_fmapKStride();
67- cachedTilingData_.weightKStride = tilingData_.conv2dApiTiling.get_weightKStride();67+ cachedTilingData_.weightKStride = tilingData_.convApiTiling.get_weightKStride();
68- cachedTilingData_.cinOffsetBlockInGM = tilingData_.conv2dApiTiling.get_cinOffsetBlockInGM();68+ cachedTilingData_.cinOffsetBlockInGM = tilingData_.convApiTiling.get_cinOffsetBlockInGM();
69- cachedTilingData_.coutOffsetBlock = tilingData_.conv2dApiTiling.get_coutOffsetBlock();69+ cachedTilingData_.coutOffsetBlock = tilingData_.convApiTiling.get_coutOffsetBlock();
70- cachedTilingData_.nL1DivBlockSize = tilingData_.conv2dApiTiling.get_nL1DivBlockSize();70+ cachedTilingData_.nL1DivBlockSize = tilingData_.convApiTiling.get_nL1DivBlockSize();
71- cachedTilingData_.iterateMNOrder = tilingData_.conv2dApiTiling.get_iterateMNOrder();71+ cachedTilingData_.iterateMNOrder = tilingData_.convApiTiling.get_iterateMNOrder();
72- cachedTilingData_.biasFullLoadFlag = tilingData_.conv2dApiTiling.get_biasFullLoadFlag();72+ cachedTilingData_.biasFullLoadFlag = tilingData_.convApiTiling.get_biasFullLoadFlag();
73- cachedTilingData_.fixpParamsFullLoadFlag = tilingData_.conv2dApiTiling.get_fixpParamsFullLoadFlag();73+ cachedTilingData_.fixpParamsFullLoadFlag = tilingData_.convApiTiling.get_fixpParamsFullLoadFlag();
74- cachedTilingData_.hf32Enable = tilingData_.conv2dApiTiling.get_hf32Enable();74+ cachedTilingData_.hf32Enable = tilingData_.convApiTiling.get_hf32Enable();
75- cachedTilingData_.hf32TransMode = tilingData_.conv2dApiTiling.get_hf32TransMode();75+ cachedTilingData_.hf32TransMode = tilingData_.convApiTiling.get_hf32TransMode();
76- cachedTilingData_.batchDim = tilingData_.conv2dRunInfo.get_batchDim();76+ cachedTilingData_.batchDim = tilingData_.convRunInfo.get_batchDim();
77- cachedTilingData_.groupDim = tilingData_.conv2dRunInfo.get_groupDim();77+ cachedTilingData_.groupDim = tilingData_.convRunInfo.get_groupDim();
78- cachedTilingData_.nDim = tilingData_.conv2dRunInfo.get_nDim();78+ cachedTilingData_.nDim = tilingData_.convRunInfo.get_nDim();
79- cachedTilingData_.hoDim = tilingData_.conv2dRunInfo.get_hoDim();79+ cachedTilingData_.hoDim = tilingData_.convRunInfo.get_hoDim();
80- cachedTilingData_.woDim = tilingData_.conv2dRunInfo.get_woDim();80+ cachedTilingData_.woDim = tilingData_.convRunInfo.get_woDim();
81- cachedTilingData_.cinOpt = tilingData_.conv2dRunInfo.get_cinOpt();81+ cachedTilingData_.cinOpt = tilingData_.convRunInfo.get_cinOpt();
82- cachedTilingData_.coutOpt = tilingData_.conv2dRunInfo.get_coutOpt();82+ cachedTilingData_.coutOpt = tilingData_.convRunInfo.get_coutOpt();
83- cachedTilingData_.groupOpt = tilingData_.conv2dRunInfo.get_groupOpt();83+ cachedTilingData_.groupOpt = tilingData_.convRunInfo.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}87}
87 88 
88void Conv2dBaseTiling::GetCachedTilingData()89void Conv2dBaseTiling::GetCachedTilingData()
89{90{
90- cachedTilingData_.singleCoreBatch = tilingData_.conv2dApiTiling.get_singleCoreBatch();91+ cachedTilingData_.singleCoreBatch = tilingData_.convApiTiling.get_singleCoreBatch();
91- cachedTilingData_.singleCoreHo = tilingData_.conv2dApiTiling.get_singleCoreHo();92+ cachedTilingData_.singleCoreHo = tilingData_.convApiTiling.get_singleCoreHo();
92- cachedTilingData_.singleCoreWo = tilingData_.conv2dApiTiling.get_singleCoreWo();93+ cachedTilingData_.singleCoreWo = tilingData_.convApiTiling.get_singleCoreWo();
93- cachedTilingData_.singleCoreCi = tilingData_.conv2dApiTiling.get_singleCoreCi();94+ cachedTilingData_.singleCoreCi = tilingData_.convApiTiling.get_singleCoreCi();
94- cachedTilingData_.singleCoreCo = tilingData_.conv2dApiTiling.get_singleCoreCo();95+ cachedTilingData_.singleCoreCo = tilingData_.convApiTiling.get_singleCoreCo();
95- cachedTilingData_.hoL1 = tilingData_.conv2dApiTiling.get_hoL1();96+ cachedTilingData_.hoL1 = tilingData_.convApiTiling.get_hoL1();
96- cachedTilingData_.woL1 = tilingData_.conv2dApiTiling.get_woL1();97+ cachedTilingData_.woL1 = tilingData_.convApiTiling.get_woL1();
97- cachedTilingData_.kAL1 = tilingData_.conv2dApiTiling.get_kAL1();98+ cachedTilingData_.kAL1 = tilingData_.convApiTiling.get_kAL1();
98- cachedTilingData_.kBL1 = tilingData_.conv2dApiTiling.get_kBL1();99+ cachedTilingData_.kBL1 = tilingData_.convApiTiling.get_kBL1();
99- cachedTilingData_.khL1 = tilingData_.conv2dApiTiling.get_khL1();100+ cachedTilingData_.khL1 = tilingData_.convApiTiling.get_khL1();
100- cachedTilingData_.kwL1 = tilingData_.conv2dApiTiling.get_kwL1();101+ cachedTilingData_.kwL1 = tilingData_.convApiTiling.get_kwL1();
101- cachedTilingData_.nBL1 = tilingData_.conv2dApiTiling.get_nBL1();102+ cachedTilingData_.nBL1 = tilingData_.convApiTiling.get_nBL1();
102- cachedTilingData_.hoL0 = tilingData_.conv2dApiTiling.get_hoL0();103+ cachedTilingData_.hoL0 = tilingData_.convApiTiling.get_hoL0();
103- cachedTilingData_.woL0 = tilingData_.conv2dApiTiling.get_woL0();104+ cachedTilingData_.woL0 = tilingData_.convApiTiling.get_woL0();
104- cachedTilingData_.kL0 = tilingData_.conv2dApiTiling.get_kL0();105+ cachedTilingData_.kL0 = tilingData_.convApiTiling.get_kL0();
105- cachedTilingData_.nL0 = tilingData_.conv2dApiTiling.get_nL0();106+ cachedTilingData_.nL0 = tilingData_.convApiTiling.get_nL0();
106- cachedTilingData_.pBufferFlag = tilingData_.conv2dApiTiling.get_pBufferFlag();107+ cachedTilingData_.pBufferFlag = tilingData_.convApiTiling.get_pBufferFlag();
107- cachedTilingData_.enlarge = tilingData_.conv2dApiTiling.get_enlarge();108+ cachedTilingData_.enlarge = tilingData_.convApiTiling.get_enlarge();
108- cachedTilingData_.singleCoreGroups = tilingData_.conv2dApiTiling.get_singleCoreGroups();109+ cachedTilingData_.singleCoreGroups = tilingData_.convApiTiling.get_singleCoreGroups();
109- cachedTilingData_.singleCoreGroupOpt = tilingData_.conv2dApiTiling.get_singleCoreGroupOpt();110+ cachedTilingData_.singleCoreGroupOpt = tilingData_.convApiTiling.get_singleCoreGroupOpt();
110- cachedTilingData_.bUbNStep = tilingData_.conv2dApiTiling.get_bUbNStep();111+ cachedTilingData_.bUbNStep = tilingData_.convApiTiling.get_bUbNStep();
111- cachedTilingData_.bUbKStep = tilingData_.conv2dApiTiling.get_bUbKStep();112+ cachedTilingData_.bUbKStep = tilingData_.convApiTiling.get_bUbKStep();
112- cachedTilingData_.khUb = tilingData_.conv2dApiTiling.get_khUb();113+ cachedTilingData_.khUb = tilingData_.convApiTiling.get_khUb();
113- cachedTilingData_.kwUb = tilingData_.conv2dApiTiling.get_kwUb();114+ cachedTilingData_.kwUb = tilingData_.convApiTiling.get_kwUb();
114- cachedTilingData_.orgHixWi = tilingData_.conv2dApiTiling.get_orgHixWi();115+ cachedTilingData_.orgHixWi = tilingData_.convApiTiling.get_orgHixWi();
115- cachedTilingData_.kernelHxkernelW = tilingData_.conv2dApiTiling.get_kernelHxkernelW();116+ cachedTilingData_.kernelHxkernelW = tilingData_.convApiTiling.get_kernelHxkernelW();
116- cachedTilingData_.kernelHxkernelWxkernelD = tilingData_.conv2dApiTiling.get_kernelHxkernelWxkernelD();117+ cachedTilingData_.kernelHxkernelWxkernelD = tilingData_.convApiTiling.get_kernelHxkernelWxkernelD();
117- cachedTilingData_.aL1SpaceSize = tilingData_.conv2dApiTiling.get_aL1SpaceSize();118+ cachedTilingData_.aL1SpaceSize = tilingData_.convApiTiling.get_aL1SpaceSize();
118- cachedTilingData_.multiNBL1 = tilingData_.conv2dApiTiling.get_multiNBL1();119+ cachedTilingData_.multiNBL1 = tilingData_.convApiTiling.get_multiNBL1();
119- cachedTilingData_.cinAInCore = tilingData_.conv2dApiTiling.get_cinAInCore();120+ cachedTilingData_.cinAInCore = tilingData_.convApiTiling.get_cinAInCore();
120- cachedTilingData_.cinATailInCore = tilingData_.conv2dApiTiling.get_cinATailInCore();121+ cachedTilingData_.cinATailInCore = tilingData_.convApiTiling.get_cinATailInCore();
121- cachedTilingData_.cinBInCore = tilingData_.conv2dApiTiling.get_cinBInCore();122+ cachedTilingData_.cinBInCore = tilingData_.convApiTiling.get_cinBInCore();
122- cachedTilingData_.cinBTailInCore = tilingData_.conv2dApiTiling.get_cinBTailInCore();123+ cachedTilingData_.cinBTailInCore = tilingData_.convApiTiling.get_cinBTailInCore();
123- cachedTilingData_.mStep = tilingData_.conv2dApiTiling.get_mStep();124+ cachedTilingData_.mStep = tilingData_.convApiTiling.get_mStep();
124- cachedTilingData_.kStep = tilingData_.conv2dApiTiling.get_kStep();125+ cachedTilingData_.kStep = tilingData_.convApiTiling.get_kStep();
125- cachedTilingData_.nStep = tilingData_.conv2dApiTiling.get_nStep();126+ cachedTilingData_.nStep = tilingData_.convApiTiling.get_nStep();
126- cachedTilingData_.innerBatch = tilingData_.conv2dApiTiling.get_innerBatch();127+ cachedTilingData_.innerBatch = tilingData_.convApiTiling.get_innerBatch();
127 GetCachedTilingDataAux();128 GetCachedTilingDataAux();
128}129}
129 130 
@@ -187,115 +188,116 @@ void Conv2dBaseTiling::GetCacheTilingInputArgs()
187 188 
188void Conv2dBaseTiling::TranslateCachedRunInfo()189void Conv2dBaseTiling::TranslateCachedRunInfo()
189{190{
190- tilingData_.conv2dRunInfo.set_batch(cacheInputArgs_.inputShapeN);191+ tilingData_.convRunInfo.set_batch(cacheInputArgs_.inputShapeN);
191- tilingData_.conv2dRunInfo.set_hin(cacheInputArgs_.inputShapeH);192+ tilingData_.convRunInfo.set_hin(cacheInputArgs_.inputShapeH);
192- tilingData_.conv2dRunInfo.set_win(cacheInputArgs_.inputShapeW);193+ tilingData_.convRunInfo.set_win(cacheInputArgs_.inputShapeW);
193- tilingData_.conv2dRunInfo.set_batchDim(cachedTilingData_.batchDim);194+ tilingData_.convRunInfo.set_batchDim(cachedTilingData_.batchDim);
194- tilingData_.conv2dRunInfo.set_hoDim(cachedTilingData_.hoDim);195+ tilingData_.convRunInfo.set_hoDim(cachedTilingData_.hoDim);
195- tilingData_.conv2dRunInfo.set_woDim(cachedTilingData_.woDim);196+ tilingData_.convRunInfo.set_woDim(cachedTilingData_.woDim);
196- tilingData_.conv2dRunInfo.set_nDim(cachedTilingData_.nDim);197+ tilingData_.convRunInfo.set_nDim(cachedTilingData_.nDim);
197- tilingData_.conv2dRunInfo.set_cin(shapeInfo_.ci);198+ tilingData_.convRunInfo.set_cin(shapeInfo_.ci);
198- tilingData_.conv2dRunInfo.set_cout(cacheInputArgs_.weightShapeN);199+ tilingData_.convRunInfo.set_cout(cacheInputArgs_.weightShapeN);
199- tilingData_.conv2dRunInfo.set_kh(cacheInputArgs_.weightShapeH);200+ tilingData_.convRunInfo.set_kh(cacheInputArgs_.weightShapeH);
200- tilingData_.conv2dRunInfo.set_kw(cacheInputArgs_.weightShapeW);201+ tilingData_.convRunInfo.set_kw(cacheInputArgs_.weightShapeW);
201- tilingData_.conv2dRunInfo.set_hout(cacheInputArgs_.outputShapeH);202+ tilingData_.convRunInfo.set_hout(cacheInputArgs_.outputShapeH);
202- tilingData_.conv2dRunInfo.set_wout(cacheInputArgs_.outputShapeW);203+ tilingData_.convRunInfo.set_wout(cacheInputArgs_.outputShapeW);
203- tilingData_.conv2dRunInfo.set_strideH(cacheInputArgs_.strideH);204+ tilingData_.convRunInfo.set_strideH(cacheInputArgs_.strideH);
204- tilingData_.conv2dRunInfo.set_strideW(cacheInputArgs_.strideW);205+ tilingData_.convRunInfo.set_strideW(cacheInputArgs_.strideW);
205- tilingData_.conv2dRunInfo.set_dilationH(cacheInputArgs_.dilationH);206+ tilingData_.convRunInfo.set_dilationH(cacheInputArgs_.dilationH);
206- tilingData_.conv2dRunInfo.set_dilationW(cacheInputArgs_.dilationW);207+ tilingData_.convRunInfo.set_dilationW(cacheInputArgs_.dilationW);
207- tilingData_.conv2dRunInfo.set_padTop(cacheInputArgs_.padTop);208+ tilingData_.convRunInfo.set_padTop(cacheInputArgs_.padTop);
208- tilingData_.conv2dRunInfo.set_padLeft(cacheInputArgs_.padLeft);209+ tilingData_.convRunInfo.set_padLeft(cacheInputArgs_.padLeft);
209- tilingData_.conv2dRunInfo.set_hasBias(cacheInputArgs_.biasFlag);210+ tilingData_.convRunInfo.set_hasBias(cacheInputArgs_.biasFlag);
210- tilingData_.conv2dRunInfo.set_groups(cacheInputArgs_.groups);211+ tilingData_.convRunInfo.set_groups(cacheInputArgs_.groups);
211- tilingData_.conv2dRunInfo.set_cinOpt(cachedTilingData_.cinOpt);212+ tilingData_.convRunInfo.set_cinOpt(cachedTilingData_.cinOpt);
212- tilingData_.conv2dRunInfo.set_coutOpt(cachedTilingData_.coutOpt);213+ tilingData_.convRunInfo.set_coutOpt(cachedTilingData_.coutOpt);
213- tilingData_.conv2dRunInfo.set_groupOpt(cachedTilingData_.groupOpt);214+ tilingData_.convRunInfo.set_groupOpt(cachedTilingData_.groupOpt);
214- tilingData_.conv2dRunInfo.set_enlarge(cachedTilingData_.enlarge);215+ tilingData_.convRunInfo.set_enlarge(cachedTilingData_.enlarge);
215- tilingData_.conv2dRunInfo.set_groupDim(cachedTilingData_.groupDim);216+ tilingData_.convRunInfo.set_groupDim(cachedTilingData_.groupDim);
216}217}
217 218 
218void Conv2dBaseTiling::TranslateCachedApiTilingPartTwo()219void Conv2dBaseTiling::TranslateCachedApiTilingPartTwo()
219{220{
220- tilingData_.conv2dApiTiling.set_strideH(cacheInputArgs_.strideH);221+ tilingData_.convApiTiling.set_strideH(cacheInputArgs_.strideH);
221- tilingData_.conv2dApiTiling.set_strideW(cacheInputArgs_.strideW);222+ tilingData_.convApiTiling.set_strideW(cacheInputArgs_.strideW);
222- tilingData_.conv2dApiTiling.set_dilationH(cacheInputArgs_.dilationH);223+ tilingData_.convApiTiling.set_dilationH(cacheInputArgs_.dilationH);
223- tilingData_.conv2dApiTiling.set_dilationW(cacheInputArgs_.dilationW);224+ tilingData_.convApiTiling.set_dilationW(cacheInputArgs_.dilationW);
224- tilingData_.conv2dApiTiling.set_padTop(cacheInputArgs_.padTop);225+ tilingData_.convApiTiling.set_padTop(cacheInputArgs_.padTop);
225- tilingData_.conv2dApiTiling.set_padBottom(cacheInputArgs_.padBottom);226+ tilingData_.convApiTiling.set_padBottom(cacheInputArgs_.padBottom);
226- tilingData_.conv2dApiTiling.set_padLeft(cacheInputArgs_.padLeft);227+ tilingData_.convApiTiling.set_padLeft(cacheInputArgs_.padLeft);
227- tilingData_.conv2dApiTiling.set_padRight(cacheInputArgs_.padRight);228+ tilingData_.convApiTiling.set_padRight(cacheInputArgs_.padRight);
228- tilingData_.conv2dApiTiling.set_iterateMNOrder(cachedTilingData_.iterateMNOrder);229+ tilingData_.convApiTiling.set_iterateMNOrder(cachedTilingData_.iterateMNOrder);
229- tilingData_.conv2dApiTiling.set_biasFullLoadFlag(cachedTilingData_.biasFullLoadFlag);230+ tilingData_.convApiTiling.set_biasFullLoadFlag(cachedTilingData_.biasFullLoadFlag);
230- tilingData_.conv2dApiTiling.set_fixpParamsFullLoadFlag(cachedTilingData_.fixpParamsFullLoadFlag);231+ tilingData_.convApiTiling.set_fixpParamsFullLoadFlag(cachedTilingData_.fixpParamsFullLoadFlag);
231- tilingData_.conv2dApiTiling.set_hf32Enable(cachedTilingData_.hf32Enable);232+ tilingData_.convApiTiling.set_hf32Enable(cachedTilingData_.hf32Enable);
232- tilingData_.conv2dApiTiling.set_hf32TransMode(cachedTilingData_.hf32TransMode);233+ tilingData_.convApiTiling.set_hf32TransMode(cachedTilingData_.hf32TransMode);
233- tilingData_.conv2dApiTiling.set_hasBias(cacheInputArgs_.biasFlag);234+ tilingData_.convApiTiling.set_hasBias(cacheInputArgs_.biasFlag);
234- tilingData_.conv2dApiTiling.set_hasScale(static_cast<uint8_t>(flagInfo_.quantFlag || flagInfo_.extendConvFlag));235+ tilingData_.convApiTiling.set_hasScale(static_cast<uint8_t>(flagInfo_.quantFlag || flagInfo_.extendConvFlag));
235- tilingData_.conv2dApiTiling.set_offsetx(attrInfo_.offsetx);236+ tilingData_.convApiTiling.set_offsetx(attrInfo_.offsetx);
236- tilingData_.conv2dApiTiling.set_roundMode(attrInfo_.roundMode);237+ tilingData_.convApiTiling.set_roundMode(attrInfo_.roundMode);
237- tilingData_.conv2dApiTiling.set_dualOutput(fixpipeInfo_.dualOutput);238+ tilingData_.convApiTiling.set_dualOutput(fixpipeInfo_.dualOutput);
238- tilingData_.conv2dApiTiling.set_quantMode0(fixpipeInfo_.quantMode0);239+ tilingData_.convApiTiling.set_quantMode0(fixpipeInfo_.quantMode0);
239- tilingData_.conv2dApiTiling.set_reluMode0(fixpipeInfo_.reluMode0);240+ tilingData_.convApiTiling.set_reluMode0(fixpipeInfo_.reluMode0);
240- tilingData_.conv2dApiTiling.set_clipMode0(fixpipeInfo_.clipMode0);241+ tilingData_.convApiTiling.set_clipMode0(fixpipeInfo_.clipMode0);
241- tilingData_.conv2dApiTiling.set_quantMode1(fixpipeInfo_.quantMode1);242+ tilingData_.convApiTiling.set_quantMode1(fixpipeInfo_.quantMode1);
242- tilingData_.conv2dApiTiling.set_reluMode1(fixpipeInfo_.reluMode1);243+ tilingData_.convApiTiling.set_reluMode1(fixpipeInfo_.reluMode1);
243- tilingData_.conv2dApiTiling.set_clipMode1(fixpipeInfo_.clipMode1);244+ tilingData_.convApiTiling.set_clipMode1(fixpipeInfo_.clipMode1);
244- tilingData_.conv2dApiTiling.set_innerBatch(cachedTilingData_.innerBatch);245+ tilingData_.convApiTiling.set_innerBatch(cachedTilingData_.innerBatch);
246+ tilingData_.convApiTiling.set_unionDataXt(cachedTilingData_.unionDataXt);
245}247}
246 248 
247void Conv2dBaseTiling::TranslateCachedApiTilingPartOne()249void Conv2dBaseTiling::TranslateCachedApiTilingPartOne()
248{250{
249- tilingData_.conv2dApiTiling.set_orgHi(cacheInputArgs_.inputShapeH);251+ tilingData_.convApiTiling.set_orgHi(cacheInputArgs_.inputShapeH);
250- tilingData_.conv2dApiTiling.set_orgWi(cacheInputArgs_.inputShapeW);252+ tilingData_.convApiTiling.set_orgWi(cacheInputArgs_.inputShapeW);
251- tilingData_.conv2dApiTiling.set_orgHo(cacheInputArgs_.outputShapeH);253+ tilingData_.convApiTiling.set_orgHo(cacheInputArgs_.outputShapeH);
252- tilingData_.conv2dApiTiling.set_orgWo(cacheInputArgs_.outputShapeW);254+ tilingData_.convApiTiling.set_orgWo(cacheInputArgs_.outputShapeW);
253- tilingData_.conv2dApiTiling.set_singleCoreBatch(cachedTilingData_.singleCoreBatch);255+ tilingData_.convApiTiling.set_singleCoreBatch(cachedTilingData_.singleCoreBatch);
254- tilingData_.conv2dApiTiling.set_singleCoreHo(cachedTilingData_.singleCoreHo);256+ tilingData_.convApiTiling.set_singleCoreHo(cachedTilingData_.singleCoreHo);
255- tilingData_.conv2dApiTiling.set_singleCoreWo(cachedTilingData_.singleCoreWo);257+ tilingData_.convApiTiling.set_singleCoreWo(cachedTilingData_.singleCoreWo);
256- tilingData_.conv2dApiTiling.set_orgCi(shapeInfo_.ci);258+ tilingData_.convApiTiling.set_orgCi(shapeInfo_.ci);
257- tilingData_.conv2dApiTiling.set_orgCo(cacheInputArgs_.weightShapeN);259+ tilingData_.convApiTiling.set_orgCo(cacheInputArgs_.weightShapeN);
258- tilingData_.conv2dApiTiling.set_singleCoreCi(cachedTilingData_.singleCoreCi);260+ tilingData_.convApiTiling.set_singleCoreCi(cachedTilingData_.singleCoreCi);
259- tilingData_.conv2dApiTiling.set_singleCoreCo(cachedTilingData_.singleCoreCo);261+ tilingData_.convApiTiling.set_singleCoreCo(cachedTilingData_.singleCoreCo);
260- tilingData_.conv2dApiTiling.set_hoL1(cachedTilingData_.hoL1);262+ tilingData_.convApiTiling.set_hoL1(cachedTilingData_.hoL1);
261- tilingData_.conv2dApiTiling.set_woL1(cachedTilingData_.woL1);263+ tilingData_.convApiTiling.set_woL1(cachedTilingData_.woL1);
262- tilingData_.conv2dApiTiling.set_kAL1(cachedTilingData_.kAL1);264+ tilingData_.convApiTiling.set_kAL1(cachedTilingData_.kAL1);
263- tilingData_.conv2dApiTiling.set_kBL1(cachedTilingData_.kBL1);265+ tilingData_.convApiTiling.set_kBL1(cachedTilingData_.kBL1);
264- tilingData_.conv2dApiTiling.set_khL1(cachedTilingData_.khL1);266+ tilingData_.convApiTiling.set_khL1(cachedTilingData_.khL1);
265- tilingData_.conv2dApiTiling.set_kwL1(cachedTilingData_.kwL1);267+ tilingData_.convApiTiling.set_kwL1(cachedTilingData_.kwL1);
266- tilingData_.conv2dApiTiling.set_nBL1(cachedTilingData_.nBL1);268+ tilingData_.convApiTiling.set_nBL1(cachedTilingData_.nBL1);
267- tilingData_.conv2dApiTiling.set_hoL0(cachedTilingData_.hoL0);269+ tilingData_.convApiTiling.set_hoL0(cachedTilingData_.hoL0);
268- tilingData_.conv2dApiTiling.set_woL0(cachedTilingData_.woL0);270+ tilingData_.convApiTiling.set_woL0(cachedTilingData_.woL0);
269- tilingData_.conv2dApiTiling.set_kL0(cachedTilingData_.kL0);271+ tilingData_.convApiTiling.set_kL0(cachedTilingData_.kL0);
270- tilingData_.conv2dApiTiling.set_nL0(cachedTilingData_.nL0);272+ tilingData_.convApiTiling.set_nL0(cachedTilingData_.nL0);
271- tilingData_.conv2dApiTiling.set_pBufferFlag(cachedTilingData_.pBufferFlag);273+ tilingData_.convApiTiling.set_pBufferFlag(cachedTilingData_.pBufferFlag);
272- tilingData_.conv2dApiTiling.set_groups(cacheInputArgs_.groups);274+ tilingData_.convApiTiling.set_groups(cacheInputArgs_.groups);
273- tilingData_.conv2dApiTiling.set_enlarge(cachedTilingData_.enlarge);275+ tilingData_.convApiTiling.set_enlarge(cachedTilingData_.enlarge);
274- tilingData_.conv2dApiTiling.set_singleCoreGroups(cachedTilingData_.singleCoreGroups);276+ tilingData_.convApiTiling.set_singleCoreGroups(cachedTilingData_.singleCoreGroups);
275- tilingData_.conv2dApiTiling.set_singleCoreGroupOpt(cachedTilingData_.singleCoreGroupOpt);277+ tilingData_.convApiTiling.set_singleCoreGroupOpt(cachedTilingData_.singleCoreGroupOpt);
276- tilingData_.conv2dApiTiling.set_bUbNStep(cachedTilingData_.bUbNStep);278+ tilingData_.convApiTiling.set_bUbNStep(cachedTilingData_.bUbNStep);
277- tilingData_.conv2dApiTiling.set_bUbKStep(cachedTilingData_.bUbKStep);279+ tilingData_.convApiTiling.set_bUbKStep(cachedTilingData_.bUbKStep);
278- tilingData_.conv2dApiTiling.set_khUb(cachedTilingData_.khUb);280+ tilingData_.convApiTiling.set_khUb(cachedTilingData_.khUb);
279- tilingData_.conv2dApiTiling.set_kwUb(cachedTilingData_.kwUb);281+ tilingData_.convApiTiling.set_kwUb(cachedTilingData_.kwUb);
280- tilingData_.conv2dApiTiling.set_orgHixWi(cachedTilingData_.orgHixWi);282+ tilingData_.convApiTiling.set_orgHixWi(cachedTilingData_.orgHixWi);
281- tilingData_.conv2dApiTiling.set_kernelHxkernelW(cachedTilingData_.kernelHxkernelW);283+ tilingData_.convApiTiling.set_kernelHxkernelW(cachedTilingData_.kernelHxkernelW);
282- tilingData_.conv2dApiTiling.set_kernelHxkernelWxkernelD(cachedTilingData_.kernelHxkernelWxkernelD);284+ tilingData_.convApiTiling.set_kernelHxkernelWxkernelD(cachedTilingData_.kernelHxkernelWxkernelD);
283- tilingData_.conv2dApiTiling.set_aL1SpaceSize(cachedTilingData_.aL1SpaceSize);285+ tilingData_.convApiTiling.set_aL1SpaceSize(cachedTilingData_.aL1SpaceSize);
284- tilingData_.conv2dApiTiling.set_multiNBL1(cachedTilingData_.multiNBL1);286+ tilingData_.convApiTiling.set_multiNBL1(cachedTilingData_.multiNBL1);
285- tilingData_.conv2dApiTiling.set_cinAInCore(cachedTilingData_.cinAInCore);287+ tilingData_.convApiTiling.set_cinAInCore(cachedTilingData_.cinAInCore);
286- tilingData_.conv2dApiTiling.set_cinATailInCore(cachedTilingData_.cinATailInCore);288+ tilingData_.convApiTiling.set_cinATailInCore(cachedTilingData_.cinATailInCore);
287- tilingData_.conv2dApiTiling.set_cinBInCore(cachedTilingData_.cinBInCore);289+ tilingData_.convApiTiling.set_cinBInCore(cachedTilingData_.cinBInCore);
288- tilingData_.conv2dApiTiling.set_cinBTailInCore(cachedTilingData_.cinBTailInCore);290+ tilingData_.convApiTiling.set_cinBTailInCore(cachedTilingData_.cinBTailInCore);
289- tilingData_.conv2dApiTiling.set_mStep(cachedTilingData_.mStep);291+ tilingData_.convApiTiling.set_mStep(cachedTilingData_.mStep);
290- tilingData_.conv2dApiTiling.set_kStep(cachedTilingData_.kStep);292+ tilingData_.convApiTiling.set_kStep(cachedTilingData_.kStep);
291- tilingData_.conv2dApiTiling.set_nStep(cachedTilingData_.nStep);293+ tilingData_.convApiTiling.set_nStep(cachedTilingData_.nStep);
292- tilingData_.conv2dApiTiling.set_fmapKStride(cachedTilingData_.fmapKStride);294+ tilingData_.convApiTiling.set_fmapKStride(cachedTilingData_.fmapKStride);
293- tilingData_.conv2dApiTiling.set_weightKStride(cachedTilingData_.weightKStride);295+ tilingData_.convApiTiling.set_weightKStride(cachedTilingData_.weightKStride);
294- tilingData_.conv2dApiTiling.set_cinOffsetBlockInGM(cachedTilingData_.cinOffsetBlockInGM);296+ tilingData_.convApiTiling.set_cinOffsetBlockInGM(cachedTilingData_.cinOffsetBlockInGM);
295- tilingData_.conv2dApiTiling.set_coutOffsetBlock(cachedTilingData_.coutOffsetBlock);297+ tilingData_.convApiTiling.set_coutOffsetBlock(cachedTilingData_.coutOffsetBlock);
296- tilingData_.conv2dApiTiling.set_nL1DivBlockSize(cachedTilingData_.nL1DivBlockSize);298+ tilingData_.convApiTiling.set_nL1DivBlockSize(cachedTilingData_.nL1DivBlockSize);
297- tilingData_.conv2dApiTiling.set_kernelH(cacheInputArgs_.weightShapeH);299+ tilingData_.convApiTiling.set_kernelH(cacheInputArgs_.weightShapeH);
298- tilingData_.conv2dApiTiling.set_kernelW(cacheInputArgs_.weightShapeW);300+ tilingData_.convApiTiling.set_kernelW(cacheInputArgs_.weightShapeW);
299}301}
300 302 
301void Conv2dBaseTiling::TranslateCachedTilingData()303void Conv2dBaseTiling::TranslateCachedTilingData()
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_check_attrs.cpp+1-0
@@ -404,6 +404,7 @@ ge::graphStatus Conv2dBaseTiling::CheckExtendConv2dReluWeightAndClipValue(const
404 if (reluWeightShapeLen == 1) {404 if (reluWeightShapeLen == 1) {
405 reluMode = static_cast<uint8_t>(ReluMode::SCALARRELU);405 reluMode = static_cast<uint8_t>(ReluMode::SCALARRELU);
406 } else if (reluWeightShapeLen > 1) {406 } else if (reluWeightShapeLen > 1) {
407+ fixpipeInfo_.channelWiseCoeff += FLOAT_DTYPE_SIZE_COMPARE_FP16;
407 reluMode = static_cast<uint8_t>(ReluMode::VECTORRELU);408 reluMode = static_cast<uint8_t>(ReluMode::VECTORRELU);
408 }409 }
409 410 
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_fast_tiling.cpp+30-30
@@ -243,12 +243,12 @@ ge::graphStatus Conv2dBaseTiling::GetConv2dApiTiling()
243 Conv2dApiTilingSetShape();243 Conv2dApiTilingSetShape();
244 244 
245 if (flagInfo_.mBasicBlockFlag) {245 if (flagInfo_.mBasicBlockFlag) {
246- if (!conv2dApiTiling_.GetTiling(conv2dBasicBlockInfo_, tilingData_.conv2dApiTiling)) {246+ if (!conv2dApiTiling_.GetTiling(conv2dBasicBlockInfo_, tilingData_.convApiTiling)) {
247 OP_LOGE(context_->GetNodeName(), "%s AscendC: get api tiling wrong", paramInfo_.nodeType.c_str());247 OP_LOGE(context_->GetNodeName(), "%s AscendC: get api tiling wrong", paramInfo_.nodeType.c_str());
248 return ge::GRAPH_FAILED;248 return ge::GRAPH_FAILED;
249 }249 }
250 } else {250 } else {
251- if (conv2dApiTiling_.GetTiling(tilingData_.conv2dApiTiling) == -1) {251+ if (conv2dApiTiling_.GetTiling(tilingData_.convApiTiling) == -1) {
252 OP_LOGE(context_->GetNodeName(), "%s AscendC: get api tiling wrong", paramInfo_.nodeType.c_str());252 OP_LOGE(context_->GetNodeName(), "%s AscendC: get api tiling wrong", paramInfo_.nodeType.c_str());
253 return ge::GRAPH_FAILED;253 return ge::GRAPH_FAILED;
254 }254 }
@@ -326,39 +326,39 @@ void Conv2dBaseTiling::Conv2dOpTilingSetAttr()
326 326 
327ge::graphStatus Conv2dBaseTiling::GetConv2dOpsTiling()327ge::graphStatus Conv2dBaseTiling::GetConv2dOpsTiling()
328{328{
329- tilingData_.conv2dRunInfo.set_hin(static_cast<uint64_t>(shapeInfo_.hi));329+ tilingData_.convRunInfo.set_hin(static_cast<uint64_t>(shapeInfo_.hi));
330- tilingData_.conv2dRunInfo.set_win(static_cast<uint64_t>(shapeInfo_.wi));330+ tilingData_.convRunInfo.set_win(static_cast<uint64_t>(shapeInfo_.wi));
331- tilingData_.conv2dRunInfo.set_hout(static_cast<uint64_t>(shapeInfo_.ho));331+ tilingData_.convRunInfo.set_hout(static_cast<uint64_t>(shapeInfo_.ho));
332- tilingData_.conv2dRunInfo.set_wout(static_cast<uint64_t>(shapeInfo_.wo));332+ tilingData_.convRunInfo.set_wout(static_cast<uint64_t>(shapeInfo_.wo));
333- tilingData_.conv2dRunInfo.set_batch(static_cast<uint32_t>(shapeInfo_.batch));333+ tilingData_.convRunInfo.set_batch(static_cast<uint32_t>(shapeInfo_.batch));
334- tilingData_.conv2dRunInfo.set_cin(static_cast<uint32_t>(shapeInfo_.ci));334+ tilingData_.convRunInfo.set_cin(static_cast<uint32_t>(shapeInfo_.ci));
335- tilingData_.conv2dRunInfo.set_cout(static_cast<uint32_t>(shapeInfo_.co));335+ tilingData_.convRunInfo.set_cout(static_cast<uint32_t>(shapeInfo_.co));
336- tilingData_.conv2dRunInfo.set_kh(static_cast<uint32_t>(shapeInfo_.kh));336+ tilingData_.convRunInfo.set_kh(static_cast<uint32_t>(shapeInfo_.kh));
337- tilingData_.conv2dRunInfo.set_kw(static_cast<uint32_t>(shapeInfo_.kw));337+ tilingData_.convRunInfo.set_kw(static_cast<uint32_t>(shapeInfo_.kw));
338- tilingData_.conv2dRunInfo.set_strideH(static_cast<uint32_t>(attrInfo_.strideH));338+ tilingData_.convRunInfo.set_strideH(static_cast<uint32_t>(attrInfo_.strideH));
339- tilingData_.conv2dRunInfo.set_strideW(static_cast<uint32_t>(attrInfo_.strideW));339+ tilingData_.convRunInfo.set_strideW(static_cast<uint32_t>(attrInfo_.strideW));
340- tilingData_.conv2dRunInfo.set_dilationH(static_cast<uint32_t>(attrInfo_.dilationH));340+ tilingData_.convRunInfo.set_dilationH(static_cast<uint32_t>(attrInfo_.dilationH));
341- tilingData_.conv2dRunInfo.set_dilationW(static_cast<uint32_t>(attrInfo_.dilationW));341+ tilingData_.convRunInfo.set_dilationW(static_cast<uint32_t>(attrInfo_.dilationW));
342- tilingData_.conv2dRunInfo.set_padTop(static_cast<uint32_t>(attrInfo_.padTop));342+ tilingData_.convRunInfo.set_padTop(static_cast<uint32_t>(attrInfo_.padTop));
343- tilingData_.conv2dRunInfo.set_padLeft(static_cast<uint32_t>(attrInfo_.padLeft));343+ tilingData_.convRunInfo.set_padLeft(static_cast<uint32_t>(attrInfo_.padLeft));
344- tilingData_.conv2dRunInfo.set_hasBias(static_cast<uint8_t>(flagInfo_.hasBias));344+ tilingData_.convRunInfo.set_hasBias(static_cast<uint8_t>(flagInfo_.hasBias));
345- tilingData_.conv2dRunInfo.set_batchDim(static_cast<uint32_t>(numBlocksRes.batchDim));345+ tilingData_.convRunInfo.set_batchDim(static_cast<uint32_t>(numBlocksRes.batchDim));
346- tilingData_.conv2dRunInfo.set_nDim(static_cast<uint32_t>(numBlocksRes.nDim));346+ tilingData_.convRunInfo.set_nDim(static_cast<uint32_t>(numBlocksRes.nDim));
347- tilingData_.conv2dRunInfo.set_groupDim(static_cast<uint32_t>(numBlocksRes.groupDim));347+ tilingData_.convRunInfo.set_groupDim(static_cast<uint32_t>(numBlocksRes.groupDim));
348- tilingData_.conv2dRunInfo.set_groups(static_cast<uint32_t>(attrInfo_.groups));348+ tilingData_.convRunInfo.set_groups(static_cast<uint32_t>(attrInfo_.groups));
349 if (flagInfo_.convGroupType == ConvGroupType::OPT_GROUP_CONV) {349 if (flagInfo_.convGroupType == ConvGroupType::OPT_GROUP_CONV) {
350- tilingData_.conv2dRunInfo.set_cinOpt(static_cast<uint32_t>(optGroupInfo_.cinOpt));350+ tilingData_.convRunInfo.set_cinOpt(static_cast<uint32_t>(optGroupInfo_.cinOpt));
351- tilingData_.conv2dRunInfo.set_coutOpt(static_cast<uint32_t>(optGroupInfo_.coutOpt));351+ tilingData_.convRunInfo.set_coutOpt(static_cast<uint32_t>(optGroupInfo_.coutOpt));
352- tilingData_.conv2dRunInfo.set_groupOpt(static_cast<uint32_t>(optGroupInfo_.groupOpt));352+ tilingData_.convRunInfo.set_groupOpt(static_cast<uint32_t>(optGroupInfo_.groupOpt));
353- tilingData_.conv2dRunInfo.set_enlarge(static_cast<uint32_t>(optGroupInfo_.enlarge));353+ tilingData_.convRunInfo.set_enlarge(static_cast<uint32_t>(optGroupInfo_.enlarge));
354 }354 }
355 355 
356 if (flagInfo_.mSplitModeFlag) {356 if (flagInfo_.mSplitModeFlag) {
357- tilingData_.conv2dRunInfo.set_hoDim(static_cast<uint32_t>(numBlocksRes.mDim));357+ tilingData_.convRunInfo.set_hoDim(static_cast<uint32_t>(numBlocksRes.mDim));
358- tilingData_.conv2dRunInfo.set_woDim(static_cast<uint32_t>(1));358+ tilingData_.convRunInfo.set_woDim(static_cast<uint32_t>(1));
359 } else {359 } else {
360- tilingData_.conv2dRunInfo.set_hoDim(static_cast<uint32_t>(numBlocksRes.hoDim));360+ tilingData_.convRunInfo.set_hoDim(static_cast<uint32_t>(numBlocksRes.hoDim));
361- tilingData_.conv2dRunInfo.set_woDim(static_cast<uint32_t>(numBlocksRes.woDim));361+ tilingData_.convRunInfo.set_woDim(static_cast<uint32_t>(numBlocksRes.woDim));
362 }362 }
363 return ge::GRAPH_SUCCESS;363 return ge::GRAPH_SUCCESS;
364}364}
Mconv/conv2d_v2/op_host/op_tiling/arch35/conv2d_v2_base_tiling_print.cpp+96-95
@@ -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_.conv2dRunInfo.get_hin() << ", win: " << tilingData_.conv2dRunInfo.get_win()26+ ss << "hin: " << tilingData_.convRunInfo.get_hin() << ", win: " << tilingData_.convRunInfo.get_win()
27- << ", hout: " << tilingData_.conv2dRunInfo.get_hout() << ", wout: " << tilingData_.conv2dRunInfo.get_wout()27+ << ", hout: " << tilingData_.convRunInfo.get_hout() << ", wout: " << tilingData_.convRunInfo.get_wout()
28- << ", batch: " << tilingData_.conv2dRunInfo.get_batch() << ", cin: " << tilingData_.conv2dRunInfo.get_cin()28+ << ", batch: " << tilingData_.convRunInfo.get_batch() << ", cin: " << tilingData_.convRunInfo.get_cin()
29- << ", cout: " << tilingData_.conv2dRunInfo.get_cout() << ", kh: " << tilingData_.conv2dRunInfo.get_kh()29+ << ", cout: " << tilingData_.convRunInfo.get_cout() << ", kh: " << tilingData_.convRunInfo.get_kh()
30- << ", kw: " << tilingData_.conv2dRunInfo.get_kw() << ", batchDim: " << tilingData_.conv2dRunInfo.get_batchDim()30+ << ", kw: " << tilingData_.convRunInfo.get_kw() << ", batchDim: " << tilingData_.convRunInfo.get_batchDim()
31- << ", hoDim: " << tilingData_.conv2dRunInfo.get_hoDim() << ", woDim: " << tilingData_.conv2dRunInfo.get_woDim()31+ << ", hoDim: " << tilingData_.convRunInfo.get_hoDim() << ", woDim: " << tilingData_.convRunInfo.get_woDim()
32- << ", nDim: " << tilingData_.conv2dRunInfo.get_nDim()32+ << ", nDim: " << tilingData_.convRunInfo.get_nDim()
33- << ", strideH: " << tilingData_.conv2dRunInfo.get_strideH()33+ << ", strideH: " << tilingData_.convRunInfo.get_strideH()
34- << ", strideW: " << tilingData_.conv2dRunInfo.get_strideW()34+ << ", strideW: " << tilingData_.convRunInfo.get_strideW()
35- << ", dilationH: " << tilingData_.conv2dRunInfo.get_dilationH()35+ << ", dilationH: " << tilingData_.convRunInfo.get_dilationH()
36- << ", dilationW: " << tilingData_.conv2dRunInfo.get_dilationW()36+ << ", dilationW: " << tilingData_.convRunInfo.get_dilationW()
37- << ", padTop: " << tilingData_.conv2dRunInfo.get_padTop()37+ << ", padTop: " << tilingData_.convRunInfo.get_padTop()
38- << ", padLeft: " << tilingData_.conv2dRunInfo.get_padLeft()38+ << ", padLeft: " << tilingData_.convRunInfo.get_padLeft()
39- << ", groups: " << tilingData_.conv2dRunInfo.get_groups()39+ << ", groups: " << tilingData_.convRunInfo.get_groups()
40- << ", cinOpt: " << tilingData_.conv2dRunInfo.get_cinOpt()40+ << ", cinOpt: " << tilingData_.convRunInfo.get_cinOpt()
41- << ", coutOpt: " << tilingData_.conv2dRunInfo.get_coutOpt()41+ << ", coutOpt: " << tilingData_.convRunInfo.get_coutOpt()
42- << ", groupOpt: " << tilingData_.conv2dRunInfo.get_groupOpt()42+ << ", groupOpt: " << tilingData_.convRunInfo.get_groupOpt()
43- << ", enlarge: " << tilingData_.conv2dRunInfo.get_enlarge()43+ << ", enlarge: " << tilingData_.convRunInfo.get_enlarge()
44- << ", groupDim: " << tilingData_.conv2dRunInfo.get_groupDim()44+ << ", groupDim: " << tilingData_.convRunInfo.get_groupDim()
45- << ", hasBias: " << static_cast<uint32_t>(tilingData_.conv2dRunInfo.get_hasBias());45+ << ", hasBias: " << static_cast<uint32_t>(tilingData_.convRunInfo.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_.conv2dApiTiling.get_singleCoreHo()76+ ss << "singleCoreHo: " << tilingData_.convApiTiling.get_singleCoreHo()
77- << ", singleCoreWo: " << tilingData_.conv2dApiTiling.get_singleCoreWo()77+ << ", singleCoreWo: " << tilingData_.convApiTiling.get_singleCoreWo()
78- << ", singleCoreBatch: " << tilingData_.conv2dApiTiling.get_singleCoreBatch()78+ << ", singleCoreBatch: " << tilingData_.convApiTiling.get_singleCoreBatch()
79- << ", orgHi: " << tilingData_.conv2dApiTiling.get_orgHi()79+ << ", orgHi: " << tilingData_.convApiTiling.get_orgHi()
80- << ", orgWi: " << tilingData_.conv2dApiTiling.get_orgWi()80+ << ", orgWi: " << tilingData_.convApiTiling.get_orgWi()
81- << ", orgHo: " << tilingData_.conv2dApiTiling.get_orgHo()81+ << ", orgHo: " << tilingData_.convApiTiling.get_orgHo()
82- << ", orgWo: " << tilingData_.conv2dApiTiling.get_orgWo()82+ << ", orgWo: " << tilingData_.convApiTiling.get_orgWo()
83- << ", groups: " << tilingData_.conv2dApiTiling.get_groups()83+ << ", groups: " << tilingData_.convApiTiling.get_groups()
84- << ", orgCi: " << tilingData_.conv2dApiTiling.get_orgCi()84+ << ", orgCi: " << tilingData_.convApiTiling.get_orgCi()
85- << ", orgCo: " << tilingData_.conv2dApiTiling.get_orgCo()85+ << ", orgCo: " << tilingData_.convApiTiling.get_orgCo()
86- << ", kernelH: " << tilingData_.conv2dApiTiling.get_kernelH()86+ << ", kernelH: " << tilingData_.convApiTiling.get_kernelH()
87- << ", kernelW: " << tilingData_.conv2dApiTiling.get_kernelW()87+ << ", kernelW: " << tilingData_.convApiTiling.get_kernelW()
88- << ", singleCoreCi: " << tilingData_.conv2dApiTiling.get_singleCoreCi()88+ << ", singleCoreCi: " << tilingData_.convApiTiling.get_singleCoreCi()
89- << ", singleCoreCo: " << tilingData_.conv2dApiTiling.get_singleCoreCo()89+ << ", singleCoreCo: " << tilingData_.convApiTiling.get_singleCoreCo()
90- << ", hoL1: " << tilingData_.conv2dApiTiling.get_hoL1()90+ << ", hoL1: " << tilingData_.convApiTiling.get_hoL1()
91- << ", woL1: " << tilingData_.conv2dApiTiling.get_woL1()91+ << ", woL1: " << tilingData_.convApiTiling.get_woL1()
92- << ", kAL1: " << tilingData_.conv2dApiTiling.get_kAL1()92+ << ", kAL1: " << tilingData_.convApiTiling.get_kAL1()
93- << ", kBL1: " << tilingData_.conv2dApiTiling.get_kBL1()93+ << ", kBL1: " << tilingData_.convApiTiling.get_kBL1()
94- << ", nBL1: " << tilingData_.conv2dApiTiling.get_nBL1()94+ << ", nBL1: " << tilingData_.convApiTiling.get_nBL1()
95- << ", hoL0: " << tilingData_.conv2dApiTiling.get_hoL0()95+ << ", hoL0: " << tilingData_.convApiTiling.get_hoL0()
96- << ", woL0: " << tilingData_.conv2dApiTiling.get_woL0()96+ << ", woL0: " << tilingData_.convApiTiling.get_woL0()
97- << ", kL0: " << tilingData_.conv2dApiTiling.get_kL0()97+ << ", kL0: " << tilingData_.convApiTiling.get_kL0()
98- << ", nL0: " << tilingData_.conv2dApiTiling.get_nL0()98+ << ", nL0: " << tilingData_.convApiTiling.get_nL0()
99- << ", pBufferFlag: " << tilingData_.conv2dApiTiling.get_pBufferFlag()99+ << ", pBufferFlag: " << tilingData_.convApiTiling.get_pBufferFlag()
100- << ", multiNBL1: " << tilingData_.conv2dApiTiling.get_multiNBL1()100+ << ", multiNBL1: " << tilingData_.convApiTiling.get_multiNBL1()
101- << ", strideH: " << tilingData_.conv2dApiTiling.get_strideH()101+ << ", strideH: " << tilingData_.convApiTiling.get_strideH()
102- << ", strideW: " << tilingData_.conv2dApiTiling.get_strideW()102+ << ", strideW: " << tilingData_.convApiTiling.get_strideW()
103- << ", dilationH: " << tilingData_.conv2dApiTiling.get_dilationH()103+ << ", dilationH: " << tilingData_.convApiTiling.get_dilationH()
104- << ", dilationW: " << tilingData_.conv2dApiTiling.get_dilationW()104+ << ", dilationW: " << tilingData_.convApiTiling.get_dilationW()
105- << ", padTop: " << tilingData_.conv2dApiTiling.get_padTop()105+ << ", padTop: " << tilingData_.convApiTiling.get_padTop()
106- << ", padBottom: " << tilingData_.conv2dApiTiling.get_padBottom()106+ << ", padBottom: " << tilingData_.convApiTiling.get_padBottom()
107- << ", padLeft: " << tilingData_.conv2dApiTiling.get_padLeft()107+ << ", padLeft: " << tilingData_.convApiTiling.get_padLeft()
108- << ", padRight: " << tilingData_.conv2dApiTiling.get_padRight()108+ << ", padRight: " << tilingData_.convApiTiling.get_padRight()
109- << ", aL1SpaceSize: " << tilingData_.conv2dApiTiling.get_aL1SpaceSize()109+ << ", aL1SpaceSize: " << tilingData_.convApiTiling.get_aL1SpaceSize()
110- << ", singleCoreGroups: " << tilingData_.conv2dApiTiling.get_singleCoreGroups()110+ << ", singleCoreGroups: " << tilingData_.convApiTiling.get_singleCoreGroups()
111- << ", singleCoreGroupOpt: " << tilingData_.conv2dApiTiling.get_singleCoreGroupOpt()111+ << ", singleCoreGroupOpt: " << tilingData_.convApiTiling.get_singleCoreGroupOpt()
112- << ", enlarge: " << tilingData_.conv2dApiTiling.get_enlarge()112+ << ", enlarge: " << tilingData_.convApiTiling.get_enlarge()
113- << ", bUbNStep: " << tilingData_.conv2dApiTiling.get_bUbNStep()113+ << ", bUbNStep: " << tilingData_.convApiTiling.get_bUbNStep()
114- << ", iterateMNOrder: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_iterateMNOrder())114+ << ", iterateMNOrder: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_iterateMNOrder())
115- << ", biasFullLoadFlag: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_biasFullLoadFlag())115+ << ", biasFullLoadFlag: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_biasFullLoadFlag())
116- << ", fixpParamsFullLoadFlag: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_fixpParamsFullLoadFlag())116+ << ", fixpParamsFullLoadFlag: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_fixpParamsFullLoadFlag())
117- << ", hf32Enable: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_hf32Enable())117+ << ", hf32Enable: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_hf32Enable())
118- << ", hf32TransMode: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_hf32TransMode())118+ << ", hf32TransMode: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_hf32TransMode())
119- << ", hasBias: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_hasBias())119+ << ", hasBias: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_hasBias())
120- << ", hasScale: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_hasScale())120+ << ", hasScale: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_hasScale())
121- << ", offsetx: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_offsetx())121+ << ", offsetx: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_offsetx())
122- << ", roundMode: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_roundMode())122+ << ", roundMode: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_roundMode())
123- << ", innerBatch: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_innerBatch());123+ << ", innerBatch: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_innerBatch());
124}124}
125 125 
126 126 
@@ -133,33 +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_.conv2dApiTiling.get_bUbKStep()136+ ssPartTwo << "bUbKStep: " << tilingData_.convApiTiling.get_bUbKStep()
137- << ", orgHixWi: " << tilingData_.conv2dApiTiling.get_orgHixWi()137+ << ", orgHixWi: " << tilingData_.convApiTiling.get_orgHixWi()
138- << ", kernelHxkernelW: " << tilingData_.conv2dApiTiling.get_kernelHxkernelW()138+ << ", kernelHxkernelW: " << tilingData_.convApiTiling.get_kernelHxkernelW()
139- << ", kernelHxkernelWxkernelD: " << tilingData_.conv2dApiTiling.get_kernelHxkernelWxkernelD()139+ << ", kernelHxkernelWxkernelD: " << tilingData_.convApiTiling.get_kernelHxkernelWxkernelD()
140- << ", cinAInCore: " << tilingData_.conv2dApiTiling.get_cinAInCore()140+ << ", cinAInCore: " << tilingData_.convApiTiling.get_cinAInCore()
141- << ", cinATailInCore: " << tilingData_.conv2dApiTiling.get_cinATailInCore()141+ << ", cinATailInCore: " << tilingData_.convApiTiling.get_cinATailInCore()
142- << ", cinBInCore: " << tilingData_.conv2dApiTiling.get_cinBInCore()142+ << ", cinBInCore: " << tilingData_.convApiTiling.get_cinBInCore()
143- << ", cinBTailInCore: " << tilingData_.conv2dApiTiling.get_cinBTailInCore()143+ << ", cinBTailInCore: " << tilingData_.convApiTiling.get_cinBTailInCore()
144- << ", mStep: " << tilingData_.conv2dApiTiling.get_mStep()144+ << ", mStep: " << tilingData_.convApiTiling.get_mStep()
145- << ", kStep: " << tilingData_.conv2dApiTiling.get_kStep()145+ << ", kStep: " << tilingData_.convApiTiling.get_kStep()
146- << ", nStep: " << tilingData_.conv2dApiTiling.get_nStep()146+ << ", nStep: " << tilingData_.convApiTiling.get_nStep()
147- << ", fmapKStride: " << tilingData_.conv2dApiTiling.get_fmapKStride()147+ << ", fmapKStride: " << tilingData_.convApiTiling.get_fmapKStride()
148- << ", weightKStride: " << tilingData_.conv2dApiTiling.get_weightKStride()148+ << ", weightKStride: " << tilingData_.convApiTiling.get_weightKStride()
149- << ", cinOffsetBlockInGM: " << tilingData_.conv2dApiTiling.get_cinOffsetBlockInGM()149+ << ", cinOffsetBlockInGM: " << tilingData_.convApiTiling.get_cinOffsetBlockInGM()
150- << ", coutOffsetBlock: " << tilingData_.conv2dApiTiling.get_coutOffsetBlock()150+ << ", coutOffsetBlock: " << tilingData_.convApiTiling.get_coutOffsetBlock()
151- << ", nL1DivBlockSize: " << tilingData_.conv2dApiTiling.get_nL1DivBlockSize()151+ << ", nL1DivBlockSize: " << tilingData_.convApiTiling.get_nL1DivBlockSize()
152- << ", dualOutput: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_dualOutput())152+ << ", dualOutput: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_dualOutput())
153- << ", quantMode0: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_quantMode0())153+ << ", quantMode0: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_quantMode0())
154- << ", reluMode0: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_reluMode0())154+ << ", reluMode0: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_reluMode0())
155- << ", clipMode0: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_clipMode0())155+ << ", clipMode0: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_clipMode0())
156- << ", quantMode1: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_quantMode1())156+ << ", quantMode1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_quantMode1())
157- << ", reluMode1: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_reluMode1())157+ << ", reluMode1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_reluMode1())
158- << ", clipMode1: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_clipMode1())158+ << ", clipMode1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_clipMode1())
159- << ", khL1: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_khL1())159+ << ", khL1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_khL1())
160- << ", kwL1: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_kwL1())160+ << ", kwL1: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_kwL1())
161- << ", khUb: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_khUb())161+ << ", khUb: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_khUb())
162- << ", kwUb: " << static_cast<uint32_t>(tilingData_.conv2dApiTiling.get_kwUb());162+ << ", kwUb: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_kwUb())
163+ << ", unionDataXt: " << static_cast<uint32_t>(tilingData_.convApiTiling.get_unionDataXt());
163 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());
164 OP_LOGD(context_->GetNodeName(), "%s", ssPartTwo.str().c_str());165 OP_LOGD(context_->GetNodeName(), "%s", ssPartTwo.str().c_str());
165}166}