Star1296
1410
代码
代码
Issues108
Pull Requests213
Actions
项目讨论
Wiki
分析
Star1296
1410
已合并
【bugfix】修复unfoldGrad的同步问题 #2967
majiajian_hw创建于 5月27日
【bugfix】修复unfoldGrad的同步问题 #2967
已合并
majiajian_hw创建于 5月27日
从nunnons2/ops-mathbugfix_unfoldgrad_sync合入到cann/ops-mathmaster
共 4 个文件变更+67-23
Mconversion/unfold_grad/op_kernel/unfold_grad.h+21-8
@@ -93,6 +93,14 @@ public:
93 params.paramsOut = {blockCount, blockLenOut, 0, 0, 0};93 params.paramsOut = {blockCount, blockLenOut, 0, 0, 0};
94 }94 }
95 95 
96+ template <AscendC::HardEvent hardEvent>
97+ __aicore__ inline void PipeSync()
98+ {
99+ int32_t eventID = static_cast<int32_t>(GetTPipePtr()->FetchEventID(hardEvent));
100+ AscendC::SetFlag<hardEvent>(eventID);
101+ AscendC::WaitFlag<hardEvent>(eventID);
102+ }
103+ 
96 __aicore__ inline void CopyToOutBigShapeOnePage(int64_t inPageIdx, int64_t outPageIdx, sDataCopyExtParams& params)104 __aicore__ inline void CopyToOutBigShapeOnePage(int64_t inPageIdx, int64_t outPageIdx, sDataCopyExtParams& params)
97 {105 {
98 int64_t inOffset = inPageIdx * workspaceLen_;106 int64_t inOffset = inPageIdx * workspaceLen_;
@@ -111,15 +119,16 @@ public:
111 AscendC::Duplicate<T1>(zeroLocal, 0, zeroNum);119 AscendC::Duplicate<T1>(zeroLocal, 0, zeroNum);
112 uint32_t blockLen = static_cast<uint32_t>(T2DstDataSize);120 uint32_t blockLen = static_cast<uint32_t>(T2DstDataSize);
113 AscendC::DataCopyExtParams copyParams = {1, blockLen, 0, 0, 0};121 AscendC::DataCopyExtParams copyParams = {1, blockLen, 0, 0, 0};
114- AscendC::PipeBarrier<PIPE_ALL>();122+ PipeSync<AscendC::HardEvent::V_MTE3>();
115 for (int i = 0; i < loop; i++) {123 for (int i = 0; i < loop; i++) {
116 AscendC::DataCopyPad(dstGlobal[dstGlobalStart + i * zeroNum], zeroLocal, copyParams);124 AscendC::DataCopyPad(dstGlobal[dstGlobalStart + i * zeroNum], zeroLocal, copyParams);
125+ AscendC::PipeBarrier<PIPE_MTE3>();
117 }126 }
118 if (tail > 0) {127 if (tail > 0) {
119 copyParams.blockLen = static_cast<uint32_t>(tail * typeSizeT1);128 copyParams.blockLen = static_cast<uint32_t>(tail * typeSizeT1);
120 AscendC::DataCopyPad(dstGlobal[dstGlobalStart + loop * zeroNum], zeroLocal, copyParams);129 AscendC::DataCopyPad(dstGlobal[dstGlobalStart + loop * zeroNum], zeroLocal, copyParams);
121 }130 }
122- AscendC::PipeBarrier<PIPE_ALL>();131+ PipeSync<AscendC::HardEvent::MTE3_S>();
123 computeOutQueueDst.FreeTensor(zeroLocal);132 computeOutQueueDst.FreeTensor(zeroLocal);
124 }133 }
125 134 
@@ -132,15 +141,16 @@ public:
132 AscendC::Duplicate<T2>(zeroLocal, 0, zeroNum);141 AscendC::Duplicate<T2>(zeroLocal, 0, zeroNum);
133 uint32_t blockLen = static_cast<uint32_t>(T2DstDataSize);142 uint32_t blockLen = static_cast<uint32_t>(T2DstDataSize);
134 AscendC::DataCopyExtParams copyParams = {1, blockLen, 0, 0, 0};143 AscendC::DataCopyExtParams copyParams = {1, blockLen, 0, 0, 0};
135- AscendC::PipeBarrier<PIPE_ALL>();144+ PipeSync<AscendC::HardEvent::V_MTE3>();
136 for (int i = 0; i < loop; i++) {145 for (int i = 0; i < loop; i++) {
137 AscendC::DataCopyPad(workspaceT2SumRes[dstGlobalStart + i * zeroNum], zeroLocal, copyParams);146 AscendC::DataCopyPad(workspaceT2SumRes[dstGlobalStart + i * zeroNum], zeroLocal, copyParams);
147+ AscendC::PipeBarrier<PIPE_MTE3>();
138 }148 }
139 if (tail > 0) {149 if (tail > 0) {
140 copyParams.blockLen = static_cast<uint32_t>(tail * typeSizeT2);150 copyParams.blockLen = static_cast<uint32_t>(tail * typeSizeT2);
141 AscendC::DataCopyPad(workspaceT2SumRes[dstGlobalStart + loop * zeroNum], zeroLocal, copyParams);151 AscendC::DataCopyPad(workspaceT2SumRes[dstGlobalStart + loop * zeroNum], zeroLocal, copyParams);
142 }152 }
143- AscendC::PipeBarrier<PIPE_ALL>();153+ PipeSync<AscendC::HardEvent::MTE3_S>();
144 computeOutQueueDst.FreeTensor(zeroLocal);154 computeOutQueueDst.FreeTensor(zeroLocal);
145 }155 }
146 156 
@@ -171,22 +181,25 @@ public:
171 {181 {
172 AscendC::DataCopyPadExtParams<T2> padParmsT2{false, 0, 0, 0};182 AscendC::DataCopyPadExtParams<T2> padParmsT2{false, 0, 0, 0};
173 auto inLocalT2 = computeOutQueueDst.AllocTensor<T2>();183 auto inLocalT2 = computeOutQueueDst.AllocTensor<T2>();
174- AscendC::PipeBarrier<PIPE_ALL>();184+ PipeSync<AscendC::HardEvent::S_MTE2>();
185+ PipeSync<AscendC::HardEvent::V_MTE2>();
186+ PipeSync<AscendC::HardEvent::MTE3_MTE2>();
175 AscendC::DataCopyPad(inLocalT2, srcGM[inCopyParams.offset], inCopyParams.dcParams, padParmsT2);187 AscendC::DataCopyPad(inLocalT2, srcGM[inCopyParams.offset], inCopyParams.dcParams, padParmsT2);
176- AscendC::PipeBarrier<PIPE_ALL>();188+ PipeSync<AscendC::HardEvent::MTE2_S>();
177 computeOutQueueDst.EnQue(inLocalT2);189 computeOutQueueDst.EnQue(inLocalT2);
178 190 
179 inLocalT2 = computeOutQueueDst.DeQue<T2>();191 inLocalT2 = computeOutQueueDst.DeQue<T2>();
180 auto inLocal = inQueueSrc.AllocTensor<T1>();192 auto inLocal = inQueueSrc.AllocTensor<T1>();
193+ PipeSync<AscendC::HardEvent::MTE2_V>();
181 AscendC::Cast(inLocal, inLocalT2, AscendC::RoundMode::CAST_RINT, inCopyParams.dcParams.blockLen / typeSizeT2);194 AscendC::Cast(inLocal, inLocalT2, AscendC::RoundMode::CAST_RINT, inCopyParams.dcParams.blockLen / typeSizeT2);
182 inQueueSrc.EnQue(inLocal);195 inQueueSrc.EnQue(inLocal);
183 computeOutQueueDst.FreeTensor(inLocalT2);196 computeOutQueueDst.FreeTensor(inLocalT2);
184 197 
185 inLocal = inQueueSrc.DeQue<T1>();198 inLocal = inQueueSrc.DeQue<T1>();
186- AscendC::PipeBarrier<PIPE_ALL>();199+ PipeSync<AscendC::HardEvent::V_MTE3>();
200+ PipeSync<AscendC::HardEvent::S_MTE3>();
187 AscendC::DataCopyPad(dstGM[outCopyParams.offset], inLocal, outCopyParams.dcParams);201 AscendC::DataCopyPad(dstGM[outCopyParams.offset], inLocal, outCopyParams.dcParams);
188 inQueueSrc.FreeTensor(inLocal);202 inQueueSrc.FreeTensor(inLocal);
189- AscendC::PipeBarrier<PIPE_ALL>();
190 }203 }
191 204 
192 __aicore__ inline void ParseTilingData(const UnfoldGradTilingData* tilingData)205 __aicore__ inline void ParseTilingData(const UnfoldGradTilingData* tilingData)
Mconversion/unfold_grad/op_kernel/unfold_grad_final_axe.h+19-3
@@ -43,6 +43,14 @@ public:
43 this->workspaceT2SumRes.SetGlobalBuffer(reinterpret_cast<__gm__ T2*>(workspace) + gradInBlockOffset);43 this->workspaceT2SumRes.SetGlobalBuffer(reinterpret_cast<__gm__ T2*>(workspace) + gradInBlockOffset);
44 }44 }
45 45 
46+ template <AscendC::HardEvent hardEvent>
47+ __aicore__ inline void PipeSync()
48+ {
49+ int32_t eventID = static_cast<int32_t>(GetTPipePtr()->FetchEventID(hardEvent));
50+ AscendC::SetFlag<hardEvent>(eventID);
51+ AscendC::WaitFlag<hardEvent>(eventID);
52+ }
53+ 
46 __aicore__ inline void ProcessFinalAxes(int curSrcStart, int curDstStart)54 __aicore__ inline void ProcessFinalAxes(int curSrcStart, int curDstStart)
47 {55 {
48 this->tasksOnce = this->tasksOnceMaxPerCore;56 this->tasksOnce = this->tasksOnceMaxPerCore;
@@ -78,7 +86,6 @@ public:
78 this->SetGMtoZero(this->outputNumPerCore, dstStart);86 this->SetGMtoZero(this->outputNumPerCore, dstStart);
79 }87 }
80 88 
81- AscendC::PipeBarrier<PIPE_ALL>(); // 尾轴情况
82 ProcessFinalAxes(srcStart, dstStart);89 ProcessFinalAxes(srcStart, dstStart);
83 90 
84 if constexpr (ISCAST) {91 if constexpr (ISCAST) {
@@ -86,7 +93,8 @@ public:
86 this->CalculateOutParms(params);93 this->CalculateOutParms(params);
87 this->CopyToOutBigShapeOnePage(batchIdx, batchIdx, params);94 this->CopyToOutBigShapeOnePage(batchIdx, batchIdx, params);
88 }95 }
89- AscendC::PipeBarrier<PIPE_ALL>();96+ PipeSync<AscendC::HardEvent::MTE3_V>();
97+ PipeSync<AscendC::HardEvent::S_V>();
90 }98 }
91 }99 }
92 100 
@@ -97,6 +105,8 @@ private:
97 ISCAST ? this->inQueueSrc.template AllocTensor<T1>() : this->computeInQueueSrc.template AllocTensor<T1>();105 ISCAST ? this->inQueueSrc.template AllocTensor<T1>() : this->computeInQueueSrc.template AllocTensor<T1>();
98 AscendC::DataCopyPadExtParams<T1> padParams{false, 0, 0, 0};106 AscendC::DataCopyPadExtParams<T1> padParams{false, 0, 0, 0};
99 AscendC::PipeBarrier<PIPE_V>();107 AscendC::PipeBarrier<PIPE_V>();
108+ PipeSync<AscendC::HardEvent::MTE3_V>();
109+ PipeSync<AscendC::HardEvent::S_V>();
100 T1 zeroVal(0.0);110 T1 zeroVal(0.0);
101 int srcDataSize = ISCAST ? this->ubSizeT1 : this->T2SrcDataSize;111 int srcDataSize = ISCAST ? this->ubSizeT1 : this->T2SrcDataSize;
102 AscendC::Duplicate<T1>(srcLocal, zeroVal, srcDataSize / this->typeSizeT1);112 AscendC::Duplicate<T1>(srcLocal, zeroVal, srcDataSize / this->typeSizeT1);
@@ -109,7 +119,7 @@ private:
109 uint32_t srcStride = 0;119 uint32_t srcStride = 0;
110 uint32_t dstStride =120 uint32_t dstStride =
111 (colNumSpace - (colHandleNum * this->typeSizeT1 + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE) / BLOCK_SIZE;121 (colNumSpace - (colHandleNum * this->typeSizeT1 + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE) / BLOCK_SIZE;
112- AscendC::PipeBarrier<PIPE_ALL>();122+ PipeSync<AscendC::HardEvent::V_MTE2>();
113 AscendC::DataCopyExtParams copyParamsIn{123 AscendC::DataCopyExtParams copyParamsIn{
114 static_cast<uint16_t>(blockCount - 1), blockLen, srcStride, dstStride, 0};124 static_cast<uint16_t>(blockCount - 1), blockLen, srcStride, dstStride, 0};
115 if (blockCount > 1) {125 if (blockCount > 1) {
@@ -120,6 +130,7 @@ private:
120 padParams.paddingValue = 0;130 padParams.paddingValue = 0;
121 AscendC::DataCopyPad(131 AscendC::DataCopyPad(
122 srcLocal, this->srcGlobal[curSrcStart + index * this->tasksOnceMaxPerCore], copyParamsIn, padParams);132 srcLocal, this->srcGlobal[curSrcStart + index * this->tasksOnceMaxPerCore], copyParamsIn, padParams);
133+ AscendC::PipeBarrier<PIPE_MTE2>();
123 }134 }
124 copyParamsIn.blockCount = 1;135 copyParamsIn.blockCount = 1;
125 copyParamsIn.blockLen = (curHandleNum - (blockCount - 1) * colHandleNum) * this->typeSizeT1;136 copyParamsIn.blockLen = (curHandleNum - (blockCount - 1) * colHandleNum) * this->typeSizeT1;
@@ -137,6 +148,7 @@ private:
137 // fp16转fp32148 // fp16转fp32
138 srcLocal = this->inQueueSrc.template DeQue<T1>();149 srcLocal = this->inQueueSrc.template DeQue<T1>();
139 AscendC::LocalTensor<T2> computeSrcLocal = this->computeInQueueSrc.template AllocTensor<T2>();150 AscendC::LocalTensor<T2> computeSrcLocal = this->computeInQueueSrc.template AllocTensor<T2>();
151+ AscendC::PipeBarrier<PIPE_V>();
140 AscendC::Cast(152 AscendC::Cast(
141 computeSrcLocal, srcLocal, AscendC::RoundMode::CAST_NONE, colNumSpace * blockCount / this->typeSizeT1);153 computeSrcLocal, srcLocal, AscendC::RoundMode::CAST_NONE, colNumSpace * blockCount / this->typeSizeT1);
142 this->computeInQueueSrc.template EnQue(computeSrcLocal);154 this->computeInQueueSrc.template EnQue(computeSrcLocal);
@@ -174,6 +186,7 @@ private:
174 .GetPhyAddr());186 .GetPhyAddr());
175 }187 }
176 AscendC::TransDataTo5HD<T2>(dstLocalList, srcLocalList, transDataTo5HDParams.transDataParams);188 AscendC::TransDataTo5HD<T2>(dstLocalList, srcLocalList, transDataTo5HDParams.transDataParams);
189+ AscendC::PipeBarrier<PIPE_V>();
177 }190 }
178 191 
179 __aicore__ inline void TransDataForUnfold1(192 __aicore__ inline void TransDataForUnfold1(
@@ -239,6 +252,7 @@ private:
239 .GetPhyAddr());252 .GetPhyAddr());
240 }253 }
241 AscendC::TransDataTo5HD<T2>(dstLocalList, srcLocalList, transDataTo5HDParams.transDataParams);254 AscendC::TransDataTo5HD<T2>(dstLocalList, srcLocalList, transDataTo5HDParams.transDataParams);
255+ AscendC::PipeBarrier<PIPE_V>();
242 }256 }
243 257 
244 __aicore__ inline void TransDataForUnfold2(258 __aicore__ inline void TransDataForUnfold2(
@@ -302,6 +316,7 @@ private:
302 srcOffset = srcStart * TRANS_BLOCK;316 srcOffset = srcStart * TRANS_BLOCK;
303 dstOffset = dstStart * TRANS_BLOCK;317 dstOffset = dstStart * TRANS_BLOCK;
304 AscendC::Add<T2>(srcLocal[dstOffset], srcLocal[dstOffset], dstLocal[srcOffset], TRANS_BLOCK * this->size);318 AscendC::Add<T2>(srcLocal[dstOffset], srcLocal[dstOffset], dstLocal[srcOffset], TRANS_BLOCK * this->size);
319+ AscendC::PipeBarrier<PIPE_V>();
305 srcStart += this->size;320 srcStart += this->size;
306 dstStart += this->step;321 dstStart += this->step;
307 }322 }
@@ -319,6 +334,7 @@ private:
319 AscendC::PipeBarrier<PIPE_V>();334 AscendC::PipeBarrier<PIPE_V>();
320 T2 zeroVal(0.0);335 T2 zeroVal(0.0);
321 AscendC::Duplicate<T2>(computeSrcLocal, zeroVal, this->T2SrcDataSize / this->typeSizeT2);336 AscendC::Duplicate<T2>(computeSrcLocal, zeroVal, this->T2SrcDataSize / this->typeSizeT2);
337+ AscendC::PipeBarrier<PIPE_V>();
322 338 
323 // 累加计算339 // 累加计算
324 AccumulateFinalAxes(computeSrcLocal, computeDstLocal, curHandleNum);340 AccumulateFinalAxes(computeSrcLocal, computeDstLocal, curHandleNum);
Mconversion/unfold_grad/op_kernel/unfold_grad_final_axe_big_size.h+14-10
@@ -42,6 +42,14 @@ public:
42 this->workspaceT2SumRes.SetGlobalBuffer(reinterpret_cast<__gm__ T2*>(workspace) + gradInBlockOffset);42 this->workspaceT2SumRes.SetGlobalBuffer(reinterpret_cast<__gm__ T2*>(workspace) + gradInBlockOffset);
43 }43 }
44 44 
45+ template <AscendC::HardEvent hardEvent>
46+ __aicore__ inline void PipeSync()
47+ {
48+ int32_t eventID = static_cast<int32_t>(GetTPipePtr()->FetchEventID(hardEvent));
49+ AscendC::SetFlag<hardEvent>(eventID);
50+ AscendC::WaitFlag<hardEvent>(eventID);
51+ }
52+ 
45 __aicore__ inline void ProcessFinalAxeBigSize(int curSrcStart, int curDstStart)53 __aicore__ inline void ProcessFinalAxeBigSize(int curSrcStart, int curDstStart)
46 {54 {
47 this->tasksOnce = this->tasksOnceMaxPerCore;55 this->tasksOnce = this->tasksOnceMaxPerCore;
@@ -65,7 +73,7 @@ public:
65 }73 }
66 }74 }
67 }75 }
68- 76+ 
69 __aicore__ inline void Process()77 __aicore__ inline void Process()
70 {78 {
71 for (int batchIdx = 0; batchIdx < curCoreBatchNum; batchIdx++) {79 for (int batchIdx = 0; batchIdx < curCoreBatchNum; batchIdx++) {
@@ -77,7 +85,6 @@ public:
77 this->SetGMtoZero(this->outputNumPerCore, dstStart);85 this->SetGMtoZero(this->outputNumPerCore, dstStart);
78 }86 }
79 87 
80- AscendC::PipeBarrier<PIPE_ALL>();
81 ProcessFinalAxeBigSize(srcStart, dstStart);88 ProcessFinalAxeBigSize(srcStart, dstStart);
82 89 
83 if constexpr (ISCAST) {90 if constexpr (ISCAST) {
@@ -85,24 +92,20 @@ public:
85 this->CalculateOutParms(params);92 this->CalculateOutParms(params);
86 this->CopyToOutBigShapeOnePage(batchIdx, batchIdx, params);93 this->CopyToOutBigShapeOnePage(batchIdx, batchIdx, params);
87 }94 }
88- AscendC::PipeBarrier<PIPE_ALL>();95+ PipeSync<AscendC::HardEvent::MTE3_V>();
96+ PipeSync<AscendC::HardEvent::S_V>();
89 }97 }
90 }98 }
91 99 
92private:100private:
93- template <AscendC::HardEvent hardEvent>
94- __aicore__ inline void PipeSync()
95- {
96- int32_t eventID = static_cast<int32_t>(GetTPipePtr()->FetchEventID(hardEvent));
97- AscendC::SetFlag<hardEvent>(eventID);
98- AscendC::WaitFlag<hardEvent>(eventID);
99- }
100 __aicore__ inline void CopyInFinalAxeBigSize(int64_t curSrcStart, int64_t index, int64_t curHandleNum)101 __aicore__ inline void CopyInFinalAxeBigSize(int64_t curSrcStart, int64_t index, int64_t curHandleNum)
101 {102 {
102 AscendC::LocalTensor<T1> srcLocal =103 AscendC::LocalTensor<T1> srcLocal =
103 ISCAST ? this->inQueueSrc.template AllocTensor<T1>() : this->computeOutQueueDst.template AllocTensor<T1>();104 ISCAST ? this->inQueueSrc.template AllocTensor<T1>() : this->computeOutQueueDst.template AllocTensor<T1>();
104 AscendC::DataCopyPadExtParams<T1> padParams{false, 0, 0, 0};105 AscendC::DataCopyPadExtParams<T1> padParams{false, 0, 0, 0};
105 AscendC::PipeBarrier<PIPE_V>();106 AscendC::PipeBarrier<PIPE_V>();
107+ PipeSync<AscendC::HardEvent::MTE3_V>();
108+ PipeSync<AscendC::HardEvent::S_V>();
106 T1 zeroVal(0.0);109 T1 zeroVal(0.0);
107 int srcDataSize = ISCAST ? this->ubSizeT1 : this->T2SrcDataSize;110 int srcDataSize = ISCAST ? this->ubSizeT1 : this->T2SrcDataSize;
108 AscendC::Duplicate<T1>(srcLocal, zeroVal, srcDataSize / this->typeSizeT1);111 AscendC::Duplicate<T1>(srcLocal, zeroVal, srcDataSize / this->typeSizeT1);
@@ -130,6 +133,7 @@ private:
130 // fp16转fp32133 // fp16转fp32
131 srcLocal = this->inQueueSrc.template DeQue<T1>();134 srcLocal = this->inQueueSrc.template DeQue<T1>();
132 AscendC::LocalTensor<T2> computeDstLocal = this->computeOutQueueDst.template AllocTensor<T2>();135 AscendC::LocalTensor<T2> computeDstLocal = this->computeOutQueueDst.template AllocTensor<T2>();
136+ AscendC::PipeBarrier<PIPE_V>();
133 AscendC::Cast(137 AscendC::Cast(
134 computeDstLocal, srcLocal, AscendC::RoundMode::CAST_NONE, srcDataSize / this->typeSizeT1);138 computeDstLocal, srcLocal, AscendC::RoundMode::CAST_NONE, srcDataSize / this->typeSizeT1);
135 this->computeOutQueueDst.template EnQue(computeDstLocal);139 this->computeOutQueueDst.template EnQue(computeDstLocal);
Mconversion/unfold_grad/op_kernel/unfold_grad_final_second_axe.h+13-2
@@ -43,6 +43,14 @@ public:
43 this->workspaceT2SumRes.SetGlobalBuffer(reinterpret_cast<__gm__ T2*>(workspace) + gradInBlockOffset);43 this->workspaceT2SumRes.SetGlobalBuffer(reinterpret_cast<__gm__ T2*>(workspace) + gradInBlockOffset);
44 }44 }
45 45 
46+ template <AscendC::HardEvent hardEvent>
47+ __aicore__ inline void PipeSync()
48+ {
49+ int32_t eventID = static_cast<int32_t>(GetTPipePtr()->FetchEventID(hardEvent));
50+ AscendC::SetFlag<hardEvent>(eventID);
51+ AscendC::WaitFlag<hardEvent>(eventID);
52+ }
53+ 
46 __aicore__ inline void ProcessFinalSecondAxes(int curSrcStart, int curDstStart)54 __aicore__ inline void ProcessFinalSecondAxes(int curSrcStart, int curDstStart)
47 {55 {
48 for (int k = 0; k < this->iterationNumPerCore; k++) {56 for (int k = 0; k < this->iterationNumPerCore; k++) {
@@ -86,7 +94,6 @@ public:
86 this->SetGMtoZero(this->outputNumPerCore, dstStart);94 this->SetGMtoZero(this->outputNumPerCore, dstStart);
87 }95 }
88 96 
89- AscendC::PipeBarrier<PIPE_ALL>();
90 ProcessFinalSecondAxes(srcStart, dstStart);97 ProcessFinalSecondAxes(srcStart, dstStart);
91 98 
92 if constexpr (ISCAST) {99 if constexpr (ISCAST) {
@@ -94,7 +101,8 @@ public:
94 this->CalculateOutParms(params);101 this->CalculateOutParms(params);
95 this->CopyToOutBigShapeOnePage(batchIdx, batchIdx, params);102 this->CopyToOutBigShapeOnePage(batchIdx, batchIdx, params);
96 }103 }
97- AscendC::PipeBarrier<PIPE_ALL>();104+ PipeSync<AscendC::HardEvent::MTE3_V>();
105+ PipeSync<AscendC::HardEvent::S_V>();
98 }106 }
99 }107 }
100 108 
@@ -107,6 +115,9 @@ private:
107 uint32_t blockLen = curHandleNum * this->size * this->typeSizeT1;115 uint32_t blockLen = curHandleNum * this->size * this->typeSizeT1;
108 AscendC::DataCopyExtParams copyParamsIn{116 AscendC::DataCopyExtParams copyParamsIn{
109 1, blockLen, 0, 0, 0}; // 处理tasksOnce个数需要从srcGM中取((tasksOnce-1) * size + 1)个数117 1, blockLen, 0, 0, 0}; // 处理tasksOnce个数需要从srcGM中取((tasksOnce-1) * size + 1)个数
118+ PipeSync<AscendC::HardEvent::MTE3_MTE2>();
119+ PipeSync<AscendC::HardEvent::S_MTE2>();
120+ PipeSync<AscendC::HardEvent::V_MTE2>();
110 AscendC::DataCopyPad(121 AscendC::DataCopyPad(
111 srcLocal, this->srcGlobal[curSrcStart + index * this->tasksOnceMaxPerCore * this->size], copyParamsIn,122 srcLocal, this->srcGlobal[curSrcStart + index * this->tasksOnceMaxPerCore * this->size], copyParamsIn,
112 padParams); // k * size * inputSizeLastDim + j + index * tasksOnceMaxPerCore * size123 padParams); // k * size * inputSizeLastDim + j + index * tasksOnceMaxPerCore * size