Star1329
1446
代码
代码
Issues117
Pull Requests254
Actions
项目讨论
Wiki
分析
Star1329
1446
已合并
fix transpose 021 profiling #3705
jzj007创建于 7月2日
fix transpose 021 profiling #3705
已合并
jzj007创建于 7月2日
从jzj007/ops-math0701_transpose合入到cann/ops-mathmaster
共 2 个文件变更+188-166
Mconversion/transpose/op_host/arch35/transpose_tiling_arch35.cpp+175-153
@@ -30,21 +30,19 @@ ge::graphStatus TransposeNddmaTiling::Init(const int64_t& coreNum, const int64_t
30{30{
31 OP_LOGD(tilingContext_->GetNodeName(), "Start init TransposeNddmaTiling.");31 OP_LOGD(tilingContext_->GetNodeName(), "Start init TransposeNddmaTiling.");
32 coreNum_ = coreNum;32 coreNum_ = coreNum;
33- OP_CHECK_IF(33+ OP_CHECK_IF((coreNum_ <= 0), OP_LOGE(tilingContext_->GetNodeName(), "Failed to get core num."),
34- (coreNum_ <= 0), OP_LOGE(tilingContext_->GetNodeName(), "Failed to get core num."), return ge::GRAPH_FAILED);34+ return ge::GRAPH_FAILED);
35 ubSize_ = ubSize;35 ubSize_ = ubSize;
36- OP_CHECK_IF(36+ OP_CHECK_IF((ubSize_ <= 0), OP_LOGE(tilingContext_->GetNodeName(), "Failed to get ub size."),
37- (ubSize_ <= 0), OP_LOGE(tilingContext_->GetNodeName(), "Failed to get ub size."), return ge::GRAPH_FAILED);37+ return ge::GRAPH_FAILED);
38 38 
39 cacheLineSize_ = Ops::Base::GetCacheLineSize(tilingContext_);39 cacheLineSize_ = Ops::Base::GetCacheLineSize(tilingContext_);
40- OP_CHECK_IF(40+ OP_CHECK_IF((cacheLineSize_ <= 0), OP_LOGE(tilingContext_->GetNodeName(), "Failed to get cache line size."),
41- (cacheLineSize_ <= 0), OP_LOGE(tilingContext_->GetNodeName(), "Failed to get cache line size."),41+ return ge::GRAPH_FAILED);
42- return ge::GRAPH_FAILED);
43 42 
44 ubBlockSize_ = Ops::Base::GetUbBlockSize(tilingContext_);43 ubBlockSize_ = Ops::Base::GetUbBlockSize(tilingContext_);
45- OP_CHECK_IF(44+ OP_CHECK_IF((ubBlockSize_ <= 0), OP_LOGE(tilingContext_->GetNodeName(), "Failed to get ub block size."),
46- (ubBlockSize_ <= 0), OP_LOGE(tilingContext_->GetNodeName(), "Failed to get ub block size."),45+ return ge::GRAPH_FAILED);
47- return ge::GRAPH_FAILED);
48 return ge::GRAPH_SUCCESS;46 return ge::GRAPH_SUCCESS;
49}47}
50 48 
@@ -52,9 +50,8 @@ ge::graphStatus TransposeNddmaTiling::RunTranposelTiling()
52{50{
53 OP_LOGD(tilingContext_->GetNodeName(), "Start running Tiling4Transpose.");51 OP_LOGD(tilingContext_->GetNodeName(), "Start running Tiling4Transpose.");
54 if (!isReleatedTranspsoe_) {52 if (!isReleatedTranspsoe_) {
55- OP_CHECK_IF(53+ OP_CHECK_IF(GetShapeInfo() != ge::GRAPH_SUCCESS,
56- GetShapeInfo() != ge::GRAPH_SUCCESS, OP_LOGE(tilingContext_->GetNodeName(), "Failed to get shape info!"),54+ OP_LOGE(tilingContext_->GetNodeName(), "Failed to get shape info!"), return ge::GRAPH_FAILED);
57- return ge::GRAPH_FAILED);
58 }55 }
59 56 
60 auto ret = CheckShapeInfo();57 auto ret = CheckShapeInfo();
@@ -68,17 +65,15 @@ ge::graphStatus TransposeNddmaTiling::RunTranposelTiling()
68 CHECK_RET_SUCC(ret);65 CHECK_RET_SUCC(ret);
69 66 
70 CalcTotalVolumeActual();67 CalcTotalVolumeActual();
71- OP_CHECK_IF(68+ OP_CHECK_IF(TryVCONVTiling() == ge::GRAPH_SUCCESS, OP_LOGD(tilingContext_->GetNodeName(), "Do convTiling success"),
72- TryVCONVTiling() == ge::GRAPH_SUCCESS, OP_LOGD(tilingContext_->GetNodeName(), "Do convTiling success"),69+ return ge::GRAPH_SUCCESS);
73- return ge::GRAPH_SUCCESS);
74 70 
75 SetIsLastAxisTranspose();71 SetIsLastAxisTranspose();
76 if (!isReleatedTranspsoe_ && shapeInfo_.isLastAxisTranspose) {72 if (!isReleatedTranspsoe_ && shapeInfo_.isLastAxisTranspose) {
77 TransWithGather::PlatInfo platInfo{coreNum_, ubSize_, cacheLineSize_, ubBlockSize_};73 TransWithGather::PlatInfo platInfo{coreNum_, ubSize_, cacheLineSize_, ubBlockSize_};
78 TransWithGather::TransposeGatherTiling gatherTiling(tilingContext_, platInfo, shapeInfo_);74 TransWithGather::TransposeGatherTiling gatherTiling(tilingContext_, platInfo, shapeInfo_);
79- OP_CHECK_IF(75+ OP_CHECK_IF(gatherTiling.DoTiling() == ge::GRAPH_SUCCESS,
80- gatherTiling.DoTiling() == ge::GRAPH_SUCCESS,76+ OP_LOGD(tilingContext_->GetNodeName(), "Do gather tiling done!"), return ge::GRAPH_SUCCESS);
81- OP_LOGD(tilingContext_->GetNodeName(), "Do gather tiling done!"), return ge::GRAPH_SUCCESS);
82 }77 }
83 78 
84 // ensure tiling template79 // ensure tiling template
@@ -119,29 +114,56 @@ ge::graphStatus TransposeNddmaTiling::TryVCONVTiling()
119 shapeInfo_.eleLenInBytes == VCONV_DSIZE && shapeInfo_.reducedInShape[0] > DIM_FIVE) {114 shapeInfo_.eleLenInBytes == VCONV_DSIZE && shapeInfo_.reducedInShape[0] > DIM_FIVE) {
120 TransposeWithVCONV::PlatInfo platInfo{coreNum_, ubSize_};115 TransposeWithVCONV::PlatInfo platInfo{coreNum_, ubSize_};
121 TransposeWithVCONV::TransposeVCONVTiling vconvTiling(tilingContext_, platInfo, shapeInfo_);116 TransposeWithVCONV::TransposeVCONVTiling vconvTiling(tilingContext_, platInfo, shapeInfo_);
122- OP_CHECK_IF(117+ OP_CHECK_IF(vconvTiling.DoTiling() == ge::GRAPH_SUCCESS,
123- vconvTiling.DoTiling() == ge::GRAPH_SUCCESS,118+ OP_LOGD(tilingContext_->GetNodeName(), "Do convTiling done"), return ge::GRAPH_SUCCESS);
124- OP_LOGD(tilingContext_->GetNodeName(), "Do convTiling done"), return ge::GRAPH_SUCCESS);
125 }119 }
126- if (shapeInfo_.reducedPerm[0] == 0 && shapeInfo_.reducedPerm[DIM_ONE] == VCONV_DIM_NUM && shapeInfo_.reducedPerm[DIM_TWO] == 1 &&120+ if (Is021VConvValid()) {
127- shapeInfo_.dim == DIM_THREE // 021 transpose
128- && (shapeInfo_.eleLenInBytes == B8_BYTES || shapeInfo_.eleLenInBytes == B16_BYTES ||
129- shapeInfo_.eleLenInBytes == B32_BYTES) // support B8、B16、B32
130- && shapeInfo_.reducedInShape[DIM_ONE] > DIM_EIGHT &&
131- shapeInfo_.reducedInShape[DIM_TWO] > DIM_EIGHT // not support very little W or H
132- && shapeInfo_.totalVolumeActual * shapeInfo_.eleLenInBytes >=
133- SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102_021 // shape size bigger than 7w
134- ) {
135 Transpose021WithVCONV::PlatInfo platInfo{coreNum_, ubSize_};121 Transpose021WithVCONV::PlatInfo platInfo{coreNum_, ubSize_};
136 Transpose021WithVCONV::Transpose021VCONVTiling vconv021Tiling(tilingContext_, platInfo, shapeInfo_);122 Transpose021WithVCONV::Transpose021VCONVTiling vconv021Tiling(tilingContext_, platInfo, shapeInfo_);
137- OP_CHECK_IF(123+ OP_CHECK_IF(vconv021Tiling.DoTiling() == ge::GRAPH_SUCCESS,
138- vconv021Tiling.DoTiling() == ge::GRAPH_SUCCESS,124+ OP_LOGD(tilingContext_->GetNodeName(), "Do 021 convTiling done"), return ge::GRAPH_SUCCESS);
139- OP_LOGD(tilingContext_->GetNodeName(), "Do 021 convTiling done"), return ge::GRAPH_SUCCESS);
140 }125 }
141 }126 }
142 return ge::GRAPH_FAILED;127 return ge::GRAPH_FAILED;
143}128}
144 129 
130+bool TransposeNddmaTiling::Is021VConvValid()
131+{
132+ // check perm: 021 transpose
133+ if (!(shapeInfo_.reducedPerm[0] == 0 && shapeInfo_.reducedPerm[DIM_ONE] == VCONV_DIM_NUM &&
134+ shapeInfo_.reducedPerm[DIM_TWO] == 1)) {
135+ return false;
136+ }
137+ // check dim
138+ if (shapeInfo_.dim != DIM_THREE) {
139+ return false;
140+ }
141+ // check dtype: support B8、B16、B32
142+ if (!(shapeInfo_.eleLenInBytes == B8_BYTES || shapeInfo_.eleLenInBytes == B16_BYTES ||
143+ shapeInfo_.eleLenInBytes == B32_BYTES)) {
144+ return false;
145+ }
146+ // check HW shape
147+ int64_t H = shapeInfo_.reducedInShape[DIM_ONE];
148+ int64_t W = shapeInfo_.reducedInShape[DIM_TWO];
149+ if (H <= DIM_EIGHT || W <= DIM_EIGHT) {
150+ return false;
151+ }
152+ if (H * W < HW_MIN_PRODUCT) {
153+ return false;
154+ }
155+ int64_t hAlign = Ops::Base::CeilDiv(H, HW_ALIGN) * HW_ALIGN;
156+ int64_t wAlign = Ops::Base::CeilDiv(W, HW_ALIGN) * HW_ALIGN;
157+ if (H * W <= hAlign * wAlign / DIM_TWO) {
158+ return false;
159+ }
160+ // check total volume
161+ if (shapeInfo_.totalVolumeActual * shapeInfo_.eleLenInBytes < SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102_021) {
162+ return false;
163+ }
164+ return true;
165+}
166+ 
145template <typename T>167template <typename T>
146bool TransposeNddmaTiling::GetPerm(const gert::Tensor* permTensor)168bool TransposeNddmaTiling::GetPerm(const gert::Tensor* permTensor)
147{169{
@@ -191,9 +213,8 @@ ge::graphStatus TransposeNddmaTiling::GetShapeInfo()
191 return ge::GRAPH_FAILED;213 return ge::GRAPH_FAILED;
192 }214 }
193 } else {215 } else {
194- OP_LOGE_FOR_INVALID_DTYPE(216+ OP_LOGE_FOR_INVALID_DTYPE(tilingContext_->GetNodeName(), "perm",
195- tilingContext_->GetNodeName(), "perm", ge::TypeUtils::DataTypeToSerialString(permDtype).c_str(),217+ ge::TypeUtils::DataTypeToSerialString(permDtype).c_str(), "int32 or int64");
196- "int32 or int64");
197 return ge::GRAPH_FAILED;218 return ge::GRAPH_FAILED;
198 }219 }
199 220 
@@ -225,23 +246,20 @@ ge::graphStatus TransposeNddmaTiling::CheckShapeDims()
225 int64_t inDims = shapeInfo_.inShapeSize;246 int64_t inDims = shapeInfo_.inShapeSize;
226 int64_t outDims = shapeInfo_.outShapeSize;247 int64_t outDims = shapeInfo_.outShapeSize;
227 int64_t permDims = shapeInfo_.permSize;248 int64_t permDims = shapeInfo_.permSize;
228- OP_CHECK_IF(249+ OP_CHECK_IF(inDims < 1,
229- inDims < 1,250+ OP_LOGE_FOR_INVALID_SHAPEDIM_WITH_REASON(tilingContext_->GetNodeName(), "x",
230- OP_LOGE_FOR_INVALID_SHAPEDIM_WITH_REASON(251+ std::to_string(inDims).c_str(), "positive"),
231- tilingContext_->GetNodeName(), "x", std::to_string(inDims).c_str(), "positive"),252+ return ge::GRAPH_FAILED);
232- return ge::GRAPH_FAILED);253+ OP_CHECK_IF(inDims != outDims,
233- OP_CHECK_IF(254+ OP_LOGE_FOR_INVALID_SHAPEDIMS_WITH_REASON(tilingContext_->GetNodeName(), "x and y",
234- inDims != outDims,255+ Ops::Math::Join(inDims, outDims).c_str(),
235- OP_LOGE_FOR_INVALID_SHAPEDIMS_WITH_REASON(256+ "The shape dims of x and y must be the same"),
236- tilingContext_->GetNodeName(), "x and y", Ops::Math::Join(inDims, outDims).c_str(),257+ return ge::GRAPH_FAILED);
237- "The shape dims of x and y must be the same"),258+ OP_CHECK_IF(inDims != permDims,
238- return ge::GRAPH_FAILED);259+ OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON(
239- OP_CHECK_IF(260+ tilingContext_->GetNodeName(), "perm", std::to_string(permDims).c_str(),
240- inDims != permDims,261+ "The total number of elements of perm must be equal to the shape dim of x"),
241- OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON(262+ return ge::GRAPH_FAILED);
242- tilingContext_->GetNodeName(), "perm", std::to_string(permDims).c_str(),
243- "The total number of elements of perm must be equal to the shape dim of x"),
244- return ge::GRAPH_FAILED);
245 return ge::GRAPH_SUCCESS;263 return ge::GRAPH_SUCCESS;
246}264}
247 265 
@@ -252,9 +270,9 @@ ge::graphStatus TransposeNddmaTiling::CheckShapeInfo()
252 270 
253 for (int64_t i = 0; i < shapeInfo_.inShapeSize; i++) {271 for (int64_t i = 0; i < shapeInfo_.inShapeSize; i++) {
254 if (shapeInfo_.perm[i] >= shapeInfo_.inShapeSize) {272 if (shapeInfo_.perm[i] >= shapeInfo_.inShapeSize) {
255- OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(273+ OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(tilingContext_->GetNodeName(), "perm",
256- tilingContext_->GetNodeName(), "perm", std::to_string(shapeInfo_.perm[i]).c_str(),274+ std::to_string(shapeInfo_.perm[i]).c_str(),
257- "The value of perm must be less than shape dim of x");275+ "The value of perm must be less than shape dim of x");
258 return ge::GRAPH_FAILED;276 return ge::GRAPH_FAILED;
259 }277 }
260 if (shapeInfo_.inShape[shapeInfo_.perm[i]] != shapeInfo_.outShape[i]) {278 if (shapeInfo_.inShape[shapeInfo_.perm[i]] != shapeInfo_.outShape[i]) {
@@ -273,15 +291,15 @@ ge::graphStatus TransposeNddmaTiling::CheckShapeInfo()
273 291 
274 for (int64_t i = 0; i < shapeInfo_.inShapeSize; i++) {292 for (int64_t i = 0; i < shapeInfo_.inShapeSize; i++) {
275 if (shapeInfo_.inShape[i] <= 0) {293 if (shapeInfo_.inShape[i] <= 0) {
276- OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON(294+ OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON(tilingContext_->GetNodeName(), "x",
277- tilingContext_->GetNodeName(), "x", std::to_string(shapeInfo_.inShape[i]).c_str(),295+ std::to_string(shapeInfo_.inShape[i]).c_str(),
278- "All axes of x must be positive numbers");296+ "All axes of x must be positive numbers");
279 return ge::GRAPH_FAILED;297 return ge::GRAPH_FAILED;
280 }298 }
281 if (shapeInfo_.outShape[i] <= 0) {299 if (shapeInfo_.outShape[i] <= 0) {
282- OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON(300+ OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON(tilingContext_->GetNodeName(), "y",
283- tilingContext_->GetNodeName(), "y", std::to_string(shapeInfo_.outShape[i]).c_str(),301+ std::to_string(shapeInfo_.outShape[i]).c_str(),
284- "All axes of y must be positive numbers");302+ "All axes of y must be positive numbers");
285 return ge::GRAPH_FAILED;303 return ge::GRAPH_FAILED;
286 }304 }
287 }305 }
@@ -300,9 +318,8 @@ ge::graphStatus TransposeNddmaTiling::CheckReducedShapeInfo()
300 318 
301 for (int64_t i = 0; i < dim; i++) {319 for (int64_t i = 0; i < dim; i++) {
302 if (shapeInfo_.reducedInShape[i] <= 0 || shapeInfo_.reducedOutShape[i] <= 0) {320 if (shapeInfo_.reducedInShape[i] <= 0 || shapeInfo_.reducedOutShape[i] <= 0) {
303- OP_LOGE(321+ OP_LOGE(tilingContext_->GetNodeName(), "Invalid shape, index is %ld, inShape is %ld, outShape is %ld", i,
304- tilingContext_->GetNodeName(), "Invalid shape, index is %ld, inShape is %ld, outShape is %ld", i,322+ shapeInfo_.reducedInShape[i], shapeInfo_.reducedOutShape[i]);
305- shapeInfo_.reducedInShape[i], shapeInfo_.reducedOutShape[i]);
306 return ge::GRAPH_FAILED;323 return ge::GRAPH_FAILED;
307 }324 }
308 }325 }
@@ -320,8 +337,8 @@ int64_t TransposeNddmaTiling::DoSplitUBInput()
320 splitInfo_.inUbFactor = splitInfo_.inUbElement;337 splitInfo_.inUbFactor = splitInfo_.inUbElement;
321 splitInfo_.inTailFactor = currentShapeDim % splitInfo_.inUbFactor;338 splitInfo_.inTailFactor = currentShapeDim % splitInfo_.inUbFactor;
322 splitInfo_.inUbActual *= splitInfo_.inUbElement;339 splitInfo_.inUbActual *= splitInfo_.inUbElement;
323- remainingTotalElment =340+ remainingTotalElment = remainingTotalElment / currentShapeDim *
324- remainingTotalElment / currentShapeDim * Ops::Base::CeilDiv(currentShapeDim, splitInfo_.inUbElement);341+ Ops::Base::CeilDiv(currentShapeDim, splitInfo_.inUbElement);
325 break;342 break;
326 } else {343 } else {
327 splitInfo_.inUbElement /= currentShapeDim;344 splitInfo_.inUbElement /= currentShapeDim;
@@ -354,8 +371,8 @@ bool TransposeNddmaTiling::UbOutOfBoundCheck(int64_t currentSplitIndex, int64_t
354 if (calcIn && shapeInfo_.reducedPerm[currentSplitIndex] == splitInfo_.inCutIndex) {371 if (calcIn && shapeInfo_.reducedPerm[currentSplitIndex] == splitInfo_.inCutIndex) {
355 burstLenBlockAlign *= splitInfo_.inUbFactor;372 burstLenBlockAlign *= splitInfo_.inUbFactor;
356 }373 }
357- burstLenBlockAlign =374+ burstLenBlockAlign = Ops::Base::CeilAlign(burstLenBlockAlign * shapeInfo_.eleLenInBytes, ubBlockSize_) /
358- Ops::Base::CeilAlign(burstLenBlockAlign * shapeInfo_.eleLenInBytes, ubBlockSize_) / shapeInfo_.eleLenInBytes;375+ shapeInfo_.eleLenInBytes;
359 int64_t inUbElements = burstLenBlockAlign;376 int64_t inUbElements = burstLenBlockAlign;
360 for (int64_t i = 0; i < currentSplitIndex; i++) {377 for (int64_t i = 0; i < currentSplitIndex; i++) {
361 if (shapeInfo_.reducedPerm[i] > splitInfo_.inCutIndex) {378 if (shapeInfo_.reducedPerm[i] > splitInfo_.inCutIndex) {
@@ -378,8 +395,8 @@ bool TransposeNddmaTiling::UbOutOfBoundCheckNLast(int64_t currentSplitIndex, int
378 } else {395 } else {
379 burstLenBlockAlign = shapeInfo_.reducedInShape[shapeInfo_.dim - 1];396 burstLenBlockAlign = shapeInfo_.reducedInShape[shapeInfo_.dim - 1];
380 }397 }
381- burstLenBlockAlign =398+ burstLenBlockAlign = Ops::Base::CeilAlign(burstLenBlockAlign * shapeInfo_.eleLenInBytes, ubBlockSize_) /
382- Ops::Base::CeilAlign(burstLenBlockAlign * shapeInfo_.eleLenInBytes, ubBlockSize_) / shapeInfo_.eleLenInBytes;399+ shapeInfo_.eleLenInBytes;
383 int64_t inUbElements = burstLenBlockAlign;400 int64_t inUbElements = burstLenBlockAlign;
384 for (int64_t i = currentSplitIndex; i < shapeInfo_.dim - 1; i++) {401 for (int64_t i = currentSplitIndex; i < shapeInfo_.dim - 1; i++) {
385 if (i == currentSplitIndex) {402 if (i == currentSplitIndex) {
@@ -394,8 +411,8 @@ bool TransposeNddmaTiling::UbOutOfBoundCheckNLast(int64_t currentSplitIndex, int
394 return false;411 return false;
395}412}
396 413 
397-void TransposeNddmaTiling::FindSplitFactorByRateNLast(414+void TransposeNddmaTiling::FindSplitFactorByRateNLast(int64_t currentSplitIndex, int64_t currentInShapeDim,
398- int64_t currentSplitIndex, int64_t currentInShapeDim, int64_t remainingTotalElment)415+ int64_t remainingTotalElment)
399{416{
400 splitInfo_.inCutIndex = currentSplitIndex;417 splitInfo_.inCutIndex = currentSplitIndex;
401 splitInfo_.inUbFactor = 1;418 splitInfo_.inUbFactor = 1;
@@ -412,8 +429,8 @@ void TransposeNddmaTiling::FindSplitFactorByRateNLast(
412 }429 }
413}430}
414 431 
415-void TransposeNddmaTiling::FindSplitFactorByMultiplesLast(432+void TransposeNddmaTiling::FindSplitFactorByMultiplesLast(int64_t currentSplitIndex, int64_t currentShapeDim,
416- int64_t currentSplitIndex, int64_t currentShapeDim, int64_t remainingTotalElment, int64_t coreNumMultiples)433+ int64_t remainingTotalElment, int64_t coreNumMultiples)
417{434{
418 splitInfo_.outCutIndex = currentSplitIndex;435 splitInfo_.outCutIndex = currentSplitIndex;
419 int64_t bestI = 1;436 int64_t bestI = 1;
@@ -435,8 +452,8 @@ void TransposeNddmaTiling::FindSplitFactorByMultiplesLast(
435 splitInfo_.outUbActual *= bestI;452 splitInfo_.outUbActual *= bestI;
436}453}
437 454 
438-void TransposeNddmaTiling::FindSplitFactorByMultiplesNLast(455+void TransposeNddmaTiling::FindSplitFactorByMultiplesNLast(int64_t currentSplitIndex, int64_t currentInShapeDim,
439- int64_t currentSplitIndex, int64_t currentInShapeDim, int64_t remainingTotalElment, int64_t coreNumMultiples)456+ int64_t remainingTotalElment, int64_t coreNumMultiples)
440{457{
441 splitInfo_.inCutIndex = currentSplitIndex;458 splitInfo_.inCutIndex = currentSplitIndex;
442 for (int64_t i = splitInfo_.inUbElement; i >= 1; i--) {459 for (int64_t i = splitInfo_.inUbElement; i >= 1; i--) {
@@ -471,16 +488,16 @@ void TransposeNddmaTiling::DoSplitUB()
471 }488 }
472 int64_t currentShapeDim = shapeInfo_.reducedOutShape[currentSplitIndex];489 int64_t currentShapeDim = shapeInfo_.reducedOutShape[currentSplitIndex];
473 if (shapeInfo_.reducedPerm[currentSplitIndex] == splitInfo_.inCutIndex) {490 if (shapeInfo_.reducedPerm[currentSplitIndex] == splitInfo_.inCutIndex) {
474- currentShapeDim =491+ currentShapeDim = Ops::Base::CeilDiv(shapeInfo_.reducedInShape[splitInfo_.inCutIndex],
475- Ops::Base::CeilDiv(shapeInfo_.reducedInShape[splitInfo_.inCutIndex], splitInfo_.inUbFactor);492+ splitInfo_.inUbFactor);
476 }493 }
477 remainingTotalElment /= currentShapeDim;494 remainingTotalElment /= currentShapeDim;
478 int64_t coreNumTmp = remainingTotalElment * Ops::Base::CeilDiv(currentShapeDim, splitInfo_.outUbElement);495 int64_t coreNumTmp = remainingTotalElment * Ops::Base::CeilDiv(currentShapeDim, splitInfo_.outUbElement);
479 if (splitInfo_.outUbElement < currentShapeDim) {496 if (splitInfo_.outUbElement < currentShapeDim) {
480 if (coreNumTmp > coreNum_) { // use full coreNum497 if (coreNumTmp > coreNum_) { // use full coreNum
481 int64_t coreNumMultiples = Ops::Base::FloorDiv(coreNumTmp, coreNum_);498 int64_t coreNumMultiples = Ops::Base::FloorDiv(coreNumTmp, coreNum_);
482- FindSplitFactorByMultiplesLast(499+ FindSplitFactorByMultiplesLast(currentSplitIndex, currentShapeDim, remainingTotalElment,
483- currentSplitIndex, currentShapeDim, remainingTotalElment, coreNumMultiples);500+ coreNumMultiples);
484 } else {501 } else {
485 splitInfo_.outCutIndex = currentSplitIndex;502 splitInfo_.outCutIndex = currentSplitIndex;
486 splitInfo_.outUbFactor = splitInfo_.outUbElement;503 splitInfo_.outUbFactor = splitInfo_.outUbElement;
@@ -593,7 +610,8 @@ void TransposeNddmaTiling::EntryTilingTemplate()
593 auto platformInfo = tilingContext_->GetPlatformInfo();610 auto platformInfo = tilingContext_->GetPlatformInfo();
594 auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);611 auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);
595 auto arch = ascendcPlatform.GetCurNpuArch();612 auto arch = ascendcPlatform.GetCurNpuArch();
596- if (arch == NpuArch::DAV_5102 && !shapeInfo_.isLastAxisTranspose) {613+ if (arch == NpuArch::DAV_5102 && !shapeInfo_.isLastAxisTranspose &&
614+ shapeInfo_.reducedInShape[shapeInfo_.dim - 1] >= MOVEALIGN_LAST_MIN_ELE) {
597 SMALL_SHAPE_BYTES_THRES_HOLD = SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102_NLAST;615 SMALL_SHAPE_BYTES_THRES_HOLD = SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102_NLAST;
598 }616 }
599 if (shapeInfo_.totalVolumeActual * shapeInfo_.eleLenInBytes >= SMALL_SHAPE_BYTES_THRES_HOLD) {617 if (shapeInfo_.totalVolumeActual * shapeInfo_.eleLenInBytes >= SMALL_SHAPE_BYTES_THRES_HOLD) {
@@ -653,8 +671,8 @@ void TransposeNddmaTiling::CalcBlockSplitInfoForTensorMove()
653 }671 }
654}672}
655 673 
656-int64_t TransposeNddmaTiling::CalcBlockSplitInfoForNoCutForMultiCore(674+int64_t TransposeNddmaTiling::CalcBlockSplitInfoForNoCutForMultiCore(int64_t i, int64_t shapeSizeByte,
657- int64_t i, int64_t shapeSizeByte, int64_t& totalElment)675+ int64_t& totalElment)
658{676{
659 for (int64_t j = 2; j <= shapeInfo_.reducedOutShape[i]; j++) {677 for (int64_t j = 2; j <= shapeInfo_.reducedOutShape[i]; j++) {
660 if ((shapeInfo_.reducedOutShape[i] % j == 0) &&678 if ((shapeInfo_.reducedOutShape[i] % j == 0) &&
@@ -662,8 +680,8 @@ int64_t TransposeNddmaTiling::CalcBlockSplitInfoForNoCutForMultiCore(
662 if (j == shapeInfo_.reducedOutShape[i] && i == 0) {680 if (j == shapeInfo_.reducedOutShape[i] && i == 0) {
663 // 素数且切到了最后，正常切681 // 素数且切到了最后，正常切
664 splitInfo_.outCutIndex = i;682 splitInfo_.outCutIndex = i;
665- splitInfo_.outUbFactor =683+ splitInfo_.outUbFactor = Ops::Base::CeilDiv(cacheLineSize_ + 1,
666- Ops::Base::CeilDiv(cacheLineSize_ + 1, shapeSizeByte / shapeInfo_.reducedOutShape[i]);684+ shapeSizeByte / shapeInfo_.reducedOutShape[i]);
667 splitInfo_.outTailFactor = shapeInfo_.reducedOutShape[i] % splitInfo_.outUbFactor;685 splitInfo_.outTailFactor = shapeInfo_.reducedOutShape[i] % splitInfo_.outUbFactor;
668 totalElment *= Ops::Base::CeilDiv(shapeInfo_.reducedOutShape[i], splitInfo_.outUbFactor);686 totalElment *= Ops::Base::CeilDiv(shapeInfo_.reducedOutShape[i], splitInfo_.outUbFactor);
669 break;687 break;
@@ -698,12 +716,12 @@ void TransposeNddmaTiling::CalcBlockSplitInfoForSmallShape()
698 }716 }
699 // simt every core elemets align to 128Byte717 // simt every core elemets align to 128Byte
700 int64_t blkFactor = totalElements / coreNum_;718 int64_t blkFactor = totalElements / coreNum_;
701- int64_t ceilAlignFactor =719+ int64_t ceilAlignFactor = Ops::Base::CeilDiv(blkFactor * shapeInfo_.eleLenInBytes,
702- Ops::Base::CeilDiv(blkFactor * shapeInfo_.eleLenInBytes, SMALL_SHAPE_SPLIT_BYTES_ALIGN_SIZE) *720+ SMALL_SHAPE_SPLIT_BYTES_ALIGN_SIZE) *
703- SMALL_SHAPE_SPLIT_BYTES_ALIGN_SIZE / shapeInfo_.eleLenInBytes;721+ SMALL_SHAPE_SPLIT_BYTES_ALIGN_SIZE / shapeInfo_.eleLenInBytes;
704- int64_t floorAlignFactor =722+ int64_t floorAlignFactor = Ops::Base::FloorDiv(blkFactor * shapeInfo_.eleLenInBytes,
705- Ops::Base::FloorDiv(blkFactor * shapeInfo_.eleLenInBytes, SMALL_SHAPE_SPLIT_BYTES_ALIGN_SIZE) *723+ SMALL_SHAPE_SPLIT_BYTES_ALIGN_SIZE) *
706- SMALL_SHAPE_SPLIT_BYTES_ALIGN_SIZE / shapeInfo_.eleLenInBytes;724+ SMALL_SHAPE_SPLIT_BYTES_ALIGN_SIZE / shapeInfo_.eleLenInBytes;
707 if (totalElements - floorAlignFactor * (coreNum_ - 1) <= floorAlignFactor) {725 if (totalElements - floorAlignFactor * (coreNum_ - 1) <= floorAlignFactor) {
708 realCoreNum_ = coreNum_;726 realCoreNum_ = coreNum_;
709 blkFactor_ = floorAlignFactor;727 blkFactor_ = floorAlignFactor;
@@ -735,11 +753,11 @@ void TransposeNddmaTiling::CalcBlockSplitInfoForNLastTranspose()
735 break;753 break;
736 } else { // use full coreNum754 } else { // use full coreNum
737 int64_t coreNumMultiples = Ops::Base::FloorDiv(coreNumTmp, coreNum_);755 int64_t coreNumMultiples = Ops::Base::FloorDiv(coreNumTmp, coreNum_);
738- FindSplitFactorByMultiplesNLast(756+ FindSplitFactorByMultiplesNLast(currentSplitIndex, currentInShapeDim, remainingTotalElment,
739- currentSplitIndex, currentInShapeDim, remainingTotalElment, coreNumMultiples);757+ coreNumMultiples);
740 /* 检查inUbFactor是否合法，不合法则执行退轴逻辑 */758 /* 检查inUbFactor是否合法，不合法则执行退轴逻辑 */
741- CheckInUbFactorValid(759+ CheckInUbFactorValid(currentSplitIndex, currentInShapeDim, remainingTotalElment, coreNumMultiples,
742- currentSplitIndex, currentInShapeDim, remainingTotalElment, coreNumMultiples, solvedTotalElment);760+ solvedTotalElment);
743 break;761 break;
744 }762 }
745 } else if (coreNumTmp < coreNum_) { // use at least VEC_CORE_USED_THRES_HOLD * coreNum763 } else if (coreNumTmp < coreNum_) { // use at least VEC_CORE_USED_THRES_HOLD * coreNum
@@ -757,9 +775,9 @@ void TransposeNddmaTiling::CalcBlockSplitInfoForNLastTranspose()
757 SetRealCoreNumAndBlkFactor(coreNum);775 SetRealCoreNumAndBlkFactor(coreNum);
758}776}
759 777 
760-void TransposeNddmaTiling::CheckInUbFactorValid(778+void TransposeNddmaTiling::CheckInUbFactorValid(int64_t& currentSplitIndex, int64_t& currentInShapeDim,
761- int64_t& currentSplitIndex, int64_t& currentInShapeDim, int64_t& remainingTotalElment, int64_t& coreNumMultiples,779+ int64_t& remainingTotalElment, int64_t& coreNumMultiples,
762- int64_t* solvedTotalElment)780+ int64_t* solvedTotalElment)
763{781{
764 if (splitInfo_.inUbFactor == 0 && currentSplitIndex < shapeInfo_.dim - 1) {782 if (splitInfo_.inUbFactor == 0 && currentSplitIndex < shapeInfo_.dim - 1) {
765 while (currentSplitIndex < shapeInfo_.dim - 1) {783 while (currentSplitIndex < shapeInfo_.dim - 1) {
@@ -768,8 +786,8 @@ void TransposeNddmaTiling::CheckInUbFactorValid(
768 splitInfo_.inUbElement = shapeInfo_.reducedInShape[currentSplitIndex];786 splitInfo_.inUbElement = shapeInfo_.reducedInShape[currentSplitIndex];
769 remainingTotalElment = shapeInfo_.totalVolumeActual / solvedTotalElment[currentSplitIndex];787 remainingTotalElment = shapeInfo_.totalVolumeActual / solvedTotalElment[currentSplitIndex];
770 coreNumMultiples = remainingTotalElment;788 coreNumMultiples = remainingTotalElment;
771- FindSplitFactorByMultiplesNLast(789+ FindSplitFactorByMultiplesNLast(currentSplitIndex, currentInShapeDim, remainingTotalElment,
772- currentSplitIndex, currentInShapeDim, remainingTotalElment, coreNumMultiples);790+ coreNumMultiples);
773 if (splitInfo_.inUbFactor > 0) {791 if (splitInfo_.inUbFactor > 0) {
774 break;792 break;
775 }793 }
@@ -851,8 +869,8 @@ void TransposeNddmaTiling::CalcBlockSplitInfoForCutTwice()
851 outAxiseExceptSplitInAxis *= shapeInfo_.reducedOutShape[i];869 outAxiseExceptSplitInAxis *= shapeInfo_.reducedOutShape[i];
852 }870 }
853 }871 }
854- outAxiseExceptSplitInAxis *=872+ outAxiseExceptSplitInAxis *= Ops::Base::CeilDiv(shapeInfo_.reducedOutShape[splitInfo_.outCutIndex],
855- Ops::Base::CeilDiv(shapeInfo_.reducedOutShape[splitInfo_.outCutIndex], splitInfo_.outUbFactor);873+ splitInfo_.outUbFactor);
856 int64_t inUbAxis = Ops::Base::CeilDiv(shapeInfo_.reducedInShape[splitInfo_.inCutIndex], splitInfo_.inUbFactor);874 int64_t inUbAxis = Ops::Base::CeilDiv(shapeInfo_.reducedInShape[splitInfo_.inCutIndex], splitInfo_.inUbFactor);
857 if (outAxiseExceptSplitInAxis * inUbAxis < coreNum_) {875 if (outAxiseExceptSplitInAxis * inUbAxis < coreNum_) {
858 // use at least VEC_CORE_USED_THRES_HOLD * coreNum876 // use at least VEC_CORE_USED_THRES_HOLD * coreNum
@@ -1064,14 +1082,14 @@ void TransposeNddmaTiling::CalcInUbShapeInfoForCutTwice()
1064 inUbOutputTailSrcShape_[idx] = inUbInputTailSrcShape_[idx];1082 inUbOutputTailSrcShape_[idx] = inUbInputTailSrcShape_[idx];
1065 }1083 }
1066 inUbOutputTailSrcShape_[splitInfo_.inCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim] = splitInfo_.inUbFactor;1084 inUbOutputTailSrcShape_[splitInfo_.inCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim] = splitInfo_.inUbFactor;
1067- inUbOutputTailSrcShape_[expandedPerm_[splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim]] =1085+ inUbOutputTailSrcShape_[expandedPerm_[splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM -
1068- splitInfo_.outTailFactor;1086+ shapeInfo_.dim]] = splitInfo_.outTailFactor;
1069- inUbMainSrcShape_[expandedPerm_[splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim]] =1087+ inUbMainSrcShape_[expandedPerm_[splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim]] = splitInfo_
1070- splitInfo_.outUbFactor;1088+ .outUbFactor;
1071- inUbInputTailSrcShape_[expandedPerm_[splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim]] =1089+ inUbInputTailSrcShape_[expandedPerm_[splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM -
1072- splitInfo_.outUbFactor;1090+ shapeInfo_.dim]] = splitInfo_.outUbFactor;
1073- inUbTailSrcShape_[expandedPerm_[splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim]] =1091+ inUbTailSrcShape_[expandedPerm_[splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim]] = splitInfo_
1074- splitInfo_.outTailFactor;1092+ .outTailFactor;
1075 for (int64_t idx = splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim + 1; idx < NDDMA_MAX_DIM_NUM;1093 for (int64_t idx = splitInfo_.outCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim + 1; idx < NDDMA_MAX_DIM_NUM;
1076 idx++) {1094 idx++) {
1077 if (expandedPerm_[idx] == splitInfo_.inCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim) {1095 if (expandedPerm_[idx] == splitInfo_.inCutIndex + NDDMA_MAX_DIM_NUM - shapeInfo_.dim) {
@@ -1110,29 +1128,37 @@ void TransposeNddmaTiling::GetIntervalInfoForCutTwice()
1110 1128 
1111 if (splitInfo_.inTailFactor != 0 && splitInfo_.outTailFactor != 0) {1129 if (splitInfo_.inTailFactor != 0 && splitInfo_.outTailFactor != 0) {
1112 offsetRangeInputTail_.start = offsetRangeMain_.end + 1;1130 offsetRangeInputTail_.start = offsetRangeMain_.end + 1;
1113- offsetRangeInputTail_.end =1131+ offsetRangeInputTail_.end = offsetRangeInputTail_.start +
1114- offsetRangeInputTail_.start +1132+ (expandedInputShape_[inputOutputCutIndex] /
1115- (expandedInputShape_[inputOutputCutIndex] / inUbMainSrcShape_[inputOutputCutIndex]) * outUbLoop - 1;1133+ inUbMainSrcShape_[inputOutputCutIndex]) *
1134+ outUbLoop -
1135+ 1;
1116 offsetRangeOutputTail_.start = offsetRangeInputTail_.end + 1;1136 offsetRangeOutputTail_.start = offsetRangeInputTail_.end + 1;
1117- offsetRangeOutputTail_.end =1137+ offsetRangeOutputTail_.end = offsetRangeOutputTail_.start +
1118- offsetRangeOutputTail_.start +1138+ (expandedInputShape_[expandedInputCutIndex] /
1119- (expandedInputShape_[expandedInputCutIndex] / inUbMainSrcShape_[expandedInputCutIndex]) * outUbLoop - 1;1139+ inUbMainSrcShape_[expandedInputCutIndex]) *
1140+ outUbLoop -
1141+ 1;
1120 offsetRangeTail_.start = offsetRangeOutputTail_.end + 1;1142 offsetRangeTail_.start = offsetRangeOutputTail_.end + 1;
1121 offsetRangeTail_.end = offsetRangeTail_.start + outUbLoop - 1;1143 offsetRangeTail_.end = offsetRangeTail_.start + outUbLoop - 1;
1122 }1144 }
1123 1145 
1124 if (splitInfo_.inTailFactor != 0 && splitInfo_.outTailFactor == 0) {1146 if (splitInfo_.inTailFactor != 0 && splitInfo_.outTailFactor == 0) {
1125 offsetRangeInputTail_.start = offsetRangeMain_.end + 1;1147 offsetRangeInputTail_.start = offsetRangeMain_.end + 1;
1126- offsetRangeInputTail_.end =1148+ offsetRangeInputTail_.end = offsetRangeInputTail_.start +
1127- offsetRangeInputTail_.start +1149+ (expandedInputShape_[inputOutputCutIndex] /
1128- (expandedInputShape_[inputOutputCutIndex] / inUbMainSrcShape_[inputOutputCutIndex]) * outUbLoop - 1;1150+ inUbMainSrcShape_[inputOutputCutIndex]) *
1151+ outUbLoop -
1152+ 1;
1129 }1153 }
1130 1154 
1131 if (splitInfo_.inTailFactor == 0 && splitInfo_.outTailFactor != 0) {1155 if (splitInfo_.inTailFactor == 0 && splitInfo_.outTailFactor != 0) {
1132 offsetRangeOutputTail_.start = offsetRangeMain_.end + 1;1156 offsetRangeOutputTail_.start = offsetRangeMain_.end + 1;
1133- offsetRangeOutputTail_.end =1157+ offsetRangeOutputTail_.end = offsetRangeOutputTail_.start +
1134- offsetRangeOutputTail_.start +1158+ (expandedInputShape_[expandedInputCutIndex] /
1135- (expandedInputShape_[expandedInputCutIndex] / inUbMainSrcShape_[expandedInputCutIndex]) * outUbLoop - 1;1159+ inUbMainSrcShape_[expandedInputCutIndex]) *
1160+ outUbLoop -
1161+ 1;
1136 }1162 }
1137}1163}
1138 1164 
@@ -1185,8 +1211,8 @@ void TransposeNddmaTiling::FillTilingData()
1185 tilingData_.transposeOpTiling.set_inUbTailDstShape(inUbTailDstShape_);1211 tilingData_.transposeOpTiling.set_inUbTailDstShape(inUbTailDstShape_);
1186 1212 
1187 if (!isReleatedTranspsoe_) {1213 if (!isReleatedTranspsoe_) {
1188- tilingData_.SaveToBuffer(1214+ tilingData_.SaveToBuffer(tilingContext_->GetRawTilingData()->GetData(),
1189- tilingContext_->GetRawTilingData()->GetData(), tilingContext_->GetRawTilingData()->GetCapacity());1215+ tilingContext_->GetRawTilingData()->GetCapacity());
1190 tilingContext_->GetRawTilingData()->SetDataSize(tilingData_.GetDataSize());1216 tilingContext_->GetRawTilingData()->SetDataSize(tilingData_.GetDataSize());
1191 }1217 }
1192}1218}
@@ -1195,33 +1221,31 @@ void TransposeNddmaTiling::PrintTilingData()
1195{1221{
1196 OP_LOGI(tilingContext_->GetNodeName(), "Entering PrintTilingData.");1222 OP_LOGI(tilingContext_->GetNodeName(), "Entering PrintTilingData.");
1197 for (int64_t i = 0; i < shapeInfo_.dim; i++) {1223 for (int64_t i = 0; i < shapeInfo_.dim; i++) {
1198- OP_LOGI(1224+ OP_LOGI(tilingContext_->GetNodeName(),
1199- tilingContext_->GetNodeName(),1225+ "reducedInShape[%ld] is:%ld, reducedOutShape[%ld]:%ld, reducedPerm[%ld]:%ld. \
1200- "reducedInShape[%ld] is:%ld, reducedOutShape[%ld]:%ld, reducedPerm[%ld]:%ld. \
1201 baseInShape[%ld] is:%ld",1226 baseInShape[%ld] is:%ld",
1202- i, inputShape_[i], i, outputShape_[i], i, perm_[i], i, baseInShape_[i]);1227+ i, inputShape_[i], i, outputShape_[i], i, perm_[i], i, baseInShape_[i]);
1203 }1228 }
1204 for (int64_t i = 0; i < NDDMA_MAX_DIM_NUM; i++) {1229 for (int64_t i = 0; i < NDDMA_MAX_DIM_NUM; i++) {
1205- OP_LOGI(1230+ OP_LOGI(tilingContext_->GetNodeName(), "baseNddmaShape_[%ld] is:%ld, nddmaIdx_[%ld]:%ld", i, baseNddmaShape_[i],
1206- tilingContext_->GetNodeName(), "baseNddmaShape_[%ld] is:%ld, nddmaIdx_[%ld]:%ld", i, baseNddmaShape_[i], i,1231+ i, nddmaIdx_[i]);
1207- nddmaIdx_[i]);
1208 }1232 }
1209- OP_LOGI(1233+ OP_LOGI(tilingContext_->GetNodeName(),
1210- tilingContext_->GetNodeName(),1234+ "tilingData is permSize:%ld, inCutIndex:%ld, outCutIndex:%ld, inUbFactor:%ld, outUbFactor:%ld, \
1211- "tilingData is permSize:%ld, inCutIndex:%ld, outCutIndex:%ld, inUbFactor:%ld, outUbFactor:%ld, \
1212 inTailFactor:%ld, outTailFactor:%ld, realCoreNum:%ld, blkFactor:%ld, blkTailFactor:%ld, \1235 inTailFactor:%ld, outTailFactor:%ld, realCoreNum:%ld, blkFactor:%ld, blkTailFactor:%ld, \
1213 ubSize:%ld, totalNddmaNum:%ld, Tiling4Transpose ends. ",1236 ubSize:%ld, totalNddmaNum:%ld, Tiling4Transpose ends. ",
1214- tilingData_.transposeOpTiling.get_permSize(), tilingData_.transposeOpTiling.get_inCutIndex(),1237+ tilingData_.transposeOpTiling.get_permSize(), tilingData_.transposeOpTiling.get_inCutIndex(),
1215- tilingData_.transposeOpTiling.get_outCutIndex(), tilingData_.transposeOpTiling.get_inUbFactor(),1238+ tilingData_.transposeOpTiling.get_outCutIndex(), tilingData_.transposeOpTiling.get_inUbFactor(),
1216- tilingData_.transposeOpTiling.get_outUbFactor(), tilingData_.transposeOpTiling.get_inTailFactor(),1239+ tilingData_.transposeOpTiling.get_outUbFactor(), tilingData_.transposeOpTiling.get_inTailFactor(),
1217- tilingData_.transposeOpTiling.get_outTailFactor(), tilingData_.transposeOpTiling.get_realCoreNum(),1240+ tilingData_.transposeOpTiling.get_outTailFactor(), tilingData_.transposeOpTiling.get_realCoreNum(),
1218- tilingData_.transposeOpTiling.get_blkFactor(), tilingData_.transposeOpTiling.get_blkTailFactor(),1241+ tilingData_.transposeOpTiling.get_blkFactor(), tilingData_.transposeOpTiling.get_blkTailFactor(),
1219- tilingData_.transposeOpTiling.get_ubSize(), tilingData_.transposeOpTiling.get_totalNddmaNum());1242+ tilingData_.transposeOpTiling.get_ubSize(), tilingData_.transposeOpTiling.get_totalNddmaNum());
1220}1243}
1221 1244 
1222-ge::graphStatus TransposeNddmaTiling::TilingForReleatedTranspose(1245+ge::graphStatus TransposeNddmaTiling::TilingForReleatedTranspose(gert::TilingContext* context,
1223- gert::TilingContext* context, TransposeOpTilingData* tilingData, TransposeCompilerInfo* compilerInfo,1246+ TransposeOpTilingData* tilingData,
1224- ShapeInfo& opInput)1247+ TransposeCompilerInfo* compilerInfo,
1248+ ShapeInfo& opInput)
1225{1249{
1226 OP_LOGD(context->GetNodeName(), "Start TilingForReleatedTranspose.");1250 OP_LOGD(context->GetNodeName(), "Start TilingForReleatedTranspose.");
1227 TransposeNddmaTiling tilingObject(context);1251 TransposeNddmaTiling tilingObject(context);
@@ -1259,17 +1283,15 @@ ge::graphStatus TilingPrepareTransposeForAscendC(gert::TilingParseContext* conte
1259 OP_CHECK_NULL_WITH_CONTEXT(context, platformInfo);1283 OP_CHECK_NULL_WITH_CONTEXT(context, platformInfo);
1260 auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);1284 auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);
1261 ci->coreNum = ascendcPlatform.GetCoreNumAiv();1285 ci->coreNum = ascendcPlatform.GetCoreNumAiv();
1262- OP_CHECK_IF(1286+ OP_CHECK_IF((ci->coreNum <= 0),
1263- (ci->coreNum <= 0),1287+ OP_LOGE(context->GetNodeName(), "Transpose Op GetHardwareInfo Failed, coreNum:%ld.", ci->coreNum),
1264- OP_LOGE(context->GetNodeName(), "Transpose Op GetHardwareInfo Failed, coreNum:%ld.", ci->coreNum),1288+ return ge::GRAPH_FAILED);
1265- return ge::GRAPH_FAILED);
1266 uint64_t ubSize;1289 uint64_t ubSize;
1267 ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSize);1290 ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSize);
1268 ci->ubSize = static_cast<int64_t>(ubSize);1291 ci->ubSize = static_cast<int64_t>(ubSize);
1269- OP_CHECK_IF(1292+ OP_CHECK_IF((ci->ubSize <= 0),
1270- (ci->ubSize <= 0),1293+ OP_LOGE(context->GetNodeName(), "Transpose Op GetHardwareInfo Failed, ubSize:%ld.", ci->ubSize),
1271- OP_LOGE(context->GetNodeName(), "Transpose Op GetHardwareInfo Failed, ubSize:%ld.", ci->ubSize),1294+ return ge::GRAPH_FAILED);
1272- return ge::GRAPH_FAILED);
1273 1295 
1274 OP_LOGD(context->GetNodeName(), "Transpose Op get coreNum:%ld, ubSize:%ld.", ci->coreNum, ci->ubSize);1296 OP_LOGD(context->GetNodeName(), "Transpose Op get coreNum:%ld, ubSize:%ld.", ci->coreNum, ci->ubSize);
1275 return ge::GRAPH_SUCCESS;1297 return ge::GRAPH_SUCCESS;
Mconversion/transpose/op_host/arch35/transpose_tiling_arch35.h+13-13
@@ -59,6 +59,8 @@ constexpr int64_t DIM_FOUR = 4;
59constexpr int64_t DIM_FIVE = 5;59constexpr int64_t DIM_FIVE = 5;
60constexpr int64_t DIM_SIX = 6;60constexpr int64_t DIM_SIX = 6;
61constexpr int64_t DIM_EIGHT = 8;61constexpr int64_t DIM_EIGHT = 8;
62+constexpr int64_t HW_ALIGN = 16;
63+constexpr int64_t HW_MIN_PRODUCT = 448;
62constexpr int64_t SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102 = 1000000;64constexpr int64_t SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102 = 1000000;
63constexpr int64_t SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102_NLAST = 400000;65constexpr int64_t SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102_NLAST = 400000;
64constexpr int64_t SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102_021 = 70000;66constexpr int64_t SMALL_SHAPE_BYTES_THRES_HOLD_DAV_5102_021 = 70000;
@@ -107,8 +109,7 @@ TILING_DATA_FIELD_DEF_STRUCT(TransposeOpTilingData, transposeOpTiling);
107END_TILING_DATA_DEF;109END_TILING_DATA_DEF;
108REGISTER_TILING_DATA_CLASS(Transpose, TransposeTilingData);110REGISTER_TILING_DATA_CLASS(Transpose, TransposeTilingData);
109 111 
110-enum class SplitMode : int64_t112+enum class SplitMode : int64_t {
111-{
112 TENSOR_MOVE = 10000, // only one axis after fuse113 TENSOR_MOVE = 10000, // only one axis after fuse
113 SMALL_SHAPE = 10001, // UB is enough114 SMALL_SHAPE = 10001, // UB is enough
114 CUT_ONCE = 10002, // cut one axis and last transpose115 CUT_ONCE = 10002, // cut one axis and last transpose
@@ -158,12 +159,11 @@ ge::graphStatus TilingPrepareTransposeForAscendC(gert::TilingParseContext* conte
158 159 
159class TransposeNddmaTiling {160class TransposeNddmaTiling {
160public:161public:
161- explicit TransposeNddmaTiling(gert::TilingContext* context) : tilingContext_(context) {};162+ explicit TransposeNddmaTiling(gert::TilingContext* context) : tilingContext_(context){};
162 ge::graphStatus Init(const int64_t& coreNum, const int64_t& ubSize);163 ge::graphStatus Init(const int64_t& coreNum, const int64_t& ubSize);
163 ge::graphStatus RunTranposelTiling();164 ge::graphStatus RunTranposelTiling();
164- ge::graphStatus TilingForReleatedTranspose(165+ ge::graphStatus TilingForReleatedTranspose(gert::TilingContext* context, TransposeOpTilingData* tilingData,
165- gert::TilingContext* context, TransposeOpTilingData* tilingData, TransposeCompilerInfo* compilerInfo,166+ TransposeCompilerInfo* compilerInfo, ShapeInfo& opInput);
166- ShapeInfo& opInput);
167 167 
168private:168private:
169 template <typename T>169 template <typename T>
@@ -175,6 +175,7 @@ private:
175 ge::graphStatus CheckShapeInfo();175 ge::graphStatus CheckShapeInfo();
176 ge::graphStatus CheckReducedShapeInfo();176 ge::graphStatus CheckReducedShapeInfo();
177 ge::graphStatus TryVCONVTiling();177 ge::graphStatus TryVCONVTiling();
178+ bool Is021VConvValid();
178 void FlushBaseNumForBigDim();179 void FlushBaseNumForBigDim();
179 void EntryTilingTemplate();180 void EntryTilingTemplate();
180 void CalcUBSplitInfo();181 void CalcUBSplitInfo();
@@ -194,14 +195,13 @@ private:
194 int64_t FindOutIndex(int64_t index);195 int64_t FindOutIndex(int64_t index);
195 bool UbOutOfBoundCheck(int64_t currentSplitIndex, int64_t currentSplitValue, bool calcIn);196 bool UbOutOfBoundCheck(int64_t currentSplitIndex, int64_t currentSplitValue, bool calcIn);
196 bool UbOutOfBoundCheckNLast(int64_t currentSplitIndex, int64_t currentSplitValue);197 bool UbOutOfBoundCheckNLast(int64_t currentSplitIndex, int64_t currentSplitValue);
197- void FindSplitFactorByMultiplesLast(198+ void FindSplitFactorByMultiplesLast(int64_t currentSplitIndex, int64_t currentInShapeDim,
198- int64_t currentSplitIndex, int64_t currentInShapeDim, int64_t remainingTotalElment, int64_t coreNumMultiples);199+ int64_t remainingTotalElment, int64_t coreNumMultiples);
199 void FindSplitFactorByRateNLast(int64_t currentSplitIndex, int64_t currentInShapeDim, int64_t remainingTotalElment);200 void FindSplitFactorByRateNLast(int64_t currentSplitIndex, int64_t currentInShapeDim, int64_t remainingTotalElment);
200- void FindSplitFactorByMultiplesNLast(201+ void FindSplitFactorByMultiplesNLast(int64_t currentSplitIndex, int64_t currentInShapeDim,
201- int64_t currentSplitIndex, int64_t currentInShapeDim, int64_t remainingTotalElment, int64_t coreNumMultiples);202+ int64_t remainingTotalElment, int64_t coreNumMultiples);
202- void CheckInUbFactorValid(203+ void CheckInUbFactorValid(int64_t& currentSplitIndex, int64_t& currentInShapeDim, int64_t& remainingTotalElment,
203- int64_t& currentSplitIndex, int64_t& currentInShapeDim, int64_t& remainingTotalElment,204+ int64_t& coreNumMultiples, int64_t* solvedTotalElment);
204- int64_t& coreNumMultiples, int64_t* solvedTotalElment);
205 void DoSplitUBBigDim();205 void DoSplitUBBigDim();
206 void NDDMADimExpand();206 void NDDMADimExpand();
207 void GetInUbShapeInfo();207 void GetInUbShapeInfo();