ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
2312eef6c36d2811b1f57c85c8ae4c58a595be9e
composable_kernel/include/ck_tile/ops/gemm/warp/warp_gemm.hpp
Anton Gorenko 2312eef6c3 [rocm-libraries] ROCm/rocm-libraries#4368 (commit 17f7dfc) 
[CK_TILE][FMHA] Support microscaling (mxfp8 and mxfp4) on
 gfx950 (#4368)

## Motivation

Microscaling types (mxfp8 and mxfp4) for fwd qr pipeline

## Technical Details

The microscaling is used when quant scale mode is
`BlockAttentionQuantScaleEnum::MX` and `Q/K/P/VDataType` are
fp8/bf8/fp4.

Supported features:
* only "qr" pipeline is implemented
* hdim 128 and 256 (smaller hdim are not possible due to restrictions of
"qr" pipeline, but they can be computed using instances with padding)
 * both 32x32x64 and 16x16x128 scale MFMAs are supported
 * Q and K scales are applied in hdim, V scales - in seqlen dimension
 * column-major V only
 * batch and group mode
 * bias, Alibi (tested but no instances by default, just like fp8)
 * masking etc.

Aiter PR with new API args: https://github.com/ROCm/aiter/pull/2008

## Test Plan

```
ninja test_ck_tile_fmha_fwd_mxfp8 && bin/test_ck_tile_fmha_fwd_mxfp8
ninja test_ck_tile_fmha_fwd_mxfp4 && bin/test_ck_tile_fmha_fwd_mxfp4
```

## Test Result

The tests must pass.

## Submission Checklist

- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
2026-03-11 10:00:52 +00:00

505 lines
22 KiB
C++
Raw Blame History

// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_impl.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_smfmac_impl.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_attribute_smfmac.hpp"
namespace ck_tile {
// fp32
using WarpGemmMfmaF32F32F32M16N16K4 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplF32F32F32M16N16K4<WGAttrCtlEnum::Default_>>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF32F32F32M16N16K16 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplF32F32F32M16N16K4<WGAttrCtlEnum::Default_>,
4,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF32F32F32M16N16K8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplF32F32F32M16N16K4<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF32F32F32M32N32K8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplF32F32F32M32N32K2<WGAttrCtlEnum::Default_>,
4,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF32F32F32M32N32K4 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplF32F32F32M32N32K2<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF32F32F32M16N16K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplF32F32F32M16N16K4<WGAttrCtlEnum::Default_>,
4,
AttrNumAccess>>;
// fp16
using WarpGemmMfmaF16F16F32M32N32K8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M16N16K16 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF16F16F32M32N32K16 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K16<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF16F16F32M32N32K16 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#endif
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF16F16F32M16N16K32 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K32<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF16F16F32M16N16K32 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#endif
using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAttributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
1>>;
using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl<WarpGemmAttributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M32N32K16<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#endif
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M16N16K32<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#endif
#if defined(__gfx950__)
using WarpGemmMfmaF16F16F32M16N16K32SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplF16F16F32M16N16K32<WGAttrCtlEnum::Default_>,
1>>;
using WarpGemmMfmaBf16Bf16F32M16N16K32SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K32<WGAttrCtlEnum::Default_>,
1>>;
#endif
using WarpGemmMfmaF16F16F32M32N32K8SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
#if defined(__gfx950__)
using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplF16F16F32M32N32K16<WGAttrCtlEnum::Default_>>>;
#else
using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
#endif
using WarpGemmMfmaF16F16F32M4N64K16 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplF16F16F32M4N64K4<WGAttrCtlEnum::Default_>,
4>>;
using WarpGemmMfmaF16F16F32M64N4K16 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplF16F16F32M64N4K4<WGAttrCtlEnum::Default_>,
4>>;
// fp16 2:4 structured sparsity
using WarpGemmSmfmacF16F16F32M32N32K16 = WarpGemmSmfmacImpl<WarpGemmAttributeSmfmac<
WarpGemmAttributeSmfmacImplF16F16F32M32N32K16<WGAttrCtlEnum::Default_>>>;
using WarpGemmSmfmacF16F16F32M16N16K32 = WarpGemmSmfmacImpl<WarpGemmAttributeSmfmac<
WarpGemmAttributeSmfmacImplF16F16F32M16N16K32<WGAttrCtlEnum::Default_>>>;
// bf16
using WarpGemmMfmaBf16Bf16F32M32N32K8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaBf16Bf16F32M32N32K16 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K16<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaBf16Bf16F32M32N32K16 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#endif
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccessA = WGAttrNumAccessEnum::Single,
WGAttrNumAccessEnum AttrNumAccessB = AttrNumAccessA>
using WarpGemmMfmaBf16Bf16F32M16N16K32 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K32<WGAttrCtlEnum::Default_>,
AttrNumAccessA,
AttrNumAccessB>>;
template <WGAttrNumAccessEnum AttrNumAccessA = WGAttrNumAccessEnum::Single,
WGAttrNumAccessEnum AttrNumAccessB = AttrNumAccessA>
using WarpGemmMfmaBf16Bf16F32M16N16K64 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K32<WGAttrCtlEnum::Default_>,
2,
AttrNumAccessA,
AttrNumAccessB>>;
#else
template <WGAttrNumAccessEnum AttrNumAccessA = WGAttrNumAccessEnum::Single,
WGAttrNumAccessEnum AttrNumAccessB = AttrNumAccessA>
using WarpGemmMfmaBf16Bf16F32M16N16K32 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
2,
AttrNumAccessA>>;
template <WGAttrNumAccessEnum AttrNumAccessA = WGAttrNumAccessEnum::Single,
WGAttrNumAccessEnum AttrNumAccessB = AttrNumAccessA>
using WarpGemmMfmaBf16Bf16F32M16N16K64 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
4,
AttrNumAccessA,
AttrNumAccessB>>;
#endif
using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAttributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
1>>;
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA =
WarpGemmImpl<WarpGemmAttributeMfmaIterateK_SwizzleA<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K16<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#endif
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K32<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#endif
using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
#if defined(__gfx950__)
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K16<WGAttrCtlEnum::Default_>>>;
#else
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
2>>;
#endif
using WarpGemmMfmaBf16Bf16F32M4N64K16 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M4N64K4<WGAttrCtlEnum::Default_>,
4>>;
using WarpGemmMfmaBf16Bf16F32M64N4K16 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImplBf16Bf16F32M64N4K4<WGAttrCtlEnum::Default_>,
4>>;
// fp8
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x32_fp8_fp8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x32_fp8_fp8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>,
2>>;
#endif
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_16x16x64_fp8_fp8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_fp8<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_16x16x64_fp8_fp8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_fp8<WGAttrCtlEnum::Default_>,
2>>;
#endif
#if defined(__gfx950__)
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x32_bf8_bf8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>,
2,
AttrNumAccess>>;
#else
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x32_bf8_bf8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>,
2>>;
#endif
using WarpGemmMfma_f32_32x32x16_fp8_fp8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_16x16x32_fp8_bf8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x32_fp8_bf8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfma_f32_16x16x32_fp8_fp8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_16x16x32_fp8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_16x16x32_bf8_bf8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_16x16x32_bf8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_16x16x32_bf8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x32_bf8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_16x16x64_bf8_bf8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_16x16x32_bf8_bf8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfma_f32_16x16x64_fp8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_fp8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfma_f32_16x16x64_bf8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x32_bf8_bf8<WGAttrCtlEnum::Default_>,
2>>;
template <typename A, typename B, WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_16x16x128_f8f6f4 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_16x16x128_f8f6f4<A, B>, AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_16x16x128_fp8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x128_f8f6f4<fp8_t, fp8_t>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_16x16x128_fp8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x128_f8f6f4<fp8_t, bf8_t>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_16x16x128_bf8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x128_f8f6f4<bf8_t, fp8_t>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_16x16x128_bf8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x128_f8f6f4<bf8_t, bf8_t>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_16x16x128_fp4_fp4_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_16x16x128_f8f6f4<pk_fp4_t, pk_fp4_t>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_fp8_fp8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_fp8_fp8<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_fp8_bf8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_fp8_bf8<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_bf8_fp8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_bf8_fp8<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_bf8_bf8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_bf8_bf8<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_fp8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x64_fp8_fp8<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_fp8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x64_fp8_bf8<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_bf8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x64_bf8_fp8<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_bf8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x64_bf8_bf8<WGAttrCtlEnum::Default_>,
AttrNumAccess>>;
template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
using WarpGemmMfma_f32_32x32x64_fp4_fp4_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x64_f8f6f4<pk_fp4_t, pk_fp4_t>,
AttrNumAccess>>;
using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
template <index_t swizzle_factor = 2>
using WarpGemmMfmaFp8Fp8F32M32N32K32SwizzleBTransposedCDistribution =
WarpGemmImpl<WarpGemmAttributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>,
2,
swizzle_factor>>;
// int8
using WarpGemmMfma_i32_32x32x16_i8_i8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_i32_32x32x16_i8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_i32_32x32x16_i8_i8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_i32_32x32x16_i8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_i32_16x16x32_i8_i8 = WarpGemmImpl<
WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_i32_16x16x32_i8<WGAttrCtlEnum::Default_>>>;
using WarpGemmMfma_i32_16x16x32_i8_i8_CTransposed =
WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
WarpGemmAttributeMfmaImpl_i32_16x16x32_i8<WGAttrCtlEnum::Default_>>>;
} // namespace ck_tile
Reference in New Issue View Git Blame Copy Permalink