mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 13:11:25 +00:00
Ck tile grouped GEMM example (#1713)
* Ck-tile, impl. grouped gemm * Workspace is allocated by user, and is passed to the function * Prepare test to new api design * Unify GemTransKernelArgs, removing N0 param * Add 1 to dim3 in paritioner * Typo: gem - > gemm --------- Co-authored-by: Adam Osewski <19374865+aosewski@users.noreply.github.com>
This commit is contained in:
Mateusz Ozga
@@ -25,6 +25,7 @@
|
||||
#include "ck_tile/ops/gemm/block/block_universal_gemm_as_bs_cr.hpp"
|
||||
#include "ck_tile/ops/gemm/kernel/gemm_kernel.hpp"
|
||||
#include "ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp"
|
||||
#include "ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp"
|
||||
#include "ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp"
|
||||
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp"
|
||||
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp"
|
||||
|
||||
@@ -35,4 +35,40 @@ struct GemmTilePartitioner
|
||||
return make_tuple(iM, iN);
|
||||
}
|
||||
};
|
||||
|
||||
template <typename BlockGemmShape_>
|
||||
struct GemmTile1DPartitioner
|
||||
{
|
||||
using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
|
||||
|
||||
static constexpr index_t MPerBlock = BlockGemmShape::kM;
|
||||
static constexpr index_t NPerBlock = BlockGemmShape::kN;
|
||||
static constexpr index_t KPerBlock = BlockGemmShape::kK;
|
||||
|
||||
CK_TILE_HOST static constexpr auto GridSize(index_t M, index_t N)
|
||||
{
|
||||
index_t GridDimX = (M + MPerBlock - 1) / MPerBlock;
|
||||
index_t GridDimY = (N + NPerBlock - 1) / NPerBlock;
|
||||
return dim3(GridDimX * GridDimY, 1, 1);
|
||||
}
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetNBlock(index_t N)
|
||||
{
|
||||
return integer_divide_ceil(N, NPerBlock);
|
||||
}
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetLoopNum(index_t K)
|
||||
{
|
||||
return integer_divide_ceil(K, KPerBlock);
|
||||
}
|
||||
|
||||
CK_TILE_DEVICE auto operator()(index_t blockOffset, index_t NBlockSize)
|
||||
{
|
||||
index_t iM = __builtin_amdgcn_readfirstlane((blockIdx.x - blockOffset) /
|
||||
GetNBlock(NBlockSize) * MPerBlock);
|
||||
index_t iN = __builtin_amdgcn_readfirstlane((blockIdx.x - blockOffset) %
|
||||
GetNBlock(NBlockSize) * NPerBlock);
|
||||
return make_tuple(iM, iN);
|
||||
}
|
||||
};
|
||||
} // namespace ck_tile
|
||||
|
||||
310
include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp
Normal file
310
include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp
Normal file
@@ -0,0 +1,310 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
#include <string>
|
||||
|
||||
#include "ck_tile/core/numeric/math.hpp"
|
||||
#include "ck_tile/core/utility/literals.hpp"
|
||||
#include "ck_tile/core/utility/amd_address_space.hpp"
|
||||
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/common.hpp"
|
||||
#include "ck_tile/host.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
struct GroupedGemmHostArgs
|
||||
{
|
||||
const void* a_ptr;
|
||||
const void* b_ptr;
|
||||
void* c_ptr;
|
||||
index_t M;
|
||||
index_t N;
|
||||
index_t K;
|
||||
index_t stride_A;
|
||||
index_t stride_B;
|
||||
index_t stride_C;
|
||||
};
|
||||
|
||||
template <typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
|
||||
struct GroupedGemmKernel
|
||||
{
|
||||
using TilePartitioner = remove_cvref_t<TilePartitioner_>;
|
||||
using GemmPipeline = remove_cvref_t<GemmPipeline_>;
|
||||
using EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>;
|
||||
using ALayout = remove_cvref_t<typename GemmPipeline::ALayout>;
|
||||
using BLayout = remove_cvref_t<typename GemmPipeline::BLayout>;
|
||||
using CLayout = remove_cvref_t<typename GemmPipeline::CLayout>;
|
||||
static constexpr index_t KernelBlockSize = GemmPipeline::BlockSize;
|
||||
|
||||
using ADataType = remove_cvref_t<typename GemmPipeline::ADataType>;
|
||||
using BDataType = remove_cvref_t<typename GemmPipeline::BDataType>;
|
||||
using CDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>;
|
||||
|
||||
struct GemmTransKernelArg
|
||||
{
|
||||
GroupedGemmHostArgs group_karg;
|
||||
ck_tile::index_t block_start;
|
||||
ck_tile::index_t block_end;
|
||||
|
||||
GemmTransKernelArg() = default;
|
||||
GemmTransKernelArg(GroupedGemmHostArgs&& karg, index_t bl_start, index_t bl_end)
|
||||
: group_karg{karg}, block_start{bl_start}, block_end{bl_end}
|
||||
{
|
||||
}
|
||||
};
|
||||
|
||||
__host__ static size_t GetWorkSpaceSize(const std::vector<GroupedGemmHostArgs>& gemm_descs)
|
||||
{
|
||||
return gemm_descs.size() * sizeof(GemmTransKernelArg);
|
||||
}
|
||||
|
||||
__host__ static constexpr auto BlockSize() { return dim3(KernelBlockSize); }
|
||||
|
||||
using Hargs = GroupedGemmHostArgs;
|
||||
|
||||
__host__ static constexpr auto GridSize(const std::vector<Hargs>& gemm_descs)
|
||||
{
|
||||
index_t grid_size = 0;
|
||||
for(const auto& it_desc : gemm_descs)
|
||||
{
|
||||
const auto dim3 = TilePartitioner::GridSize(it_desc.M, it_desc.N);
|
||||
grid_size += dim3.x * dim3.y * 1;
|
||||
}
|
||||
return dim3(grid_size, 1, 1);
|
||||
}
|
||||
|
||||
CK_TILE_HOST static auto MakeKargs(const std::vector<Hargs>& gemm_descs)
|
||||
{
|
||||
std::vector<GemmTransKernelArg> gemm_kernel_args_;
|
||||
index_t group_count = ck_tile::type_convert<ck_tile::index_t>(gemm_descs.size());
|
||||
index_t grid_size = 0;
|
||||
gemm_kernel_args_.reserve(group_count);
|
||||
|
||||
for(std::size_t i = 0; i < gemm_descs.size(); ++i)
|
||||
{
|
||||
const index_t M = gemm_descs[i].M;
|
||||
const index_t N = gemm_descs[i].N;
|
||||
const index_t K = gemm_descs[i].K;
|
||||
|
||||
if(M == 0 || N == 0 || K == 0)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
const index_t stride_a = gemm_descs[i].stride_A;
|
||||
const index_t stride_b = gemm_descs[i].stride_B;
|
||||
const index_t stride_c = gemm_descs[i].stride_C;
|
||||
|
||||
const auto dim3 = TilePartitioner::GridSize(M, N);
|
||||
const index_t grid_size_grp = dim3.x * 1 * 1;
|
||||
|
||||
const index_t block_start = grid_size;
|
||||
const index_t block_end = grid_size + grid_size_grp;
|
||||
|
||||
grid_size += grid_size_grp;
|
||||
|
||||
auto karg = GroupedGemmHostArgs{type_convert<const ADataType*>(gemm_descs[i].a_ptr),
|
||||
type_convert<const BDataType*>(gemm_descs[i].b_ptr),
|
||||
type_convert<CDataType*>(gemm_descs[i].c_ptr),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_a,
|
||||
stride_b,
|
||||
stride_c};
|
||||
|
||||
gemm_kernel_args_.emplace_back(std::move(karg), block_start, block_end);
|
||||
}
|
||||
|
||||
return gemm_kernel_args_;
|
||||
}
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
|
||||
{
|
||||
return max(GemmPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
|
||||
}
|
||||
|
||||
CK_TILE_DEVICE void Run(const Hargs& kargs, const index_t block_start) const
|
||||
{
|
||||
const auto [i_m, i_n] = TilePartitioner{}(block_start, kargs.N);
|
||||
// options
|
||||
const ADataType* a_start = static_cast<const ADataType*>(kargs.a_ptr);
|
||||
const BDataType* b_start = static_cast<const BDataType*>(kargs.b_ptr);
|
||||
// Convert pointers to tensor views
|
||||
auto a_tensor_view = [&]() {
|
||||
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return make_naive_tensor_view<address_space_enum::global>(
|
||||
a_start,
|
||||
make_tuple(kargs.M, kargs.K),
|
||||
make_tuple(kargs.stride_A, 1),
|
||||
number<GemmPipeline::VectorSizeA>{},
|
||||
number<1>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return make_naive_tensor_view<address_space_enum::global>(
|
||||
a_start,
|
||||
make_tuple(kargs.M, kargs.K),
|
||||
make_tuple(1, kargs.stride_A),
|
||||
number<1>{},
|
||||
number<1>{});
|
||||
}
|
||||
}();
|
||||
|
||||
auto b_tensor_view = [&]() {
|
||||
if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return make_naive_tensor_view<address_space_enum::global>(
|
||||
b_start,
|
||||
make_tuple(kargs.N, kargs.K),
|
||||
make_tuple(1, kargs.stride_B),
|
||||
number<1>{},
|
||||
number<1>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return make_naive_tensor_view<address_space_enum::global>(
|
||||
b_start,
|
||||
make_tuple(kargs.N, kargs.K),
|
||||
make_tuple(kargs.stride_B, 1),
|
||||
number<GemmPipeline::VectorSizeB>{},
|
||||
number<1>{});
|
||||
}
|
||||
}();
|
||||
|
||||
auto a_pad_view = [&]() {
|
||||
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return pad_tensor_view(a_tensor_view,
|
||||
make_tuple(number<TilePartitioner::MPerBlock>{},
|
||||
number<TilePartitioner::KPerBlock>{}),
|
||||
sequence<false, GemmPipeline::kPadK>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return pad_tensor_view(a_tensor_view,
|
||||
make_tuple(number<TilePartitioner::MPerBlock>{},
|
||||
number<TilePartitioner::KPerBlock>{}),
|
||||
sequence<GemmPipeline::kPadM, false>{});
|
||||
}
|
||||
}();
|
||||
// clang-format on
|
||||
|
||||
auto a_block_window = make_tile_window(
|
||||
a_pad_view,
|
||||
make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::KPerBlock>{}),
|
||||
{i_m, 0});
|
||||
|
||||
auto b_pad_view = [&]() {
|
||||
if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
|
||||
{
|
||||
return pad_tensor_view(b_tensor_view,
|
||||
make_tuple(number<TilePartitioner::NPerBlock>{},
|
||||
number<TilePartitioner::KPerBlock>{}),
|
||||
sequence<false, GemmPipeline::kPadK>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return pad_tensor_view(b_tensor_view,
|
||||
make_tuple(number<TilePartitioner::NPerBlock>{},
|
||||
number<TilePartitioner::KPerBlock>{}),
|
||||
sequence<GemmPipeline::kPadN, false>{});
|
||||
}
|
||||
}();
|
||||
|
||||
auto b_block_window = make_tile_window(
|
||||
b_pad_view,
|
||||
make_tuple(number<TilePartitioner::NPerBlock>{}, number<TilePartitioner::KPerBlock>{}),
|
||||
{i_n, 0});
|
||||
|
||||
// allocate LDS
|
||||
__shared__ char smem_ptr[GetSmemSize()];
|
||||
|
||||
const index_t num_loop = TilePartitioner::GetLoopNum(kargs.K);
|
||||
|
||||
// Run GEMM cooperatively by whole wokrgroup.
|
||||
auto c_block_tile =
|
||||
GemmPipeline{}.template operator()(a_block_window, b_block_window, num_loop, smem_ptr);
|
||||
|
||||
CDataType* c_start = static_cast<CDataType*>(kargs.c_ptr);
|
||||
auto c_tensor_view = [&]() {
|
||||
if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return make_naive_tensor_view<address_space_enum::global>(
|
||||
c_start,
|
||||
make_tuple(kargs.M, kargs.N),
|
||||
make_tuple(kargs.stride_C, 1),
|
||||
number<GemmPipeline::VectorSizeC>{},
|
||||
number<1>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return make_naive_tensor_view<address_space_enum::global>(
|
||||
c_start,
|
||||
make_tuple(kargs.M, kargs.N),
|
||||
make_tuple(1, kargs.stride_C),
|
||||
number<1>{},
|
||||
number<1>{});
|
||||
}
|
||||
}();
|
||||
|
||||
auto c_pad_view = [&]() {
|
||||
if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
|
||||
{
|
||||
return pad_tensor_view(c_tensor_view,
|
||||
make_tuple(number<TilePartitioner::MPerBlock>{},
|
||||
number<TilePartitioner::NPerBlock>{}),
|
||||
sequence<false, GemmPipeline::kPadN>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
return pad_tensor_view(c_tensor_view,
|
||||
make_tuple(number<TilePartitioner::MPerBlock>{},
|
||||
number<TilePartitioner::NPerBlock>{}),
|
||||
sequence<GemmPipeline::kPadM, false>{});
|
||||
}
|
||||
}();
|
||||
auto CBlockWindow_pad = make_tile_window(
|
||||
c_pad_view,
|
||||
make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::NPerBlock>{}),
|
||||
{i_m, i_n});
|
||||
|
||||
EpiloguePipeline{}(CBlockWindow_pad, c_block_tile);
|
||||
}
|
||||
|
||||
CK_TILE_DEVICE void operator()(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
|
||||
int group_count) const
|
||||
{
|
||||
const index_t block_id = ck_tile::get_block_1d_id();
|
||||
const auto gemm_desc_ptr = reinterpret_cast<const GemmTransKernelArg*>(
|
||||
cast_pointer_to_generic_address_space(gemm_descs_const));
|
||||
|
||||
index_t left = 0;
|
||||
index_t right = group_count;
|
||||
index_t group_id = index_t((left + right) / 2);
|
||||
|
||||
while((!(block_id >= gemm_desc_ptr[group_id].block_start &&
|
||||
block_id < gemm_desc_ptr[group_id].block_end)) &&
|
||||
left <= right)
|
||||
{
|
||||
if(block_id < gemm_desc_ptr[group_id].block_start)
|
||||
{
|
||||
right = group_id;
|
||||
}
|
||||
else
|
||||
{
|
||||
left = group_id;
|
||||
}
|
||||
group_id = index_t((left + right) / 2);
|
||||
}
|
||||
|
||||
Run(gemm_desc_ptr[group_id].group_karg, gemm_desc_ptr[group_id].block_start);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
Reference in New Issue
Block a user