Skip to content

GitLab

"sgl-kernel/csrc/moe/moe_align_kernel.cu" did not exist on "adeee15204a364559b20b0cfe1314f3090a97516"
Commit b7a6f810 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into fix_threadwise_copy_error_in_reduction

parents b29dfd70 827301d9
Show whitespace changes
Inline Side-by-side
Showing with 318 additions and 302 deletions
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f64_f64_f64.cpp
View file @ b7a6f810
......@@ -6,37 +6,37 @@ namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | ReduceDims
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 0, 1, 2); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 2, 1); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 0, 1, 2); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 2, 1); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 0, 1, 2); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 2, 1); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 0, 1, 2); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 2, 1); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 0, 1, 2); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 2, 1); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 0, 1, 2); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1); //
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 0, 1, 2); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 2, 1); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 2, 1);
// Will be moved to use MultiBlockAtomicAdd
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 0, 1, 2); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 2, 1); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 0, 1, 2); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 0); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1); //
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.cpp
View file @ b7a6f810
......@@ -6,25 +6,25 @@ 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, 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, 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, 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, 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, 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, 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 @ b7a6f810
......@@ -6,16 +6,16 @@ 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, 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, 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, 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 @ b7a6f810
......@@ -6,34 +6,34 @@ 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, 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, 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, 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, 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, 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, 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, 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, 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, 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 @ b7a6f810
......@@ -6,16 +6,16 @@ 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, 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, 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, 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 @ b7a6f810
......@@ -6,34 +6,34 @@ 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, 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, 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, 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, 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, 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, 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, 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, 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, 1);
ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
......
profiler/include/profile_reduce_impl.hpp
View file @ b7a6f810
......@@ -9,54 +9,52 @@ namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
template <int Rank, typename ReduceDims, int ReduceOpId, int NanOpt, int IndicesOpt>
template <int Rank, int NumReduceDim, int ReduceOpId, int NanOpt, int IndicesOpt>
struct ReduceDescription
{
static constexpr int Rank_ = Rank;
static constexpr int NumReduceDim_ = NumReduceDim;
static constexpr int ReduceOpId_ = ReduceOpId;
static constexpr int NanOpt_ = NanOpt;
static constexpr int IndicesOpt_ = IndicesOpt;
using ReduceDims_ = ReduceDims;
};
using reduce_description_instances =
std::tuple<ReduceDescription<4, Sequence<0, 1, 2>, 0, 0, 0>, // for ADD
ReduceDescription<4, Sequence<0>, 0, 0, 0>,
ReduceDescription<2, Sequence<1>, 0, 0, 0>,
ReduceDescription<4, Sequence<0, 1, 2>, 5, 0, 0>, // for AVG
ReduceDescription<4, Sequence<0>, 5, 0, 0>,
ReduceDescription<2, Sequence<1>, 5, 0, 0>,
ReduceDescription<4, Sequence<0, 1, 2>, 7, 0, 0>, // for NORM2
ReduceDescription<4, Sequence<0>, 7, 0, 0>,
ReduceDescription<2, Sequence<1>, 7, 0, 0>,
ReduceDescription<4, Sequence<0, 1, 2>, 2, 0, 0>, // for MIN
ReduceDescription<4, Sequence<0>, 2, 0, 0>,
ReduceDescription<2, Sequence<1>, 2, 0, 0>,
ReduceDescription<4, Sequence<0, 1, 2>, 3, 0, 0>, // for MAX
ReduceDescription<4, Sequence<0>, 3, 0, 0>,
ReduceDescription<2, Sequence<1>, 3, 0, 0>,
ReduceDescription<4, Sequence<0, 1, 2>, 4, 0, 0>, // for AMAX
ReduceDescription<4, Sequence<0>, 4, 0, 0>,
ReduceDescription<2, Sequence<1>, 4, 0, 0>,
ReduceDescription<4, Sequence<0, 1, 2>, 2, 0, 1>, // for MIN
ReduceDescription<4, Sequence<0>, 2, 0, 1>,
ReduceDescription<2, Sequence<1>, 2, 0, 1>,
ReduceDescription<4, Sequence<0, 1, 2>, 3, 0, 1>, // for MAX
ReduceDescription<4, Sequence<0>, 3, 0, 1>,
ReduceDescription<2, Sequence<1>, 3, 0, 1>,
ReduceDescription<4, Sequence<0, 1, 2>, 4, 0, 1>, // for AMAX
ReduceDescription<4, Sequence<0>, 4, 0, 1>,
ReduceDescription<2, Sequence<1>, 4, 0, 1>>;
using reduce_description_instances = std::tuple<ReduceDescription<4, 3, 0, 0, 0>, // for ADD
ReduceDescription<4, 1, 0, 0, 0>,
ReduceDescription<2, 1, 0, 0, 0>,
ReduceDescription<4, 3, 5, 0, 0>, // for AVG
ReduceDescription<4, 1, 5, 0, 0>,
ReduceDescription<2, 1, 5, 0, 0>,
ReduceDescription<4, 3, 7, 0, 0>, // for NORM2
ReduceDescription<4, 1, 7, 0, 0>,
ReduceDescription<2, 1, 7, 0, 0>,
ReduceDescription<4, 3, 2, 0, 0>, // for MIN
ReduceDescription<4, 1, 2, 0, 0>,
ReduceDescription<2, 1, 2, 0, 0>,
ReduceDescription<4, 3, 3, 0, 0>, // for MAX
ReduceDescription<4, 1, 3, 0, 0>,
ReduceDescription<2, 1, 3, 0, 0>,
ReduceDescription<4, 3, 4, 0, 0>, // for AMAX
ReduceDescription<4, 1, 4, 0, 0>,
ReduceDescription<2, 1, 4, 0, 0>,
ReduceDescription<4, 3, 2, 0, 1>, // for MIN
ReduceDescription<4, 1, 2, 0, 1>,
ReduceDescription<2, 1, 2, 0, 1>,
ReduceDescription<4, 3, 3, 0, 1>, // for MAX
ReduceDescription<4, 1, 3, 0, 1>,
ReduceDescription<2, 1, 3, 0, 1>,
ReduceDescription<4, 3, 4, 0, 1>, // for AMAX
ReduceDescription<4, 1, 4, 0, 1>,
ReduceDescription<2, 1, 4, 0, 1>>;
template <typename DescriptionType>
bool description_match(const DescriptionType& description,
int Rank,
const std::vector<int>& ReduceDims,
const std::vector<int>& reduceDims,
ReduceTensorOp_t ReduceOpId,
NanPropagation_t NanOpt,
ReduceTensorIndices_t IndicesOpt)
......@@ -66,16 +64,11 @@ bool description_match(const DescriptionType& description,
description.IndicesOpt_ != static_cast<int>(IndicesOpt))
return (false);
if(DescriptionType::ReduceDims_::Size() != ReduceDims.size())
if(DescriptionType::NumReduceDim_ != reduceDims.size())
return (false);
bool result = true;
static_for<0, DescriptionType::ReduceDims_::Size(), 1>{}([&](auto i) {
if(DescriptionType::ReduceDims_::At(i) != ReduceDims[i])
result = false;
});
return (result);
};
......@@ -87,33 +80,29 @@ bool description_match(const DescriptionType& description,
namespace ck {
namespace profiler {
template <int Rank, typename ReduceDims>
static std::vector<int> get_reduce_dims()
template <index_t Rank, index_t NumReduceDim>
static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduceDims)
{
std::vector<int> resDims;
assert(NumReduceDim == reduceDims.size());
static_for<0, ReduceDims::Size(), 1>{}([&](auto i) { resDims.push_back(ReduceDims::At(i)); });
int reduceFlag = 0;
return (resDims);
};
template <int Rank, typename ReduceDims>
static std::vector<int> get_invariant_dims()
{
std::vector<int> resDims;
unsigned int incFlag = 0;
// flag the bits for the reduceDims
for(int i = 0; i < NumReduceDim; i++)
{
reduceFlag |= 1 << reduceDims[i];
};
static_for<0, ReduceDims::Size(), 1>{}(
[&](auto i) { incFlag = incFlag | (0x1 << ReduceDims::At(i)); });
std::vector<int> invariantDims;
for(int dim = 0; dim < Rank; dim++)
// collect invariant dimensions
for(int i = 0; i < Rank; i++)
if((reduceFlag & (1 << i)) == 0)
{
if(incFlag & (0x1 << dim))
continue;
resDims.push_back(dim);
invariantDims.push_back(i);
};
return (resDims);
return invariantDims;
};
template <typename T>
......@@ -149,7 +138,7 @@ template <typename InDataType,
typename AccDataType,
typename OutDataType,
int Rank,
typename ReduceDims_,
int NumReduceDim,
ReduceTensorOp_t ReduceOpId,
NanPropagation_t NanOpt,
ReduceTensorIndices_t IndicesOpt>
......@@ -159,6 +148,7 @@ void profile_reduce_impl_impl(bool do_verification,
bool do_dumpout,
int nrepeat,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
float alpha,
float beta)
{
......@@ -203,15 +193,14 @@ void profile_reduce_impl_impl(bool do_verification,
{
Tensor<InDataType> in(inLengths);
const std::vector<int> OuterDims = get_invariant_dims<Rank, ReduceDims_>();
const std::vector<int> ReduceDims = get_reduce_dims<Rank, ReduceDims_>();
std::vector<size_t> outLengths;
if(OuterDims.empty())
const auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
if(reduceDims.size() == Rank)
outLengths.push_back(1);
else
for(auto dim : OuterDims)
for(auto dim : invariantDims)
outLengths.push_back(inLengths[dim]);
Tensor<OutDataType> out_ref(outLengths);
......@@ -302,7 +291,7 @@ void profile_reduce_impl_impl(bool do_verification,
AccDataType,
OutDataType,
Rank,
ReduceDims_,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
......@@ -311,7 +300,7 @@ void profile_reduce_impl_impl(bool do_verification,
AccDataType,
OutDataType,
Rank,
ReduceDims_,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
......@@ -321,7 +310,7 @@ void profile_reduce_impl_impl(bool do_verification,
AccDataType,
OutDataType,
Rank,
ReduceDims_,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
......@@ -330,7 +319,7 @@ void profile_reduce_impl_impl(bool do_verification,
AccDataType,
OutDataType,
Rank,
ReduceDims_,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce1_ptrs);
......@@ -341,7 +330,7 @@ void profile_reduce_impl_impl(bool do_verification,
AccDataType,
OutDataType,
Rank,
ReduceDims_,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce2_ptrs);
......@@ -358,7 +347,7 @@ void profile_reduce_impl_impl(bool do_verification,
using hCompType = typename type_mapping<AccDataType>::outDataType;
ReductionHost<hInType, hCompType, hOutType, ReduceOpId, PropagateNan, NeedIndices>
hostReduce(in.mDesc, out_ref.mDesc, OuterDims, ReduceDims);
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(alpha,
reinterpret_cast<const hInType*>(in.mData.data()),
......@@ -383,6 +372,7 @@ void profile_reduce_impl_impl(bool do_verification,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
......@@ -464,6 +454,7 @@ void profile_reduce_impl_impl(bool do_verification,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
......@@ -496,6 +487,7 @@ void profile_reduce_impl_impl(bool do_verification,
inStrides2,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
ws_dev.GetDeviceBuffer(),
......@@ -584,7 +576,7 @@ void profile_reduce_impl(bool do_verification,
bool do_dumpout,
int nrepeat,
const std::vector<size_t>& inLengths,
const std::vector<int>& ReduceDims,
const std::vector<int>& reduceDims,
ReduceTensorOp_t ReduceOpId,
NanPropagation_t NanOpt,
ReduceTensorIndices_t IndicesOpt,
......@@ -605,18 +597,26 @@ void profile_reduce_impl(bool do_verification,
using descType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
if(!description_match(
descType{}, inLengths.size(), ReduceDims, ReduceOpId, NanOpt, IndicesOpt))
descType{}, inLengths.size(), reduceDims, ReduceOpId, NanOpt, IndicesOpt))
return;
profile_reduce_impl_impl<InDataType,
AccDataType,
OutDataType,
descType::Rank_,
typename descType::ReduceDims_,
descType::NumReduceDim_,
static_cast<ReduceTensorOp_t>(descType::ReduceOpId_),
static_cast<NanPropagation_t>(descType::NanOpt_),
static_cast<ReduceTensorIndices_t>(descType::IndicesOpt_)>(
do_verification, init_method, do_log, do_dumpout, nrepeat, inLengths, alpha, beta);
do_verification,
init_method,
do_log,
do_dumpout,
nrepeat,
inLengths,
reduceDims,
alpha,
beta);
matched = true;
});
......
profiler/src/profile_reduce.cpp
View file @ b7a6f810
......@@ -25,7 +25,7 @@ using ck::ReduceTensorIndices_t;
using ck::ReduceTensorOp_t;
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
{"toReduceDims", required_argument, nullptr, 'R'},
{"reduceDims", required_argument, nullptr, 'R'},
{"reduceOp", required_argument, nullptr, 'O'},
{"compType", required_argument, nullptr, 'C'},
{"outType", required_argument, nullptr, 'W'},
......@@ -93,9 +93,9 @@ typedef enum
appDouble = 6,
} appDataType_t;
static void check_reduce_dims(const int rank, const std::vector<int>& toReduceDims)
static void check_reduce_dims(const int rank, const std::vector<int>& reduceDims)
{
for(auto dim : toReduceDims)
for(auto dim : reduceDims)
{
if(dim < 0 || dim >= rank)
throw std::runtime_error("Invalid dimension index specified for Reducing");
......@@ -103,7 +103,7 @@ static void check_reduce_dims(const int rank, const std::vector<int>& toReduceDi
unsigned int flag = 0;
for(auto dim : toReduceDims)
for(auto dim : reduceDims)
{
if(flag & (0x1 << dim))
throw std::runtime_error("All toReduce dimensions should be different!");
......@@ -122,7 +122,7 @@ class AppArgs
std::vector<size_t> inLengths;
std::vector<size_t> outLengths;
std::vector<int> toReduceDims;
std::vector<int> reduceDims;
std::vector<float> scales;
......@@ -152,7 +152,7 @@ class AppArgs
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
<< std::endl;
std::cout << "--toReduceDims or -R, comma separated list of to-reduce dimensions"
std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
<< std::endl;
std::cout << "--reduceOp or -O, enum value indicating the reduction operations"
<< std::endl;
......@@ -201,7 +201,7 @@ class AppArgs
if(!optarg)
throw std::runtime_error("Invalid option format!");
toReduceDims = getTypeValuesFromString<int>(optarg);
reduceDims = getTypeValuesFromString<int>(optarg);
break;
case 'O':
if(!optarg)
......@@ -321,7 +321,7 @@ int profile_reduce(int argc, char* argv[])
int rank = args.inLengths.size();
check_reduce_dims(rank, args.toReduceDims);
check_reduce_dims(rank, args.reduceDims);
if(args.reduceOp == ReduceTensorOp_t::MUL || args.reduceOp == ReduceTensorOp_t::NORM1)
throw std::runtime_error("MUL and NORM1 are not supported by composable kernel!");
......@@ -345,7 +345,7 @@ int profile_reduce(int argc, char* argv[])
args.do_dumpout,
args.nrepeat,
args.inLengths,
args.toReduceDims,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
......@@ -360,7 +360,7 @@ int profile_reduce(int argc, char* argv[])
args.do_dumpout,
args.nrepeat,
args.inLengths,
args.toReduceDims,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
......@@ -378,7 +378,7 @@ int profile_reduce(int argc, char* argv[])
args.do_dumpout,
args.nrepeat,
args.inLengths,
args.toReduceDims,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
......@@ -395,7 +395,7 @@ int profile_reduce(int argc, char* argv[])
args.do_dumpout,
args.nrepeat,
args.inLengths,
args.toReduceDims,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
......@@ -410,7 +410,7 @@ int profile_reduce(int argc, char* argv[])
args.do_dumpout,
args.nrepeat,
args.inLengths,
args.toReduceDims,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
......
script/profile_reduce_no_index.sh
View file @ b7a6f810
#!/bin/bash
PRECISION= ##--half
PRECISION=
##PRECISION=--half
##PRECISION=--double
if test -n $PRECISION && test "$PRECISION" = "--half"; then
CTYPE="-C 1"
ACCTYPE="-C 1"
else
CTYPE=""
ACCTYPE=""
fi
WTYPE=
driver="./bin/ckProfiler"
VERIFY="-v $1"
INIT=$2
NREPEAT=$3
if [ $# -ge 1 ] ; then
NREPEAT=$1
else
NREPEAT=1
fi
Operation=7
#### 0 - ADD, 5 - AVG, 7 - NORM2
Operations="0 5 7"
## for generic validation
for op in $Operation; do
for op in $Operations; do
set -x
./bin/ckProfiler reduce $PRECISION -D 64,4,280,82 -R 0 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 4,64,280,82 -R 0 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 280,4,64,82 -R 0 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 64,4,280,82 -R 0,1,2 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 4,64,280,82 -R 0,1,2 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 64,280,82,4 -R 0,1,2 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 700,8192 -R 1 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 700,1024 -R 1 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 700,4 -R 1 -O $op $CTYPE -v 1 1 $NREPEAT
####### datatype layout reduce dims op acctype verify init repeats
$driver reduce $PRECISION -D 64,4,280,82 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 1,2,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,2,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,22960 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,22960 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 4,1469440 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 4,1469440 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
set +x
done
Operation=5
#### 0 - ADD, 5 - AVG, 7 - NORM2
Operations=5
## for performance evaluation (resnet50 NHWC => C)
for op in $Operation; do
for op in $Operations; do
set -x
./bin/ckProfiler reduce $PRECISION -D 256,14,14,1024 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,28,28,128 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,58,58,128 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,7,7,2048 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,14,14,256 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,30,30,256 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,56,56,256 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,16,16,512 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,28,28,512 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,7,7,512 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,56,56,64 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,230,230,3 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,14,14,1024 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,28,28,128 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,58,58,128 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,7,7,2048 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,14,14,256 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,30,30,256 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,56,56,256 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,16,16,512 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,28,28,512 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,7,7,512 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,56,56,64 -R 0,1,2 -O $op $CTYPE $WTYPE -v 1 1 $NREPEAT
####### datatype layout reduce dims op acctype verify init repeats
$driver reduce $PRECISION -D 256,14,14,1024 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,28,28,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,58,58,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,7,7,2048 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,14,14,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,30,30,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,56,56,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,16,16,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,28,28,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,7,7,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,56,56,64 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,230,230,3 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,14,14,1024 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,28,28,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,58,58,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,7,7,2048 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,14,14,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,30,30,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,56,56,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,16,16,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,28,28,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,7,7,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,56,56,64 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
set +x
done
script/profile_reduce_with_index.sh
View file @ b7a6f810
#!/bin/bash
PRECISION= ##--half
PRECISION=
##PRECISION=--half
##PRECISION=--double
if [ $# -ge 1 ] ; then
NREPEAT=$1
else
NREPEAT=1
fi
driver="./bin/ckProfiler"
Operation=4
VERIFY="-v $1"
INIT=$2
NREPEAT=$3
LENGTHS=64,4,280,82
#### 2 - MIN, 3 - MAX, 4 - AMAX
Operations="2 4"
## for generic validation
for op in $Operation; do
for op in $Operations; do
for use_idx in 0 1; do
set -x
./bin/ckProfiler reduce $PRECISION -D 64,4,280,82 -R 0 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 4,64,280,82 -R 0 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 280,4,64,82 -R 0 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 64,4,280,82 -R 0,1,2 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 4,64,280,82 -R 0,1,2 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 64,280,82,4 -R 0,1,2 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 700,8192 -R 1 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 700,1024 -R 1 -O $op $CTYPE -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 700,4 -R 1 -O $op $CTYPE -v 1 1 $NREPEAT
####### datatype layout reduce dims op use index verify init repeats
$driver reduce $PRECISION -D 64,4,280,82 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 1,2,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,2,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,22960 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,22960 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 4,1469440 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 4,1469440 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
set +x
done
done
Operations=2
## for performance evaluation (resnet50 NHWC => C)
for op in $Operation; do
for op in $Operations; do
for use_idx in 0 1; do
set -x
./bin/ckProfiler reduce $PRECISION -D 256,14,14,1024 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,28,28,128 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,58,58,128 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,7,7,2048 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,14,14,256 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,30,30,256 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,56,56,256 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,16,16,512 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,28,28,512 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,7,7,512 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,56,56,64 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 256,230,230,3 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,14,14,1024 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,28,28,128 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,58,58,128 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,7,7,2048 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,14,14,256 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,30,30,256 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,56,56,256 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,16,16,512 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,28,28,512 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,7,7,512 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
./bin/ckProfiler reduce $PRECISION -D 128,56,56,64 -R 0,1,2 -O $op -I $use_idx -v 1 1 $NREPEAT
####### datatype layout reduce dims op use index verify init repeats
$driver reduce $PRECISION -D 256,14,14,1024 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,28,28,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,58,58,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,7,7,2048 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,14,14,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,30,30,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,56,56,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,16,16,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,28,28,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,7,7,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,56,56,64 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,230,230,3 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,14,14,1024 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,28,28,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,58,58,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,7,7,2048 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,14,14,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,30,30,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,56,56,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,16,16,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,28,28,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,7,7,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,56,56,64 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
set +x
done
done
......
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