mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-02 04:31:25 +00:00
introducing ck_tile! (#1216)
* enable gfx940
* switch between intrinsic mfma routines on mi100/200 and mi300
* fix mfma_int8 on MI300
* disable 2 int8 examples on MI300
* Update cmake-ck-dev.sh
* restore gitignore file
* modify Jenkinsfile to the internal repo
* Bump rocm-docs-core from 0.24.0 to 0.29.0 in /docs/sphinx
Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.24.0 to 0.29.0.
- [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases)
- [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/RadeonOpenCompute/rocm-docs-core/compare/v0.24.0...v0.29.0)
---
updated-dependencies:
- dependency-name: rocm-docs-core
dependency-type: direct:production
update-type: version-update:semver-minor
...
Signed-off-by: dependabot[bot] <support@github.com>
* initial enablement of gfx950
* fix clang format
* disable examples 31 and 41 int8 on gfx950
* add code
* fix build wip
* fix xx
* now can build
* naming
* minor fix
* wip fix
* fix macro for exp2; fix warpgemm a/b in transposedC
* unify as tuple_array
* Update the required Python version to 3.9
* Update executable name in test scripts
* re-structure tuple/array to avoid spill
* Merge function templates
* Fix format
* Add constraint to array<> ctor
* Re-use function
* Some minor changes
* remove wrong code in store_raw()
* fix compile issue in transpose
* Rename enum
Rename 'cood_transform_enum' to 'coord_transform_enum'
* let more integral_constant->constant, and formating
* make sure thread_buffer can be tuple/array
* temp fix buffer_store spill
* not using custom data type by default, now we can have ISA-level same code as opt_padding
* fix compile error, fp8 not ready now
* fix fp8 duplicated move/shift/and/or problem
* Default use CK_TILE_FLOAT_TO_FP8_STOCHASTIC rounding mode
* fix scratch in fp8 kernel
* update some readme
* fix merge from upstream
* sync with upstream
* sync upstream again
* sync 22
* remove unused
* fix clang-format
* update README of ck_tile example
* fix several issue
* let python version to be 3.8 as minimal
* remove ck_tile example from default cmake target like all/install/check
* remove mistake
* 1).support receipe in generate.py 2).use simplified mask type 3).change left/right to pass into karg
* fix some bug in group-mode masking and codegen. update README
* F8 quantization for FMHA forward (#1224)
* Add SAccElementFunction, PComputeElementFunction, OAccElementFunction in pipeline
* Add element function to fmha api
* Adjust P elementwise function
* Fix bug of elementwise op, our elementwise op is not inout
* Add some elementwise op, prepare to quantization
* Let generate.py can generate different elementwise function
* To prevent compiler issue, remove the elementwise function we have not used.
* Remove f8 pipeline, we should share the same pipeline even in f8
* Remove remove_cvref_t
* Avoid warning
* Fix wrong fp8 QK/KV block gemm setting
* Check fp8 rounding error in check_err()
* Set fp8 rounding error for check_err()
* Use CK_TILE_FLOAT_TO_FP8_STANDARD as default fp8 rounding mode
* 1. codgen the f8 api and kernel
2. f8 host code
* prevent warning in filter mode
* Remove not-in-use elementwise function kargs
* Remove more not-in-use elementwise function kargs
* Small refinements in C++ source files
* Use conditional_t<> to simplify code
* Support heterogeneous argument for binary function types
* Re-use already-existing scales<> functor template
* Fix wrong value produced by saturating
* Generalize the composes<> template
* Unify saturates<> implementation
* Fix type errors in composes<>
* Extend less_equal<>
* Reuse the existing template less_equal<> in check_err()
* Add equal<float> & equal<double>
* Rename check_err() parameter
* Rename check_err() parameter
* Add FIXME comment for adding new macro in future
* Remove unnecessary cast to void
* Eliminate duplicated code
* Avoid dividing api pool into more than 2 groups
* Use more clear variable names
* Use affirmative condition in if stmt
* Remove blank lines
* Donot perfect forwarding in composes<>
* To fix compile error, revert generate.py back to 4439cc107d
* Fix bug of p element function
* Add compute element op to host softmax
* Remove element function in api interface
* Extract user parameter
* Rename pscale and oscale variable
* rename f8 to fp8
* rename more f8 to fp8
* Add pipeline::operator() without element_functor
* 1. Remove deprecated pipeline enum
2. Refine host code parameter
* Use quantization range as input
* 1. Rename max_dtype to dtype_max.
2. Rename scale to scale_s
3.Add init description
* Refine description
* prevent early return
* unify _squant kernel name in cpp, update README
* Adjust the default range.
* Refine error message and bias range
* Add fp8 benchmark and smoke test
* fix fp8 swizzle_factor=4 case
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: carlushuang <carlus.huang@amd.com>
---------
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: illsilin <Illia.Silin@amd.com>
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: Jing Zhang <jizha@amd.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Po-Yen, Chen <PoYen.Chen@amd.com>
Co-authored-by: rocking <ChunYu.Lai@amd.com>
This commit is contained in:
carlushuang
@@ -0,0 +1,200 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// A Tile Window: global memory
|
||||
// B Tile Window: global memory
|
||||
// C Distributed tensor: register
|
||||
template <typename Problem, typename Policy = BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy>
|
||||
struct BlockGemmPipelineAGmemBGmemCRegV1
|
||||
{
|
||||
using ADataType = remove_cvref_t<typename Problem::ADataType>;
|
||||
using BDataType = remove_cvref_t<typename Problem::BDataType>;
|
||||
using CDataType = remove_cvref_t<typename Problem::CDataType>;
|
||||
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
|
||||
|
||||
static constexpr index_t kBlockSize = Problem::kBlockSize;
|
||||
|
||||
static constexpr index_t kMPerBlock = BlockGemmShape::kM;
|
||||
static constexpr index_t kNPerBlock = BlockGemmShape::kN;
|
||||
static constexpr index_t kKPerBlock = BlockGemmShape::kK;
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetStaticLdsSize()
|
||||
{
|
||||
return ck_tile::integer_divide_ceil(
|
||||
sizeof(ADataType) *
|
||||
Policy::template MakeALdsBlockDescriptor<Problem>().get_element_space_size(),
|
||||
16) *
|
||||
16 +
|
||||
sizeof(BDataType) *
|
||||
Policy::template MakeBLdsBlockDescriptor<Problem>().get_element_space_size();
|
||||
}
|
||||
|
||||
template <typename ADramBlockWindowTmp,
|
||||
typename BDramBlockWindowTmp,
|
||||
typename AElementFunction,
|
||||
typename BElementFunction>
|
||||
CK_TILE_HOST_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
|
||||
const AElementFunction& a_element_func,
|
||||
const BDramBlockWindowTmp& b_dram_block_window_tmp,
|
||||
const BElementFunction& b_element_func,
|
||||
index_t num_loop,
|
||||
void* p_smem) const
|
||||
{
|
||||
static_assert(
|
||||
std::is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
|
||||
std::is_same_v<BDataType, remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
|
||||
"wrong!");
|
||||
|
||||
static_assert(kMPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
|
||||
kNPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
|
||||
kKPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[number<1>{}],
|
||||
"wrong!");
|
||||
|
||||
// A tile in LDS
|
||||
ADataType* p_a_lds = static_cast<ADataType*>(p_smem);
|
||||
|
||||
constexpr auto a_lds_block_desc = Policy::template MakeALdsBlockDescriptor<Problem>();
|
||||
|
||||
auto a_lds_block = make_tensor_view<address_space_enum::lds>(p_a_lds, a_lds_block_desc);
|
||||
|
||||
constexpr index_t a_lds_block_space_size_aligned =
|
||||
integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16) *
|
||||
16;
|
||||
|
||||
// B tile in LDS
|
||||
BDataType* p_b_lds = static_cast<BDataType*>(
|
||||
static_cast<void*>(static_cast<char*>(p_smem) + a_lds_block_space_size_aligned));
|
||||
|
||||
constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
|
||||
|
||||
auto b_lds_block = make_tensor_view<address_space_enum::lds>(p_b_lds, b_lds_block_desc);
|
||||
|
||||
// A DRAM tile window for load
|
||||
auto a_copy_dram_window =
|
||||
make_tile_window(a_dram_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}),
|
||||
a_dram_block_window_tmp.get_window_origin(),
|
||||
Policy::template MakeADramTileDistribution<Problem>());
|
||||
|
||||
// A LDS tile window for store
|
||||
auto a_copy_lds_window =
|
||||
make_tile_window(a_lds_block,
|
||||
make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}),
|
||||
{0, 0},
|
||||
a_copy_dram_window.get_tile_distribution());
|
||||
|
||||
// B DRAM tile window for load
|
||||
auto b_copy_dram_window =
|
||||
make_tile_window(b_dram_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}),
|
||||
b_dram_block_window_tmp.get_window_origin(),
|
||||
Policy::template MakeBDramTileDistribution<Problem>());
|
||||
|
||||
// B LDS tile window for store
|
||||
auto b_copy_lds_window =
|
||||
make_tile_window(b_lds_block,
|
||||
make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}),
|
||||
{0, 0},
|
||||
b_copy_dram_window.get_tile_distribution());
|
||||
|
||||
// A LDS tile for block GEMM
|
||||
auto a_lds_gemm_window = make_tile_window(
|
||||
a_lds_block, make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}), {0, 0});
|
||||
|
||||
// B LDS tile for block GEMM
|
||||
auto b_lds_gemm_window = make_tile_window(
|
||||
b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0});
|
||||
|
||||
// Block GEMM
|
||||
constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>();
|
||||
|
||||
// Acc register tile
|
||||
auto c_block_tile = decltype(block_gemm(a_lds_gemm_window, b_lds_gemm_window)){};
|
||||
|
||||
// prefetch
|
||||
// global read 0
|
||||
auto a_block_tile = load_tile(a_copy_dram_window);
|
||||
auto b_block_tile = load_tile(b_copy_dram_window);
|
||||
|
||||
{
|
||||
// move to 1
|
||||
move_tile_window(a_copy_dram_window, {0, kKPerBlock});
|
||||
move_tile_window(b_copy_dram_window, {0, kKPerBlock});
|
||||
|
||||
// initialize C
|
||||
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
|
||||
|
||||
// LDS write 0
|
||||
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
|
||||
store_tile(a_copy_lds_window, a_block_tile_tmp);
|
||||
|
||||
// LDS write 0
|
||||
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
|
||||
store_tile(b_copy_lds_window, b_block_tile_tmp);
|
||||
}
|
||||
|
||||
index_t iCounter = num_loop - 1;
|
||||
|
||||
do
|
||||
{
|
||||
// global read i + 1
|
||||
a_block_tile = load_tile(a_copy_dram_window);
|
||||
b_block_tile = load_tile(b_copy_dram_window);
|
||||
|
||||
block_sync_lds();
|
||||
|
||||
// GEMM i
|
||||
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
|
||||
|
||||
block_sync_lds();
|
||||
|
||||
// move to i + 2
|
||||
move_tile_window(a_copy_dram_window, {0, kKPerBlock});
|
||||
move_tile_window(b_copy_dram_window, {0, kKPerBlock});
|
||||
|
||||
// LDS write i + 1
|
||||
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
|
||||
store_tile(a_copy_lds_window, a_block_tile_tmp);
|
||||
|
||||
// LDS write i + 1
|
||||
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
|
||||
store_tile(b_copy_lds_window, b_block_tile_tmp);
|
||||
|
||||
iCounter--;
|
||||
|
||||
} while(iCounter > 0);
|
||||
|
||||
// tail
|
||||
{
|
||||
block_sync_lds();
|
||||
|
||||
// GEMM num_loop - 1
|
||||
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
|
||||
}
|
||||
|
||||
return c_block_tile;
|
||||
}
|
||||
|
||||
template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
|
||||
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
|
||||
const BDramBlockWindowTmp& b_dram_block_window_tmp,
|
||||
index_t num_loop,
|
||||
void* p_smem) const
|
||||
{
|
||||
return operator()(
|
||||
a_dram_block_window_tmp,
|
||||
[](const ADataType& a) { return a; },
|
||||
b_dram_block_window_tmp,
|
||||
[](const BDataType& b) { return b; },
|
||||
num_loop,
|
||||
p_smem);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,251 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// Default policy for BlockGemmPipelineAGmemBGmemCRegV1
|
||||
// Default policy class should not be templated, put template on member functions instead
|
||||
struct BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy
|
||||
{
|
||||
#if 0
|
||||
// 2d
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
|
||||
{
|
||||
using namespace ck_tile;
|
||||
|
||||
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
constexpr auto a_lds_block_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(kMPerBlock, kKPerBlock), number<32>{});
|
||||
|
||||
return a_lds_block_desc;
|
||||
}
|
||||
|
||||
// 2d
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
|
||||
{
|
||||
using namespace ck_tile;
|
||||
|
||||
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
constexpr auto b_lds_block_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(kNPerBlock, kKPerBlock), number<32>{});
|
||||
|
||||
return b_lds_block_desc;
|
||||
}
|
||||
#elif 1
|
||||
// 3d + padding
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
|
||||
{
|
||||
using namespace ck_tile;
|
||||
|
||||
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
|
||||
make_tuple(number<kKPerBlock / 8>{}, number<kMPerBlock>{}, number<8>{}),
|
||||
make_tuple(number<(kMPerBlock + 1) * 8>{}, number<8>{}, number<1>{}),
|
||||
number<8>{},
|
||||
number<1>{});
|
||||
|
||||
constexpr auto a_lds_block_desc = transform_tensor_descriptor(
|
||||
a_lds_block_desc_0,
|
||||
make_tuple(make_pass_through_transform(kMPerBlock),
|
||||
make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
|
||||
make_tuple(sequence<1>{}, sequence<0, 2>{}),
|
||||
make_tuple(sequence<0>{}, sequence<1>{}));
|
||||
|
||||
return a_lds_block_desc;
|
||||
}
|
||||
|
||||
// 3d + padding
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
|
||||
{
|
||||
using namespace ck_tile;
|
||||
|
||||
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
|
||||
make_tuple(number<kKPerBlock / 8>{}, number<kNPerBlock>{}, number<8>{}),
|
||||
make_tuple(number<(kNPerBlock + 1) * 8>{}, number<8>{}, number<1>{}),
|
||||
number<8>{},
|
||||
number<1>{});
|
||||
|
||||
constexpr auto b_lds_block_desc = transform_tensor_descriptor(
|
||||
b_lds_block_desc_0,
|
||||
make_tuple(make_pass_through_transform(kNPerBlock),
|
||||
make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
|
||||
make_tuple(sequence<1>{}, sequence<0, 2>{}),
|
||||
make_tuple(sequence<0>{}, sequence<1>{}));
|
||||
|
||||
return b_lds_block_desc;
|
||||
}
|
||||
#elif 1
|
||||
// fake XOR
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
|
||||
{
|
||||
using namespace ck_tile;
|
||||
|
||||
using ADataType = remove_cvref_t<typename Problem::ADataType>;
|
||||
|
||||
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
constexpr auto a_lds_block_desc_d1_d2_d3 = make_naive_tensor_descriptor_packed(
|
||||
make_tuple(number<kMPerBlock / 2>{}, number<2>{}, number<kKPerBlock>{}),
|
||||
number<kKPerBlock>{});
|
||||
|
||||
constexpr index_t kK1 = 16 / sizeof(ADataType);
|
||||
|
||||
constexpr auto a_lds_block_desc_d4_d5_d6 = transform_tensor_descriptor(
|
||||
a_lds_block_desc_d1_d2_d3,
|
||||
make_tuple(
|
||||
make_xor_transform(make_tuple(number<kMPerBlock / 2>{}, number<kKPerBlock>{}), kK1),
|
||||
make_pass_through_transform(2)),
|
||||
make_tuple(sequence<0, 2>{}, sequence<1>{}),
|
||||
make_tuple(sequence<0, 2>{}, sequence<1>{}));
|
||||
|
||||
constexpr auto a_lds_block_desc_m_k = transform_tensor_descriptor(
|
||||
a_lds_block_desc_d4_d5_d6,
|
||||
make_tuple(make_merge_transform(make_tuple(number<kMPerBlock / 2>{}, number<2>{})),
|
||||
make_pass_through_transform(kKPerBlock)),
|
||||
make_tuple(sequence<0, 1>{}, sequence<2>{}),
|
||||
make_tuple(sequence<0>{}, sequence<1>{}));
|
||||
|
||||
return a_lds_block_desc_m_k;
|
||||
}
|
||||
|
||||
// fake XOR
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
|
||||
{
|
||||
using namespace ck_tile;
|
||||
|
||||
using BDataType = remove_cvref_t<typename Problem::BDataType>;
|
||||
|
||||
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
constexpr auto b_lds_block_desc_d1_d2_d3 = make_naive_tensor_descriptor_packed(
|
||||
make_tuple(number<kNPerBlock / 2>{}, number<2>{}, number<kKPerBlock>{}),
|
||||
number<kKPerBlock>{});
|
||||
|
||||
constexpr index_t kK1 = 16 / sizeof(BDataType);
|
||||
|
||||
constexpr auto b_lds_block_desc_d4_d5_d6 = transform_tensor_descriptor(
|
||||
b_lds_block_desc_d1_d2_d3,
|
||||
make_tuple(
|
||||
make_xor_transform(make_tuple(number<kNPerBlock / 2>{}, number<kKPerBlock>{}), kK1),
|
||||
make_pass_through_transform(2)),
|
||||
make_tuple(sequence<0, 2>{}, sequence<1>{}),
|
||||
make_tuple(sequence<0, 2>{}, sequence<1>{}));
|
||||
|
||||
constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
|
||||
b_lds_block_desc_d4_d5_d6,
|
||||
make_tuple(make_merge_transform(make_tuple(number<kNPerBlock / 2>{}, number<2>{})),
|
||||
make_pass_through_transform(kKPerBlock)),
|
||||
make_tuple(sequence<0, 1>{}, sequence<2>{}),
|
||||
make_tuple(sequence<0>{}, sequence<1>{}));
|
||||
|
||||
return b_lds_block_desc_n_k;
|
||||
}
|
||||
#endif
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
|
||||
{
|
||||
using ADataType = remove_cvref_t<typename Problem::ADataType>;
|
||||
|
||||
constexpr index_t kBlockSize = Problem::kBlockSize;
|
||||
|
||||
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
constexpr index_t K1 = 16 / sizeof(ADataType);
|
||||
constexpr index_t K0 = kKPerBlock / K1;
|
||||
constexpr index_t M2 = get_warp_size() / K0;
|
||||
#if 1 // coalesce reading for each blocks
|
||||
constexpr index_t M1 = kBlockSize / get_warp_size();
|
||||
constexpr index_t M0 = kMPerBlock / (M2 * M1);
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<sequence<1>,
|
||||
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
|
||||
tuple<sequence<1>, sequence<1, 2>>,
|
||||
tuple<sequence<1>, sequence<2, 0>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 1>>{});
|
||||
#else // coalesce reading for each warps
|
||||
constexpr index_t M0 = kBlockSize / get_warp_size();
|
||||
constexpr index_t M1 = kMPerBlock / (M2 * M0);
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<sequence<1>,
|
||||
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
|
||||
tuple<sequence<1>, sequence<1, 2>>,
|
||||
tuple<sequence<0>, sequence<2, 0>>,
|
||||
sequence<1, 2>,
|
||||
sequence<1, 1>>{});
|
||||
#endif
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
|
||||
{
|
||||
using BDataType = remove_cvref_t<typename Problem::BDataType>;
|
||||
|
||||
constexpr index_t kBlockSize = Problem::kBlockSize;
|
||||
|
||||
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
constexpr index_t K1 = 16 / sizeof(BDataType);
|
||||
constexpr index_t K0 = kKPerBlock / K1;
|
||||
constexpr index_t N2 = get_warp_size() / K0;
|
||||
#if 1 // coalesce reading for each blocks
|
||||
constexpr index_t N1 = kBlockSize / get_warp_size();
|
||||
constexpr index_t N0 = kNPerBlock / (N2 * N1);
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<sequence<1>,
|
||||
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
|
||||
tuple<sequence<1>, sequence<1, 2>>,
|
||||
tuple<sequence<1>, sequence<2, 0>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 1>>{});
|
||||
#else // coalesce reading for each warps
|
||||
constexpr index_t N0 = kBlockSize / get_warp_size();
|
||||
constexpr index_t N1 = kNPerBlock / (N2 * N0);
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<sequence<1>,
|
||||
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
|
||||
tuple<sequence<1>, sequence<1, 2>>,
|
||||
tuple<sequence<0>, sequence<2, 0>>,
|
||||
sequence<1, 2>,
|
||||
sequence<1, 1>>{});
|
||||
#endif
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
|
||||
{
|
||||
using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1DefaultPolicy;
|
||||
|
||||
return BlockGemmASmemBSmemCRegV1<Problem, BlockGemmPolicy>{};
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,218 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// A Tile Window: global memory
|
||||
// B Tile Window: global memory
|
||||
// C Distributed tensor: register
|
||||
template <typename Problem, typename Policy = BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy>
|
||||
struct BlockGemmPipelineAGmemBGmemCRegV2
|
||||
{
|
||||
using ADataType = remove_cvref_t<typename Problem::ADataType>;
|
||||
using BDataType = remove_cvref_t<typename Problem::BDataType>;
|
||||
using CDataType = remove_cvref_t<typename Problem::CDataType>;
|
||||
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
|
||||
|
||||
static constexpr index_t kBlockSize = Problem::kBlockSize;
|
||||
|
||||
static constexpr index_t kMPerBlock = BlockGemmShape::kM;
|
||||
static constexpr index_t kNPerBlock = BlockGemmShape::kN;
|
||||
static constexpr index_t kKPerBlock = BlockGemmShape::kK;
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetStaticLdsSize()
|
||||
{
|
||||
return ck_tile::integer_divide_ceil(
|
||||
sizeof(ADataType) *
|
||||
Policy::template MakeALdsBlockDescriptor<Problem>().get_element_space_size(),
|
||||
16) *
|
||||
16 +
|
||||
sizeof(BDataType) *
|
||||
Policy::template MakeBLdsBlockDescriptor<Problem>().get_element_space_size();
|
||||
}
|
||||
|
||||
template <typename ADramBlockWindowTmp,
|
||||
typename BDramBlockWindowTmp,
|
||||
typename AElementFunction,
|
||||
typename BElementFunction>
|
||||
CK_TILE_HOST_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
|
||||
const AElementFunction& a_element_func,
|
||||
const BDramBlockWindowTmp& b_dram_block_window_tmp,
|
||||
const BElementFunction& b_element_func,
|
||||
index_t num_loop,
|
||||
void* p_smem) const
|
||||
{
|
||||
static_assert(
|
||||
std::is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
|
||||
std::is_same_v<BDataType, remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
|
||||
"wrong!");
|
||||
|
||||
static_assert(kMPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
|
||||
kNPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
|
||||
kKPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[number<1>{}],
|
||||
"wrong!");
|
||||
|
||||
// A tile in LDS
|
||||
ADataType* p_a_lds = static_cast<ADataType*>(p_smem);
|
||||
|
||||
constexpr auto a_lds_block_desc = Policy::template MakeALdsBlockDescriptor<Problem>();
|
||||
|
||||
auto a_lds_block = make_tensor_view<address_space_enum::lds>(p_a_lds, a_lds_block_desc);
|
||||
|
||||
constexpr index_t a_lds_block_space_size_aligned =
|
||||
integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16) *
|
||||
16;
|
||||
|
||||
// B tile in LDS
|
||||
BDataType* p_b_lds = static_cast<BDataType*>(
|
||||
static_cast<void*>(static_cast<char*>(p_smem) + a_lds_block_space_size_aligned));
|
||||
|
||||
constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
|
||||
|
||||
auto b_lds_block = make_tensor_view<address_space_enum::lds>(p_b_lds, b_lds_block_desc);
|
||||
|
||||
// A DRAM tile window for load
|
||||
auto a_copy_dram_window =
|
||||
make_tile_window(a_dram_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}),
|
||||
a_dram_block_window_tmp.get_window_origin(),
|
||||
Policy::template MakeADramTileDistribution<Problem>());
|
||||
|
||||
// A LDS tile window for store
|
||||
auto a_copy_lds_window =
|
||||
make_tile_window(a_lds_block,
|
||||
make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}),
|
||||
{0, 0},
|
||||
a_copy_dram_window.get_tile_distribution());
|
||||
|
||||
// B DRAM tile window for load
|
||||
auto b_copy_dram_window =
|
||||
make_tile_window(b_dram_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}),
|
||||
b_dram_block_window_tmp.get_window_origin(),
|
||||
Policy::template MakeBDramTileDistribution<Problem>());
|
||||
|
||||
// B LDS tile window for store
|
||||
auto b_copy_lds_window =
|
||||
make_tile_window(b_lds_block,
|
||||
make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}),
|
||||
{0, 0},
|
||||
b_copy_dram_window.get_tile_distribution());
|
||||
|
||||
// A LDS tile for block GEMM
|
||||
auto a_lds_gemm_window = make_tile_window(
|
||||
a_lds_block, make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}), {0, 0});
|
||||
|
||||
// B LDS tile for block GEMM
|
||||
auto b_lds_gemm_window = make_tile_window(
|
||||
b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0});
|
||||
|
||||
// Block GEMM
|
||||
constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>();
|
||||
|
||||
// Acc register tile
|
||||
auto c_block_tile = decltype(block_gemm(a_lds_gemm_window, b_lds_gemm_window)){};
|
||||
|
||||
// prefetch
|
||||
// global read 0
|
||||
auto a_block_tile = load_tile(a_copy_dram_window);
|
||||
auto b_block_tile = load_tile(b_copy_dram_window);
|
||||
|
||||
{
|
||||
// move to 1
|
||||
move_tile_window(a_copy_dram_window, {0, kKPerBlock});
|
||||
move_tile_window(b_copy_dram_window, {0, kKPerBlock});
|
||||
|
||||
// initialize C
|
||||
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
|
||||
|
||||
// LDS write 0
|
||||
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
|
||||
store_tile(a_copy_lds_window, a_block_tile_tmp);
|
||||
// global read 1
|
||||
a_block_tile = load_tile(a_copy_dram_window);
|
||||
|
||||
// LDS write 0
|
||||
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
|
||||
store_tile(b_copy_lds_window, b_block_tile_tmp);
|
||||
// global read 1
|
||||
b_block_tile = load_tile(b_copy_dram_window);
|
||||
}
|
||||
|
||||
index_t iCounter = num_loop - 2;
|
||||
|
||||
do
|
||||
{
|
||||
block_sync_lds();
|
||||
|
||||
// GEMM i
|
||||
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
|
||||
|
||||
block_sync_lds();
|
||||
|
||||
// move to i + 2
|
||||
move_tile_window(a_copy_dram_window, {0, kKPerBlock});
|
||||
move_tile_window(b_copy_dram_window, {0, kKPerBlock});
|
||||
|
||||
// LDS write i + 1
|
||||
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
|
||||
store_tile(a_copy_lds_window, a_block_tile_tmp);
|
||||
// global read i + 2
|
||||
a_block_tile = load_tile(a_copy_dram_window);
|
||||
|
||||
// LDS write i + 1
|
||||
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
|
||||
store_tile(b_copy_lds_window, b_block_tile_tmp);
|
||||
// global read i + 2
|
||||
b_block_tile = load_tile(b_copy_dram_window);
|
||||
|
||||
iCounter--;
|
||||
|
||||
} while(iCounter > 0);
|
||||
|
||||
// tail
|
||||
{
|
||||
block_sync_lds();
|
||||
|
||||
// GEMM num_loop - 2
|
||||
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
|
||||
|
||||
block_sync_lds();
|
||||
|
||||
// LDS write num_loop - 1
|
||||
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
|
||||
store_tile(a_copy_lds_window, a_block_tile_tmp);
|
||||
|
||||
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
|
||||
store_tile(b_copy_lds_window, b_block_tile_tmp);
|
||||
|
||||
block_sync_lds();
|
||||
|
||||
// GEMM num_loop - 1
|
||||
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
|
||||
}
|
||||
|
||||
return c_block_tile;
|
||||
}
|
||||
|
||||
template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
|
||||
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
|
||||
const BDramBlockWindowTmp& b_dram_block_window_tmp,
|
||||
index_t num_loop,
|
||||
void* p_smem) const
|
||||
{
|
||||
return operator()(
|
||||
a_dram_block_window_tmp,
|
||||
[](const ADataType& a) { return a; },
|
||||
b_dram_block_window_tmp,
|
||||
[](const BDataType& b) { return b; },
|
||||
num_loop,
|
||||
p_smem);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,18 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// Default policy for BlockGemmPipelineAGmemBGmemCRegV2
|
||||
// Default policy class should not be templated, put template on member functions instead
|
||||
// NOTE: policy should be binded to its corresponding operation. It's just a coincidence that
|
||||
// BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy is the same as
|
||||
// BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy
|
||||
using BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy =
|
||||
BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy;
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,25 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename ADataType_,
|
||||
typename BDataType_,
|
||||
typename CDataType_,
|
||||
index_t kBlockSize_,
|
||||
typename BlockGemmShape_>
|
||||
struct BlockGemmPipelineProblem
|
||||
{
|
||||
using ADataType = remove_cvref_t<ADataType_>;
|
||||
using BDataType = remove_cvref_t<BDataType_>;
|
||||
using CDataType = remove_cvref_t<CDataType_>;
|
||||
using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
|
||||
|
||||
static constexpr index_t kBlockSize = kBlockSize_;
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
18
include/ck_tile/ops/gemm/pipeline/tile_gemm_shape.hpp
Normal file
18
include/ck_tile/ops/gemm/pipeline/tile_gemm_shape.hpp
Normal file
@@ -0,0 +1,18 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <index_t kMPerTile, index_t kNPerTile, index_t kKPerTile>
|
||||
struct TileGemmShape
|
||||
{
|
||||
static constexpr index_t kM = kMPerTile;
|
||||
static constexpr index_t kN = kNPerTile;
|
||||
static constexpr index_t kK = kKPerTile;
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
Reference in New Issue
Block a user