昇腾论坛

怀疑是高阶配置有啥设置错了，当前开发部分代码如下，打印结果不如预期：

using ATypeQK = MatmulType<AscendC::TPosition::GM, CubeFormat::ND, bfloat16_t, false>;

using BTypeQK = MatmulType<AscendC::TPosition::GM, CubeFormat::ND, bfloat16_t, true>;

using CTypeQK = MatmulType<AscendC::TPosition::CO1, CubeFormat::NZ, float>;

using ATypeQH = MatmulType<AscendC::TPosition::GM, CubeFormat::ND, bfloat16_t, false>;

using BTypeQH = MatmulType<AscendC::TPosition::GM, CubeFormat::ND, bfloat16_t, false>;

using CTypeQH = MatmulType<AscendC::TPosition::GM, CubeFormat::ND_ALIGN, half>;

using ATypeAV = MatmulType<AscendC::TPosition::GM, CubeFormat::ND, bfloat16_t, false>;

using BTypeAV = MatmulType<AscendC::TPosition::GM, CubeFormat::ND, bfloat16_t, false>;

using CTypeAV = MatmulType<AscendC::TPosition::GM, CubeFormat::ND_ALIGN, half>;

using BIAS_TYPE = matmul::MatmulType<TPosition::GM, CubeFormat::ND, bfloat16_t>;

printf("-----------------C2-----------------\n");

printf("-----------------C2输入 q矩阵-----------------\n");

DumpTensor(gmQTensor[cube23Off.qkGmBase + cube23Off.qkHeadInner], 000, (uint32_t)200);

printf("-----------------C2输入 h矩阵-----------------\n");

DumpTensor(gmHTensor[cube23Off.hOffset], 000, (uint32_t)200);

mmQH.SetOrgShape(cube23Off.blockTokens, vHeadDim_m, kNumHead_m * kHeadDim_m, kHeadDim_m, 0);

mmQH.SetSingleShape(cube23Off.blockTokens, vHeadDim_m, kHeadDim_m);

mmQH.SetTensorA(gmQTensor[cube23Off.qkGmBase + cube23Off.qkHeadInner]);

mmQH.SetTensorB(gmHTensor[cube23Off.hOffset]);

mmQH.IterateAll(gmHWork[cube23Off.hvWorkOffset]);

mmQH.End();

printf("-----------------C2结果-----------------\n");

DumpTensor(gmHWork[cube23Off.hvWorkOffset], 000, (uint32_t)200);

预期代码的matmul配置：