ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_BUILDER] Ck Tile Grouped convolution factory (#3352)

Browse Source 
* [BUILDER] Ck Tile Grouped convolution factory

* Part 2

* Fixes after rebase

* Remove leftovers
This commit is contained in:
Bartłomiej Kocot
2025-12-08 10:32:56 +01:00
committed by GitHub
parent 8fec8054b2
commit 04612c30ce
55 changed files with 1431 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
experimental/builder/test/conv/test_ckb_conv_fwd_1d_bf16.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_1d_bf16.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_1d_fp16.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_1d_fp16.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_1d_i8.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_1d_i8.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_2d_bf16.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_2d_bf16.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_2d_bf16_scaleadd_relu.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_2d_bf16_scaleadd_relu.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_2d_dl_fp16.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_2d_dl_fp16.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_2d_fp16.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_2d_fp16.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_2d_fp32.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_2d_fp32.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_2d_fp8.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_2d_fp8.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_2d_large_tensor_fp16.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_2d_large_tensor_fp16.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_3d_bf16.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_3d_bf16.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_3d_fp16.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_3d_fp16.cpp
View File

0
experimental/builder/test/conv/test_ckb_conv_fwd_3d_fp32.cpp → experimental/builder/test/conv/ck/test_ckb_conv_fwd_3d_fp32.cpp
View File

0
experimental/builder/test/conv/test_conv_traits.cpp → experimental/builder/test/conv/ck/test_conv_traits.cpp
View File

52
experimental/builder/test/conv/ck_tile/test_ckb_conv_bwd_data_2d_fp16_v3.cpp Normal file
View File

@@ -0,0 +1,52 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "utils/ckb_conv_tile_test_configs.hpp"
#include "utils/ckb_conv_test_utils.hpp"
namespace {
using namespace ck_tile::builder::test_utils;
TEST(FwdConvInstances, Create_ConvAlgorithm_Tile_GroupedConvolutionKernel_2D_FP16_NHWGC)
{
constexpr ConvSignature FwdConvSignature{.spatial_dim = 2,
.direction = ConvDirection::BACKWARD_DATA,
.data_type = DataType::FP16,
.accumulation_data_type = DataType::FP32,
.input = {.config = {.layout = TensorLayout::NHWGC}},
.weight = {.config = {.layout = TensorLayout::GKYXC}},
.output = {.config = {.layout = TensorLayout::NHWGK}}};
constexpr auto FwdConvAlgorithm =
ConvAlgorithm_Tile_GroupedConvolutionKernel{}
.with_tile_specializations(TileConvSpecialization::DEFAULT)
.with_tile_thread_block(FwdTileThreadBlock_64x64x64)
.with_tile_block_gemm(TileBlockGemmDesc_16x16_v3_intrawave)
.with_tile_transfer(FwdTileTransfer_4x4x4)
.with_tile_optimizations(TileOptimizations{
.num_groups_to_merge = 1, .split_image = false, .explicit_gemm = false});
using Builder = ConvBuilder<FwdConvSignature, FwdConvAlgorithm>;
run_ck_tile_test<Builder>({
"grouped_convolution_backward_data",
"fp16",
"NHWGC_GKYXC_NHWGK",
"64x64x64",
"2x2",
"16x16x16",
// "4x4x4", // TODO: Enable this check
"Default",
"Intrawave",
"CShuffleEpilogue",
"set",
"pipeline_AgBgCrCompV3",
"DoubleSmemBuffer_0",
"NumWaveGroups_1",
"MergedGroups_1",
"SplitImage_0",
"ExplicitGemm_0",
});
}
} // namespace

52
experimental/builder/test/conv/ck_tile/test_ckb_conv_bwd_weight_2d_fp16_v3.cpp Normal file
View File

@@ -0,0 +1,52 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "utils/ckb_conv_tile_test_configs.hpp"
#include "utils/ckb_conv_test_utils.hpp"
namespace {
using namespace ck_tile::builder::test_utils;
TEST(FwdConvInstances, Create_ConvAlgorithm_Tile_GroupedConvolutionKernel_2D_FP16_NHWGC)
{
constexpr ConvSignature FwdConvSignature{.spatial_dim = 2,
.direction = ConvDirection::BACKWARD_WEIGHT,
.data_type = DataType::FP16,
.accumulation_data_type = DataType::FP32,
.input = {.config = {.layout = TensorLayout::NHWGC}},
.weight = {.config = {.layout = TensorLayout::GKYXC}},
.output = {.config = {.layout = TensorLayout::NHWGK}}};
constexpr auto FwdConvAlgorithm =
ConvAlgorithm_Tile_GroupedConvolutionKernel{}
.with_tile_specializations(TileConvSpecialization::DEFAULT)
.with_tile_thread_block(FwdTileThreadBlock_64x64x64)
.with_tile_block_gemm(TileBlockGemmDesc_16x16_v3_intrawave)
.with_tile_transfer(FwdTileTransfer_4x4x4)
.with_tile_optimizations(TileOptimizations{
.num_groups_to_merge = 1, .split_image = false, .explicit_gemm = false});
using Builder = ConvBuilder<FwdConvSignature, FwdConvAlgorithm>;
run_ck_tile_test<Builder>({
"grouped_convolution_backward_weight",
"fp16",
"NHWGC_GKYXC_NHWGK",
"64x64x64",
"2x2",
"16x16x16",
// "4x4x4", // TODO: Enable this check
"Default",
"Intrawave",
"CShuffleEpilogue",
"set",
"pipeline_AgBgCrCompV3",
"DoubleSmemBuffer_0",
"NumWaveGroups_1",
"MergedGroups_1",
"SplitImage_0",
"ExplicitGemm_0",
});
}
} // namespace

52
experimental/builder/test/conv/ck_tile/test_ckb_conv_fwd_2d_fp16_v3.cpp Normal file
View File

@@ -0,0 +1,52 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "utils/ckb_conv_tile_test_configs.hpp"
#include "utils/ckb_conv_test_utils.hpp"
namespace {
using namespace ck_tile::builder::test_utils;
TEST(FwdConvInstances, Create_ConvAlgorithm_Tile_GroupedConvolutionKernel_2D_FP16_NHWGC)
{
constexpr ConvSignature FwdConvSignature{.spatial_dim = 2,
.direction = ConvDirection::FORWARD,
.data_type = DataType::FP16,
.accumulation_data_type = DataType::FP32,
.input = {.config = {.layout = TensorLayout::NHWGC}},
.weight = {.config = {.layout = TensorLayout::GKYXC}},
.output = {.config = {.layout = TensorLayout::NHWGK}}};
constexpr auto FwdConvAlgorithm =
ConvAlgorithm_Tile_GroupedConvolutionKernel{}
.with_tile_specializations(TileConvSpecialization::DEFAULT)
.with_tile_thread_block(FwdTileThreadBlock_64x64x64)
.with_tile_block_gemm(TileBlockGemmDesc_16x16_v3_intrawave)
.with_tile_transfer(FwdTileTransfer_4x4x4)
.with_tile_optimizations(TileOptimizations{
.num_groups_to_merge = 1, .split_image = false, .explicit_gemm = false});
using Builder = ConvBuilder<FwdConvSignature, FwdConvAlgorithm>;
run_ck_tile_test<Builder>({
"grouped_convolution_forward",
"fp16",
"NHWGC_GKYXC_NHWGK",
"64x64x64",
"2x2",
"16x16x16",
// "4x4x4", // TODO: Enable this check
"Default",
"Intrawave",
"CShuffleEpilogue",
"set",
"pipeline_AgBgCrCompV3",
"DoubleSmemBuffer_0",
"NumWaveGroups_1",
"MergedGroups_1",
"SplitImage_0",
"ExplicitGemm_0",
});
}
} // namespace