Star904
1821
代码
代码
Issues224
Pull Requests298
Actions
项目讨论
Wiki
分析
Star904
1821
已合并
修复gemmV3在16in32out场景下精度问题 #5602
HKFLYE创建于 6月2日
修复gemmV3在16in32out场景下精度问题 #5602
已合并
HKFLYE创建于 6月2日
从HuangKun8682/ops-nn_hkflyepersonal_addmm_support_16in32out_gemmV3_branch合入到cann/ops-nnmaster
共 9 个文件变更+10-646
Mmatmul/batch_mat_mul_v3/op_host/config/ascend910_93/batch_mat_mul_v3_binary.json+0-160
@@ -722,166 +722,6 @@
722 }722 }
723 ]723 ]
724 },724 },
725- {
726- "bin_filename": "BatchMatMulV3_ND_ND_ND_ND_FP16_FP16_FP32_FP32",
727- "simplified_key": "diy,2/2/2/2/1/1/0/0",
728- "inputs": [
729- {
730- "name": "x1",
731- "index": 0,
732- "dtype": "float16",
733- "format": "ND",
734- "paramType": "required",
735- "shape": [
736- -2
737- ]
738- },
739- {
740- "name": "x2",
741- "index": 1,
742- "dtype": "float16",
743- "format": "ND",
744- "paramType": "required",
745- "shape": [
746- -2
747- ]
748- },
749- {
750- "name": "bias",
751- "index": 2,
752- "dtype": "float32",
753- "format": "ND",
754- "paramType": "optional",
755- "shape": [
756- -2
757- ]
758- },
759- {
760- "name": "offset_w",
761- "index": 3,
762- "dtype": "int8",
763- "format": "ND",
764- "paramType": "optional",
765- "shape": [
766- -2
767- ]
768- }
769- ],
770- "outputs": [
771- {
772- "name": "y",
773- "index": 0,
774- "dtype": "float32",
775- "format": "ND",
776- "paramType": "required",
777- "shape": [
778- -2
779- ]
780- }
781- ],
782- "attrs": [
783- {
784- "name": "adj_x1",
785- "dtype": "bool",
786- "value": false
787- },
788- {
789- "name": "adj_x2",
790- "dtype": "bool",
791- "value": false
792- },
793- {
794- "name": "offset_x",
795- "dtype": "int",
796- "value": 0
797- },
798- {
799- "name": "enable_hf32",
800- "dtype": "bool",
801- "value": false
802- }
803- ]
804- },
805- {
806- "bin_filename": "BatchMatMulV3_ND_ND_ND_ND_BF16_BF16_FP32_FP32",
807- "simplified_key": "diy,2/2/2/2/27/27/0/0",
808- "inputs": [
809- {
810- "name": "x1",
811- "index": 0,
812- "dtype": "bfloat16",
813- "format": "ND",
814- "paramType": "required",
815- "shape": [
816- -2
817- ]
818- },
819- {
820- "name": "x2",
821- "index": 1,
822- "dtype": "bfloat16",
823- "format": "ND",
824- "paramType": "required",
825- "shape": [
826- -2
827- ]
828- },
829- {
830- "name": "bias",
831- "index": 2,
832- "dtype": "float32",
833- "format": "ND",
834- "paramType": "optional",
835- "shape": [
836- -2
837- ]
838- },
839- {
840- "name": "offset_w",
841- "index": 3,
842- "dtype": "int8",
843- "format": "ND",
844- "paramType": "optional",
845- "shape": [
846- -2
847- ]
848- }
849- ],
850- "outputs": [
851- {
852- "name": "y",
853- "index": 0,
854- "dtype": "float32",
855- "format": "ND",
856- "paramType": "required",
857- "shape": [
858- -2
859- ]
860- }
861- ],
862- "attrs": [
863- {
864- "name": "adj_x1",
865- "dtype": "bool",
866- "value": false
867- },
868- {
869- "name": "adj_x2",
870- "dtype": "bool",
871- "value": false
872- },
873- {
874- "name": "offset_x",
875- "dtype": "int",
876- "value": 0
877- },
878- {
879- "name": "enable_hf32",
880- "dtype": "bool",
881- "value": false
882- }
883- ]
884- },
885 {725 {
886 "bin_filename": "BatchMatMulV3_ND_NZ_ND_ND_FP16_FP16_FP16_FP32",726 "bin_filename": "BatchMatMulV3_ND_NZ_ND_ND_FP16_FP16_FP16_FP32",
887 "simplified_key": "diy,2/29/2/2/1/1/1/0",727 "simplified_key": "diy,2/29/2/2/1/1/1/0",
Mmatmul/batch_mat_mul_v3/op_host/config/ascend910b/batch_mat_mul_v3_binary.json+0-160
@@ -722,166 +722,6 @@
722 }722 }
723 ]723 ]
724 },724 },
725- {
726- "bin_filename": "BatchMatMulV3_ND_ND_ND_ND_FP16_FP16_FP32_FP32",
727- "simplified_key": "diy,2/2/2/2/1/1/0/0",
728- "inputs": [
729- {
730- "name": "x1",
731- "index": 0,
732- "dtype": "float16",
733- "format": "ND",
734- "paramType": "required",
735- "shape": [
736- -2
737- ]
738- },
739- {
740- "name": "x2",
741- "index": 1,
742- "dtype": "float16",
743- "format": "ND",
744- "paramType": "required",
745- "shape": [
746- -2
747- ]
748- },
749- {
750- "name": "bias",
751- "index": 2,
752- "dtype": "float32",
753- "format": "ND",
754- "paramType": "optional",
755- "shape": [
756- -2
757- ]
758- },
759- {
760- "name": "offset_w",
761- "index": 3,
762- "dtype": "int8",
763- "format": "ND",
764- "paramType": "optional",
765- "shape": [
766- -2
767- ]
768- }
769- ],
770- "outputs": [
771- {
772- "name": "y",
773- "index": 0,
774- "dtype": "float32",
775- "format": "ND",
776- "paramType": "required",
777- "shape": [
778- -2
779- ]
780- }
781- ],
782- "attrs": [
783- {
784- "name": "adj_x1",
785- "dtype": "bool",
786- "value": false
787- },
788- {
789- "name": "adj_x2",
790- "dtype": "bool",
791- "value": false
792- },
793- {
794- "name": "offset_x",
795- "dtype": "int",
796- "value": 0
797- },
798- {
799- "name": "enable_hf32",
800- "dtype": "bool",
801- "value": false
802- }
803- ]
804- },
805- {
806- "bin_filename": "BatchMatMulV3_ND_ND_ND_ND_BF16_BF16_FP32_FP32",
807- "simplified_key": "diy,2/2/2/2/27/27/0/0",
808- "inputs": [
809- {
810- "name": "x1",
811- "index": 0,
812- "dtype": "bfloat16",
813- "format": "ND",
814- "paramType": "required",
815- "shape": [
816- -2
817- ]
818- },
819- {
820- "name": "x2",
821- "index": 1,
822- "dtype": "bfloat16",
823- "format": "ND",
824- "paramType": "required",
825- "shape": [
826- -2
827- ]
828- },
829- {
830- "name": "bias",
831- "index": 2,
832- "dtype": "float32",
833- "format": "ND",
834- "paramType": "optional",
835- "shape": [
836- -2
837- ]
838- },
839- {
840- "name": "offset_w",
841- "index": 3,
842- "dtype": "int8",
843- "format": "ND",
844- "paramType": "optional",
845- "shape": [
846- -2
847- ]
848- }
849- ],
850- "outputs": [
851- {
852- "name": "y",
853- "index": 0,
854- "dtype": "float32",
855- "format": "ND",
856- "paramType": "required",
857- "shape": [
858- -2
859- ]
860- }
861- ],
862- "attrs": [
863- {
864- "name": "adj_x1",
865- "dtype": "bool",
866- "value": false
867- },
868- {
869- "name": "adj_x2",
870- "dtype": "bool",
871- "value": false
872- },
873- {
874- "name": "offset_x",
875- "dtype": "int",
876- "value": 0
877- },
878- {
879- "name": "enable_hf32",
880- "dtype": "bool",
881- "value": false
882- }
883- ]
884- },
885 {725 {
886 "bin_filename": "BatchMatMulV3_ND_NZ_ND_ND_FP16_FP16_FP16_FP32",726 "bin_filename": "BatchMatMulV3_ND_NZ_ND_ND_FP16_FP16_FP16_FP32",
887 "simplified_key": "diy,2/29/2/2/1/1/1/0",727 "simplified_key": "diy,2/29/2/2/1/1/1/0",
Mmatmul/batch_mat_mul_v3/op_host/op_api/aclnn_baddbmm.cpp+2-1
@@ -275,7 +275,8 @@ public:
275 bool needBroadcast = CheckAddmmTensorShapeNeedBroadcast(matA, matB, bias);275 bool needBroadcast = CheckAddmmTensorShapeNeedBroadcast(matA, matB, bias);
276 // A2/A3上对于 16in32out,且不需要broadcast场景 直接走gemmV3276 // A2/A3上对于 16in32out,且不需要broadcast场景 直接走gemmV3
277 if (CheckGemmV3WithAlphaBeta(bias, matA, matB, cubeMathType) || (enable16In32Out && !needBroadcast)) {277 if (CheckGemmV3WithAlphaBeta(bias, matA, matB, cubeMathType) || (enable16In32Out && !needBroadcast)) {
278- const aclTensor* bmmOut = ExecGemmV3WithAlphaBetaOp(bias, matA, matB, alpha, beta, executor);278+ const aclTensor* bmmOut = ExecGemmV3WithAlphaBetaOp(bias, matA, matB,
279+ alpha, beta, executor, enable16In32Out);
279 CHECK_RET(bmmOut != nullptr, ACLNN_ERR_INNER_NULLPTR);280 CHECK_RET(bmmOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
280 convOut = bmmOut;281 convOut = bmmOut;
281 return ACLNN_SUCCESS;282 return ACLNN_SUCCESS;
Mmatmul/common/op_host/op_api/batch_matmul_util.cpp+1-1
@@ -405,7 +405,7 @@ const aclTensor* GetBatchMatmulOp(
405 enable16In32Out, executor);405 enable16In32Out, executor);
406 }406 }
407 OP_LOGI("Hit batch_mat_mul_v3 scenario.");407 OP_LOGI("Hit batch_mat_mul_v3 scenario.");
408- if (enable16In32Out) {408+ if (enable16In32Out || enable2201Baddbmm16In32Out) {
409 OP_LOGI("Hit batch_mat_mul_v3 fp16/bf16 in - fp32 out scenario.");409 OP_LOGI("Hit batch_mat_mul_v3 fp16/bf16 in - fp32 out scenario.");
410 bmmOpOut = l0op::BatchMatMulV3NdFp16Bf162Fp32(410 bmmOpOut = l0op::BatchMatMulV3NdFp16Bf162Fp32(
411 selfTransdata, mat2Transdata, bias, nullptr, adjX1, adjX2, offsetX, matmulOpInfo.enableHf32, executor);411 selfTransdata, mat2Transdata, bias, nullptr, adjX1, adjX2, offsetX, matmulOpInfo.enableHf32, executor);
Mmatmul/common/op_host/op_api/matmul_util.cpp+6-1
@@ -479,6 +479,11 @@ static const aclTensor* GetMatMulOp(
479 bool supportNdNz = mmOpInfo.support_info.self_format == ge::FORMAT_ND &&479 bool supportNdNz = mmOpInfo.support_info.self_format == ge::FORMAT_ND &&
480 mmOpInfo.support_info.mat2_format == ge::FORMAT_FRACTAL_NZ;480 mmOpInfo.support_info.mat2_format == ge::FORMAT_FRACTAL_NZ;
481 bool addmm16In32Out = enable16In32Out && (bias != nullptr || supportNdNz);481 bool addmm16In32Out = enable16In32Out && (bias != nullptr || supportNdNz);
482+ bool isFp32Out = mmOpInfo.support_info.output_dtype == DataType::DT_FLOAT;
483+ bool bothMatFp16Bf16 = (mmOpInfo.support_info.self_dtype == DataType::DT_FLOAT16 &&
484+ mmOpInfo.support_info.mat2_dtype == DataType::DT_FLOAT16) ||
485+ (mmOpInfo.support_info.self_dtype == DataType::DT_BF16 &&
486+ mmOpInfo.support_info.mat2_dtype == DataType::DT_BF16);
482 if (CheckMatmulV3Support(x1, x2, bias, mmOpInfo, transposeX1, transposeX2, opImplModeEnum) ||487 if (CheckMatmulV3Support(x1, x2, bias, mmOpInfo, transposeX1, transposeX2, opImplModeEnum) ||
483 (CheckMMV3NzNzNdSupport(mmOpInfo) && CheckSupportInfoFormatNzNzNd(mmOpInfo)) || addmm16In32Out) {488 (CheckMMV3NzNzNdSupport(mmOpInfo) && CheckSupportInfoFormatNzNzNd(mmOpInfo)) || addmm16In32Out) {
484 OP_LOGI("Hit matmul_v3 scenario.");489 OP_LOGI("Hit matmul_v3 scenario.");
@@ -520,7 +525,7 @@ static const aclTensor* GetMatMulOp(
520 const aclTensor* mmOut =525 const aclTensor* mmOut =
521 l0op::MatMulV3Nd(x1, x2, bias, transposeX1, transposeX2, offsetX, opImplModeEnum, executor);526 l0op::MatMulV3Nd(x1, x2, bias, transposeX1, transposeX2, offsetX, opImplModeEnum, executor);
522 return mmOut;527 return mmOut;
523- } else if (enable16In32Out && bias == nullptr) {528+ } else if ((enable16In32Out || (isFp32Out && bothMatFp16Bf16)) && bias == nullptr) {
唐
唐唐超6月5日

这个应该是1980的分支，基本都不维护的，你确定应该在这里写吗

likedislike
HKFLYE
HKFLYE
6月5日 评论：
524 // This is Split K Mode; Check if MatMul using Nd in Nd Out529 // This is Split K Mode; Check if MatMul using Nd in Nd Out
525 OP_LOGI("hit matmulv2 fp16/bp16 in fp32 out case.");530 OP_LOGI("hit matmulv2 fp16/bp16 in fp32 out case.");
526 const aclTensor* mmOut =531 const aclTensor* mmOut =
Mmatmul/gemm_v3/op_kernel/gemm_v3_base_kernel.h+0-2
@@ -636,7 +636,6 @@ GemmV3BaseKernel<swizzleDirect, transA, transB, InDtype, BiasDtype, OutDtype, Ac
636 WaitFlag<HardEvent::MTE2_V>(EVENT_ID0);636 WaitFlag<HardEvent::MTE2_V>(EVENT_ID0);
637 for (uint32_t i = 0; i < count; ++i) {637 for (uint32_t i = 0; i < count; ++i) {
638 if constexpr (!std::is_same_v<BiasDtype, AccumDtype>) {638 if constexpr (!std::is_same_v<BiasDtype, AccumDtype>) {
639- AscendC::printf("need cast ubc dtype.\n");
640 AscendC::Cast<AccumDtype, BiasDtype, false>(ubSum_[i * VEC_ITER_NUMEL], // dst639 AscendC::Cast<AccumDtype, BiasDtype, false>(ubSum_[i * VEC_ITER_NUMEL], // dst
641 ubC_[i * VEC_ITER_NUMEL], // src640 ubC_[i * VEC_ITER_NUMEL], // src
642 AscendC::RoundMode::CAST_NONE, // mode641 AscendC::RoundMode::CAST_NONE, // mode
@@ -679,7 +678,6 @@ GemmV3BaseKernel<swizzleDirect, transA, transB, InDtype, BiasDtype, OutDtype, Ac
679 AscendC::UnaryRepeatParams(1, 1, 8, 8));678 AscendC::UnaryRepeatParams(1, 1, 8, 8));
680 AscendC::PipeBarrier<PIPE_V>();679 AscendC::PipeBarrier<PIPE_V>();
681 if constexpr (!std::is_same_v<OutDtype, AccumDtype>) {680 if constexpr (!std::is_same_v<OutDtype, AccumDtype>) {
682- AscendC::printf("need cast uby dtype.\n");
683 AscendC::Cast<OutDtype, AccumDtype, false>(ubY_[i * VEC_ITER_NUMEL], // dst681 AscendC::Cast<OutDtype, AccumDtype, false>(ubY_[i * VEC_ITER_NUMEL], // dst
684 ubSum_[i * VEC_ITER_NUMEL], // src682 ubSum_[i * VEC_ITER_NUMEL], // src
685 AscendC::RoundMode::CAST_RINT, // mode683 AscendC::RoundMode::CAST_RINT, // mode
Mmatmul/mat_mul_v3/op_host/config/ascend910_93/mat_mul_v3_binary.json+0-160
@@ -1762,166 +1762,6 @@
1762 }1762 }
1763 ]1763 ]
1764 },1764 },
1765- {
1766- "bin_filename": "MatMulV3_ND_NZ_ND_ND_FP16_FP16_FP32_FP32",
1767- "simplified_key": "diy,2/29/2/2/1/1/0/0",
1768- "inputs": [
1769- {
1770- "name": "x1",
1771- "index": 0,
1772- "dtype": "float16",
1773- "format": "ND",
1774- "paramType": "required",
1775- "shape": [
1776- -2
1777- ]
1778- },
1779- {
1780- "name": "x2",
1781- "index": 1,
1782- "dtype": "float16",
1783- "format": "FRACTAL_NZ",
1784- "paramType": "required",
1785- "shape": [
1786- -2
1787- ]
1788- },
1789- {
1790- "name": "bias",
1791- "index": 2,
1792- "dtype": "float32",
1793- "format": "ND",
1794- "paramType": "optional",
1795- "shape": [
1796- -2
1797- ]
1798- },
1799- {
1800- "name": "offset_w",
1801- "index": 3,
1802- "dtype": "int8",
1803- "format": "ND",
1804- "paramType": "optional",
1805- "shape": [
1806- -2
1807- ]
1808- }
1809- ],
1810- "outputs": [
1811- {
1812- "name": "y",
1813- "index": 0,
1814- "dtype": "float32",
1815- "format": "ND",
1816- "paramType": "required",
1817- "shape": [
1818- -2
1819- ]
1820- }
1821- ],
1822- "attrs": [
1823- {
1824- "name": "transpose_x1",
1825- "dtype": "bool",
1826- "value": false
1827- },
1828- {
1829- "name": "transpose_x2",
1830- "dtype": "bool",
1831- "value": false
1832- },
1833- {
1834- "name": "offset_x",
1835- "dtype": "int",
1836- "value": 0
1837- },
1838- {
1839- "name": "opImplMode",
1840- "dtype": "int",
1841- "value": 1
1842- }
1843- ]
1844- },
1845- {
1846- "bin_filename": "MatMulV3_ND_NZ_ND_ND_BF16_BF16_FP32_FP32",
1847- "simplified_key": "diy,2/29/2/2/27/27/0/0",
1848- "inputs": [
1849- {
1850- "name": "x1",
1851- "index": 0,
1852- "dtype": "bfloat16",
1853- "format": "ND",
1854- "paramType": "required",
1855- "shape": [
1856- -2
1857- ]
1858- },
1859- {
1860- "name": "x2",
1861- "index": 1,
1862- "dtype": "bfloat16",
1863- "format": "FRACTAL_NZ",
1864- "paramType": "required",
1865- "shape": [
1866- -2
1867- ]
1868- },
1869- {
1870- "name": "bias",
1871- "index": 2,
1872- "dtype": "float32",
1873- "format": "ND",
1874- "paramType": "optional",
1875- "shape": [
1876- -2
1877- ]
1878- },
1879- {
1880- "name": "offset_w",
1881- "index": 3,
1882- "dtype": "int8",
1883- "format": "ND",
1884- "paramType": "optional",
1885- "shape": [
1886- -2
1887- ]
1888- }
1889- ],
1890- "outputs": [
1891- {
1892- "name": "y",
1893- "index": 0,
1894- "dtype": "float32",
1895- "format": "ND",
1896- "paramType": "required",
1897- "shape": [
1898- -2
1899- ]
1900- }
1901- ],
1902- "attrs": [
1903- {
1904- "name": "transpose_x1",
1905- "dtype": "bool",
1906- "value": false
1907- },
1908- {
1909- "name": "transpose_x2",
1910- "dtype": "bool",
1911- "value": false
1912- },
1913- {
1914- "name": "offset_x",
1915- "dtype": "int",
1916- "value": 0
1917- },
1918- {
1919- "name": "opImplMode",
1920- "dtype": "int",
1921- "value": 1
1922- }
1923- ]
1924- },
1925 {1765 {
1926 "bin_filename": "MatMulV3_ND_NZ_ND_ND_FP16_FP16_FP16_FP32",1766 "bin_filename": "MatMulV3_ND_NZ_ND_ND_FP16_FP16_FP16_FP32",
1927 "simplified_key": "diy,2/29/2/2/1/1/1/0",1767 "simplified_key": "diy,2/29/2/2/1/1/1/0",
Mmatmul/mat_mul_v3/op_host/config/ascend910b/mat_mul_v3_binary.json+0-160
@@ -1762,166 +1762,6 @@
1762 }1762 }
1763 ]1763 ]
1764 },1764 },
1765- {
1766- "bin_filename": "MatMulV3_ND_NZ_ND_ND_FP16_FP16_FP32_FP32",
1767- "simplified_key": "diy,2/29/2/2/1/1/0/0",
1768- "inputs": [
1769- {
1770- "name": "x1",
1771- "index": 0,
1772- "dtype": "float16",
1773- "format": "ND",
1774- "paramType": "required",
1775- "shape": [
1776- -2
1777- ]
1778- },
1779- {
1780- "name": "x2",
1781- "index": 1,
1782- "dtype": "float16",
1783- "format": "FRACTAL_NZ",
1784- "paramType": "required",
1785- "shape": [
1786- -2
1787- ]
1788- },
1789- {
1790- "name": "bias",
1791- "index": 2,
1792- "dtype": "float32",
1793- "format": "ND",
1794- "paramType": "optional",
1795- "shape": [
1796- -2
1797- ]
1798- },
1799- {
1800- "name": "offset_w",
1801- "index": 3,
1802- "dtype": "int8",
1803- "format": "ND",
1804- "paramType": "optional",
1805- "shape": [
1806- -2
1807- ]
1808- }
1809- ],
1810- "outputs": [
1811- {
1812- "name": "y",
1813- "index": 0,
1814- "dtype": "float32",
1815- "format": "ND",
1816- "paramType": "required",
1817- "shape": [
1818- -2
1819- ]
1820- }
1821- ],
1822- "attrs": [
1823- {
1824- "name": "transpose_x1",
1825- "dtype": "bool",
1826- "value": false
1827- },
1828- {
1829- "name": "transpose_x2",
1830- "dtype": "bool",
1831- "value": false
1832- },
1833- {
1834- "name": "offset_x",
1835- "dtype": "int",
1836- "value": 0
1837- },
1838- {
1839- "name": "opImplMode",
1840- "dtype": "int",
1841- "value": 1
1842- }
1843- ]
1844- },
1845- {
1846- "bin_filename": "MatMulV3_ND_NZ_ND_ND_BF16_BF16_FP32_FP32",
1847- "simplified_key": "diy,2/29/2/2/27/27/0/0",
1848- "inputs": [
1849- {
1850- "name": "x1",
1851- "index": 0,
1852- "dtype": "bfloat16",
1853- "format": "ND",
1854- "paramType": "required",
1855- "shape": [
1856- -2
1857- ]
1858- },
1859- {
1860- "name": "x2",
1861- "index": 1,
1862- "dtype": "bfloat16",
1863- "format": "FRACTAL_NZ",
1864- "paramType": "required",
1865- "shape": [
1866- -2
1867- ]
1868- },
1869- {
1870- "name": "bias",
1871- "index": 2,
1872- "dtype": "float32",
1873- "format": "ND",
1874- "paramType": "optional",
1875- "shape": [
1876- -2
1877- ]
1878- },
1879- {
1880- "name": "offset_w",
1881- "index": 3,
1882- "dtype": "int8",
1883- "format": "ND",
1884- "paramType": "optional",
1885- "shape": [
1886- -2
1887- ]
1888- }
1889- ],
1890- "outputs": [
1891- {
1892- "name": "y",
1893- "index": 0,
1894- "dtype": "float32",
1895- "format": "ND",
1896- "paramType": "required",
1897- "shape": [
1898- -2
1899- ]
1900- }
1901- ],
1902- "attrs": [
1903- {
1904- "name": "transpose_x1",
1905- "dtype": "bool",
1906- "value": false
1907- },
1908- {
1909- "name": "transpose_x2",
1910- "dtype": "bool",
1911- "value": false
1912- },
1913- {
1914- "name": "offset_x",
1915- "dtype": "int",
1916- "value": 0
1917- },
1918- {
1919- "name": "opImplMode",
1920- "dtype": "int",
1921- "value": 1
1922- }
1923- ]
1924- },
1925 {1765 {
1926 "bin_filename": "MatMulV3_ND_NZ_ND_ND_FP16_FP16_FP16_FP32",1766 "bin_filename": "MatMulV3_ND_NZ_ND_ND_FP16_FP16_FP16_FP32",
1927 "simplified_key": "diy,2/29/2/2/1/1/1/0",1767 "simplified_key": "diy,2/29/2/2/1/1/1/0",
Mmatmul/mat_mul_v3/op_host/op_api/aclnn_addmm.cpp+1-1
@@ -508,7 +508,7 @@ public:
508 // A2/A3上对于 16in32out,且不需要broadcast场景 直接走gemmV3508 // A2/A3上对于 16in32out,且不需要broadcast场景 直接走gemmV3
509 if ((CheckGemmV3WithAlphaBeta(bias, matA, matB, cubeMathType) ||509 if ((CheckGemmV3WithAlphaBeta(bias, matA, matB, cubeMathType) ||
510 (enable16In32Out && !needBroadcast)) && isSupportNpuArch) {510 (enable16In32Out && !needBroadcast)) && isSupportNpuArch) {
511- auto outGemmV3 = ExecGemmV3WithAlphaBetaOp(bias, matA, matB, alpha, beta, executor);511+ auto outGemmV3 = ExecGemmV3WithAlphaBetaOp(bias, matA, matB, alpha, beta, executor, enable16In32Out);
512 CHECK_RET(outGemmV3 != nullptr, ACLNN_ERR_INNER_NULLPTR);512 CHECK_RET(outGemmV3 != nullptr, ACLNN_ERR_INNER_NULLPTR);
513 convOut = outGemmV3;513 convOut = outGemmV3;
514 return ACLNN_SUCCESS;514 return ACLNN_SUCCESS;