ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

[rocm-libraries] ROCm/rocm-libraries#4406 (commit 61f9f90)

Browse Source 
[CK] CK Tile grouped convolution direct load

## Motivation

CK Tile grouped convolution forward direct load support.

## Technical Details

Basic pipeline for direct load and new instances for forward for v1 and
v4 pipelines.

## Test Plan

test_grouped_convnd_fwd_tile

## Test Result

CI pending

## Submission Checklist

- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
AICK-130
This commit is contained in:
Bartłomiej Kocot
2026-02-09 21:09:42 +00:00
committed by assistant-librarian[bot]
parent 0cafa68b6f
commit 27e0a34e0f
29 changed files with 739 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
experimental/builder/include/ck_tile/builder/factory/helpers/ck/conv_tuning_params.hpp
View File

@@ -58,6 +58,8 @@ consteval BlockGemmSpec SetBlockGemm()
case PipelineVersion::V3: version = ck::BlockGemmPipelineVersion::v3; break;
case PipelineVersion::V4: version = ck::BlockGemmPipelineVersion::v4; break;
case PipelineVersion::V5: version = ck::BlockGemmPipelineVersion::v5; break;
case PipelineVersion::ASYNC_V1: throw "PipelineVersion::ASYNC can be used only for CK TILE.";
case PipelineVersion::ASYNC_V4: throw "PipelineVersion::ASYNC can be used only for CK TILE.";
case PipelineVersion::V6: throw "PipelineVersion::V6 is supported only for CK Tile.";
case PipelineVersion::WEIGHT_ONLY:
throw "PipelineVersion::WEIGHT_ONLY is not supported for block GEMM.";
@@ -93,6 +95,8 @@ consteval ck::PipelineVersion SetGridwiseGemmPipelineVersion()
case PipelineVersion::V3: throw "PipelineVersion::V3 is used only for stream-K.";
case PipelineVersion::V4: return ck_pipeline::v4;
case PipelineVersion::V5: throw "PipelineVersion::V5 cannot be used for gridwise GEMM.";
case PipelineVersion::ASYNC_V1: throw "PipelineVersion::ASYNC can be used only for CK TILE.";
case PipelineVersion::ASYNC_V4: throw "PipelineVersion::ASYNC can be used only for CK TILE.";
case PipelineVersion::V6: throw "PipelineVersion::V6 can be used only for CK TILE.";
case PipelineVersion::WEIGHT_ONLY: return ck_pipeline::weight_only;
default: throw "Unknown GridwiseGemmPipelineVersion";
@@ -139,6 +143,8 @@ consteval ck::BlockGemmPipelineVersion SetBlockGemmPipelineVersion()
case PipelineVersion::V3: return ck_pipeline::v3;
case PipelineVersion::V4: return ck_pipeline::v4;
case PipelineVersion::V5: return ck_pipeline::v5;
case PipelineVersion::ASYNC_V1: throw "PipelineVersion::ASYNC can be used only for CK TILE.";
case PipelineVersion::ASYNC_V4: throw "PipelineVersion::ASYNC can be used only for CK TILE.";
case PipelineVersion::V6: throw "PipelineVersion::V6 is supported only for CK Tile.";
case PipelineVersion::WEIGHT_ONLY:
throw "PipelineVersion::WEIGHT_ONLY is not supported for block GEMM pipeline version.";

16
experimental/builder/include/ck_tile/builder/factory/helpers/ck_tile/conv_tile_tuning_params.hpp
View File

@@ -98,6 +98,20 @@ struct TilePipelineType<ck_tile::GemmPipeline::COMPUTE_V6>
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV6<PipelineProblem>;
};
template <>
struct TilePipelineType<ck_tile::GemmPipeline::COMPUTE_ASYNC>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompAsync<PipelineProblem>;
};
template <>
struct TilePipelineType<ck_tile::GemmPipeline::BASIC_ASYNC_V1>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegAsyncV1<PipelineProblem>;
};
template <ConvAlgorithmDescriptor auto ALGORITHM>
consteval ck_tile::GemmPipeline SetTileBlockGemmPipelineVersion()
{
@@ -111,6 +125,8 @@ consteval ck_tile::GemmPipeline SetTileBlockGemmPipelineVersion()
case PipelineVersion::V4: return ck_tile_pipeline::COMPUTE_V4;
case PipelineVersion::V5: return ck_tile_pipeline::COMPUTE_V5;
case PipelineVersion::V6: return ck_tile_pipeline::COMPUTE_V6;
case PipelineVersion::ASYNC_V1: return ck_tile_pipeline::BASIC_ASYNC_V1;
case PipelineVersion::ASYNC_V4: return ck_tile_pipeline::COMPUTE_ASYNC;
case PipelineVersion::WEIGHT_ONLY:
throw "PipelineVersion::WEIGHT_ONLY is not supported for block GEMM pipeline version.";
default: throw "Unknown block GEMM PipelineVersion";

5
experimental/builder/include/ck_tile/builder/testing/tensor_initialization.hpp
View File

@@ -54,14 +54,15 @@ void init_tensor_buffer_uniform_int(void* buf,
// we might be asked to generate int values on fp data types that don't have the required
// precision
if(static_cast<ck_type>(max_value - 1) == static_cast<ck_type>(min_value))
if(static_cast<ck_type>(max_value - 1) <= static_cast<ck_type>(min_value) &&
static_cast<ck_type>(max_value - 1) >= static_cast<ck_type>(min_value))
{
throw std::runtime_error("Error while filling device tensor with random integer data: "
"insufficient precision in specified range");
}
size_t packed_size = ck::packed_size_v<ck_type>;
fill_tensor_uniform_rand_int_values<<<256, 256>>>(
static_cast<ck_type>(buf), min_value, max_value, (size * packed_size) / sizeof(ck_type));
static_cast<ck_type*>(buf), min_value, max_value, (size * packed_size) / sizeof(ck_type));
}
/// @brief Initialize tensor data with a uniform float distribution

4
experimental/builder/include/ck_tile/builder/types.hpp
View File

@@ -158,6 +158,8 @@ enum class PipelineVersion
V4,
V5,
V6,
ASYNC_V1,
ASYNC_V4,
WEIGHT_ONLY
};
@@ -330,6 +332,8 @@ inline std::string_view to_string(PipelineVersion ver)
case V4: return "V4";
case V5: return "V5";
case V6: return "V6";
case ASYNC_V1: return "ASYNC_V1";
case ASYNC_V4: return "ASYNC_V4";
case WEIGHT_ONLY: return "WEIGHT_ONLY";
default: return "Unknown";
}

28
experimental/grouped_convolution_tile_instances/configs/profiler/ndhwgc_bf16.conf
View File

@@ -234,4 +234,30 @@ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 64, 64, Filter1x1Stride
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 16, 128, 64, Filter1x1Stride1Pad0, 16, 16, 1, 4, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 128, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 16, 256, 64, Filter1x1Stride1Pad0, 16, 16, 1, 4, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 32, 256, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 32, 256, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 64, 64, Default, 16, 16, 1, 2, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 256, 32, 64, Default, 32, 32, 2, 1, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 128, 64, Default, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 32, 16, 64, Default, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 64, 64, Filter1x1Stride1Pad0, 16, 16, 1, 2, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 128, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 32, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>

28
experimental/grouped_convolution_tile_instances/configs/profiler/ndhwgc_fp16.conf
View File

@@ -225,4 +225,30 @@ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 64, 64, Filter1x1Stride
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 16, 128, 64, Filter1x1Stride1Pad0, 16, 16, 1, 4, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 128, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 16, 256, 64, Filter1x1Stride1Pad0, 16, 16, 1, 4, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 32, 256, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 32, 256, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 64, 64, Default, 16, 16, 1, 2, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 256, 32, 64, Default, 32, 32, 2, 1, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 128, 64, Default, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 32, 16, 64, Default, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 64, 64, Filter1x1Stride1Pad0, 16, 16, 1, 2, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 128, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 32, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>

28
experimental/grouped_convolution_tile_instances/configs/profiler/nhwgc_bf16.conf
View File

@@ -234,4 +234,30 @@ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 64, 64, Filter1x1Stride
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 16, 128, 64, Filter1x1Stride1Pad0, 16, 16, 1, 4, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 128, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 16, 256, 64, Filter1x1Stride1Pad0, 16, 16, 1, 4, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 32, 256, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 32, 256, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 64, 64, Default, 16, 16, 1, 2, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 256, 32, 64, Default, 32, 32, 2, 1, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 128, 64, Default, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 32, 16, 64, Default, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 64, 64, Filter1x1Stride1Pad0, 16, 16, 1, 2, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 128, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 32, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>

28
experimental/grouped_convolution_tile_instances/configs/profiler/nhwgc_fp16.conf
View File

@@ -225,4 +225,30 @@ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 64, 64, Filter1x1Stride
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 16, 128, 64, Filter1x1Stride1Pad0, 16, 16, 1, 4, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 128, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 16, 256, 64, Filter1x1Stride1Pad0, 16, 16, 1, 4, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 32, 256, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 32, 256, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 64, 64, Default, 16, 16, 1, 2, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 256, 32, 64, Default, 32, 32, 2, 1, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 128, 64, Default, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Default, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 32, 16, 64, Default, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 64, 64, Filter1x1Stride1Pad0, 16, 16, 1, 2, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 128, 64, Filter1x1Stride1Pad0, 32, 32, 1, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 16, 32, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<128, 32, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>

10
experimental/grouped_convolution_tile_instances/configs/tests/ndhwgc_bf16.conf
View File

@@ -38,4 +38,12 @@ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Default, 32, 3
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 16, 64, Filter1x1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>

10
experimental/grouped_convolution_tile_instances/configs/tests/ndhwgc_fp16.conf
View File

@@ -38,4 +38,12 @@ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Default, 32, 3
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 16, 64, Filter1x1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>

10
experimental/grouped_convolution_tile_instances/configs/tests/nhwgc_bf16.conf
View File

@@ -38,4 +38,12 @@ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Default, 32, 3
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 16, 64, Filter1x1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>

10
experimental/grouped_convolution_tile_instances/configs/tests/nhwgc_fp16.conf
View File

@@ -38,4 +38,12 @@ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Default, 32, 3
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 16, 64, Filter1x1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<128, 32, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 2, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<256, 256, 32, 64, Filter1x1Stride1Pad0, 32, 32, 2, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Default, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Default, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 8, 8, 8, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 128, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<64, 16, 16, 64, Filter1x1Stride1Pad0, 16, 16, 1, 1, 8, 8, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3_DirectLoad<256, 64, 64, 64, Filter1x1Stride1Pad0, 16, 16, 2, 2, 2, 2, 4, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1>

13
experimental/grouped_convolution_tile_instances/generate_instances.py
View File

@@ -60,7 +60,6 @@ class ConvInstanceTemplateParams:
def get_block_gemm_desc(self):
double_smem_buffer = "true" if self.double_smem_buffer else "false"
pipeline_version = self.pipeline_version[-1:]
scheduler = (
"INTRAWAVE" if self.scheduler.find("Intrawave") != -1 else "INTERWAVE"
)
@@ -69,7 +68,7 @@ class ConvInstanceTemplateParams:
.warp_tile = {{.m = {self.warp_tile[0]}, .n = {self.warp_tile[1]}, .k = {self.warp_tile[2]}}},
.double_smem_buffer = {double_smem_buffer},
.num_wave_groups = {self.num_wave_groups},
.pipeline_version = ckb::PipelineVersion::V{pipeline_version},
.pipeline_version = ckb::PipelineVersion::{self.pipeline_version},
.scheduler = ckb::PipelineScheduler::{scheduler}}}"""
def get_block_transfer(self):
@@ -180,6 +179,16 @@ def parse_fwd_instances(instances, problem_name):
pipeline_version = (
"v1" if instance.find("BlkGemmPipelineVersion") == -1 else args[15]
)
# Replace pipeline if Direct Load
if instance.find("DirectLoad") != -1:
if instance.find("BlkGemmPipelineVersion: v1") != -1:
pipeline_version = "ASYNC_V1"
elif instance.find("BlkGemmPipelineVersion: v4") != -1:
pipeline_version = "ASYNC_V4"
else:
raise RuntimeError("not supported pipeline for direct load")
else:
pipeline_version = f"""V{pipeline_version[-1:]}"""
m_warp = int(m_per_block / (m_per_xdl * m_xdl_per_wave))
n_warp = int(n_per_block / (n_per_xdl * n_xdl_per_wave))