Moritz Pflanzer | 69d3341 | 2017-08-09 11:45:15 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "FullyConnectedLayer.h" |
| 25 | |
Moritz Pflanzer | a09de0c | 2017-09-01 20:41:12 +0100 | [diff] [blame^] | 26 | #include "tests/validation/FixedPoint.h" |
| 27 | #include "tests/validation/half.h" |
Moritz Pflanzer | 69d3341 | 2017-08-09 11:45:15 +0100 | [diff] [blame] | 28 | |
| 29 | #include <numeric> |
| 30 | |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace test |
| 34 | { |
| 35 | namespace validation |
| 36 | { |
| 37 | namespace reference |
| 38 | { |
| 39 | namespace |
| 40 | { |
| 41 | // Vector matrix multiply for floating point |
| 42 | template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0> |
| 43 | void vector_matrix_multiply(const T *src, const T *weights, const T *bias, T *dst, int cols_weights, int rows_weights, uint8_t fixed_point_position) |
| 44 | { |
| 45 | ARM_COMPUTE_UNUSED(fixed_point_position); |
| 46 | |
| 47 | for(int y = 0; y < rows_weights; ++y) |
| 48 | { |
| 49 | dst[y] = std::inner_product(src, src + cols_weights, weights, static_cast<T>(0)) + bias[y]; |
| 50 | weights += cols_weights; |
| 51 | } |
| 52 | } |
| 53 | |
| 54 | // Vector matrix multiply for fixed point type |
| 55 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0> |
| 56 | void vector_matrix_multiply(const T *src, const T *weights, const T *bias, T *dst, int cols_weights, int rows_weights, uint8_t fixed_point_position) |
| 57 | { |
| 58 | using namespace fixed_point_arithmetic; |
| 59 | using promoted_type = fixed_point_arithmetic::traits::promote_t<T>; |
| 60 | |
| 61 | for(int y = 0; y < rows_weights; ++y) |
| 62 | { |
| 63 | // Reset accumulator |
| 64 | fixed_point<promoted_type> acc(0, fixed_point_position); |
| 65 | |
| 66 | for(int x = 0; x < cols_weights; ++x) |
| 67 | { |
| 68 | const fixed_point<promoted_type> i_value(src[x], fixed_point_position, true); |
| 69 | const fixed_point<promoted_type> w_value(weights[x], fixed_point_position, true); |
| 70 | acc = acc + i_value * w_value; |
| 71 | } |
| 72 | |
| 73 | // Get the bias |
| 74 | const fixed_point<T> b(bias[y], fixed_point_position, true); |
| 75 | |
| 76 | // Convert back and accumulate the bias |
| 77 | fixed_point<T> res(acc); |
| 78 | res = res + b; |
| 79 | |
| 80 | // Store the result |
| 81 | dst[y] = res.raw(); |
| 82 | |
| 83 | weights += cols_weights; |
| 84 | } |
| 85 | } |
| 86 | } // namespace |
| 87 | |
| 88 | template <typename T> |
| 89 | SimpleTensor<T> fully_connected_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<T> &bias, const TensorShape &dst_shape) |
| 90 | { |
| 91 | // Create reference |
| 92 | SimpleTensor<T> dst{ TensorShape{ dst_shape }, src.data_type(), 1, src.fixed_point_position() }; |
| 93 | |
| 94 | // Sanity checks |
| 95 | const int num_batch_dimensions = std::max(0, static_cast<int>(dst_shape.num_dimensions()) - 1); |
| 96 | const int num_input_dimensions = src.shape().num_dimensions() - num_batch_dimensions; |
| 97 | const unsigned int linear_input_size = src.shape().total_size_lower(num_input_dimensions); |
| 98 | |
| 99 | ARM_COMPUTE_UNUSED(num_batch_dimensions); |
| 100 | ARM_COMPUTE_UNUSED(num_input_dimensions); |
| 101 | ARM_COMPUTE_UNUSED(linear_input_size); |
| 102 | ARM_COMPUTE_ERROR_ON(weights.shape().x() != linear_input_size); |
| 103 | ARM_COMPUTE_ERROR_ON(weights.shape().y() != bias.shape().x()); |
| 104 | ARM_COMPUTE_ERROR_ON(weights.shape().y() != dst.shape().x()); |
| 105 | |
| 106 | // Compute reference |
| 107 | const int cols_weights = weights.shape().x(); |
| 108 | const int rows_weights = weights.shape().y(); |
| 109 | const int num_batches = dst_shape.total_size_upper(1); |
| 110 | |
| 111 | for(int k = 0; k < num_batches; ++k) |
| 112 | { |
| 113 | vector_matrix_multiply<T>(src.data() + k * cols_weights, |
| 114 | weights.data(), |
| 115 | bias.data(), |
| 116 | dst.data() + k * rows_weights, |
| 117 | cols_weights, |
| 118 | rows_weights, |
| 119 | src.fixed_point_position()); |
| 120 | } |
| 121 | |
| 122 | return dst; |
| 123 | } |
| 124 | |
| 125 | template SimpleTensor<float> fully_connected_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, const TensorShape &dst_shape); |
| 126 | template SimpleTensor<half_float::half> fully_connected_layer(const SimpleTensor<half_float::half> &src, const SimpleTensor<half_float::half> &weights, const SimpleTensor<half_float::half> &bias, |
| 127 | const TensorShape &dst_shape); |
| 128 | template SimpleTensor<qint8_t> fully_connected_layer(const SimpleTensor<qint8_t> &src, const SimpleTensor<qint8_t> &weights, const SimpleTensor<qint8_t> &bias, const TensorShape &dst_shape); |
| 129 | template SimpleTensor<qint16_t> fully_connected_layer(const SimpleTensor<qint16_t> &src, const SimpleTensor<qint16_t> &weights, const SimpleTensor<qint16_t> &bias, const TensorShape &dst_shape); |
| 130 | } // namespace reference |
| 131 | } // namespace validation |
| 132 | } // namespace test |
| 133 | } // namespace arm_compute |