添加bmmv3 vector kernel拦截条件-ops-nn-AtomGit

Mmatmul/batch_mat_mul_v3/op_host/op_tiling/batch_mat_mul_v3_base_tiling.cpp+17-4

	@@ -621,8 +621,8 @@ bool BatchMatmulV3BaseTiling::DoMultiBatchOutTiling()
621	void BatchMatmulV3BaseTiling::DoMultiBatchTiling()	621	void BatchMatmulV3BaseTiling::DoMultiBatchTiling()
622	{	622	{
623	bool isEqualBatch = batchInfo_.batchA0 == batchInfo_.batchB0 && batchInfo_.batchA1 == batchInfo_.batchB1 &&	623	bool isEqualBatch = batchInfo_.batchA0 == batchInfo_.batchB0 && batchInfo_.batchA1 == batchInfo_.batchB1 &&
624	- batchInfo_.batchA2 == batchInfo_.batchB2 && batchInfo_.batchA3 == batchInfo_.batchB3; //广播	624	+ batchInfo_.batchA2 == batchInfo_.batchB2 && batchInfo_.batchA3 == batchInfo_.batchB3; // 广播
625	- if (!isEqualBatch \|\| (args_.hasBias && !batchInfo_.biasWithBatch)) { //不支持broadcast\非多batch bias	625	+ if (!isEqualBatch \|\| (args_.hasBias && !batchInfo_.biasWithBatch)) { // 不支持broadcast\非多batch bias
626	return;	626	return;
627	}	627	}
628	uint64_t shapeM = ops::CeilAlign(static_cast<uint64_t>(bmmTilingData_.matmulTiling.matmulTiling.M), BLOCK_CUBE);	628	uint64_t shapeM = ops::CeilAlign(static_cast<uint64_t>(bmmTilingData_.matmulTiling.matmulTiling.M), BLOCK_CUBE);
	@@ -735,7 +735,7 @@ void BatchMatmulV3BaseTiling::DoMultiBatchL1FullLoadTilingImpl()
735	void BatchMatmulV3BaseTiling::DoMultiBatchL1FullLoadTiling()	735	void BatchMatmulV3BaseTiling::DoMultiBatchL1FullLoadTiling()
736	{	736	{
737	bool isEqualBatch = batchInfo_.batchA0 == batchInfo_.batchB0 && batchInfo_.batchA1 == batchInfo_.batchB1 &&	737	bool isEqualBatch = batchInfo_.batchA0 == batchInfo_.batchB0 && batchInfo_.batchA1 == batchInfo_.batchB1 &&
738	- batchInfo_.batchA2 == batchInfo_.batchB2 && batchInfo_.batchA3 == batchInfo_.batchB3; //广播	738	+ batchInfo_.batchA2 == batchInfo_.batchB2 && batchInfo_.batchA3 == batchInfo_.batchB3; // 广播
739	if (!isEqualBatch \|\| args_.hasBias) { // 暂时不支持bias	739	if (!isEqualBatch \|\| args_.hasBias) { // 暂时不支持bias
740	return;	740	return;
741	}	741	}
	@@ -1159,6 +1159,16 @@ bool BatchMatmulV3BaseTiling::CheckVectorNpuArch()
1159	return true;	1159	return true;
1160	}	1160	}
1161		1161
		1162	+bool BatchMatmulV3BaseTiling::CheckVectorTranspose()
		1163	+{
		1164	+ if (args_.isATrans \|\| !args_.isBTrans) {
		1165	+ OP_LOGD(args_.opName, "BatchMatmulV3BaseTiling: transA should be false, transB should be true. "
		1166	+ "Bmm vector opt version not supported.");
		1167	+ return false;
		1168	+ }
		1169	+ return true;
		1170	+}
		1171	+
1162	bool BatchMatmulV3BaseTiling::CheckVectorShapeDims()	1172	bool BatchMatmulV3BaseTiling::CheckVectorShapeDims()
1163	{	1173	{
1164	auto aShape = context_->GetInputShape(0)->GetOriginShape();	1174	auto aShape = context_->GetInputShape(0)->GetOriginShape();
	@@ -1251,6 +1261,9 @@ bool BatchMatmulV3BaseTiling::CheckVectorComputationCondition()
1251	if (!CheckVectorNpuArch()) {	1261	if (!CheckVectorNpuArch()) {
1252	return false;	1262	return false;
1253	}	1263	}
		1264	+ if (!CheckVectorTranspose()) {
		1265	+ return false;
		1266	+ }
1254	if (!CheckVectorShapeDims()) {	1267	if (!CheckVectorShapeDims()) {
1255	return false;	1268	return false;
1256	}	1269	}
	@@ -1328,4 +1341,4 @@ void BatchMatmulV3BaseTiling::DoTilingKeyCustom()
1328	}	1341	}
1329		1342
1330	} // namespace optiling	1343	} // namespace optiling
1331	-}	1344	+}

Mmatmul/batch_mat_mul_v3/op_host/op_tiling/batch_mat_mul_v3_base_tiling.h+7-6

	@@ -40,7 +40,7 @@ struct BatchShapeInfo {
40	bool biasWithBatch = false;	40	bool biasWithBatch = false;
41	};	41	};
42		42
43	-enum class TilingCalcSelect //选择不同的计算Tiling的方法	43	+enum class TilingCalcSelect // 选择不同的计算Tiling的方法
44	{	44	{
45	ALL = 0,	45	ALL = 0,
46	COMMON = 1,	46	COMMON = 1,
	@@ -51,14 +51,14 @@ enum class TilingEnableMultiBatchL1FullLoad : int32_t // 互斥flag, 对应不
51	{	51	{
52	IS_FALSE = 0,	52	IS_FALSE = 0,
53	IS_TRUE = 1,	53	IS_TRUE = 1,
54	- MAX = 10 //模板类别不能超过10个	54	+ MAX = 10 // 模板类别不能超过10个
55	};	55	};
56		56
57	enum class TilingEnableMultiBatch : int32_t // 互斥flag, 对应不同全载模板选择	57	enum class TilingEnableMultiBatch : int32_t // 互斥flag, 对应不同全载模板选择
58	{	58	{
59	IS_FALSE = 0,	59	IS_FALSE = 0,
60	IS_TRUE = 1,	60	IS_TRUE = 1,
61	- MAX = 10 //模板类别不能超过10个	61	+ MAX = 10 // 模板类别不能超过10个
62	};	62	};
63		63
64	enum class TilingEnableLoadMode : int32_t // 互斥flag, 对应不同全载模板选择	64	enum class TilingEnableLoadMode : int32_t // 互斥flag, 对应不同全载模板选择
	@@ -67,21 +67,21 @@ enum class TilingEnableLoadMode : int32_t // 互斥flag, 对应不同全载模
67	AL1_FULL_LOAD = 1,	67	AL1_FULL_LOAD = 1,
68	BL1_FULL_LOAD = 2,	68	BL1_FULL_LOAD = 2,
69	VECTOR_FULL_LOAD = 3,	69	VECTOR_FULL_LOAD = 3,
70	- MAX = 10 //模板类别不能超过10个	70	+ MAX = 10 // 模板类别不能超过10个
71	};	71	};
72		72
73	enum class TilingEnableMultiBatchOut : int32_t // 互斥flag, 对应不同全载模板选择	73	enum class TilingEnableMultiBatchOut : int32_t // 互斥flag, 对应不同全载模板选择
74	{	74	{
75	IS_FALSE = 0,	75	IS_FALSE = 0,
76	IS_TRUE = 1,	76	IS_TRUE = 1,
77	- MAX = 10 //模板类别不能超过10个	77	+ MAX = 10 // 模板类别不能超过10个
78	};	78	};
79		79
80	enum class TilingEnableMixNd2Nz : int32_t // 互斥flag, 对应不同全载模板选择	80	enum class TilingEnableMixNd2Nz : int32_t // 互斥flag, 对应不同全载模板选择
81	{	81	{
82	IS_TRUE = 0,	82	IS_TRUE = 0,
83	IS_FALSE = 1,	83	IS_FALSE = 1,
84	- MAX = 10 //模板类别不能超过10个	84	+ MAX = 10 // 模板类别不能超过10个
85	};	85	};
86		86
87	struct TilingEnable {	87	struct TilingEnable {
	@@ -157,6 +157,7 @@ protected:
157	private:	157	private:
158	// CheckVectorComputationCondition的子检查函数	158	// CheckVectorComputationCondition的子检查函数
159	bool CheckVectorNpuArch();	159	bool CheckVectorNpuArch();
		160	+ bool CheckVectorTranspose();
160	bool CheckVectorShapeDims();	161	bool CheckVectorShapeDims();
161	bool CheckVectorDtypeAndKAxis();	162	bool CheckVectorDtypeAndKAxis();
162	bool CheckVectorBatchBroadcast();	163	bool CheckVectorBatchBroadcast();

Mmatmul/common/op_host/op_api/batch_matmul_util.cpp+2-1

	@@ -249,7 +249,8 @@ static bool CheckAscendCScenario(const aclTensor* x1, const aclTensor* x2, const
249	int64_t mDim = x1->GetViewShape().GetDim(x1DimNum - 2);	249	int64_t mDim = x1->GetViewShape().GetDim(x1DimNum - 2);
250	int64_t kDim = x1->GetViewShape().GetDim(x1DimNum - 1);	250	int64_t kDim = x1->GetViewShape().GetDim(x1DimNum - 1);
251	int64_t MAX_MK_DIM = 8;	251	int64_t MAX_MK_DIM = 8;
252	- if (npuArch == NpuArch::DAV_2201 && adjX2 == 1 && nDim == 1 && mDim <= MAX_MK_DIM && kDim <= MAX_MK_DIM) {	252	+ if (npuArch == NpuArch::DAV_2201 && adjX1 == 0 && adjX2 == 1 && nDim == 1 && mDim <= MAX_MK_DIM &&
		253	+ kDim <= MAX_MK_DIM) {
253	OP_LOGI("Hit batch_mat_mul_v3 vector kernel scenario: effective N dimension is 1.");	254	OP_LOGI("Hit batch_mat_mul_v3 vector kernel scenario: effective N dimension is 1.");
254	return true;	255	return true;
255	}	256	}