Merged
fix(matmul_v3): 按输出 dtype 选择 fixpipe 1V1/1V2 #14100
fix(matmul_v3): 按输出 dtype 选择 fixpipe 1V1/1V2 #14100
Merged
link164735Add on 12 days ago
Total: 2 files changed+49-4
@@ -180,13 +180,14 @@ MatMulV3L0C2Out GetL0C2OutDav3510(const MatmulV3CompileInfo& compileInfo, const
180 if (!fixpipeBound || (compileInfo.aivNum != (compileInfo.aicNum * NUM_TWO)) || isSplitK) {180 if (!fixpipeBound || (compileInfo.aivNum != (compileInfo.aicNum * NUM_TWO)) || isSplitK) {
181 return MatMulV3L0C2Out::ON_THE_FLY;181 return MatMulV3L0C2Out::ON_THE_FLY;
182 }182 }
183+ if (args.cType == ge::DT_FLOAT) {
184+ runInfo.splitM = true;
185+ return MatMulV3L0C2Out::ND_FIXPIPE_1_2;
186+ }
atomgit-bot
atomgit-botatomgit-bot12 days ago

🟡 Medium Priority

改动行 183-186:新增 if (args.cType == ge::DT_FLOAT) { runInfo.splitM = true; return ND_FIXPIPE_1_2; },使 fp16/bf16 输入 + fp32 输出(dtype 支持表中 {FLOAT16,FLOAT16,FLOAT,...} / {BF16,BF16,FLOAT,...},属最常见组合)由原来的 ND_FIXPIPE_1_1 变为 ND_FIXPIPE_1_2。

证据链:

  1. GetL0C2Out 的结果经 MatMulV3BasicAswtTiling::DoOpTiling(matmul_v3_basic_aswt_tiling.cpp:341)进入 tiling key 的 L0C2OUT 字段(同文件 383 行)。
  2. CheckBL1FullLoad()(matmul_v3_basic_aswt_tiling.cpp:102-102 起)没有 l0C2Out_ != ON_THE_FLY 的拦截(该拦截只存在于 CheckAL1FullLoad(),见 68-71 行),因此满足 B 全载条件的算例会形成 (TENSOR_LEVEL, BASIC, B_FULL_LOAD, 1V2) 的 key。
  3. kernel 侧 dispatch 中可见的分支为:B_FULL_LOAD + BASIC + 1V1(mat_mul_v3.cpp:168-174,注释即“Fixpipe B全载fp16场景”)、NO_FULL_LOAD + BASIC + 1V1、NO_FULL_LOAD + SLICE + 1V1、NO_FULL_LOAD + BASIC + 1V2(190-193)以及 streamK/splitK+1V2;而 tiling key 的模板 SEL 又把 (B_FULL_LOAD, 1V2) 注册为合法组合(mat_mul_tiling_key.h:174-177)。

失败模式:若 kernel 的 if constexpr 链确实没有 (B_FULL_LOAD, BASIC, 1V2) 分支,改动后“fp16/bf16 输入 + fp32 输出 + 满足 B 全载”的算例会命中一个无 kernel 实现的 key,落到无分支/空实现,输出为未初始化数据(静默精度/正确性错误),而改动前该算例走的是已实现的 B全载+1V1 分支。触发条件明确但需要确认 dispatch 完整体(本次审查无法读取 op_kernel 目录全文)。

likedislike
Incorrect?
183 if (args.aType == ge::DT_FLOAT16 || args.aType == ge::DT_BF16) {187 if (args.aType == ge::DT_FLOAT16 || args.aType == ge::DT_BF16) {
184 return MatMulV3L0C2Out::ND_FIXPIPE_1_1;188 return MatMulV3L0C2Out::ND_FIXPIPE_1_1;
185 }189 }
186- if (args.aType == ge::DT_FLOAT) {190+ return MatMulV3L0C2Out::ON_THE_FLY;
187- runInfo.splitM = true;
188- }
189- return MatMulV3L0C2Out::ND_FIXPIPE_1_2;
190}191}
191 192 
192// 全 dtype 统一走 ND_FIXPIPE_1_1193// 全 dtype 统一走 ND_FIXPIPE_1_1
@@ -4721,6 +4721,50 @@ TEST_F(MatMulV3StreamKTilingDirectTest, GetL0C2Out_Non3510Arch_DefaultOnTheFly)
4721 MatMulV3L0C2Out::ON_THE_FLY);4721 MatMulV3L0C2Out::ON_THE_FLY);
4722}4722}
4723 4723 
4724+// GetL0C2OutDav3510:fixpipe Bound 场景按输出 dtype 选择 1V1/1V2,splitM 仅在 FP32 输出时置位
4725+TEST_F(MatMulV3StreamKTilingDirectTest, GetL0C2Out_Dav3510_Fp32Out_Fixpipe1V2WithSplitM)
4726+{
4727+ compileInfo_.npuArch = Ops::Base::DAV_3510;
4728+ SetShape(256UL, 300UL, 256UL);
4729+ args_.aType = ge::DT_FLOAT;
4730+ args_.bType = ge::DT_FLOAT;
4731+ args_.cType = ge::DT_FLOAT;
4732+ optiling::matmul_v3_advanced::MatMulV3RunInfo runInfo{};
4733+ runInfo.singleCoreM = 128UL;
4734+ runInfo.singleCoreN = 64UL;
4735+ EXPECT_EQ(
4736+ optiling::matmul_v3_advanced::MatMulV3TilingHelper::GetL0C2Out(compileInfo_, args_, runInfo, false, nullptr),
4737+ MatMulV3L0C2Out::ND_FIXPIPE_1_2);
4738+ EXPECT_TRUE(runInfo.splitM);
4739+}
4740+ 
4741+TEST_F(MatMulV3StreamKTilingDirectTest, GetL0C2Out_Dav3510_Fp16InFp32Out_Fixpipe1V2WithSplitM)
4742+{
4743+ compileInfo_.npuArch = Ops::Base::DAV_3510;
4744+ SetShape(256UL, 300UL, 256UL);
4745+ args_.cType = ge::DT_FLOAT;
4746+ optiling::matmul_v3_advanced::MatMulV3RunInfo runInfo{};
4747+ runInfo.singleCoreM = 128UL;
4748+ runInfo.singleCoreN = 64UL;
4749+ EXPECT_EQ(
4750+ optiling::matmul_v3_advanced::MatMulV3TilingHelper::GetL0C2Out(compileInfo_, args_, runInfo, false, nullptr),
4751+ MatMulV3L0C2Out::ND_FIXPIPE_1_2);
4752+ EXPECT_TRUE(runInfo.splitM);
4753+}
4754+ 
4755+TEST_F(MatMulV3StreamKTilingDirectTest, GetL0C2Out_Dav3510_Fp16Out_Fixpipe1V1WithoutSplitM)
4756+{
4757+ compileInfo_.npuArch = Ops::Base::DAV_3510;
4758+ SetShape(256UL, 300UL, 256UL);
4759+ optiling::matmul_v3_advanced::MatMulV3RunInfo runInfo{};
4760+ runInfo.singleCoreM = 128UL;
4761+ runInfo.singleCoreN = 64UL;
4762+ EXPECT_EQ(
4763+ optiling::matmul_v3_advanced::MatMulV3TilingHelper::GetL0C2Out(compileInfo_, args_, runInfo, false, nullptr),
4764+ MatMulV3L0C2Out::ND_FIXPIPE_1_1);
4765+ EXPECT_FALSE(runInfo.splitM);
4766+}
4767+ 
4724TEST_F(MatMulV3TilingAdvancedDirectTest, GetFormat_NdFormats)4768TEST_F(MatMulV3TilingAdvancedDirectTest, GetFormat_NdFormats)
4725{4769{
4726 gert::StorageShape a = {{2, 32}, {2, 32}};4770 gert::StorageShape a = {{2, 32}, {2, 32}};