mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-01 12:11:19 +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 +1,25 @@
|
||||
// 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 {
|
||||
// Problem Description for BlockGemmARegBGmemCReg
|
||||
template <typename ADataType_,
|
||||
typename BDataType_,
|
||||
typename CDataType_,
|
||||
index_t kBlockSize_,
|
||||
typename BlockGemmShape_>
|
||||
struct BlockGemmARegBGmemCRegProblem
|
||||
{
|
||||
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
|
||||
135
include/ck_tile/ops/gemm/block/block_gemm_areg_bgmem_creg_v1.hpp
Normal file
135
include/ck_tile/ops/gemm/block/block_gemm_areg_bgmem_creg_v1.hpp
Normal file
@@ -0,0 +1,135 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_default_policy.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// A is block distributed tensor
|
||||
// B is block window on global memory
|
||||
// C is block distributed tensor
|
||||
// This will:
|
||||
// 1. load B from global memory into shared memory and then
|
||||
// 2. Call BlockGemmARegSGmemCRegV1
|
||||
template <typename Problem_, typename Policy_ = BlockGemmARegBGmemCRegV1DefaultPolicy>
|
||||
struct BlockGemmARegBGmemCRegV1
|
||||
{
|
||||
using Problem = remove_cvref_t<Problem_>;
|
||||
using Policy = remove_cvref_t<Policy_>;
|
||||
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;
|
||||
|
||||
// use BlockGemmARegBSmemCRegV1 as the underlying block-GEMM implementation
|
||||
using BlockGemmARegBSmemCRegImpl = BlockGemmARegBSmemCRegV1<
|
||||
BlockGemmARegBSmemCRegProblem<ADataType, BDataType, CDataType, kBlockSize, BlockGemmShape>,
|
||||
BlockGemmARegBSmemCRegV1DefaultPolicy>;
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetStaticLdsSize()
|
||||
{
|
||||
return sizeof(BDataType) *
|
||||
Policy::template MakeBSmemBlockDescriptor<Problem>().get_element_space_size();
|
||||
}
|
||||
|
||||
// C += A * B
|
||||
template <typename CBlockTensor, typename ABlockTensor, typename BBlockGmemWindowTmp>
|
||||
CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
|
||||
const ABlockTensor& a_block_tensor,
|
||||
const BBlockGmemWindowTmp& b_block_gmem_window_tmp,
|
||||
void* smem_ptr) const
|
||||
{
|
||||
static_assert(
|
||||
std::is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
|
||||
std::is_same_v<BDataType, remove_cv_t<typename BBlockGmemWindowTmp::DataType>> &&
|
||||
std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
|
||||
"wrong!");
|
||||
|
||||
constexpr index_t MPerBlock = ABlockTensor{}.get_lengths()[number<0>{}];
|
||||
constexpr index_t NPerBlock = BBlockGmemWindowTmp{}.get_window_lengths()[number<0>{}];
|
||||
constexpr index_t KPerBlock = ABlockTensor{}.get_lengths()[number<1>{}];
|
||||
|
||||
static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
|
||||
KPerBlock == BlockGemmShape::kK,
|
||||
"wrong!");
|
||||
|
||||
const auto b_block_gmem_window =
|
||||
make_tile_window(b_block_gmem_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
|
||||
b_block_gmem_window_tmp.get_window_origin(),
|
||||
Policy::template MakeBGmemTileDistribution<Problem>());
|
||||
|
||||
// B LDS and LDS window
|
||||
auto b_block_smem = make_tensor_view<address_space_enum::lds>(
|
||||
reinterpret_cast<BDataType*>(smem_ptr),
|
||||
Policy::template MakeBSmemBlockDescriptor<Problem>());
|
||||
|
||||
auto b_block_smem_window = make_tile_window(
|
||||
b_block_smem, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
|
||||
|
||||
// load B tile from global mem
|
||||
const auto b_block_tile = load_tile(b_block_gmem_window);
|
||||
|
||||
// store B tile into shared mem
|
||||
store_tile(b_block_smem_window, b_block_tile);
|
||||
|
||||
// wait for store_tile to finish
|
||||
block_sync_lds();
|
||||
|
||||
// block GEMM
|
||||
BlockGemmARegBSmemCRegImpl{}(c_block_tensor, a_block_tensor, b_block_smem_window);
|
||||
}
|
||||
|
||||
// C = A * B
|
||||
template <typename ABlockTensor, typename BBlockGmemWindowTmp>
|
||||
CK_TILE_DEVICE auto operator()(const ABlockTensor& a_block_tensor,
|
||||
const BBlockGmemWindowTmp& b_block_gmem_window_tmp,
|
||||
void* smem_ptr) const
|
||||
{
|
||||
static_assert(
|
||||
std::is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
|
||||
std::is_same_v<BDataType, remove_cv_t<typename BBlockGmemWindowTmp::DataType>>,
|
||||
"wrong!");
|
||||
|
||||
constexpr index_t MPerBlock = ABlockTensor{}.get_lengths()[number<0>{}];
|
||||
constexpr index_t NPerBlock = BBlockGmemWindowTmp{}.get_window_lengths()[number<0>{}];
|
||||
constexpr index_t KPerBlock = ABlockTensor{}.get_lengths()[number<1>{}];
|
||||
|
||||
static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
|
||||
KPerBlock == BlockGemmShape::kK,
|
||||
"wrong!");
|
||||
|
||||
const auto b_block_gmem_window =
|
||||
make_tile_window(b_block_gmem_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
|
||||
b_block_gmem_window_tmp.get_window_origin(),
|
||||
Policy::template MakeBGmemTileDistribution<Problem>());
|
||||
|
||||
// B LDS and LDS window
|
||||
auto b_block_smem = make_tensor_view<address_space_enum::lds>(
|
||||
reinterpret_cast<BDataType*>(smem_ptr),
|
||||
Policy::template MakeBSmemBlockDescriptor<Problem>());
|
||||
|
||||
auto b_block_smem_window = make_tile_window(
|
||||
b_block_smem, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
|
||||
|
||||
// load B tile from global mem
|
||||
const auto b_block_tile = load_tile(b_block_gmem_window);
|
||||
|
||||
// store B tile into shared mem
|
||||
store_tile(b_block_smem_window, b_block_tile);
|
||||
|
||||
// wait for store_tile to finish
|
||||
block_sync_lds();
|
||||
|
||||
// block GEMM
|
||||
return BlockGemmARegBSmemCRegImpl{}(a_block_tensor, b_block_smem_window);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,110 @@
|
||||
// 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 BlockGemmARegBGmemCRegV1
|
||||
// Default policy class should not be templated, put template on member functions instead
|
||||
struct BlockGemmARegBGmemCRegV1DefaultPolicy
|
||||
{
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeBGmemTileDistribution()
|
||||
{
|
||||
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;
|
||||
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>>{});
|
||||
}
|
||||
|
||||
#if 0
|
||||
// 2d
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
|
||||
{
|
||||
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 0
|
||||
// 3d + padding
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto MakeBSmemBlockDescriptor()
|
||||
{
|
||||
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 MakeBSmemBlockDescriptor()
|
||||
{
|
||||
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
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,26 @@
|
||||
// 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 {
|
||||
|
||||
// Problem Description for BlockGemmARegBSmemCReg
|
||||
template <typename ADataType_,
|
||||
typename BDataType_,
|
||||
typename CDataType_,
|
||||
index_t kBlockSize_,
|
||||
typename BlockGemmShape_>
|
||||
struct BlockGemmARegBSmemCRegProblem
|
||||
{
|
||||
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
|
||||
340
include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1.hpp
Normal file
340
include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1.hpp
Normal file
@@ -0,0 +1,340 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/gemm/block/block_gemm_areg_bgmem_creg_v1_default_policy.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// A is block distributed tensor
|
||||
// B is block window on shared memory
|
||||
// C is block distributed tensor
|
||||
template <typename Problem_, typename Policy_ = BlockGemmARegBSmemCRegV1DefaultPolicy>
|
||||
struct BlockGemmARegBSmemCRegV1
|
||||
{
|
||||
using Problem = remove_cvref_t<Problem_>;
|
||||
using Policy = remove_cvref_t<Policy_>;
|
||||
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;
|
||||
|
||||
// C += A * B
|
||||
template <typename CBlockTensor, typename ABlockTensorTmp, typename BBlockWindowTmp>
|
||||
CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
|
||||
const ABlockTensorTmp& a_block_tensor_tmp,
|
||||
const BBlockWindowTmp& b_block_window_tmp) const
|
||||
{
|
||||
static_assert(
|
||||
std::is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
|
||||
std::is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>> &&
|
||||
std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
|
||||
"wrong!");
|
||||
|
||||
constexpr index_t MPerBlock = ABlockTensorTmp{}.get_lengths()[number<0>{}];
|
||||
constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
|
||||
constexpr index_t KPerBlock = ABlockTensorTmp{}.get_lengths()[number<1>{}];
|
||||
|
||||
static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
|
||||
KPerBlock == BlockGemmShape::kK,
|
||||
"wrong!");
|
||||
|
||||
constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
|
||||
|
||||
using WG = remove_cvref_t<decltype(config.template at<0>())>;
|
||||
|
||||
constexpr index_t MWarp = config.template at<1>();
|
||||
constexpr index_t NWarp = config.template at<2>();
|
||||
|
||||
constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
|
||||
constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
|
||||
constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
|
||||
|
||||
constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
|
||||
constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
|
||||
|
||||
const index_t iNWarp = get_warp_id() % NWarp;
|
||||
|
||||
constexpr auto a_block_outer_dstr_encoding =
|
||||
tile_distribution_encoding<sequence<NWarp>,
|
||||
tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
|
||||
tuple<sequence<1, 0>>,
|
||||
tuple<sequence<1, 0>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{};
|
||||
|
||||
constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
|
||||
sequence<>,
|
||||
tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
|
||||
tuple<sequence<1, 2>>,
|
||||
tuple<sequence<1, 1>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{};
|
||||
|
||||
constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
|
||||
a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
|
||||
|
||||
constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
|
||||
c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
|
||||
|
||||
constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
|
||||
|
||||
// constrcut from A-block-tensor from A-Block-tensor-tmp
|
||||
// FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
|
||||
// distribution
|
||||
auto a_block_tensor =
|
||||
make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(a_block_dstr);
|
||||
|
||||
a_block_tensor.get_thread_buffer() = a_block_tensor_tmp.get_thread_buffer();
|
||||
|
||||
// construct B-warp-window
|
||||
auto b_warp_window_tmp = make_tile_window(
|
||||
b_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<WG::kN>{}, number<WG::kK>{}),
|
||||
b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
|
||||
make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
|
||||
|
||||
#if 0 // FIXME: using array will cause register spill
|
||||
array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
|
||||
{b_warp_window_tmp}};
|
||||
|
||||
for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
|
||||
{
|
||||
for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
|
||||
{
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
}
|
||||
}
|
||||
#else
|
||||
statically_indexed_array<
|
||||
statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
});
|
||||
#endif
|
||||
|
||||
// check C-block-distribution
|
||||
static_assert(
|
||||
std::is_same_v<remove_cvref_t<decltype(c_block_dstr_encode)>,
|
||||
remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
|
||||
.get_static_tile_distribution_encoding())>>,
|
||||
"wrong!");
|
||||
|
||||
using AWarpDstr = typename WG::AWarpDstr;
|
||||
using CWarpDstr = typename WG::CWarpDstr;
|
||||
|
||||
using AWarpTensor = typename WG::AWarpTensor;
|
||||
using CWarpTensor = typename WG::CWarpTensor;
|
||||
|
||||
constexpr auto a_warp_y_lengths =
|
||||
to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
|
||||
constexpr auto c_warp_y_lengths =
|
||||
to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
|
||||
|
||||
constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
// C = A * B
|
||||
template <typename ABlockTensorTmp, typename BBlockWindowTmp>
|
||||
CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
|
||||
const BBlockWindowTmp& b_block_window_tmp) const
|
||||
{
|
||||
static_assert(
|
||||
std::is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
|
||||
std::is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>>,
|
||||
"wrong!");
|
||||
|
||||
constexpr index_t MPerBlock = ABlockTensorTmp{}.get_lengths()[number<0>{}];
|
||||
constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
|
||||
constexpr index_t KPerBlock = ABlockTensorTmp{}.get_lengths()[number<1>{}];
|
||||
|
||||
static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
|
||||
KPerBlock == BlockGemmShape::kK,
|
||||
"wrong!");
|
||||
|
||||
constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
|
||||
|
||||
using WG = remove_cvref_t<decltype(config.template at<0>())>;
|
||||
|
||||
constexpr index_t MWarp = config.template at<1>();
|
||||
constexpr index_t NWarp = config.template at<2>();
|
||||
|
||||
constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
|
||||
constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
|
||||
constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
|
||||
|
||||
constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
|
||||
constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
|
||||
|
||||
const index_t iNWarp = get_warp_id() % NWarp;
|
||||
|
||||
constexpr auto a_block_outer_dstr_encoding =
|
||||
tile_distribution_encoding<sequence<NWarp>,
|
||||
tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
|
||||
tuple<sequence<1, 0>>,
|
||||
tuple<sequence<1, 0>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{};
|
||||
|
||||
constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
|
||||
sequence<>,
|
||||
tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
|
||||
tuple<sequence<1, 2>>,
|
||||
tuple<sequence<1, 1>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{};
|
||||
|
||||
constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
|
||||
a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
|
||||
|
||||
constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
|
||||
c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
|
||||
|
||||
constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
|
||||
constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
|
||||
|
||||
// constrcut from A-block-tensor from A-Block-tensor-tmp
|
||||
// FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
|
||||
// distribution
|
||||
auto a_block_tensor =
|
||||
make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(a_block_dstr);
|
||||
|
||||
a_block_tensor.get_thread_buffer() = a_block_tensor_tmp.get_thread_buffer();
|
||||
|
||||
// construct B-warp-window
|
||||
auto b_warp_window_tmp = make_tile_window(
|
||||
b_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<WG::kN>{}, number<WG::kK>{}),
|
||||
b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
|
||||
make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
|
||||
|
||||
#if 0 // FIXME: using array will cause register spill
|
||||
array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
|
||||
{b_warp_window_tmp}};
|
||||
|
||||
for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
|
||||
{
|
||||
for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
|
||||
{
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
}
|
||||
}
|
||||
#else
|
||||
statically_indexed_array<
|
||||
statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
});
|
||||
#endif
|
||||
|
||||
// Construct C-Block-HostTensor
|
||||
auto c_block_tensor = make_static_distributed_tensor<CDataType>(c_block_dstr);
|
||||
|
||||
using AWarpDstr = typename WG::AWarpDstr;
|
||||
using CWarpDstr = typename WG::CWarpDstr;
|
||||
|
||||
using AWarpTensor = typename WG::AWarpTensor;
|
||||
using CWarpTensor = typename WG::CWarpTensor;
|
||||
|
||||
constexpr auto a_warp_y_lengths =
|
||||
to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
|
||||
constexpr auto c_warp_y_lengths =
|
||||
to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
|
||||
|
||||
constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
return c_block_tensor;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,36 @@
|
||||
// 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 AType_,
|
||||
typename BType_,
|
||||
typename CType_,
|
||||
typename BlockWarps_,
|
||||
typename WarpGemm_>
|
||||
struct BlockGemmARegBSmemCRegV1CustomPolicy
|
||||
{
|
||||
using AType = remove_cvref_t<AType_>;
|
||||
using BType = remove_cvref_t<BType_>;
|
||||
using CType = remove_cvref_t<CType_>;
|
||||
|
||||
using BlockWarps = remove_cvref_t<BlockWarps_>;
|
||||
|
||||
static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
|
||||
static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
|
||||
static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
|
||||
|
||||
using WarpGemm = remove_cvref_t<WarpGemm_>;
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
|
||||
{
|
||||
return make_tuple(WarpGemm{}, kMWarps, kNWarps);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,56 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/gemm/warp/warp_gemm.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// Default policy for BlockGemmARegBSmemCRegV1
|
||||
// Default policy class should not be templated, put template on member functions instead
|
||||
struct BlockGemmARegBSmemCRegV1DefaultPolicy
|
||||
{
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
|
||||
{
|
||||
if constexpr(std::is_same_v<typename Problem::ADataType, half_t> &&
|
||||
std::is_same_v<typename Problem::BDataType, half_t> &&
|
||||
std::is_same_v<typename Problem::CDataType, float>)
|
||||
{
|
||||
#if 0
|
||||
constexpr index_t kBlockSize = Problem::kBlockSize;
|
||||
|
||||
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
|
||||
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
|
||||
|
||||
constexpr index_t NumWarp = kBlockSize / get_warp_size();
|
||||
|
||||
// FIXME
|
||||
if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
|
||||
kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
|
||||
{
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
|
||||
}
|
||||
else
|
||||
{
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
|
||||
}
|
||||
#else
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
|
||||
#endif
|
||||
}
|
||||
else if constexpr(std::is_same_v<typename Problem::ADataType, bf16_t> &&
|
||||
std::is_same_v<typename Problem::BDataType, bf16_t> &&
|
||||
std::is_same_v<typename Problem::CDataType, float>)
|
||||
{
|
||||
return make_tuple(WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution{}, 4, 1);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
227
include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2.hpp
Normal file
227
include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2.hpp
Normal file
@@ -0,0 +1,227 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2_default_policy.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// A is block distributed tensor
|
||||
// B is block window on shared memory
|
||||
// C is block distributed tensor
|
||||
template <typename Problem_, typename Policy_ = BlockGemmARegBSmemCRegV2DefaultPolicy>
|
||||
struct BlockGemmARegBSmemCRegV2
|
||||
{
|
||||
using Problem = remove_cvref_t<Problem_>;
|
||||
using Policy = remove_cvref_t<Policy_>;
|
||||
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;
|
||||
|
||||
// C += A * B
|
||||
template <typename CBlockTensor, typename ABlockTensorTmp, typename BBlockWindowTmp>
|
||||
CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
|
||||
const ABlockTensorTmp& a_block_tensor_tmp,
|
||||
const BBlockWindowTmp& b_block_window_tmp) const
|
||||
{
|
||||
static_assert(
|
||||
std::is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
|
||||
std::is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>> &&
|
||||
std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
|
||||
"wrong!");
|
||||
|
||||
constexpr index_t MPerBlock = ABlockTensorTmp{}.get_lengths()[number<0>{}];
|
||||
constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
|
||||
constexpr index_t KPerBlock = ABlockTensorTmp{}.get_lengths()[number<1>{}];
|
||||
|
||||
static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
|
||||
KPerBlock == BlockGemmShape::kK,
|
||||
"wrong!");
|
||||
|
||||
constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
|
||||
|
||||
using WG = remove_cvref_t<decltype(config.template at<0>())>;
|
||||
|
||||
constexpr index_t MWarp = config.template at<1>();
|
||||
constexpr index_t NWarp = config.template at<2>();
|
||||
|
||||
constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
|
||||
constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
|
||||
constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
|
||||
|
||||
constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
|
||||
constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
|
||||
|
||||
const index_t iNWarp = get_warp_id() % NWarp;
|
||||
|
||||
constexpr auto a_block_outer_dstr_encoding =
|
||||
tile_distribution_encoding<sequence<NWarp>,
|
||||
tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
|
||||
tuple<sequence<1, 0>>,
|
||||
tuple<sequence<1, 0>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{};
|
||||
|
||||
constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
|
||||
sequence<>,
|
||||
tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
|
||||
tuple<sequence<1, 2>>,
|
||||
tuple<sequence<1, 1>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{};
|
||||
|
||||
constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
|
||||
a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
|
||||
|
||||
constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
|
||||
c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
|
||||
|
||||
constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
|
||||
|
||||
// constrcut from A-block-tensor from A-Block-tensor-tmp
|
||||
// FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
|
||||
// distribution
|
||||
auto a_block_tensor =
|
||||
make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(a_block_dstr);
|
||||
|
||||
a_block_tensor.get_thread_buffer() = a_block_tensor_tmp.get_thread_buffer();
|
||||
|
||||
// construct B-warp-window
|
||||
auto b_warp_window_tmp = make_tile_window(
|
||||
b_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<WG::kN>{}, number<WG::kK>{}),
|
||||
b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
|
||||
make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
|
||||
|
||||
#if 0 // FIXME: using array will cause register spill
|
||||
array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
|
||||
{b_warp_window_tmp}};
|
||||
|
||||
for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
|
||||
{
|
||||
for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
|
||||
{
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
}
|
||||
}
|
||||
#else
|
||||
statically_indexed_array<
|
||||
statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
});
|
||||
#endif
|
||||
|
||||
// check C-block-distribution
|
||||
static_assert(
|
||||
std::is_same_v<remove_cvref_t<decltype(c_block_dstr_encode)>,
|
||||
remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
|
||||
.get_static_tile_distribution_encoding())>>,
|
||||
"wrong!");
|
||||
|
||||
using AWarpDstr = typename WG::AWarpDstr;
|
||||
using CWarpDstr = typename WG::CWarpDstr;
|
||||
|
||||
using AWarpTensor = typename WG::AWarpTensor;
|
||||
using CWarpTensor = typename WG::CWarpTensor;
|
||||
|
||||
constexpr auto a_warp_y_lengths =
|
||||
to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
|
||||
constexpr auto c_warp_y_lengths =
|
||||
to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
|
||||
|
||||
constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block tensor
|
||||
AWarpTensor a_warp_tensor;
|
||||
|
||||
a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
// WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor_array[nIter]);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
CK_TILE_DEVICE constexpr auto MakeCBlockTile() const
|
||||
{
|
||||
constexpr index_t MPerBlock = BlockGemmShape::kM;
|
||||
constexpr index_t NPerBlock = BlockGemmShape::kN;
|
||||
|
||||
constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
|
||||
|
||||
using WG = remove_cvref_t<decltype(config.template at<0>())>;
|
||||
|
||||
constexpr index_t MWarp = config.template at<1>();
|
||||
constexpr index_t NWarp = config.template at<2>();
|
||||
|
||||
constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
|
||||
constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
|
||||
// constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
|
||||
|
||||
constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
|
||||
sequence<>,
|
||||
tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
|
||||
tuple<sequence<1, 2>>,
|
||||
tuple<sequence<1, 1>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{};
|
||||
|
||||
constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
|
||||
c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
|
||||
constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
|
||||
auto c_block_tensor = make_static_distributed_tensor<CDataType>(c_block_dstr);
|
||||
return c_block_tensor;
|
||||
}
|
||||
|
||||
// C = A * B
|
||||
template <typename ABlockTensorTmp, typename BBlockWindowTmp>
|
||||
CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
|
||||
const BBlockWindowTmp& b_block_window_tmp) const
|
||||
{
|
||||
auto c_block_tensor = MakeCBlockTile();
|
||||
operator()(c_block_tensor, a_block_tensor_tmp, b_block_window_tmp);
|
||||
return c_block_tensor;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,36 @@
|
||||
// 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 AType_,
|
||||
typename BType_,
|
||||
typename CType_,
|
||||
typename BlockWarps_,
|
||||
typename WarpGemm_>
|
||||
struct BlockGemmARegBSmemCRegV2CustomPolicy
|
||||
{
|
||||
using AType = remove_cvref_t<AType_>;
|
||||
using BType = remove_cvref_t<BType_>;
|
||||
using CType = remove_cvref_t<CType_>;
|
||||
|
||||
using BlockWarps = remove_cvref_t<BlockWarps_>;
|
||||
|
||||
static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
|
||||
static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
|
||||
static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
|
||||
|
||||
using WarpGemm = remove_cvref_t<WarpGemm_>;
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
|
||||
{
|
||||
return make_tuple(WarpGemm{}, kMWarps, kNWarps);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,46 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/gemm/warp/warp_gemm.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// Default policy for BlockGemmARegBSmemCRegV2
|
||||
// Default policy class should not be templated, put template on member functions instead
|
||||
struct BlockGemmARegBSmemCRegV2DefaultPolicy
|
||||
{
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
|
||||
{
|
||||
|
||||
#if 0
|
||||
constexpr index_t kBlockSize = Problem::kBlockSize;
|
||||
|
||||
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
|
||||
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
|
||||
|
||||
constexpr index_t NumWarp = kBlockSize / get_warp_size();
|
||||
|
||||
// FIXME
|
||||
if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
|
||||
kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
|
||||
{
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
|
||||
}
|
||||
else
|
||||
{
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
|
||||
}
|
||||
#else
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
|
||||
#endif
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,26 @@
|
||||
// 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 {
|
||||
|
||||
// Problem Description for BlockGemmASmemBSmemCRegV1
|
||||
template <typename ADataType_,
|
||||
typename BDataType_,
|
||||
typename CDataType_,
|
||||
index_t kBlockSize_,
|
||||
typename BlockGemmShape_>
|
||||
struct BlockGemmASmemBSmemCRegProblem
|
||||
{
|
||||
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
|
||||
@@ -0,0 +1,213 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// A is block window on shared memory
|
||||
// B is block window on shared memory
|
||||
// C is block distributed tensor
|
||||
template <typename Problem_, typename Policy_ = BlockGemmASmemBSmemCRegV1DefaultPolicy>
|
||||
struct BlockGemmASmemBSmemCRegV1
|
||||
{
|
||||
using Problem = remove_cvref_t<Problem_>;
|
||||
using Policy = remove_cvref_t<Policy_>;
|
||||
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;
|
||||
|
||||
// C += A * B
|
||||
template <typename CBlockTensor, typename ABlockWindowTmp, typename BBlockWindowTmp>
|
||||
CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
|
||||
const ABlockWindowTmp& a_block_window_tmp,
|
||||
const BBlockWindowTmp& b_block_window_tmp) const
|
||||
{
|
||||
static_assert(std::is_same_v<ADataType, typename ABlockWindowTmp::DataType> &&
|
||||
std::is_same_v<BDataType, typename BBlockWindowTmp::DataType> &&
|
||||
std::is_same_v<CDataType, typename CBlockTensor::DataType>,
|
||||
"wrong!");
|
||||
|
||||
constexpr index_t MPerBlock = ABlockWindowTmp{}.get_window_lengths()[number<0>{}];
|
||||
constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
|
||||
constexpr index_t KPerBlock = ABlockWindowTmp{}.get_window_lengths()[number<1>{}];
|
||||
|
||||
static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
|
||||
KPerBlock == BlockGemmShape::kK,
|
||||
"wrong!");
|
||||
|
||||
constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
|
||||
|
||||
using WG = remove_cvref_t<decltype(config.template at<0>())>;
|
||||
|
||||
constexpr index_t MWarp = config.template at<1>();
|
||||
constexpr index_t NWarp = config.template at<2>();
|
||||
|
||||
constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
|
||||
constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
|
||||
constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
|
||||
|
||||
constexpr index_t MPerBlockPerIter = MPerBlock / MIterPerWarp;
|
||||
constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
|
||||
constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
|
||||
|
||||
const index_t iMWarp = get_warp_id() / NWarp;
|
||||
const index_t iNWarp = get_warp_id() % NWarp;
|
||||
|
||||
// construct A-warp-window
|
||||
auto a_warp_window_tmp = make_tile_window(
|
||||
a_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<WG::kM>{}, number<WG::kK>{}),
|
||||
a_block_window_tmp.get_window_origin() + multi_index<2>{iMWarp * WG::kM, 0},
|
||||
make_static_tile_distribution(typename WG::AWarpDstrEncoding{}));
|
||||
|
||||
#if 0 // FIXME: using array will cause register spill
|
||||
array<array<decltype(a_warp_window_tmp), KIterPerWarp>, MIterPerWarp> a_warp_windows{
|
||||
{a_warp_window_tmp}};
|
||||
|
||||
for(index_t mIter = 0; mIter < MIterPerWarp; mIter++)
|
||||
{
|
||||
for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
|
||||
{
|
||||
move_tile_window(a_warp_windows(mIter)(kIter),
|
||||
{mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
}
|
||||
}
|
||||
#else
|
||||
statically_indexed_array<
|
||||
statically_indexed_array<decltype(a_warp_window_tmp), KIterPerWarp>,
|
||||
MIterPerWarp>
|
||||
a_warp_windows;
|
||||
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
|
||||
|
||||
move_tile_window(a_warp_windows(mIter)(kIter),
|
||||
{mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
});
|
||||
#endif
|
||||
|
||||
// construct B-warp-window
|
||||
auto b_warp_window_tmp = make_tile_window(
|
||||
b_block_window_tmp.get_bottom_tensor_view(),
|
||||
make_tuple(number<WG::kN>{}, number<WG::kK>{}),
|
||||
b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
|
||||
make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
|
||||
|
||||
#if 0 // FIXME: using array will cause register spill
|
||||
array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
|
||||
{b_warp_window_tmp}};
|
||||
|
||||
for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
|
||||
{
|
||||
for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
|
||||
{
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
}
|
||||
}
|
||||
#else
|
||||
statically_indexed_array<
|
||||
statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
|
||||
NIterPerWarp>
|
||||
b_warp_windows;
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
|
||||
|
||||
move_tile_window(b_warp_windows(nIter)(kIter),
|
||||
{nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
|
||||
});
|
||||
});
|
||||
#endif
|
||||
|
||||
using CWarpDstr = typename WG::CWarpDstr;
|
||||
using CWarpTensor = typename WG::CWarpTensor;
|
||||
|
||||
constexpr auto c_warp_y_lengths =
|
||||
to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
|
||||
constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
|
||||
|
||||
// hot loop:
|
||||
static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
|
||||
static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
|
||||
// read A warp tensor from A block window
|
||||
const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
|
||||
|
||||
static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
|
||||
// read B warp tensor from B Block window
|
||||
const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
|
||||
|
||||
// read C warp tensor from C block tensor
|
||||
CWarpTensor c_warp_tensor;
|
||||
|
||||
c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
|
||||
|
||||
// warp GEMM
|
||||
WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
|
||||
|
||||
// write C warp tensor into C block tensor
|
||||
c_block_tensor.set_y_sliced_thread_data(
|
||||
merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
|
||||
merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
|
||||
c_warp_tensor.get_thread_buffer());
|
||||
});
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
CK_TILE_DEVICE constexpr auto MakeCBlockTile() const
|
||||
{
|
||||
constexpr index_t MPerBlock = BlockGemmShape::kM;
|
||||
constexpr index_t NPerBlock = BlockGemmShape::kN;
|
||||
|
||||
constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
|
||||
|
||||
using WG = remove_cvref_t<decltype(config.template at<0>())>;
|
||||
|
||||
constexpr index_t MWarp = config.template at<1>();
|
||||
constexpr index_t NWarp = config.template at<2>();
|
||||
|
||||
constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
|
||||
constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
|
||||
|
||||
constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
|
||||
sequence<>,
|
||||
tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
|
||||
tuple<sequence<1, 2>>,
|
||||
tuple<sequence<1, 1>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{};
|
||||
|
||||
constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
|
||||
c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
|
||||
|
||||
constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
|
||||
|
||||
auto c_block_tensor = make_static_distributed_tensor<CDataType>(c_block_dstr);
|
||||
return c_block_tensor;
|
||||
}
|
||||
|
||||
// C = A * B
|
||||
template <typename ABlockTensorTmp, typename BBlockWindowTmp>
|
||||
CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
|
||||
const BBlockWindowTmp& b_block_window_tmp) const
|
||||
{
|
||||
auto c_block_tensor = MakeCBlockTile();
|
||||
operator()(c_block_tensor, a_block_tensor_tmp, b_block_window_tmp);
|
||||
return c_block_tensor;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,38 @@
|
||||
// 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 BlockGemmASmemBSmemCRegV1
|
||||
// Default policy class should not be templated, put template on member functions instead
|
||||
template <typename AType_,
|
||||
typename BType_,
|
||||
typename CType_,
|
||||
typename BlockWarps_,
|
||||
typename WarpGemm_>
|
||||
struct BlockGemmASmemBSmemCRegV1CustomPolicy
|
||||
{
|
||||
using AType = remove_cvref_t<AType_>;
|
||||
using BType = remove_cvref_t<BType_>;
|
||||
using CType = remove_cvref_t<CType_>;
|
||||
|
||||
using BlockWarps = remove_cvref_t<BlockWarps_>;
|
||||
|
||||
static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
|
||||
static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
|
||||
static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
|
||||
|
||||
using WarpGemm = remove_cvref_t<WarpGemm_>;
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
|
||||
{
|
||||
return make_tuple(WarpGemm{}, kMWarps, kNWarps);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,55 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/gemm/warp/warp_gemm.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// Default policy for BlockGemmASmemBSmemCRegV1
|
||||
// Default policy class should not be templated, put template on member functions instead
|
||||
struct BlockGemmASmemBSmemCRegV1DefaultPolicy
|
||||
{
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
|
||||
{
|
||||
if constexpr(std::is_same_v<typename Problem::ADataType, half_t> &&
|
||||
std::is_same_v<typename Problem::BDataType, half_t> &&
|
||||
std::is_same_v<typename Problem::CDataType, float>)
|
||||
{
|
||||
#if 0
|
||||
constexpr index_t kBlockSize = Problem::kBlockSize;
|
||||
|
||||
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
|
||||
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
|
||||
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
|
||||
|
||||
static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
|
||||
|
||||
constexpr index_t NumWarp = kBlockSize / get_warp_size();
|
||||
|
||||
if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
|
||||
kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
|
||||
{
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K16{}, 2, 2);
|
||||
}
|
||||
else
|
||||
{
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K16{}, 2, 2);
|
||||
}
|
||||
#else
|
||||
return make_tuple(WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution{}, 4, 1);
|
||||
#endif
|
||||
}
|
||||
else if constexpr(std::is_same_v<typename Problem::ADataType, bf16_t> &&
|
||||
std::is_same_v<typename Problem::BDataType, bf16_t> &&
|
||||
std::is_same_v<typename Problem::CDataType, float>)
|
||||
{
|
||||
return make_tuple(WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution{}, 4, 1);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
Reference in New Issue
Block a user