COMPMID-934: Asymmetric padding support.
Change-Id: Ibe7a679e4c053a088b8c893e495c97cb24bf7272
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/121298
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
index 536a667..4dc186a 100644
--- a/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
@@ -274,6 +274,7 @@
ARM_COMPUTE_ERROR_ON(input->info()->dimension(Window::DimX) > small_tensor_size_optim);
ARM_COMPUTE_ERROR_ON(input->info()->dimension(Window::DimY) > small_tensor_size_optim);
+ const int input_stride_x = input->info()->strides_in_bytes().x();
const int input_stride_y = input->info()->strides_in_bytes().y();
const int input_stride_z = input->info()->strides_in_bytes().z();
const int output_stride_y = output->info()->strides_in_bytes().y();
@@ -284,6 +285,8 @@
const int range_z = window.z().end() - window.z().start();
const int kernel_depth = weights->info()->dimension(Window::DimZ);
const unsigned int conv_stride_y = std::get<1>(conv_info.stride());
+ const unsigned int conv_pad_left = conv_info.pad_left();
+ const unsigned int conv_pad_top = conv_info.pad_top();
// setup output window for the iterator
Window window_out = window;
@@ -307,7 +310,7 @@
execute_window_loop(window_out, [&](const Coordinates & id)
{
- const uint8_t *input_ptr = in.ptr();
+ const uint8_t *input_ptr = in.ptr() - conv_pad_left * input_stride_x - conv_pad_top * input_stride_y;
uint8_t *out_ptr = out.ptr();
int ih = 0;
int oh = 0;
@@ -351,6 +354,7 @@
static void convolve(const Window &window, unsigned int num_elems_read_per_iteration, unsigned int num_elems_written_per_iteration,
const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info)
{
+ const int input_stride_x = input->info()->strides_in_bytes().x();
const int input_stride_y = input->info()->strides_in_bytes().y();
const int input_stride_z = input->info()->strides_in_bytes().z();
const int output_stride_y = output->info()->strides_in_bytes().y();
@@ -362,6 +366,8 @@
const int range_z = window.z().end() - window.z().start();
const int kernel_depth = weights->info()->dimension(Window::DimZ);
const unsigned int conv_stride_y = std::get<1>(conv_info.stride());
+ const unsigned int conv_pad_left = conv_info.pad_left();
+ const unsigned int conv_pad_top = conv_info.pad_top();
const int fixed_point_position = input->info()->fixed_point_position();
// setup output window for the iterator
@@ -389,7 +395,7 @@
/*
For a detailed explanation on how the algorithm works refer to template <> class convolver_3x3<1>
*/
- const uint8_t *input_ptr = in.ptr();
+ const uint8_t *input_ptr = in.ptr() - conv_pad_left * input_stride_x - conv_pad_top * input_stride_y;
uint8_t *out_ptr = out.ptr();
int ih = 0;
int oh = 0;
@@ -680,8 +686,8 @@
const int delta_input = get_input_num_elems_processed<stridex>(num_elems_written_per_iteration);
const int kernel_depth = weights->info()->dimension(Window::DimZ);
const unsigned int conv_stride_y = std::get<1>(conv_info.stride());
- const unsigned int conv_pad_x = std::get<0>(conv_info.pad());
- const unsigned int conv_pad_y = std::get<1>(conv_info.pad());
+ const unsigned int conv_pad_left = conv_info.pad_left();
+ const unsigned int conv_pad_top = conv_info.pad_top();
const int fixed_point_position = input->info()->fixed_point_position();
// setup output window for the iterator
@@ -707,7 +713,7 @@
execute_window_loop(window_out, [&](const Coordinates & id)
{
- const uint8_t *input_ptr = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y;
+ const uint8_t *input_ptr = in.ptr() - conv_pad_left * input_stride_x - conv_pad_top * input_stride_y;
uint8_t *out_ptr = out.ptr();
int ih = 0;
int oh = 0;
@@ -804,8 +810,8 @@
const int delta_input = get_input_num_elems_processed<stridex>(num_elems_written_per_iteration);
const int kernel_depth = weights->info()->dimension(Window::DimZ);
const unsigned int conv_stride_y = std::get<1>(conv_info.stride());
- const unsigned int conv_pad_x = std::get<0>(conv_info.pad());
- const unsigned int conv_pad_y = std::get<1>(conv_info.pad());
+ const unsigned int conv_pad_left = conv_info.pad_left();
+ const unsigned int conv_pad_top = conv_info.pad_top();
const int fixed_point_position = input->info()->fixed_point_position();
// setup output window for the iterator
@@ -831,7 +837,7 @@
execute_window_loop(window_out, [&](const Coordinates & id)
{
- const uint8_t *input_ptr = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y;
+ const uint8_t *input_ptr = in.ptr() - conv_pad_left * input_stride_x - conv_pad_top * input_stride_y;
uint8_t *out_ptr = out.ptr();
int ih = 0;
int oh = 0;
@@ -1016,13 +1022,6 @@
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
- ARM_COMPUTE_RETURN_ERROR_ON(!conv_info.padding_is_symmetric());
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) == 1 && (std::get<0>(conv_info.pad()) || std::get<1>(conv_info.pad())),
- "Pad > 0 not supported for 1x1 weights");
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) == 3 && (std::get<0>(conv_info.pad()) > 1 || std::get<1>(conv_info.pad()) > 1),
- "Pad > 1 not supported for 3x3 weights");
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) == 5 && (std::get<0>(conv_info.pad()) > 2 || std::get<1>(conv_info.pad()) > 2),
- "Pad > 2 not supported for 5x5 weights");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(std::get<0>(conv_info.stride()) > 3, "Strides larger than 3 not supported.");
ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) != input->dimension(2));
@@ -1204,7 +1203,7 @@
unsigned int num_weight_elems_read_per_row = 0;
unsigned int num_elems_read_per_iteration = 0;
unsigned int num_elems_written_per_iteration = 0;
- BorderSize border_size(conv_info.pad().first, conv_info.pad().second);
+ BorderSize border_size = {};
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, output, conv_info));
ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(),
weights->clone().get(),