mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-16 19:09:59 +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
[ROCm/composable_kernel commit: 0e91d32c61]
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;
|
||||
}
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
@@ -9,6 +9,30 @@
|
||||
#include "ck_tile/host/kernel_launch.hpp"
|
||||
#include "ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp"
|
||||
|
||||
#define CK_TILE_PIPELINE_COMPUTE_V3 1
|
||||
#define CK_TILE_PIPELINE_MEMORY 2
|
||||
#define CK_TILE_PIPELINE_COMPUTE_V4 3
|
||||
|
||||
#ifndef CK_TILE_PIPELINE_DEFAULT
|
||||
#define CK_TILE_PIPELINE_DEFAULT CK_TILE_PIPELINE_COMPUTE_V3
|
||||
#endif
|
||||
|
||||
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
|
||||
#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrMem
|
||||
#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrMem
|
||||
#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Interwave
|
||||
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
|
||||
#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrCompV3
|
||||
#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrCompV3
|
||||
#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Intrawave
|
||||
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
|
||||
#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrCompV4
|
||||
#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrCompV4
|
||||
#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Intrawave
|
||||
#else
|
||||
#error "unsupported CK_TILE_PIPELINE_DEFAULT value"
|
||||
#endif
|
||||
|
||||
template <typename DataType>
|
||||
struct BatchedGemmTypeConfig;
|
||||
|
||||
@@ -32,19 +56,19 @@ using CDataType = Types::CDataType;
|
||||
auto create_args(int argc, char* argv[])
|
||||
{
|
||||
ck_tile::ArgParser arg_parser;
|
||||
arg_parser.insert("m", "256", "m dimension")
|
||||
.insert("n", "128", "n dimension")
|
||||
.insert("k", "128", "k dimension")
|
||||
arg_parser.insert("m", "512", "m dimension")
|
||||
.insert("n", "1024", "n dimension")
|
||||
.insert("k", "2048", "k dimension")
|
||||
.insert("stride_a", "0", "Tensor A stride")
|
||||
.insert("stride_b", "0", "Tensor B stride")
|
||||
.insert("stride_c", "0", "Tensor C stride")
|
||||
.insert("a_layout", "R", "A tensor data layout - Row by default")
|
||||
.insert("b_layout", "C", "B tensor data layout - Row by default")
|
||||
.insert("c_layout", "R", "C tensor data layout - Row by default")
|
||||
.insert("batch_stride_a", "32768", "Batch A stride")
|
||||
.insert("batch_stride_b", "16384", "Batch B stride")
|
||||
.insert("batch_stride_c", "32768", "Batch C stride")
|
||||
.insert("batch_count", "16", "Batch count")
|
||||
.insert("batch_stride_a", "1048576", "Batch A stride")
|
||||
.insert("batch_stride_b", "2097152", "Batch B stride")
|
||||
.insert("batch_stride_c", "524288", "Batch C stride")
|
||||
.insert("batch_count", "8", "Batch count")
|
||||
.insert("v", "2", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
|
||||
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
|
||||
.insert("warmup", "50", "number of iterations before benchmark the kernel")
|
||||
|
||||
@@ -185,7 +185,6 @@ int run_batched_gemm_example_with_layouts(int argc,
|
||||
kbatch,
|
||||
n_warmup,
|
||||
n_repeat);
|
||||
|
||||
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
|
||||
bool pass = true;
|
||||
|
||||
|
||||
@@ -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;
|
||||
}
|
||||
|
||||
|
||||
@@ -9,6 +9,30 @@
|
||||
#include "ck_tile/host/kernel_launch.hpp"
|
||||
#include "ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp"
|
||||
|
||||
#define CK_TILE_PIPELINE_COMPUTE_V3 1
|
||||
#define CK_TILE_PIPELINE_MEMORY 2
|
||||
#define CK_TILE_PIPELINE_COMPUTE_V4 3
|
||||
|
||||
#ifndef CK_TILE_PIPELINE_DEFAULT
|
||||
#define CK_TILE_PIPELINE_DEFAULT CK_TILE_PIPELINE_COMPUTE_V3
|
||||
#endif
|
||||
|
||||
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
|
||||
#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrMem
|
||||
#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrMem
|
||||
#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Interwave
|
||||
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
|
||||
#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrCompV3
|
||||
#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrCompV3
|
||||
#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Intrawave
|
||||
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
|
||||
#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrCompV4
|
||||
#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrCompV4
|
||||
#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Intrawave
|
||||
#else
|
||||
#error "unsupported CK_TILE_PIPELINE_DEFAULT value"
|
||||
#endif
|
||||
|
||||
template <typename DataType>
|
||||
struct GemmTypeConfig;
|
||||
|
||||
@@ -29,7 +53,7 @@ using BDataType = Types::BDataType;
|
||||
using AccDataType = Types::AccDataType;
|
||||
using CDataType = Types::CDataType;
|
||||
|
||||
using grouped_gemm_kargs = ck_tile::GroupedGemmHostArgs;
|
||||
using grouped_gemm_kargs = ck_tile::GemmHostArgs;
|
||||
|
||||
auto create_args(int argc, char* argv[])
|
||||
{
|
||||
@@ -46,7 +70,7 @@ auto create_args(int argc, char* argv[])
|
||||
.insert("validate", "1", "0. No validation, 1. Validation on CPU.")
|
||||
.insert("warmup", "10", "number of iterations before benchmark the kernel.")
|
||||
.insert("repeat", "100", "number of iterations to benchmark the kernel.")
|
||||
.insert("group_count", "16", "group count.");
|
||||
.insert("group_count", "8", "group count.");
|
||||
|
||||
bool result = arg_parser.parse(argc, argv);
|
||||
return std::make_tuple(result, arg_parser);
|
||||
|
||||
@@ -101,8 +101,8 @@ int run_grouped_gemm_example_with_layouts(int argc,
|
||||
for(int i = 0; i < group_count; i++)
|
||||
{
|
||||
Ms.push_back(256 + 256 * i);
|
||||
Ns.push_back(128 + 128 * i);
|
||||
Ks.push_back(128 + 64 * i);
|
||||
Ns.push_back(256 + 512 * i);
|
||||
Ks.push_back(256 + 64 * i);
|
||||
|
||||
stride_As.push_back(Ks[i]);
|
||||
stride_Bs.push_back(Ks[i]);
|
||||
@@ -169,7 +169,10 @@ int run_grouped_gemm_example_with_layouts(int argc,
|
||||
const void* p_b = b_k_n_dev_buf[i]->GetDeviceBuffer();
|
||||
void* p_c = c_m_n_dev_buf[i]->GetDeviceBuffer();
|
||||
|
||||
gemm_descs.push_back({p_a, p_b, p_c, M, N, K, stride_As[i], stride_Bs[i], stride_Cs[i]});
|
||||
// TODO Add support for kbatch > 1 in grouped gemm
|
||||
static constexpr ck_tile::index_t k_batch = 1;
|
||||
gemm_descs.push_back(
|
||||
{p_a, p_b, p_c, k_batch, M, N, K, stride_As[i], stride_Bs[i], stride_Cs[i]});
|
||||
}
|
||||
|
||||
invoke_gemm<ALayout, BLayout, CLayout>(warmup, repeat, group_count, gemm_descs);
|
||||
|
||||
Reference in New Issue
Block a user