Skip to content

GitLab

"configs/_base_/datasets/scannet-3d.py" did not exist on "aee606fc3d89245dc5fa6bb4c70da0f2e6f74933"
Commit dd6a8de4 authored by Jehandad Khan's avatar Jehandad Khan
Browse files

Merge branch 'develop' into jd/dev_pkg

parents 0aa899aa abf4bdb9
Hide whitespace changes
Inline Side-by-side
Showing with 1668 additions and 118 deletions
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i8_i8.cpp 0 → 100644
View file @ dd6a8de4
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16.cpp 0 → 100644
View file @ dd6a8de4
#include "device_reduce_instance_threadwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 3); // for ADD
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 3); // for AVG
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 3); // for NORM2
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.cpp
View file @ dd6a8de4
......@@ -6,25 +6,31 @@ namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | ReduceDims
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 0, 1, 2); // for MIN
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 0, 1, 2); // for MAX
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 0, 1, 2); // for AMAX
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 0, 1, 2); // for MIN
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 2, 1); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 0, 1, 2); // for MAX
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 2, 1); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 0, 1, 2); // for AMAX
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 2, 1); //
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16.cpp
View file @ dd6a8de4
......@@ -6,16 +6,19 @@ namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | ReduceDims
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 0, 1, 2); // for ADD
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 0);
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 3); // for ADD
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 0, 1, 2); // for AVG
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 0, 1, 2); // for NORM2
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 3); // for AVG
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 3); // for NORM2
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32.cpp
View file @ dd6a8de4
......@@ -6,34 +6,43 @@ namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | ReduceDims
ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 4, 0, 1, 2); // for ADD
ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 4, 0);
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 4, 3); // for ADD
ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 4, 0, 1, 2); // for AVG
ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 0, 1, 2); // for NORM2
ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 0, 1, 2); // for MIN
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 0, 1, 2); // for MAX
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 0, 1, 2); // for AMAX
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 0, 1, 2); // for MIN
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 0, 1, 2); // for MAX
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 0, 1, 2); // for AMAX
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 4, 3); // for AVG
ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 3); // for NORM2
ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32.cpp
View file @ dd6a8de4
......@@ -6,16 +6,19 @@ namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | ReduceDims
ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 4, 0, 1, 2); // for ADD
ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 4, 0);
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 4, 3); // for ADD
ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 4, 0, 1, 2); // for AVG
ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 0, 1, 2); // for NORM2
ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 4, 3); // for AVG
ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 3); // for NORM2
ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64.cpp
View file @ dd6a8de4
......@@ -6,34 +6,43 @@ namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | ReduceDims
ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 0, 1, 2); // for ADD
ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 0);
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 3); // for ADD
ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 0, 1, 2); // for AVG
ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 0, 1, 2); // for NORM2
ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 0, 1, 2); // for MIN
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 0, 1, 2); // for MAX
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 0, 1, 2); // for AMAX
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 0); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 2, 1); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 0, 1, 2); // for MIN
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 2, 1); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 0, 1, 2); // for MAX
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 2, 1); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 0, 1, 2); // for AMAX
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 0); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1); //
ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 3); // for AVG
ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 3); // for NORM2
ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8.cpp 0 → 100644
View file @ dd6a8de4
#include "device_reduce_instance_threadwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
// clang-format on
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8.cpp 0 → 100644
View file @ dd6a8de4
#include "device_reduce_instance_threadwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 4);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 2, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 3); // for MIN
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 3); // for MAX
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 2, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 3); // for AMAX
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 4);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);
ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ dd6a8de4
......@@ -15,6 +15,7 @@ include_directories(BEFORE
${PROJECT_SOURCE_DIR}/library/include/ck/library/tensor_operation_instance/gpu/reduce
${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/cpu
${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/gpu
${PROJECT_SOURCE_DIR}/library/include/ck/library/utility
${PROJECT_SOURCE_DIR}/profiler/include
${PROJECT_SOURCE_DIR}/external/include/half
)
......@@ -26,18 +27,23 @@ set(PROFILER_SOURCE
src/profile_gemm_bias_2d.cpp
src/profile_gemm_bias_relu.cpp
src/profile_gemm_bias_relu_add.cpp
src/profile_gemm_reduce.cpp
src/profile_batched_gemm.cpp
src/profile_conv_fwd.cpp
src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp
src/profile_conv_fwd_bias_relu_atomic_add.cpp
src/profile_conv_bwd_data.cpp
src/profile_convnd_bwd_data.cpp
src/profile_reduce.cpp
src/profile_grouped_gemm.cpp
src/profile_conv_bwd_weight.cpp
src/profile_batched_gemm_reduce.cpp
)
add_executable(ckProfiler ${PROFILER_SOURCE})
target_link_libraries(ckProfiler PRIVATE host_tensor)
target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)
......@@ -47,7 +53,11 @@ target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_atomic_add_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
set_target_properties(ckProfiler PROPERTIES EXCLUDE_FROM_ALL 1)
\ No newline at end of file
set_target_properties(ckProfiler PROPERTIES EXCLUDE_FROM_ALL 1)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
profiler/src/README.md → profiler/README.md
View file @ dd6a8de4
## Docker script
```bash
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```ckProfiler```
```bash
mkdir build && cd build
```
```bash
# Need to Specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
```bash
make -j ckProfiler
```
## Profile GEMM kernels
```bash
#arg1: tensor operation (gemm=GEMM)
......@@ -42,8 +9,8 @@ cmake \
#arg7: run kernel # of times (>1)
#arg8 to 13: M, N, K, StrideA, StrideB, StrideC
##################### op datatype layout verify init log repeat M___ N___ K___ StrideA StrideB StrideC
./profiler/ckProfiler gemm 1 1 1 1 0 5 3840 4096 4096 4096 4096 4096
################ op datatype layout verify init log repeat M___ N___ K___ StrideA StrideB StrideC
./bin/ckProfiler gemm 1 1 1 1 0 5 3840 4096 4096 4096 4096 4096
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
......@@ -55,7 +22,7 @@ c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
```
## Profile forward convolution kernels
## Profile 2d forward convolution kernels
```bash
#arg1: tensor operation (conv=Convolution)
#arg2: data type (0=fp32, 1=fp16)
......@@ -67,8 +34,8 @@ Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
#arg8: print matrix value (0=no, 1=yes)
#arg9: run kernel # of times (>1)
#arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
##################### op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
./profiler/ckProfiler conv 1 1 1 1 1 1 0 5 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
################ op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
./bin/ckProfiler conv2d_fwd 1 1 1 1 1 1 0 5 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
......
profiler/include/profile_batched_gemm_impl.hpp
View file @ dd6a8de4
#pragma once
#include <memory>
#include "check_err.hpp"
#include "config.hpp"
#include "element_wise_operation.hpp"
#include "tensor_layout.hpp"
#include "device.hpp"
#include "host_tensor_generator.hpp"
#include "device_gemm.hpp"
#include "reference_batched_gemm.hpp"
namespace ck {
......@@ -11,10 +21,30 @@ using DeviceGemmNoOpPtr =
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(
std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_batched_gemm_instance
} // namespace device
......@@ -30,7 +60,7 @@ template <typename ADataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_batched_gemm_impl(int do_verification,
bool profile_batched_gemm_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
......@@ -42,6 +72,8 @@ void profile_batched_gemm_impl(int do_verification,
int StrideC,
int BatchCount = 1)
{
bool pass = true;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
......@@ -65,12 +97,14 @@ void profile_batched_gemm_impl(int do_verification,
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
std::unique_ptr<Tensor<float>> c_f32_g_m_n_host_result = nullptr;
std::unique_ptr<Tensor<float>> c_f32_g_m_n_device_result = nullptr;
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency();
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
......@@ -95,21 +129,56 @@ void profile_batched_gemm_impl(int do_verification,
if(do_verification)
{
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
is_same<BDataType, ck::bhalf_t>::value &&
is_same<CDataType, ck::bhalf_t>::value)
{
Tensor<float> a_f32_g_m_k(
f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
Tensor<float> b_f32_g_k_n(
f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
c_f32_g_m_n_host_result = std::make_unique<Tensor<float>>(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
c_f32_g_m_n_device_result = std::make_unique<Tensor<float>>(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
bf16_to_f32_(a_g_m_k, a_f32_g_m_k);
bf16_to_f32_(b_g_k_n, b_f32_g_k_n);
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
ReferenceBatchedGemm<float, float, float, AElementOp, BElementOp, CElementOp>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument = ref_batched_gemm.MakeArgument(a_f32_g_m_k,
b_f32_g_k_n,
*c_f32_g_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
}
else
{
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
ref_invoker.Run(ref_argument);
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
......@@ -156,6 +225,102 @@ void profile_batched_gemm_impl(int do_verification,
add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ADataType, bhalf_t>::value && is_same<BDataType, bhalf_t>::value &&
is_same<CDataType, bhalf_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
is_same<CDataType, float>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
is_same<CDataType, int8_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_batched_gemm_instance::
add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
......@@ -218,7 +383,20 @@ void profile_batched_gemm_impl(int do_verification,
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
check_error(c_g_m_n_host_result, c_g_m_n_device_result);
if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
is_same<BDataType, ck::bhalf_t>::value &&
is_same<CDataType, ck::bhalf_t>::value)
{
bf16_to_f32_(c_g_m_n_device_result, *c_f32_g_m_n_device_result);
float err = check_error(*c_f32_g_m_n_host_result, *c_f32_g_m_n_device_result);
pass = pass && (err < 1E-6);
}
else
{
float err = check_error(c_g_m_n_host_result, c_g_m_n_device_result);
pass = pass && (err < 1E-6);
}
if(do_log)
{
......@@ -241,6 +419,8 @@ void profile_batched_gemm_impl(int do_verification,
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
return pass;
}
} // namespace profiler
......
profiler/include/profile_batched_gemm_reduce_impl.hpp 0 → 100644
View file @ dd6a8de4
#pragma once
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "device_gemm_reduce.hpp"
#include "reference_batched_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::ReduceSum,
ck::tensor_operation::element_wise::ReduceSquareSum>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename ALayout,
typename BLayout,
typename CLayout>
bool profile_batched_gemm_reduce_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int BatchCount)
{
bool pass = true;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({row * stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({col * stride, 1, stride}));
}
};
Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
std::cout << "d0_g_m: " << d0_g_m_host_result.mDesc << std::endl;
std::cout << "d1_g_m: " << d1_g_m_host_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method)
{
case 0: break;
case 1:
std::srand(0);
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
std::srand(0);
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::tensor_operation::element_wise::ReduceSum;
using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
if(do_verification)
{
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
for(int batch = 0; batch < BatchCount; ++batch)
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReduceZeroValue();
float d1_acc = d1_reduce_op.GetReduceZeroValue();
for(int n = 0; n < N; ++n)
{
d0_reduce_op.Reduce(d0_acc, c_g_m_n_host_result(batch, m, n));
d1_reduce_op.Reduce(d1_acc, c_g_m_n_host_result(batch, m, n));
}
d0_g_m_host_result(batch, m) = d0_acc;
d1_g_m_host_result(batch, m) = d1_acc;
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmReduceNoOpPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
d0_reduce_op,
d1_reduce_op,
BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// warm up
invoker_ptr->Run(argument_ptr.get());
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker_ptr->Run(argument_ptr.get());
timer.End();
total_time += timer.GetElapsedTime();
}
float ave_time = total_time / nrepeat;
std::string gemm_name = gemm_ptr->GetTypeString();
std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
sizeof(BDataType) * BatchCount * K * N +
sizeof(CDataType) * BatchCount * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
float c_error = check_error(c_g_m_n_host_result, c_g_m_n_device_result);
float d0_error = check_error(d0_g_m_host_result, d0_g_m_device_result);
float d1_error = check_error(d1_g_m_host_result, d1_g_m_device_result);
pass = pass && (c_error < 1E-6);
pass = pass && (d0_error < 1E-6);
pass = pass && (d1_error < 1E-6);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_host: ", d0_g_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "d0_device: ", d0_g_m_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_host: ", d1_g_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "d1_device: ", d1_g_m_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_conv_bwd_data_impl.hpp
View file @ dd6a8de4
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -11,7 +13,7 @@
using F16 = ck::half_t;
using F32 = float;
using BF16 = ushort;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
......@@ -42,6 +44,7 @@ template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
......@@ -123,6 +126,7 @@ void profile_conv_bwd_data_impl(int do_verification,
ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
......@@ -172,9 +176,9 @@ void profile_conv_bwd_data_impl(int do_verification,
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ushort> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ushort> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ushort>)
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::bhalf_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
......@@ -251,7 +255,8 @@ void profile_conv_bwd_data_impl(int do_verification,
{
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
check_error(in_n_c_hi_wi_host_result, in_n_c_hi_wi_device_result);
ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData);
if(do_log)
{
......
profiler/include/profile_conv_bwd_weight_impl.hpp 0 → 100644
View file @ dd6a8de4
#pragma once
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_backward_weight.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_backward_weight.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_weight_instance {
using DeviceConvBwdWeightNoOpPtr =
DeviceConvBwdWeightPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdWeightNoOpPtr>&);
void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdWeightNoOpPtr>&);
} // namespace device_conv2d_bwd_weight_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
bool profile_conv_bwd_weight_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
ck::index_t split_k)
{
const ck::index_t Y = filter_spatial_lengths[0];
const ck::index_t X = filter_spatial_lengths[1];
const ck::index_t Hi = input_spatial_lengths[0];
const ck::index_t Wi = input_spatial_lengths[1];
const ck::index_t Ho = output_spatial_lengths[0];
const ck::index_t Wo = output_spatial_lengths[1];
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x_host_result(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<WeiDataType> wei_k_c_y_x_device_result(
f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
}
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
using ReferenceConvBwdWeightInstance =
ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvBwdWeightInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x_host_result,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) *
wei_k_c_y_x_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvBwdWeightNoOpPtr =
ck::tensor_operation::device::DeviceConvBwdWeightPtr<PassThrough, PassThrough, PassThrough>;
// add device Conv instances
std::vector<DeviceConvBwdWeightNoOpPtr> conv_ptrs;
if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
ck::tensor_operation::device::device_conv2d_bwd_weight_instance::
add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_weight_instance::
add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
bool pass = true;
for(auto& conv_ptr : conv_ptrs)
{
// using atomic, so need to reset input
if(split_k > 1)
{
wei_device_buf.SetZero();
}
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op,
split_k);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << conv_name << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
float max_error = check_error(wei_k_c_y_x_host_result, wei_k_c_y_x_device_result);
if(max_error > 8)
{
pass = false;
std::cout << "Fail info:" << conv_ptr->GetTypeString() << std::endl;
}
if(do_log)
{
LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "wei_device: ", wei_k_c_y_x_device_result.mData, ",")
<< std::endl;
}
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp
View file @ dd6a8de4
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -245,7 +247,8 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
{
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
check_error(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result);
ck::utils::check_err(out_n_k_ho_wo_device_result.mData,
out_n_k_ho_wo_host_result.mData);
if(do_log)
{
......
profiler/include/profile_conv_fwd_bias_relu_atomic_add_impl.hpp
View file @ dd6a8de4
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -301,7 +302,8 @@ void profile_conv_fwd_bias_relu_atomic_add_impl(int do_verification,
{
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
check_error(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result);
ck::utils::check_err(out_n_k_ho_wo_device_result.mData,
out_n_k_ho_wo_host_result.mData);
if(do_log)
{
......
profiler/include/profile_conv_fwd_bias_relu_impl.hpp
View file @ dd6a8de4
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -233,7 +234,8 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
{
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
check_error(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result);
ck::utils::check_err(out_n_k_ho_wo_device_result.mData,
out_n_k_ho_wo_host_result.mData);
if(do_log)
{
......
profiler/include/profile_conv_fwd_impl.hpp
View file @ dd6a8de4
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -253,7 +255,8 @@ void profile_conv_fwd_impl(int do_verification,
{
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
check_error(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result);
ck::utils::check_err(out_n_k_ho_wo_device_result.mData,
out_n_k_ho_wo_host_result.mData);
if(do_log)
{
......
profiler/include/profile_convnd_bwd_data_impl.hpp 0 → 100644
View file @ dd6a8de4
#pragma once
#include "config.hpp"
#include "device.hpp"
#include "conv_fwd_util.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_bwd_data.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_bwd_data.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace device_conv2d_bwd_data_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
using DeviceConvBwdDataNoOpPtr =
ck::tensor_operation::device::device_conv2d_bwd_data_instance::DeviceConvBwdDataNoOpPtr;
template <typename InLayout>
HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, InLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename WeiLayout>
HostTensorDescriptor get_filters_host_tensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, WeiLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename OutLayout>
HostTensorDescriptor get_output_host_ensor_descriptor(const std::vector<std::size_t>& dims,
int num_dim_spatial = 2)
{
namespace tl = ck::tensor_layout::convolution;
switch(num_dim_spatial)
{
case 3: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
case 2: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
case 1: {
return ck::utils::conv::get_host_tensor_descriptor(dims, OutLayout{});
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
template <typename InDataType, typename WeiDataType, typename OutDataType>
void get_device_conv_bwd_data_op_ptr(
InDataType, WeiDataType, OutDataType, std::vector<DeviceConvBwdDataNoOpPtr>&, int)
{
std::cout << "can not find device conv bwd data" << std::endl;
exit(1);
}
template <>
void get_device_conv_bwd_data_op_ptr(
F32, F32, F32, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
F16, F16, F16, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
BF16, BF16, BF16, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(conv_ptrs);
break;
default: break;
}
}
template <>
void get_device_conv_bwd_data_op_ptr(
INT8, INT8, INT8, std::vector<DeviceConvBwdDataNoOpPtr>& conv_ptrs, int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 1:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(conv_ptrs);
break;
case 2:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
break;
case 3:
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(conv_ptrs);
break;
default: break;
}
}
template <typename T>
static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
{
float max_diff = 1e-6;
for(int i = 0; i < ref.mData.size(); ++i)
{
float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
if(max_diff < diff)
{
return false;
}
}
return true;
}
template <typename DataType>
void show_data_nhwc_layout(Tensor<DataType>& nhwc)
{
std::cout << "[";
for(int n = 0; n < nhwc.mDesc.GetLengths()[0]; n++)
{
std::cout << "[";
for(int hi = 0; hi < nhwc.mDesc.GetLengths()[2]; hi++)
{
std::cout << "[";
for(int wi = 0; wi < nhwc.mDesc.GetLengths()[3]; wi++)
{
std::cout << "[";
for(int c = 0; c < nhwc.mDesc.GetLengths()[1]; c++)
{
std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << " ";
}
std::cout << "]";
}
std::cout << "]";
}
std::cout << "]";
}
std::cout << "]";
}
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
bool profile_convnd_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
ck::index_t N,
ck::index_t K,
ck::index_t C,
const std::vector<ck::index_t>& input_spatial_lengths,
const std::vector<ck::index_t>& filter_spatial_lengths,
const std::vector<ck::index_t>& output_spatial_lengths,
const std::vector<ck::index_t>& conv_filter_strides,
const std::vector<ck::index_t>& conv_filter_dilations,
const std::vector<ck::index_t>& input_left_pads,
const std::vector<ck::index_t>& input_right_pads)
{
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(C)};
input_dims.insert(
std::end(input_dims), std::begin(input_spatial_lengths), std::end(input_spatial_lengths));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(K), static_cast<std::size_t>(C)};
filter_dims.insert(std::end(filter_dims),
std::begin(filter_spatial_lengths),
std::end(filter_spatial_lengths));
std::vector<std::size_t> output_dims{static_cast<std::size_t>(N), static_cast<std::size_t>(K)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
Tensor<InDataType> input_host_result(
get_input_host_tensor_descriptor<InLayout>(input_dims, NDimSpatial));
Tensor<InDataType> input_device_result(
get_input_host_tensor_descriptor<InLayout>(input_dims, NDimSpatial));
Tensor<WeiDataType> weights(
get_filters_host_tensor_descriptor<WeiLayout>(filter_dims, NDimSpatial));
Tensor<OutDataType> output(
get_output_host_ensor_descriptor<OutLayout>(output_dims, NDimSpatial));
std::cout << "input: " << input_host_result.mDesc << std::endl;
std::cout << "weights: " << weights.mDesc << std::endl;
std::cout << "output: " << output.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
output.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
weights.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * input_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * weights.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpace());
out_device_buf.ToDevice(output.mData.data());
wei_device_buf.ToDevice(weights.mData.data());
// reset input to zero
in_device_buf.SetZero();
if(do_verification)
{
auto RunReference = [&](auto& ref_conv) {
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(input_host_result,
weights,
output,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
};
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NDimSpatial>();
RunReference(ref_conv);
}
// add device Conv instances
std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
get_device_conv_bwd_data_op_ptr(
InDataType{}, WeiDataType{}, OutDataType{}, conv_ptrs, NDimSpatial);
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
bool success = true;
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop =
ck::utils::conv::get_flops(N, C, K, filter_spatial_lengths, output_spatial_lengths);
std::size_t num_btype =
ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
N, C, K, input_spatial_lengths, filter_spatial_lengths, output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s" << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
in_device_buf.FromDevice(input_device_result.mData.data());
if(!check_out(input_host_result, input_device_result))
{
std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
success = false;
}
else
{
std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
}
check_error(input_host_result, input_device_result);
if(do_log)
{
std::cout << "in : ";
show_data_nhwc_layout(output);
std::cout << std::endl;
std::cout << "wei: ";
show_data_nhwc_layout(weights);
std::cout << std::endl;
std::cout << "out_host : ";
show_data_nhwc_layout(input_host_result);
std::cout << std::endl;
std::cout << "out_device: ";
show_data_nhwc_layout(input_device_result);
std::cout << std::endl;
}
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
return success;
}
} // namespace profiler
} // namespace ck
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