Skip to content

GitLab

Commit b3e5048f authored by aska-0096's avatar aska-0096
Browse files

tempsave

parent e8c19535
Hide whitespace changes
Inline Side-by-side
Showing with 190 additions and 158 deletions
example/65_gemm_multiply_multiply/CMakeLists.txt
View file @ b3e5048f
add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp)
add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp) add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp) add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp)
\ No newline at end of file target_compile_options(example_gemm_multiply_multiply_xdl_int8 PRIVATE -Wno-gnu-line-marker -save-temps=$PWD)
\ No newline at end of file
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_int8.cpp
View file @ b3e5048f
...@@ -37,8 +37,8 @@ using A0DataType = I8; ...@@ -37,8 +37,8 @@ using A0DataType = I8;
using B0DataType = I8; using B0DataType = I8;
using AccDataType = I32; using AccDataType = I32;
using CShuffleDataType = I32; using CShuffleDataType = I32;
using D0DataType = F32; using D0DataType = F16;
using D1DataType = F32; using D1DataType = F16;
using DsDataType = ck::Tuple<D0DataType, D1DataType>; using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = F16; using EDataType = F16;
...@@ -74,6 +74,16 @@ struct MultiplyMultiply ...@@ -74,6 +74,16 @@ struct MultiplyMultiply
e = ck::type_convert<ck::half_t>(x0_f); e = ck::type_convert<ck::half_t>(x0_f);
} }
template <>
__host__ __device__ constexpr void operator()<ck::half_t, int, ck::half_t, ck::half_t>(
ck::half_t& e, const int& c, const ck::half_t& d0, const ck::half_t& d1) const
{
const ck::half_t x0_f =
ck::type_convert<ck::half_t>(c) * d0* d1;
e = x0_f;
}
template <> template <>
__host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>( __host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>(
ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const
...@@ -91,7 +101,7 @@ using AElementOp = PassThrough; ...@@ -91,7 +101,7 @@ using AElementOp = PassThrough;
using BElementOp = PassThrough; using BElementOp = PassThrough;
using CDEElementOp = MultiplyMultiply; using CDEElementOp = MultiplyMultiply;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding; static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3 using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
// clang-format off // clang-format off
...@@ -102,7 +112,17 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu ...@@ -102,7 +112,17 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu
///###### RRR ///###### RRR
///< Row, Row, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, I8>; ///< Row, Row, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, I8>;
///###### RCR ///###### RCR
< Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 256, 128, 64, 16, 16, 32, 32, 4, 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>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, I8>; < Row, Col, DsLayout, ELayout,
A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, 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>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, I8>;
// clang-format on // clang-format on
int main(int argc, char* argv[]) int main(int argc, char* argv[])
...@@ -251,7 +271,7 @@ int main(int argc, char* argv[]) ...@@ -251,7 +271,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem"); "not support this GEMM problem");
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 20, 50, true, 50});
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 = std::size_t num_btype =
......
include/ck/host_utility/flush_cache.hpp
View file @ b3e5048f
...@@ -93,8 +93,11 @@ struct RotatingMemWrapperMultiD ...@@ -93,8 +93,11 @@ struct RotatingMemWrapperMultiD
} }
void Print() void Print()
{ {
std::cout << "RotatingMemWrapperMultiD: { size_a: " << size_a << ", size_b: " << size_b std::cout << "RotatingMemWrapperMultiD: { size_a: " << size_a << ", size_b: " << size_b;
<< ", rotating_count: " << rotating_count << "}" << std::endl; static_for<0, NumDs, 1>{}([&](auto j) {
std::cout << ", size_d" <<j.value<<": "<< size_ds[j];
});
std::cout << ", rotating_count: " << rotating_count << "}" << std::endl;
} }
~RotatingMemWrapperMultiD() ~RotatingMemWrapperMultiD()
{ {
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ b3e5048f
...@@ -66,6 +66,12 @@ struct PassThrough ...@@ -66,6 +66,12 @@ struct PassThrough
y = type_convert<half_t>(x); y = type_convert<half_t>(x);
} }
template <>
__host__ __device__ void operator()<half_t, int32_t>(half_t& y, const int32_t& x) const
{
y = type_convert<half_t>(x);
}
template <> template <>
__host__ __device__ void operator()<bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x) const __host__ __device__ void operator()<bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x) const
{ {
......
profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
View file @ b3e5048f
...@@ -188,7 +188,9 @@ bool profile_gemm_multiply_multiply_impl(int do_verification, ...@@ -188,7 +188,9 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
// profile device GEMM instances // profile device GEMM instances
for(auto& op_ptr : op_ptrs) for(auto& op_ptr : op_ptrs)
{ {
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38}; // Seems like when performance measurement has bug when spiltK is large
// std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
std::vector<int> kbatch_list = {1, 2, 4};
if(KBatch > 0) if(KBatch > 0)
{ {
......
profiler/src/CMakeLists.txt
View file @ b3e5048f
# 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)
...@@ -94,87 +94,87 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) ...@@ -94,87 +94,87 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
endif() endif()
target_link_libraries(${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)
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