wqbmmv3 do not support nz in dav3510-ops-nn-AtomGit

已合并

wqbmmv3 do not support nz in dav3510 #4100

zhangquanxin创建于 4月22日从zhangquanxin/ops-nn： wqbmmv3_not_support_nz 合入到 cann/ops-nn： master

wqbmmv3 do not support nz in dav3510 #4100

已合并

zhangquanxin创建于 4月22日

从zhangquanxin/ops-nn wqbmmv3_not_support_nz合入到cann/ops-nn master

共 2 个文件变更+17-12

Rmatmul/quant_batch_matmul_v3/examples/arch35/test_aclnn_quant_matmul_weight_nz a8w4.cpp→matmul/quant_batch_matmul_v3/examples/arch35/test_aclnn_quant_matmul_weight_nz_a8w4.cpp+2-2

	@@ -126,8 +126,8 @@ int AclnnQuantMatmulWeightNzTest(int32_t deviceId, aclrtStream& stream)
126	int64_t groupSize = 32;	126	int64_t groupSize = 32;
127	std::vector<int64_t> x1Shape = {m, k};	127	std::vector<int64_t> x1Shape = {m, k};
128	std::vector<int64_t> x2Shape = {n, k};	128	std::vector<int64_t> x2Shape = {n, k};
129	- std::vector<int64_t> x1ScaleShape = {m, k / groupSize};	129	+ std::vector<int64_t> x1ScaleShape = {m, k / groupSize / 2, 2};
130	- std::vector<int64_t> x2ScaleShape = {n, k / groupSize};	130	+ std::vector<int64_t> x2ScaleShape = {n, k / groupSize / 2, 2};
131	std::vector<int64_t> outShape = {m, n};	131	std::vector<int64_t> outShape = {m, n};
132	void* x1DeviceAddr = nullptr;	132	void* x1DeviceAddr = nullptr;
133	void* x2DeviceAddr = nullptr;	133	void* x2DeviceAddr = nullptr;

Mmatmul/weight_quant_batch_matmul_v2/op_host/op_api/aclnn_weight_quant_batch_matmul_v2.cpp+15-10

	@@ -832,6 +832,19 @@ static bool CheckNotNull(
832	return true;	832	return true;
833	}	833	}
834		834
		835	+static bool CheckWeightFormat(const aclTensor* weight)
		836	+{
		837	+ if (op::GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510) {
		838	+ if (weight->GetStorageFormat() != op::Format::FORMAT_ND) {
		839	+ OP_LOGE(ACLNN_ERR_PARAM_INVALID,
		840	+ "In DAV_3510, aclnnWeightQuantBatchMatmulV2 and aclnnWeightQuantBatchMatmulV3 weight format only "
		841	+ "support FORMAT_ND.");
		842	+ return false;
		843	+ }
		844	+ }
		845	+ return true;
		846	+}
		847	+
835	static bool CheckOptionalNotNull(const aclTensor* quantScaleOptional, const aclTensor* quantOffsetOptional)	848	static bool CheckOptionalNotNull(const aclTensor* quantScaleOptional, const aclTensor* quantOffsetOptional)
836	{	849	{
837	if (GetCurrentPlatformInfo().GetSocVersion() == SocVersion::ASCEND310P) {	850	if (GetCurrentPlatformInfo().GetSocVersion() == SocVersion::ASCEND310P) {
	@@ -1738,15 +1751,7 @@ aclnnStatus aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
1738	CHECK_RET(socRes == ACLNN_SUCCESS, socRes);	1751	CHECK_RET(socRes == ACLNN_SUCCESS, socRes);
1739	CHECK_RET(CheckNotNull(x, weight, antiquantScale, y), ACLNN_ERR_PARAM_NULLPTR);	1752	CHECK_RET(CheckNotNull(x, weight, antiquantScale, y), ACLNN_ERR_PARAM_NULLPTR);
1740	CHECK_RET(CheckOptionalNotNull(quantScaleOptional, quantOffsetOptional), ACLNN_ERR_PARAM_NULLPTR);	1753	CHECK_RET(CheckOptionalNotNull(quantScaleOptional, quantOffsetOptional), ACLNN_ERR_PARAM_NULLPTR);
1741	- if (op::GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510) {	1754	+ CHECK_RET(CheckWeightFormat(weight), ACLNN_ERR_PARAM_INVALID);
1742	- OP_CHECK(
1743	- weight->GetStorageFormat() == op::Format::FORMAT_ND,
1744	- OP_LOGE(
1745	- ACLNN_ERR_PARAM_INVALID,
1746	- "In DAV_3510, aclnnWeightQuantBatchMatmulV2 not support FORMAT_FRACTAL_NZ and only support "
1747	- "FORMAT_ND."),
1748	- return ACLNN_ERR_PARAM_INVALID);
1749	- }
1750	const aclTensor* tensorWeight = weight;	1755	const aclTensor* tensorWeight = weight;
1751	const aclTensor* antiquantScaleRef = antiquantScale;	1756	const aclTensor* antiquantScaleRef = antiquantScale;
1752	const aclTensor* tensorQuantScaleOptional = quantScaleOptional;	1757	const aclTensor* tensorQuantScaleOptional = quantScaleOptional;
	@@ -1798,7 +1803,7 @@ aclnnStatus aclnnWeightQuantBatchMatmulV3GetWorkspaceSize(
1798	CHECK_RET(socRes == ACLNN_SUCCESS, socRes);	1803	CHECK_RET(socRes == ACLNN_SUCCESS, socRes);
1799	CHECK_RET(CheckNotNull(x, weight, antiquantScale, y), ACLNN_ERR_PARAM_NULLPTR);	1804	CHECK_RET(CheckNotNull(x, weight, antiquantScale, y), ACLNN_ERR_PARAM_NULLPTR);
1800	CHECK_RET(CheckOptionalNotNull(quantScaleOptional, quantOffsetOptional), ACLNN_ERR_PARAM_NULLPTR);	1805	CHECK_RET(CheckOptionalNotNull(quantScaleOptional, quantOffsetOptional), ACLNN_ERR_PARAM_NULLPTR);
1801	-	1806	+ CHECK_RET(CheckWeightFormat(weight), ACLNN_ERR_PARAM_INVALID);
1802	const aclTensor* tensorWeight = weight;	1807	const aclTensor* tensorWeight = weight;
1803	const aclTensor* antiquantScaleRef = antiquantScale;	1808	const aclTensor* antiquantScaleRef = antiquantScale;
1804	const aclTensor* tensorQuantScaleOptional = quantScaleOptional;	1809	const aclTensor* tensorQuantScaleOptional = quantScaleOptional;