mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +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
@@ -12,282 +12,32 @@
|
||||
#include "ck_tile/host.hpp"
|
||||
#include "gemm_utils.hpp"
|
||||
#include "run_gemm_example.inc"
|
||||
#include "run_gemm_example_common.hpp"
|
||||
#include "universal_gemm_invoker.hpp"
|
||||
|
||||
template <typename GemmConfig,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename DsDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
bool Persistent,
|
||||
typename CDEElementWise>
|
||||
float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
|
||||
|
||||
{
|
||||
using GemmShape = ck_tile::TileGemmShape<
|
||||
ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
|
||||
ck_tile::sequence<GemmConfig::M_Warp, GemmConfig::N_Warp, GemmConfig::K_Warp>,
|
||||
ck_tile::
|
||||
sequence<GemmConfig::M_Warp_Tile, GemmConfig::N_Warp_Tile, GemmConfig::K_Warp_Tile>,
|
||||
GemmConfig::PermuteA,
|
||||
GemmConfig::PermuteB>;
|
||||
|
||||
using TilePartitioner =
|
||||
ck_tile::GemmSpatiallyLocalTilePartitioner<GemmShape,
|
||||
GemmConfig::TileParitionerGroupNum,
|
||||
GemmConfig::TileParitionerM01>;
|
||||
|
||||
using Traits = ck_tile::TileGemmTraits<GemmConfig::kPadM,
|
||||
GemmConfig::kPadN,
|
||||
GemmConfig::kPadK,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ELayout,
|
||||
GemmConfig::NumWaveGroups>;
|
||||
|
||||
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<GemmConfig::kPadM,
|
||||
GemmConfig::kPadN,
|
||||
GemmConfig::kPadK,
|
||||
GemmConfig::DoubleSmemBuffer,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ELayout,
|
||||
GemmConfig::TransposeC,
|
||||
GemmConfig::UseStructuredSparsity,
|
||||
Persistent,
|
||||
GemmConfig::NumWaveGroups,
|
||||
GemmConfig::Preshuffle>;
|
||||
using GemmPipelineProblem =
|
||||
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
|
||||
|
||||
using BaseGemmPipeline = typename PipelineTypeTraits<
|
||||
GemmConfig::Pipeline>::template UniversalGemmPipeline<GemmPipelineProblem>;
|
||||
|
||||
const ck_tile::index_t k_grain = args.k_batch * GemmConfig::K_Tile;
|
||||
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * GemmConfig::K_Tile;
|
||||
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
|
||||
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
|
||||
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
|
||||
float ave_time{0};
|
||||
|
||||
const auto Run = [&](const auto has_hot_loop_,
|
||||
const auto tail_number_,
|
||||
const auto memory_operation_) {
|
||||
constexpr bool has_hot_loop_v = has_hot_loop_.value;
|
||||
constexpr auto tail_number_v = tail_number_.value;
|
||||
constexpr auto scheduler = GemmConfig::Scheduler;
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
|
||||
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
scheduler,
|
||||
has_hot_loop_v,
|
||||
tail_number_v>;
|
||||
|
||||
using GemmPipeline = typename PipelineTypeTraits<
|
||||
GemmConfig::Pipeline>::template GemmPipeline<UniversalGemmProblem>;
|
||||
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
CDEElementWise,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
GemmConfig::M_Warp,
|
||||
GemmConfig::N_Warp,
|
||||
GemmConfig::M_Warp_Tile,
|
||||
GemmConfig::N_Warp_Tile,
|
||||
GemmConfig::K_Warp_Tile,
|
||||
UniversalGemmProblem::TransposeC,
|
||||
memory_operation,
|
||||
GemmConfig::NumWaveGroups>>;
|
||||
|
||||
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKernelArgs(args);
|
||||
|
||||
dim3 grids;
|
||||
if constexpr(Persistent)
|
||||
{
|
||||
grids = Kernel::MaxOccupancyGridSize(s);
|
||||
}
|
||||
else
|
||||
{
|
||||
grids = Kernel::GridSize(args.M, args.N, args.k_batch);
|
||||
}
|
||||
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: " << GemmShape::GetName() << '\n'
|
||||
<< "problem: " << UniversalGemmProblem::GetName() << '\n'
|
||||
<< "pipeline: " << GemmPipeline::GetName() << '\n'
|
||||
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
|
||||
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
|
||||
<< std::endl;
|
||||
}
|
||||
if(s.flush_cache_)
|
||||
{
|
||||
std::cout << "Flushing cache..." << std::endl;
|
||||
|
||||
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
|
||||
args.M, args.K, args.stride_A, is_row_major(ALayout{})));
|
||||
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
|
||||
args.K, args.N, args.stride_B, is_row_major(BLayout{})));
|
||||
|
||||
auto size_a_buffer = a_m.get_element_space_size_in_bytes();
|
||||
auto size_b_buffer = b_n.get_element_space_size_in_bytes();
|
||||
|
||||
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
|
||||
kargs.as_ptr[0], kargs.bs_ptr[0], s.rotating_count_, size_a_buffer, size_b_buffer);
|
||||
rotating_mem.Print();
|
||||
|
||||
auto run_flush_cache = [&]() {
|
||||
// flush icache
|
||||
ck_tile::flush_icache();
|
||||
// rotating mem
|
||||
rotating_mem.Next();
|
||||
// clear c mem
|
||||
if(args.k_batch > 1)
|
||||
hipGetErrorString(hipMemsetAsync(
|
||||
args.e_ptr, 0, args.M * args.N * sizeof(CDataType), s.stream_id_));
|
||||
};
|
||||
ave_time = ck_tile::launch_kernel_time_mask(
|
||||
s,
|
||||
run_flush_cache,
|
||||
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
|
||||
}
|
||||
else
|
||||
{
|
||||
ave_time = ck_tile::launch_kernel(
|
||||
s,
|
||||
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
|
||||
}
|
||||
return ave_time;
|
||||
};
|
||||
|
||||
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
|
||||
if(args.k_batch == 1)
|
||||
{
|
||||
Run(has_hot_loop_, tail_number_, MemoryOpSet{});
|
||||
}
|
||||
else
|
||||
{
|
||||
Run(has_hot_loop_, tail_number_, MemoryOpAtomicAdd{});
|
||||
}
|
||||
};
|
||||
|
||||
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
template <typename GemmConfig,
|
||||
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;
|
||||
bool preshuffle = GemmConfig::Preshuffle;
|
||||
|
||||
if(preshuffle && std::is_same_v<BPrecType, ck_tile::pk_int4_t>)
|
||||
{
|
||||
throw std::runtime_error("Preshuffle is not supported for this int4 datatype!");
|
||||
}
|
||||
|
||||
if(preshuffle && a_layout != "R" && b_layout != "C")
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"Preshuffle is supported only for A(Row major), B(column major) input matrices!");
|
||||
}
|
||||
|
||||
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<GemmConfig, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Row{}, Col{}, Row{});
|
||||
}
|
||||
else if(a_layout == "C" && b_layout == "C")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfig, 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 == "R")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfig, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(a_layout == "R" && b_layout == "C")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfig, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Row{}, Col{}, Row{});
|
||||
}
|
||||
else if(a_layout == "C" && b_layout == "R")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfig, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Col{}, Row{}, Row{});
|
||||
}
|
||||
else if(a_layout == "C" && b_layout == "C")
|
||||
{
|
||||
return run_gemm_example_with_layouts<GemmConfig, APrecType, BPrecType, CPrecType>(
|
||||
arg_parser, Col{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
throw std::runtime_error("Unsupported memory layout for the input matrices!");
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <template <typename PreType> typename GemmConfig>
|
||||
template <template <typename PrecType> typename GemmConfig>
|
||||
int run_gemm_example(ck_tile::ArgParser& arg_parser)
|
||||
{
|
||||
std::string data_type = arg_parser.get_str("prec");
|
||||
std::string a_layout = arg_parser.get_str("a_layout");
|
||||
std::string b_layout = arg_parser.get_str("b_layout");
|
||||
|
||||
using Invoker = UniversalInvoker;
|
||||
|
||||
if(data_type == "fp16")
|
||||
{
|
||||
return run_gemm_example_prec_type<GemmConfig<ck_tile::half_t>, ck_tile::half_t>(
|
||||
return run_gemm_example_prec_type<GemmConfig<ck_tile::half_t>, Invoker, ck_tile::half_t>(
|
||||
a_layout, b_layout, arg_parser);
|
||||
}
|
||||
else if(data_type == "bf16")
|
||||
{
|
||||
return run_gemm_example_prec_type<GemmConfig<ck_tile::half_t>, ck_tile::bf16_t>(
|
||||
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf16_t>, Invoker, ck_tile::bf16_t>(
|
||||
a_layout, b_layout, arg_parser);
|
||||
}
|
||||
else if(data_type == "fp8")
|
||||
{
|
||||
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
|
||||
Invoker,
|
||||
ck_tile::fp8_t,
|
||||
ck_tile::fp8_t,
|
||||
ck_tile::half_t>(a_layout, b_layout, arg_parser);
|
||||
@@ -295,6 +45,7 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
|
||||
else if(data_type == "bf8")
|
||||
{
|
||||
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
|
||||
Invoker,
|
||||
ck_tile::bf8_t,
|
||||
ck_tile::bf8_t,
|
||||
ck_tile::half_t>(a_layout, b_layout, arg_parser);
|
||||
@@ -302,6 +53,7 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
|
||||
else if(data_type == "int8")
|
||||
{
|
||||
return run_gemm_example_prec_type<GemmConfig<ck_tile::int8_t>,
|
||||
Invoker,
|
||||
ck_tile::int8_t,
|
||||
ck_tile::int8_t,
|
||||
ck_tile::int32_t>(a_layout, b_layout, arg_parser);
|
||||
@@ -312,6 +64,7 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
|
||||
if constexpr(GemmConfig<ck_tile::half_t>::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
|
||||
{
|
||||
return run_gemm_example_prec_type<GemmConfig<ck_tile::half_t>,
|
||||
Invoker,
|
||||
ck_tile::half_t,
|
||||
ck_tile::pk_int4_t,
|
||||
ck_tile::half_t>(a_layout, b_layout, arg_parser);
|
||||
@@ -329,7 +82,9 @@ 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