ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-12 18:17:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add more instances.

Browse Source
This commit is contained in:
Ville Pietilä
2025-10-16 14:18:40 +00:00
parent 6c5531a4ae
commit c0b68c8a85
2 changed files with 11 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_bf16_instances.hpp
View File

@@ -9,8 +9,6 @@ namespace ck_tile {
namespace ops {
using BF16 = ck_tile::bfloat16_t;
using F16 = ck_tile::half_t;
using F32 = float;
template <ck_tile::index_t NDimSpatial,
typename ALayout,
@@ -24,12 +22,19 @@ using tile_grouped_conv_bwd_weight_bf16_instances = std::tuple<
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 64, 64, 64, 2, 2, 1, 32, 32, 16, 2, 4, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 64, 64, 64, 2, 2, 1, 32, 32, 8, 2, 4, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 64, 64, 64, 2, 2, 1, 16, 16, 32, 2, 4, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
// GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 8, 128, 64, 2, 2, 1, 4, 64, 16, 2, 4, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
// GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 128, 8, 64, 2, 2, 1, 64, 4, 16, 2, 4, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 64, 64, 64, 2, 2, 1, 16, 16, 32, 4, 4, 4, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 64, 64, 64, 2, 2, 1, 16, 16, 32, 8, 8, 8, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 1, 64, 64, 64, 1, 1, 1, 32, 32, 8, 2, 2, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 1, 64, 64, 64, 1, 1, 1, 32, 32, 16, 2, 2, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 1, 64, 64, 64, 1, 1, 1, 16, 16, 32, 2, 2, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>
// clang-format on
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 1, 64, 64, 64, 1, 1, 1, 16, 16, 32, 2, 2, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 1, 16, 16, 64, 1, 1, 1, 16, 16, 32, 8, 8, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 16, 16, 64, 1, 1, 1, 16, 16, 32, 8, 8, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 1, 16, 16, 64, 1, 1, 1, 16, 16, 32, 4, 4, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 16, 16, 64, 1, 1, 1, 16, 16, 32, 4, 4, 2, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 1, 16, 16, 64, 1, 1, 1, 16, 16, 32, 4, 4, 4, false, CK_TILE_PIPELINE_COMPUTE_V3>,
GroupedConvolutionBackwardWeightInvoker<NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough, 2, 16, 16, 64, 1, 1, 1, 16, 16, 32, 4, 4, 4, false, CK_TILE_PIPELINE_COMPUTE_V3>
// clang-format on
>;
} // namespace ops

2
library/include/ck_tile/library/tensor_operation_instance/gpu/tile_grouped_conv_bwd_weight_fp16_instances.hpp
View File

@@ -8,9 +8,7 @@
namespace ck_tile {
namespace ops {
using BF16 = ck_tile::bfloat16_t;
using F16 = ck_tile::half_t;
using F32 = float;
template <ck_tile::index_t NDimSpatial,
typename ALayout,