Skip to content

GitLab

"git@developer.sourcefind.cn:cnjsdfcy/simbricks.git" did not exist on "8b8280b78cd9fa1ef72557930e20519b7cf8bd10"
Commit 3afb2f74 authored by root's avatar root
Browse files

add multi_d deviceop

parent 489599ba
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 854 additions and 147 deletions
example/01_gemm/gemm_xdl_fp16_v3.cpp
View file @ 3afb2f74
...@@ -25,7 +25,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa ...@@ -25,7 +25,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
using DeviceGemmV2Instance = using DeviceGemmV2Instance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3< ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
ALayout, BLayout, CLayout, ALayout, BLayout, CLayout,
ADataType, BDataType, ck::Tuple<>, CDataType, AccDataType, CShuffleDataType, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,
PassThrough, PassThrough, PassThrough, GemmDefault, PassThrough, PassThrough, PassThrough, GemmDefault,
256, 256,
224, 256, 224, 256,
......
example/01_gemm/run_gemm_example_v2.inc
View file @ 3afb2f74
...@@ -133,12 +133,10 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -133,12 +133,10 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
#ifdef BUILD_INT4_EXAMPLE #ifdef BUILD_INT4_EXAMPLE
static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()), static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()), static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
{},
static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()), static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#else #else
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()), static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()), static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
{},
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()), static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#endif #endif
M, M,
...@@ -146,7 +144,6 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config) ...@@ -146,7 +144,6 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
K, K,
StrideA, StrideA,
StrideB, StrideB,
{},
StrideC, StrideC,
KBatch, KBatch,
a_element_op, a_element_op,
......
include/ck/tensor_operation/gpu/device/device_gemm_v2.hpp
View file @ 3afb2f74
...@@ -14,26 +14,21 @@ template <typename ALayout, ...@@ -14,26 +14,21 @@ template <typename ALayout,
typename CLayout, typename CLayout,
typename ADataType, typename ADataType,
typename BDataType, typename BDataType,
typename DsDataType,
typename CDataType, typename CDataType,
typename AElementwiseOperation, typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation> typename CElementwiseOperation>
struct DeviceGemmV2 : public BaseOperator struct DeviceGemmV2 : public BaseOperator
{ {
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument> virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a, MakeArgumentPointer(const void* p_a,
const void* p_b, const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_c, void* p_c,
ck::index_t M, ck::index_t M,
ck::index_t N, ck::index_t N,
ck::index_t K, ck::index_t K,
ck::index_t StrideA, ck::index_t StrideA,
ck::index_t StrideB, ck::index_t StrideB,
std::array<ck::index_t, NumDTensor> StrideDs,
ck::index_t StrideC, ck::index_t StrideC,
ck::index_t KSplit, ck::index_t KSplit,
AElementwiseOperation a_element_op, AElementwiseOperation a_element_op,
......
profiler/src/CMakeLists.txt
View file @ 3afb2f74
# 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_pool3d_fwd.cpp #profile_max_pool3d_fwd.cpp
profile_avg_pool3d_bwd.cpp #profile_avg_pool3d_bwd.cpp
profile_max_pool3d_bwd.cpp #profile_max_pool3d_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
profile_gemm_universal.cpp
) )
if(GPU_TARGETS MATCHES "gfx9") #if(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)
endif() # endif()
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
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_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) #
#endif()
endif() #
#if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx9")
if(GPU_TARGETS MATCHES "gfx11" OR 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)
...@@ -77,76 +77,76 @@ add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES}) ...@@ -77,76 +77,76 @@ add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors) target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
target_link_libraries(${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_pool3d_fwd_instance) #target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
target_link_libraries(${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(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
#
if(GPU_TARGETS MATCHES "gfx9") #if(GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${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(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
endif() # endif()
target_link_libraries(${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(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
target_link_libraries(${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_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(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
endif() #endif()
#
if(GPU_TARGETS MATCHES "gfx9" OR GPU_TARGETS MATCHES "gfx11") #if(GPU_TARGETS MATCHES "gfx9" OR GPU_TARGETS MATCHES "gfx11")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${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