ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-02 04:31:25 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
4c626aeaa6c4e4df27210ab21ed1b2a3290b0f5c
composable_kernel/include/ck_tile/ops/gemm/warp/warp_gemm.hpp
Enrico Degregori 4c626aeaa6 [rocm-libraries] ROCm/rocm-libraries#4267 (commit 3c5d95e) 
[CK_TILE] Extend support of mix precision microscaling BQuant
 (#4267)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

## Proposed changes

Supported types combinations using BQuant=e8m0:
 - A=bf16
 - B=bf16,bf8,fp4

Summary:
- remove usage of `pk_fp4_raw_t`: consistent with other implementations
and avoid taking into account of the packed size explicitly. In general,
the raw type should not be used because CK Tile internally takes care of
the PackedSize, so using the raw type adds unnecessary complexity to the
implementation
- handle microscaling by checking for `e8m0` type for BQuant (previous
implementation was inconsistent)
 - add support for scaling instructions in `DequantPack8`
 - mx pipeline:
   - extend existing pipeline to support different B types
- add support to scale and cast before writing to LDS or after reading
from LDS (this can be defined in the `Problem` by the user)
 - block gemm:
   - mx pipeline is now using block gemm BQuant
- block gemm BQuant can now load from LDS and apply scale and then call
block gemm universal operator. This adds new functionalities and remove
code duplication
 - warp gemm:
- add case to support 128bit ds_read/write for both A and B when A=16bit
and B=8bit
- add examples and tests: note that some tests for bf16/fp4 already
existed but were removed during previous tests refactoring. I added them
again and other relevant tests for new types combinations

## Checklist

Please put an `x` into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.

- [ ] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers
with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [ ] I have run `clang-format` on all changed files
- [ ] Any dependent changes have been merged

## Discussion

If this is a relatively large or complex change, feel free to start a
discussion by explaining why you chose the solution you did and what
alternatives you considered
2026-02-24 17:57:02 +00:00

437 lines
19 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 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
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_fp8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>,
2>>;
using WarpGemmMfma_f32_32x32x32_bf8_bf8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>,
2>>;
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_fp8_fp8 = WarpGemmImpl<WarpGemmAttributeMfmaIterateK<
WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_fp8<WGAttrCtlEnum::Default_>,
2>>;
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_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>>;
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