[CK_TILE] fused-moe first version (#1634)

* moe pipeline

* update code

* compile OK

* update

* update cpu reference

* update pipeline_gemm0

* compiler ok

* update pipeline

* rename to ex pipeline

* block-asm

* update

* update

* update first gemm ok

* compute correct

* update file structure

* update README

* update

* update

* update code

* update API

* return unsupport case

* add comment

* update readme

* update

* uncomment

* update

* fix build err

---------

Co-authored-by: valarLip <340077269@qq.com>
This commit is contained in:
carlushuang
2024-11-26 11:14:56 +08:00
committed by GitHub
parent 645fe812f6
commit 440e28b08f
66 changed files with 8066 additions and 308 deletions

View File

@@ -10,114 +10,134 @@
namespace ck_tile {
// fp16
using WarpGemmMfmaF16F16F32M32N32K8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8>>;
using WarpGemmMfmaF16F16F32M16N16K16 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16>>;
using WarpGemmMfmaF16F16F32M32N32K8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M32N32K16 =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 2>>;
using WarpGemmMfmaF16F16F32M16N16K16 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M16N16K32 =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplF16F16F32M16N16K16, 2>>;
using WarpGemmMfmaF16F16F32M32N32K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl<
WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 1>>;
using WarpGemmMfmaF16F16F32M16N16K32 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl<
WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 2>>;
using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
1>>;
using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplF16F16F32M32N32K8>>;
using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplF16F16F32M16N16K16>>;
using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8,
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16,
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8,
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
// bf16
using WarpGemmMfmaBf16Bf16F32M32N32K8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16 =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 2>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M16N16K32 =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16, 2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl<
WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 1>>;
using WarpGemmMfmaBf16Bf16F32M16N16K32 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA = WarpGemmImpl<
WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
1>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8,
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16,
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8,
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
// fp8
using WarpGemmMfma_f32_32x32x16_fp8_fp8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8>>;
using WarpGemmMfma_f32_32x32x16_fp8_fp8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8 =
WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8 = WarpGemmImpl<
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8>>;
using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed = WarpGemmImpl<
WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
template <index_t swizzle_factor = 2>
using WarpGemmMfmaFp8Fp8F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t>,
WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t, WGAttrCtlEnum::Default_>,
2,
swizzle_factor>>;