mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-05 22:22:27 +00:00
[CK Tile] gemm splitk two stage (#2697)
* Fix a typo * Use std::variant to call run_gemm_example_with_layouts with the available layout variant combinations * Use a unified run_gemm_example_prec_type for basic gemm and universal gemm * Factor out run_gemm_example_prec_type * Refactor argument parsing in gemm_splitk_two_stage_reduce.cpp * Parse arguments outside of create_args * Move the gemm operators to separate structs to facilitate their reuse * Move the invokers to separate files to facilitate their reuse * Rename the invoker files for consistency with the examples that use them * Add fp32 support to the elementwise examples, and produce an error message for unsupported types * Get rid of four unused variables * Make two variables const * Add support for different input-output type combinations in elementwise examples * Test support for different input and output types in elementwise examples * Add support for different operations in the elementwise unary tests * Add support for UnaryConvert in the elementwise unary tests * Add support for bf16 in elementwise examples, excluding unsupported type combinations * Make some operator parameters const in ElementWiseKernel * Remove some unnecessary include statements * Implement a two-stage GEMM that does a type conversion in the second stage using the elementwise kernel * Clear workspace instead of output when flushing the cache in SplitKTwoStageInvoker::gemm * Fix formatting issues reported by clang * Add back CK_TILE_USE_WMMA related changes * Use the right prec type for bf16 in the universal GEMM and two stage split K examples * Add some brackets * Add some brackets * Separate the clearing of the GEMM output memory from the cache flushing in the universal GEMM example * Separate the clearing of the GEMM output memory from the cache flushing in the split K two stage example * Fix formatting * No need to call SetZero on ws_m_n_dev_buf here, as clear_gemm_output now does this as part of the kernel preprocessing * Add fp16 data type to splitk two stage example * Add preprocessing with optional cache flushing and clearing of output for k_batch > 1 to the basic GEMM example
This commit is contained in:
SamiAario-AMD
@@ -2,185 +2,9 @@
|
||||
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "gemm_utils.hpp"
|
||||
|
||||
template <typename GemmConfig,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename DsDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename CLayout,
|
||||
bool Persistent,
|
||||
typename CDEElementWise>
|
||||
float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
|
||||
|
||||
{
|
||||
if constexpr(Persistent)
|
||||
std::cout << "WARNING: Ignoring persistent kernel option for basic gemm." << std::endl;
|
||||
|
||||
// This part comes from the Codegen
|
||||
constexpr ck_tile::index_t M_Tile = 256;
|
||||
constexpr ck_tile::index_t N_Tile = 256;
|
||||
constexpr ck_tile::index_t K_Tile = 64;
|
||||
|
||||
#if CK_TILE_USE_WMMA
|
||||
constexpr ck_tile::index_t M_Warp = 4;
|
||||
constexpr ck_tile::index_t N_Warp = 2;
|
||||
constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
constexpr ck_tile::index_t M_Warp_Tile = 16;
|
||||
constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
constexpr ck_tile::index_t K_Warp_Tile = 16;
|
||||
#else
|
||||
constexpr ck_tile::index_t M_Warp = 2;
|
||||
constexpr ck_tile::index_t N_Warp = 2;
|
||||
constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t K_Warp_Tile = 16;
|
||||
#endif
|
||||
|
||||
using CodegenGemmShape =
|
||||
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
|
||||
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
|
||||
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
|
||||
|
||||
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
|
||||
|
||||
using CodegenGemmTraits = ck_tile::TileGemmTraits<GemmConfig::kPadM,
|
||||
GemmConfig::kPadN,
|
||||
GemmConfig::kPadK,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>;
|
||||
|
||||
using CodegenPipelineProblem = ck_tile::
|
||||
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
|
||||
|
||||
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
|
||||
|
||||
const auto Run = [&](const auto memory_operation_) {
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
ck_tile::tuple<>,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ck_tile::tuple<>,
|
||||
CLayout,
|
||||
ck_tile::element_wise::PassThrough,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
M_Warp,
|
||||
N_Warp,
|
||||
M_Warp_Tile,
|
||||
N_Warp_Tile,
|
||||
K_Warp_Tile,
|
||||
CodegenPipelineProblem::TransposeC,
|
||||
memory_operation>>;
|
||||
|
||||
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
|
||||
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
|
||||
using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKernelArgs(args);
|
||||
|
||||
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
|
||||
const dim3 blocks = Kernel::BlockSize();
|
||||
|
||||
if(!Kernel::IsSupportedArgument(kargs))
|
||||
{
|
||||
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
|
||||
}
|
||||
|
||||
if(s.log_level_ > 0)
|
||||
{
|
||||
std::cout << "Launching kernel with args: " << Kernel::GetName() << '\n'
|
||||
<< "shape: " << CodegenGemmShape::GetName() << '\n'
|
||||
<< "problem: " << CodegenPipelineProblem::GetName() << '\n'
|
||||
<< "pipeline: " << CodegenGemmPipeline::GetName() << '\n'
|
||||
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
|
||||
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
|
||||
<< std::endl;
|
||||
}
|
||||
|
||||
float ave_time = ck_tile::launch_kernel(
|
||||
s, ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
|
||||
|
||||
return ave_time;
|
||||
};
|
||||
|
||||
if(args.k_batch == 1)
|
||||
{
|
||||
return Run(MemoryOpSet{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return Run(MemoryOpAtomicAdd{});
|
||||
}
|
||||
}
|
||||
|
||||
#include "run_gemm_example.inc"
|
||||
|
||||
template <typename APrecType, typename BPrecType = APrecType, typename CPrecType = APrecType>
|
||||
int run_gemm_example_prec_type(std::string a_layout,
|
||||
std::string b_layout,
|
||||
ck_tile::ArgParser& arg_parser)
|
||||
{
|
||||
using Row = ck_tile::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
if constexpr(std::is_same_v<BPrecType, ck_tile::pk_int4_t>)
|
||||
{
|
||||
if(a_layout == "R" && b_layout == "C")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfigBase, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Row{}, Col{}, Row{});
|
||||
}
|
||||
else if(a_layout == "C" && b_layout == "C")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfigBase, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Col{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Unsupported memory layout for the input matrices when "
|
||||
"BPrecType is ck_tile::pk_int4_t!");
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
if(a_layout == "R" && b_layout == "C")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfigBase, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Row{}, Col{}, Row{});
|
||||
}
|
||||
else if(a_layout == "R" && b_layout == "R")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfigBase, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(a_layout == "C" && b_layout == "R")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfigBase, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Col{}, Row{}, Row{});
|
||||
}
|
||||
else if(a_layout == "C" && b_layout == "C")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfigBase, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Col{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Unsupported memory layout for the input matrices!");
|
||||
}
|
||||
}
|
||||
}
|
||||
#include "run_gemm_example_common.hpp"
|
||||
#include "gemm_basic_invoker.hpp"
|
||||
|
||||
int run_gemm_example(ck_tile::ArgParser& arg_parser)
|
||||
{
|
||||
@@ -188,36 +12,53 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
|
||||
std::string a_layout = arg_parser.get_str("a_layout");
|
||||
std::string b_layout = arg_parser.get_str("b_layout");
|
||||
|
||||
using GemmConfig = GemmConfigBase;
|
||||
using Invoker = BasicInvoker;
|
||||
|
||||
if(data_type == "fp16")
|
||||
{
|
||||
return run_gemm_example_prec_type<ck_tile::half_t>(a_layout, b_layout, arg_parser);
|
||||
return run_gemm_example_prec_type<GemmConfig, Invoker, ck_tile::half_t>(
|
||||
a_layout, b_layout, arg_parser);
|
||||
}
|
||||
else if(data_type == "bf16")
|
||||
{
|
||||
return run_gemm_example_prec_type<ck_tile::bf16_t>(a_layout, b_layout, arg_parser);
|
||||
return run_gemm_example_prec_type<GemmConfig, Invoker, ck_tile::bf16_t>(
|
||||
a_layout, b_layout, arg_parser);
|
||||
}
|
||||
else if(data_type == "fp8")
|
||||
{
|
||||
return run_gemm_example_prec_type<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>(
|
||||
a_layout, b_layout, arg_parser);
|
||||
return run_gemm_example_prec_type<GemmConfig,
|
||||
Invoker,
|
||||
ck_tile::fp8_t,
|
||||
ck_tile::fp8_t,
|
||||
ck_tile::half_t>(a_layout, b_layout, arg_parser);
|
||||
}
|
||||
else if(data_type == "bf8")
|
||||
{
|
||||
return run_gemm_example_prec_type<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t>(
|
||||
a_layout, b_layout, arg_parser);
|
||||
return run_gemm_example_prec_type<GemmConfig,
|
||||
Invoker,
|
||||
ck_tile::bf8_t,
|
||||
ck_tile::bf8_t,
|
||||
ck_tile::half_t>(a_layout, b_layout, arg_parser);
|
||||
}
|
||||
else if(data_type == "i8")
|
||||
{
|
||||
return run_gemm_example_prec_type<ck_tile::int8_t, ck_tile::int8_t, int32_t>(
|
||||
a_layout, b_layout, arg_parser);
|
||||
return run_gemm_example_prec_type<GemmConfig,
|
||||
Invoker,
|
||||
ck_tile::int8_t,
|
||||
ck_tile::int8_t,
|
||||
int32_t>(a_layout, b_layout, arg_parser);
|
||||
}
|
||||
else if(data_type == "pk_int4_t")
|
||||
{
|
||||
// TODO: Add support for bhalf_t ADataType
|
||||
if constexpr(GemmConfigBase::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
|
||||
if constexpr(GemmConfig::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
|
||||
{
|
||||
return run_gemm_example_prec_type<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>(
|
||||
a_layout, b_layout, arg_parser);
|
||||
return run_gemm_example_prec_type<GemmConfig,
|
||||
Invoker,
|
||||
ck_tile::half_t,
|
||||
ck_tile::pk_int4_t,
|
||||
ck_tile::half_t>(a_layout, b_layout, arg_parser);
|
||||
}
|
||||
else
|
||||
{
|
||||
@@ -232,7 +73,9 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
auto [result, arg_parser] = create_args(argc, argv);
|
||||
auto arg_parser = create_args();
|
||||
auto result = arg_parser.parse(argc, argv);
|
||||
|
||||
if(!result)
|
||||
return -1;
|
||||
|
||||
|
||||
Reference in New Issue
Block a user