Skip to content

GitLab

Commit 83074f4c authored by Astha Rai's avatar Astha Rai
Browse files

Merge branch 'develop' into codegen_hiprtc

parents 7c56cd01 5fb150db
Hide whitespace changes
Inline Side-by-side
Showing with 213 additions and 20 deletions
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd_dynamic_op/xdl/device_grouped_conv2d_fwd_xdl_dynamic_op_nhwgc_gkyxc_nhwgk_int8_instance.cpp
View file @ 83074f4c
......@@ -30,6 +30,8 @@ void add_device_grouped_conv2d_fwd_xdl_dynamic_op_nhwgc_gkyxc_nhwgk_int8_instanc
Tuple<>,
NHWGK,
ConvFwdDefault>{});
#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
// long compilation time)
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_dynamic_op_int8_instances<2,
......@@ -46,6 +48,7 @@ void add_device_grouped_conv2d_fwd_xdl_dynamic_op_nhwgc_gkyxc_nhwgk_int8_instanc
Tuple<>,
NHWGK,
ConvFwd1x1S1P0>{});
#endif
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
View file @ 83074f4c
......@@ -15,6 +15,10 @@ set(GROUPED_CONV3D_BWD_WEIGHT
xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev2_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev5_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instance.cpp
xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp
)
if(DL_KERNELS)
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instance.cpp 0 → 100644
View file @ 83074f4c
// SPDX-License-Identifier: MIT
// 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_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_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_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_instance.cpp
View file @ 83074f4c
......@@ -25,7 +25,7 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf1
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances<
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instances<
3,
NDHWGC,
GKZYXC,
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_instance.cpp
View file @ 83074f4c
......@@ -25,7 +25,7 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf1
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances<
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instances<
3,
NDHWGC,
GKZYXC,
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instance.cpp 0 → 100644
View file @ 83074f4c
// SPDX-License-Identifier: MIT
// 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_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_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_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instance.cpp
View file @ 83074f4c
......@@ -25,7 +25,7 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances<
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances<
3,
NDHWGC,
GKZYXC,
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_instance.cpp
View file @ 83074f4c
......@@ -25,7 +25,7 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances<
device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances<
3,
NDHWGC,
GKZYXC,
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp 0 → 100644
View file @ 83074f4c
// SPDX-License-Identifier: MIT
// 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_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_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_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NGCDHW,
GKZYXC,
NGKDHW,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_instances<
3,
NGCDHW,
GKZYXC,
NGKDHW,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp 0 → 100644
View file @ 83074f4c
// SPDX-License-Identifier: MIT
// 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_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_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_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NGCDHW,
GKZYXC,
NGKDHW,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_instances<
3,
NGCDHW,
GKZYXC,
NGKDHW,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
View file @ 83074f4c
......@@ -31,6 +31,8 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_bf16_inst
Tuple<>,
NDHWGK,
ConvFwdDefault>{});
#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
// long compilation time)
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_dynamic_op_bf16_instances<3,
......@@ -47,6 +49,7 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_bf16_inst
Tuple<>,
NDHWGK,
ConvFwd1x1S1P0>{});
#endif
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
View file @ 83074f4c
......@@ -31,6 +31,8 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f16_insta
Tuple<>,
NDHWGK,
ConvFwdDefault>{});
#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
// long compilation time)
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_dynamic_op_f16_instances<3,
......@@ -47,6 +49,7 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f16_insta
Tuple<>,
NDHWGK,
ConvFwd1x1S1P0>{});
#endif
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
View file @ 83074f4c
......@@ -31,6 +31,8 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f32_insta
Tuple<>,
NDHWGK,
ConvFwdDefault>{});
#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
// long compilation time)
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_dynamic_op_f32_instances<3,
......@@ -47,6 +49,7 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f32_insta
Tuple<>,
NDHWGK,
ConvFwd1x1S1P0>{});
#endif
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
View file @ 83074f4c
......@@ -30,6 +30,8 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_int8_inst
Tuple<>,
NDHWGK,
ConvFwdDefault>{});
#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
// long compilation time)
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_dynamic_op_int8_instances<3,
......@@ -46,6 +48,7 @@ void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_int8_inst
Tuple<>,
NDHWGK,
ConvFwd1x1S1P0>{});
#endif
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/mha/CMakeLists.txt
View file @ 83074f4c
......@@ -27,11 +27,6 @@ rocm_install(FILES ${MHA_HEADERS} DESTINATION include/ck_tile/ops)
# headers for building lib
file(COPY ${MHA_HEADERS} DESTINATION ${FMHA_CPP_FOLDER})
# Delete the blob file if it exists to avoid append of old content.
if(EXISTS ${FMHA_CPP_FOLDER}/blob_list.txt)
file(REMOVE ${FMHA_CPP_FOLDER}/blob_list.txt)
endif()
set(FMHA_KNOWN_APIS "fwd,fwd_splitkv,fwd_appendkv,bwd")
# generate a list of kernels, but not actually emit files at config stage
......
profiler/include/profiler/profile_pool3d_fwd_impl.hpp
View file @ 83074f4c
......@@ -240,6 +240,19 @@ bool profile_pool3d_fwd_impl(PoolFwdInputParams& in_params, PoolFwdKernelParams&
{
out_device_buf.FromDevice(out_n_c_do_ho_wo_device.mData.data());
auto number_of_accumulations = 1;
static_assert(
ReduceOpId == ck::ReduceTensorOp::AVG || ReduceOpId == ck::ReduceTensorOp::MAX,
"Warning: Unhandled ReduceOpId for setting up the number of accumulations!");
if constexpr(ReduceOpId == ck::ReduceTensorOp::AVG)
{
for(size_t i = 0; i < kernel_params.window_spatial_lengths.size(); ++i)
{
number_of_accumulations *= kernel_params.window_spatial_lengths.at(i);
}
}
auto absolute_error_threshold = 1.0;
switch(in_params.init_method)
{
......@@ -250,9 +263,10 @@ bool profile_pool3d_fwd_impl(PoolFwdInputParams& in_params, PoolFwdKernelParams&
absolute_error_threshold =
ck::utils::get_absolute_threshold<ComputeDataType, OutDataType>(
absolute_error_threshold);
absolute_error_threshold, number_of_accumulations);
auto relative_error_threshold =
ck::utils::get_relative_threshold<ComputeDataType, OutDataType>();
ck::utils::get_relative_threshold<ComputeDataType, OutDataType>(
number_of_accumulations);
bool pass = ck::utils::check_err(out_n_c_do_ho_wo_device.mData,
out_n_c_do_ho_wo_host.mData,
......
profiler/src/profile_layernorm_fwd.cpp
View file @ 83074f4c
......@@ -85,7 +85,7 @@ int profile_layernorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F32, false, rank>(
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F16, false, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Float)
......
script/process_perf_data.py
View file @ 83074f4c
......@@ -133,12 +133,12 @@ def parse_logfile(logfile):
if 'Best Perf' in line:
lst=line.split()
res.append(lst[4])
elif 'onnx_gemm' in logfile or 'mixed_gemm' in logfile:
elif 'onnx_gemm' in logfile:
for line in open(logfile):
if 'Best Perf' in line:
lst=line.split()
res.append(lst[33])
elif 'splitK_gemm' in logfile:
elif 'splitK_gemm' in logfile or 'mixed_gemm' in logfile:
for line in open(logfile):
if 'Best Perf' in line:
lst=line.split()
......
script/process_qa_data.sh
View file @ 83074f4c
......@@ -22,6 +22,7 @@ python3 process_perf_data.py perf_gemm_bilinear.log
python3 process_perf_data.py perf_reduction.log
python3 process_perf_data.py perf_splitK_gemm.log
python3 process_perf_data.py perf_onnx_gemm.log
python3 process_perf_data.py perf_mixed_gemm.log
file=./perf_fmha_fwd_gfx942.log
if [ -e "$file" ]; then
......
test/CMakeLists.txt
View file @ 83074f4c
......@@ -64,11 +64,11 @@ function(add_test_executable TEST_NAME)
#only continue if there are some source files left on the list
if(ARGN)
if(ARGN MATCHES "_xdl")
list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(ARGN MATCHES "_wmma")
list(REMOVE_ITEM TEST_TARGETS gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
elseif(ARGN MATCHES "_smfmac")
list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201)
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
set_source_files_properties(${ARGN} PROPERTIES LANGUAGE HIP)
add_executable(${TEST_NAME} ${ARGN})
......@@ -141,11 +141,11 @@ function(add_gtest_executable TEST_NAME)
#only continue if there are some source files left on the list
if(ARGN)
if(ARGN MATCHES "_xdl")
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(ARGN MATCHES "_wmma")
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
elseif(ARGN MATCHES "_smfmac")
list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201)
list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
set_source_files_properties(${ARGN} PROPERTIES LANGUAGE HIP)
add_executable(${TEST_NAME} ${ARGN})
......
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