mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 05:01:25 +00:00
1441a0a7ee
* something khushbu can help with * v1 v2 works with flatmm develop * v0 v1 v2 numerical error gone * Fixing numerical error, and interchange preshuffle configs to match with flatmm * Refactor GEMM pipeline configurations and integrate preshuffle support - Updated preshuffle pipeline definitions to include multiple versions (V1, V2, V3). - Changed the pipeline constant from CK_TILE_PIPELINE_PRESHUFFLE to CK_TILE_PIPELINE_PRESHUFFLE_V3 in relevant configurations. - Removed obsolete code and comments * clang format * fix vectorloadsize bug * add the Preshuffle3 * update kwarp calculation in gemm utils * update vector size A and B correctly in V2 pipeline; Added few more changes to align with dteng's branch * fix: add CK_GFX950_SUPPORT macro for gfx950 detection * default disable rotating buffer * docs(CHANGELOG): update changelog for rocm 7.0 * Revert "docs(CHANGELOG): update changelog for rocm 7.0" This reverts commit2bc16fff84. * Remove unused Preshuffle V3 pipeline and related code; update gemm function to use Preshuffle V2; clean up comments and formatting in various files. * revert example/ck_tile/flatmm to its original state * remove comment added by second author * switch to xor ALDSDescriptor * modify the MakeALdsDescriptor() * temporary profiling script * getting rid of line marker compiler error * UniversalWeightPreshufflePipelineAgBgCrPolicy now derives from UniversalGemmBasePolicy * add a minor fix for the config * typo fix * Fix formatting in lambda function for WeightPreshufflePipelineAGmemBGmemCRegV2 * revert change in include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1.hpp * revert change in include/ck_tile/core/arch/amd_buffer_addressing.hpp * reenable the GemmSpatiallyLocalTilePartitioner * make GemmConfigPreshuffle_1 for v1 pipeline, GemmConfigPreshuffle_2 for v2 pipeline * remove hardcoded true for preshuffle bool template argument * rename script * remove gemm_profilie.sh script * merge conflict resolve * clang formatted * typo fix * Remove duplicate include of block_gemm_areg_bsmem_creg_v2r1.hpp in gemm.hpp * Remove commented-out code in UniversalWeightPreshufflePipelineAgBgCrPolicy * Fix missing newline at end of file in run_gemm_example.inc * Remove unused barrier call in BlockWeightPreshuffleASmemBSmemCRegV1 * addressing review comments * removing debug code * addressing review comments * Revert "addressing review comments" This reverts commit29c45192ba. * updating tile_engine code * addressing review comments --------- Co-authored-by: amd-khushbu <khuagarw@amd.com> Co-authored-by: ThomasNing <thomas.ning@amd.com>
492 lines
17 KiB
C++
492 lines
17 KiB
C++
// SPDX-License-Identifier: MIT
|
|
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
#pragma once
|
|
|
|
#include <string>
|
|
|
|
#include "ck_tile/core.hpp"
|
|
#include "ck_tile/host/kernel_launch.hpp"
|
|
#include "ck_tile/ops/epilogue.hpp"
|
|
#include "ck_tile/ops/gemm.hpp"
|
|
|
|
#define CK_TILE_PIPELINE_COMPUTE_V3 1
|
|
#define CK_TILE_PIPELINE_MEMORY 2
|
|
#define CK_TILE_PIPELINE_COMPUTE_V4 3
|
|
#define CK_TILE_PIPELINE_COMPUTE_V5 4
|
|
#define CK_TILE_PIPELINE_PRESHUFFLE_V1 5
|
|
#define CK_TILE_PIPELINE_PRESHUFFLE_V2 6
|
|
|
|
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
|
|
constexpr ck_tile::index_t get_k_warp_tile()
|
|
{
|
|
#if defined(CK_GFX950_SUPPORT)
|
|
constexpr bool is_8bit_float =
|
|
std::is_same_v<PrecType, ck_tile::fp8_t> || std::is_same_v<PrecType, ck_tile::bf8_t>;
|
|
if constexpr(M_Warp_Tile == 32)
|
|
return is_8bit_float ? 64 : 16;
|
|
else
|
|
return is_8bit_float ? 128 : 32;
|
|
#else
|
|
if constexpr(M_Warp_Tile == 32)
|
|
return 16;
|
|
else
|
|
return 32;
|
|
#endif
|
|
}
|
|
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
|
|
constexpr ck_tile::index_t get_k_warp_tile_flatmm()
|
|
{
|
|
#if defined(CK_GFX950_SUPPORT)
|
|
if constexpr(M_Warp_Tile == 32)
|
|
return sizeof(PrecType) == 2 ? 16 : 64;
|
|
else
|
|
return sizeof(PrecType) == 2 ? 32 : 128;
|
|
#else
|
|
if constexpr(M_Warp_Tile == 32)
|
|
return sizeof(PrecType) == 2 ? 16 : 32;
|
|
else
|
|
return sizeof(PrecType) == 2 ? 32 : 64;
|
|
#endif
|
|
}
|
|
|
|
struct GemmConfigBase
|
|
{
|
|
static constexpr bool kPadM = false;
|
|
static constexpr bool kPadN = false;
|
|
static constexpr bool kPadK = false;
|
|
|
|
static constexpr bool PermuteA = false;
|
|
static constexpr bool PermuteB = false;
|
|
|
|
static constexpr bool TransposeC = false;
|
|
static constexpr bool UseStructuredSparsity = false;
|
|
|
|
static constexpr int kBlockPerCu = 1;
|
|
static constexpr ck_tile::index_t TileParitionerGroupNum = 8;
|
|
static constexpr ck_tile::index_t TileParitionerM01 = 4;
|
|
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Intrawave;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V3;
|
|
static constexpr ck_tile::index_t NumWaveGroups = 1;
|
|
static constexpr bool Preshuffle = false;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigMemoryInterwave : public GemmConfigBase
|
|
{
|
|
// Memory friendly for Interwave scheduler
|
|
static constexpr ck_tile::index_t M_Tile = 128;
|
|
static constexpr ck_tile::index_t N_Tile = 32;
|
|
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 4;
|
|
static constexpr ck_tile::index_t N_Warp = 1;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(PrecType) == 2 ? 8 : 16;
|
|
|
|
static constexpr bool DoubleSmemBuffer = false;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_MEMORY;
|
|
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Interwave;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigMemoryIntrawave : public GemmConfigBase
|
|
{
|
|
static constexpr ck_tile::index_t M_Tile = 128;
|
|
static constexpr ck_tile::index_t N_Tile = 32;
|
|
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 4;
|
|
static constexpr ck_tile::index_t N_Warp = 1;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(PrecType) == 2 ? 8 : 16;
|
|
|
|
static constexpr bool DoubleSmemBuffer = false;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_MEMORY;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigComputeV3 : public GemmConfigBase
|
|
{
|
|
// Compute V3 only support Intrawave scheduler
|
|
static constexpr ck_tile::index_t M_Tile = 16;
|
|
static constexpr ck_tile::index_t N_Tile = 64;
|
|
static constexpr ck_tile::index_t K_Tile = 256 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 1;
|
|
static constexpr ck_tile::index_t N_Warp = 4;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
|
|
|
|
static constexpr bool DoubleSmemBuffer = false;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V3;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigComputeV3_1 : public GemmConfigBase
|
|
{
|
|
static constexpr ck_tile::index_t M_Tile = 256;
|
|
static constexpr ck_tile::index_t N_Tile = 256;
|
|
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 2;
|
|
static constexpr ck_tile::index_t N_Warp = 2;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
|
|
|
|
static constexpr bool DoubleSmemBuffer = false;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V3;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigComputeV3_2 : public GemmConfigBase
|
|
{
|
|
static constexpr ck_tile::index_t M_Tile = 128;
|
|
static constexpr ck_tile::index_t N_Tile = 128;
|
|
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 2;
|
|
static constexpr ck_tile::index_t N_Warp = 2;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
|
|
|
|
static constexpr bool DoubleSmemBuffer = false;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V3;
|
|
|
|
static constexpr int kBlockPerCu = 2;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigComputeV4 : public GemmConfigBase
|
|
{
|
|
// Compute V4 only support Intrawave scheduler
|
|
// Using the ping pong reader in the lds level
|
|
static constexpr ck_tile::index_t M_Tile = 256;
|
|
static constexpr ck_tile::index_t N_Tile = 256;
|
|
static constexpr ck_tile::index_t K_Tile = 64 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 2;
|
|
static constexpr ck_tile::index_t N_Warp = 2;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
|
|
|
|
static constexpr bool DoubleSmemBuffer = true;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V4;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigComputeV4_1 : public GemmConfigBase
|
|
{
|
|
static constexpr ck_tile::index_t M_Tile = 256;
|
|
static constexpr ck_tile::index_t N_Tile = 256;
|
|
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 2;
|
|
static constexpr ck_tile::index_t N_Warp = 2;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
|
|
|
|
static constexpr bool DoubleSmemBuffer = true;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V4;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigComputeV5 : public GemmConfigBase
|
|
{
|
|
static constexpr ck_tile::index_t M_Tile = 128;
|
|
static constexpr ck_tile::index_t N_Tile = 128;
|
|
static constexpr ck_tile::index_t K_Tile = 64 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 1;
|
|
static constexpr ck_tile::index_t N_Warp = 1;
|
|
static constexpr ck_tile::index_t K_Warp = 2;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
|
|
|
|
static constexpr bool DoubleSmemBuffer = false;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V5;
|
|
static constexpr ck_tile::index_t NumWaNumWaveGroups = 2;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigPreshuffle_1 : public GemmConfigBase
|
|
{
|
|
static constexpr ck_tile::index_t M_Tile = 128;
|
|
static constexpr ck_tile::index_t N_Tile = 128;
|
|
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 1;
|
|
static constexpr ck_tile::index_t N_Warp = 4;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile_flatmm<PrecType, M_Warp_Tile>();
|
|
|
|
static constexpr int kBlockPerCu = 2;
|
|
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_PRESHUFFLE_V1;
|
|
static constexpr bool Preshuffle = true;
|
|
static constexpr bool DoubleSmemBuffer = false;
|
|
};
|
|
|
|
template <typename PrecType>
|
|
struct GemmConfigPreshuffle_2 : public GemmConfigBase
|
|
{
|
|
static constexpr ck_tile::index_t M_Tile = 128;
|
|
static constexpr ck_tile::index_t N_Tile = 128;
|
|
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
|
|
|
|
static constexpr ck_tile::index_t M_Warp = 1;
|
|
static constexpr ck_tile::index_t N_Warp = 4;
|
|
static constexpr ck_tile::index_t K_Warp = 1;
|
|
|
|
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
|
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
|
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile_flatmm<PrecType, M_Warp_Tile>();
|
|
|
|
static constexpr int kBlockPerCu = 2;
|
|
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
|
|
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_PRESHUFFLE_V2;
|
|
static constexpr bool Preshuffle = true;
|
|
static constexpr bool DoubleSmemBuffer = true;
|
|
};
|
|
|
|
template <typename ADataType, typename BDataType = ADataType, typename CDataType = ADataType>
|
|
struct GemmTypeConfig;
|
|
|
|
template <>
|
|
struct GemmTypeConfig<ck_tile::half_t>
|
|
{
|
|
using ADataType = ck_tile::half_t;
|
|
using BDataType = ck_tile::half_t;
|
|
using AccDataType = float;
|
|
using CDataType = ck_tile::half_t;
|
|
// ToDo: Add more bias config to support different categories of GEMM.
|
|
};
|
|
|
|
template <>
|
|
struct GemmTypeConfig<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t>
|
|
{
|
|
using ADataType = ck_tile::bf16_t;
|
|
using BDataType = ck_tile::bf16_t;
|
|
using AccDataType = float;
|
|
using CDataType = ck_tile::bf16_t;
|
|
};
|
|
|
|
template <>
|
|
struct GemmTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>
|
|
{
|
|
using ADataType = ck_tile::fp8_t;
|
|
using BDataType = ck_tile::fp8_t;
|
|
using AccDataType = float;
|
|
using CDataType = ck_tile::half_t;
|
|
};
|
|
|
|
template <>
|
|
struct GemmTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t>
|
|
{
|
|
using ADataType = ck_tile::bf8_t;
|
|
using BDataType = ck_tile::bf8_t;
|
|
using AccDataType = float;
|
|
using CDataType = ck_tile::half_t;
|
|
};
|
|
|
|
template <>
|
|
struct GemmTypeConfig<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>
|
|
{
|
|
using ADataType = ck_tile::half_t;
|
|
using BDataType = ck_tile::pk_int4_t;
|
|
using AccDataType = float;
|
|
using CDataType = ck_tile::half_t;
|
|
};
|
|
|
|
template <>
|
|
struct GemmTypeConfig<ck_tile::int8_t, ck_tile::int8_t, int32_t>
|
|
{
|
|
using ADataType = ck_tile::int8_t;
|
|
using BDataType = ck_tile::int8_t;
|
|
using AccDataType = int32_t;
|
|
using CDataType = int32_t;
|
|
};
|
|
|
|
template <typename T>
|
|
struct DataTypeTraits;
|
|
|
|
template <>
|
|
struct DataTypeTraits<float>
|
|
{
|
|
static constexpr const char* name = "fp32";
|
|
};
|
|
|
|
template <>
|
|
struct DataTypeTraits<double>
|
|
{
|
|
static constexpr const char* name = "fp64";
|
|
};
|
|
|
|
template <>
|
|
struct DataTypeTraits<int32_t>
|
|
{
|
|
static constexpr const char* name = "int32";
|
|
};
|
|
|
|
template <>
|
|
struct DataTypeTraits<ck_tile::half_t>
|
|
{
|
|
static constexpr const char* name = "fp16";
|
|
};
|
|
|
|
template <>
|
|
struct DataTypeTraits<ck_tile::bf16_t>
|
|
{
|
|
static constexpr const char* name = "bf16";
|
|
};
|
|
|
|
template <>
|
|
struct DataTypeTraits<ck_tile::fp8_t>
|
|
{
|
|
static constexpr const char* name = "fp8";
|
|
};
|
|
|
|
template <>
|
|
struct DataTypeTraits<ck_tile::bf8_t>
|
|
{
|
|
static constexpr const char* name = "bf8";
|
|
};
|
|
|
|
template <>
|
|
struct DataTypeTraits<ck_tile::pk_int4_t>
|
|
{
|
|
static constexpr const char* name = "pk_int4_t";
|
|
};
|
|
|
|
template <>
|
|
struct DataTypeTraits<ck_tile::int8_t>
|
|
{
|
|
static constexpr const char* name = "int8";
|
|
};
|
|
|
|
template <ck_tile::index_t PipelineId>
|
|
struct PipelineTypeTraits;
|
|
|
|
template <>
|
|
struct PipelineTypeTraits<CK_TILE_PIPELINE_MEMORY>
|
|
{
|
|
template <typename PipelineProblem>
|
|
using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<PipelineProblem>;
|
|
template <typename PipelineProblem>
|
|
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<PipelineProblem>;
|
|
};
|
|
|
|
template <>
|
|
struct PipelineTypeTraits<CK_TILE_PIPELINE_COMPUTE_V3>
|
|
{
|
|
template <typename PipelineProblem>
|
|
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV3<PipelineProblem>;
|
|
template <typename PipelineProblem>
|
|
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV3<PipelineProblem>;
|
|
};
|
|
|
|
template <>
|
|
struct PipelineTypeTraits<CK_TILE_PIPELINE_COMPUTE_V4>
|
|
{
|
|
template <typename PipelineProblem>
|
|
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV4<PipelineProblem>;
|
|
template <typename PipelineProblem>
|
|
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV4<PipelineProblem>;
|
|
};
|
|
|
|
template <>
|
|
struct PipelineTypeTraits<CK_TILE_PIPELINE_COMPUTE_V5>
|
|
{
|
|
template <typename PipelineProblem>
|
|
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV5<PipelineProblem>;
|
|
template <typename PipelineProblem>
|
|
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV5<PipelineProblem>;
|
|
};
|
|
|
|
template <>
|
|
struct PipelineTypeTraits<CK_TILE_PIPELINE_PRESHUFFLE_V1>
|
|
{
|
|
template <typename PipelineProblem>
|
|
using GemmPipeline = ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV1<PipelineProblem>;
|
|
template <typename PipelineProblem>
|
|
using UniversalGemmPipeline =
|
|
ck_tile::BaseWeightPreshufflePipelineAGmemBGmemCRegV1<PipelineProblem>;
|
|
};
|
|
|
|
template <>
|
|
struct PipelineTypeTraits<CK_TILE_PIPELINE_PRESHUFFLE_V2>
|
|
{
|
|
template <typename PipelineProblem>
|
|
using GemmPipeline = ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV2<PipelineProblem>;
|
|
template <typename PipelineProblem>
|
|
using UniversalGemmPipeline =
|
|
ck_tile::BaseWeightPreshufflePipelineAGmemBGmemCRegV2<PipelineProblem>;
|
|
};
|
|
|
|
auto create_args(int argc, char* argv[])
|
|
{
|
|
ck_tile::ArgParser arg_parser;
|
|
arg_parser.insert("m", "3840", "m dimension")
|
|
.insert("n", "4096", "n dimension")
|
|
.insert("k", "2048", "k dimension")
|
|
.insert("a_layout", "R", "A tensor data layout - Row by default")
|
|
.insert("b_layout", "C", "B tensor data layout - Column by default")
|
|
.insert("c_layout", "R", "C tensor data layout - Row by default")
|
|
.insert("stride_a", "0", "Tensor A stride")
|
|
.insert("stride_b", "0", "Tensor B stride")
|
|
.insert("stride_c", "0", "Tensor C stride")
|
|
.insert("v", "2", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
|
|
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
|
|
.insert("warmup", "50", "number of iterations before benchmark the kernel")
|
|
.insert("repeat", "100", "number of iterations to benchmark the kernel")
|
|
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
|
|
.insert("split_k", "1", "splitK value")
|
|
.insert("init", "0", "0:random, 1:linear, 2:constant(1)")
|
|
.insert("persistent", "0", "0:non-persistent, 1:persistent")
|
|
.insert("flush_cache", "true", "flush cache before running the kernel, defaults to true")
|
|
.insert("rotating_count", "1", "rotating count, defaults to 1");
|
|
|
|
bool result = arg_parser.parse(argc, argv);
|
|
return std::make_tuple(result, arg_parser);
|
|
}
|
|
|
|
// host API
|
|
template <typename ADataType,
|
|
typename BDataType,
|
|
typename DsDataType,
|
|
typename AccDataType,
|
|
typename CDataType,
|
|
typename ALayout,
|
|
typename BLayout,
|
|
typename DsLayout,
|
|
typename CLayout,
|
|
bool Persistent = false,
|
|
typename CDEElementWise>
|
|
float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s);
|