mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-11 00:40:09 +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
This commit is contained in:
Max Podkorytov
@@ -123,8 +123,6 @@ class TestCkTileGroupedGemmPreshuffle : public ::testing::Test
|
||||
using TilePartitioner = ck_tile::
|
||||
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
|
||||
|
||||
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
|
||||
|
||||
// for testing purposes, we can hardcode the values here as we what is compatible with
|
||||
// pipeline
|
||||
using GemmUniversalTraits =
|
||||
@@ -140,58 +138,37 @@ class TestCkTileGroupedGemmPreshuffle : public ::testing::Test
|
||||
/*Persistent*/ false,
|
||||
/*NumWaveGroups*/ 1,
|
||||
/*Preshuffle*/ true>;
|
||||
using GemmPipelineProblem =
|
||||
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
|
||||
|
||||
using BaseGemmPipeline =
|
||||
ck_tile::BaseWeightPreshufflePipelineAGmemBGmemCRegV2<GemmPipelineProblem>;
|
||||
using UniversalGemmProblem =
|
||||
ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
ck_tile::GemmPipelineScheduler::Default>;
|
||||
using GemmPipeline =
|
||||
ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV2<UniversalGemmProblem>;
|
||||
|
||||
const ck_tile::index_t k_grain = gemm_descs[0].k_batch * K_Tile;
|
||||
const ck_tile::index_t K_split = (gemm_descs[0].K + k_grain - 1) / k_grain * K_Tile;
|
||||
const ck_tile::index_t num_loop =
|
||||
ck_tile::GemmSpatiallyLocalTilePartitioner<GemmShape,
|
||||
TileParitionerGroupNum,
|
||||
TileParitionerM01>::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;
|
||||
const auto Run = [&](const auto memory_operation_) {
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
using UniversalGemmProblem =
|
||||
ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
ck_tile::GemmPipelineScheduler::Default,
|
||||
has_hot_loop_v,
|
||||
tail_number_v>;
|
||||
using GemmPipeline =
|
||||
ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV2<UniversalGemmProblem>;
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
ck_tile::element_wise::PassThrough,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
M_Warp,
|
||||
N_Warp,
|
||||
M_Warp_Tile,
|
||||
N_Warp_Tile,
|
||||
K_Warp_Tile,
|
||||
UniversalGemmProblem::TransposeC,
|
||||
memory_operation>>;
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
ck_tile::element_wise::PassThrough,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
M_Warp,
|
||||
N_Warp,
|
||||
M_Warp_Tile,
|
||||
N_Warp_Tile,
|
||||
K_Warp_Tile,
|
||||
UniversalGemmProblem::TransposeC,
|
||||
memory_operation>>;
|
||||
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKargs(gemm_descs);
|
||||
EXPECT_TRUE(Kernel::IsSupportedArgument(kargs));
|
||||
@@ -204,7 +181,7 @@ class TestCkTileGroupedGemmPreshuffle : public ::testing::Test
|
||||
hipMemcpyHostToDevice,
|
||||
s.stream_id_));
|
||||
|
||||
ave_time = ck_tile::launch_kernel(
|
||||
return ck_tile::launch_kernel(
|
||||
s,
|
||||
ck_tile::make_kernel<kBlockPerCu>(
|
||||
Kernel{},
|
||||
@@ -213,25 +190,18 @@ class TestCkTileGroupedGemmPreshuffle : public ::testing::Test
|
||||
0,
|
||||
ck_tile::cast_pointer_to_constant_address_space(kargs_ptr),
|
||||
gemm_descs.size()));
|
||||
return ave_time;
|
||||
};
|
||||
|
||||
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
|
||||
if(gemm_descs[0].k_batch == 1)
|
||||
{
|
||||
Run(has_hot_loop_,
|
||||
tail_number_,
|
||||
ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::set>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
// EXPECT TO FAIL because splitk is not supported
|
||||
EXPECT_FALSE(true);
|
||||
}
|
||||
};
|
||||
|
||||
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
|
||||
if(gemm_descs[0].k_batch == 1)
|
||||
{
|
||||
Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::set>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
// EXPECT TO FAIL because splitk is not supported
|
||||
EXPECT_FALSE(true);
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
@@ -247,8 +217,6 @@ class TestCkTileGroupedGemmPreshuffle : public ::testing::Test
|
||||
using TilePartitioner = ck_tile::
|
||||
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
|
||||
|
||||
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
|
||||
|
||||
// Enable persistent mode for preshuffle
|
||||
using GemmUniversalTraits =
|
||||
ck_tile::TileGemmUniversalTraits</*kPadM*/ true,
|
||||
@@ -263,58 +231,36 @@ class TestCkTileGroupedGemmPreshuffle : public ::testing::Test
|
||||
/*Persistent*/ true, // Enable persistent mode
|
||||
/*NumWaveGroups*/ 1,
|
||||
/*Preshuffle*/ true>;
|
||||
using GemmPipelineProblem =
|
||||
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
|
||||
|
||||
using BaseGemmPipeline =
|
||||
ck_tile::BaseWeightPreshufflePipelineAGmemBGmemCRegV2<GemmPipelineProblem>;
|
||||
|
||||
const ck_tile::index_t k_grain = gemm_descs[0].k_batch * K_Tile;
|
||||
const ck_tile::index_t K_split = (gemm_descs[0].K + k_grain - 1) / k_grain * K_Tile;
|
||||
const ck_tile::index_t num_loop =
|
||||
ck_tile::GemmSpatiallyLocalTilePartitioner<GemmShape,
|
||||
TileParitionerGroupNum,
|
||||
TileParitionerM01>::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;
|
||||
using UniversalGemmProblem =
|
||||
ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
ck_tile::GemmPipelineScheduler::Default>;
|
||||
using GemmPipeline =
|
||||
ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV2<UniversalGemmProblem>;
|
||||
const auto Run = [&](const auto memory_operation_) {
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
using UniversalGemmProblem =
|
||||
ck_tile::UniversalGemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
GemmShape,
|
||||
GemmUniversalTraits,
|
||||
ck_tile::GemmPipelineScheduler::Default,
|
||||
has_hot_loop_v,
|
||||
tail_number_v>;
|
||||
using GemmPipeline =
|
||||
ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV2<UniversalGemmProblem>;
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
ck_tile::element_wise::PassThrough,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
M_Warp,
|
||||
N_Warp,
|
||||
M_Warp_Tile,
|
||||
N_Warp_Tile,
|
||||
K_Warp_Tile,
|
||||
UniversalGemmProblem::TransposeC,
|
||||
memory_operation>>;
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<ADataType,
|
||||
BDataType,
|
||||
DsDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
ck_tile::element_wise::PassThrough,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
M_Warp,
|
||||
N_Warp,
|
||||
M_Warp_Tile,
|
||||
N_Warp_Tile,
|
||||
K_Warp_Tile,
|
||||
UniversalGemmProblem::TransposeC,
|
||||
memory_operation>>;
|
||||
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKargs(gemm_descs);
|
||||
EXPECT_TRUE(Kernel::IsSupportedArgument(kargs));
|
||||
@@ -327,7 +273,7 @@ class TestCkTileGroupedGemmPreshuffle : public ::testing::Test
|
||||
hipMemcpyHostToDevice,
|
||||
s.stream_id_));
|
||||
|
||||
ave_time = ck_tile::launch_kernel(
|
||||
return ck_tile::launch_kernel(
|
||||
s,
|
||||
ck_tile::make_kernel<kBlockPerCu>(
|
||||
Kernel{},
|
||||
@@ -336,25 +282,18 @@ class TestCkTileGroupedGemmPreshuffle : public ::testing::Test
|
||||
0,
|
||||
ck_tile::cast_pointer_to_constant_address_space(kargs_ptr),
|
||||
gemm_descs.size()));
|
||||
return ave_time;
|
||||
};
|
||||
|
||||
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
|
||||
if(gemm_descs[0].k_batch == 1)
|
||||
{
|
||||
Run(has_hot_loop_,
|
||||
tail_number_,
|
||||
ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::set>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
// EXPECT TO FAIL because splitk is not supported
|
||||
EXPECT_FALSE(true);
|
||||
}
|
||||
};
|
||||
|
||||
BaseGemmPipeline::TailHandler(RunSplitk, has_hot_loop, tail_num);
|
||||
if(gemm_descs[0].k_batch == 1)
|
||||
{
|
||||
Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
|
||||
ck_tile::memory_operation_enum::set>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
// EXPECT TO FAIL because splitk is not supported
|
||||
EXPECT_FALSE(true);
|
||||
}
|
||||
}
|
||||
|
||||
public:
|
||||
|
||||
Reference in New Issue
Block a user