Skip to content

GitLab

Commit 39002e9e authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'develop' into amd-develop

parents b26bdd61 d52ec016
Hide whitespace changes
Inline Side-by-side
Showing with 271 additions and 560 deletions
library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
View file @ 39002e9e
...@@ -5,6 +5,7 @@ ...@@ -5,6 +5,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
#include "ck/utility/data_type.hpp" #include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
...@@ -43,6 +44,32 @@ using device_normalization_f16_instances = ...@@ -43,6 +44,32 @@ using device_normalization_f16_instances =
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f16_instances =
// clang-format off
std::tuple <
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>, // irregular size
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_generic_instance = std::tuple< using device_normalization_f16_generic_instance = std::tuple<
// clang-format off // clang-format off
...@@ -76,6 +103,32 @@ using device_normalization_f32_instances = std::tuple< ...@@ -76,6 +103,32 @@ using device_normalization_f32_instances = std::tuple<
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f32_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f32_generic_instance = std::tuple< using device_normalization_f32_generic_instance = std::tuple<
// clang-format off // clang-format off
...@@ -109,6 +162,32 @@ using device_normalization_f16_f32_f32_f16_instances = std::tuple< ...@@ -109,6 +162,32 @@ using device_normalization_f16_f32_f32_f16_instances = std::tuple<
// clang-format on // clang-format on
>; >;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_splitk_f16_f32_f32_f16_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>, // irregular size
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce> template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple< using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple<
// clang-format off // clang-format off
......
library/src/tensor_operation_instance/gpu/pool_fwd/CMakeLists.txt → library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
View file @ 39002e9e
set(DEVICE_POOL_FWD_INSTANCES) set(DEVICE_POOL3D_FWD_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_POOL_FWD_INSTANCES device_avg_pool2d_fwd_nhwc_f16_instance.cpp list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
device_max_pool2d_fwd_nhwc_f16_instance.cpp
device_max_pool3d_fwd_ndhwc_f16_instance.cpp) device_max_pool3d_fwd_ndhwc_f16_instance.cpp)
endif() endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES) if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_POOL_FWD_INSTANCES device_avg_pool2d_fwd_nhwc_f32_instance.cpp list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
device_max_pool2d_fwd_nhwc_f32_instance.cpp
device_max_pool3d_fwd_ndhwc_f32_instance.cpp) device_max_pool3d_fwd_ndhwc_f32_instance.cpp)
endif() endif()
add_instance_library(device_pool_fwd_instance ${DEVICE_POOL_FWD_INSTANCES}) add_instance_library(device_pool3d_fwd_instance ${DEVICE_POOL3D_FWD_INSTANCES})
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
View file @ 39002e9e
...@@ -11,7 +11,9 @@ namespace instance { ...@@ -11,7 +11,9 @@ namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG; static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool3d_fwd_ndhwc_f16_instances( void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, false>>>& instances) std::vector<
std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{}); instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{});
......
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
View file @ 39002e9e
...@@ -11,7 +11,9 @@ namespace instance { ...@@ -11,7 +11,9 @@ namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG; static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool3d_fwd_ndhwc_f32_instances( void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, false>>>& instances) std::vector<
std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{}); instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
......
library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp
View file @ 39002e9e
...@@ -11,14 +11,18 @@ namespace instance { ...@@ -11,14 +11,18 @@ namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX; static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
void add_device_pool3d_fwd_ndhwc_f16_instances( void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, false>>>& instances) std::vector<
std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, false>{}); instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, false>{});
} }
void add_device_pool3d_fwd_ndhwc_index_f16_instances( void add_device_pool3d_fwd_ndhwc_index_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, true>>>& instances) std::vector<
std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, true>{}); instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, true>{});
......
library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp
View file @ 39002e9e
...@@ -11,14 +11,18 @@ namespace instance { ...@@ -11,14 +11,18 @@ namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX; static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
void add_device_pool3d_fwd_ndhwc_f32_instances( void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, false>>>& instances) std::vector<
std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{}); instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
} }
void add_device_pool3d_fwd_ndhwc_index_f32_instances( void add_device_pool3d_fwd_ndhwc_index_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, true>>>& instances) std::vector<
std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
instances)
{ {
add_device_operation_instances( add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{}); instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});
......
library/src/tensor_operation_instance/gpu/pool_fwd/pool_fwd_instance_common.hpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp
View file @ 39002e9e
...@@ -15,24 +15,10 @@ namespace tensor_operation { ...@@ -15,24 +15,10 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
using I32 = int32_t; using I32 = int32_t;
using F16 = ck::half_t; using F16 = ck::half_t;
using F32 = float; using F32 = float;
using NDHWC = ck::tensor_layout::convolution::NDHWC;
template <typename InDataType,
typename OutDataType,
typename IndexDataType,
typename ComputeDataType,
ReduceTensorOp ReduceOpId,
bool OutputIndex>
using device_pool2d_fwd_nhwc_instances =
// clang-format off
std::tuple <
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
// clang-format on
>;
template <typename InDataType, template <typename InDataType,
typename OutDataType, typename OutDataType,
...@@ -43,9 +29,9 @@ template <typename InDataType, ...@@ -43,9 +29,9 @@ template <typename InDataType,
using device_pool3d_fwd_ndhwc_instances = using device_pool3d_fwd_ndhwc_instances =
// clang-format off // clang-format off
std::tuple < std::tuple <
DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>, DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>, DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4> DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
// clang-format on // clang-format on
>; >;
......
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool2d_fwd_nhwc_f32_instance.cpp deleted 100644 → 0
View file @ b26bdd61
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool2d_fwd_nhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool2d_fwd_nhwc_f16_instance.cpp deleted 100644 → 0
View file @ b26bdd61
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
void add_device_pool2d_fwd_nhwc_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F16, ReduceOpId, false>{});
}
void add_device_pool2d_fwd_nhwc_index_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, ReduceOpId, true>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F16, ReduceOpId, true>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool2d_fwd_nhwc_f32_instance.cpp deleted 100644 → 0
View file @ b26bdd61
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
void add_device_pool2d_fwd_nhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
}
void add_device_pool2d_fwd_nhwc_index_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, true>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/include/profiler/profile_groupnorm_impl.hpp
View file @ 39002e9e
...@@ -139,6 +139,10 @@ bool profile_groupnorm_impl(int do_verification, ...@@ -139,6 +139,10 @@ bool profile_groupnorm_impl(int do_verification,
continue; continue;
} }
size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = inst_ptr->MakeInvokerPointer(); auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......
profiler/include/profiler/profile_layernorm_impl.hpp
View file @ 39002e9e
...@@ -155,6 +155,10 @@ bool profile_layernorm_impl(int do_verification, ...@@ -155,6 +155,10 @@ bool profile_layernorm_impl(int do_verification,
continue; continue;
} }
size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = inst_ptr->MakeInvokerPointer(); auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......
profiler/include/profiler/profile_pool2d_fwd_impl.hpp deleted 100644 → 0
View file @ b26bdd61
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/pool2d_fwd.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
namespace ck {
namespace profiler {
template <typename InDataType,
typename OutDataType,
typename ComputeDataType,
typename IndexDataType,
ck::ReduceTensorOp ReduceOpId,
bool PropagateNan,
bool OutputIndex>
bool profile_pool2d_fwd_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> in_length, // NCHW
std::vector<index_t> window_spatial_lengths,
std::vector<index_t> window_strides,
std::vector<index_t> input_left_pads,
std::vector<index_t> input_right_pads)
{
constexpr index_t InOutRank = 4;
constexpr index_t WindowRank = 2;
if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
window_strides.size() != WindowRank || input_left_pads.size() != WindowRank ||
input_right_pads.size() != WindowRank)
return false;
std::vector<index_t> out_length(InOutRank);
int N = in_length[0];
int C = in_length[1];
out_length[0] = N;
out_length[1] = C;
// Calculate Ho, Wo
for(int i = 2; i < InOutRank; ++i)
{
auto pad1 = input_left_pads[i - 2];
auto pad2 = input_right_pads[i - 2];
auto windows_size = window_spatial_lengths[i - 2];
auto windows_stride = window_strides[i - 2];
out_length[i] = (in_length[i] + pad1 + pad2 - windows_size) / windows_stride + 1;
}
int Hi = in_length[2];
int Wi = in_length[3];
int Ho = out_length[2];
int Wo = out_length[3];
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
using namespace ck::literals;
return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
Tensor<IndexDataType> out_indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
Tensor<IndexDataType> out_indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
switch(init_method)
{
case 0: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); break;
case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) *
out_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
out_indices_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
// add device normalization instances
using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
if(do_verification)
{
using ReferenceInstance = ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
ComputeDataType,
IndexDataType,
ReduceOpId,
PropagateNan,
OutputIndex>;
ReferenceInstance ref;
auto ref_argument = ref.MakeArgument(in_n_c_hi_wi,
out_n_c_ho_wo_host,
out_indices_n_c_ho_wo_host,
window_spatial_lengths,
window_strides,
input_left_pads,
input_right_pads);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
in_length,
window_spatial_lengths,
out_length,
{C * Hi * Wi, 1, Wi * C, C},
{C * Ho * Wo, 1, Wo * C, C},
{C * Ho * Wo, 1, Wo * C, C},
window_strides,
input_left_pads,
input_right_pads,
{2, 3});
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
++num_kernel;
}
else
{
if(time_kernel)
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = ", in_length, ", ") << std::endl;
}
continue;
}
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes = in_n_c_hi_wi.mDesc.GetElementSize() * sizeof(InDataType) +
out_n_c_ho_wo_host.mDesc.GetElementSize() * sizeof(OutDataType);
if constexpr(OutputIndex)
num_bytes += out_indices_n_c_ho_wo_host.mDesc.GetElementSize() * sizeof(IndexDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
bool pass = ck::utils::check_err(out_n_c_ho_wo_device.mData,
out_n_c_ho_wo_host.mData,
"Error: Incorrect results",
1e-3,
1e-3);
if constexpr(OutputIndex)
{
out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
pass = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device,
out_indices_n_c_ho_wo_host);
}
if(do_log)
{
LogRangeAsType<float>(std::cout << "in_n_c_hi_wi : ", in_n_c_hi_wi.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_n_c_ho_wo_host : ", out_n_c_ho_wo_host.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_n_c_ho_wo_device : ", out_n_c_ho_wo_device.mData, ",")
<< std::endl;
if constexpr(OutputIndex)
LogRangeAsType<float>(std::cout << "out_indices_n_c_ho_wo_device : ",
out_indices_n_c_ho_wo_device.mData,
",")
<< std::endl;
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "lengths = [", in_length, ", ") << "]." << std::endl;
return false;
}
else
{
if(time_kernel)
std::cout << "pass" << std::endl;
}
}
}
if(time_kernel)
{
LogRange(std::cout << "length = ", in_length, ",") << std::endl;
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_instance_name << std::endl;
}
if(num_kernel == 0)
{
std::cout << "Error: No kernel is applicable" << std::endl;
return false;
}
return true;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_pool3d_fwd_impl.hpp
View file @ 39002e9e
...@@ -21,6 +21,8 @@ template <typename InDataType, ...@@ -21,6 +21,8 @@ template <typename InDataType,
typename OutDataType, typename OutDataType,
typename ComputeDataType, typename ComputeDataType,
typename IndexDataType, typename IndexDataType,
typename InLayout,
typename OutLayout,
ck::ReduceTensorOp ReduceOpId, ck::ReduceTensorOp ReduceOpId,
bool PropagateNan, bool PropagateNan,
bool OutputIndex> bool OutputIndex>
...@@ -31,6 +33,7 @@ bool profile_pool3d_fwd_impl(int do_verification, ...@@ -31,6 +33,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
std::vector<index_t> in_length, // NCDHW std::vector<index_t> in_length, // NCDHW
std::vector<index_t> window_spatial_lengths, std::vector<index_t> window_spatial_lengths,
std::vector<index_t> window_strides, std::vector<index_t> window_strides,
std::vector<index_t> window_dilations,
std::vector<index_t> input_left_pads, std::vector<index_t> input_left_pads,
std::vector<index_t> input_right_pads) std::vector<index_t> input_right_pads)
{ {
...@@ -38,8 +41,8 @@ bool profile_pool3d_fwd_impl(int do_verification, ...@@ -38,8 +41,8 @@ bool profile_pool3d_fwd_impl(int do_verification,
constexpr index_t WindowRank = 3; constexpr index_t WindowRank = 3;
if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank || if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
window_strides.size() != WindowRank || input_left_pads.size() != WindowRank || window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
input_right_pads.size() != WindowRank) input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
return false; return false;
std::vector<index_t> out_length(InOutRank); std::vector<index_t> out_length(InOutRank);
...@@ -53,11 +56,13 @@ bool profile_pool3d_fwd_impl(int do_verification, ...@@ -53,11 +56,13 @@ bool profile_pool3d_fwd_impl(int do_verification,
// Calculate Do, Ho, Wo // Calculate Do, Ho, Wo
for(int i = 2; i < InOutRank; ++i) for(int i = 2; i < InOutRank; ++i)
{ {
auto pad1 = input_left_pads[i - 2]; auto pad1 = input_left_pads[i - 2];
auto pad2 = input_right_pads[i - 2]; auto pad2 = input_right_pads[i - 2];
auto windows_size = window_spatial_lengths[i - 2]; auto windows_size = window_spatial_lengths[i - 2];
auto windows_stride = window_strides[i - 2]; auto windows_stride = window_strides[i - 2];
out_length[i] = (in_length[i] + pad1 + pad2 - windows_size) / windows_stride + 1; auto windows_dilation = window_dilations[i - 2];
auto eff = (windows_size - 1) * windows_dilation + 1;
out_length[i] = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
} }
int Di = in_length[2]; int Di = in_length[2];
...@@ -104,6 +109,8 @@ bool profile_pool3d_fwd_impl(int do_verification, ...@@ -104,6 +109,8 @@ bool profile_pool3d_fwd_impl(int do_verification,
InDataType, InDataType,
OutDataType, OutDataType,
IndexDataType, IndexDataType,
InLayout,
OutLayout,
ReduceOpId, ReduceOpId,
OutputIndex>; OutputIndex>;
...@@ -136,6 +143,7 @@ bool profile_pool3d_fwd_impl(int do_verification, ...@@ -136,6 +143,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
out_indices_n_c_do_ho_wo_host, out_indices_n_c_do_ho_wo_host,
window_spatial_lengths, window_spatial_lengths,
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads); input_right_pads);
auto ref_invoker = ref.MakeInvoker(); auto ref_invoker = ref.MakeInvoker();
...@@ -157,6 +165,7 @@ bool profile_pool3d_fwd_impl(int do_verification, ...@@ -157,6 +165,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
{Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C}, {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
{Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C}, {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
window_strides, window_strides,
window_dilations,
input_left_pads, input_left_pads,
input_right_pads, input_right_pads,
{2, 3, 4}); {2, 3, 4});
......
profiler/src/CMakeLists.txt
View file @ 39002e9e
...@@ -17,7 +17,6 @@ set(PROFILER_SOURCES ...@@ -17,7 +17,6 @@ set(PROFILER_SOURCES
profile_reduce.cpp profile_reduce.cpp
profile_groupnorm.cpp profile_groupnorm.cpp
profile_layernorm.cpp profile_layernorm.cpp
profile_avg_pool2d_fwd.cpp
profile_max_pool3d_fwd.cpp profile_max_pool3d_fwd.cpp
profile_softmax.cpp profile_softmax.cpp
profile_batchnorm_fwd.cpp profile_batchnorm_fwd.cpp
...@@ -74,7 +73,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) ...@@ -74,7 +73,7 @@ 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_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(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
target_link_libraries(${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)
if(DL_KERNELS) if(DL_KERNELS)
......
profiler/src/profile_avg_pool2d_fwd.cpp deleted 100644 → 0
View file @ b26bdd61
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_pool2d_fwd_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct avgPoolFwdArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {
{"length", {}}, {"wsize", {}}, {"wstride", {}}, {"pad1", {}}, {"pad2", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help_avg_pool2d_fwd()
{
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
<< "arg2: verification (0: no; 1: yes)\n"
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: input tensor length for NDHW(e.g, --length 2 32 30 30) \n"
<< "--wsize: window size for YX (e.g, --wsize 2 2) \n"
<< "--wstride: window stride for HW (e.g, --wstride 2 2) \n"
<< "--pad1: left side of padding in HW (e.g, --pad1 1 1) \n"
<< "--pad2: right side of padding in HW (e.g, --pad2 1 1) \n"
<< "eg: ckProfiler avg_pool2d_fwd 0 1 2 0 1 0 --length 2 32 30 30 --wsize 2 2 "
"--wstride 2 2 --pad1 1 1 --pad2 1 1"
<< std::endl;
}
int profile_avg_pool2d_fwd(int argc, char* argv[])
{
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
bool do_verification = true;
int init_method = 0;
bool do_log = false;
bool time_kernel = true;
std::vector<index_t> in_length = {2, 32, 30, 30};
std::vector<index_t> wsize = {2, 2};
std::vector<index_t> wstride = {2, 2};
std::vector<index_t> pad1 = {1, 1};
std::vector<index_t> pad2 = {1, 1};
if(argc != 2 && argc != 25)
{
print_help_avg_pool2d_fwd();
return 0;
}
else if(argc == 25)
{
data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
do_verification = std::stoi(argv[3]);
init_method = std::stoi(argv[4]);
do_log = std::stoi(argv[5]);
time_kernel = std::stoi(argv[6]);
// parse the long options
avgPoolFwdArgParser arg_parser;
arg_parser(argc, argv);
in_length = arg_parser.long_opts["length"];
wsize = arg_parser.long_opts["wsize"];
wstride = arg_parser.long_opts["wstride"];
pad1 = arg_parser.long_opts["pad1"];
pad2 = arg_parser.long_opts["pad2"];
}
using F16 = ck::half_t;
using F32 = float;
using I32 = int32_t;
constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_pool2d_fwd_impl<F16, F16, F32, I32, ReduceOpId, false, false>(
do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
pad1,
pad2);
}
else if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_pool2d_fwd_impl<F32, F32, F32, I32, ReduceOpId, false, false>(
do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
pad1,
pad2);
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
REGISTER_PROFILER_OPERATION("avg_pool2d_fwd", "avg_pool2d fwd", profile_avg_pool2d_fwd);
profiler/src/profile_max_pool3d_fwd.cpp
View file @ 39002e9e
...@@ -13,8 +13,12 @@ using ck::index_t; ...@@ -13,8 +13,12 @@ using ck::index_t;
struct maxPoolFwdArgParser struct maxPoolFwdArgParser
{ {
std::unordered_map<std::string, std::vector<int>> long_opts = { std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
{"length", {}}, {"wsize", {}}, {"wstride", {}}, {"pad1", {}}, {"pad2", {}}}; {"wsize", {}},
{"wstride", {}},
{"wdilation", {}},
{"pad1", {}},
{"pad2", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i) bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{ {
...@@ -56,10 +60,11 @@ void print_help_max_pool3d_fwd() ...@@ -56,10 +60,11 @@ void print_help_max_pool3d_fwd()
<< "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n" << "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
<< "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n" << "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
<< "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n" << "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
<< "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
<< "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n" << "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
<< "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n" << "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
<< "eg: ckProfiler max_pool3d_fwd 0 1 2 0 1 0 --length 2 32 30 30 30 --wsize 2 2 2 " << "eg: ckProfiler max_pool3d_fwd 0 1 2 0 1 0 --length 2 32 30 30 30 --wsize 2 2 2 "
"--wstride 2 2 2 --pad1 1 1 1 --pad2 1 1 1" "--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
<< std::endl; << std::endl;
} }
...@@ -75,15 +80,16 @@ int profile_max_pool3d_fwd(int argc, char* argv[]) ...@@ -75,15 +80,16 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
std::vector<index_t> in_length = {2, 32, 30, 30, 30}; std::vector<index_t> in_length = {2, 32, 30, 30, 30};
std::vector<index_t> wsize = {2, 2, 2}; std::vector<index_t> wsize = {2, 2, 2};
std::vector<index_t> wstride = {2, 2, 2}; std::vector<index_t> wstride = {2, 2, 2};
std::vector<index_t> wdilation = {1, 1, 1};
std::vector<index_t> pad1 = {1, 1, 1}; std::vector<index_t> pad1 = {1, 1, 1};
std::vector<index_t> pad2 = {1, 1, 1}; std::vector<index_t> pad2 = {1, 1, 1};
if(argc != 2 && argc != 30) if(argc != 2 && argc != 34)
{ {
print_help_max_pool3d_fwd(); print_help_max_pool3d_fwd();
return 0; return 0;
} }
else if(argc == 30) else if(argc == 34)
{ {
data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2])); data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
do_verification = std::stoi(argv[3]); do_verification = std::stoi(argv[3]);
...@@ -98,64 +104,79 @@ int profile_max_pool3d_fwd(int argc, char* argv[]) ...@@ -98,64 +104,79 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
in_length = arg_parser.long_opts["length"]; in_length = arg_parser.long_opts["length"];
wsize = arg_parser.long_opts["wsize"]; wsize = arg_parser.long_opts["wsize"];
wstride = arg_parser.long_opts["wstride"]; wstride = arg_parser.long_opts["wstride"];
wdilation = arg_parser.long_opts["wdilation"];
pad1 = arg_parser.long_opts["pad1"]; pad1 = arg_parser.long_opts["pad1"];
pad2 = arg_parser.long_opts["pad2"]; pad2 = arg_parser.long_opts["pad2"];
} }
using F16 = ck::half_t; using F16 = ck::half_t;
using F32 = float; using F32 = float;
using I32 = int32_t; using I32 = int32_t;
using NDHWC = ck::tensor_layout::convolution::NDHWC;
#if 1
constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX; constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
#else
constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
#endif
if(data_type == ck::DataTypeEnum::Half) if(data_type == ck::DataTypeEnum::Half)
{ {
if(return_index) if(return_index)
ck::profiler::profile_pool3d_fwd_impl<F16, F16, F16, I32, ReduceOpId, false, true>( ck::profiler::
do_verification, profile_pool3d_fwd_impl<F16, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false, true>(
init_method, do_verification,
do_log, init_method,
time_kernel, do_log,
in_length, time_kernel,
wsize, in_length,
wstride, wsize,
pad1, wstride,
pad2); wdilation,
pad1,
pad2);
else else
ck::profiler::profile_pool3d_fwd_impl<F16, F16, F16, I32, ReduceOpId, false, false>( ck::profiler::
do_verification, profile_pool3d_fwd_impl<F16, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false, false>(
init_method, do_verification,
do_log, init_method,
time_kernel, do_log,
in_length, time_kernel,
wsize, in_length,
wstride, wsize,
pad1, wstride,
pad2); wdilation,
pad1,
pad2);
} }
else if(data_type == ck::DataTypeEnum::Float) else if(data_type == ck::DataTypeEnum::Float)
{ {
if(return_index) if(return_index)
ck::profiler::profile_pool3d_fwd_impl<F32, F32, F32, I32, ReduceOpId, false, true>( ck::profiler::
do_verification, profile_pool3d_fwd_impl<F32, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false, true>(
init_method, do_verification,
do_log, init_method,
time_kernel, do_log,
in_length, time_kernel,
wsize, in_length,
wstride, wsize,
pad1, wstride,
pad2); wdilation,
pad1,
pad2);
else else
ck::profiler::profile_pool3d_fwd_impl<F32, F32, F32, I32, ReduceOpId, false, false>( ck::profiler::
do_verification, profile_pool3d_fwd_impl<F32, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false, false>(
init_method, do_verification,
do_log, init_method,
time_kernel, do_log,
in_length, time_kernel,
wsize, in_length,
wstride, wsize,
pad1, wstride,
pad2); wdilation,
pad1,
pad2);
} }
else else
{ {
......
test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
View file @ 39002e9e
...@@ -100,6 +100,9 @@ TYPED_TEST(TestGroupedConvndBwdWeight1d, Test1D) ...@@ -100,6 +100,9 @@ TYPED_TEST(TestGroupedConvndBwdWeight1d, Test1D)
this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}}); this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
this->conv_params.push_back({1, 2, 32, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}}); this->conv_params.push_back({1, 2, 32, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}}); this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
this->conv_params.push_back({1, 1, 1, 1, 32, {3}, {32}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 1, 1, 64, 3, {3}, {32}, {1}, {1}, {1}, {1}});
this->conv_params.push_back({1, 1, 1, 1, 1, {3}, {32}, {1}, {1}, {1}, {1}});
this->Run(); this->Run();
} }
...@@ -112,6 +115,9 @@ TYPED_TEST(TestGroupedConvndBwdWeight2d, Test2D) ...@@ -112,6 +115,9 @@ TYPED_TEST(TestGroupedConvndBwdWeight2d, Test2D)
{2, 2, 4, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); {2, 2, 4, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back( this->conv_params.push_back(
{2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}}); {2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run(); this->Run();
} }
...@@ -124,5 +130,11 @@ TYPED_TEST(TestGroupedConvndBwdWeight3d, Test3D) ...@@ -124,5 +130,11 @@ TYPED_TEST(TestGroupedConvndBwdWeight3d, Test3D)
{3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}); {3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back( this->conv_params.push_back(
{3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}}); {3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
this->conv_params.push_back(
{3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->Run(); this->Run();
} }
test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
View file @ 39002e9e
...@@ -22,6 +22,8 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv1dFwdGNWC) ...@@ -22,6 +22,8 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv1dFwdGNWC)
conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}}); conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
conv_params.push_back({1, 2, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}}); conv_params.push_back({1, 2, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}}); conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
conv_params.push_back({1, 1, 1, 1, 32, {3}, {32}, {1}, {1}, {1}, {1}});
conv_params.push_back({1, 1, 1, 64, 3, {3}, {32}, {1}, {1}, {1}, {1}});
for(auto& param : conv_params) for(auto& param : conv_params)
{ {
...@@ -96,6 +98,9 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdGNHWC) ...@@ -96,6 +98,9 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdGNHWC)
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}}); conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}}); conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
for(auto& param : conv_params) for(auto& param : conv_params)
{ {
...@@ -173,6 +178,12 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv3dFwdGNDHWC) ...@@ -173,6 +178,12 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv3dFwdGNDHWC)
{3, 2, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}); {3, 2, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
conv_params.push_back( conv_params.push_back(
{3, 2, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}}); {3, 2, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
conv_params.push_back(
{3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
this->conv_params.push_back(
{3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
conv_params.push_back(
{3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
for(auto& param : conv_params) for(auto& param : conv_params)
{ {
...@@ -247,6 +258,9 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC) ...@@ -247,6 +258,9 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC)
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}}); conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}}); conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
for(auto& param : conv_params) for(auto& param : conv_params)
{ {
...@@ -255,7 +269,7 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC) ...@@ -255,7 +269,7 @@ TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC)
// fp16 // fp16
pass = ck::profiler::profile_grouped_conv_fwd_impl<2, pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
ck::tensor_layout::convolution::NHWGC, ck::tensor_layout::convolution::NHWGC,
ck::tensor_layout::convolution::KYXGC, ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::NHWGK, ck::tensor_layout::convolution::NHWGK,
ck::half_t, ck::half_t,
ck::half_t, ck::half_t,
......
test/pool_fwd/CMakeLists.txt
View file @ 39002e9e
add_custom_target(test_pool_fwd) add_custom_target(test_pool_fwd)
add_gtest_executable(test_avg_pool2d_fwd test_avg_pool2d_fwd.cpp)
add_gtest_executable(test_avg_pool3d_fwd test_avg_pool3d_fwd.cpp) add_gtest_executable(test_avg_pool3d_fwd test_avg_pool3d_fwd.cpp)
add_gtest_executable(test_max_pool2d_fwd test_max_pool2d_fwd.cpp)
add_gtest_executable(test_max_pool3d_fwd test_max_pool3d_fwd.cpp) add_gtest_executable(test_max_pool3d_fwd test_max_pool3d_fwd.cpp)
target_link_libraries(test_avg_pool2d_fwd PRIVATE utility device_pool_fwd_instance) target_link_libraries(test_avg_pool3d_fwd PRIVATE utility device_pool3d_fwd_instance)
target_link_libraries(test_avg_pool3d_fwd PRIVATE utility device_pool_fwd_instance) target_link_libraries(test_max_pool3d_fwd PRIVATE utility device_pool3d_fwd_instance)
target_link_libraries(test_max_pool2d_fwd PRIVATE utility device_pool_fwd_instance)
target_link_libraries(test_max_pool3d_fwd PRIVATE utility device_pool_fwd_instance)
add_dependencies(test_pool_fwd test_avg_pool2d_fwd)
add_dependencies(test_pool_fwd test_avg_pool3d_fwd) add_dependencies(test_pool_fwd test_avg_pool3d_fwd)
add_dependencies(test_pool_fwd test_max_pool2d_fwd)
add_dependencies(test_pool_fwd test_max_pool3d_fwd) add_dependencies(test_pool_fwd test_max_pool3d_fwd)
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