修复连续使用initGlobalMemory初始化不同值之间的同步问题-ops-nn-AtomGit

已合并

修复连续使用initGlobalMemory初始化不同值之间的同步问题 #2939

zhang-wenbo-beat创建于 3月19日从zhang-wenbo-beat/ops-nn： 9.0.0 合入到 cann/ops-nn： 9.0.0

修复连续使用initGlobalMemory初始化不同值之间的同步问题 #2939

已合并

zhang-wenbo-beat创建于 3月19日

从zhang-wenbo-beat/ops-nn 9.0.0合入到cann/ops-nn 9.0.0

共 4 个文件变更+40-24

Mindex/inplace_index_add/op_kernel/arch35/inplace_index_add_determinstic.h+3-0

	@@ -159,6 +159,9 @@ __aicore__ inline void InplaceIndexAddDeterminstic<VAR_T, IDX_T>::Init(
159	GetBlockIdx() * tilingData_.eachCoreIndexCount, tilingData_.eachCoreIndexCount);	159	GetBlockIdx() * tilingData_.eachCoreIndexCount, tilingData_.eachCoreIndexCount);
160		160
161	InitGlobalMemory(updateSumWsGm_, sumWsSize_, (float)(0));	161	InitGlobalMemory(updateSumWsGm_, sumWsSize_, (float)(0));
		162	+ auto vWaitMte3EventID = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE3_V));
		163	+ SetFlag<HardEvent::MTE3_V>(vWaitMte3EventID);
		164	+ WaitFlag<HardEvent::MTE3_V>(vWaitMte3EventID);
162	InitGlobalMemory(updateSumIdxWsGm_, tilingData_.eachCoreIndexCount, (IDX_T)(-1));	165	InitGlobalMemory(updateSumIdxWsGm_, tilingData_.eachCoreIndexCount, (IDX_T)(-1));
163	AscendC::SyncAll();	166	AscendC::SyncAll();
164	}	167	}

Mindex/scatter_add/op_kernel/arch35/scatter_add_deterministic.h+3-0

	@@ -168,6 +168,9 @@ template<typename T, typename U, uint32_t scatterOp>
168	__aicore__ inline void ScatterAddDeterministicImpl<T, U, scatterOp>::InitWspZero()	168	__aicore__ inline void ScatterAddDeterministicImpl<T, U, scatterOp>::InitWspZero()
169	{	169	{
170	InitGlobalMemory(workspaceCoreLogicCoreSumId_, tilingData_.perCoreHandleIndices, (U)(-1));	170	InitGlobalMemory(workspaceCoreLogicCoreSumId_, tilingData_.perCoreHandleIndices, (U)(-1));
		171	+ auto vWaitMte3EventID = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE3_V));
		172	+ SetFlag<HardEvent::MTE3_V>(vWaitMte3EventID);
		173	+ WaitFlag<HardEvent::MTE3_V>(vWaitMte3EventID);
171	InitGlobalMemory(workspaceLogicCoreSumValue_, tilingData_.perCoreHandleIndices * postVarAlignSizeFp32_, (float)(0));	174	InitGlobalMemory(workspaceLogicCoreSumValue_, tilingData_.perCoreHandleIndices * postVarAlignSizeFp32_, (float)(0));
172	InitGlobalMemory(workspaceMaxValueCount_, tilingData_.varShape[0], (uint32_t)(0));	175	InitGlobalMemory(workspaceMaxValueCount_, tilingData_.varShape[0], (uint32_t)(0));
173	InitGlobalMemory(workspaceMaxValue_, tilingData_.varShape[1] * tilingData_.varShape[0], (float)(0));	176	InitGlobalMemory(workspaceMaxValue_, tilingData_.varShape[1] * tilingData_.varShape[0], (float)(0));

Mindex/scatter_nd/op_kernel/arch35/scatter_nd_deterministic.h+31-24

	@@ -171,6 +171,34 @@ __aicore__ inline void ScatterNdDeterministicImpl<T, U>::Init(GM_ADDR indices, G
171	// 量化分支	171	// 量化分支
172	shiftOffset_ = UB_AGLIN_VALUE / sizeof(U);	172	shiftOffset_ = UB_AGLIN_VALUE / sizeof(U);
173	indicesUbLoop_ = GetBlockIdx() == tilingData_.logicCoreNum - 1 ? tilingData_.tailCoreIndicesLoopSize : tilingData_.indicesLoopSize;	173	indicesUbLoop_ = GetBlockIdx() == tilingData_.logicCoreNum - 1 ? tilingData_.tailCoreIndicesLoopSize : tilingData_.indicesLoopSize;
		174	+
		175	+ indicesGm_.SetGlobalBuffer((__gm__ U )(indices) + GetBlockIdx() tilingData_.perCoreHandleIndices * tilingData_.rankSize);
		176	+ updatesGm_.SetGlobalBuffer((__gm__ T )(updates) + GetBlockIdx() tilingData_.perCoreHandleIndices * tilingData_.afterAxis);
		177	+
		178	+ workspaceMaxValueCount_.SetGlobalBuffer((__gm__ uint32_t *)(workspace));
		179	+
		180	+ workspaceMaxValue_.SetGlobalBuffer((__gm__ float *)workspace + tilingData_.rankFusedAxis);
		181	+ workspaceMaxValueInit_.SetGlobalBuffer((__gm__ float )workspace + tilingData_.rankFusedAxis + GetBlockIdx() initPerCore);
		182	+
		183	+ workspaceInt32Res_.SetGlobalBuffer((__gm__ int32_t *)workspace + tilingData_.rankFusedAxis +
		184	+ tilingData_.rankFusedAxis * tilingData_.afterAxis);
		185	+ workspaceInt32ResInit_.SetGlobalBuffer((__gm__ int32_t *)workspace + tilingData_.rankFusedAxis +
		186	+ tilingData_.rankFusedAxis * tilingData_.afterAxis + GetBlockIdx() * initPerCore);
		187	+
		188	+ workspaceLogicCoreSumValue_.SetGlobalBuffer((__gm__ float*)workspace + tilingData_.rankFusedAxis +
		189	+ tilingData_.rankFusedAxis * tilingData_.afterAxis * DOUBLE + GetBlockIdx() * tilingData_.perCoreHandleIndices * postVarAlignSizeFp32_);
		190	+
		191	+ __gm__ float* workspaceCoreLogicCoreSumIdStart = (__gm__ float*)workspace + tilingData_.rankFusedAxis +
		192	+ tilingData_.rankFusedAxis * tilingData_.afterAxis * DOUBLE +
		193	+ tilingData_.logicCoreNum * tilingData_.perCoreHandleIndices * postVarAlignSizeFp32_;
		194	+ workspaceCoreLogicCoreSumId_.SetGlobalBuffer((__gm__ U)workspaceCoreLogicCoreSumIdStart + GetBlockIdx() tilingData_.perCoreHandleIndices);
		195	+
		196	+ InitWspZero();
		197	+ if (tilingData_.isIdxSplit == 1) {
		198	+ InitGlobalMemory(yGmInit_, initCoreReal, (T)0);
		199	+ }
		200	+ SyncAll();
		201	+
174	pipe_.InitBuffer(indicesQue_, DOUBLE_BUF, ops::CeilAlign(tilingData_.indicesUbFactor * tilingData_.rankSize * sizeof(U), UB_AGLIN_VALUE));	202	pipe_.InitBuffer(indicesQue_, DOUBLE_BUF, ops::CeilAlign(tilingData_.indicesUbFactor * tilingData_.rankSize * sizeof(U), UB_AGLIN_VALUE));
175	pipe_.InitBuffer(outOfstBuf_, tilingData_.indicesUbFactor * sizeof(U));	203	pipe_.InitBuffer(outOfstBuf_, tilingData_.indicesUbFactor * sizeof(U));
176	pipe_.InitBuffer(calcBuf_, MAX_RANK_COUNT * sizeof(U));	204	pipe_.InitBuffer(calcBuf_, MAX_RANK_COUNT * sizeof(U));
	@@ -179,30 +207,15 @@ __aicore__ inline void ScatterNdDeterministicImpl<T, U>::Init(GM_ADDR indices, G
179	pipe_.InitBuffer(updateSumIdxQueue_, DOUBLE_BUF, ops::CeilAlign(tilingData_.indicesUbFactor * sizeof(U), UB_AGLIN_VALUE));	207	pipe_.InitBuffer(updateSumIdxQueue_, DOUBLE_BUF, ops::CeilAlign(tilingData_.indicesUbFactor * sizeof(U), UB_AGLIN_VALUE));
180	pipe_.InitBuffer(updatesQueue_, DOUBLE_BUF, tilingData_.indicesUbFactor * postVarAlignSize_ * sizeof(T));	208	pipe_.InitBuffer(updatesQueue_, DOUBLE_BUF, tilingData_.indicesUbFactor * postVarAlignSize_ * sizeof(T));
181	pipe_.InitBuffer(updateSumQue_, DOUBLE_BUF, tilingData_.indicesUbFactor * postVarAlignSizeFp32_ * sizeof(float));	209	pipe_.InitBuffer(updateSumQue_, DOUBLE_BUF, tilingData_.indicesUbFactor * postVarAlignSizeFp32_ * sizeof(float));
182	-
183	- indicesGm_.SetGlobalBuffer((__gm__ U )(indices) + GetBlockIdx() tilingData_.perCoreHandleIndices * tilingData_.rankSize);
184	- updatesGm_.SetGlobalBuffer((__gm__ T )(updates) + GetBlockIdx() tilingData_.perCoreHandleIndices * tilingData_.afterAxis);
185	-
186	- workspaceMaxValueCount_.SetGlobalBuffer((__gm__ uint32_t *)(workspace));
187	- workspaceMaxValue_.SetGlobalBuffer((__gm__ float *)workspace + tilingData_.rankFusedAxis);
188	- workspaceMaxValueInit_.SetGlobalBuffer((__gm__ float )workspace + tilingData_.rankFusedAxis + GetBlockIdx() initPerCore);
189	- workspaceInt32Res_.SetGlobalBuffer((__gm__ int32_t *)workspace + tilingData_.rankFusedAxis +
190	- tilingData_.rankFusedAxis * tilingData_.afterAxis);
191	- workspaceInt32ResInit_.SetGlobalBuffer((__gm__ int32_t *)workspace + tilingData_.rankFusedAxis +
192	- tilingData_.rankFusedAxis * tilingData_.afterAxis + GetBlockIdx() * initPerCore);
193	- workspaceLogicCoreSumValue_.SetGlobalBuffer((__gm__ float*)workspace + tilingData_.rankFusedAxis +
194	- tilingData_.rankFusedAxis * tilingData_.afterAxis * DOUBLE + GetBlockIdx() * tilingData_.perCoreHandleIndices * postVarAlignSizeFp32_);
195	-
196	- __gm__ float* workspaceCoreLogicCoreSumIdStart = (__gm__ float*)workspace + tilingData_.rankFusedAxis +
197	- tilingData_.rankFusedAxis * tilingData_.afterAxis * DOUBLE +
198	- tilingData_.logicCoreNum * tilingData_.perCoreHandleIndices * postVarAlignSizeFp32_;
199	- workspaceCoreLogicCoreSumId_.SetGlobalBuffer((__gm__ U)workspaceCoreLogicCoreSumIdStart + GetBlockIdx() tilingData_.perCoreHandleIndices);
200	}	210	}
201		211
202	template<typename T, typename U>	212	template<typename T, typename U>
203	__aicore__ inline void ScatterNdDeterministicImpl<T, U>::InitWspZero()	213	__aicore__ inline void ScatterNdDeterministicImpl<T, U>::InitWspZero()
204	{	214	{
205	InitGlobalMemory(workspaceCoreLogicCoreSumId_, tilingData_.perCoreHandleIndices, (U)(-1));	215	InitGlobalMemory(workspaceCoreLogicCoreSumId_, tilingData_.perCoreHandleIndices, (U)(-1));
		216	+ auto vWaitMte3EventID = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE3_V));
		217	+ SetFlag<HardEvent::MTE3_V>(vWaitMte3EventID);
		218	+ WaitFlag<HardEvent::MTE3_V>(vWaitMte3EventID);
206	InitGlobalMemory(workspaceLogicCoreSumValue_, tilingData_.perCoreHandleIndices * postVarAlignSizeFp32_, (float)(0));	219	InitGlobalMemory(workspaceLogicCoreSumValue_, tilingData_.perCoreHandleIndices * postVarAlignSizeFp32_, (float)(0));
207	InitGlobalMemory(workspaceMaxValueCount_, tilingData_.rankFusedAxis, (uint32_t)(0));	220	InitGlobalMemory(workspaceMaxValueCount_, tilingData_.rankFusedAxis, (uint32_t)(0));
208	InitGlobalMemory(workspaceMaxValueInit_, initCoreReal, (float)(0));	221	InitGlobalMemory(workspaceMaxValueInit_, initCoreReal, (float)(0));
	@@ -913,12 +926,6 @@ __aicore__ inline void ScatterNdDeterministicImpl<T, U>::Process()
913	return;	926	return;
914	}	927	}
915		928
916	- InitWspZero();
917	- if (tilingData_.isIdxSplit == 1) {
918	- InitGlobalMemory(yGmInit_, initCoreReal, (T)0);
919	- }
920	- SyncAll();
921	-
922	LocalTensor<U> calcLocal = calcBuf_.Get<U>();	929	LocalTensor<U> calcLocal = calcBuf_.Get<U>();
923	for (int32_t i = 0; i < MAX_RANK_COUNT; i++) {	930	for (int32_t i = 0; i < MAX_RANK_COUNT; i++) {
924	calcLocal(i) = tilingData_.strideList[i];	931	calcLocal(i) = tilingData_.strideList[i];

Mindex/scatter_nd_add/op_kernel/arch35/scatter_nd_add_determinstic.h+3-0

	@@ -160,6 +160,9 @@ __aicore__ inline void ScatterNdAddDeterministic<T, U, CAST_T, castType>::Init(G
160	GetBlockIdx() * tilingData_.eachCoreIndexCount, tilingData_.eachCoreIndexCount);	160	GetBlockIdx() * tilingData_.eachCoreIndexCount, tilingData_.eachCoreIndexCount);
161		161
162	InitGlobalMemory(updateSumWsGm_, sumWsSize_, (float)(0));	162	InitGlobalMemory(updateSumWsGm_, sumWsSize_, (float)(0));
		163	+ auto vWaitMte3EventID = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE3_V));
		164	+ SetFlag<HardEvent::MTE3_V>(vWaitMte3EventID);
		165	+ WaitFlag<HardEvent::MTE3_V>(vWaitMte3EventID);
163	InitGlobalMemory(updateSumIdxWsGm_, tilingData_.eachCoreIndexCount, (U)(-1));	166	InitGlobalMemory(updateSumIdxWsGm_, tilingData_.eachCoreIndexCount, (U)(-1));
164	AscendC::SyncAll();	167	AscendC::SyncAll();
165	}	168	}