diff options
Diffstat (limited to 'arm_compute/core/Types.h')
| -rw-r--r-- | arm_compute/core/Types.h | 1134 |
1 files changed, 472 insertions, 662 deletions
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h index 48c87cd8ac..f2f60c150e 100644 --- a/arm_compute/core/Types.h +++ b/arm_compute/core/Types.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2021 Arm Limited. + * Copyright (c) 2016-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,17 +21,52 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TYPES_H -#define ARM_COMPUTE_TYPES_H - +#ifndef ACL_ARM_COMPUTE_CORE_TYPES_H +#define ACL_ARM_COMPUTE_CORE_TYPES_H + +/** The following symbols have been moved to: + * half + * PermutationVector + * Format + * DataType + * DataLayout + * DataLayoutDimension + * PadStrideInfo + * WeightFormat + * Channel + * DimensionRoundingType + */ +#include "arm_compute/core/CoreTypes.h" +/** The following symbols have been moved to: + * ActivationFunction + * ActivationLayerInfo + */ +#include "arm_compute/function_info/ActivationLayerInfo.h" +/** The following symbols have been moved to: + * ConvolutionInfo + */ +#include "arm_compute/function_info/ConvolutionInfo.h" +/** The following symbols have been moved to: + * FullyConnectedLayerInfo + */ +#include "arm_compute/function_info/FullyConnectedLayerInfo.h" +/** The following symbols have been moved to: + * GEMMLowpOutputStageType + * GEMMLowpOutputStageInfo + * GEMMInfo + */ +#include "arm_compute/function_info/GEMMInfo.h" +/** The following symbols have been moved to: + * MatMulInfo + */ #include "arm_compute/core/Coordinates.h" -#include "arm_compute/core/QuantizationInfo.h" #include "arm_compute/core/Size2D.h" -#include "arm_compute/core/Strides.h" +#include "arm_compute/core/Size3D.h" #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/utils/misc/Macros.h" +#include "arm_compute/function_info/MatMulInfo.h" + #include "support/Bfloat16.h" -#include "support/Half.h" #include <cmath> #include <cstddef> @@ -42,62 +77,9 @@ namespace arm_compute { -/** 16-bit floating point type */ -using half = half_float::half; - -/** Permutation vector */ -using PermutationVector = Strides; /** Bidirectional strides */ using BiStrides = Coordinates; -/** Image colour formats */ -enum class Format -{ - UNKNOWN, /**< Unknown image format */ - U8, /**< 1 channel, 1 U8 per channel */ - S16, /**< 1 channel, 1 S16 per channel */ - U16, /**< 1 channel, 1 U16 per channel */ - S32, /**< 1 channel, 1 S32 per channel */ - U32, /**< 1 channel, 1 U32 per channel */ - BFLOAT16, /**< 16-bit brain floating-point number */ - F16, /**< 1 channel, 1 F16 per channel */ - F32, /**< 1 channel, 1 F32 per channel */ - UV88, /**< 2 channel, 1 U8 per channel */ - RGB888, /**< 3 channels, 1 U8 per channel */ - RGBA8888, /**< 4 channels, 1 U8 per channel */ - YUV444, /**< A 3 plane of 8 bit 4:4:4 sampled Y, U, V planes */ - YUYV422, /**< A single plane of 32-bit macro pixel of Y0, U0, Y1, V0 bytes */ - NV12, /**< A 2 plane YUV format of Luma (Y) and interleaved UV data at 4:2:0 sampling */ - NV21, /**< A 2 plane YUV format of Luma (Y) and interleaved VU data at 4:2:0 sampling */ - IYUV, /**< A 3 plane of 8-bit 4:2:0 sampled Y, U, V planes */ - UYVY422 /**< A single plane of 32-bit macro pixel of U0, Y0, V0, Y1 byte */ -}; - -/** Available data types */ -enum class DataType -{ - UNKNOWN, /**< Unknown data type */ - U8, /**< unsigned 8-bit number */ - S8, /**< signed 8-bit number */ - QSYMM8, /**< quantized, symmetric fixed-point 8-bit number */ - QASYMM8, /**< quantized, asymmetric fixed-point 8-bit number unsigned */ - QASYMM8_SIGNED, /**< quantized, asymmetric fixed-point 8-bit number signed */ - QSYMM8_PER_CHANNEL, /**< quantized, symmetric per channel fixed-point 8-bit number */ - U16, /**< unsigned 16-bit number */ - S16, /**< signed 16-bit number */ - QSYMM16, /**< quantized, symmetric fixed-point 16-bit number */ - QASYMM16, /**< quantized, asymmetric fixed-point 16-bit number */ - U32, /**< unsigned 32-bit number */ - S32, /**< signed 32-bit number */ - U64, /**< unsigned 64-bit number */ - S64, /**< signed 64-bit number */ - BFLOAT16, /**< 16-bit brain floating-point number */ - F16, /**< 16-bit floating-point number */ - F32, /**< 32-bit floating-point number */ - F64, /**< 64-bit floating-point number */ - SIZET /**< size_t */ -}; - /** Available Sampling Policies */ enum class SamplingPolicy { @@ -105,32 +87,13 @@ enum class SamplingPolicy TOP_LEFT /**< Samples are taken at pixel top left corner */ }; -/** [DataLayout enum definition] **/ - -/** Supported tensor data layouts */ -enum class DataLayout -{ - UNKNOWN, /**< Unknown data layout */ - NCHW, /**< Num samples, channels, height, width */ - NHWC /**< Num samples, height, width, channels */ -}; -/** [DataLayout enum definition] **/ - -/** Supported tensor data layout dimensions */ -enum class DataLayoutDimension -{ - CHANNEL, /**< channel */ - HEIGHT, /**< height */ - WIDTH, /**< width */ - BATCHES /**< batches */ -}; - /** Available ConvolutionMethod*/ enum class ConvolutionMethod { GEMM, /**< Convolution using GEMM */ GEMM_CONV2D, /**< Direct 2D GEMM convolution */ DIRECT, /**< Direct convolution */ + INDIRECT, /**< Indirect convolution */ WINOGRAD, /**< Convolution using Winograd */ FFT /**< Convolution using FFT */ }; @@ -145,8 +108,9 @@ enum class DepthwiseConvolutionFunction /** Available DeconvolutionMethod*/ enum class DeconvolutionMethod { - GEMM, /**< Deconvolution using GEMM */ - DIRECT, /**< Direct deconvolution */ + GEMM, /**< Deconvolution using GEMM */ + DIRECT, /**< Direct deconvolution */ + UPSCALE_CONV2D /**< Deconvolution with Upscaling */ }; /** Available FuseBatchNormalizationType*/ @@ -179,8 +143,7 @@ enum class ComparisonOperation struct ValidRegion { /** Default constructor */ - ValidRegion() - : anchor{}, shape{} + ValidRegion() : anchor{}, shape{} { } @@ -201,8 +164,7 @@ struct ValidRegion * @param[in] a_shape Shape of the valid region. * */ - ValidRegion(const Coordinates &an_anchor, const TensorShape &a_shape) - : anchor{ an_anchor }, shape{ a_shape } + ValidRegion(const Coordinates &an_anchor, const TensorShape &a_shape) : anchor{an_anchor}, shape{a_shape} { anchor.set_num_dimensions(std::max(anchor.num_dimensions(), shape.num_dimensions())); } @@ -215,7 +177,7 @@ struct ValidRegion * */ ValidRegion(const Coordinates &an_anchor, const TensorShape &a_shape, size_t num_dimensions) - : anchor{ an_anchor }, shape{ a_shape } + : anchor{an_anchor}, shape{a_shape} { ARM_COMPUTE_ERROR_ON(num_dimensions < std::max(anchor.num_dimensions(), shape.num_dimensions())); anchor.set_num_dimensions(num_dimensions); @@ -248,9 +210,22 @@ struct ValidRegion return *this; } + /** Check whether two valid regions are equal. + * + * @param[in] lhs LHS valid region + * @param[in] rhs RHS valid region + * + * @return True if the valid regions are the same. + */ + inline friend bool operator==(const ValidRegion &lhs, const ValidRegion &rhs); + Coordinates anchor; /**< Anchor for the start of the valid region. */ TensorShape shape; /**< Shape of the valid region. */ }; +inline bool operator==(const ValidRegion &lhs, const ValidRegion &rhs) +{ + return (lhs.anchor == rhs.anchor) && (lhs.shape == rhs.shape); +} /** Methods available to handle borders */ enum class BorderMode @@ -264,32 +239,24 @@ enum class BorderMode struct BorderSize { /** Empty border, i.e. no border */ - constexpr BorderSize() noexcept - : top{ 0 }, - right{ 0 }, - bottom{ 0 }, - left{ 0 } + constexpr BorderSize() noexcept : top{0}, right{0}, bottom{0}, left{0} { } /** Border with equal size around the 2D plane */ - explicit constexpr BorderSize(unsigned int size) noexcept - : top{ size }, - right{ size }, - bottom{ size }, - left{ size } + explicit constexpr BorderSize(unsigned int size) noexcept : top{size}, right{size}, bottom{size}, left{size} { } /** Border with same size for top/bottom and left/right */ constexpr BorderSize(unsigned int top_bottom, unsigned int left_right) - : top{ top_bottom }, right{ left_right }, bottom{ top_bottom }, left{ left_right } + : top{top_bottom}, right{left_right}, bottom{top_bottom}, left{left_right} { } /** Border with different sizes */ constexpr BorderSize(unsigned int top, unsigned int right, unsigned int bottom, unsigned int left) - : top{ top }, right{ right }, bottom{ bottom }, left{ left } + : top{top}, right{right}, bottom{bottom}, left{left} { } @@ -341,7 +308,7 @@ struct BorderSize * * @return true if they are equal */ - bool operator==(const BorderSize &rhs) + bool operator==(const BorderSize &rhs) const { return (top == rhs.top) && (right == rhs.right) && (bottom == rhs.bottom) && (left == rhs.left); } @@ -352,7 +319,7 @@ struct BorderSize * * @return true if they are different */ - bool operator!=(const BorderSize &rhs) + bool operator!=(const BorderSize &rhs) const { return !(*this == rhs); } @@ -378,7 +345,11 @@ struct BorderSize /** Container for 2D padding size */ using PaddingSize = BorderSize; -/** Policy to handle overflow */ +/** Policy to handle integer overflow + * @note: This is ignored by floating point operations where the overflow behavior adheres to the IEEE-754 standard + * which states that in case of overflow ±infinity is returned for the round-to-nearest modes (and follows the + * rounding rules for the directed rounding modes) by default. + */ enum class ConvertPolicy { WRAP, /**< Wrap around */ @@ -390,7 +361,7 @@ enum class InterpolationPolicy { NEAREST_NEIGHBOR, /**< Output values are defined to match the source pixel whose center is nearest to the sample position */ BILINEAR, /**< Output values are defined by bilinear interpolation between the pixels */ - AREA, /**< Output values are determined by averaging the source pixels whose areas fall under the area of the destination pixel, projected onto the source image */ + AREA, /**< Output values are determined by averaging the source pixels whose areas fall under the area of the destination pixel, projected onto the source image */ }; /** Bilinear Interpolation method used by LKTracker */ @@ -433,23 +404,6 @@ using PaddingList = std::vector<PaddingInfo>; /** Information to produce a tiled version of a Tensor */ using Multiples = std::vector<uint32_t>; -/** Available channels */ -enum class Channel -{ - UNKNOWN, /** Unknown channel format */ - C0, /**< First channel (used by formats with unknown channel types). */ - C1, /**< Second channel (used by formats with unknown channel types). */ - C2, /**< Third channel (used by formats with unknown channel types). */ - C3, /**< Fourth channel (used by formats with unknown channel types). */ - R, /**< Red channel. */ - G, /**< Green channel. */ - B, /**< Blue channel. */ - A, /**< Alpha channel. */ - Y, /**< Luma channel. */ - U, /**< Cb/U channel. */ - V /**< Cr/V/Value channel. */ -}; - /** Available reduction operations */ enum class ReductionOperation { @@ -514,21 +468,12 @@ enum class NormType */ struct DetectionWindow { - uint16_t x{ 0 }; /**< Top-left x coordinate */ - uint16_t y{ 0 }; /**< Top-left y coordinate */ - uint16_t width{ 0 }; /**< Width of the detection window */ - uint16_t height{ 0 }; /**< Height of the detection window */ - uint16_t idx_class{ 0 }; /**< Index of the class */ - float score{ 0.f }; /**< Confidence value for the detection window */ -}; - -/** Dimension rounding type when down-scaling on CNNs - * @note Used in pooling and convolution layer - */ -enum class DimensionRoundingType -{ - FLOOR, /**< Floor rounding */ - CEIL /**< Ceil rounding */ + uint16_t x{0}; /**< Top-left x coordinate */ + uint16_t y{0}; /**< Top-left y coordinate */ + uint16_t width{0}; /**< Width of the detection window */ + uint16_t height{0}; /**< Height of the detection window */ + uint16_t idx_class{0}; /**< Index of the class */ + float score{0.f}; /**< Confidence value for the detection window */ }; /** Available pooling types */ @@ -565,12 +510,28 @@ public: * @param[in] im_width (Optional) Boxes whose centers (on the x axis) is beyond im_width will be filtered. Defaults to 1 * @param[in] im_height (Optional) Boxes whose centers (on the y axis) is beyond im_height will be filtered. Defaults to 1 */ - BoxNMSLimitInfo(float score_thresh = 0.05f, float nms = 0.3f, - int detections = 100, bool soft_nms_enabled = false, - NMSType soft_nms_method = NMSType::LINEAR, - float soft_nms_sigma = 0.5f, float soft_nms_min_score_thres = 0.001f, bool suppress_size = false, float min_size = 1.0f, float im_width = 1.0f, float im_height = 1.0f) - : _score_thresh(score_thresh), _nms(nms), _detections_per_im(detections), _soft_nms_enabled(soft_nms_enabled), _soft_nms_method(soft_nms_method), _soft_nms_sigma(soft_nms_sigma), - _soft_nms_min_score_thres(soft_nms_min_score_thres), _suppress_size(suppress_size), _min_size(min_size), _im_width(im_width), _im_height(im_height) + BoxNMSLimitInfo(float score_thresh = 0.05f, + float nms = 0.3f, + int detections = 100, + bool soft_nms_enabled = false, + NMSType soft_nms_method = NMSType::LINEAR, + float soft_nms_sigma = 0.5f, + float soft_nms_min_score_thres = 0.001f, + bool suppress_size = false, + float min_size = 1.0f, + float im_width = 1.0f, + float im_height = 1.0f) + : _score_thresh(score_thresh), + _nms(nms), + _detections_per_im(detections), + _soft_nms_enabled(soft_nms_enabled), + _soft_nms_method(soft_nms_method), + _soft_nms_sigma(soft_nms_sigma), + _soft_nms_min_score_thres(soft_nms_min_score_thres), + _suppress_size(suppress_size), + _min_size(min_size), + _im_width(im_width), + _im_height(im_height) { } /** Get the score threshold */ @@ -644,120 +605,42 @@ private: }; /** Padding and stride information class */ -class PadStrideInfo +/** Padding information for 2D operations like Conv2d */ +struct Padding2D { -public: - /** Constructor - * - * @param[in] stride_x (Optional) Stride, in elements, across x. Defaults to 1. - * @param[in] stride_y (Optional) Stride, in elements, across y. Defaults to 1. - * @param[in] pad_x (Optional) Padding, in elements, across x. Defaults to 0. - * @param[in] pad_y (Optional) Padding, in elements, across y. Defaults to 0. - * @param[in] round (Optional) Dimensions rounding. Defaults to @ref FLOOR. - */ - PadStrideInfo(unsigned int stride_x = 1, unsigned int stride_y = 1, - unsigned int pad_x = 0, unsigned int pad_y = 0, - DimensionRoundingType round = DimensionRoundingType::FLOOR) - : _stride(std::make_pair(stride_x, stride_y)), - _pad_left(pad_x), - _pad_top(pad_y), - _pad_right(pad_x), - _pad_bottom(pad_y), - _round_type(round) - { - } - /** Constructor - * - * @param[in] stride_x Stride, in elements, across x. - * @param[in] stride_y Stride, in elements, across y. - * @param[in] pad_left Padding across x on the left, in elements. - * @param[in] pad_top Padding across y on the top, in elements. - * @param[in] pad_right Padding across x on the right, in elements. - * @param[in] pad_bottom Padding across y on the bottom, in elements. - * @param[in] round Dimensions rounding. - */ - PadStrideInfo(unsigned int stride_x, unsigned int stride_y, - unsigned int pad_left, unsigned int pad_right, - unsigned int pad_top, unsigned int pad_bottom, - DimensionRoundingType round) - : _stride(std::make_pair(stride_x, stride_y)), - _pad_left(pad_left), - _pad_top(pad_top), - _pad_right(pad_right), - _pad_bottom(pad_bottom), - _round_type(round) - { - } - /** Get the stride. - * - * @return a pair: stride x, stride y. - */ - std::pair<unsigned int, unsigned int> stride() const - { - return _stride; - } - /** Check whether the padding is symmetric. - * - * @return True if the padding is symmetric. - */ - bool padding_is_symmetric() const - { - return (_pad_left == _pad_right) && (_pad_top == _pad_bottom); - } - /** Get the padding. - * - * @note This should only be used when the padding is symmetric. - * - * @return a pair: padding left/right, padding top/bottom - */ - std::pair<unsigned int, unsigned int> pad() const + Padding2D() = default; + Padding2D(size_t left, size_t right, size_t top, size_t bottom) : left(left), right(right), top(top), bottom(bottom) { - //this accessor should be used only when padding is symmetric - ARM_COMPUTE_ERROR_ON(!padding_is_symmetric()); - return std::make_pair(_pad_left, _pad_top); } + size_t left = {0}; /**< Padding across the width dimension on the left, in elements. */ + size_t right = {0}; /**< Padding across the width dimension on the right, in elements. */ + size_t top = {0}; /**< Padding across the height dimension on the top, in elements. */ + size_t bottom = {0}; /**< Padding across the height dimension on the bottom, in elements. */ +}; - /** Get the left padding */ - unsigned int pad_left() const - { - return _pad_left; - } - /** Get the right padding */ - unsigned int pad_right() const - { - return _pad_right; - } - /** Get the top padding */ - unsigned int pad_top() const - { - return _pad_top; - } - /** Get the bottom padding */ - unsigned int pad_bottom() const +/** Padding information for 3D operations like Conv3d */ +struct Padding3D +{ + Padding3D() noexcept { - return _pad_bottom; } - /** Get the rounding type */ - DimensionRoundingType round() const + Padding3D(size_t pad_x, size_t pad_y, size_t pad_z) + : left(pad_x), right(pad_x), top(pad_y), bottom(pad_y), front(pad_z), back(pad_z) { - return _round_type; } - /** Check whether this has any padding */ - bool has_padding() const + Padding3D(size_t left, size_t right, size_t top, size_t bottom, size_t front, size_t back) + : left(left), right(right), top(top), bottom(bottom), front(front), back(back) { - return (_pad_left != 0 || _pad_top != 0 || _pad_right != 0 || _pad_bottom != 0); } -private: - std::pair<unsigned int, unsigned int> _stride; - unsigned int _pad_left; - unsigned int _pad_top; - unsigned int _pad_right; - unsigned int _pad_bottom; - - DimensionRoundingType _round_type; + size_t left = {0}; /**< Padding across the width dimenstion on the left, in elements. */ + size_t right = {0}; /**< Padding across the width dimenstion on the right, in elements. */ + size_t top = {0}; /**< Padding across the height dimenstion on the top, in elements. */ + size_t bottom = {0}; /**< Padding across the height dimenstion on the bottom, in elements. */ + size_t front = {0}; /**< Padding across the depth dimenstion on the front, in elements. */ + size_t back = {0}; /**< Padding across the depth dimenstion on the back, in elements. */ }; /** PriorBox layer info */ @@ -789,9 +672,15 @@ public: * @param[in] img_size (Optional) Image size. * @param[in] steps (Optional) Step values. */ - PriorBoxLayerInfo(const std::vector<float> &min_sizes, const std::vector<float> &variances, float offset, bool flip = true, bool clip = false, - const std::vector<float> &max_sizes = {}, const std::vector<float> &aspect_ratios = {}, - const Coordinates2D &img_size = Coordinates2D{ 0, 0 }, const std::array<float, 2> &steps = { { 0.f, 0.f } }) + PriorBoxLayerInfo(const std::vector<float> &min_sizes, + const std::vector<float> &variances, + float offset, + bool flip = true, + bool clip = false, + const std::vector<float> &max_sizes = {}, + const std::vector<float> &aspect_ratios = {}, + const Coordinates2D &img_size = Coordinates2D{0, 0}, + const std::array<float, 2> &steps = {{0.f, 0.f}}) : _min_sizes(min_sizes), _variances(variances), _offset(offset), @@ -803,22 +692,22 @@ public: _steps(steps) { _aspect_ratios.push_back(1.); - for(unsigned int i = 0; i < aspect_ratios.size(); ++i) + for (unsigned int i = 0; i < aspect_ratios.size(); ++i) { float ar = aspect_ratios[i]; bool already_exist = false; - for(auto ar_new : _aspect_ratios) + for (auto ar_new : _aspect_ratios) { - if(fabs(ar - ar_new) < 1e-6) + if (fabs(ar - ar_new) < 1e-6) { already_exist = true; break; } } - if(!already_exist) + if (!already_exist) { _aspect_ratios.push_back(ar); - if(flip) + if (flip) { _aspect_ratios.push_back(1.f / ar); } @@ -872,14 +761,14 @@ public: } private: - std::vector<float> _min_sizes; - std::vector<float> _variances; - float _offset; - bool _flip; - bool _clip; - std::vector<float> _max_sizes; - std::vector<float> _aspect_ratios; - Coordinates2D _img_size; + std::vector<float> _min_sizes; + std::vector<float> _variances; + float _offset; + bool _flip; + bool _clip; + std::vector<float> _max_sizes; + std::vector<float> _aspect_ratios; + Coordinates2D _img_size; std::array<float, 2> _steps; }; @@ -930,8 +819,16 @@ public: * @param[in] variance_encoded_in_target (Optional) If true, variance is encoded in target. Otherwise we need to adjust the predicted offset accordingly.Default set to false. * @param[in] eta (Optional) Eta. */ - DetectionOutputLayerInfo(int num_classes, bool share_location, DetectionOutputLayerCodeType code_type, int keep_top_k, float nms_threshold, int top_k = -1, int background_label_id = -1, - float confidence_threshold = std::numeric_limits<float>::lowest(), bool variance_encoded_in_target = false, float eta = 1) + DetectionOutputLayerInfo(int num_classes, + bool share_location, + DetectionOutputLayerCodeType code_type, + int keep_top_k, + float nms_threshold, + int top_k = -1, + int background_label_id = -1, + float confidence_threshold = std::numeric_limits<float>::lowest(), + bool variance_encoded_in_target = false, + float eta = 1) : _num_classes(num_classes), _share_location(share_location), _code_type(code_type), @@ -1045,8 +942,15 @@ public: * @param[in] detection_per_class (Optional) Number of detection per class. Used in the Regular Non-Max-Suppression. Defaults to 100. * @param[in] dequantize_scores (Optional) If the scores need to be dequantized. Defaults to true. */ - DetectionPostProcessLayerInfo(unsigned int max_detections, unsigned int max_classes_per_detection, float nms_score_threshold, float iou_threshold, unsigned int num_classes, - std::array<float, 4> scales_values, bool use_regular_nms = false, unsigned int detection_per_class = 100, bool dequantize_scores = true) + DetectionPostProcessLayerInfo(unsigned int max_detections, + unsigned int max_classes_per_detection, + float nms_score_threshold, + float iou_threshold, + unsigned int num_classes, + std::array<float, 4> scales_values, + bool use_regular_nms = false, + unsigned int detection_per_class = 100, + bool dequantize_scores = true) : _max_detections(max_detections), _max_classes_per_detection(max_classes_per_detection), _nms_score_threshold(nms_score_threshold), @@ -1124,15 +1028,15 @@ public: } private: - unsigned int _max_detections; - unsigned int _max_classes_per_detection; - float _nms_score_threshold; - float _iou_threshold; - unsigned int _num_classes; + unsigned int _max_detections; + unsigned int _max_classes_per_detection; + float _nms_score_threshold; + float _iou_threshold; + unsigned int _num_classes; std::array<float, 4> _scales_values; - bool _use_regular_nms; - unsigned int _detection_per_class; - bool _dequantize_scores; + bool _use_regular_nms; + unsigned int _detection_per_class; + bool _dequantize_scores; }; /** Pooling Layer Information struct*/ @@ -1146,7 +1050,9 @@ struct PoolingLayerInfo pad_stride_info(PadStrideInfo()), exclude_padding(false), is_global_pooling(false), - fp_mixed_precision(false) + fp_mixed_precision(false), + use_inf_as_limit(true), + use_kernel_indices(false) { } /** Constructor @@ -1159,20 +1065,26 @@ struct PoolingLayerInfo * True will exclude padding while false will not (Used in AVG/L2 pooling to determine the pooling area). * Defaults to false; * @param[in] fp_mixed_precision (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy. + * @param[in] use_inf_as_limit (Optional) Use inf to represent the limits of datatypes range, instead of using "lowest" property of the data type. + * @param[in] use_kernel_indices (Optional) Use kernel indices instead of using source indices while computing indices tensor. */ explicit PoolingLayerInfo(PoolingType pool_type, unsigned int pool_size, DataLayout data_layout, PadStrideInfo pad_stride_info = PadStrideInfo(), bool exclude_padding = false, - bool fp_mixed_precision = false) + bool fp_mixed_precision = false, + bool use_inf_as_limit = true, + bool use_kernel_indices = false) : pool_type(pool_type), pool_size(Size2D(pool_size, pool_size)), data_layout(data_layout), pad_stride_info(pad_stride_info), exclude_padding(exclude_padding), is_global_pooling(false), - fp_mixed_precision(fp_mixed_precision) + fp_mixed_precision(fp_mixed_precision), + use_inf_as_limit(use_inf_as_limit), + use_kernel_indices(use_kernel_indices) { } @@ -1186,20 +1098,26 @@ struct PoolingLayerInfo * True will exclude padding while false will not (Used in AVG/L2 pooling to determine the pooling area). * Defaults to false; * @param[in] fp_mixed_precision (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy. + * @param[in] use_inf_as_limit (Optional) Use inf to represent the limits of datatypes range, instead of using "lowest" property of the data type. + * @param[in] use_kernel_indices (Optional) Use kernel indices instead of using source indices while computing indices tensor. */ explicit PoolingLayerInfo(PoolingType pool_type, Size2D pool_size, DataLayout data_layout, PadStrideInfo pad_stride_info = PadStrideInfo(), bool exclude_padding = false, - bool fp_mixed_precision = false) + bool fp_mixed_precision = false, + bool use_inf_as_limit = true, + bool use_kernel_indices = false) : pool_type(pool_type), pool_size(pool_size), data_layout(data_layout), pad_stride_info(pad_stride_info), exclude_padding(exclude_padding), is_global_pooling(false), - fp_mixed_precision(fp_mixed_precision) + fp_mixed_precision(fp_mixed_precision), + use_inf_as_limit(use_inf_as_limit), + use_kernel_indices(use_kernel_indices) { } @@ -1217,7 +1135,9 @@ struct PoolingLayerInfo pad_stride_info(PadStrideInfo(1, 1, 0, 0)), exclude_padding(false), is_global_pooling(true), - fp_mixed_precision(false) + fp_mixed_precision(false), + use_inf_as_limit(true), + use_kernel_indices(false) { } @@ -1228,6 +1148,111 @@ struct PoolingLayerInfo bool exclude_padding; bool is_global_pooling; bool fp_mixed_precision; + bool use_inf_as_limit; + bool use_kernel_indices; +}; + +/** Pooling Layer Information struct*/ +struct Pooling3dLayerInfo +{ + /** Default Constructor */ + Pooling3dLayerInfo() noexcept + : pool_type(PoolingType::MAX), + pool_size(Size3D()), + stride(Size3D()), + padding(Padding3D()), + exclude_padding(false), + is_global_pooling(false), + fp_mixed_precision(false), + round_type(DimensionRoundingType::FLOOR) + { + } + /** Constructor + * + * @param[in] pool_type Pooling type @ref PoolingType. + * @param[in] pool_size Pooling size, in elements, across x, y and z. + * @param[in] stride (Optional) stride information @ref Size3D + * @param[in] padding (Optional) padding information @ref Padding3D + * @param[in] exclude_padding (Optional) Strategy when accounting padding in calculations. + * True will exclude padding while false will not (Used in AVG/L2 pooling to determine the pooling area). + * Defaults to false; + * @param[in] fp_mixed_precision (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy. + * @param[in] round_type (Optional) Dimensions rounding. Defaults to @ref DimensionRoundingType::FLOOR + */ + explicit Pooling3dLayerInfo(PoolingType pool_type, + unsigned int pool_size, + Size3D stride = Size3D(1U, 1U, 1U), + Padding3D padding = Padding3D(), + bool exclude_padding = false, + bool fp_mixed_precision = false, + DimensionRoundingType round_type = DimensionRoundingType::FLOOR) + : pool_type(pool_type), + pool_size(Size3D(pool_size, pool_size, pool_size)), + stride(stride), + padding(padding), + exclude_padding(exclude_padding), + is_global_pooling(false), + fp_mixed_precision(fp_mixed_precision), + round_type(round_type) + { + } + + /** Constructor + * + * @param[in] pool_type Pooling type @ref PoolingType. + * @param[in] pool_size Pooling size, in elements, across x, y and z. + * @param[in] stride (Optional) stride information @ref Size3D + * @param[in] padding (Optional) padding information @ref Padding3D + * @param[in] exclude_padding (Optional) Strategy when accounting padding in calculations. + * True will exclude padding while false will not (Used in AVG/L2 pooling to determine the pooling area). + * Defaults to false; + * @param[in] fp_mixed_precision (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy. + * @param[in] round_type (Optional) Dimensions rounding. Defaults to @ref DimensionRoundingType::FLOOR + */ + explicit Pooling3dLayerInfo(PoolingType pool_type, + Size3D pool_size, + Size3D stride = Size3D(1U, 1U, 1U), + Padding3D padding = Padding3D(), + bool exclude_padding = false, + bool fp_mixed_precision = false, + DimensionRoundingType round_type = DimensionRoundingType::FLOOR) + : pool_type(pool_type), + pool_size(pool_size), + stride(stride), + padding(padding), + exclude_padding(exclude_padding), + is_global_pooling(false), + fp_mixed_precision(fp_mixed_precision), + round_type(round_type) + { + } + + /** Constructor + * + * @note This constructor is used for global pooling + * + * @param[in] pool_type Pooling type @ref PoolingType. + */ + explicit Pooling3dLayerInfo(PoolingType pool_type) + : pool_type(pool_type), + pool_size(Size3D()), + stride(Size3D(1U, 1U, 1U)), + padding(Padding3D(0, 0, 0)), + exclude_padding(false), + is_global_pooling(true), + fp_mixed_precision(false), + round_type(DimensionRoundingType::FLOOR) + { + } + + PoolingType pool_type; + Size3D pool_size; + Size3D stride; + Padding3D padding; + bool exclude_padding; + bool is_global_pooling; + bool fp_mixed_precision; + DimensionRoundingType round_type; }; /** ROI Pooling Layer Information class */ @@ -1241,8 +1266,14 @@ public: * @param[in] spatial_scale Spatial scale to be applied to the ROI coordinates and dimensions. * @param[in] sampling_ratio Number of samples to include in each pooling region (if set to zero, a ceil(roi_dims/pooling_dims)) */ - ROIPoolingLayerInfo(unsigned int pooled_width, unsigned int pooled_height, float spatial_scale, unsigned int sampling_ratio = 0) - : _pooled_width(pooled_width), _pooled_height(pooled_height), _spatial_scale(spatial_scale), _sampling_ratio(sampling_ratio) + ROIPoolingLayerInfo(unsigned int pooled_width, + unsigned int pooled_height, + float spatial_scale, + unsigned int sampling_ratio = 0) + : _pooled_width(pooled_width), + _pooled_height(pooled_height), + _spatial_scale(spatial_scale), + _sampling_ratio(sampling_ratio) { } /** Get the pooled width of the layer */ @@ -1289,10 +1320,24 @@ public: * @param[in] min_size (Optional)Size used to validate the anchors produced. Defaults to 16. * @param[in] values_per_roi (Optional)Values used to represent a ROI(Region of interest). Defaults to 4. */ - GenerateProposalsInfo(float im_width, float im_height, float im_scale, float spatial_scale = 1.0, int pre_nms_topN = 6000, int post_nms_topN = 300, float nms_thres = 0.7, float min_size = 16.0, + GenerateProposalsInfo(float im_width, + float im_height, + float im_scale, + float spatial_scale = 1.0, + int pre_nms_topN = 6000, + int post_nms_topN = 300, + float nms_thres = 0.7, + float min_size = 16.0, size_t values_per_roi = 4) - : _im_height(im_height), _im_width(im_width), _im_scale(im_scale), _spatial_scale(spatial_scale), _pre_nms_topN(pre_nms_topN), _post_nms_topN(post_nms_topN), _nms_thres(nms_thres), - _min_size(min_size), _values_per_roi(values_per_roi) + : _im_height(im_height), + _im_width(im_width), + _im_scale(im_scale), + _spatial_scale(spatial_scale), + _pre_nms_topN(pre_nms_topN), + _post_nms_topN(post_nms_topN), + _nms_thres(nms_thres), + _min_size(min_size), + _values_per_roi(values_per_roi) { } @@ -1418,11 +1463,20 @@ public: * @param[in] correct_transform_coords (Optional)Correct bounding box transform coordinates. Defaults to false * @param[in] bbox_xform_clip (Optional)Minimum bounding box width and height after bounding box transformation in log-space. Defaults to log(1000/16) */ - BoundingBoxTransformInfo(float img_width, float img_height, float scale, bool apply_scale = false, const std::array<float, 4> weights = { { 1.f, 1.f, 1.f, 1.f } }, bool correct_transform_coords = - false, - float bbox_xform_clip = - 4.135166556742356f) - : _img_width(img_width), _img_height(img_height), _scale(scale), _apply_scale(apply_scale), _correct_transform_coords(correct_transform_coords), _weights(weights), _bbox_xform_clip(bbox_xform_clip) + BoundingBoxTransformInfo(float img_width, + float img_height, + float scale, + bool apply_scale = false, + const std::array<float, 4> weights = {{1.f, 1.f, 1.f, 1.f}}, + bool correct_transform_coords = false, + float bbox_xform_clip = 4.135166556742356f) + : _img_width(img_width), + _img_height(img_height), + _scale(scale), + _apply_scale(apply_scale), + _correct_transform_coords(correct_transform_coords), + _weights(weights), + _bbox_xform_clip(bbox_xform_clip) { } @@ -1462,114 +1516,13 @@ public: } private: - float _img_width; - float _img_height; - float _scale; - bool _apply_scale; - bool _correct_transform_coords; + float _img_width; + float _img_height; + float _scale; + bool _apply_scale; + bool _correct_transform_coords; std::array<float, 4> _weights; - float _bbox_xform_clip; -}; - -/** Activation Layer Information class */ -class ActivationLayerInfo -{ -public: - /** Available activation functions */ - enum class ActivationFunction - { - LOGISTIC, /**< Logistic ( \f$ f(x) = \frac{1}{1 + e^{-x}} \f$ ) */ - TANH, /**< Hyperbolic tangent ( \f$ f(x) = a \cdot tanh(b \cdot x) \f$ ) */ - RELU, /**< Rectifier ( \f$ f(x) = max(0,x) \f$ ) */ - BOUNDED_RELU, /**< Upper Bounded Rectifier ( \f$ f(x) = min(a, max(0,x)) \f$ ) */ - LU_BOUNDED_RELU, /**< Lower and Upper Bounded Rectifier ( \f$ f(x) = min(a, max(b,x)) \f$ ) */ - LEAKY_RELU, /**< Leaky Rectifier ( \f$ f(x) = \begin{cases} \alpha x & \quad \text{if } x \text{ < 0}\\ x & \quad \text{if } x \geq \text{ 0 } \end{cases} \f$ ) */ - SOFT_RELU, /**< Soft Rectifier ( \f$ f(x)= log(1+e^x) \f$ ) */ - ELU, /**< Exponential Linear Unit ( \f$ f(x) = \begin{cases} \alpha (exp(x) - 1) & \quad \text{if } x \text{ < 0}\\ x & \quad \text{if } x \geq \text{ 0 } \end{cases} \f$ ) */ - ABS, /**< Absolute ( \f$ f(x)= |x| \f$ ) */ - SQUARE, /**< Square ( \f$ f(x)= x^2 \f$ )*/ - SQRT, /**< Square root ( \f$ f(x) = \sqrt{x} \f$ )*/ - LINEAR, /**< Linear ( \f$ f(x)= ax + b \f$ ) */ - IDENTITY, /**< Identity ( \f$ f(x)= x \f$ ) */ - HARD_SWISH /**< Hard-swish ( \f$ f(x) = (x * relu6(x+3))/6 \f$ ) */ - }; - - ActivationLayerInfo() = default; - /** Default Constructor - * - * @param[in] f The activation function to use. - * @param[in] a (Optional) The alpha parameter used by some activation functions - * (@ref ActivationFunction::BOUNDED_RELU, @ref ActivationFunction::LU_BOUNDED_RELU, @ref ActivationFunction::LINEAR, @ref ActivationFunction::TANH). - * @param[in] b (Optional) The beta parameter used by some activation functions (@ref ActivationFunction::LINEAR, @ref ActivationFunction::LU_BOUNDED_RELU, @ref ActivationFunction::TANH). - */ - ActivationLayerInfo(ActivationFunction f, float a = 0.0f, float b = 0.0f) - : _act(f), _a(a), _b(b), _enabled(true) - { - } - /** Get the type of activation function */ - ActivationFunction activation() const - { - return _act; - } - /** Get the alpha value */ - float a() const - { - return _a; - } - /** Get the beta value */ - float b() const - { - return _b; - } - /** Check if initialised */ - bool enabled() const - { - return _enabled; - } - -private: - ActivationFunction _act = { ActivationLayerInfo::ActivationFunction::IDENTITY }; - float _a = {}; - float _b = {}; - bool _enabled = { false }; -}; - -/** Fully connected layer info */ -struct FullyConnectedLayerInfo -{ - /* Fused-activation parameters */ - ActivationLayerInfo activation_info{}; /**< Fused activation to apply after the matrix multiplication. */ - /* Information about weights */ - DataLayout weights_trained_layout{ DataLayout::NCHW }; /**< Layout that the weights have been trained with. */ - bool transpose_weights{ true }; /**< Transpose weights if true. */ - bool are_weights_reshaped{ false }; /**< Reshape the weights tensor if false. */ - bool retain_internal_weights{ false }; /**< Retain internal reshaped weights. */ - bool constant_weights{ true }; /**< If false, weights can vary between runs. */ - /* Other parameters */ - bool fp_mixed_precision{ false }; /**< Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy. */ - - /** Sets the weights trained data layout - * - * @param[in] layout Data layout that the weights were trained with - * - * @return Updated object - */ - FullyConnectedLayerInfo &set_weights_trained_layout(DataLayout layout) - { - weights_trained_layout = layout; - return *this; - } - /** Sets the transpose weights flag - * - * @param[in] should_transpose_weights Boolean flag indicating if weights should be transposed - * - * @return Updated object - */ - FullyConnectedLayerInfo &set_transpose_weights(bool should_transpose_weights) - { - transpose_weights = should_transpose_weights; - return *this; - } + float _bbox_xform_clip; }; /** Normalization Layer Information class */ @@ -1586,7 +1539,12 @@ public: * @param[in] is_scaled (Optional) Boolean that specifies if alpha will be scaled by the normalization size or not. * Should be false to follow [Krichevksy 2012]. */ - NormalizationLayerInfo(NormType type, uint32_t norm_size = 5, float alpha = 0.0001f, float beta = 0.5f, float kappa = 1.f, bool is_scaled = true) + NormalizationLayerInfo(NormType type, + uint32_t norm_size = 5, + float alpha = 0.0001f, + float beta = 0.5f, + float kappa = 1.f, + bool is_scaled = true) : _type(type), _norm_size(norm_size), _alpha(alpha), _beta(beta), _kappa(kappa), _is_scaled(is_scaled) { } @@ -1690,13 +1648,36 @@ private: int32_t _shrink_axis_mask; }; +// OHWIo<interleave_by>i<block_by> +inline int interleave_by(const WeightFormat wf) +{ + return (static_cast<int>(wf) >> 8) & 0xFFF; +} +inline int block_by(const WeightFormat wf) +{ + return (static_cast<int>(wf) >> 20) & 0xF; +} +inline bool is_fixed_format(const WeightFormat &wf) +{ + return wf != WeightFormat::UNSPECIFIED && wf != WeightFormat::ANY; +} +inline bool is_fixed_format_fast_math(const WeightFormat &wf) +{ + return (static_cast<int>(wf) >> 4) & 0x1; +} + /** Convolution Layer Weights Information class. This class stores the necessary information to compute convolution layer when the weights are already reshaped */ class WeightsInfo { public: /** Default constructor */ WeightsInfo() - : _are_reshaped(false), _kernel_width(0), _kernel_height(0), _num_kernels(0), _retain_internal_weights(false) + : _are_reshaped(false), + _kernel_width(0), + _kernel_height(0), + _num_kernels(0), + _retain_internal_weights(false), + _weight_format(arm_compute::WeightFormat::UNSPECIFIED) { } /** Constructor @@ -1706,9 +1687,20 @@ public: * @param[in] kernel_height Kernel height. * @param[in] num_kernels Number of convolution kernels. * @param[in] retain_internal_weights (Optional) True if internal reshaped weights must be retained. Used for reconfiguration purposes. Default is false. + * @param[in] weight_format (Optional) arm_gemm:WeightFormat enumeration requested by the user. Default is arm_compute::WeightFormat::UNSPECIFIED. */ - WeightsInfo(bool are_reshaped, unsigned int kernel_width, unsigned int kernel_height, unsigned int num_kernels, bool retain_internal_weights = false) - : _are_reshaped(are_reshaped), _kernel_width(kernel_width), _kernel_height(kernel_height), _num_kernels(num_kernels), _retain_internal_weights(retain_internal_weights) + WeightsInfo(bool are_reshaped, + unsigned int kernel_width, + unsigned int kernel_height, + unsigned int num_kernels, + bool retain_internal_weights = false, + arm_compute::WeightFormat weight_format = arm_compute::WeightFormat::UNSPECIFIED) + : _are_reshaped(are_reshaped), + _kernel_width(kernel_width), + _kernel_height(kernel_height), + _num_kernels(num_kernels), + _retain_internal_weights(retain_internal_weights), + _weight_format(weight_format) { } /** Flag which specifies if the weights tensor has been reshaped. @@ -1739,21 +1731,39 @@ public: { return _retain_internal_weights; } + arm_compute::WeightFormat weight_format() const + { + return _weight_format; + } + void set_weight_format(arm_compute::WeightFormat weight_format) + { + _weight_format = weight_format; + } + + unsigned int kernel_width() const + { + return _kernel_width; + } + unsigned int kernel_height() const + { + return _kernel_height; + } private: - bool _are_reshaped; - unsigned int _kernel_width; - unsigned int _kernel_height; - unsigned int _num_kernels; - bool _retain_internal_weights; + bool _are_reshaped; + unsigned int _kernel_width; + unsigned int _kernel_height; + unsigned int _num_kernels; + bool _retain_internal_weights; + arm_compute::WeightFormat _weight_format; }; /** GEMM reshape information class. This class stores the necessary information about matrix A and matrix B reshape. * - * The matrix A can only be reshaped through @ref opencl::kernels::ClGemmReshapeLhsMatrixKernel or @ref NEGEMMInterleave4x4Kernel + * The matrix A can only be reshaped through @ref opencl::kernels::ClGemmReshapeLhsMatrixKernel or @ref cpu::kernels::CpuGemmInterleave4x4Kernel * Note: Optionally just for @ref opencl::kernels::ClGemmReshapeLhsMatrixKernel is it possible to set mult_interleave4x4_height, the multiplication factor for the height of the 4x4 interleaved block * - * The matrix B can only be reshaped through @ref opencl::kernels::ClGemmReshapeRhsMatrixKernel or @ref NEGEMMTranspose1xWKernel + * The matrix B can only be reshaped through @ref opencl::kernels::ClGemmReshapeRhsMatrixKernel or @ref cpu::kernels::CpuGemmTranspose1xWKernel * Note: Optionally just for @ref opencl::kernels::ClGemmReshapeRhsMatrixKernel is it possible to set mult_transpose1xW_width, the multiplication factor for the width of the 1xW transposed block * */ @@ -1762,7 +1772,14 @@ class GEMMReshapeInfo final public: /** Default constructor */ GEMMReshapeInfo() - : _m(1), _n(1), _k(1), _mult_transpose1xW_width(1), _mult_interleave4x4_height(1), _depth_output_gemm3d(0), _reinterpret_input_as_3d(false), _broadcast_bias(false) + : _m(1), + _n(1), + _k(1), + _mult_transpose1xW_width(1), + _mult_interleave4x4_height(1), + _depth_output_gemm3d(0), + _reinterpret_input_as_3d(false), + _broadcast_bias(false) { } /** Constructor @@ -1778,9 +1795,22 @@ public: * to perform 1x1 convolutions with the NHWC data layout) * @param[in] broadcast_bias (Optional) Broadcast the shape of the bias tensor from a vector to a matrix. */ - GEMMReshapeInfo(int m, int n, int k, int mult_transpose1xW_width = 1, int mult_interleave4x4_height = 1, int depth_output_gemm3d = 0, bool reinterpret_input_as_3d = false, bool broadcast_bias = false) - : _m(m), _n(n), _k(k), _mult_transpose1xW_width(mult_transpose1xW_width), _mult_interleave4x4_height(mult_interleave4x4_height), _depth_output_gemm3d(depth_output_gemm3d), - _reinterpret_input_as_3d(reinterpret_input_as_3d), _broadcast_bias(broadcast_bias) + GEMMReshapeInfo(int m, + int n, + int k, + int mult_transpose1xW_width = 1, + int mult_interleave4x4_height = 1, + int depth_output_gemm3d = 0, + bool reinterpret_input_as_3d = false, + bool broadcast_bias = false) + : _m(m), + _n(n), + _k(k), + _mult_transpose1xW_width(mult_transpose1xW_width), + _mult_interleave4x4_height(mult_interleave4x4_height), + _depth_output_gemm3d(depth_output_gemm3d), + _reinterpret_input_as_3d(reinterpret_input_as_3d), + _broadcast_bias(broadcast_bias) { } /** Number of matrix A rows @@ -1862,44 +1892,6 @@ private: bool _broadcast_bias; }; -struct ConvolutionInfo -{ - ConvolutionInfo() = default; - ConvolutionInfo(const PadStrideInfo &pad_stride_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation) - : pad_stride_info(pad_stride_info), depth_multiplier(depth_multiplier), act_info(act_info), dilation(dilation) - { - } - PadStrideInfo pad_stride_info{}; /**< Convolution info (Pads, strides,...) */ - unsigned int depth_multiplier{ 1 }; /**< Multiplier to apply to input's depth to retrieve the output depth. Defaults to 1 */ - ActivationLayerInfo act_info{}; /**< Fused activation to apply after convolution. */ - Size2D dilation{ Size2D(1, 1) }; /**< Dilation, in elements, across x and y. Defaults to (1, 1). */ -}; - -/** GEMMLowp output stage type */ -enum class GEMMLowpOutputStageType -{ - NONE, /**< No quantization */ - QUANTIZE_DOWN, /**< Quantize using an integer multiplication */ - QUANTIZE_DOWN_FIXEDPOINT, /**< Quantize using a fixed point multiplication */ - QUANTIZE_DOWN_FLOAT /**< Quantize using a floating point multiplication */ -}; - -/** GEMMLowp output stage info */ -struct GEMMLowpOutputStageInfo -{ - GEMMLowpOutputStageType type{ GEMMLowpOutputStageType::NONE }; /**< GEMMLowp output stage type */ - int32_t gemmlowp_offset{ 0 }; /**< GEMMLowp output stage offset used for quantizing to QASYMM8 */ - int32_t gemmlowp_multiplier{ 0 }; /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */ - int32_t gemmlowp_shift{ 0 }; /**< GEMMLowp output stage shift used for quantizing to uint8 */ - int32_t gemmlowp_min_bound{ std::numeric_limits<int32_t>::lowest() }; /**< GEMMLowp min value used to saturate down the output result before converting back to QASYMM8 */ - int32_t gemmlowp_max_bound{ std::numeric_limits<int32_t>::max() }; /**< GEMMLowp max value used to saturate down the output result before converting back to QASYMM8 */ - std::vector<int32_t> gemmlowp_multipliers{}; /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */ - std::vector<int32_t> gemmlowp_shifts{}; /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */ - float gemmlowp_real_multiplier{ 0 }; /**< GEMMLowp output stage real multiplier used for quantizing to QASYMM8 */ - bool is_quantized_per_channel{ false }; /**< GEMMLowp quantized per-channel flag */ - DataType output_data_type{ DataType::UNKNOWN }; /**< Output tensor data type to use if the output is not initialized */ -}; - /** GEMM LHS (Left Hand Side) matrix information */ struct GEMMLHSMatrixInfo { @@ -1908,11 +1900,11 @@ struct GEMMLHSMatrixInfo : m0(m), k0(k), v0(v), transpose(trans), interleave(inter) { } - unsigned int m0{ 1 }; /**< Number of rows processed by the matrix multiplication */ - unsigned int k0{ 1 }; /**< Number of partial accumulations performed by the matrix multiplication */ - unsigned int v0{ 1 }; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */ - bool transpose{ true }; /**< True if the (m0xk0) block has to be transposed before been stored */ - bool interleave{ true }; /**< True if the v0 (m0xk0) blocks have to be interleaved in the output row */ + unsigned int m0{1}; /**< Number of rows processed by the matrix multiplication */ + unsigned int k0{1}; /**< Number of partial accumulations performed by the matrix multiplication */ + unsigned int v0{1}; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */ + bool transpose{true}; /**< True if the (m0xk0) block has to be transposed before been stored */ + bool interleave{true}; /**< True if the v0 (m0xk0) blocks have to be interleaved in the output row */ }; /** GEMM RHS (Right Hand Side) matrix information */ @@ -1923,208 +1915,16 @@ struct GEMMRHSMatrixInfo : n0(n), k0(k), h0(h), transpose(trans), interleave(inter), export_to_cl_image(export_to_cl_img) { } - unsigned int n0{ 1 }; /**< Number of columns processed by the matrix multiplication */ - unsigned int k0{ 1 }; /**< Number of partial accumulations performed by the matrix multiplication */ - unsigned int h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ - bool transpose{ true }; /**< True if the (k0xn0) block has to be transposed before been stored */ - bool interleave{ true }; /**< True if the h0 (k0xn0) blocks have to be interleaved in the output row */ - bool export_to_cl_image{ false }; /**< True if the reshaped rhs has to be exported to cl_image. n0 must be equal to 4 */ + unsigned int n0{1}; /**< Number of columns processed by the matrix multiplication */ + unsigned int k0{1}; /**< Number of partial accumulations performed by the matrix multiplication */ + unsigned int h0{1}; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ + bool transpose{true}; /**< True if the (k0xn0) block has to be transposed before been stored */ + bool interleave{true}; /**< True if the h0 (k0xn0) blocks have to be interleaved in the output row */ + bool export_to_cl_image{ + false}; /**< True if the reshaped rhs has to be exported to cl_image. n0 must be equal to 4 */ }; -/** GEMM information class. This class stores the necessary information to compute GEMM functions - * - * This object also contains the information about how matrix A and matrix B have been reshaped - * - */ -class GEMMInfo -{ -public: - /** Default constructor */ - GEMMInfo() noexcept - : _is_a_reshaped(false), - _is_b_reshaped(false), - _reshape_b_only_on_first_run(true), - _depth_output_gemm3d(0), - _reinterpret_input_as_3d(false), - _retain_internal_weights(false), - _gemmlowp_output_stage(), - _fp_mixed_precision(false), - _broadcast_bias(false), - _pretranpose_B(true), - _activation_info(), - _constant_weights(true) - { - } - /** Constructor - * - * @param[in] is_a_reshaped True if the matrix A has been reshaped - * @param[in] is_b_reshaped True if the matrix B has been reshaped - * @param[in] reshape_b_only_on_first_run Reshape matrix B only for the first run - * @param[in] depth_output_gemm3d (Optional) Depth (third dimension) of the output tensor to be used with the GEMM3D kernel - * If 0 the output will not be reinterpreted as 3D. Default 0 - * @param[in] reinterpret_input_as_3d (Optional) Reinterpret the input as 3D tensor. (i.e. this flag should be set to true when GEMM is used - * to perform 1x1 convolutions with the NHWC data layout) - * @param[in] retain_internal_weights (Optional) Retain the weights tensor from previous run - * @param[in] gemmlowp_output_stage (Optional) GEMMLowp Output stage info - * @param[in] fp_mixed_precision (Optional) Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy. - * @param[in] broadcast_bias (Optional) Broadcast the shape of the bias tensor from a vector to a matrix. - * @param[in] activation_info (Optional) Activation to apply after the matrix multiplication - * @param[in] constant_weights (Optional) Weights have constant values throughout multiple executions - */ - GEMMInfo(bool is_a_reshaped, bool is_b_reshaped, bool reshape_b_only_on_first_run, int depth_output_gemm3d = 0, bool reinterpret_input_as_3d = false, bool retain_internal_weights = false, - GEMMLowpOutputStageInfo gemmlowp_output_stage = GEMMLowpOutputStageInfo(), bool fp_mixed_precision = false, bool broadcast_bias = false, - const ActivationLayerInfo &activation_info = ActivationLayerInfo(), bool constant_weights = true) noexcept - : _is_a_reshaped(is_a_reshaped), - _is_b_reshaped(is_b_reshaped), - _reshape_b_only_on_first_run(reshape_b_only_on_first_run), - _depth_output_gemm3d(depth_output_gemm3d), - _reinterpret_input_as_3d(reinterpret_input_as_3d), - _retain_internal_weights(retain_internal_weights), - _gemmlowp_output_stage(gemmlowp_output_stage), - _fp_mixed_precision(fp_mixed_precision), - _broadcast_bias(broadcast_bias), - _pretranpose_B(reshape_b_only_on_first_run), - _activation_info(activation_info), - _constant_weights(constant_weights) - { - } - /** Flag which specifies if the matrix A has been reshaped - * - * @return True if the matrix A has been reshaped - */ - bool is_a_reshaped() const - { - return _is_a_reshaped; - }; - /** Flag which specifies if the matrix B has been reshaped - * - * @return True if the matrix B has been reshaped - */ - bool is_b_reshaped() const - { - return _is_b_reshaped; - }; - /** Flag which specifies if the reshape of matrix B should executed only for the first - * - * @note This flag could be set to TRUE when GEMM is used to accelerate convolution layer - * - * @return True if the reshaped of matrix B happens only for the first run - */ - bool reshape_b_only_on_first_run() const - { - return _reshape_b_only_on_first_run; - }; - /** Depth of the output when GEMM output is reinterpreted as 3D tensor - * - * @return the depth of the output tensor - */ - int depth_output_gemm3d() const - { - return _depth_output_gemm3d; - }; - /** Flag which specifies if the input tensor has to be reinterpreted as 3D - * - * @return True if the input tensor has to be reinterpreted as 3D tensor - */ - bool reinterpret_input_as_3d() const - { - return _reinterpret_input_as_3d; - }; - /** Flag which specifies if the weights tensor has to be retained from previous run - * - * @return True if the weights tensor has to be retained - */ - bool retain_internal_weights() const - { - return _retain_internal_weights; - }; - /** GEMMLowp output stage - * - * @return the GEMMLowp output stage info - */ - GEMMLowpOutputStageInfo gemmlowp_output_stage() const - { - return _gemmlowp_output_stage; - }; - /** Sets GEMMLowp output stage - * - * @param[in] output_stage Output stage to set - */ - void set_gemmlowp_output_stage(GEMMLowpOutputStageInfo &output_stage) - { - _gemmlowp_output_stage = output_stage; - }; - /** Flag which specifies if a wider accumulator should be used. - * - * @return True if a wider accumulator has to be used - */ - bool fp_mixed_precision() const - { - return _fp_mixed_precision; - }; - /** Flag which specifies whether to broadcast the shape of the bias tensor. - * - * @return True if the shape of the bias tensor is to be broadcasted. - */ - bool broadcast_bias() const - { - return _broadcast_bias; - }; - /** Flag which specifies whether b should be pre-transposed if supported. - * - * @return True if b should be pre-transposed else false. - */ - bool pretranpose_B() const - { - return _pretranpose_B; - }; - /** Set pre-transpose b flag - * - * @param[in] flag Flag to set - */ - void set_pretranpose_B(bool flag) - { - _pretranpose_B = flag; - } - /** Activation layer to apply after the matrix multiplication - * - * @return ActivationLayerInfo object - */ - ActivationLayerInfo activation_info() const - { - return _activation_info; - } - /** Set activation layer info - * - * @param[in] activation_info ActivationLayerInfo object to set - */ - void set_activation_info(const ActivationLayerInfo &activation_info) - { - _activation_info = activation_info; - } - /** Flag which specifies if the values of the weights tensor are constant throughout multiple executions or not - * - * @return True if the weights tensor is constant - */ - bool constant_weights() const - { - return _constant_weights; - }; - -private: - bool _is_a_reshaped; - bool _is_b_reshaped; - bool _reshape_b_only_on_first_run; - int _depth_output_gemm3d; - bool _reinterpret_input_as_3d; - bool _retain_internal_weights; - GEMMLowpOutputStageInfo _gemmlowp_output_stage; - bool _fp_mixed_precision; - bool _broadcast_bias; - bool _pretranpose_B; - ActivationLayerInfo _activation_info; - bool _constant_weights; -}; +class ITensorInfo; /** Winograd information */ struct WinogradInfo @@ -2137,16 +1937,23 @@ struct WinogradInfo * @param[in] conv_info Convolution info (Pads, strides) * @param[in] data_layout Data layout to use for the output tensor once the convolution has been applied */ - WinogradInfo(Size2D output_tile_sz, Size2D kernel_sz, Size2D input_dims, PadStrideInfo conv_info, DataLayout data_layout) - : output_tile_size(output_tile_sz), kernel_size(kernel_sz), input_dimensions(input_dims), convolution_info(conv_info), output_data_layout(data_layout) - { - } - - Size2D output_tile_size{}; /**< Width and height of the output tile */ - Size2D kernel_size{}; /**< Width and height of the kernel*/ - Size2D input_dimensions{}; /**< Width and height of the input tensor before the convolution is applied */ - PadStrideInfo convolution_info{}; /**< Convolution info (Pads, strides,...) */ - DataLayout output_data_layout{ DataLayout::NCHW }; /**< Data layout to use for the output tensor once the convolution has been applied (NCHW or NHWC) */ + WinogradInfo( + Size2D output_tile_sz, Size2D kernel_sz, Size2D input_dims, PadStrideInfo conv_info, DataLayout data_layout) + : output_tile_size(output_tile_sz), + kernel_size(kernel_sz), + input_dimensions(input_dims), + convolution_info(conv_info), + output_data_layout(data_layout) + { + } + + Size2D output_tile_size{}; /**< Width and height of the output tile */ + Size2D kernel_size{}; /**< Width and height of the kernel*/ + Size2D input_dimensions{}; /**< Width and height of the input tensor before the convolution is applied */ + PadStrideInfo convolution_info{}; /**< Convolution info (Pads, strides,...) */ + DataLayout output_data_layout{ + DataLayout:: + NCHW}; /**< Data layout to use for the output tensor once the convolution has been applied (NCHW or NHWC) */ }; /** IO formatting information class*/ @@ -2205,5 +2012,8 @@ struct IOFormatInfo /** Align columns */ bool align_columns; }; + +/** Class for holding information related to cropping */ +using CropInfo = Padding2D; } // namespace arm_compute -#endif /* ARM_COMPUTE_TYPES_H */ +#endif // ACL_ARM_COMPUTE_CORE_TYPES_H |