Skip to content

GitLab

".github/vscode:/vscode.git/clone" did not exist on "976bc302e52b12d1d2e581cc5d8a952ac1c6b0a4"
Commit 0eb75e21 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into ck_tile/moe

parents 1b4b640b c8b6b642
Hide whitespace changes
Inline Side-by-side
Showing with 476 additions and 290 deletions
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
View file @ 0eb75e21
......@@ -9,11 +9,11 @@ add_instance_library(device_grouped_conv2d_fwd_instance
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
# merged groups
# large tensor
# NHWGC, GKYXC, NHWGK
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instance.cpp
#mem
# NHWGC, GKYXC, NHWGK
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instance.cpp → library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
View file @ 0eb75e21
......@@ -2,44 +2,35 @@
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
BF16,
BF16,
Empty_Tuple,
F16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f16_instances<2,
device_grouped_conv_fwd_xdl_large_tensor_bf16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwd3x3>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp → library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instance.cpp
View file @ 0eb75e21
......@@ -2,44 +2,35 @@
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
F16,
F16,
Empty_Tuple,
BF16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwd3x3>{});
device_grouped_conv_fwd_xdl_large_tensor_f16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instance.cpp → library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instance.cpp
View file @ 0eb75e21
......@@ -2,14 +2,14 @@
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instances(
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
......@@ -25,21 +25,12 @@ void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_insta
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f32_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f32_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwd3x3>{});
device_grouped_conv_fwd_xdl_large_tensor_f32_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
View file @ 0eb75e21
......@@ -9,9 +9,9 @@ set(GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/mem/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp
xdl/mem/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_mem_inter_instance.cpp
......@@ -48,12 +48,12 @@ if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
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)
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)
list(APPEND GROUPED_CONV3D_FWD
xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf8_instance.cpp)
endif()
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp → library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
......@@ -9,36 +9,28 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instances(
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F32,
F32,
BF16,
BF16,
Empty_Tuple,
F32,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f32_instances<3,
device_grouped_conv_fwd_xdl_large_tensor_bf16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f32_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwd3x3>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp → library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
......@@ -9,7 +9,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instances(
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
......@@ -25,20 +25,12 @@ void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_in
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwd3x3>{});
device_grouped_conv_fwd_xdl_large_tensor_f16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp → library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
......@@ -9,36 +9,28 @@ namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
F32,
F32,
Empty_Tuple,
BF16,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwd3x3>{});
device_grouped_conv_fwd_xdl_large_tensor_f32_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/mha/CMakeLists.txt 0 → 100644
View file @ 0eb75e21
set(FMHA_CPP_FOLDER ${CMAKE_CURRENT_BINARY_DIR})
set(FMHA_SRC_FOLDER ${CMAKE_SOURCE_DIR}/example/ck_tile/01_fmha/)
set(CK_TILE_SRC_FOLDER ${CMAKE_SOURCE_DIR}/include/ck_tile/)
# python stuff
find_package(PythonInterp 3 REQUIRED)
rocm_install(DIRECTORY ${CK_TILE_SRC_FOLDER} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck_tile)
rocm_install(FILES
"${FMHA_SRC_FOLDER}/fmha_fwd.hpp"
"${FMHA_SRC_FOLDER}/bias.hpp"
"${FMHA_SRC_FOLDER}/mask.hpp"
DESTINATION include/ck_tile/ops
)
# header for building lib
file(COPY ${FMHA_SRC_FOLDER}/fmha_fwd.hpp DESTINATION ${FMHA_CPP_FOLDER})
file(COPY ${FMHA_SRC_FOLDER}/bias.hpp DESTINATION ${FMHA_CPP_FOLDER})
file(COPY ${FMHA_SRC_FOLDER}/mask.hpp DESTINATION ${FMHA_CPP_FOLDER})
# generate a list of kernels, but not actually emit files at config stage
execute_process(
COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_SOURCE_DIR}/example/ck_tile/01_fmha/generate.py
--list_blobs ${FMHA_CPP_FOLDER}/blob_list.txt
)
file(STRINGS ${FMHA_CPP_FOLDER}/blob_list.txt FMHA_FWD_GEN_BLOBS)
# actually generate the cpp files
add_custom_command(
OUTPUT ${FMHA_FWD_GEN_BLOBS}
COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_SOURCE_DIR}/example/ck_tile/01_fmha/generate.py
--output_dir ${FMHA_CPP_FOLDER}
COMMENT "Generating mha kernel (cpp) files now ..."
VERBATIM
)
# This is done to remove path info and just
# have filename. Since, it was cauing the cmake
# to throw "File name too long"
set(device_files)
foreach(filepath IN LISTS FMHA_FWD_GEN_BLOBS)
get_filename_component(filename ${filepath} NAME)
# Append the filename to the device_files list
list(APPEND device_files ${filename})
endforeach()
add_custom_target(generate_cpp_files DEPENDS ${FMHA_FWD_GEN_BLOBS})
add_instance_library(device_mha_instance ${device_files})
if (TARGET device_mha_instance)
add_dependencies(device_mha_instance generate_cpp_files)
endif()
library/src/utility/convolution_parameter.cpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/host_utility/io.hpp"
......@@ -20,6 +20,63 @@ ConvParam::ConvParam(ck::index_t n_dim,
const std::vector<ck::index_t>& dilations,
const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads)
: num_dim_spatial_(static_cast<ck::long_index_t>(n_dim)),
G_(static_cast<ck::long_index_t>(group_count)),
N_(static_cast<ck::long_index_t>(n_batch)),
K_(static_cast<ck::long_index_t>(n_out_channels)),
C_(static_cast<ck::long_index_t>(n_in_channels)),
filter_spatial_lengths_(num_dim_spatial_),
input_spatial_lengths_(num_dim_spatial_),
output_spatial_lengths_(num_dim_spatial_),
conv_filter_strides_(num_dim_spatial_),
conv_filter_dilations_(num_dim_spatial_),
input_left_pads_(num_dim_spatial_),
input_right_pads_(num_dim_spatial_)
{
if(static_cast<ck::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
static_cast<ck::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
static_cast<ck::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
static_cast<ck::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
static_cast<ck::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
static_cast<ck::index_t>(input_right_pads_.size()) != num_dim_spatial_)
{
throw(
std::runtime_error("ConvParam::ConvParam: "
"parameter size is different from number of declared dimensions!"));
}
for(ck::index_t i = 0; i < num_dim_spatial_; ++i)
{
filter_spatial_lengths_[i] = static_cast<ck::long_index_t>(filters_len[i]);
input_spatial_lengths_[i] = static_cast<ck::long_index_t>(input_len[i]);
conv_filter_strides_[i] = static_cast<ck::long_index_t>(strides[i]);
conv_filter_dilations_[i] = static_cast<ck::long_index_t>(dilations[i]);
input_left_pads_[i] = static_cast<ck::long_index_t>(left_pads[i]);
input_right_pads_[i] = static_cast<ck::long_index_t>(right_pads[i]);
// XEff = (X - 1) * conv_dilation_w + 1;
// Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const ck::long_index_t x_eff =
(filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
output_spatial_lengths_[i] =
(input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - x_eff) /
conv_filter_strides_[i] +
1;
}
}
ConvParam::ConvParam(ck::long_index_t n_dim,
ck::long_index_t group_count,
ck::long_index_t n_batch,
ck::long_index_t n_out_channels,
ck::long_index_t n_in_channels,
const std::vector<ck::long_index_t>& filters_len,
const std::vector<ck::long_index_t>& input_len,
const std::vector<ck::long_index_t>& strides,
const std::vector<ck::long_index_t>& dilations,
const std::vector<ck::long_index_t>& left_pads,
const std::vector<ck::long_index_t>& right_pads)
: num_dim_spatial_(n_dim),
G_(group_count),
N_(n_batch),
......@@ -49,7 +106,8 @@ ConvParam::ConvParam(ck::index_t n_dim,
{
// XEff = (X - 1) * conv_dilation_w + 1;
// Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const ck::index_t x_eff = (filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
const ck::long_index_t x_eff =
(filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
output_spatial_lengths_[i] =
(input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - x_eff) /
......@@ -63,7 +121,7 @@ ConvParam::ConvParam()
{
}
std::vector<ck::index_t> ConvParam::GetOutputSpatialLengths() const
std::vector<ck::long_index_t> ConvParam::GetOutputSpatialLengths() const
{
return output_spatial_lengths_;
}
......@@ -97,46 +155,46 @@ std::string get_conv_param_parser_helper_msg()
ck::utils::conv::ConvParam parse_conv_param(int num_dim_spatial, int arg_idx, char* const argv[])
{
const ck::index_t G = std::stoi(argv[arg_idx++]);
const ck::index_t N = std::stoi(argv[arg_idx++]);
const ck::index_t K = std::stoi(argv[arg_idx++]);
const ck::index_t C = std::stoi(argv[arg_idx++]);
std::vector<ck::index_t> filter_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> input_spatial_lengths(num_dim_spatial);
std::vector<ck::index_t> conv_filter_strides(num_dim_spatial);
std::vector<ck::index_t> conv_filter_dilations(num_dim_spatial);
std::vector<ck::index_t> input_left_pads(num_dim_spatial);
std::vector<ck::index_t> input_right_pads(num_dim_spatial);
const ck::long_index_t G = std::stol(argv[arg_idx++]);
const ck::long_index_t N = std::stol(argv[arg_idx++]);
const ck::long_index_t K = std::stol(argv[arg_idx++]);
const ck::long_index_t C = std::stol(argv[arg_idx++]);
std::vector<ck::long_index_t> filter_spatial_lengths(num_dim_spatial);
std::vector<ck::long_index_t> input_spatial_lengths(num_dim_spatial);
std::vector<ck::long_index_t> conv_filter_strides(num_dim_spatial);
std::vector<ck::long_index_t> conv_filter_dilations(num_dim_spatial);
std::vector<ck::long_index_t> input_left_pads(num_dim_spatial);
std::vector<ck::long_index_t> input_right_pads(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
filter_spatial_lengths[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
input_spatial_lengths[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
conv_filter_strides[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
conv_filter_dilations[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_left_pads[i] = std::stoi(argv[arg_idx++]);
input_left_pads[i] = std::stol(argv[arg_idx++]);
}
for(int i = 0; i < num_dim_spatial; ++i)
{
input_right_pads[i] = std::stoi(argv[arg_idx++]);
input_right_pads[i] = std::stol(argv[arg_idx++]);
}
return ck::utils::conv::ConvParam{num_dim_spatial,
......
profiler/include/profiler/profile_conv_bwd_data_impl.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -82,6 +82,29 @@ bool profile_conv_bwd_data_impl(int do_verification,
Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
Tensor<OutDataType> output(out_g_n_k_wos_desc);
std::vector<ck::index_t> input_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> filter_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> output_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> conv_filter_strides_i32(NDimSpatial);
std::vector<ck::index_t> conv_filter_dilations_i32(NDimSpatial);
std::vector<ck::index_t> input_left_pads_i32(NDimSpatial);
std::vector<ck::index_t> input_right_pads_i32(NDimSpatial);
for(ck::index_t d = 0; d < NDimSpatial; d++)
{
input_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.input_spatial_lengths_[d]);
filter_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.filter_spatial_lengths_[d]);
output_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.GetOutputSpatialLengths()[d]);
conv_filter_strides_i32[d] = static_cast<ck::index_t>(conv_param.conv_filter_strides_[d]);
conv_filter_dilations_i32[d] =
static_cast<ck::index_t>(conv_param.conv_filter_dilations_[d]);
input_left_pads_i32[d] = static_cast<ck::index_t>(conv_param.input_left_pads_[d]);
input_right_pads_i32[d] = static_cast<ck::index_t>(conv_param.input_right_pads_[d]);
}
std::cout << "input: " << input_host_result.mDesc << std::endl;
std::cout << "weight: " << weight.mDesc << std::endl;
std::cout << "output: " << output.mDesc << std::endl;
......@@ -161,16 +184,16 @@ bool profile_conv_bwd_data_impl(int do_verification,
op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
conv_param.N_,
conv_param.K_,
conv_param.C_,
conv_param.input_spatial_lengths_,
conv_param.filter_spatial_lengths_,
conv_param.output_spatial_lengths_,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
static_cast<ck::index_t>(conv_param.N_),
static_cast<ck::index_t>(conv_param.K_),
static_cast<ck::index_t>(conv_param.C_),
input_spatial_lengths_i32,
filter_spatial_lengths_i32,
output_spatial_lengths_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
in_element_op,
wei_element_op,
out_element_op);
......
profiler/include/profiler/profile_conv_fwd_impl.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -60,6 +60,29 @@ bool profile_conv_fwd_impl(int do_verification,
Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
std::vector<ck::index_t> input_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> filter_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> output_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> conv_filter_strides_i32(NDimSpatial);
std::vector<ck::index_t> conv_filter_dilations_i32(NDimSpatial);
std::vector<ck::index_t> input_left_pads_i32(NDimSpatial);
std::vector<ck::index_t> input_right_pads_i32(NDimSpatial);
for(ck::index_t d = 0; d < NDimSpatial; d++)
{
input_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.input_spatial_lengths_[d]);
filter_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.filter_spatial_lengths_[d]);
output_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.GetOutputSpatialLengths()[d]);
conv_filter_strides_i32[d] = static_cast<ck::index_t>(conv_param.conv_filter_strides_[d]);
conv_filter_dilations_i32[d] =
static_cast<ck::index_t>(conv_param.conv_filter_dilations_[d]);
input_left_pads_i32[d] = static_cast<ck::index_t>(conv_param.input_left_pads_[d]);
input_right_pads_i32[d] = static_cast<ck::index_t>(conv_param.input_right_pads_[d]);
}
std::cout << "input: " << input.mDesc << std::endl;
std::cout << "weight: " << weight.mDesc << std::endl;
std::cout << "output: " << host_output.mDesc << std::endl;
......@@ -143,16 +166,16 @@ bool profile_conv_fwd_impl(int do_verification,
op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
conv_param.N_,
conv_param.K_,
conv_param.C_,
conv_param.input_spatial_lengths_,
conv_param.filter_spatial_lengths_,
conv_param.GetOutputSpatialLengths(),
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
static_cast<ck::index_t>(conv_param.N_),
static_cast<ck::index_t>(conv_param.K_),
static_cast<ck::index_t>(conv_param.C_),
input_spatial_lengths_i32,
filter_spatial_lengths_i32,
output_spatial_lengths_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
in_element_op,
wei_element_op,
out_element_op);
......
profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
View file @ 0eb75e21
......@@ -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 @ 0eb75e21
......@@ -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/include/profiler/profile_grouped_conv_fwd_impl.hpp
View file @ 0eb75e21
......@@ -33,7 +33,8 @@ template <ck::index_t NDimSpatial,
typename WeiDataType,
typename OutDataType,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
typename BComputeType = AComputeType,
typename IndexType = ck::index_t>
bool profile_grouped_conv_fwd_impl(int do_verification,
int init_method,
bool do_log,
......@@ -57,16 +58,16 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
std::array<IndexType, NDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<IndexType, NDimSpatial + 3> a_g_n_c_wis_strides{};
std::array<IndexType, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<IndexType, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<IndexType, NDimSpatial + 3> e_g_n_k_wos_lengths{};
std::array<IndexType, NDimSpatial + 3> e_g_n_k_wos_strides{};
std::array<IndexType, NDimSpatial> conv_filter_strides{};
std::array<IndexType, NDimSpatial> conv_filter_dilations{};
std::array<IndexType, NDimSpatial> input_left_pads{};
std::array<IndexType, NDimSpatial> input_right_pads{};
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
......
profiler/src/CMakeLists.txt
View file @ 0eb75e21
......@@ -46,8 +46,10 @@ if(GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
endif()
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
if(GPU_TARGETS MATCHES "gfx94")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
endif()
list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
......@@ -82,6 +84,11 @@ set(PROFILER_EXECUTABLE ckProfiler)
add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
# flags to compress the library
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
message("Adding --offload-compress flag for ${PROFILER_EXECUTABLE}")
target_compile_options(${PROFILER_EXECUTABLE} PRIVATE --offload-compress)
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
......@@ -123,8 +130,10 @@ if(GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
if(GPU_TARGETS MATCHES "gfx94")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
......
profiler/src/profile_gemm_multiply_multiply.cpp
View file @ 0eb75e21
......@@ -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 @ 0eb75e21
......@@ -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_conv_fwd.cpp
View file @ 0eb75e21
......@@ -29,6 +29,12 @@ enum struct ConvDataType
BF8_F8_F8, // 7
};
enum struct IndexType
{
INDEX_T, // 0
LONG_INDEX_T, // 1
};
#define OP_NAME "grouped_conv_fwd"
#define OP_DESC "Grouped Convolution Forward"
......@@ -45,12 +51,13 @@ static void print_helper_msg()
<< " 5: Input bf8, Weight bf8, Output fp8\n"
<< " 6: Input fp8, Weight bf8, Output fp8\n"
<< " 7: Input bf8, Weight fp8, Output fp8)\n"
<< "arg3: tensor layout (0: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]\n"
<< "arg3: indexing data type (0: 32-bit, 1: 64-bit)\n"
<< "arg4: tensor layout (0: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]\n"
<< " 1: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K])\n"
<< "arg4: verification (0: no, 1: yes)\n"
<< "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg5: verification (0: no, 1: yes)\n"
<< "arg6: initialization (0: no init, 1: integer value, 2: decimal value)\n"
<< "arg7: print tensor value (0: no; 1: yes)\n"
<< "arg8: time kernel (0: no, 1: yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
// clang-format on
}
......@@ -60,7 +67,7 @@ static void print_helper_msg()
int profile_grouped_conv_fwd(int argc, char* argv[])
{
// 8 for control, 1 for num_dim_spatial
if(argc < 9)
if(argc < 10)
{
print_helper_msg();
return 1;
......@@ -68,20 +75,21 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
const auto data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const auto layout = static_cast<ConvLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int num_dim_spatial = std::stoi(argv[8]);
const auto index_type = static_cast<IndexType>(std::stoi(argv[4]));
const bool do_verification = std::stoi(argv[5]);
const int init_method = std::stoi(argv[6]);
const bool do_log = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[8]);
const int num_dim_spatial = std::stoi(argv[9]);
// 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
// 9 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
if(argc != 9 + 1 + 4 + 6 * num_dim_spatial)
{
print_helper_msg();
return 1;
}
const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 10, argv);
using F32 = float;
using F16 = ck::half_t;
......@@ -138,18 +146,43 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
using AComputeType = decltype(a_compute_type);
using BComputeType = decltype(b_compute_type);
bool pass = ck::profiler::profile_grouped_conv_fwd_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
AComputeType,
BComputeType>(
do_verification, init_method, do_log, time_kernel, params);
if(index_type == IndexType::INDEX_T)
{
bool pass = ck::profiler::profile_grouped_conv_fwd_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
AComputeType,
BComputeType,
ck::index_t>(
do_verification, init_method, do_log, time_kernel, params);
return pass ? 0 : 1;
}
else if(index_type == IndexType::LONG_INDEX_T)
{
bool pass = ck::profiler::profile_grouped_conv_fwd_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
AComputeType,
BComputeType,
ck::long_index_t>(
do_verification, init_method, do_log, time_kernel, params);
return pass ? 0 : 1;
return pass ? 0 : 1;
}
else
{
std::cout << "this indexing data type is not implemented" << std::endl;
return 1;
}
};
// GNHWC_GKYXC_GNHWK
......
profiler/src/profile_grouped_gemm_fixed_nk.cpp
View file @ 0eb75e21
......@@ -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;
......
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