Michele Di Giorgio | d02d5ed | 2021-01-22 09:47:04 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 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 | |
| 25 | #pragma once |
| 26 | |
| 27 | #include "depthwise_depthfirst_multiplier.hpp" |
| 28 | |
| 29 | namespace arm_conv { |
| 30 | namespace depthwise { |
| 31 | |
| 32 | template <class strategy> |
| 33 | class DepthwiseDepthfirstWithMultiplierQuantized : |
| 34 | public DepthwiseCommon<typename strategy::input_type, |
| 35 | typename strategy::weight_type, |
| 36 | typename strategy::return_type> |
| 37 | { |
| 38 | using Parent = DepthwiseCommon<typename strategy::input_type, |
| 39 | typename strategy::weight_type, |
| 40 | typename strategy::return_type>; |
| 41 | using TInput = typename strategy::input_type; |
| 42 | using TWeight = typename strategy::weight_type; |
| 43 | using TOutput = typename strategy::return_type; |
| 44 | |
| 45 | const arm_gemm::Requantize32 m_qp; |
| 46 | |
| 47 | size_t sizeof_output_buffer(unsigned int n_channels) const |
| 48 | { |
| 49 | const unsigned int vl = arm_gemm::utils::get_vector_length<typename strategy::return_type>(strategy::vl_type); |
| 50 | const auto rounded_channels = arm_gemm::roundup(n_channels, vl); |
| 51 | return sizeof(typename strategy::return_type) * rounded_channels; |
| 52 | } |
| 53 | |
| 54 | public: |
| 55 | DepthwiseDepthfirstWithMultiplierQuantized(const DepthwiseArgs &args, const arm_gemm::Requantize32 &qp) |
| 56 | : Parent(args), m_qp(qp) |
| 57 | { |
| 58 | } |
| 59 | |
| 60 | DepthwiseDepthfirstWithMultiplierQuantized(DepthwiseDepthfirstWithMultiplierQuantized &) = delete; |
| 61 | DepthwiseDepthfirstWithMultiplierQuantized &operator=(DepthwiseDepthfirstWithMultiplierQuantized &) = delete; |
| 62 | |
| 63 | size_t get_storage_size(void) const override |
| 64 | { |
| 65 | // We produce VL<int32_t> channels at a time, for each of these blocks of |
| 66 | // channels we store a vector of biases, weights (complicated) and |
| 67 | // requantize parameters. |
| 68 | const unsigned int iter_length = |
| 69 | arm_gemm::utils::get_vector_length<int32_t>(strategy::vl_type); |
| 70 | const unsigned int n_iters = |
| 71 | this->m_args.input_channels * arm_gemm::iceildiv(this->m_args.channel_multiplier, iter_length); |
| 72 | |
| 73 | // Compute the cost of storing the weights |
| 74 | const unsigned int n_dots_per_kernel_row = arm_gemm::iceildiv(strategy::kernel_cols, 4u); |
| 75 | |
| 76 | return n_iters * iter_length * ( |
| 77 | sizeof(int32_t) + // Bias |
| 78 | 4 * n_dots_per_kernel_row * strategy::kernel_rows * sizeof(TWeight) + // Weights |
| 79 | 2 * sizeof(int32_t) // Requantisation parameters |
| 80 | ); |
| 81 | } |
| 82 | |
| 83 | // We'll want an optimised version of this, but for now a C++ implementation |
| 84 | // is probably sufficient. |
| 85 | void pack_parameters(void *_buffer, const void *_biases, const void *_weights, size_t ld_weight_col, size_t ld_weight_row) override |
| 86 | { |
| 87 | auto buffer = static_cast<uint8_t *>(_buffer); |
| 88 | auto biases = static_cast<const int32_t *>(_biases); |
| 89 | auto weights = static_cast<const TWeight *>(_weights); |
| 90 | auto requant_muls = m_qp.per_channel_muls; |
| 91 | auto requant_shifts = m_qp.per_channel_right_shifts; |
| 92 | |
| 93 | const unsigned int iter_length = |
| 94 | arm_gemm::utils::get_vector_length<int32_t>(strategy::vl_type); |
| 95 | const unsigned int n_iters_per_input_channel = |
| 96 | arm_gemm::iceildiv(this->m_args.channel_multiplier, iter_length); |
| 97 | |
| 98 | const unsigned int n_dots_per_kernel_row = arm_gemm::iceildiv(strategy::kernel_cols, 4u); |
| 99 | |
| 100 | const size_t iter_stride = iter_length * ( |
| 101 | sizeof(int32_t) + // Bias |
| 102 | 4 * n_dots_per_kernel_row * strategy::kernel_rows * sizeof(int8_t) + // Weights |
| 103 | 2 * sizeof(int32_t) // Requantisation parameters |
| 104 | ); |
| 105 | |
| 106 | ld_weight_col = (ld_weight_col == 0) ? this->m_args.input_channels * this->m_args.channel_multiplier : ld_weight_col; |
| 107 | ld_weight_row = (ld_weight_row == 0) ? this->m_args.kernel_cols * ld_weight_col : ld_weight_row; |
| 108 | |
| 109 | for (unsigned int input_channel = 0; input_channel < this->m_args.input_channels; input_channel++) |
| 110 | { |
| 111 | auto buffer_input_channel = buffer + input_channel * n_iters_per_input_channel * iter_stride; |
| 112 | auto weights_input_channel = weights + input_channel * this->m_args.channel_multiplier; |
| 113 | |
| 114 | for (unsigned int iter = 0; iter < n_iters_per_input_channel; iter++) |
| 115 | { |
| 116 | // Get a pointer to the start of this portion of the buffer; consequently |
| 117 | // derive pointers to the bias, weight and requantisation portions of |
| 118 | // this frame. |
| 119 | auto buffer_base = buffer_input_channel + iter_stride * iter; |
| 120 | auto buffer_biases = reinterpret_cast<int32_t *>(buffer_base); |
| 121 | auto buffer_weights = buffer_base + sizeof(int32_t) * iter_length; |
| 122 | auto buffer_requant_mul = reinterpret_cast<int32_t *>( |
| 123 | buffer_weights + strategy::kernel_rows * n_dots_per_kernel_row * 4 * iter_length); |
| 124 | auto buffer_requant_shift = buffer_requant_mul + iter_length; |
| 125 | auto weights_base = weights_input_channel + iter * iter_length; |
| 126 | |
| 127 | // Hence work through the data for this iteration, on a |
| 128 | // channel-by-channel basis. |
| 129 | const auto this_iter_length = std::min<unsigned int>( |
| 130 | iter_length, this->m_args.channel_multiplier - iter * iter_length |
| 131 | ); |
| 132 | for (unsigned int i = 0; i < this_iter_length; i++) |
| 133 | { |
| 134 | auto weights_channel = weights_base + i; |
| 135 | |
| 136 | // Read the bias value, we modify this as we read the weights. |
| 137 | auto bias_value = biases == nullptr ? 0 : *(biases++); |
| 138 | int32_t elements_sum = 0; |
| 139 | |
| 140 | // Read through the kernel; for each row, marshal together as many dot |
| 141 | // product terms as are required. |
| 142 | for (unsigned int ki = 0; ki < strategy::kernel_rows; ki++) |
| 143 | { |
| 144 | auto buffer_row = buffer_weights + i*4 + ki * 4 * n_dots_per_kernel_row * iter_length; |
| 145 | auto weights_row = weights_channel + ki * ld_weight_row; |
| 146 | |
| 147 | unsigned int kj = 0; |
| 148 | for (; kj < strategy::kernel_cols; kj++) |
| 149 | { |
| 150 | // Determine which element to which we're writing |
| 151 | const auto dot = kj / 4; |
| 152 | const auto elem = kj % 4; |
| 153 | |
| 154 | // Copy the value; include in the sum |
| 155 | const auto val = weights_row[kj * ld_weight_col]; |
| 156 | buffer_row[dot * 4 * iter_length + elem] = val; |
| 157 | elements_sum += val; |
| 158 | } |
| 159 | for (; kj < 4 * n_dots_per_kernel_row; kj++) |
| 160 | { |
| 161 | const auto dot = kj / 4; |
| 162 | const auto elem = kj % 4; |
| 163 | buffer_row[dot * 4 * iter_length + elem] = 0; |
| 164 | } |
| 165 | |
| 166 | buffer_row += 4 * n_dots_per_kernel_row * iter_length; |
| 167 | } |
| 168 | |
| 169 | // Write back the bias and offset values |
| 170 | *(buffer_biases++) = |
| 171 | bias_value - m_qp.a_offset * elements_sum + |
| 172 | strategy::kernel_rows * strategy::kernel_cols * m_qp.a_offset * m_qp.b_offset; |
| 173 | |
| 174 | // Write out the requantisation parameters |
| 175 | *(buffer_requant_mul++) = m_qp.per_channel_requant ? *(requant_muls++) : m_qp.per_layer_mul; |
| 176 | *(buffer_requant_shift++) = m_qp.per_channel_requant ? *(requant_shifts++) : m_qp.per_layer_right_shift; |
| 177 | } |
| 178 | } |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | size_t get_working_size(const unsigned int n_threads, const unsigned int n_channels) const override |
| 183 | { |
| 184 | const unsigned int n_output_channels = n_channels * this->m_args.channel_multiplier; |
| 185 | return n_threads * sizeof_output_buffer(n_output_channels); |
| 186 | } |
| 187 | |
| 188 | using Parent::execute; |
| 189 | void execute( |
| 190 | const unsigned int batches, |
| 191 | const unsigned int input_height, |
| 192 | const unsigned int input_width, |
| 193 | const unsigned int input_channels, |
| 194 | const PaddingValues &padding, |
| 195 | const void *const _input, |
| 196 | const size_t ld_input_col, |
| 197 | const size_t ld_input_row, |
| 198 | const size_t ld_input_batch, |
| 199 | const void *const parameters, |
| 200 | const unsigned int output_height, |
| 201 | const unsigned int output_width, |
| 202 | void *const _output, |
| 203 | const size_t ld_output_col, |
| 204 | const size_t ld_output_row, |
| 205 | const size_t ld_output_batch, |
| 206 | void *const _working_space, |
| 207 | const unsigned int thread_id, |
| 208 | const unsigned int n_threads |
| 209 | ) const override |
| 210 | { |
| 211 | strategy strat(this->m_args.cpu_info); |
| 212 | #ifdef CYCLE_PROFILING |
| 213 | arm_gemm::profiler prof; |
| 214 | #endif |
| 215 | |
| 216 | auto executefn = [strat, this] ( |
| 217 | const TInput *const *const inptrs, |
| 218 | TOutput *const *const outptr_array, |
| 219 | const void *const params |
| 220 | ) { |
| 221 | strat.kernel(inptrs, outptr_array, params, this->m_args.channel_multiplier, m_qp); |
| 222 | }; |
| 223 | |
| 224 | // Get working space for this thread |
| 225 | uint8_t *const working_space = static_cast<uint8_t *>(_working_space) + get_working_size(1, input_channels) * thread_id; |
| 226 | |
| 227 | // Determine the stride across blocks of parameters |
| 228 | const unsigned int iter_length = |
| 229 | arm_gemm::utils::get_vector_length<int32_t>(strategy::vl_type); |
| 230 | const unsigned int n_iters_per_input_channel = arm_gemm::iceildiv(this->m_args.channel_multiplier, iter_length); |
| 231 | const unsigned int n_dots_per_kernel_row = arm_gemm::iceildiv(strategy::kernel_cols, 4u); |
| 232 | const size_t param_stride = n_iters_per_input_channel * iter_length * ( |
| 233 | sizeof(int32_t) + // Bias |
| 234 | 4 * n_dots_per_kernel_row * strategy::kernel_rows * sizeof(int8_t) + // Weights |
| 235 | 2 * sizeof(int32_t) // Requantisation parameters |
| 236 | ); |
| 237 | |
| 238 | common::depthwise_multiplier_execute<strategy>( |
| 239 | executefn, m_qp.a_offset, this->m_args, |
| 240 | batches, input_height, input_width, input_channels, padding, |
| 241 | _input, ld_input_col, ld_input_row, ld_input_batch, |
| 242 | parameters, param_stride, |
| 243 | output_height, output_width, |
| 244 | _output, ld_output_col, ld_output_row, ld_output_batch, |
| 245 | working_space, thread_id, n_threads |
| 246 | ); |
| 247 | } |
| 248 | }; |
| 249 | |
| 250 | } // namespace depthwise |
| 251 | } // namespace arm_conv |