Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
gaoqiong
composable_kernel_ROCM
Commits
09852d3b
"example/01_gemm/gemm_wmma_fp16.cpp" did not exist on "304802889728707c2a162322ce18686169e732ea"
Commit
09852d3b
authored
Oct 21, 2024
by
chenjun
Browse files
uncomment
parent
7fb0b322
Changes
2
Show whitespace changes
Inline
Side-by-side
Showing
2 changed files
with
151 additions
and
151 deletions
+151
-151
profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
.../include/profiler/profile_gemm_multiply_multiply_impl.hpp
+7
-7
profiler/src/CMakeLists.txt
profiler/src/CMakeLists.txt
+144
-144
No files found.
profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
View file @
09852d3b
...
@@ -84,12 +84,12 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
...
@@ -84,12 +84,12 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
std
::
min
(
n_iter
,
std
::
min
(
n_iter
,
static_cast
<
int
>
(
std
::
ceil
(
static_cast
<
double
>
(
rotating
)
/
total_gemm_needed
))));
static_cast
<
int
>
(
std
::
ceil
(
static_cast
<
double
>
(
rotating
)
/
total_gemm_needed
))));
//
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std
::
cout
<<
"a_m_k: "
<<
a_m_k
.
mDesc
<<
std
::
endl
;
//
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std
::
cout
<<
"b_k_n: "
<<
b_k_n
.
mDesc
<<
std
::
endl
;
//
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std
::
cout
<<
"d0_m_n: "
<<
d0_m_n
.
mDesc
<<
std
::
endl
;
//
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
std
::
cout
<<
"d1_m_n: "
<<
d1_m_n
.
mDesc
<<
std
::
endl
;
//
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
std
::
cout
<<
"e_m_n: "
<<
e_m_n_device_result
.
mDesc
<<
std
::
endl
;
//
std::cout << "rotating count: " << rotating_count << std::endl;
std
::
cout
<<
"rotating count: "
<<
rotating_count
<<
std
::
endl
;
switch
(
init_method
)
switch
(
init_method
)
{
{
...
@@ -146,7 +146,7 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
...
@@ -146,7 +146,7 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
const
auto
op_ptrs
=
ck
::
tensor_operation
::
device
::
instance
::
DeviceOperationInstanceFactory
<
const
auto
op_ptrs
=
ck
::
tensor_operation
::
device
::
instance
::
DeviceOperationInstanceFactory
<
DeviceOp
>::
GetInstances
();
DeviceOp
>::
GetInstances
();
//
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std
::
cout
<<
"found "
<<
op_ptrs
.
size
()
<<
" instances"
<<
std
::
endl
;
// Run reference GEMM
// Run reference GEMM
if
(
do_verification
)
if
(
do_verification
)
...
...
profiler/src/CMakeLists.txt
View file @
09852d3b
# ckProfiler
# ckProfiler
set
(
PROFILER_SOURCES
set
(
PROFILER_SOURCES
profiler.cpp
profiler.cpp
#
profile_gemm.cpp
profile_gemm.cpp
#
profile_reduce.cpp
profile_reduce.cpp
#
profile_groupnorm_bwd_data.cpp
profile_groupnorm_bwd_data.cpp
#
profile_groupnorm_fwd.cpp
profile_groupnorm_fwd.cpp
#
profile_layernorm_bwd_data.cpp
profile_layernorm_bwd_data.cpp
#
profile_layernorm_bwd_gamma_beta.cpp
profile_layernorm_bwd_gamma_beta.cpp
#
profile_groupnorm_bwd_gamma_beta.cpp
profile_groupnorm_bwd_gamma_beta.cpp
#
profile_layernorm_fwd.cpp
profile_layernorm_fwd.cpp
#
profile_max_pool2d_fwd.cpp
profile_max_pool2d_fwd.cpp
#
profile_pool3d_fwd.cpp
profile_pool3d_fwd.cpp
#
profile_avg_pool3d_bwd.cpp
profile_avg_pool3d_bwd.cpp
#
profile_max_pool3d_bwd.cpp
profile_max_pool3d_bwd.cpp
#
profile_avg_pool2d_bwd.cpp
profile_avg_pool2d_bwd.cpp
#
profile_max_pool2d_bwd.cpp
profile_max_pool2d_bwd.cpp
#
profile_softmax.cpp
profile_softmax.cpp
#
profile_batchnorm_fwd.cpp
profile_batchnorm_fwd.cpp
#
profile_batchnorm_bwd.cpp
profile_batchnorm_bwd.cpp
#
profile_batchnorm_infer.cpp
profile_batchnorm_infer.cpp
#
profile_conv_tensor_rearrange.cpp
profile_conv_tensor_rearrange.cpp
#
profile_transpose.cpp
profile_transpose.cpp
#
profile_permute_scale.cpp
profile_permute_scale.cpp
)
)
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx9"
)
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx9"
)
#
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
if
(
DTYPES MATCHES
"fp32"
OR DTYPES MATCHES
"fp64"
OR NOT DEFINED DTYPES
)
#
list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
list
(
APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp
)
#
list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
list
(
APPEND PROFILER_SOURCES profile_contraction_scale.cpp
)
#
endif()
endif
()
#
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
#
list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_reduce.cpp
)
#
list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
list
(
APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp
)
#
list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
list
(
APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_add.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_gemm.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_streamk.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp
)
#
endif()
endif
()
#
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp
)
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx94"
)
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx94"
)
list
(
APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp
)
list
(
APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp
)
endif
()
endif
()
#
list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
list
(
APPEND PROFILER_SOURCES profile_batched_gemm.cpp
)
#
list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
list
(
APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_splitk.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_universal.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp
)
#
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp
)
#
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
list
(
APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp
)
#
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
list
(
APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp
)
#
list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp)
list
(
APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp
)
#
list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp)
list
(
APPEND PROFILER_SOURCES profile_conv_fwd.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp
)
endif
()
endif
()
#
if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx11"
OR SUPPORTED_GPU_TARGETS MATCHES
"gfx12"
OR SUPPORTED_GPU_TARGETS MATCHES
"gfx9"
)
#
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
#
list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
list
(
APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp
)
#
endif()
endif
()
#
list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp
)
#
endif()
endif
()
#
if(DL_KERNELS)
if
(
DL_KERNELS
)
#
list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
list
(
APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp
)
#
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
list
(
APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp
)
#
endif()
endif
()
set
(
PROFILER_EXECUTABLE ckProfiler
)
set
(
PROFILER_EXECUTABLE ckProfiler
)
...
@@ -95,86 +95,86 @@ endif()
...
@@ -95,86 +95,86 @@ endif()
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE utility getopt::getopt
)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE utility getopt::getopt
)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_instance
)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_normalization_fwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_normalization_bwd_data_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_normalization_bwd_gamma_beta_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_softmax_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_reduce_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_batchnorm_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_pool2d_fwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_pool3d_fwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_avg_pool2d_bwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_avg_pool3d_bwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_max_pool_bwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_image_to_column_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_column_to_image_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_transpose_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_permute_scale_instance
)
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx9"
)
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx9"
)
#
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
if
(
DTYPES MATCHES
"fp32"
OR DTYPES MATCHES
"fp64"
OR NOT DEFINED DTYPES
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_contraction_bilinear_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_contraction_scale_instance
)
#
endif()
endif
()
#
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_add_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_add_add_fastgelu_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_fastgelu_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_batched_gemm_gemm_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_batched_gemm_add_relu_gemm_add_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_gemm_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_streamk_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_add_fastgelu_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_add_relu_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_add_silu_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_add_relu_add_layernorm_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_gemm_fixed_nk_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_gemm_fastgelu_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_gemm_tile_loop_instance
)
#
endif()
endif
()
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_batched_gemm_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_batched_gemm_reduce_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_multiply_add_instance
)
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx94"
)
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx94"
)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_multiply_multiply_instance
)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_multiply_multiply_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_ab_scale_instance
)
endif
()
endif
()
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_splitk_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_universal_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_universal_reduce_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_universal_streamk_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_add_multiply_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_reduce_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_bias_add_reduce_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_conv2d_fwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_conv2d_fwd_bias_relu_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_conv2d_fwd_bias_relu_add_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv1d_fwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_conv1d_bwd_data_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_conv3d_bwd_data_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_conv2d_bwd_data_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv1d_bwd_weight_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv2d_bwd_weight_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv3d_fwd_convscale_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv3d_fwd_convinvscale_instance
)
endif
()
endif
()
#
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
if
(
SUPPORTED_GPU_TARGETS MATCHES
"gfx9"
OR SUPPORTED_GPU_TARGETS MATCHES
"gfx11"
OR SUPPORTED_GPU_TARGETS MATCHES
"gfx12"
)
#
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_gemm_bilinear_instance
)
#
endif()
endif
()
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv3d_fwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv2d_bwd_data_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv3d_bwd_data_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv2d_fwd_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv3d_bwd_weight_instance
)
#
endif()
endif
()
#
if(DL_KERNELS)
if
(
DL_KERNELS
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_batched_gemm_multi_d_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv1d_bwd_weight_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv2d_bwd_weight_instance
)
#
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
target_link_libraries
(
${
PROFILER_EXECUTABLE
}
PRIVATE device_grouped_conv3d_bwd_weight_instance
)
#
endif()
endif
()
rocm_install
(
TARGETS
${
PROFILER_EXECUTABLE
}
COMPONENT profiler
)
rocm_install
(
TARGETS
${
PROFILER_EXECUTABLE
}
COMPONENT profiler
)
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment
