mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-11 17:00:18 +00:00
[CK_TILE] Switch to universal gemm for batched and grouped gemms (#1919)
* switch to universal gemm for batched and grouped gemms * added reviewer comments * fixed grouped gemm tests
This commit is contained in:
jakpiase
@@ -18,16 +18,42 @@
|
||||
template <typename ALayout, typename BLayout, typename CLayout>
|
||||
float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stream_config& s)
|
||||
{
|
||||
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
|
||||
constexpr bool kPadM = false;
|
||||
constexpr bool kPadN = false;
|
||||
constexpr bool kPadK = false;
|
||||
|
||||
constexpr int kBlockPerCu = 1;
|
||||
|
||||
// This part comes from the Codegen
|
||||
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
|
||||
// Memory friendly for Interwave scheduler
|
||||
constexpr ck_tile::index_t M_Tile = 128;
|
||||
constexpr ck_tile::index_t N_Tile = 128;
|
||||
constexpr ck_tile::index_t N_Tile = 32;
|
||||
constexpr ck_tile::index_t K_Tile = 64;
|
||||
|
||||
constexpr ck_tile::index_t M_Warp = 4;
|
||||
constexpr ck_tile::index_t N_Warp = 1;
|
||||
constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t K_Warp_Tile = 8;
|
||||
|
||||
constexpr bool DoubleSmemBuffer = false;
|
||||
#endif
|
||||
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
|
||||
// Compute friendly for Intrawave scheduler
|
||||
constexpr ck_tile::index_t M_Tile = 256;
|
||||
constexpr ck_tile::index_t N_Tile = 256;
|
||||
constexpr ck_tile::index_t K_Tile = 64;
|
||||
|
||||
constexpr ck_tile::index_t M_Warp = 2;
|
||||
constexpr ck_tile::index_t N_Warp = 2;
|
||||
constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t K_Warp_Tile = 16;
|
||||
|
||||
constexpr bool DoubleSmemBuffer = false;
|
||||
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
|
||||
// Compute friendly for Intrawave scheduler
|
||||
// Using the ping pong reader in the lds level
|
||||
constexpr ck_tile::index_t M_Tile = 256;
|
||||
constexpr ck_tile::index_t N_Tile = 256;
|
||||
constexpr ck_tile::index_t K_Tile = 32;
|
||||
|
||||
constexpr ck_tile::index_t M_Warp = 2;
|
||||
@@ -36,61 +62,232 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
|
||||
|
||||
constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
constexpr ck_tile::index_t K_Warp_Tile = 8;
|
||||
constexpr ck_tile::index_t K_Warp_Tile = 16;
|
||||
|
||||
using CodegenGemmShape =
|
||||
constexpr bool DoubleSmemBuffer = true;
|
||||
#endif
|
||||
|
||||
constexpr bool kPadM = false;
|
||||
constexpr bool kPadN = false;
|
||||
constexpr bool kPadK = false;
|
||||
|
||||
constexpr bool TransposeC = false;
|
||||
|
||||
constexpr int kBlockPerCu = 1;
|
||||
constexpr ck_tile::index_t TileParitionerGroupNum = 8;
|
||||
constexpr ck_tile::index_t TileParitionerM01 = 4;
|
||||
|
||||
using GemmShape =
|
||||
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
|
||||
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
|
||||
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
|
||||
using TilePartitioner = ck_tile::
|
||||
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
|
||||
|
||||
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
|
||||
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
|
||||
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<kPadM,
|
||||
kPadN,
|
||||
kPadK,
|
||||
DoubleSmemBuffer,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout,
|
||||
TransposeC>;
|
||||
using GemmPipelineProblem =
|
||||
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
|
||||
|
||||
using CodegenGemmTraits =
|
||||
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
|
||||
using CodegenPipelineProblem = ck_tile::
|
||||
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
|
||||
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<
|
||||
ck_tile::CShuffleEpilogueProblem<AccDataType,
|
||||
CDataType,
|
||||
CLayout,
|
||||
CodegenPipelineProblem::kBlockSize,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
M_Warp,
|
||||
N_Warp,
|
||||
M_Warp_Tile,
|
||||
N_Warp_Tile,
|
||||
K_Warp_Tile,
|
||||
CodegenPipelineProblem::TransposeC>>;
|
||||
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
|
||||
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
|
||||
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
|
||||
using BaseGemmPipeline = UNIVERSAL_GEMM_PIPELINE<GemmPipelineProblem>;
|
||||
|
||||
auto kargs = Kernel::MakeKernelArgs(args);
|
||||
const ck_tile::index_t k_grain = args.k_batch * K_Tile;
|
||||
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * 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);
|
||||
|
||||
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch, args.batch_count);
|
||||
constexpr dim3 blocks = Kernel::BlockSize();
|
||||
float ave_time{0};
|
||||
|
||||
if(!Kernel::IsSupportedArgument(kargs))
|
||||
const auto Run = [&](const auto has_hot_loop_, const auto tail_number_) {
|
||||
constexpr bool has_hot_loop_v = has_hot_loop_.value;
|
||||
constexpr auto tail_number_v = tail_number_.value;
|
||||
constexpr auto scheduler = GEMM_PIPELINE_SCHEDULER;
|
||||
|
||||
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,
|
||||
M_Warp,
|
||||
N_Warp,
|
||||
M_Warp_Tile,
|
||||
N_Warp_Tile,
|
||||
K_Warp_Tile,
|
||||
UniversalGemmProblem::TransposeC>>;
|
||||
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKernelArgs(args);
|
||||
|
||||
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch, args.batch_count);
|
||||
constexpr 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: " << GemmPipelineProblem::GetName() << '\n'
|
||||
<< "pipeline: " << GemmPipeline::GetName() << '\n'
|
||||
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
|
||||
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
|
||||
<< std::endl;
|
||||
}
|
||||
|
||||
ave_time = ck_tile::launch_kernel(
|
||||
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
|
||||
return ave_time;
|
||||
};
|
||||
|
||||
if(has_hot_loop)
|
||||
{
|
||||
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
|
||||
}
|
||||
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
|
||||
if(tail_num == ck_tile::TailNumber::Full)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
|
||||
}
|
||||
else if(tail_num == ck_tile::TailNumber::Odd)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Odd>{});
|
||||
}
|
||||
else if(tail_num == ck_tile::TailNumber::Even)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Even>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
std::ostringstream err;
|
||||
err << "Incorrect tail_num for compv3 pipeline! Expected Full, Odd or Even, but got "
|
||||
<< tail_num << "\nPrefetchStages: " << BaseGemmPipeline::PrefetchStages
|
||||
<< "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
|
||||
throw std::runtime_error(err.str());
|
||||
}
|
||||
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
|
||||
// Tail pipeline One to Seven
|
||||
if(tail_num == ck_tile::TailNumber::One)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::One>{});
|
||||
}
|
||||
else if(tail_num == ck_tile::TailNumber::Full)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
|
||||
}
|
||||
|
||||
if(s.log_level_ > 0)
|
||||
if constexpr(BaseGemmPipeline::PrefetchStages > 2)
|
||||
{
|
||||
if(tail_num == ck_tile::TailNumber::Two)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Two>{});
|
||||
}
|
||||
}
|
||||
if constexpr(BaseGemmPipeline::PrefetchStages > 3)
|
||||
{
|
||||
if(tail_num == ck_tile::TailNumber::Three)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Three>{});
|
||||
}
|
||||
}
|
||||
if constexpr(BaseGemmPipeline::PrefetchStages > 4)
|
||||
{
|
||||
if(tail_num == ck_tile::TailNumber::Four)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Four>{});
|
||||
}
|
||||
}
|
||||
if constexpr(BaseGemmPipeline::PrefetchStages > 5)
|
||||
{
|
||||
if(tail_num == ck_tile::TailNumber::Five)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Five>{});
|
||||
}
|
||||
}
|
||||
if constexpr(BaseGemmPipeline::PrefetchStages > 6)
|
||||
{
|
||||
if(tail_num == ck_tile::TailNumber::Six)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Six>{});
|
||||
}
|
||||
}
|
||||
if constexpr(BaseGemmPipeline::PrefetchStages > 7)
|
||||
{
|
||||
if(tail_num == ck_tile::TailNumber::Seven)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Seven>{});
|
||||
}
|
||||
}
|
||||
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
|
||||
if(tail_num == ck_tile::TailNumber::Three)
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Three>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
Run(ck_tile::bool_constant<true>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Two>{});
|
||||
}
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "Launching kernel with args: " << Kernel::GetName() << '\n'
|
||||
<< "shape: " << CodegenGemmShape::GetName() << '\n'
|
||||
<< "problem: " << CodegenPipelineProblem::GetName() << '\n'
|
||||
<< "pipeline: " << CodegenGemmPipeline::GetName() << '\n'
|
||||
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
|
||||
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
|
||||
<< std::endl;
|
||||
if(tail_num == ck_tile::TailNumber::Full)
|
||||
{
|
||||
Run(ck_tile::bool_constant<false>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
|
||||
}
|
||||
else if(tail_num == ck_tile::TailNumber::Odd)
|
||||
{
|
||||
Run(ck_tile::bool_constant<false>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Odd>{});
|
||||
}
|
||||
else if(tail_num == ck_tile::TailNumber::Even)
|
||||
{
|
||||
Run(ck_tile::bool_constant<false>{},
|
||||
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Odd>{});
|
||||
}
|
||||
std::ostringstream err;
|
||||
err << "Incorrect tail_num for pipeline without hotloop, expected Full, Odd or Even, but "
|
||||
"got "
|
||||
<< tail_num << "\n PrefetchStages: " << BaseGemmPipeline::PrefetchStages
|
||||
<< "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
|
||||
throw std::runtime_error(err.str());
|
||||
}
|
||||
|
||||
float ave_time = ck_tile::launch_kernel(
|
||||
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
|
||||
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user