mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 05:01:25 +00:00
e03293ebce
* updates to support int8 in 03_gemm example * added comments, using aliases, helper functions * test(gemm_universal): add test cases for int8 gemm pipeline * fix(test_gemm): fix for failing test unit test for int8 * test(ck_tile): add int8 unit test for gemm universal * refactor(gemm_universal): GPU reference verification for GEMM code improved * style(gemm_universal): removed extra comments and did clang format * merging recent changes to universal gemm to tile_engine * ck tile engine integration work * feat(tile_engine): add int8 support to tile engine ops/gemm * feat(tile_engine): added 32 32 16 mfma instances to tile engine for int8 * style: Format code with clang-format-12 * refactor(tile_engine): address review comments * style: removed unhelpful comments & unused variables. * build: tile engine uses default config * feat: add int8 support for CK_TILE GEMM * style: added trailing commas to codegen_utils.py * refactor: tile engine * refactor: formatting and code review * refactor: code formatting for python files * fix: suppress build warning * add support for gfx950 * refactor:KWarpTile size in gemms util * Fix the branch and wrap up the k warp tile * Add bf8 integration * refactor: clang format and rebase --------- Co-authored-by: zjli2013 <leezhengjiang@gmail.com> Co-authored-by: AviralGoelAMD <aviral.goel@amd.com> Co-authored-by: Khushbu Agarwal <khuagarw@amd.com>
301 lines
13 KiB
C++
301 lines
13 KiB
C++
// SPDX-License-Identifier: MIT
|
|
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
#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 {
|
|
|
|
// fp16
|
|
|
|
using WarpGemmMfmaF16F16F32M32N32K8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfmaF16F16F32M16N16K16 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaF16F16F32M32N32K16 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K16<WGAttrCtlEnum::Default_>>>;
|
|
|
|
#else
|
|
using WarpGemmMfmaF16F16F32M32N32K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaF16F16F32M16N16K32 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K32<WGAttrCtlEnum::Default_>>>;
|
|
#else
|
|
using WarpGemmMfmaF16F16F32M16N16K32 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
|
|
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
1>>;
|
|
|
|
using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
|
|
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
|
|
using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplF16F16F32M32N32K16<WGAttrCtlEnum::Default_>>>;
|
|
#else
|
|
using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplF16F16F32M16N16K32<WGAttrCtlEnum::Default_>>>;
|
|
#else
|
|
using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaF16F16F32M16N16K32SwizzleBTransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB<
|
|
WarpGemmAttributeMfmaImplF16F16F32M16N16K32<WGAttrCtlEnum::Default_>,
|
|
1>>;
|
|
|
|
using WarpGemmMfmaBf16Bf16F32M16N16K32SwizzleBTransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K32<WGAttrCtlEnum::Default_>,
|
|
1>>;
|
|
#endif
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB<
|
|
WarpGemmAttributeMfmaImplF16F16F32M32N32K16<WGAttrCtlEnum::Default_>>>;
|
|
#else
|
|
using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
|
|
WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
using WarpGemmMfmaF16F16F32M4N64K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImplF16F16F32M4N64K4<WGAttrCtlEnum::Default_>,
|
|
4>>;
|
|
|
|
using WarpGemmMfmaF16F16F32M64N4K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
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<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfmaBf16Bf16F32M16N16K16 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K16 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K16<WGAttrCtlEnum::Default_>>>;
|
|
|
|
#else
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaBf16Bf16F32M16N16K32 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K32<WGAttrCtlEnum::Default_>>>;
|
|
#else
|
|
using WarpGemmMfmaBf16Bf16F32M16N16K32 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
1>>;
|
|
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K16<WGAttrCtlEnum::Default_>>>;
|
|
#else
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K32<WGAttrCtlEnum::Default_>>>;
|
|
#else
|
|
using WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
#if defined(__gfx950__)
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
|
|
#else
|
|
using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
#endif
|
|
|
|
using WarpGemmMfmaBf16Bf16F32M4N64K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M4N64K4<WGAttrCtlEnum::Default_>,
|
|
4>>;
|
|
|
|
using WarpGemmMfmaBf16Bf16F32M64N4K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImplBf16Bf16F32M64N4K4<WGAttrCtlEnum::Default_>,
|
|
4>>;
|
|
|
|
// fp8
|
|
|
|
using WarpGemmMfma_f32_32x32x16_fp8_fp8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x16_fp8_bf8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x16_bf8_fp8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x16_bf8_bf8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x32_fp8_fp8 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x32_bf8_bf8 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
|
|
using WarpGemmMfma_f32_16x16x32_fp8_fp8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_fp8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_16x16x32_bf8_bf8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_16x16x32_bf8_bf8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_16x16x64_fp8_fp8 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImpl_f32_16x16x32_fp8_fp8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
|
|
using WarpGemmMfma_f32_16x16x64_bf8_bf8 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
|
|
WarpGemmAttributeMfmaImpl_f32_16x16x32_bf8_bf8<WGAttrCtlEnum::Default_>,
|
|
2>>;
|
|
|
|
using WarpGemmMfma_f32_16x16x128_fp8_fp8 = WarpGemmImpl<WarpGemmAtrributeMfma<
|
|
WarpGemmAttributeMfmaImpl_f32_16x16x128_fp8_fp8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_16x16x128_fp8_bf8 = WarpGemmImpl<WarpGemmAtrributeMfma<
|
|
WarpGemmAttributeMfmaImpl_f32_16x16x128_fp8_bf8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_16x16x128_bf8_fp8 = WarpGemmImpl<WarpGemmAtrributeMfma<
|
|
WarpGemmAttributeMfmaImpl_f32_16x16x128_bf8_fp8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_16x16x128_bf8_bf8 = WarpGemmImpl<WarpGemmAtrributeMfma<
|
|
WarpGemmAttributeMfmaImpl_f32_16x16x128_bf8_bf8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x64_fp8_fp8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_fp8_fp8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x64_fp8_bf8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_fp8_bf8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x64_bf8_fp8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_bf8_fp8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x64_bf8_bf8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_bf8_bf8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
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, WGAttrCtlEnum::Default_>,
|
|
2,
|
|
swizzle_factor>>;
|
|
|
|
// int8
|
|
using WarpGemmMfma_i32_32x32x16_i8_i8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_i32_32x32x16_i8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_i32_32x32x16_i8_i8_CTransposed =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImpl_i32_32x32x16_i8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_i32_16x16x32_i8_i8 = WarpGemmImpl<
|
|
WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_i32_16x16x32_i8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
using WarpGemmMfma_i32_16x16x32_i8_i8_CTransposed =
|
|
WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
|
|
WarpGemmAttributeMfmaImpl_i32_16x16x32_i8<WGAttrCtlEnum::Default_>>>;
|
|
|
|
} // namespace ck_tile
|