diff options
Diffstat (limited to 'tests/validation/Helpers.cpp')
-rw-r--r-- | tests/validation/Helpers.cpp | 326 |
1 files changed, 249 insertions, 77 deletions
diff --git a/tests/validation/Helpers.cpp b/tests/validation/Helpers.cpp index eb8bdcf5a7..560460fd33 100644 --- a/tests/validation/Helpers.cpp +++ b/tests/validation/Helpers.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,9 +22,12 @@ * SOFTWARE. */ #include "tests/validation/Helpers.h" +#include "tests/framework/Asserts.h" #include <algorithm> #include <cmath> +#include <cstdint> +#include <tuple> namespace arm_compute { @@ -32,82 +35,6 @@ namespace test { namespace validation { -void fill_mask_from_pattern(uint8_t *mask, int cols, int rows, MatrixPattern pattern) -{ - unsigned int v = 0; - std::mt19937 gen(library->seed()); - std::bernoulli_distribution dist(0.5); - - for(int r = 0; r < rows; ++r) - { - for(int c = 0; c < cols; ++c, ++v) - { - uint8_t val = 0; - - switch(pattern) - { - case MatrixPattern::BOX: - val = 255; - break; - case MatrixPattern::CROSS: - val = ((r == (rows / 2)) || (c == (cols / 2))) ? 255 : 0; - break; - case MatrixPattern::DISK: - val = (((r - rows / 2.0f + 0.5f) * (r - rows / 2.0f + 0.5f)) / ((rows / 2.0f) * (rows / 2.0f)) + ((c - cols / 2.0f + 0.5f) * (c - cols / 2.0f + 0.5f)) / ((cols / 2.0f) * - (cols / 2.0f))) <= 1.0f ? 255 : 0; - break; - case MatrixPattern::OTHER: - val = (dist(gen) ? 0 : 255); - break; - default: - return; - } - - mask[v] = val; - } - } - - if(pattern == MatrixPattern::OTHER) - { - std::uniform_int_distribution<uint8_t> distribution_u8(0, ((cols * rows) - 1)); - mask[distribution_u8(gen)] = 255; - } -} - -HarrisCornersParameters harris_corners_parameters() -{ - HarrisCornersParameters params; - - std::mt19937 gen(library->seed()); - std::uniform_real_distribution<float> threshold_dist(0.f, 0.001f); - std::uniform_real_distribution<float> sensitivity(0.04f, 0.15f); - std::uniform_real_distribution<float> euclidean_distance(0.f, 30.f); - std::uniform_int_distribution<uint8_t> int_dist(0, 255); - - params.threshold = threshold_dist(gen); - params.sensitivity = sensitivity(gen); - params.min_dist = euclidean_distance(gen); - params.constant_border_value = int_dist(gen); - - return params; -} - -CannyEdgeParameters canny_edge_parameters() -{ - CannyEdgeParameters params; - - std::mt19937 gen(library->seed()); - std::uniform_int_distribution<uint8_t> int_dist(0, 255); - std::uniform_int_distribution<uint8_t> threshold_dist(2, 255); - - params.constant_border_value = int_dist(gen); - params.upper_thresh = threshold_dist(gen); // upper_threshold >= 2 - threshold_dist = std::uniform_int_distribution<uint8_t>(1, params.upper_thresh - 1); - params.lower_thresh = threshold_dist(gen); // lower_threshold >= 1 && lower_threshold < upper_threshold - - return params; -} - template <> SimpleTensor<float> convert_from_asymmetric(const SimpleTensor<uint8_t> &src) { @@ -401,6 +328,249 @@ std::pair<int, int> get_symm_quantized_per_channel_bounds(const QuantizationInfo return std::pair<int, int> { min_bound, max_bound }; } +void add_padding_x(std::initializer_list<ITensor *> tensors, const DataLayout &data_layout, bool only_right_pad) +{ + if(data_layout == DataLayout::NHWC) + { + constexpr unsigned int lower = 1U; + constexpr unsigned int upper = 16U; + + std::uniform_int_distribution<unsigned int> distribution(lower, upper); + size_t seed_offset = 0; + + for(ITensor *tensor : tensors) + { + ARM_COMPUTE_ERROR_ON(!tensor->info()->is_resizable()); + + std::mt19937 gen(library->seed() + seed_offset++); + + const unsigned int right = distribution(gen); + const unsigned int left = only_right_pad ? 0 : distribution(gen); + + tensor->info()->extend_padding(PaddingSize(0U, right, 0U, left)); + } + } +} + +QuantizationHint suggest_conv_dst_q_info_and_bias(const QuantizationInfo &in_q_info, + const QuantizationInfo &weight_q_info, + int32_t height, + int32_t width, + int32_t channels, + DataType data_type, + float bias_fraction) +{ + /** Quantization Setup of convolution + * + * Just like any other multiply-accummulate, convolution (2D) operation + * multiplies and accumulates the input and weight tensors. This operation + * takes place in three dimensions: height, width and channels. All of them + * belong to the weight tensor. + * + * The formula for simple convolution can be written as: + * C = sum_h sum_w sum_c(I[h_offset + h, w_offset + w, c] * W[h, w, c]) + * + * Here, h_offset and w_offset are the starting positions in the image. Effects + * of paddings are ignored. This accumulation reduces to something like + * + * C = sum_m(I_index * W_hwc) + * where m is height x width x channels. + * + * Non-unit strides and/or dilations do not change the probabilistic nature of + * this sum because we always iterate as the size of the weight tensor. + * + * Paddings may affect this summation, but it's a boundary condition and so is + * neglected for brevity. + */ + + return suggest_mac_dst_q_info_and_bias(in_q_info, weight_q_info, height * width * channels, data_type, bias_fraction); +} + +QuantizationHint suggest_matmul_dst_q_info_and_bias(const QuantizationInfo &lhs_q_info, + const QuantizationInfo &rhs_q_info, + int32_t m, int32_t n, int32_t k, DataType data_type, + float bias_fraction) +{ + ARM_COMPUTE_UNUSED(m, n); + + /** Quantization Setup of matrix multiplication + * + * We have a matrix multiplication of the form C = A * B + D + * where A is (m X k), B is (k x n) and C is therefore (m x n). + * The bias, D is (1 x n). + * + * If we have some distributional statistics of A, B and D, i.e. mean and variance, + * we can estimate the mean and variance of a single value in C matrix and pick + * good scale and offset values for the output and have non-saturated tests. + * + * Each element in the output matrix can be calculated as follows: + * C_ij = sum_k(A_ik * B_kj) + D_j + * + * Note: All possible A_ik, B_kj, D_j random variables are assumed mutually independent. + * Note: In quantized operators, bias is an integer. But, its quantization scale is + * assumed to be equal to lhs_scale * rhs_scale, and offset equal to 0. + * Note: Since, bias is an integer that should be given as input, we need to pick responsible + * values when adding it on top of the summation. This is where "bias_fraction" comes + * into play. Based on the fraction given, we also return suggested bias range (min/max) + * for not saturating the output. + * + * Because all random variables are mutually independent, any C_ij has the same statistics, + * which is why we return a single destination quantization info object; which is why we can + * resort to a more general calculation explained in suggest_mac_dst_q_info_and_bias(). + * + * From a probabilistic perspective, the above calculation reduces to + * c = sum_k (a_k * b_k) + d + */ + + return suggest_mac_dst_q_info_and_bias(lhs_q_info, rhs_q_info, k, data_type, bias_fraction); +} + +QuantizationHint suggest_mac_dst_q_info_and_bias( + const QuantizationInfo &a_q_info, const QuantizationInfo &b_q_info, int32_t K, DataType data_type, float bias_fraction, int num_sd) +{ + QuantizationInfo c_q_info; + + ARM_COMPUTE_ASSERT(data_type == DataType::QASYMM8 || data_type == DataType::QASYMM8_SIGNED); + + const int32_t t_max = static_cast<int32_t>(data_type == DataType::QASYMM8 ? std::numeric_limits<uint8_t>::max() : std::numeric_limits<int8_t>::max()); + const int32_t t_min = static_cast<int32_t>(data_type == DataType::QASYMM8 ? std::numeric_limits<uint8_t>::min() : std::numeric_limits<int8_t>::min()); + + /** Quantization Setup of multiply-accummulate + * + * Expression (in float): + * C = sum_k ( A_k * B_k ) + D + * + * Lemma: An affine transformation (i.e. aX + b) to a discrete uniform random variable + * creates another discrete uniform random variable. + * + * Terminology: + * E[X]: Mean of the random variable X (sometimes referred as mu_x) + * var(X): Variance of the random variable X (someimes referred as sigma^2_x) + * std(X): sqrt(var(X)), standard deviation of X + * + * 1) Calculate the mean: + * E[C] = sum_k( E[A_k] * E[B_k] ) + D = K * mean_a * mean_b + mean_d + * + * Since elements of A and B are uniformly distributed random variables, we have + * mean_a = (max_a + min_a) / 2, mean_b = (max_b + min_b ) / 2 + * max_a and min_a can be calculated with the scale_a/b and offset_a/b + * by replacing data type minimum and maximums in the equations + * + * We don't know mean_d because we have to choose it based on bias_fraction. If we call + * the summation as M_int, similar to above, we have: + * + * E[C_int] = sum_k( E[A_k_int] * E[B_k_int] ) + E[D_int] = K * mean_a_int * mean_b_int + mean_d_int + * \___________________________/ + * E[M_int] + * + * We choose a bias mean proportional to the integer summation. This proportion is "bias_fraction". + * So, we have D_int = f * M_int (f: fraction), and + * E[D_int] = mean_d_int = f * E[M_int] + * + * This also means, for floating point value of D, the following: + * E[D] = mean_d = E[D_int] * a_scale * b_scale + * + * 2) Calculate the variance: + * var(C) = sum_k( var(A_k * B_k) ) + var(D) + * = sum_k ( E[A_k^2 * B_k^2] - E[A_k]^2E[B_k^2] ) + * = ... + * = K * (var_a * var_b + var_a * mean^2_b + var_b * mean^2_a) + var_d + * + * Similarly, due to uniform random variable properties, we have + * var_a = (max_a - min_a)^2 / 12 + * var_b = (max_b - min_b)^2 / 12 + * + * Again, we don't know var_d as we don't know the bias. As set out in the previous section, we have + * var(D_int) = var(f * M_int) = f^2 * var(M_int) + * + * Using the same expression, we can find var(M_int): + * var(C_int) = sum_k( var(A_k_int * B_k_int) ) + var(D_int) + * = sum_k ( E[A_k_int^2 * B_k_int^2] - E[A_k_int]^2E[B_k_int^2] ) + * = ... + * = K * (var_a_int * var_b_int + var_a_int * mean^2_b_int + var_b_int * mean^2_a_int) + var_d_int + * \_______________________________________________________________________________/ + * var(M_int) + * + * Now, we know mean and variance of D_int, we can return a suitable bias range with + * [mean_d_int +/- 2 * std_d_int] + * + * This also means, for floating point value of D, the following: + * var(D) = var_d = var(D_int) * a_scale^2 * b_scale^2 + * + * E[D] and var(D) calculated in steps (1) and (2) can be substituted into E[C] and var(C) calculatons. + * + * 3) Now, we have an idea of what would an average C will look like and how much deviation + * is present around it. The exact distribution of C is difficult to come up with dependent on K. + * But, as K increases, due to Central Limit Theorem, it'll look more like a bell shaped figure, + * approaching normal distribution. + * + * This is useful because, in normal distribution, we know that values +- 2 std_deviation around + * the mean constitute 95% of the values. Therefore, setting a plausible range for us: + * C_range = [C_min, C_max] = [mean_c - 2 * std_c, mean_c + 2 * std_c] + * + * 4) + * If we map this [C_min, C_max] to [0, 255] or [-128, 127] depending on the signedness of the + * data type, we can find a suitable scale and offset for the output. On average, it's expected + * that 5% of the output values will saturate and 95% will remain in the range. + * + * The equations to be solved for offset_c and scale_c are: + * C_min = scale_c * (type_min - offset_c) + * C_max = scale_c * (type_max - offset_c) + */ + + const int32_t a_offset = a_q_info.uniform().offset; + const float a_scale = a_q_info.uniform().scale; + const int32_t b_offset = b_q_info.uniform().offset; + const float b_scale = b_q_info.uniform().scale; + + // Integer value statistics. Valid for both Lhs/A and Rhs/B + const float mean_a_int = (t_max + t_min) / 2.f; + constexpr float var_a_int = (256 * 256 - 1) / 12.f; // Discrete uniform RV variance + const float mean_b_int = mean_a_int; // A_int and B_int has the same stats + constexpr float var_b_int = var_a_int; + + // Lhs/A stats + const float max_a = (t_max - a_offset) * a_scale; + const float min_a = (t_min - a_offset) * a_scale; + const float mean_a = (max_a + min_a) / 2; + const float var_a = (max_a - min_a) * (max_a - min_a) / 12; + + // Rhs/B stats + const float max_b = (t_max - b_offset) * b_scale; + const float min_b = (t_min - b_offset) * b_scale; + const float mean_b = (max_b + min_b) / 2; + const float var_b = (max_b - min_b) * (max_b - min_b) / 12; + + // Integer multiplication output/M stats + const float mean_m_int = K * mean_a_int * mean_b_int; + const float var_m_int = K * (var_a_int * var_b_int + mean_a_int * var_b_int + mean_b_int + var_a_int); + const float std_m_int = sqrt(var_m_int); + + // Bias/D both Int and Float statistics + const float mean_d_int = bias_fraction * mean_m_int; + const float std_d_int = bias_fraction * std_m_int; + const float mean_d = a_scale * b_scale * mean_d_int; + const float std_d = a_scale * b_scale * std_d_int; + const float var_d = std_d * std_d; + + // Also calculate the suggested bias range + const int32_t min_bias = mean_d_int - (num_sd * std_d_int); + const int32_t max_bias = mean_d_int + (num_sd * std_d_int); + + // Output/C stats + const float mean_out = K * mean_a * mean_b + mean_d; + const float var_out = K * (var_a * var_b + var_a * mean_b * mean_b + var_b * mean_a * mean_a) + var_d; + const float std_out = sqrt(var_out); + + // Output quantization setup + const float scale_out = (2 * num_sd) * std_out / 255; + const int32_t offset_out = static_cast<int32_t>(t_min - (mean_out - (num_sd * std_out)) / scale_out); + + c_q_info = QuantizationInfo(scale_out, offset_out); + + return { c_q_info, min_bias, max_bias }; +} + template void get_tile(const SimpleTensor<float> &in, SimpleTensor<float> &roi, const Coordinates &coord); template void get_tile(const SimpleTensor<half> &in, SimpleTensor<half> &roi, const Coordinates &coord); template void get_tile(const SimpleTensor<int> &in, SimpleTensor<int> &roi, const Coordinates &coord); @@ -413,6 +583,8 @@ template void transpose_matrix(const SimpleTensor<half> &in, SimpleTensor<half> template void transpose_matrix(const SimpleTensor<int> &in, SimpleTensor<int> &out); template void transpose_matrix(const SimpleTensor<short> &in, SimpleTensor<short> &out); template void transpose_matrix(const SimpleTensor<char> &in, SimpleTensor<char> &out); +template void transpose_matrix(const SimpleTensor<int8_t> &in, SimpleTensor<int8_t> &out); +template void transpose_matrix(const SimpleTensor<uint8_t> &in, SimpleTensor<uint8_t> &out); template void matrix_multiply(const SimpleTensor<float> &a, const SimpleTensor<float> &b, SimpleTensor<float> &out); template void matrix_multiply(const SimpleTensor<half> &a, const SimpleTensor<half> &b, SimpleTensor<half> &out); |