Merged
【WeightQuantBatchMamtul】【9202】 A16系列适配9202 #14232
【WeightQuantBatchMamtul】【9202】 A16系列适配9202 #14232
Merged
ruan-cjAdd on 10 days ago
Total: 11 files changed+3236-48
@@ -18,7 +18,7 @@ add_kernel_sources(
18 18 
19add_kernel_sources(19add_kernel_sources(
20 KERNEL_SRC arch35/weight_quant_batch_matmul_v2.cpp20 KERNEL_SRC arch35/weight_quant_batch_matmul_v2.cpp
21- COMPUTE_UNITS ascend950 ascend350 ascend960dt21+ COMPUTE_UNITS ascend950 ascend350 ascend960dt ascend960pr
22 AUTO_SYNC false22 AUTO_SYNC false
23 OPTIONS "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true"23 OPTIONS "--cce-no-dcache-flush" "-DENABLE_CV_COMM_VIA_SSBUF=true"
24)24)
@@ -151,7 +151,7 @@
151 {151 {
152 "name": "antiquant_offset",152 "name": "antiquant_offset",
153 "index": 3,153 "index": 3,
154- "dtype": "bfloat16",154+ "dtype": "float16",
155 "format": "ND",155 "format": "ND",
156 "paramType": "optional",156 "paramType": "optional",
157 "shape": [157 "shape": [
@@ -2659,7 +2659,7 @@
2659 {2659 {
2660 "name": "antiquant_offset",2660 "name": "antiquant_offset",
2661 "index": 3,2661 "index": 3,
2662- "dtype": "float16",2662+ "dtype": "bfloat16",
2663 "format": "ND",2663 "format": "ND",
2664 "paramType": "optional",2664 "paramType": "optional",
2665 "shape": [2665 "shape": [
@@ -779,15 +779,7 @@ void WeightQuantBatchMatmulV2TilingASWMix::CalL1Tiling()
779 OP_LOGD(opName_, "coreNum: %u", basicTiling_.usedCoreNum);779 OP_LOGD(opName_, "coreNum: %u", basicTiling_.usedCoreNum);
780 basicTiling_.baseM = adaptiveWin_.baseM;780 basicTiling_.baseM = adaptiveWin_.baseM;
781 basicTiling_.baseN = adaptiveWin_.baseN;781 basicTiling_.baseN = adaptiveWin_.baseN;
782- 782+ basicTiling_.baseK = std::min(adaptiveWin_.baseK, ops::CeilAlign<uint64_t>(matmulInfoPtr_->kSize, B4_C0_SIZE));
atomgit-bot
atomgit-botatomgit-bot10 days ago

🟡 Medium Priority

建议:仅在 4bit 权重(Is4BitDtype/HiFp4/MxFp4)时用 CeilAlign(kSize, B4_C0_SIZE),其余 dtype 继续用 min(adaptiveWin_.baseK, kSize),保证 baseK 不超过 kSize。

likedislike
Incorrect?
783- bool isA16MxFp4 = (matmulInfoPtr_->aDtype == ge::DT_FLOAT16 || matmulInfoPtr_->aDtype == ge::DT_BF16) &&
784- matmulInfoPtr_->bDtype == ge::DT_FLOAT4_E2M1 && matmulInfoPtr_->antiQuantType == QuantType::MX;
785- if (isA16MxFp4) {
786- basicTiling_.baseK = std::min(adaptiveWin_.baseK, ops::CeilAlign<uint64_t>(matmulInfoPtr_->kSize, B4_C0_SIZE));
787- } else {
788- basicTiling_.baseK = std::min(adaptiveWin_.baseK, matmulInfoPtr_->kSize);
789- }
790- 
791 basicTiling_.stepM = 1;783 basicTiling_.stepM = 1;
792 basicTiling_.stepN = 1;784 basicTiling_.stepN = 1;
793 basicTiling_.singleCoreM = std::min(matmulInfoPtr_->mSize, static_cast<uint64_t>(basicTiling_.baseM));785 basicTiling_.singleCoreM = std::min(matmulInfoPtr_->mSize, static_cast<uint64_t>(basicTiling_.baseM));
@@ -423,13 +423,14 @@ bool WeightQuantBatchMatmulV2CheckerFor920R1::CheckInputSize(const gert::Storage
423 return false);423 return false);
424 424 
425 isA16MxFp4_ = inputParams_.bDtype == ge::DT_FLOAT4_E2M1 && inputParams_.antiQuantScaleDtype == ge::DT_FLOAT8_E8M0;425 isA16MxFp4_ = inputParams_.bDtype == ge::DT_FLOAT4_E2M1 && inputParams_.antiQuantScaleDtype == ge::DT_FLOAT8_E8M0;
426+ isA16HiF4_ = inputParams_.bDtype == Ops::NN::DT_HIFLOAT4;
426 inputParams_.nSize = static_cast<uint64_t>(inputParams_.transB ? weightShape->GetOriginShape().GetDim(0) :427 inputParams_.nSize = static_cast<uint64_t>(inputParams_.transB ? weightShape->GetOriginShape().GetDim(0) :
427 weightLastDim);428 weightLastDim);
428 auto weightStorageShape = weightShape->GetStorageShape();429 auto weightStorageShape = weightShape->GetStorageShape();
429 if (inputParams_.bFormat == ge::Format::FORMAT_FRACTAL_NZ && inputParams_.transB) {430 if (inputParams_.bFormat == ge::Format::FORMAT_FRACTAL_NZ && inputParams_.transB) {
430 uint64_t n0 = weightStorageShape.GetDim(2);431 uint64_t n0 = weightStorageShape.GetDim(2);
431 uint64_t n1 = weightStorageShape.GetDim(1);432 uint64_t n1 = weightStorageShape.GetDim(1);
432- OP_TILING_CHECK(!isA16MxFp4_ && n0 * n1 != inputParams_.nSize,433+ OP_TILING_CHECK(!isA16MxFp4_ && !isA16HiF4_ && n0 * n1 != inputParams_.nSize,
433 OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(434 OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(
434 inputParams_.opName, "weight",435 inputParams_.opName, "weight",
435 Ops::NN::FormatString("%llu, %llu", static_cast<unsigned long long>(n0 * n1),436 Ops::NN::FormatString("%llu, %llu", static_cast<unsigned long long>(n0 * n1),
@@ -650,8 +651,7 @@ bool WeightQuantBatchMatmulV2CheckerFor920R1::CheckShapeAdvanced()
650 .c_str(),651 .c_str(),
651 "the K and N dimensions must be divisible by 2"),652 "the K and N dimensions must be divisible by 2"),
652 return false);653 return false);
653- } else if (inputParams_.bDtype == ge::DT_FLOAT4_E2M1 ||654+ } else if (inputParams_.bDtype == ge::DT_FLOAT4_E2M1) {
654- inputParams_.bDtype == Ops::NN::DT_HIFLOAT4) { // 其余 A16W4 仅支持 n 和 k 对 64 对齐
655 OP_TILING_CHECK(655 OP_TILING_CHECK(
656 // 64对齐656 // 64对齐
657 !isA16MxFp4_ && (inputParams_.nSize % 64 != 0 || inputParams_.kSize % 64 != 0),657 !isA16MxFp4_ && (inputParams_.nSize % 64 != 0 || inputParams_.kSize % 64 != 0),
@@ -98,7 +98,8 @@ static ge::graphStatus WeightQuantBatchMatmulV2TilingFunc(gert::TilingContext* c
98 }98 }
99 auto platformInfoPtr = context->GetPlatformInfo();99 auto platformInfoPtr = context->GetPlatformInfo();
100 if (platformInfoPtr != nullptr &&100 if (platformInfoPtr != nullptr &&
101- platform_ascendc::PlatformAscendC(platformInfoPtr).GetCurNpuArch() == NpuArch::DAV_9201) {101+ (platform_ascendc::PlatformAscendC(platformInfoPtr).GetCurNpuArch() == NpuArch::DAV_9201 ||
102+ platform_ascendc::PlatformAscendC(platformInfoPtr).GetCurNpuArch() == NpuArch::DAV_9202)) {
102 WeightQuantBatchMatmulV2CheckerFor920R1 wqbmmv2Checker(context);103 WeightQuantBatchMatmulV2CheckerFor920R1 wqbmmv2Checker(context);
103 OP_TILING_CHECK(104 OP_TILING_CHECK(
104 wqbmmv2Checker.Check() != ge::GRAPH_SUCCESS,105 wqbmmv2Checker.Check() != ge::GRAPH_SUCCESS,
@@ -615,8 +615,8 @@ public:
615 opbase::ArchConfigHelper::AddConfigByArch(this->AICore(), Ops::Base::DAV_3510, aicore_config_dav3510);615 opbase::ArchConfigHelper::AddConfigByArch(this->AICore(), Ops::Base::DAV_3510, aicore_config_dav3510);
616 this->AICore().AddConfig("ascend350", aicore_config_dav3510);616 this->AICore().AddConfig("ascend350", aicore_config_dav3510);
617 617 
618- OpAICoreConfig aicore_config_960DT;618+ OpAICoreConfig aicore_config_960;
619- aicore_config_960DT.Input("x")619+ aicore_config_960.Input("x")
620 .ParamType(REQUIRED)620 .ParamType(REQUIRED)
621 .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,621 .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,
622 ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16,622 ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16,
@@ -637,7 +637,7 @@ public:
637 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,637 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
638 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,638 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
639 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});639 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
640- aicore_config_960DT.Input("weight")640+ aicore_config_960.Input("weight")
641 .ParamType(REQUIRED)641 .ParamType(REQUIRED)
642 .DataType({ge::DT_INT8, ge::DT_INT8, ge::DT_INT8, ge::DT_INT8,642 .DataType({ge::DT_INT8, ge::DT_INT8, ge::DT_INT8, ge::DT_INT8,
643 ge::DT_INT8, ge::DT_INT4, ge::DT_INT4, ge::DT_FLOAT4_E2M1,643 ge::DT_INT8, ge::DT_INT4, ge::DT_INT4, ge::DT_FLOAT4_E2M1,
@@ -677,7 +677,7 @@ public:
677 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ,677 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ,
678 ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ,678 ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ,
679 ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ});679 ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ, ge::FORMAT_FRACTAL_NZ});
680- aicore_config_960DT.Input("antiquant_scale")680+ aicore_config_960.Input("antiquant_scale")
681 .ParamType(REQUIRED)681 .ParamType(REQUIRED)
682 .DataType({ge::DT_UINT64,682 .DataType({ge::DT_UINT64,
683 ge::DT_UINT64,683 ge::DT_UINT64,
@@ -727,7 +727,7 @@ public:
727 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,727 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
728 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,728 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
729 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});729 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
730- aicore_config_960DT.Input("antiquant_offset")730+ aicore_config_960.Input("antiquant_offset")
731 .ParamType(OPTIONAL)731 .ParamType(OPTIONAL)
732 .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,732 .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,
733 ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16,733 ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16,
@@ -748,7 +748,7 @@ public:
748 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,748 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
749 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,749 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
750 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});750 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
751- aicore_config_960DT.Input("quant_scale")751+ aicore_config_960.Input("quant_scale")
752 .ParamType(OPTIONAL)752 .ParamType(OPTIONAL)
753 .DataType({ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64,753 .DataType({ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64,
754 ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64,754 ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64, ge::DT_UINT64,
@@ -769,7 +769,7 @@ public:
769 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,769 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
770 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,770 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
771 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});771 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
772- aicore_config_960DT.Input("quant_offset")772+ aicore_config_960.Input("quant_offset")
773 .ParamType(OPTIONAL)773 .ParamType(OPTIONAL)
774 .DataType({ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT,774 .DataType({ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT,
775 ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT,775 ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT,
@@ -790,7 +790,7 @@ public:
790 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,790 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
791 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,791 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
792 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});792 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
793- aicore_config_960DT.Input("bias")793+ aicore_config_960.Input("bias")
794 .ParamType(OPTIONAL)794 .ParamType(OPTIONAL)
795 .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,795 .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,
796 ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT, ge::DT_FLOAT16, ge::DT_BF16,796 ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT, ge::DT_FLOAT16, ge::DT_BF16,
@@ -811,7 +811,7 @@ public:
811 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,811 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
812 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,812 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
813 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});813 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
814- aicore_config_960DT.Output("y")814+ aicore_config_960.Output("y")
815 .ParamType(REQUIRED)815 .ParamType(REQUIRED)
816 .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,816 .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_FLOAT16,
817 ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16,817 ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16, ge::DT_BF16, ge::DT_FLOAT16, ge::DT_BF16,
@@ -832,14 +832,15 @@ public:
832 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,832 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
833 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,833 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND,
834 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});834 ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
835- aicore_config_960DT.DynamicCompileStaticFlag(true)835+ aicore_config_960.DynamicCompileStaticFlag(true)
836 .DynamicFormatFlag(true)836 .DynamicFormatFlag(true)
837 .DynamicRankSupportFlag(true)837 .DynamicRankSupportFlag(true)
838 .DynamicShapeSupportFlag(true)838 .DynamicShapeSupportFlag(true)
839 .NeedCheckSupportFlag(false)839 .NeedCheckSupportFlag(false)
840 .PrecisionReduceFlag(true)840 .PrecisionReduceFlag(true)
841 .ExtendCfgInfo("aclnnSupport.value", "support_aclnn");841 .ExtendCfgInfo("aclnnSupport.value", "support_aclnn");
842- this->AICore().AddConfig("ascend960dt", aicore_config_960DT);842+ this->AICore().AddConfig("ascend960dt", aicore_config_960);
843+ this->AICore().AddConfig("ascend960pr", aicore_config_960);
atomgit-bot
atomgit-botatomgit-bot10 days ago

🟡 Medium Priority

变更行 834-843:新增 this->AICore().AddConfig("ascend960pr", aicore_config_960);,把同一个 AICore 配置同时注册到 ascend960dt 与 ascend960pr 两个平台。

证据链:本 op 的 kernel 编译单元在 matmul/weight_quant_batch_matmul_v2/CMakeLists.txt 行 21 声明为 COMPUTE_UNITS ascend950 ascend350 ascend960dt,本次 diff 未同步把 ascend960pr 加入该列表(diff 只改了 op_host 的 def 文件)。

失败模式:若构建系统按 COMPUTE_UNITS 逐个平台编译并打包 kernel bin,则 ascend960pr 平台上该算子只有 host 侧的 op 定义、没有对应的 kernel 二进制;图编译/下发阶段会因找不到该 soc 的 kernel 而报 tiling/bin 缺失错误,或在 fallback 到 ascend960dt 的 bin 时与 9202 的实际指令集不匹配。触发条件为在 ascend960pr 平台上运行 weight_quant_batch_matmul_v2。

建议:确认 ascend960pr 的 kernel 产物是否已由构建系统覆盖(必要时在 matmul/weight_quant_batch_matmul_v2/CMakeLists.txt 的 COMPUTE_UNITS 中补充 ascend960pr),避免只注册 host 配置而无对应 kernel bin。

likedislike
Incorrect?
843 }844 }
844};845};
845 846 
@@ -355,12 +355,9 @@ __aicore__ inline void WeightQuantBatchMatmulV2ASWMxKernel<LOCAL_TEMPLATE_FUNC_P
355 uint64_t nL1Size = Min<uint64_t>(tiling_->matmulTiling.N - block_.offset_.offsetN, block_.params_.singleCoreN);355 uint64_t nL1Size = Min<uint64_t>(tiling_->matmulTiling.N - block_.offset_.offsetN, block_.params_.singleCoreN);
356 uint64_t scaleGmOffset = (block_.offset_.offsetN * tiling_->matmulTiling.Kb + kOffset) / MX_GROUP_SIZE;356 uint64_t scaleGmOffset = (block_.offset_.offsetN * tiling_->matmulTiling.Kb + kOffset) / MX_GROUP_SIZE;
357 357 
358- // Mx非对齐场景,weight和weight scale的K轴需要对齐到C0 Size(64)358+ kL1Size = CeilAlign<uint64_t>(kL1Size, C0<wType>::value);
359- if constexpr (IsSameType<wType, fp4x2_e2m1_t>::value) {359+ uint64_t kAlignSize = CeilAlign<uint64_t>(tiling_->matmulTiling.Kb, C0<wType>::value);
360- kL1Size = CeilAlign<uint64_t>(kL1Size, C0<wType>::value);360+ scaleGmOffset = (block_.offset_.offsetN * kAlignSize + kOffset) / MX_GROUP_SIZE;
atomgit-bot
atomgit-botatomgit-bot10 days ago

🟡 Medium Priority

变更行(CopyScaleToL1,行 358-360):原代码把「kL1Size 对齐到 C0」「kAlignSize 重算」「scaleGmOffset 按对齐后的 K 重算」三件事都包在 if constexpr (IsSameType<wType, fp4x2_e2m1_t>::value) 内,本次改为对所有 wType 无条件执行。

影响范围:同一个 WeightQuantBatchMatmulV2ASWMxKernel 并非只被 fp4x2_e2m1_t 实例化——本文件行 313、367 都显式判断 IsSameType<wType, hifloat4x2_t>::value,且 op_kernel/arch35/weight_quant_batch_matmul_v2.cpp 行 339 用 defined(WEIGHT_HIFP4) 也实例化该 MX kernel,因此 wType=hifloat4x2_t(HIFP4)会走同一函数。

失败模式:对 HIFP4 而言,scaleGmOffset 由 offsetN * Kb + kOffset(旧)变成 offsetN * CeilAlign(Kb, C0<wType>::value) + kOffset(新),kL1Size 也被向上对齐。只要 host tiling 下发的 matmulTiling.Kb(或 N 侧 stride)不是 C0<hifloat4x2_t>::value 的整数倍,scale 的 GM 读偏移就会整体平移,读到相邻分块的 scale(精度错误),末块还会按对齐后的长度越界读取 scale GM。触发条件是 Kb % C0 != 0 的 HIFP4 用例;若 host 侧已保证 Kb/N 对齐,则行为等价。

likedislike
Incorrect?
361- uint64_t kAlignSize = CeilAlign<uint64_t>(tiling_->matmulTiling.Kb, C0<wType>::value);
362- scaleGmOffset = (block_.offset_.offsetN * kAlignSize + kOffset) / MX_GROUP_SIZE;
363- }
364 361 
365 uint64_t scaleGmSize;362 uint64_t scaleGmSize;
366 Dn2NzParams dn2NzParams;363 Dn2NzParams dn2NzParams;
@@ -434,13 +431,11 @@ __aicore__ inline void WeightQuantBatchMatmulV2ASWMxKernel<LOCAL_TEMPLATE_FUNC_P
434 dataCopyParams.dstStride = CeilAlign<uint64_t>(nL1Size, static_cast<uint64_t>(AscendC::BLOCK_CUBE)) - nL1Size;431 dataCopyParams.dstStride = CeilAlign<uint64_t>(nL1Size, static_cast<uint64_t>(AscendC::BLOCK_CUBE)) - nL1Size;
435 432 
436 // weight shape:(k1,n1,(n0=16,k0=64))433 // weight shape:(k1,n1,(n0=16,k0=64))
437- if constexpr (IsSameType<wType, fp4x2_e2m1_t>::value) {434+ uint64_t nAlignSize = CeilAlign<uint64_t>(tiling_->matmulTiling.N, AscendC::BLOCK_CUBE);
438- uint64_t nAlignSize = CeilAlign<uint64_t>(tiling_->matmulTiling.N, AscendC::BLOCK_CUBE);435+ weightGmOffset = (kOffset * nAlignSize + block_.offset_.offsetN * C0<wType>::value) >> 1;
439- weightGmOffset = (kOffset * nAlignSize + block_.offset_.offsetN * C0<wType>::value) >> 1;
440 436 
441- uint64_t nL1AlignSize = CeilAlign<uint64_t>(nL1Size, AscendC::BLOCK_CUBE);437+ uint64_t nL1AlignSize = CeilAlign<uint64_t>(nL1Size, AscendC::BLOCK_CUBE);
442- dataCopyParams.srcStride = nAlignSize - nL1Size;438+ dataCopyParams.srcStride = nAlignSize - nL1Size;
atomgit-bot
atomgit-botatomgit-bot10 days ago

🟡 Medium Priority

变更行(CopyWeightGm2L1,行 434-438):原来只有 wType == fp4x2_e2m1_t 时才把 nAlignSize = CeilAlign(N, BLOCK_CUBE) 用于 weightGmOffset 与 dataCopyParams.srcStride;本次改为对所有 wType 无条件覆盖(行 435、438)。

证据链:本文件行 313、367 明确存在 IsSameType<wType, hifloat4x2_t> 分支,op_kernel/arch35/weight_quant_batch_matmul_v2.cpp 行 339 的 defined(WEIGHT_HIFP4) 也会实例化该 MX kernel,说明 wType=hifloat4x2_t(HIFP4)同样走此函数。改动前 HIFP4 使用行 430 的 srcStride = N - nL1Size,改动后被行 438 覆盖为 CeilAlign(N, BLOCK_CUBE) - nL1Size,同时 weightGmOffset 的行步长从 N 变为 CeilAlign(N, 16)。

失败模式:若 host tiling 下发的 matmulTiling.N 不是 16 的整数倍,HIFP4 每次 MTE2 搬运的行步长会与实际 GM 布局错位,取到错位的 weight 数据(精度错误);若 N 已对齐则行为等价。触发条件为 HIFP4 + N % 16 != 0 的用例。

建议:核实 tiling 侧 N 是否对 HIFP4 保证 BLOCK_CUBE 对齐;未保证时让 srcStride/weightGmOffset 按 wType 分支选择 N 或 CeilAlign(N, BLOCK_CUBE),并补充 N 非 16 倍数的 HIFP4 用例。

likedislike
Incorrect?
443- }
444 439 
445 // 最后一个参数enablePreReadIgnoreSync = true表示开启PreAllocation440 // 最后一个参数enablePreReadIgnoreSync = true表示开启PreAllocation
446 AscendC::DataCopy((l1DbFlag_ == IDX_0) ? bL1PingLocalTensor_.template ReinterpretCast<uint8_t>() :441 AscendC::DataCopy((l1DbFlag_ == IDX_0) ? bL1PingLocalTensor_.template ReinterpretCast<uint8_t>() :
@@ -26,14 +26,15 @@
26#include "lib/matmul_intf.h"26#include "lib/matmul_intf.h"
27#include "weight_quant_batch_matmul_v2_arch35_tiling_key.h"27#include "weight_quant_batch_matmul_v2_arch35_tiling_key.h"
28#include "weight_quant_batch_matmul_v2_arch35_tiling_data.h"28#include "weight_quant_batch_matmul_v2_arch35_tiling_data.h"
29-#if (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201))29+#if (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202))
30#include "asw/weight_quant_batch_matmul_v2_adaptive_sliding_window.h"30#include "asw/weight_quant_batch_matmul_v2_adaptive_sliding_window.h"
31#include "asw/weight_quant_batch_matmul_v2_adaptive_sliding_window_mix.h"31#include "asw/weight_quant_batch_matmul_v2_adaptive_sliding_window_mix.h"
32#include "asw/weight_quant_batch_matmul_v2_adaptive_sliding_window_mx.h"32#include "asw/weight_quant_batch_matmul_v2_adaptive_sliding_window_mx.h"
33#include "asw/weight_quant_batch_matmul_v2_adaptive_sliding_window_general.h"33#include "asw/weight_quant_batch_matmul_v2_adaptive_sliding_window_general.h"
34#endif34#endif
35 35 
36-#if !defined(__DAV_310R6__) && !__FIXED_POINT_ONLY_CUBE_TO_L0C__ && !(defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201))36+#if !defined(__DAV_310R6__) && !__FIXED_POINT_ONLY_CUBE_TO_L0C__ && \
37+ !(defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202))
37#if defined(A16) && (defined(WQBMMV2_S4) || defined(FP4) || (defined(WQBMMV2_S8) && defined(WEIGHT_ND)))38#if defined(A16) && (defined(WQBMMV2_S4) || defined(FP4) || (defined(WQBMMV2_S8) && defined(WEIGHT_ND)))
38#include "weight_quant_batch_matmul_v2_reg_base.h"39#include "weight_quant_batch_matmul_v2_reg_base.h"
39#endif40#endif
@@ -77,7 +78,7 @@ constexpr MatmulBatchParams batchParamsNoBatchOut{false, BatchMode::BATCH_LESS_T
77 BatchOutMode::SINGLE_BATCH};78 BatchOutMode::SINGLE_BATCH};
78constexpr MatmulConfig MM_CFG_MULTI_BATCH_NO_BATCH_OUT = GetMMConfig<configMode>(batchParamsNoBatchOut);79constexpr MatmulConfig MM_CFG_MULTI_BATCH_NO_BATCH_OUT = GetMMConfig<configMode>(batchParamsNoBatchOut);
79 80 
80-#if !__FIXED_POINT_ONLY_CUBE_TO_L0C__ && !(defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201))81+#if !__FIXED_POINT_ONLY_CUBE_TO_L0C__ && !(defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202))
81static constexpr VecAntiQuantConfig VEC_ANTIQUANT_CONFIGS[] = {82static constexpr VecAntiQuantConfig VEC_ANTIQUANT_CONFIGS[] = {
82 {2, 512}, // idx 083 {2, 512}, // idx 0
83 {4, 512}, // idx 184 {4, 512}, // idx 1
@@ -148,7 +149,7 @@ static constexpr VecAntiQuantConfig VEC_ANTIQUANT_CONFIG_5 = VEC_ANTIQUANT_CONFI
148 } while (0)149 } while (0)
149#endif150#endif
150 151 
151-#if (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201))152+#if (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202))
152#define INVOKE_WEIGHT_QUANT_BMM_OP_ASW_IMPL_9201(templateClass, ...) \153#define INVOKE_WEIGHT_QUANT_BMM_OP_ASW_IMPL_9201(templateClass, ...) \
153 do { \154 do { \
154 AscendC::TPipe tPipe; \155 AscendC::TPipe tPipe; \
@@ -187,14 +188,15 @@ __global__ __aicore__ void weight_quant_batch_matmul_v2(GM_ADDR x, GM_ADDR weigh
187 GM_ADDR quantOffset, GM_ADDR bias, GM_ADDR y, GM_ADDR workspace,188 GM_ADDR quantOffset, GM_ADDR bias, GM_ADDR y, GM_ADDR workspace,
188 GM_ADDR tiling)189 GM_ADDR tiling)
189{190{
190-#if (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201))191+#if (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202))
191 GM_ADDR userWS = GetUserWorkspace(workspace);192 GM_ADDR userWS = GetUserWorkspace(workspace);
192 if (userWS == nullptr) {193 if (userWS == nullptr) {
193 return;194 return;
194 }195 }
195#endif196#endif
196 197 
197-#if !defined(__DAV_310R6__) && !__FIXED_POINT_ONLY_CUBE_TO_L0C__ && !(defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201))198+#if !defined(__DAV_310R6__) && !__FIXED_POINT_ONLY_CUBE_TO_L0C__ && \
199+ !(defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202))
198 KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_MIX_AIC_1_2);200 KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_MIX_AIC_1_2);
199 REGISTER_TILING_DEFAULT(wqbmmv2_tiling::WeightQuantBatchMatmulV2ASTilingDataParams);201 REGISTER_TILING_DEFAULT(wqbmmv2_tiling::WeightQuantBatchMatmulV2ASTilingDataParams);
200 using DtypeBias = AscendC::Std::conditional_t<IS_BIAS_FP32, float, DTYPE_X>;202 using DtypeBias = AscendC::Std::conditional_t<IS_BIAS_FP32, float, DTYPE_X>;
@@ -328,7 +330,7 @@ __global__ __aicore__ void weight_quant_batch_matmul_v2(GM_ADDR x, GM_ADDR weigh
328 VEC_ANTIQUANT_CONFIG_5);330 VEC_ANTIQUANT_CONFIG_5);
329 }331 }
330#endif332#endif
331-#elif (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201))333+#elif (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202))
332 using DtypeBias = AscendC::Std::conditional_t<IS_BIAS_FP32, float, DTYPE_X>;334 using DtypeBias = AscendC::Std::conditional_t<IS_BIAS_FP32, float, DTYPE_X>;
333 static constexpr WeightQuantBatchMatmulV2::QuantType335 static constexpr WeightQuantBatchMatmulV2::QuantType
334 AntiquantType = static_cast<WeightQuantBatchMatmulV2::QuantType>(ANTIQUANT_TYPE);336 AntiquantType = static_cast<WeightQuantBatchMatmulV2::QuantType>(ANTIQUANT_TYPE);
@@ -643,7 +643,7 @@ ASCENDC_TPL_SEL(
643#endif643#endif
644#endif644#endif
645 645 
646-#if !defined(__CCE_AICORE__) || (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201))646+#if !defined(__CCE_AICORE__) || (defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202))
647#if !defined(__CCE_AICORE__) || (defined(WEIGHT_ND))647#if !defined(__CCE_AICORE__) || (defined(WEIGHT_ND))
648#if (defined(A16) && defined(WQBMMV2_HIFP8))648#if (defined(A16) && defined(WQBMMV2_HIFP8))
649 ASCENDC_TPL_ARGS_SEL(ASCENDC_TPL_KERNEL_TYPE_SEL(ASCENDC_TPL_MIX_AIC_1_1),649 ASCENDC_TPL_ARGS_SEL(ASCENDC_TPL_KERNEL_TYPE_SEL(ASCENDC_TPL_MIX_AIC_1_1),
@@ -694,7 +694,7 @@ vector<CheckerCase> BuildCheckerCases()
694 {694 {
695 auto c = hiF4();695 auto c = hiF4();
696 c.n = 96;696 c.n = 96;
697- add("align_hif4_n_not_64_fail", false, c);697+ add("align_hif4_n_not_64_ok", true, c);
698 }698 }
699 {699 {
700 auto c = w8NzC032PerGroup();700 auto c = w8NzC032PerGroup();