ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-18 01:28:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

tmp save

Browse Source
This commit is contained in:
Jakub Piasecki
2025-10-30 13:17:06 +00:00
parent d8559798d5
commit e7696eeff1
32 changed files with 2189 additions and 93 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp
View File

@@ -392,7 +392,7 @@ struct UniversalGemmBasePolicy
}
template <typename Problem, bool IsWave32Host = false>
CK_TILE_HOST_DEVICE static constexpr auto GetVectorSizeA()
CK_TILE_HOST_DEVICE static constexpr index_t GetVectorSizeA()
{
using AsLayout = remove_cvref_t<typename Problem::AsLayoutTuple>;
using AsDataType = remove_cvref_t<typename Problem::AsDataTypeTuple>;
@@ -402,6 +402,10 @@ struct UniversalGemmBasePolicy
using ALayout = remove_cvref_t<std::tuple_element_t<number<0>{}, AsLayout>>;
using ADataType = remove_cvref_t<std::tuple_element_t<number<0>{}, AsDataType>>;
if constexpr(Problem::FixedVectorSize) {
return Problem::VectorSizeA;
}
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
return GetGlobalVectorLoadSize<Problem,
@@ -421,7 +425,7 @@ struct UniversalGemmBasePolicy
}
template <typename Problem, bool IsWave32Host = false>
CK_TILE_HOST_DEVICE static constexpr auto GetVectorSizeB()
CK_TILE_HOST_DEVICE static constexpr index_t GetVectorSizeB()
{
using BsLayout = remove_cvref_t<typename Problem::BsLayoutTuple>;
using BsDataType = remove_cvref_t<typename Problem::BsDataTypeTuple>;
@@ -431,6 +435,10 @@ struct UniversalGemmBasePolicy
using BLayout = remove_cvref_t<std::tuple_element_t<number<0>{}, BsLayout>>;
using BDataType = remove_cvref_t<std::tuple_element_t<number<0>{}, BsDataType>>;
if constexpr(Problem::FixedVectorSize) {
return Problem::VectorSizeB;
}
if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
return GetGlobalVectorLoadSize<Problem,

4
include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_backward_data_kernel.hpp
View File

@@ -887,8 +887,8 @@ struct GroupedConvolutionBackwardDataKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop = amd_wave_read_first_lane(
TilePartitioner::GetLoopNum(gemm_tile_windows.at(I0).get_length(I1)));
const index_t num_loop = amd_wave_read_first_lane(TilePartitioner::GetLoopNum(
gemm_pad_views.at(I0).get_tensor_descriptor().get_length(I1)));
// Run GEMM cooperatively by whole workgroup.
const auto& a_block_window = gemm_tile_windows.at(I0);

6
include/ck_tile/utility.hpp Normal file
View File

@@ -0,0 +1,6 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/utility/json_dump.hpp"