mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
[CK_tile] Add rotating buffer feature for universal gemm (#2200)
* Add rotating buffer feature for universal gemm * adding changes in tile_engine * Updated code to merge kernel_launch * removing comments * Enable rotating buffer changes to flatmm * Created diff launch_kernel function for rotating buffer * Simplfied calculation using macros * merge code with new changes in tile_engine * clang formatted * Redefine macros
This commit is contained in:
Khushbu Agarwal
@@ -11,6 +11,7 @@
|
||||
|
||||
#include "ck_tile/host.hpp"
|
||||
#include "gemm_utils.hpp"
|
||||
#include "run_gemm_example.inc"
|
||||
|
||||
template <typename Pipeline, ck_tile::TailNumber TN>
|
||||
void try_run(ck_tile::TailNumber tn)
|
||||
@@ -74,64 +75,102 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
|
||||
|
||||
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 = GEMM_PIPELINE_SCHEDULER;
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
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 = GEMM_PIPELINE_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 UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
scheduler,
|
||||
has_hot_loop_v,
|
||||
tail_number_v>;
|
||||
|
||||
using GemmPipeline = GEMM_PIPELINE<UniversalGemmProblem>;
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
CLayout,
|
||||
GemmPipelineProblem::kBlockSize,
|
||||
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>>;
|
||||
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKernelArgs(args);
|
||||
using GemmPipeline = GEMM_PIPELINE<UniversalGemmProblem>;
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
CLayout,
|
||||
GemmPipelineProblem::kBlockSize,
|
||||
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>>;
|
||||
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKernelArgs(args);
|
||||
|
||||
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
|
||||
constexpr dim3 blocks = Kernel::BlockSize();
|
||||
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
|
||||
constexpr dim3 blocks = Kernel::BlockSize();
|
||||
|
||||
if(!Kernel::IsSupportedArgument(kargs))
|
||||
{
|
||||
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
|
||||
}
|
||||
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:"
|
||||
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
|
||||
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
|
||||
<< std::endl;
|
||||
}
|
||||
if(s.log_level_ > 0)
|
||||
{
|
||||
std::cout << "Launching kernel with args:"
|
||||
<< " 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;
|
||||
static constexpr ck_tile::index_t APackedSize =
|
||||
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
|
||||
static constexpr ck_tile::index_t BPackedSize =
|
||||
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
|
||||
|
||||
ave_time = ck_tile::launch_kernel(s,
|
||||
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
|
||||
Kernel{}, grids, blocks, 0, kargs));
|
||||
return ave_time;
|
||||
};
|
||||
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() / APackedSize;
|
||||
auto size_b_buffer = b_n.get_element_space_size_in_bytes() / BPackedSize;
|
||||
|
||||
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
|
||||
kargs.a_ptr, kargs.b_ptr, 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.c_ptr, 0, args.M * args.N * sizeof(CDataType), s.stream_id_));
|
||||
};
|
||||
ave_time = ck_tile::launch_kernel_preprocess(
|
||||
s,
|
||||
run_flush_cache,
|
||||
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
|
||||
Kernel{}, grids, blocks, 0, kargs));
|
||||
}
|
||||
else
|
||||
{
|
||||
ave_time =
|
||||
ck_tile::launch_kernel(s,
|
||||
ck_tile::make_kernel<blocks.x, 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)
|
||||
@@ -243,8 +282,6 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
#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, int argc, char* argv[])
|
||||
{
|
||||
|
||||
Reference in New Issue
Block a user