From a2611815f278334c801094d095901d36e111c3f9 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Fri, 21 Jul 2017 10:08:48 +0100 Subject: COMPMID-417 NEON/CL MeanStdDev bugfix using FillBorderKernel Change-Id: Ic48ba7f69783d0e1e80611264e2bc67d1732436e Reviewed-on: http://mpd-gerrit.cambridge.arm.com/81293 Reviewed-by: Anthony Barbier Tested-by: Kaizen --- src/core/CL/kernels/CLMeanStdDevKernel.cpp | 10 +++++++++- src/core/NEON/kernels/NEMeanStdDevKernel.cpp | 10 +++++++++- src/core/NEON/kernels/NESoftmaxLayerKernel.cpp | 2 +- src/runtime/CL/functions/CLMeanStdDev.cpp | 6 ++++-- src/runtime/NEON/functions/NEMeanStdDev.cpp | 7 ++++--- 5 files changed, 27 insertions(+), 8 deletions(-) (limited to 'src') diff --git a/src/core/CL/kernels/CLMeanStdDevKernel.cpp b/src/core/CL/kernels/CLMeanStdDevKernel.cpp index b0b748f466..1b70d7513f 100644 --- a/src/core/CL/kernels/CLMeanStdDevKernel.cpp +++ b/src/core/CL/kernels/CLMeanStdDevKernel.cpp @@ -40,10 +40,15 @@ using namespace arm_compute; CLMeanStdDevKernel::CLMeanStdDevKernel() - : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr) + : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr), _border_size(0) { } +BorderSize CLMeanStdDevKernel::border_size() const +{ + return _border_size; +} + void CLMeanStdDevKernel::configure(const ICLImage *input, float *mean, cl::Buffer *global_sum, float *stddev, cl::Buffer *global_sum_squared) { ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input); @@ -83,6 +88,9 @@ void CLMeanStdDevKernel::configure(const ICLImage *input, float *mean, cl::Buffe constexpr unsigned int num_elems_processed_per_iteration_x = 8; const unsigned int num_elems_processed_per_iteration_y = input->info()->dimension(1); + _border_size = BorderSize(std::max(static_cast(num_elems_processed_per_iteration_x) - static_cast(input->info()->dimension(0)), + static_cast(input->info()->dimension(0) % num_elems_processed_per_iteration_x))); + Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); AccessWindowRectangle input_access(input->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); update_window_and_padding(win, input_access); diff --git a/src/core/NEON/kernels/NEMeanStdDevKernel.cpp b/src/core/NEON/kernels/NEMeanStdDevKernel.cpp index 4616203d66..1eb7e45e36 100644 --- a/src/core/NEON/kernels/NEMeanStdDevKernel.cpp +++ b/src/core/NEON/kernels/NEMeanStdDevKernel.cpp @@ -85,10 +85,15 @@ std::pair accumulate(const Window &window, Iterator &ite } // namespace NEMeanStdDevKernel::NEMeanStdDevKernel() - : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr), _mtx() + : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr), _mtx(), _border_size(0) { } +BorderSize NEMeanStdDevKernel::border_size() const +{ + return _border_size; +} + void NEMeanStdDevKernel::configure(const IImage *input, float *mean, uint64_t *global_sum, float *stddev, uint64_t *global_sum_squared) { ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input); @@ -105,6 +110,9 @@ void NEMeanStdDevKernel::configure(const IImage *input, float *mean, uint64_t *g constexpr unsigned int num_elems_processed_per_iteration = 16; + _border_size = BorderSize(std::max(static_cast(num_elems_processed_per_iteration) - static_cast(input->info()->dimension(0)), + static_cast(input->info()->dimension(0) % num_elems_processed_per_iteration))); + // Configure kernel window Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); diff --git a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp index 79fcba1dfb..176e3d688e 100644 --- a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp +++ b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp @@ -224,7 +224,7 @@ void NELogits1DMaxKernel::configure(const ITensor *input, ITensor *output) _input = input; _output = output; - _border_size = BorderSize(0, input_width % num_elems_processed_per_iteration, 0, 0); + _border_size = BorderSize(0, num_elems_processed_per_iteration - (input_width % num_elems_processed_per_iteration), 0, 0); // Configure kernel window constexpr unsigned int num_elems_written_per_row = 1; diff --git a/src/runtime/CL/functions/CLMeanStdDev.cpp b/src/runtime/CL/functions/CLMeanStdDev.cpp index 56ba146790..838f7e73d2 100644 --- a/src/runtime/CL/functions/CLMeanStdDev.cpp +++ b/src/runtime/CL/functions/CLMeanStdDev.cpp @@ -23,19 +23,19 @@ */ #include "arm_compute/runtime/CL/functions/CLMeanStdDev.h" -#include "arm_compute/core/CL/kernels/CLMeanStdDevKernel.h" #include "arm_compute/runtime/CL/CLScheduler.h" using namespace arm_compute; CLMeanStdDev::CLMeanStdDev() : _mean_stddev_kernel(), + _fill_border_kernel(), _global_sum(), _global_sum_squared() { } -void CLMeanStdDev::configure(const ICLImage *input, float *mean, float *stddev) +void CLMeanStdDev::configure(ICLImage *input, float *mean, float *stddev) { _global_sum = cl::Buffer(CLScheduler::get().context(), CL_MEM_ALLOC_HOST_PTR | CL_MEM_READ_WRITE, sizeof(cl_ulong)); @@ -45,9 +45,11 @@ void CLMeanStdDev::configure(const ICLImage *input, float *mean, float *stddev) } _mean_stddev_kernel.configure(input, mean, &_global_sum, stddev, &_global_sum_squared); + _fill_border_kernel.configure(input, _mean_stddev_kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast(0))); } void CLMeanStdDev::run() { + CLScheduler::get().enqueue(_fill_border_kernel); CLScheduler::get().enqueue(_mean_stddev_kernel); } diff --git a/src/runtime/NEON/functions/NEMeanStdDev.cpp b/src/runtime/NEON/functions/NEMeanStdDev.cpp index 47143f5e5b..ab8e72bf1d 100644 --- a/src/runtime/NEON/functions/NEMeanStdDev.cpp +++ b/src/runtime/NEON/functions/NEMeanStdDev.cpp @@ -23,19 +23,19 @@ */ #include "arm_compute/runtime/NEON/functions/NEMeanStdDev.h" -#include "arm_compute/core/NEON/kernels/NEMeanStdDevKernel.h" #include "arm_compute/runtime/NEON/NEScheduler.h" using namespace arm_compute; NEMeanStdDev::NEMeanStdDev() - : _mean_stddev_kernel(), _global_sum(0), _global_sum_squared(0) + : _mean_stddev_kernel(), _fill_border_kernel(), _global_sum(0), _global_sum_squared(0) { } -void NEMeanStdDev::configure(const IImage *input, float *mean, float *stddev) +void NEMeanStdDev::configure(IImage *input, float *mean, float *stddev) { _mean_stddev_kernel.configure(input, mean, &_global_sum, stddev, &_global_sum_squared); + _fill_border_kernel.configure(input, _mean_stddev_kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast(0))); } void NEMeanStdDev::run() @@ -43,5 +43,6 @@ void NEMeanStdDev::run() _global_sum = 0; _global_sum_squared = 0; + _fill_border_kernel.run(_fill_border_kernel.window()); NEScheduler::get().schedule(&_mean_stddev_kernel, Window::DimY); } -- cgit v1.2.1