ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-01 12:11:19 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
78ae3835a66d725dcf19b4e3877203e4e1aaf4a4
composable_kernel/example/ck_tile/03_gemm/gemm_utils.hpp
Max Podkorytov 91b4102a59 Add persistent async input scheduler for GEMM kernels (#3520) 
Add signal-based synchronization for persistent GEMM kernels where
input data becomes available incrementally. Uses modulo wraparound
(like PyTorch's AsyncMM) for chunk index calculation:
  chunk_idx = ((tile_idx + tile_idx_pivot) / tiles_per_chunk) % num_chunks

Key components:
- PersistentAsyncInputScheduler struct with tiles_per_chunk_m,
  chunk_signals, tile_idx_pivot_m, and num_chunks fields
- wait_eq_wave method using __builtin_amdgcn_s_sleep for power efficiency
- IsSupportedArgument validation for scheduler parameters
- Example demonstrating async input scheduling with simulated producer
- GTest unit tests covering all layout combinations
2026-01-20 10:37:09 -08:00

477 lines
18 KiB
C++
Raw Blame History

// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#pragma once
#include <string>
#include <variant>
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/utility/json_dump.hpp"
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::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V3;
static constexpr ck_tile::index_t NumWaveGroups = 1;
static constexpr bool Preshuffle = false;
static constexpr bool TiledMMAPermuteN = 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::GemmPipeline Pipeline = ck_tile::GemmPipeline::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::GemmPipeline Pipeline = ck_tile::GemmPipeline::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 =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::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 =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::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 =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V3;
static constexpr int kBlockPerCu = 2;
};
template <typename PrecType>
struct GemmConfigComputeV3_WMMA : 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 = 4;
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 = 16;
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::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 =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = true;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::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 =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = true;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::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 =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V5;
static constexpr ck_tile::index_t NumWaveGroups = 2;
};
template <typename PrecType>
struct GemmConfigComputeV6 : 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 = 32;
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 = 16;
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V6;
static constexpr ck_tile::index_t NumWaveGroups = 1;
};
template <typename PrecType>
struct GemmConfigPreshuffleDecode : public GemmConfigBase
{
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 =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr int kBlockPerCu = 1;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::PRESHUFFLE_V2;
static constexpr bool Preshuffle = true;
static constexpr bool DoubleSmemBuffer = true;
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
static constexpr bool TiledMMAPermuteN = N_Repeat % 2 == 0;
};
template <typename PrecType>
struct GemmConfigPreshufflePrefill : 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 =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr int kBlockPerCu = 2;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::PRESHUFFLE_V2;
static constexpr bool Preshuffle = true;
static constexpr bool DoubleSmemBuffer = true;
static constexpr int N_Repeat = N_Tile / N_Warp_Tile / N_Warp;
static constexpr bool TiledMMAPermuteN = N_Repeat % 2 == 0;
};
template <typename PrecType>
struct GemmConfigPreshufflePrefill_Wmma : public GemmConfigPreshufflePrefill<PrecType>
{
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 = 16;
};
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::fp8_t, ck_tile::pk_int4_t, ck_tile::half_t>
{
using ADataType = ck_tile::fp8_t;
using BDataType = ck_tile::pk_int4_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmTypeConfig<ck_tile::bf8_t, ck_tile::pk_int4_t, ck_tile::half_t>
{
using ADataType = ck_tile::bf8_t;
using BDataType = ck_tile::pk_int4_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 <ck_tile::GemmPipeline PipelineId>
struct PipelineTypeTraits;
template <>
struct PipelineTypeTraits<ck_tile::GemmPipeline::MEMORY>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<PipelineProblem>;
template <typename PipelineProblem>
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<PipelineProblem>;
};
template <>
struct PipelineTypeTraits<ck_tile::GemmPipeline::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::GemmPipeline::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::GemmPipeline::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::GemmPipeline::COMPUTE_V6>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV6<PipelineProblem>;
template <typename PipelineProblem>
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV6<PipelineProblem>;
};
template <>
struct PipelineTypeTraits<ck_tile::GemmPipeline::PRESHUFFLE_V2>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV2<PipelineProblem>;
template <typename PipelineProblem>
using UniversalGemmPipeline =
ck_tile::BaseWeightPreshufflePipelineAGmemBGmemCRegV2<PipelineProblem>;
};
inline auto create_args()
{
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/pk_int4_t")
.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("json", "0", "0: No Json, 1: Dump Results in Json format")
.insert("jsonfile", "gemm.json", "json file name to dump results")
.insert("flush_cache", "true", "flush cache before running the kernel, defaults to true")
.insert("rotating_count", "1000", "rotating count, defaults to 1000")
.insert("test_async", "0", "0: normal gemm, 1: test async input scheduler");
return 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);
Reference in New Issue View Git Blame Copy Permalink