Skip to content

GitLab

Unverified Commit 3049b546 authored by Haocong WANG's avatar Haocong WANG Committed by GitHub
Browse files

[GEMM] gemm_universal related optimization (#1453) 



* replace buffer_atomic with global_atomic

* fixed global_atomic_add

* added bf16 atomic_add

* format

* clang-format-12

* clean

* clean

* add guards

* Update gtest.cmake

* enabled splitk_gemm_multi_d

* format

* add ckProfiler

* format

* fixed naming

* format

* clean

* clean

* add guards

* fix clang format

* format

* add kbatch printout

* clean

* Add rocm6.2 related gemm optimization

* Limit bf16 atomic usage

* remove redundant RCR gemm_universal instance

* Add RRR fp8 gemm universal instance

* Bug fix

* Add GPU_TARGET guard to FP8/BF8 target

* bug fix

* update cmake

* remove all fp8/bf8 example if arch not support

* Enable fp8 RRR support in ckProfiler

* limit greedy-reverse flag to gemm_universal in ckProfiler

---------
Co-authored-by: default avatarJing Zhang <jizhan@fb.com>
Co-authored-by: default avatarJing Zhang <jizhan@meta.com>
Co-authored-by: default avatarzjing14 <zhangjing14@gmail.com>
Co-authored-by: default avatarIllia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: default avatarillsilin <Illia.Silin@amd.com>
parent 50c42348
Hide whitespace changes
Inline Side-by-side
Showing with 205 additions and 110 deletions
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instance.cpp 0 → 100644
View file @ 3049b546
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_instances<Interwave, GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instance.cpp 0 → 100644
View file @ 3049b546
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_instances<Interwave, GemmNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp
View file @ 3049b546
......@@ -45,8 +45,7 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_instances = std::tuple<
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 64, 16, 16, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
......@@ -72,7 +71,11 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_instances = std::tuple<
// Latency friendly
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 256, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 512, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 256, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 512, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
// Memory friendly
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 32, 128, 16, 16, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 128, 16, 16, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
......@@ -83,7 +86,11 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_instances = std::tuple<
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 16, 16, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 256, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 512, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 256, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 512, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 128, 16, 16, 16, 16, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F8, F8, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 128, 16, 16, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
View file @ 3049b546
......@@ -15,7 +15,7 @@ set(GROUPED_CONV3D_BWD_DATA
wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
list(APPEND GROUPED_CONV3D_BWD_DATA
xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp)
endif()
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
View file @ 3049b546
......@@ -30,7 +30,7 @@ list(APPEND GROUPED_CONV3D_BWD_WEIGHT
wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
list(APPEND GROUPED_CONV3D_BWD_WEIGHT
xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp)
endif()
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_bilinear/CMakeLists.txt
View file @ 3049b546
......@@ -4,7 +4,7 @@ set(GROUPED_CONV3D_BWD_WEIGHT_BILINEAR
xdl/device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
list(APPEND GROUPED_CONV3D_BWD_WEIGHT_BILINEAR
xdl/device_grouped_conv3d_bwd_weight_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp)
endif()
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/CMakeLists.txt
View file @ 3049b546
......@@ -4,7 +4,7 @@ set(GROUPED_CONV3D_BWD_WEIGHT_SCALE
xdl/device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
list(APPEND GROUPED_CONV3D_BWD_WEIGHT_SCALE
xdl/device_grouped_conv3d_bwd_weight_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp)
endif()
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
View file @ 3049b546
......@@ -43,22 +43,22 @@ set(GROUPED_CONV3D_FWD
wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp
)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
list(APPEND GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp)
endif()
if(DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
list(APPEND GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_fp8_instance.cpp)
endif()
if(DTYPES MATCHES "bf8" OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
list(APPEND GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf8_instance.cpp)
endif()
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR NOT DEFINED DTYPES)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR (NOT DEFINED DTYPES AND GPU_TARGETS MATCHES "gfx94"))
list(APPEND GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_fp8_bf8_instance.cpp)
list(APPEND GROUPED_CONV3D_FWD
......
profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
View file @ 3049b546
......@@ -48,6 +48,7 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
int StrideD0,
int StrideD1,
int StrideE,
int KBatch,
int n_warmup,
int n_iter,
uint64_t rotating = 0)
......@@ -129,17 +130,17 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
d1_device_buf.ToDevice(d1_m_n.mData.data());
using DeviceOp =
ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
AElementOp,
BElementOp,
CElementOp>;
ck::tensor_operation::device::DeviceGemmMultipleDSplitK<ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
AElementOp,
BElementOp,
CElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......@@ -182,104 +183,128 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
std::array<const void*, 2>{d0_device_buf.GetDeviceBuffer(),
d1_device_buf.GetDeviceBuffer()},
static_cast<EDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(KBatch > 0)
{
kbatch_list = {KBatch};
}
if(do_verification)
for(std::size_t i = 0; i < kbatch_list.size(); i++)
{
auto kbatch_curr = kbatch_list[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
std::array<const void*, 2>{d0_device_buf.GetDeviceBuffer(),
d1_device_buf.GetDeviceBuffer()},
static_cast<EDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
kbatch_curr,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
c_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_log)
if(do_verification)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", e_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", e_m_n_device_result.mData, ",")
<< std::endl;
c_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", e_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", e_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
std::string op_name = op_ptr->GetTypeString();
std::string op_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(
argument_ptr.get(),
StreamConfig{
nullptr, time_kernel, 0, n_warmup, n_iter, rotating_count > 1, rotating_count});
float ave_time = invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr,
time_kernel,
0,
n_warmup,
n_iter,
rotating_count > 1,
rotating_count});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
#if defined CK_ENABLE_FP8
// set softer tolerances for fp8
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
is_same_v<EDataType, f8_t>)
{
std::string msg = "Error: Incorrect results!";
double rtol = 1e-1;
double atol = 1e-1;
pass = pass & ck::utils::check_err(
e_m_n_device_result, e_m_n_host_result, msg, rtol, atol);
}
else
{
// set softer tolerances for fp8
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
is_same_v<EDataType, f8_t>)
{
std::string msg = "Error: Incorrect results!";
double rtol = 1e-1;
double atol = 1e-1;
pass = pass & ck::utils::check_err(
e_m_n_device_result, e_m_n_host_result, msg, rtol, atol);
}
else
{
#endif
pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
#if defined CK_ENABLE_FP8
}
}
#endif
if(tflops > best_tflops)
if(tflops > best_tflops && ave_time > 1e-10)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
else
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
std::cout << op_ptr->GetTypeString() << " does not support this problem"
<< std::endl;
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
}
if constexpr(is_same<EDataType, float>::value)
......@@ -318,9 +343,9 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideE = " << StrideE << " : " << best_ave_time
<< " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
<< " StrideB = " << StrideB << " StrideE = " << StrideE << " KBatch = " << best_kbatch
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
return pass;
}
......
profiler/include/profiler/profile_gemm_universal_impl.hpp
View file @ 3049b546
......@@ -152,7 +152,7 @@ bool profile_gemm_universal_impl(int do_verification,
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 19, 20, 32, 38};
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
if(KBatch > 0)
{
......@@ -249,7 +249,7 @@ bool profile_gemm_universal_impl(int do_verification,
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
if(tflops > best_tflops)
if(tflops > best_tflops && ave_time > 1e-10)
{
best_op_name = op_name;
best_tflops = tflops;
......
profiler/src/profile_gemm_multiply_multiply.cpp
View file @ 3049b546
......@@ -34,7 +34,7 @@ enum struct GemmDataType
int profile_gemm_multiply_multiply(int argc, char* argv[])
{
if(argc != 16 && argc != 19)
if(argc != 16 && argc != 20)
{
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
......@@ -50,9 +50,10 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
printf("optional:\n");
printf("arg16: number of warm-up cycles (default 1)\n");
printf("arg17: number of iterations (default 10)\n");
printf("arg18: memory for rotating buffer (default 0, size in MB)\n");
printf("arg16: number of kbatch (default 1)\n");
printf("arg17: number of warm-up cycles (default 1)\n");
printf("arg18: number of iterations (default 10)\n");
printf("arg19: memory for rotating buffer (default 0, size in MB)\n");
exit(1);
}
......@@ -76,11 +77,13 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
int n_warmup = 1;
int n_iter = 10;
uint64_t rotating = 0;
if(argc == 19)
int KBatch = 1;
if(argc == 20)
{
n_warmup = std::stoi(argv[16]);
n_iter = std::stoi(argv[17]);
rotating = std::stoull(argv[18]) * 1024 * 1024;
KBatch = std::stoi(argv[16]);
n_warmup = std::stoi(argv[17]);
n_iter = std::stoi(argv[18]);
rotating = std::stoull(argv[19]) * 1024 * 1024;
}
using F32 = float;
......@@ -146,6 +149,7 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
(StrideE < 0) ? DefaultStrideE : StrideE,
KBatch,
n_warmup,
n_iter,
rotating);
......
profiler/src/profile_gemm_universal.cpp
View file @ 3049b546
......@@ -171,6 +171,10 @@ int profile_gemm_universal(int argc, char* argv[])
{
return profile(BF16{}, BF16{}, BF16{}, F32{}, BF16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Col{}, Row{});
......
profiler/src/profile_grouped_gemm_fixed_nk.cpp
View file @ 3049b546
......@@ -85,9 +85,11 @@ int profile_grouped_gemm_fixed_nk(int argc, char* argv[])
const auto StrideCs = argToIntArray(argv[13]);
const int kbatch = argc == 15 ? std::stoi(argv[14]) : 1;
using F32 = float;
using F16 = ck::half_t;
using F8 = ck::f8_t;
using F32 = float;
using F16 = ck::half_t;
#if defined(CK_ENABLE_FP8)
using F8 = ck::f8_t;
#endif
using BF16 = ck::bhalf_t;
using I8 = int8_t;
......
test/gemm_universal/test_gemm_universal_xdl.cpp
View file @ 3049b546
......@@ -44,17 +44,22 @@ class TestGemmUniversal_MK_NK
using KernelTypes_MK_KN = ::testing::Types<
// ADataType, BDataType, ComputeDataType, CDataType
std::tuple< F16, F16, F16, F16>,
#if (defined CK_ENABLE_FP8)
std::tuple< F16, F8, F16, F16>,
std::tuple< F8, F16, F16, F16>,
std::tuple< F8, F8, F8, BF16>,
#endif
std::tuple< BF16, BF16, BF16, BF16>
>;
using KernelTypes_MK_NK = ::testing::Types<
// ADataType, BDataType, ComputeDataType, CDataType
std::tuple< F16, F16, F16, F16>,
#if (defined CK_ENABLE_FP8)
std::tuple< F16, F8, F16, F16>,
std::tuple< F8, F16, F16, F16>,
std::tuple< BF16, BF16, BF16, BF16>,
std::tuple< F8, F8, F8, BF16>
std::tuple< F8, F8, F8, BF16>,
#endif
std::tuple< BF16, BF16, BF16, BF16>
>;
// clang-format on
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment