mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-24 14:54:47 +00:00
[CK-Tile] Refactor base pipeline usage (#3251)
* initial poc
* factor out common parts in operator()
* cv4
* rest of the universal gemm pipelines
* fix test
* remove boilerplate from tile engine
* fix example
* fix example
* format
* fix tests build for gemm
* remove base pipeline codegen from gemm instance builder
* unify v3 logic with the rest of universal gemm pipelines
* fix build for multi abd test
* fix test gemm multi d
* fix build for weight preshuffle
* fix grouped gemm test
* fix grouped gemm multi d test
* fix grouped gemm preshuffle
* fix grouped gemm example except for quant
* fix gemm preshuffle
* fix splitk 2 stage example
* fix batched gemm example
* fix multid example
* fix multiabd example
* fix batched gemm test
* fixup
* fix examples build
* fix grouped gemm test build
* fix smoke builder
[ROCm/composable_kernel commit: d184eed823]
This commit is contained in:
Max Podkorytov
@@ -42,12 +42,6 @@ float grouped_gemm(const std::vector<grouped_gemm_kargs>& gemm_descs,
|
||||
GemmConfig::TileParitionerGroupNum,
|
||||
GemmConfig::TileParitionerM01>;
|
||||
|
||||
using Traits = ck_tile::TileGemmTraits<GemmConfig::kPadM,
|
||||
GemmConfig::kPadN,
|
||||
GemmConfig::kPadK,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>;
|
||||
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<GemmConfig::kPadM,
|
||||
GemmConfig::kPadN,
|
||||
GemmConfig::kPadK,
|
||||
@@ -56,109 +50,83 @@ float grouped_gemm(const std::vector<grouped_gemm_kargs>& gemm_descs,
|
||||
BLayout,
|
||||
CLayout,
|
||||
GemmConfig::TransposeC>;
|
||||
using GemmPipelineProblem =
|
||||
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
|
||||
|
||||
using BaseGemmPipeline = typename PipelineTypeTraits<
|
||||
GemmConfig::Pipeline>::template UniversalGemmPipeline<GemmPipelineProblem>;
|
||||
constexpr auto scheduler = GemmConfig::Scheduler;
|
||||
|
||||
const ck_tile::index_t k_grain = gemm_descs[0].k_batch * GemmConfig::K_Tile;
|
||||
const ck_tile::index_t K_split = (gemm_descs[0].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);
|
||||
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
scheduler>;
|
||||
|
||||
float ave_time{0};
|
||||
using GemmPipeline = typename PipelineTypeTraits<GemmConfig::Pipeline>::template GemmPipeline<
|
||||
UniversalGemmProblem>;
|
||||
const auto Run = [&](const auto memory_operation_) {
|
||||
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 = 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,
|
||||
CLayout,
|
||||
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>>;
|
||||
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKargs(gemm_descs);
|
||||
if(!Kernel::IsSupportedArgument(kargs))
|
||||
{
|
||||
throw std::runtime_error("Kernel arguments not supported!");
|
||||
}
|
||||
|
||||
const dim3 blocks = Kernel::BlockSize();
|
||||
const dim3 grids = Kernel::GridSize(gemm_descs);
|
||||
|
||||
HIP_CHECK_ERROR(hipMemcpyWithStream(kargs_ptr,
|
||||
kargs.data(),
|
||||
get_workspace_size(gemm_descs),
|
||||
hipMemcpyHostToDevice,
|
||||
s.stream_id_));
|
||||
|
||||
if(s.log_level_ > 0)
|
||||
{
|
||||
std::cout << "Launching kernel: " << Kernel::GetName()
|
||||
<< " with args:" << " grid: {" << grids.x << ", " << grids.y << ", "
|
||||
<< grids.z << "}" << ", blocks: {" << blocks.x << ", " << blocks.y << ", "
|
||||
<< blocks.z << "}" << std::endl;
|
||||
}
|
||||
|
||||
return ave_time = ck_tile::launch_kernel(
|
||||
s,
|
||||
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(
|
||||
Kernel{},
|
||||
grids,
|
||||
blocks,
|
||||
0,
|
||||
ck_tile::cast_pointer_to_constant_address_space(kargs_ptr),
|
||||
gemm_descs.size()));
|
||||
};
|
||||
|
||||
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
|
||||
if(gemm_descs[0].k_batch == 1)
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
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>>;
|
||||
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKargs(gemm_descs);
|
||||
if(!Kernel::IsSupportedArgument(kargs))
|
||||
{
|
||||
return Run(has_hot_loop_,
|
||||
tail_number_,
|
||||
ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::set>{});
|
||||
throw std::runtime_error("Kernel arguments not supported!");
|
||||
}
|
||||
else
|
||||
|
||||
const dim3 blocks = Kernel::BlockSize();
|
||||
const dim3 grids = Kernel::GridSize(gemm_descs);
|
||||
|
||||
HIP_CHECK_ERROR(hipMemcpyWithStream(kargs_ptr,
|
||||
kargs.data(),
|
||||
get_workspace_size(gemm_descs),
|
||||
hipMemcpyHostToDevice,
|
||||
s.stream_id_));
|
||||
|
||||
if(s.log_level_ > 0)
|
||||
{
|
||||
return Run(has_hot_loop_,
|
||||
tail_number_,
|
||||
ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::atomic_add>{});
|
||||
std::cout << "Launching kernel: " << Kernel::GetName() << " with args:" << " grid: {"
|
||||
<< grids.x << ", " << grids.y << ", " << grids.z << "}" << ", blocks: {"
|
||||
<< blocks.x << ", " << blocks.y << ", " << blocks.z << "}" << std::endl;
|
||||
}
|
||||
|
||||
return ck_tile::launch_kernel(
|
||||
s,
|
||||
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(
|
||||
Kernel{},
|
||||
grids,
|
||||
blocks,
|
||||
0,
|
||||
ck_tile::cast_pointer_to_constant_address_space(kargs_ptr),
|
||||
gemm_descs.size()));
|
||||
};
|
||||
|
||||
return ave_time = BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
|
||||
if(gemm_descs[0].k_batch == 1)
|
||||
{
|
||||
return Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::set>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::atomic_add>{});
|
||||
}
|
||||
}
|
||||
|
||||
template <typename GemmConfig,
|
||||
|
||||
Reference in New Issue
Block a user