mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-18 20:09:25 +00:00
[CK-Tile] move out memory operation from cshuffle epilogue class (#3359)
* 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
* hacky poc
* fix tile engine
* kill the lambda
* maybe fix test build
* more fixes
* clang-format
* save temp
* clang-format
* mostly fix examples
* clang-format
* remove dead code
* more cleanup
* fix fmha bwd build (default epilogue set/add appears to be broken)
* fix default epilogue tests but not correctness
* clang-format
* fix bquant
* clang-format
* cleanup dead code
* rearrange make windows for readability
* restore changes to IsSupportedArgument
* fix smoke-builder
* clang-format
* fixup rename class
* build fixes
* clang-format
* fix builder
* fixup
* remove set from builder tests
* fix test
* clang-format
* re-refactor the kernels
* clang-format
* fix header license
* remove memory operation from conv bwd test
* clang-format
* clang-format example,include
* clang-format test
* build fixes
* clang-format
* solve compilation error
* fix the CI
* solve compilation error
* clang format
* solve merge conflict
* solve merge conflict
* solve the gfx11 error
* solve test error
* moar build fixes
* remove AtomicAddRequiresKBatchGreaterThanOne test since the property is removed from the kernel scope
---------
Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
[ROCm/composable_kernel commit: e339101e9c]
This commit is contained in:
Max Podkorytov
@@ -62,71 +62,55 @@ float grouped_gemm(const std::vector<grouped_gemm_kargs>& gemm_descs,
|
||||
|
||||
using GemmPipeline = typename PipelineTypeTraits<GemmConfig::Pipeline>::template GemmPipeline<
|
||||
UniversalGemmProblem>;
|
||||
const auto Run = [&](const auto memory_operation_) {
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
|
||||
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 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()));
|
||||
};
|
||||
|
||||
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>>;
|
||||
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKargs(gemm_descs);
|
||||
if(!Kernel::IsSupportedArgument(kargs))
|
||||
{
|
||||
return Run(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(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()));
|
||||
}
|
||||
|
||||
template <typename GemmConfig,
|
||||
@@ -139,8 +123,7 @@ template <typename GemmConfig,
|
||||
typename CDataType>
|
||||
float grouped_gemm_tileloop(const ck_tile::stream_config& s,
|
||||
const ck_tile::index_t num_groups,
|
||||
void* kargs_ptr,
|
||||
bool splitk)
|
||||
void* kargs_ptr)
|
||||
{
|
||||
using GemmShape = ck_tile::TileGemmShape<
|
||||
ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
|
||||
@@ -161,74 +144,55 @@ float grouped_gemm_tileloop(const ck_tile::stream_config& s,
|
||||
BLayout,
|
||||
CLayout>;
|
||||
|
||||
float ave_time{0};
|
||||
constexpr auto scheduler = GemmConfig::Scheduler;
|
||||
|
||||
const auto Run = [&](const auto memory_operation_) {
|
||||
constexpr auto scheduler = GemmConfig::Scheduler;
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
// We create the GEMM pipeline without specifying hotloop or tailnumber.
|
||||
// These are automatically run inside the kernel based on the given input data.
|
||||
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
scheduler>;
|
||||
|
||||
// We create the GEMM pipeline without specifying hotloop or tailnumber.
|
||||
// These are automatically run inside the kernel based on the given input data.
|
||||
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
scheduler>;
|
||||
using GemmPipeline = typename PipelineTypeTraits<GemmConfig::Pipeline>::template GemmPipeline<
|
||||
UniversalGemmProblem>;
|
||||
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,
|
||||
GemmConfig::M_Warp,
|
||||
GemmConfig::N_Warp,
|
||||
GemmConfig::M_Warp_Tile,
|
||||
GemmConfig::N_Warp_Tile,
|
||||
GemmConfig::K_Warp_Tile,
|
||||
UniversalGemmProblem::TransposeC>>;
|
||||
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
const dim3 blocks = Kernel::BlockSize();
|
||||
const dim3 grids = Kernel::MaxOccupancyGridSize(s);
|
||||
|
||||
using GemmPipeline = typename PipelineTypeTraits<
|
||||
GemmConfig::Pipeline>::template GemmPipeline<UniversalGemmProblem>;
|
||||
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,
|
||||
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>;
|
||||
const dim3 blocks = Kernel::BlockSize();
|
||||
const dim3 grids = Kernel::MaxOccupancyGridSize(s);
|
||||
|
||||
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),
|
||||
num_groups));
|
||||
};
|
||||
|
||||
if(!splitk)
|
||||
if(s.log_level_ > 0)
|
||||
{
|
||||
return ave_time = Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::set>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return ave_time =
|
||||
Run(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),
|
||||
num_groups));
|
||||
}
|
||||
|
||||
#include "run_grouped_gemm_example.inc"
|
||||
|
||||
Reference in New Issue
Block a user