/* * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. * * SPDX-License-Identifier: Apache-2.0 * * Licensed under the Apache License, Version 2.0 (the License); you may * not use this file except in compliance with the License. * You may obtain a copy of the License at * * www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an AS IS BASIS, WITHOUT * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /* ---------------------------------------------------------------------- * Project: CMSIS NN Library * Title: arm_convolve_s8.c * Description: s8 version of convolution using symmetric quantization. * * $Date: May 29, 2020 * $Revision: V.2.0.1 * * Target Processor: Cortex-M cores * * -------------------------------------------------------------------- */ #include "cmsis/CMSIS/DSP/Include/arm_math.h" #include "cmsis/CMSIS/NN/Include/arm_nn_types.h" #include "cmsis/CMSIS/NN/Include/arm_nnfunctions.h" #include "cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" /** * @ingroup groupNN */ /** * @addtogroup NNConv * @{ */ /* * Basic s8 convolution function. * * Refer header file for details. Optimal use case for the DSP/MVE implementation is when input and output channels * are multiples of 4 or atleast greater than 4. * */ arm_status arm_convolve_s8(const cmsis_nn_context* ctx, const cmsis_nn_conv_params* conv_params, const cmsis_nn_per_channel_quant_params* quant_params, const cmsis_nn_dims* input_dims, const q7_t *input_data, const cmsis_nn_dims* filter_dims, const q7_t *filter_data, const cmsis_nn_dims* bias_dims, const int32_t *bias_data, const cmsis_nn_dims* output_dims, q7_t *output_data) { q15_t *buffer_a = (q15_t *)ctx->buf; const uint16_t input_batches = input_dims->n; const uint16_t input_x = input_dims->w; const uint16_t input_y = input_dims->h; const uint16_t input_ch = input_dims->c; const uint16_t kernel_x = filter_dims->w; const uint16_t kernel_y = filter_dims->h; const uint16_t output_x = output_dims->w; const uint16_t output_y = output_dims->h; const uint16_t output_ch = output_dims->c; const uint16_t pad_x = conv_params->padding.w; const uint16_t pad_y = conv_params->padding.h; const uint16_t stride_x = conv_params->stride.w; const uint16_t stride_y = conv_params->stride.h; const int32_t input_offset = conv_params->input_offset; const int32_t out_offset = conv_params->output_offset; const int32_t out_activation_min = conv_params->activation.min; const int32_t out_activation_max = conv_params->activation.max; int32_t *output_mult = quant_params->multiplier; int32_t *output_shift = quant_params->shift; int i_batch; for (i_batch = 0; i_batch < input_batches; i_batch++) { #if defined(ARM_MATH_MVEI) (void)bias_dims; /* Generate upto four columns from the input tensor a GEMM computation */ q7_t *im2col_buf = (q7_t *)buffer_a; q7_t *out = output_data; int32_t buffer_fill_cnt = 0; int32_t padded = 0; const int32_t num_elem = kernel_x * kernel_y * input_ch; /* This part implements the im2col function */ for (int i_out_y = 0; i_out_y < output_y; i_out_y++) { for (int i_out_x = 0; i_out_x < output_x; i_out_x++) { for (int i_ker_y = i_out_y * stride_y - pad_y; i_ker_y < i_out_y * stride_y - pad_y + kernel_y; i_ker_y++) { for (int i_ker_x = i_out_x * stride_x - pad_x; i_ker_x < i_out_x * stride_x - pad_x + kernel_x; i_ker_x++) { if (i_ker_y < 0 || i_ker_y >= input_y || i_ker_x < 0 || i_ker_x >= input_x) { memset(im2col_buf, (int8_t)-input_offset, sizeof(q7_t) * input_ch); padded = 1; } else { arm_memcpy_q7(im2col_buf, input_data + (i_ker_y * input_x + i_ker_x) * input_ch, input_ch); } im2col_buf += input_ch; } } buffer_fill_cnt++; /* Computation is filed for every 4 columns */ if (buffer_fill_cnt == 4 && (padded == 0)) { buffer_fill_cnt = 0; for (int i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) { int32_t sum_row; int32_t acc[4]; (void)arm_nn_mat_mul_core_4x_s8(num_elem, num_elem, (q7_t *)buffer_a, filter_data + num_elem * i_out_ch, &sum_row, acc); int32x4_t s_offset = vdupq_n_s32(sum_row); int32x4_t res = vldrwq_s32(acc); s_offset = vmulq_n_s32(s_offset, input_offset); res = vaddq_n_s32(res, bias_data[i_out_ch]); res = vaddq_s32(res, s_offset); res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]); res = vaddq_n_s32(res, out_offset); res = vmaxq_s32(res, vdupq_n_s32(out_activation_min)); res = vminq_s32(res, vdupq_n_s32(out_activation_max)); const uint32x4_t scatter_offset = {0, output_ch, output_ch * 2, output_ch * 3}; vstrbq_scatter_offset_s32(out, scatter_offset, res); out++; } out += (3 * output_ch); im2col_buf = (q7_t *)buffer_a; } else if (buffer_fill_cnt == 4 && (padded != 0)) { buffer_fill_cnt = 0; out = arm_nn_mat_mult_s8(filter_data, (q7_t *)buffer_a, output_ch, 4, output_shift, output_mult, out_offset, input_offset, 0, out_activation_min, out_activation_max, num_elem, bias_data, out); im2col_buf = (q7_t *)buffer_a; padded = 0; } } } /* Handle left over columns */ if (buffer_fill_cnt != 0) { out = arm_nn_mat_mult_s8(filter_data, (q7_t *)buffer_a, output_ch, buffer_fill_cnt, output_shift, output_mult, out_offset, input_offset, 0, out_activation_min, out_activation_max, num_elem, bias_data, out); } #elif defined(ARM_MATH_DSP) (void)bias_dims; int32_t i_out_y, i_out_x, i_ker_y, i_ker_x; /* Generate two columns from the input tensor a GEMM computation */ q15_t *two_column_buf = buffer_a; q7_t *out = output_data; /* This part implements the im2col function */ for (i_out_y = 0; i_out_y < output_y; i_out_y++) { for (i_out_x = 0; i_out_x < output_x; i_out_x++) { for (i_ker_y = i_out_y * stride_y - pad_y; i_ker_y < i_out_y * stride_y - pad_y + kernel_y; i_ker_y++) { for (i_ker_x = i_out_x * stride_x - pad_x; i_ker_x < i_out_x * stride_x - pad_x + kernel_x; i_ker_x++) { if (i_ker_y < 0 || i_ker_y >= input_y || i_ker_x < 0 || i_ker_x >= input_x) { /* Filling 0 for out-of-bound paddings */ memset(two_column_buf, 0, sizeof(q15_t) * input_ch); } else { /* Copying the pixel data to column */ arm_q7_to_q15_with_offset(input_data + (i_ker_y * input_x + i_ker_x) * input_ch, two_column_buf, input_ch, input_offset); } two_column_buf += input_ch; } } /* Computation is filed for every 2 columns */ if (two_column_buf == buffer_a + 2 * input_ch * kernel_y * kernel_x) { out = arm_nn_mat_mult_kernel_s8_s16(filter_data, buffer_a, output_ch, output_shift, output_mult, out_offset, out_activation_min, out_activation_max, input_ch * kernel_y * kernel_x, bias_data, out); /* counter reset */ two_column_buf = buffer_a; } } } /* left-over because odd number of output pixels */ if (two_column_buf != buffer_a) { const q7_t *ker_a = filter_data; int i; for (i = 0; i < output_ch; i++) { /* Load the accumulator with bias first */ q31_t sum = bias_data[i]; /* Point to the beginning of the im2col buffer where the input is available as a rearranged column */ const q15_t *ip_as_col = buffer_a; /* 4 multiply and accumulates are done in one loop. */ uint16_t col_count = (input_ch * kernel_y * kernel_x) >> 2; while (col_count) { q31_t ker_a1, ker_a2; q31_t ip_b1, ip_b2; ker_a = read_and_pad(ker_a, &ker_a1, &ker_a2); ip_b1 = arm_nn_read_q15x2_ia(&ip_as_col); sum = __SMLAD(ker_a1, ip_b1, sum); ip_b2 = arm_nn_read_q15x2_ia(&ip_as_col); sum = __SMLAD(ker_a2, ip_b2, sum); col_count--; } /* Handle left over mac */ col_count = input_ch * kernel_y * kernel_x & 0x3; while (col_count) { q7_t ker_a1 = *ker_a++; q15_t ip_b1 = *ip_as_col++; sum += ker_a1 * ip_b1; col_count--; } sum = arm_nn_requantize(sum, output_mult[i], output_shift[i]); sum += out_offset; sum = MAX(sum, out_activation_min); sum = MIN(sum, out_activation_max); *out++ = (q7_t)sum; } } #else /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ (void)buffer_a; int32_t i_out_ch, i_out_y, i_out_x, i_input_ch, i_ker_y, i_ker_x; int32_t conv_out; for (i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) { for (i_out_y = 0; i_out_y < output_y; i_out_y++) { for (i_out_x = 0; i_out_x < output_x; i_out_x++) { conv_out = bias_data[i_out_ch]; const int32_t base_idx_y = stride_y * i_out_y - pad_y; const int32_t base_idx_x = stride_x * i_out_x - pad_x; const int32_t ker_y_start = MAX(0, -base_idx_y); const int32_t ker_x_start = MAX(0, -base_idx_x); const int32_t ker_y_end = MIN(kernel_y, input_y - base_idx_y); const int32_t ker_x_end = MIN(kernel_x, input_x - base_idx_x); for (i_ker_y = ker_y_start; i_ker_y < ker_y_end; i_ker_y++) { for (i_ker_x = ker_x_start; i_ker_x < ker_x_end; i_ker_x++) { const int32_t in_row = base_idx_y + i_ker_y; const int32_t in_col = base_idx_x + i_ker_x; for (i_input_ch = 0; i_input_ch < input_ch; i_input_ch++) { conv_out += (input_data[(in_row * input_x + in_col) * input_ch + i_input_ch] + input_offset) * filter_data[i_out_ch * input_ch * kernel_y * kernel_x + (i_ker_y * kernel_x + i_ker_x) * input_ch + i_input_ch]; } } } conv_out = arm_nn_requantize(conv_out, output_mult[i_out_ch], output_shift[i_out_ch]); conv_out += out_offset; conv_out = MAX(conv_out, out_activation_min); conv_out = MIN(conv_out, out_activation_max); output_data[i_out_ch + (i_out_y * output_x + i_out_x) * output_ch] = (int8_t)conv_out; } } } #endif /* Advance to the next batch */ input_data += (input_x * input_y * input_ch); output_data += (output_x * output_y * output_ch); } /* Return to application */ return ARM_MATH_SUCCESS; } int32_t arm_convolve_s8_get_buffer_size(const cmsis_nn_dims* input_dims, const cmsis_nn_dims* filter_dims) { #if defined(ARM_MATH_DSP) return (2 * input_dims->c * filter_dims->w * filter_dims->h) * (int32_t)sizeof(int16_t); #else (void)input_dims; (void)filter_dims; return 0; #endif } /** * @} end of NNConv group */