From 68dd25fbe6e4d3c3513fa5993863419769aa08fc Mon Sep 17 00:00:00 2001 From: Sang-Hoon Park Date: Mon, 19 Oct 2020 16:00:11 +0100 Subject: COMPMID-3637: Move utility headers from arm_compute to src Signed-off-by: Georgios Pinitas Change-Id: If9d6fa8c900b68c4b6fd373f2fc1f9abb83ea917 Signed-off-by: Michalis Spyrou Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4145 Tested-by: Arm Jenkins Reviewed-by: Sang-Hoon Park Comments-Addressed: Arm Jenkins --- arm_compute/core/Helpers.h | 615 +-------------------------------------------- 1 file changed, 2 insertions(+), 613 deletions(-) (limited to 'arm_compute/core/Helpers.h') diff --git a/arm_compute/core/Helpers.h b/arm_compute/core/Helpers.h index 90dd6082e1..5a8d6efe9d 100644 --- a/arm_compute/core/Helpers.h +++ b/arm_compute/core/Helpers.h @@ -24,23 +24,17 @@ #ifndef ARM_COMPUTE_HELPERS_H #define ARM_COMPUTE_HELPERS_H -#include "arm_compute/core/Coordinates.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/IAccessWindow.h" -#include "arm_compute/core/Steps.h" -#include "arm_compute/core/Strides.h" -#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/ITensor.h" #include "arm_compute/core/Types.h" +#include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" -#include "support/MemorySupport.h" #include #include #include -#include #include -#include -#include namespace arm_compute { @@ -48,307 +42,6 @@ class IKernel; class ITensor; class ITensorInfo; -/** Disable bitwise operations by default */ -template -struct enable_bitwise_ops -{ - static constexpr bool value = false; /**< Disabled */ -}; - -#ifndef DOXYGEN_SKIP_THIS -template -typename std::enable_if::value, T>::type operator&(T lhs, T rhs) -{ - using underlying_type = typename std::underlying_type::type; - return static_cast(static_cast(lhs) & static_cast(rhs)); -} -#endif /* DOXYGEN_SKIP_THIS */ - -/** Helper function to create and return a unique_ptr pointed to a CL/GLES kernel object - * It also calls the kernel's configuration. - * - * @param[in] args All the arguments that need pass to kernel's configuration. - * - * @return A unique pointer pointed to a CL/GLES kernel object - */ -template -std::unique_ptr create_configure_kernel(T &&... args) -{ - std::unique_ptr k = arm_compute::support::cpp14::make_unique(); - k->configure(std::forward(args)...); - return k; -} - -/** Helper function to create and return a unique_ptr pointed to a CL/GLES kernel object - * - * @return A unique pointer pointed to a Kernel kernel object - */ -template -std::unique_ptr create_kernel() -{ - std::unique_ptr k = arm_compute::support::cpp14::make_unique(); - return k; -} - -namespace traits -{ -/** Check if a type T is contained in a tuple Tuple of types */ -template -struct is_contained; - -template -struct is_contained> : std::false_type -{ -}; - -template -struct is_contained> : std::true_type -{ -}; - -template -struct is_contained> : is_contained> -{ -}; -} - -/** Computes bilinear interpolation using the pointer to the top-left pixel and the pixel's distance between - * the real coordinates and the smallest following integer coordinates. Input must be in single channel format. - * - * @param[in] pixel_ptr Pointer to the top-left pixel value of a single channel input. - * @param[in] stride Stride to access the bottom-left and bottom-right pixel values - * @param[in] dx Pixel's distance between the X real coordinate and the smallest X following integer - * @param[in] dy Pixel's distance between the Y real coordinate and the smallest Y following integer - * - * @note dx and dy must be in the range [0, 1.0] - * - * @return The bilinear interpolated pixel value - */ -template -inline T delta_bilinear_c1(const T *pixel_ptr, size_t stride, float dx, float dy) -{ - ARM_COMPUTE_ERROR_ON(pixel_ptr == nullptr); - - const float dx1 = 1.0f - dx; - const float dy1 = 1.0f - dy; - - const T a00 = *pixel_ptr; - const T a01 = *(pixel_ptr + 1); - const T a10 = *(pixel_ptr + stride); - const T a11 = *(pixel_ptr + stride + 1); - - const float w1 = dx1 * dy1; - const float w2 = dx * dy1; - const float w3 = dx1 * dy; - const float w4 = dx * dy; - - return static_cast(a00 * w1 + a01 * w2 + a10 * w3 + a11 * w4); -} - -/** Computes bilinear interpolation for quantized input and output, using the pointer to the top-left pixel and the pixel's distance between - * the real coordinates and the smallest following integer coordinates. Input must be QASYMM8 and in single channel format. - * - * @param[in] pixel_ptr Pointer to the top-left pixel value of a single channel input. - * @param[in] stride Stride to access the bottom-left and bottom-right pixel values - * @param[in] dx Pixel's distance between the X real coordinate and the smallest X following integer - * @param[in] dy Pixel's distance between the Y real coordinate and the smallest Y following integer - * @param[in] iq_info Input QuantizationInfo - * @param[in] oq_info Output QuantizationInfo - * - * @note dx and dy must be in the range [0, 1.0] - * - * @return The bilinear interpolated pixel value - */ -inline uint8_t delta_bilinear_c1_quantized(const uint8_t *pixel_ptr, size_t stride, float dx, float dy, UniformQuantizationInfo iq_info, UniformQuantizationInfo oq_info) -{ - ARM_COMPUTE_ERROR_ON(pixel_ptr == nullptr); - - const float dx1 = 1.0f - dx; - const float dy1 = 1.0f - dy; - - const float a00 = dequantize_qasymm8(*pixel_ptr, iq_info); - const float a01 = dequantize_qasymm8(*(pixel_ptr + 1), iq_info); - const float a10 = dequantize_qasymm8(*(pixel_ptr + stride), iq_info); - const float a11 = dequantize_qasymm8(*(pixel_ptr + stride + 1), iq_info); - - const float w1 = dx1 * dy1; - const float w2 = dx * dy1; - const float w3 = dx1 * dy; - const float w4 = dx * dy; - float res = a00 * w1 + a01 * w2 + a10 * w3 + a11 * w4; - return static_cast(quantize_qasymm8(res, oq_info)); -} - -/** Computes bilinear interpolation for quantized input and output, using the pointer to the top-left pixel and the pixel's distance between - * the real coordinates and the smallest following integer coordinates. Input must be QASYMM8_SIGNED and in single channel format. - * - * @param[in] pixel_ptr Pointer to the top-left pixel value of a single channel input. - * @param[in] stride Stride to access the bottom-left and bottom-right pixel values - * @param[in] dx Pixel's distance between the X real coordinate and the smallest X following integer - * @param[in] dy Pixel's distance between the Y real coordinate and the smallest Y following integer - * @param[in] iq_info Input QuantizationInfo - * @param[in] oq_info Output QuantizationInfo - * - * @note dx and dy must be in the range [0, 1.0] - * - * @return The bilinear interpolated pixel value - */ -inline int8_t delta_bilinear_c1_quantized(const int8_t *pixel_ptr, size_t stride, float dx, float dy, UniformQuantizationInfo iq_info, UniformQuantizationInfo oq_info) -{ - ARM_COMPUTE_ERROR_ON(pixel_ptr == nullptr); - - const float dx1 = 1.0f - dx; - const float dy1 = 1.0f - dy; - - const float a00 = dequantize_qasymm8_signed(*pixel_ptr, iq_info); - const float a01 = dequantize_qasymm8_signed(*(pixel_ptr + 1), iq_info); - const float a10 = dequantize_qasymm8_signed(*(pixel_ptr + stride), iq_info); - const float a11 = dequantize_qasymm8_signed(*(pixel_ptr + stride + 1), iq_info); - - const float w1 = dx1 * dy1; - const float w2 = dx * dy1; - const float w3 = dx1 * dy; - const float w4 = dx * dy; - float res = a00 * w1 + a01 * w2 + a10 * w3 + a11 * w4; - return static_cast(quantize_qasymm8_signed(res, oq_info)); -} - -/** Computes linear interpolation using the pointer to the top pixel and the pixel's distance between - * the real coordinates and the smallest following integer coordinates. Input must be in single channel format. - * - * @param[in] pixel_ptr Pointer to the top pixel value of a single channel input. - * @param[in] stride Stride to access the bottom pixel value - * @param[in] dy Pixel's distance between the Y real coordinate and the smallest Y following integer - * - * @note dy must be in the range [0, 1.0] - * - * @return The linear interpolated pixel value - */ -template -inline T delta_linear_c1_y(const T *pixel_ptr, size_t stride, float dy) -{ - ARM_COMPUTE_ERROR_ON(pixel_ptr == nullptr); - - const float dy1 = 1.0f - dy; - - const T a00 = *pixel_ptr; - const T a10 = *(pixel_ptr + stride); - - const float w1 = dy1; - const float w3 = dy; - - return static_cast(a00 * w1 + a10 * w3); -} -/** Computes linear interpolation using the pointer to the left pixel and the pixel's distance between - * the real coordinates and the smallest following integer coordinates. Input must be in single channel format. - * - * @param[in] pixel_ptr Pointer to the left pixel value of a single channel input. - * @param[in] dx Pixel's distance between the X real coordinate and the smallest X following integer - * - * @note dx must be in the range [0, 1.0] - * - * @return The linear interpolated pixel value - */ -template -inline T delta_linear_c1_x(const T *pixel_ptr, float dx) -{ - ARM_COMPUTE_ERROR_ON(pixel_ptr == nullptr); - - const T a00 = *pixel_ptr; - const T a01 = *(pixel_ptr + 1); - - const float dx1 = 1.0f - dx; - - const float w1 = dx1; - const float w2 = dx; - - return static_cast(a00 * w1 + a01 * w2); -} -/** Return the pixel at (x,y) using bilinear interpolation. - * - * @warning Only works if the iterator was created with an IImage - * - * @param[in] first_pixel_ptr Pointer to the first pixel of a single channel input. - * @param[in] stride Stride in bytes of the image; - * @param[in] x X position of the wanted pixel - * @param[in] y Y position of the wanted pixel - * - * @return The pixel at (x, y) using bilinear interpolation. - */ -template -inline T pixel_bilinear_c1(const T *first_pixel_ptr, size_t stride, float x, float y) -{ - ARM_COMPUTE_ERROR_ON(first_pixel_ptr == nullptr); - - const int32_t xi = std::floor(x); - const int32_t yi = std::floor(y); - - const float dx = x - xi; - const float dy = y - yi; - - return delta_bilinear_c1(first_pixel_ptr + xi + yi * stride, stride, dx, dy); -} - -/** Return the pixel at (x,y) using bilinear interpolation by clamping when out of borders. The image must be single channel input - * - * @warning Only works if the iterator was created with an IImage - * - * @param[in] first_pixel_ptr Pointer to the first pixel of a single channel image. - * @param[in] stride Stride in bytes of the image - * @param[in] width Width of the image - * @param[in] height Height of the image - * @param[in] x X position of the wanted pixel - * @param[in] y Y position of the wanted pixel - * - * @return The pixel at (x, y) using bilinear interpolation. - */ -template -inline uint8_t pixel_bilinear_c1_clamp(const T *first_pixel_ptr, size_t stride, size_t width, size_t height, float x, float y) -{ - ARM_COMPUTE_ERROR_ON(first_pixel_ptr == nullptr); - - x = std::max(-1.f, std::min(x, static_cast(width))); - y = std::max(-1.f, std::min(y, static_cast(height))); - - const float xi = std::floor(x); - const float yi = std::floor(y); - - const float dx = x - xi; - const float dy = y - yi; - - if(dx == 0.0f) - { - if(dy == 0.0f) - { - return static_cast(first_pixel_ptr[static_cast(xi) + static_cast(yi) * stride]); - } - return delta_linear_c1_y(first_pixel_ptr + static_cast(xi) + static_cast(yi) * stride, stride, dy); - } - if(dy == 0.0f) - { - return delta_linear_c1_x(first_pixel_ptr + static_cast(xi) + static_cast(yi) * stride, dx); - } - return delta_bilinear_c1(first_pixel_ptr + static_cast(xi) + static_cast(yi) * stride, stride, dx, dy); -} - -/** Return the pixel at (x,y) using area interpolation by clamping when out of borders. The image must be single channel U8 - * - * @note The interpolation area depends on the width and height ration of the input and output images - * @note Currently average of the contributing pixels is calculated - * - * @param[in] first_pixel_ptr Pointer to the first pixel of a single channel U8 image. - * @param[in] stride Stride in bytes of the image - * @param[in] width Width of the image - * @param[in] height Height of the image - * @param[in] wr Width ratio among the input image width and output image width. - * @param[in] hr Height ratio among the input image height and output image height. - * @param[in] x X position of the wanted pixel - * @param[in] y Y position of the wanted pixel - * - * @return The pixel at (x, y) using area interpolation. - */ -inline uint8_t pixel_area_c1u8_clamp(const uint8_t *first_pixel_ptr, size_t stride, size_t width, size_t height, float wr, float hr, int x, int y); - /** Iterator updated by @ref execute_window_loop for each window element */ class Iterator { @@ -421,179 +114,6 @@ private: template inline void execute_window_loop(const Window &w, L &&lambda_function, Ts &&... iterators); -/** Update window and padding size for each of the access patterns. - * - * First the window size is reduced based on all access patterns that are not - * allowed to modify the padding of the underlying tensor. Then the padding of - * the remaining tensors is increased to match the window. - * - * @param[in] win Window that is used by the kernel. - * @param[in] patterns Access patterns used to calculate the final window and padding. - * - * @return True if the window has been changed. Changes to the padding do not - * influence the returned value. - */ -template -bool update_window_and_padding(Window &win, Ts &&... patterns) -{ - bool window_changed = false; - - utility::for_each([&](const IAccessWindow & w) - { - window_changed |= w.update_window_if_needed(win); - }, - patterns...); - - bool padding_changed = false; - - utility::for_each([&](IAccessWindow & w) - { - padding_changed |= w.update_padding_if_needed(win); - }, - patterns...); - - return window_changed; -} - -/** Calculate the maximum window for a given tensor shape and border setting - * - * @param[in] valid_region Valid region object defining the shape of the tensor space for which the window is created. - * @param[in] steps (Optional) Number of elements processed for each step. - * @param[in] skip_border (Optional) If true exclude the border region from the window. - * @param[in] border_size (Optional) Border size. - * - * @return The maximum window the kernel can be executed on. - */ -Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps = Steps(), bool skip_border = false, BorderSize border_size = BorderSize()); - -/** Calculate the maximum window for a given tensor shape and border setting - * - * @param[in] info Tensor info object defining the shape of the object for which the window is created. - * @param[in] steps (Optional) Number of elements processed for each step. - * @param[in] skip_border (Optional) If true exclude the border region from the window. - * @param[in] border_size (Optional) Border size. - * - * @return The maximum window the kernel can be executed on. - */ -inline Window calculate_max_window(const ITensorInfo &info, const Steps &steps = Steps(), bool skip_border = false, BorderSize border_size = BorderSize()) -{ - return calculate_max_window(info.valid_region(), steps, skip_border, border_size); -} - -/** Calculate the maximum window used by a horizontal kernel for a given tensor shape and border setting - * - * @param[in] valid_region Valid region object defining the shape of the tensor space for which the window is created. - * @param[in] steps (Optional) Number of elements processed for each step. - * @param[in] skip_border (Optional) If true exclude the border region from the window. - * @param[in] border_size (Optional) Border size. The border region will be excluded from the window. - * - * @return The maximum window the kernel can be executed on. - */ -Window calculate_max_window_horizontal(const ValidRegion &valid_region, const Steps &steps = Steps(), bool skip_border = false, BorderSize border_size = BorderSize()); - -/** Calculate the maximum window used by a horizontal kernel for a given tensor shape and border setting - * - * @param[in] info Tensor info object defining the shape of the object for which the window is created. - * @param[in] steps (Optional) Number of elements processed for each step. - * @param[in] skip_border (Optional) If true exclude the border region from the window. - * @param[in] border_size (Optional) Border size. - * - * @return The maximum window the kernel can be executed on. - */ -inline Window calculate_max_window_horizontal(const ITensorInfo &info, const Steps &steps = Steps(), bool skip_border = false, BorderSize border_size = BorderSize()) -{ - return calculate_max_window_horizontal(info.valid_region(), steps, skip_border, border_size); -} - -/** Calculate the maximum window for a given tensor shape and border setting. The window will also includes the border. - * - * @param[in] valid_region Valid region object defining the shape of the tensor space for which the window is created. - * @param[in] steps (Optional) Number of elements processed for each step. - * @param[in] border_size (Optional) Border size. The border region will be included in the window. - * - * @return The maximum window the kernel can be executed on. - */ -Window calculate_max_enlarged_window(const ValidRegion &valid_region, const Steps &steps = Steps(), BorderSize border_size = BorderSize()); - -/** Calculate the maximum window for a given tensor shape and border setting. The window will also includes the border. - * - * @param[in] info Tensor info object defining the shape of the object for which the window is created. - * @param[in] steps (Optional) Number of elements processed for each step. - * @param[in] border_size (Optional) Border size. The border region will be included in the window. - * - * @return The maximum window the kernel can be executed on. - */ -inline Window calculate_max_enlarged_window(const ITensorInfo &info, const Steps &steps = Steps(), BorderSize border_size = BorderSize()) -{ - return calculate_max_enlarged_window(info.valid_region(), steps, border_size); -} - -/** Intersect multiple valid regions. - * - * @param[in] regions Valid regions. - * - * @return Intersection of all regions. - */ -template -ValidRegion intersect_valid_regions(const Ts &... regions) -{ - auto intersect = [](const ValidRegion & r1, const ValidRegion & r2) -> ValidRegion - { - ValidRegion region; - - for(size_t d = 0; d < std::min(r1.anchor.num_dimensions(), r2.anchor.num_dimensions()); ++d) - { - region.anchor.set(d, std::max(r1.anchor[d], r2.anchor[d])); - } - - for(size_t d = 0; d < std::min(r1.shape.num_dimensions(), r2.shape.num_dimensions()); ++d) - { - region.shape.set(d, std::min(r1.shape[d], r2.shape[d])); - } - - return region; - }; - - return utility::foldl(intersect, regions...); -} - -/** Create a strides object based on the provided strides and the tensor dimensions. - * - * @param[in] info Tensor info object providing the shape of the tensor for unspecified strides. - * @param[in] stride_x Stride to be used in X dimension (in bytes). - * @param[in] fixed_strides Strides to be used in higher dimensions starting at Y (in bytes). - * - * @return Strides object based on the specified strides. Missing strides are - * calculated based on the tensor shape and the strides of lower dimensions. - */ -template -inline Strides compute_strides(const ITensorInfo &info, T stride_x, Ts &&... fixed_strides) -{ - const TensorShape &shape = info.tensor_shape(); - - // Create strides object - Strides strides(stride_x, fixed_strides...); - - for(size_t i = 1 + sizeof...(Ts); i < info.num_dimensions(); ++i) - { - strides.set(i, shape[i - 1] * strides[i - 1]); - } - - return strides; -} - -/** Create a strides object based on the tensor dimensions. - * - * @param[in] info Tensor info object used to compute the strides. - * - * @return Strides object based on element size and tensor shape. - */ -template -inline Strides compute_strides(const ITensorInfo &info) -{ - return compute_strides(info, info.element_size()); -} - /** Permutes given Dimensions according to a permutation vector * * @warning Validity of permutation is not checked @@ -629,79 +149,6 @@ inline void permute(TensorShape &shape, const PermutationVector &perm) } } -/** Auto initialize the tensor info (shape, number of channels and data type) if the current assignment is empty. - * - * @param[in,out] info Tensor info used to check and assign. - * @param[in] shape New shape. - * @param[in] num_channels New number of channels. - * @param[in] data_type New data type - * @param[in] quantization_info (Optional) New quantization info - * - * @return True if the tensor info has been initialized - */ -bool auto_init_if_empty(ITensorInfo &info, - const TensorShape &shape, - int num_channels, DataType data_type, - QuantizationInfo quantization_info = QuantizationInfo()); - -/** Auto initialize the tensor info using another tensor info. - * - * @param info_sink Tensor info used to check and assign - * @param info_source Tensor info used to assign - * - * @return True if the tensor info has been initialized - */ -bool auto_init_if_empty(ITensorInfo &info_sink, const ITensorInfo &info_source); - -/** Set the shape to the specified value if the current assignment is empty. - * - * @param[in,out] info Tensor info used to check and assign. - * @param[in] shape New shape. - * - * @return True if the shape has been changed. - */ -bool set_shape_if_empty(ITensorInfo &info, const TensorShape &shape); - -/** Set the format, data type and number of channels to the specified value if - * the current data type is unknown. - * - * @param[in,out] info Tensor info used to check and assign. - * @param[in] format New format. - * - * @return True if the format has been changed. - */ -bool set_format_if_unknown(ITensorInfo &info, Format format); - -/** Set the data type and number of channels to the specified value if - * the current data type is unknown. - * - * @param[in,out] info Tensor info used to check and assign. - * @param[in] data_type New data type. - * - * @return True if the data type has been changed. - */ -bool set_data_type_if_unknown(ITensorInfo &info, DataType data_type); - -/** Set the data layout to the specified value if - * the current data layout is unknown. - * - * @param[in,out] info Tensor info used to check and assign. - * @param[in] data_layout New data layout. - * - * @return True if the data type has been changed. - */ -bool set_data_layout_if_unknown(ITensorInfo &info, DataLayout data_layout); - -/** Set the quantization info to the specified value if - * the current quantization info is empty and the data type of asymmetric quantized type - * - * @param[in,out] info Tensor info used to check and assign. - * @param[in] quantization_info Quantization info - * - * @return True if the quantization info has been changed. - */ -bool set_quantization_info_if_empty(ITensorInfo &info, QuantizationInfo quantization_info); - /** Helper function to calculate the Valid Region for Scale. * * @param[in] src_info Input tensor info used to check. @@ -751,21 +198,6 @@ inline size_t get_data_layout_dimension_index(const DataLayout data_layout, cons */ inline DataLayoutDimension get_index_data_layout_dimension(const DataLayout data_layout, const size_t index); -/** Calculate the normalization dimension index for a given normalization type - * - * @param[in] layout Data layout of the input and output tensor - * @param[in] info Normalization info - * - * @return Normalization dimension index - */ -inline unsigned int get_normalization_dimension_index(DataLayout layout, const NormalizationLayerInfo &info) -{ - const unsigned int width_idx = get_data_layout_dimension_index(layout, DataLayoutDimension::WIDTH); - const unsigned int channel_idx = get_data_layout_dimension_index(layout, DataLayoutDimension::CHANNEL); - - return info.is_in_map() ? width_idx : channel_idx; -} - /** Calculate the number of output tiles required by Winograd Convolution layer. This utility function can be used by the Winograd input transform * to know the number of tiles on the x and y direction * @@ -814,49 +246,6 @@ inline Coordinates &convert_negative_axis(Coordinates &coords, int max_value) } return coords; } - -/** Given an integer value, this function returns the next power of two - * - * @param[in] x Input value - * - * @return the next power of two - */ -inline unsigned int get_next_power_two(unsigned int x) -{ - // Decrement by 1 - x--; - - // Shift right by 1 - x |= x >> 1u; - // Shift right by 2 - x |= x >> 2u; - // Shift right by 4 - x |= x >> 4u; - // Shift right by 8 - x |= x >> 8u; - // Shift right by 16 - x |= x >> 16u; - - // Increment by 1 - x++; - - return x; -} - -/** Given a softmax axis, this function returns the permutation vector required to put the axis to the front - * - * @note This function assumes a tensor rank <= 4 - * - * Axis selects the dimension on which softmax is performed. - * E.g. For input of shape 4x5x6 and axis=1, softmax will be applied to 4x6=24 vectors of size 5. - * Interally softmax kernels is always performed on the first dimension (front dimension), therefore permutation is - * required to put the dimension specified by @p axis to the first dimension. - * - * @param[in] axis Axis on which to perform softmax. Supported: 1, 2, 3 (0 implies no permutation needed) - * - * @return the permutation vector - */ -PermutationVector get_permutation_vector_from_softmax_axis(size_t axis); } // namespace arm_compute #include "arm_compute/core/Helpers.inl" -- cgit v1.2.1