mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 21:21:22 +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
@@ -16,85 +16,9 @@
|
||||
#include "ck_tile/host.hpp"
|
||||
#include "grouped_gemm.hpp"
|
||||
|
||||
namespace {
|
||||
|
||||
struct GroupedGemmKernelParam
|
||||
{
|
||||
static const bool kPadM = false;
|
||||
static const bool kPadN = false;
|
||||
static const bool kPadK = false;
|
||||
|
||||
static const int kBlockPerCu = 1;
|
||||
static const ck_tile::index_t M_Tile = 128;
|
||||
static const ck_tile::index_t N_Tile = 128;
|
||||
static const ck_tile::index_t K_Tile = 32;
|
||||
|
||||
static const ck_tile::index_t M_Warp = 2;
|
||||
static const ck_tile::index_t N_Warp = 2;
|
||||
static const ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static const ck_tile::index_t M_Warp_Tile = 32;
|
||||
static const ck_tile::index_t N_Warp_Tile = 32;
|
||||
static const ck_tile::index_t K_Warp_Tile = 8;
|
||||
};
|
||||
|
||||
using CodegenGemmShape =
|
||||
ck_tile::TileGemmShape<ck_tile::sequence<GroupedGemmKernelParam::M_Tile,
|
||||
GroupedGemmKernelParam::N_Tile,
|
||||
GroupedGemmKernelParam::K_Tile>,
|
||||
ck_tile::sequence<GroupedGemmKernelParam::M_Warp,
|
||||
GroupedGemmKernelParam::N_Warp,
|
||||
GroupedGemmKernelParam::K_Warp>,
|
||||
ck_tile::sequence<GroupedGemmKernelParam::M_Warp_Tile,
|
||||
GroupedGemmKernelParam::N_Warp_Tile,
|
||||
GroupedGemmKernelParam::K_Warp_Tile>>;
|
||||
|
||||
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
|
||||
|
||||
template <typename ALayout, typename BLayout, typename CLayout>
|
||||
using CodegenGemmTraits = ck_tile::TileGemmTraits<GroupedGemmKernelParam::kPadM,
|
||||
GroupedGemmKernelParam::kPadN,
|
||||
GroupedGemmKernelParam::kPadK,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>;
|
||||
|
||||
template <typename ALayout, typename BLayout, typename CLayout>
|
||||
using CodegenPipelineProblem =
|
||||
ck_tile::GemmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CodegenGemmShape,
|
||||
CodegenGemmTraits<ALayout, BLayout, CLayout>>;
|
||||
|
||||
template <typename ALayout, typename BLayout, typename CLayout>
|
||||
using CodegenGemmPipeline =
|
||||
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>>;
|
||||
|
||||
template <typename ALayout, typename BLayout, typename CLayout>
|
||||
using GemmEpilogue = ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<
|
||||
AccDataType,
|
||||
CDataType,
|
||||
CLayout,
|
||||
CodegenPipelineProblem<ALayout, BLayout, CLayout>::kBlockSize,
|
||||
TilePartitioner::MPerBlock,
|
||||
TilePartitioner::NPerBlock,
|
||||
GroupedGemmKernelParam::M_Warp,
|
||||
GroupedGemmKernelParam::N_Warp,
|
||||
GroupedGemmKernelParam::M_Warp_Tile,
|
||||
GroupedGemmKernelParam::N_Warp_Tile,
|
||||
GroupedGemmKernelParam::K_Warp_Tile,
|
||||
CodegenPipelineProblem<ALayout, BLayout, CLayout>::TransposeC>>;
|
||||
|
||||
template <typename ALayout, typename BLayout, typename CLayout>
|
||||
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner,
|
||||
CodegenGemmPipeline<ALayout, BLayout, CLayout>,
|
||||
GemmEpilogue<ALayout, BLayout, CLayout>>;
|
||||
}; // namespace
|
||||
|
||||
std::size_t get_workspace_size(const std::vector<grouped_gemm_kargs>& gemm_descs)
|
||||
{
|
||||
return ::Kernel<std::nullptr_t, std::nullptr_t, std::nullptr_t>::GetWorkSpaceSize(gemm_descs);
|
||||
return gemm_descs.size() * sizeof(ck_tile::GemmTransKernelArg);
|
||||
}
|
||||
|
||||
template <typename ALayout, typename BLayout, typename CLayout>
|
||||
@@ -102,37 +26,265 @@ float grouped_gemm(const std::vector<grouped_gemm_kargs>& gemm_descs,
|
||||
const ck_tile::stream_config& s,
|
||||
void* p_workspace_)
|
||||
{
|
||||
using GroupedGemmKernel = ::Kernel<ALayout, BLayout, CLayout>;
|
||||
#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 = 32;
|
||||
constexpr ck_tile::index_t K_Tile = 64;
|
||||
|
||||
auto arguments = GroupedGemmKernel::MakeKargs(gemm_descs);
|
||||
constexpr ck_tile::index_t M_Warp = 4;
|
||||
constexpr ck_tile::index_t N_Warp = 1;
|
||||
constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
const dim3 grids = GroupedGemmKernel::GridSize(gemm_descs);
|
||||
constexpr dim3 blocks = GroupedGemmKernel::BlockSize();
|
||||
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;
|
||||
|
||||
ck_tile::hip_check_error(hipMemcpyWithStream(
|
||||
p_workspace_,
|
||||
arguments.data(),
|
||||
arguments.size() * sizeof(typename GroupedGemmKernel::GemmTransKernelArg),
|
||||
hipMemcpyHostToDevice,
|
||||
s.stream_id_));
|
||||
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;
|
||||
|
||||
if(s.log_level_ > 0)
|
||||
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;
|
||||
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 = 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 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 BaseGemmPipeline = UNIVERSAL_GEMM_PIPELINE<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 = TilePartitioner::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_) {
|
||||
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::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
|
||||
auto kargs = Kernel::MakeKargs(gemm_descs);
|
||||
|
||||
const dim3 grids = Kernel::GridSize(gemm_descs);
|
||||
constexpr dim3 blocks = Kernel::BlockSize();
|
||||
|
||||
ck_tile::hip_check_error(hipMemcpyWithStream(p_workspace_,
|
||||
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;
|
||||
}
|
||||
|
||||
ave_time = ck_tile::launch_kernel(
|
||||
s,
|
||||
ck_tile::make_kernel<blocks.x, kBlockPerCu>(
|
||||
Kernel{},
|
||||
grids,
|
||||
blocks,
|
||||
0,
|
||||
ck_tile::cast_pointer_to_constant_address_space(p_workspace_),
|
||||
gemm_descs.size()));
|
||||
return ave_time;
|
||||
};
|
||||
|
||||
if(has_hot_loop)
|
||||
{
|
||||
std::cout << "Launching kernel: " << GroupedGemmKernel::GetName() << " with args:"
|
||||
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
|
||||
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
|
||||
<< std::endl;
|
||||
#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 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::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, GroupedGemmKernelParam::kBlockPerCu>(
|
||||
GroupedGemmKernel{},
|
||||
grids,
|
||||
blocks,
|
||||
0,
|
||||
ck_tile::cast_pointer_to_constant_address_space(p_workspace_),
|
||||
gemm_descs.size()));
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user