aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/NEReorderKernel.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/NEReorderKernel.cpp')
-rw-r--r--src/core/NEON/kernels/NEReorderKernel.cpp70
1 files changed, 43 insertions, 27 deletions
diff --git a/src/core/NEON/kernels/NEReorderKernel.cpp b/src/core/NEON/kernels/NEReorderKernel.cpp
index 1a7f58bb08..f92a4c87da 100644
--- a/src/core/NEON/kernels/NEReorderKernel.cpp
+++ b/src/core/NEON/kernels/NEReorderKernel.cpp
@@ -24,11 +24,13 @@
#if defined(__aarch64__)
#include "src/core/NEON/kernels/NEReorderKernel.h"
-#include "src/common/utils/Log.h"
-#include "src/core/NEON/kernels/arm_gemm/transform.hpp"
+
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/Validate.h"
+#include "src/common/utils/Log.h"
+#include "src/core/NEON/kernels/arm_gemm/transform.hpp"
+
namespace arm_compute
{
@@ -37,29 +39,32 @@ void NEReorderKernel::run(const Window &window, const ThreadInfo &info)
ARM_COMPUTE_UNUSED(info);
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
- switch(_input->info()->data_type())
+ switch (_input->info()->data_type())
{
case DataType::F32:
{
const int ksize_rows_elements = _xmax * _ksize;
- const int jump_rows = ksize_rows_elements * window.x().start();
- const int k_start = window.x().start() * _ksize;
- const int k_end = std::min(window.x().end() * _ksize, _kmax);
- const int stride = _kmax;
- if(k_start < k_end)
+ const int jump_rows = ksize_rows_elements * window.x().start();
+ const int k_start = window.x().start() * _ksize;
+ const int k_end = std::min(window.x().end() * _ksize, _kmax);
+ const int stride = _kmax;
+ if (k_start < k_end)
{
-
- switch(_output_wf)
+ switch (_output_wf)
{
case WeightFormat::OHWIo4:
{
- arm_gemm::Transform<4, 1, true, arm_gemm::VLType::None>(reinterpret_cast<float *>(_output->buffer()) + jump_rows, reinterpret_cast<float *>(_input->buffer()), stride, k_start, k_end, 0, _xmax);
+ arm_gemm::Transform<4, 1, true, arm_gemm::VLType::None>(
+ reinterpret_cast<float *>(_output->buffer()) + jump_rows,
+ reinterpret_cast<float *>(_input->buffer()), stride, k_start, k_end, 0, _xmax);
break;
}
#if defined(ARM_COMPUTE_ENABLE_SVE)
case WeightFormat::OHWIo8:
{
- arm_gemm::Transform<1, 1, true, arm_gemm::VLType::SVE>(reinterpret_cast<float *>(_output->buffer()) + jump_rows, reinterpret_cast<float *>(_input->buffer()), stride, k_start, k_end, 0, _xmax);
+ arm_gemm::Transform<1, 1, true, arm_gemm::VLType::SVE>(
+ reinterpret_cast<float *>(_output->buffer()) + jump_rows,
+ reinterpret_cast<float *>(_input->buffer()), stride, k_start, k_end, 0, _xmax);
break;
}
#endif /* ARM_COMPUTE_ENABLE_SVE */
@@ -78,11 +83,20 @@ void NEReorderKernel::run(const Window &window, const ThreadInfo &info)
}
NEReorderKernel::NEReorderKernel()
- : _input(nullptr), _output(nullptr), _ksize(0), _kmax(0), _xmax(0), _input_wf(WeightFormat::ANY), _output_wf(WeightFormat::ANY)
+ : _input(nullptr),
+ _output(nullptr),
+ _ksize(0),
+ _kmax(0),
+ _xmax(0),
+ _input_wf(WeightFormat::ANY),
+ _output_wf(WeightFormat::ANY)
{
}
-void NEReorderKernel::configure(const ITensor *input, ITensor *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf)
+void NEReorderKernel::configure(const ITensor *input,
+ ITensor *output,
+ arm_compute::WeightFormat input_wf,
+ arm_compute::WeightFormat output_wf)
{
ARM_COMPUTE_LOG_PARAMS(input, output, input_wf, output_wf);
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
@@ -96,7 +110,7 @@ void NEReorderKernel::configure(const ITensor *input, ITensor *output, arm_compu
// Setting parameters for transform
auto dims = input->info()->num_dimensions();
- switch(dims)
+ switch (dims)
{
case 2:
{
@@ -120,7 +134,7 @@ void NEReorderKernel::configure(const ITensor *input, ITensor *output, arm_compu
// Window size is set by rows / _ksize
Window win;
int window_size = 0;
- switch(_output_wf)
+ switch (_output_wf)
{
#if defined(ARM_COMPUTE_ENABLE_SVE)
case WeightFormat::OHWIo8:
@@ -142,7 +156,7 @@ void NEReorderKernel::configure(const ITensor *input, ITensor *output, arm_compu
break;
}
}
- if(_kmax % _ksize != 0)
+ if (_kmax % _ksize != 0)
{
window_size += 1;
}
@@ -152,11 +166,14 @@ void NEReorderKernel::configure(const ITensor *input, ITensor *output, arm_compu
INEKernel::configure(win);
}
-Status NEReorderKernel::validate(const ITensorInfo *input, const ITensorInfo *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf)
+Status NEReorderKernel::validate(const ITensorInfo *input,
+ const ITensorInfo *output,
+ arm_compute::WeightFormat input_wf,
+ arm_compute::WeightFormat output_wf)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
- if(output->tensor_shape().total_size() != 0)
+ if (output->tensor_shape().total_size() != 0)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
@@ -167,20 +184,20 @@ Status NEReorderKernel::validate(const ITensorInfo *input, const ITensorInfo *ou
int output_x_dim;
int output_k_dim;
auto dims = output->num_dimensions();
- switch(dims)
+ switch (dims)
{
case 2:
{
- input_x_dim = input->dimension(0); // Number of columns in input matrix
- input_k_dim = input->dimension(1); // Number of rows in input matrix
+ input_x_dim = input->dimension(0); // Number of columns in input matrix
+ input_k_dim = input->dimension(1); // Number of rows in input matrix
output_x_dim = output->dimension(0); // Number of columns in output matrix
output_k_dim = output->dimension(1); // Number of rows in output matrix
break;
}
case 4:
{
- input_x_dim = input->dimension(2); // Number of columns in input matrix
- input_k_dim = input->dimension(3); // Number of rows in input matrix
+ input_x_dim = input->dimension(2); // Number of columns in input matrix
+ input_k_dim = input->dimension(3); // Number of rows in input matrix
output_x_dim = output->dimension(2); // Number of columns in output matrix
output_k_dim = output->dimension(3); // Number of rows in output matrix
break;
@@ -192,7 +209,7 @@ Status NEReorderKernel::validate(const ITensorInfo *input, const ITensorInfo *ou
}
int ksize;
- switch(output_wf)
+ switch (output_wf)
{
case WeightFormat::OHWIo8:
{
@@ -216,11 +233,10 @@ Status NEReorderKernel::validate(const ITensorInfo *input, const ITensorInfo *ou
ARM_COMPUTE_RETURN_ERROR_ON(rnd_up_input_kdim != output_k_dim);
// output x_dim needs to be same as input
ARM_COMPUTE_RETURN_ERROR_ON(input_x_dim != output_x_dim);
-
}
return Status{};
}
} // namespace arm_compute
-#endif // defined(__aarch64__) \ No newline at end of file
+#endif // defined(__aarch64__)
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-09 21:09:46 +0000