From 1fbb8120aa4f5ea2e73393421611ffea803a05b4 Mon Sep 17 00:00:00 2001 From: Anthony Barbier Date: Mon, 19 Feb 2018 19:36:02 +0000 Subject: Revert "COMPMID-582: Add validation to channel_extract kernels." This reverts commit 9a0875951d43dda035f32d2e0728cf59d80cb4d3. Change-Id: I6af0bc64c656f91cf1e0357f8760defa08f2e78d Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/121190 Reviewed-by: Anthony Barbier Tested-by: Anthony Barbier --- src/core/CL/kernels/CLChannelExtractKernel.cpp | 76 ++------ src/core/NEON/kernels/NEChannelExtractKernel.cpp | 231 ++++++++++++++++------- src/runtime/CL/CLMultiImage.cpp | 14 +- src/runtime/MultiImage.cpp | 14 +- 4 files changed, 187 insertions(+), 148 deletions(-) (limited to 'src') diff --git a/src/core/CL/kernels/CLChannelExtractKernel.cpp b/src/core/CL/kernels/CLChannelExtractKernel.cpp index 65843b8d5d..d0c633555d 100644 --- a/src/core/CL/kernels/CLChannelExtractKernel.cpp +++ b/src/core/CL/kernels/CLChannelExtractKernel.cpp @@ -49,48 +49,25 @@ CLChannelExtractKernel::CLChannelExtractKernel() void CLChannelExtractKernel::configure(const ICLTensor *input, Channel channel, ICLTensor *output) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_ON(input == output); - - set_format_if_unknown(*output->info(), Format::U8); - - // Check if input tensor has a valid format ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(input, Format::RGB888, Format::RGBA8888, Format::YUYV422, Format::UYVY422); ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output, Format::U8); - ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(output); - - // Check if channel is valid for given format - const Format format = input->info()->format(); - ARM_COMPUTE_ERROR_ON_CHANNEL_NOT_IN_KNOWN_FORMAT(format, channel); - - // Half the processed elements for U,V channels due to sub-sampling of 2 - _subsampling = 1; - - if(format == Format::YUYV422 || format == Format::UYVY422) - { - // Check if the width of the tensor shape is even for formats with subsampled channels (UYVY422 and YUYV422) - ARM_COMPUTE_ERROR_ON_TENSORS_NOT_EVEN(format, input); - - if(channel != Channel::Y) - { - _subsampling = 2; - } - } - - // Calculate output tensor shape using subsampling - TensorShape output_shape = calculate_subsampled_shape(input->info()->tensor_shape(), format, channel); - set_shape_if_empty(*output->info(), output_shape); - - ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); + ARM_COMPUTE_ERROR_ON(static_cast(input) == static_cast(output)); _input = input; _output = output; + // Check format + const Format format = input->info()->format(); + ARM_COMPUTE_ERROR_ON_CHANNEL_NOT_IN_KNOWN_FORMAT(format, channel); + // Create kernel std::string kernel_name = "channel_extract_" + string_from_format(format); std::set build_opts = { ("-DCHANNEL_" + string_from_channel(channel)) }; _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); + // Half the processed elements for U,V channels due to sub-sampling of 2 + _subsampling = ((Format::YUYV422 == format || Format::UYVY422 == format) && Channel::Y != channel) ? 2 : 1; + // Configure window Window win = calculate_max_window(*input->info(), Steps(_num_elems_processed_per_iteration)); AccessWindowHorizontal input_access(input->info(), 0, _num_elems_processed_per_iteration); @@ -106,34 +83,17 @@ void CLChannelExtractKernel::configure(const ICLTensor *input, Channel channel, void CLChannelExtractKernel::configure(const ICLMultiImage *input, Channel channel, ICLImage *output) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(output); - - set_format_if_unknown(*output->info(), Format::U8); - - // Check if channel is valid for given format - const Format format = input->info()->format(); - ARM_COMPUTE_ERROR_ON_CHANNEL_NOT_IN_KNOWN_FORMAT(format, channel); - - // Get input plane from the given channel - const ICLImage *input_plane = input->cl_plane(plane_idx_from_channel(format, channel)); - ARM_COMPUTE_ERROR_ON_NULLPTR(input_plane); - - if(Channel::Y == channel && format != Format::YUV444) - { - // Check if the width of the tensor shape is even for formats with subsampled channels (UYVY422 and YUYV422) - ARM_COMPUTE_ERROR_ON_TENSORS_NOT_EVEN(format, input_plane); - } - - // Calculate 2x2 subsampled tensor shape - TensorShape output_shape = calculate_subsampled_shape(input->cl_plane(0)->info()->tensor_shape(), format, channel); - set_shape_if_empty(*output->info(), output_shape); - - ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output_shape, output->info()->tensor_shape()); - - // Check if input tensor has a valid format ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(input, Format::NV12, Format::NV21, Format::IYUV, Format::YUV444); ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output, Format::U8); + ARM_COMPUTE_ERROR_ON(static_cast(input) == static_cast(output)); + + // Get format + const Format fmt = input->info()->format(); + + // Get input plane + const ICLImage *input_plane = input->cl_plane(plane_idx_from_channel(fmt, channel)); + ARM_COMPUTE_ERROR_ON(nullptr == input_plane); _output = output; _input = input_plane; @@ -142,13 +102,13 @@ void CLChannelExtractKernel::configure(const ICLMultiImage *input, Channel chann // Create kernel std::string kernel_name; std::set build_opts; - if(Channel::Y == channel || Format::IYUV == format || Format::YUV444 == format) + if(Channel::Y == channel || Format::IYUV == fmt || Format::YUV444 == fmt) { kernel_name = "copy_plane"; } else { - kernel_name = "channel_extract_" + string_from_format(format); + kernel_name = "channel_extract_" + string_from_format(fmt); build_opts.insert(("-DCHANNEL_" + string_from_channel(channel))); } _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); diff --git a/src/core/NEON/kernels/NEChannelExtractKernel.cpp b/src/core/NEON/kernels/NEChannelExtractKernel.cpp index 98b2f280d4..e8375d1e6d 100644 --- a/src/core/NEON/kernels/NEChannelExtractKernel.cpp +++ b/src/core/NEON/kernels/NEChannelExtractKernel.cpp @@ -56,68 +56,106 @@ void NEChannelExtractKernel::configure(const ITensor *input, Channel channel, IT set_format_if_unknown(*output->info(), Format::U8); - // Check if input tensor has a valid format ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(input, Format::RGB888, Format::RGBA8888, Format::UYVY422, Format::YUYV422); ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output, Format::U8); - ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input); - ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(output); - - // Check if channel is valid for given format - const Format format = input->info()->format(); - ARM_COMPUTE_ERROR_ON_CHANNEL_NOT_IN_KNOWN_FORMAT(format, channel); + unsigned int num_elems_processed_per_iteration = 8; - unsigned int subsampling = 1; + // Check format and channel + const Format format = input->info()->format(); + const unsigned int subsampling = (format == Format::YUYV422 || format == Format::UYVY422) && channel != Channel::Y ? 2 : 1; + TensorShape output_shape; - if(format == Format::YUYV422 || format == Format::UYVY422) + switch(format) { - // Check if the width of the tensor shape is even for formats with subsampled channels (UYVY422 and YUYV422) - ARM_COMPUTE_ERROR_ON_TENSORS_NOT_EVEN(format, input); - - if(channel != Channel::Y) - { - subsampling = 2; - } + case Format::RGB888: + case Format::RGBA8888: + num_elems_processed_per_iteration = 16; + output_shape = input->info()->tensor_shape(); + + if(format == Format::RGB888) + { + _func = &NEChannelExtractKernel::extract_1C_from_3C_img; + } + else if(format == Format::RGBA8888) + { + _func = &NEChannelExtractKernel::extract_1C_from_4C_img; + } + + switch(channel) + { + case Channel::R: + _lut_index = 0; + break; + case Channel::G: + _lut_index = 1; + break; + case Channel::B: + _lut_index = 2; + break; + case Channel::A: + if(format == Format::RGBA8888) + { + _lut_index = 3; + _func = &NEChannelExtractKernel::extract_1C_from_4C_img; + break; + } + default: + ARM_COMPUTE_ERROR("Not supported channel for this format."); + break; + } + break; + case Format::YUYV422: + case Format::UYVY422: + output_shape = input->info()->tensor_shape(); + + if(channel != Channel::Y) + { + output_shape.set(0, output_shape[0] / 2); + } + + switch(channel) + { + case Channel::Y: + num_elems_processed_per_iteration = 16; + _func = &NEChannelExtractKernel::extract_1C_from_2C_img; + _lut_index = (Format::YUYV422 == format) ? 0 : 1; + break; + case Channel::U: + num_elems_processed_per_iteration = 32; + _func = &NEChannelExtractKernel::extract_YUYV_uv; + _lut_index = (Format::YUYV422 == format) ? 1 : 0; + break; + case Channel::V: + num_elems_processed_per_iteration = 32; + _func = &NEChannelExtractKernel::extract_YUYV_uv; + _lut_index = (Format::YUYV422 == format) ? 3 : 2; + break; + default: + ARM_COMPUTE_ERROR("Not supported channel for this format."); + break; + } + break; + default: + ARM_COMPUTE_ERROR("Not supported format."); + break; } - TensorShape output_shape = calculate_subsampled_shape(input->info()->tensor_shape(), format, channel); set_shape_if_empty(*output->info(), output_shape); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output_shape, output->info()->tensor_shape()); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); - _input = input; - _output = output; - _lut_index = channel_idx_from_format(format, channel); - - unsigned int num_elems_processed_per_iteration = 16; - - if(format == Format::YUYV422 || format == Format::UYVY422) - { - _func = &NEChannelExtractKernel::extract_1C_from_2C_img; - - if(channel != Channel::Y) // Channel::U or Channel::V - { - num_elems_processed_per_iteration = 32; - _func = &NEChannelExtractKernel::extract_YUYV_uv; - } - } - else // Format::RGB888 or Format::RGBA8888 - { - _func = &NEChannelExtractKernel::extract_1C_from_3C_img; - - if(format == Format::RGBA8888) - { - _func = &NEChannelExtractKernel::extract_1C_from_4C_img; - } - } - - Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); + _input = input; + _output = output; + Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration, 1, 1.f / subsampling, 1.f / subsampling); + update_window_and_padding(win, input_access, output_access); ValidRegion input_valid_region = input->info()->valid_region(); + output_access.set_valid_region(win, ValidRegion(input_valid_region.anchor, output->info()->tensor_shape())); INEKernel::configure(win); @@ -130,45 +168,94 @@ void NEChannelExtractKernel::configure(const IMultiImage *input, Channel channel set_format_if_unknown(*output->info(), Format::U8); - const Format format = input->info()->format(); - ARM_COMPUTE_ERROR_ON_CHANNEL_NOT_IN_KNOWN_FORMAT(format, channel); - - // Get input plane - const IImage *input_plane = input->plane(plane_idx_from_channel(format, channel)); - ARM_COMPUTE_ERROR_ON_NULLPTR(input_plane); - - if(Channel::Y == channel && format != Format::YUV444) + switch(input->info()->format()) { - // Check if the width of the tensor shape is even for formats with subsampled channels (UYVY422 and YUYV422) - ARM_COMPUTE_ERROR_ON_TENSORS_NOT_EVEN(format, input_plane); + case Format::NV12: + case Format::NV21: + case Format::IYUV: + switch(channel) + { + case Channel::Y: + set_shape_if_empty(*output->info(), input->plane(0)->info()->tensor_shape()); + ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input->plane(0), output); + break; + case Channel::U: + case Channel::V: + set_shape_if_empty(*output->info(), input->plane(1)->info()->tensor_shape()); + ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input->plane(1), output); + break; + default: + ARM_COMPUTE_ERROR("Unsupported channel for selected format"); + } + break; + case Format::YUV444: + set_shape_if_empty(*output->info(), input->plane(0)->info()->tensor_shape()); + ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input->plane(0), output); + break; + default: + ARM_COMPUTE_ERROR("Unsupported format"); } - // Calculate 2x2 subsampled tensor shape - TensorShape output_shape = calculate_subsampled_shape(input->plane(0)->info()->tensor_shape(), format, channel); - set_shape_if_empty(*output->info(), output_shape); - - ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output_shape, output->info()->tensor_shape()); - - // Check if input tensor has a valid format ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(input, Format::NV12, Format::NV21, Format::IYUV, Format::YUV444); ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output, Format::U8); - _input = input_plane; - _output = output; - _lut_index = channel_idx_from_format(format, channel); - unsigned int num_elems_processed_per_iteration = 32; - _func = &NEChannelExtractKernel::copy_plane; + const Format &format = input->info()->format(); - if((format == Format::NV12 || format == Format::NV21) && channel != Channel::Y) + switch(format) { - num_elems_processed_per_iteration = 16; - _func = &NEChannelExtractKernel::extract_1C_from_2C_img; + case Format::NV12: + case Format::NV21: + switch(channel) + { + case Channel::Y: + _input = input->plane(0); + _func = &NEChannelExtractKernel::copy_plane; + break; + case Channel::U: + _input = input->plane(1); + num_elems_processed_per_iteration = 16; + _func = &NEChannelExtractKernel::extract_1C_from_2C_img; + _lut_index = (Format::NV12 == format) ? 0 : 1; + break; + case Channel::V: + _input = input->plane(1); + num_elems_processed_per_iteration = 16; + _func = &NEChannelExtractKernel::extract_1C_from_2C_img; + _lut_index = (Format::NV12 == format) ? 1 : 0; + break; + default: + ARM_COMPUTE_ERROR("Not supported channel for this format."); + break; + } + break; + case Format::IYUV: + case Format::YUV444: + _func = &NEChannelExtractKernel::copy_plane; + switch(channel) + { + case Channel::Y: + _input = input->plane(0); + break; + case Channel::U: + _input = input->plane(1); + break; + case Channel::V: + _input = input->plane(2); + break; + default: + ARM_COMPUTE_ERROR("Not supported channel for this format."); + break; + } + break; + default: + ARM_COMPUTE_ERROR("Not supported format."); + break; } - Window win = calculate_max_window(*_input->info(), Steps(num_elems_processed_per_iteration)); - + _output = output; + Window win = calculate_max_window(*_input->info(), Steps(num_elems_processed_per_iteration)); AccessWindowHorizontal input_access(_input->info(), 0, num_elems_processed_per_iteration); AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); update_window_and_padding(win, input_access, output_access); diff --git a/src/runtime/CL/CLMultiImage.cpp b/src/runtime/CL/CLMultiImage.cpp index 92254f3531..aa09c7a67b 100644 --- a/src/runtime/CL/CLMultiImage.cpp +++ b/src/runtime/CL/CLMultiImage.cpp @@ -25,7 +25,6 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" #include "arm_compute/runtime/ITensorAllocator.h" using namespace arm_compute; @@ -52,8 +51,7 @@ void CLMultiImage::init_auto_padding(unsigned int width, unsigned int height, Fo void CLMultiImage::internal_init(unsigned int width, unsigned int height, Format format, bool auto_padding) { - TensorShape shape = adjust_odd_shape(TensorShape{ width, height }, format); - TensorInfo info(shape, Format::U8); + TensorInfo info(width, height, Format::U8); if(auto_padding) { @@ -74,7 +72,7 @@ void CLMultiImage::internal_init(unsigned int width, unsigned int height, Format case Format::YUYV422: case Format::UYVY422: { - TensorInfo info_full(shape, format); + TensorInfo info_full(width, height, format); if(auto_padding) { @@ -87,8 +85,7 @@ void CLMultiImage::internal_init(unsigned int width, unsigned int height, Format case Format::NV12: case Format::NV21: { - const TensorShape shape_uv88 = calculate_subsampled_shape(shape, Format::UV88); - TensorInfo info_uv88(shape_uv88, Format::UV88); + TensorInfo info_uv88(width / 2, height / 2, Format::UV88); if(auto_padding) { @@ -101,8 +98,7 @@ void CLMultiImage::internal_init(unsigned int width, unsigned int height, Format } case Format::IYUV: { - const TensorShape shape_sub2 = calculate_subsampled_shape(shape, Format::IYUV); - TensorInfo info_sub2(shape_sub2, Format::U8); + TensorInfo info_sub2(width / 2, height / 2, Format::U8); if(auto_padding) { @@ -124,7 +120,7 @@ void CLMultiImage::internal_init(unsigned int width, unsigned int height, Format break; } - _info.init(shape.x(), shape.y(), format); + _info.init(width, height, format); } void CLMultiImage::allocate() diff --git a/src/runtime/MultiImage.cpp b/src/runtime/MultiImage.cpp index 6eba71bf1d..bdafe4c9a2 100644 --- a/src/runtime/MultiImage.cpp +++ b/src/runtime/MultiImage.cpp @@ -25,7 +25,6 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" #include "arm_compute/runtime/TensorAllocator.h" using namespace arm_compute; @@ -52,8 +51,7 @@ void MultiImage::init_auto_padding(unsigned int width, unsigned int height, Form void MultiImage::internal_init(unsigned int width, unsigned int height, Format format, bool auto_padding) { - TensorShape shape = adjust_odd_shape(TensorShape{ width, height }, format); - TensorInfo info(shape, Format::U8); + TensorInfo info(width, height, Format::U8); if(auto_padding) { @@ -74,7 +72,7 @@ void MultiImage::internal_init(unsigned int width, unsigned int height, Format f case Format::YUYV422: case Format::UYVY422: { - TensorInfo info_full(shape, format); + TensorInfo info_full(width, height, format); if(auto_padding) { @@ -87,8 +85,7 @@ void MultiImage::internal_init(unsigned int width, unsigned int height, Format f case Format::NV12: case Format::NV21: { - const TensorShape shape_uv88 = calculate_subsampled_shape(shape, Format::UV88); - TensorInfo info_uv88(shape_uv88, Format::UV88); + TensorInfo info_uv88(width / 2, height / 2, Format::UV88); if(auto_padding) { @@ -101,8 +98,7 @@ void MultiImage::internal_init(unsigned int width, unsigned int height, Format f } case Format::IYUV: { - const TensorShape shape_sub2 = calculate_subsampled_shape(shape, Format::IYUV); - TensorInfo info_sub2(shape_sub2, Format::U8); + TensorInfo info_sub2(width / 2, height / 2, Format::U8); if(auto_padding) { @@ -124,7 +120,7 @@ void MultiImage::internal_init(unsigned int width, unsigned int height, Format f break; } - _info.init(shape.x(), shape.y(), format); + _info.init(width, height, format); } void MultiImage::allocate() -- cgit v1.2.1