CMSIS_5/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8.c

/*
 * Copyright (C) 2010-2021 Arm Limited or its affiliates.
 *
 * 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_depthwise_conv_s8.c
 * Description:  s8 version of depthwise convolution.
 *
 * $Date:        30. Dec 2021
 * $Revision:    V.2.7.1
 *
 * Target Processor:  Cortex-M CPUs
 *
 * -------------------------------------------------------------------- */

#include "arm_nnfunctions.h"
#include "arm_nnsupportfunctions.h"

/**
 *  @ingroup groupNN
 */

/**
 * @addtogroup NNConv
 * @{
 */

static void depthwise_conv_s8_mult_4(const int8_t *input,
                                     const int32_t input_x,
                                     const int32_t input_y,
                                     const int32_t input_ch,
                                     const int8_t *kernel,
                                     const int32_t output_ch,
                                     const int32_t ch_mult,
                                     const int32_t kernel_x,
                                     const int32_t kernel_y,
                                     const int32_t pad_x,
                                     const int32_t pad_y,
                                     const int32_t stride_x,
                                     const int32_t stride_y,
                                     const int32_t *bias,
                                     int8_t *output,
                                     const int32_t *output_shift,
                                     const int32_t *output_mult,
                                     const int32_t output_x,
                                     const int32_t output_y,
                                     const int32_t output_offset,
                                     const int32_t input_offset,
                                     const int32_t output_activation_min,
                                     const int32_t output_activation_max)
{
    for (int32_t in_h = -pad_y, out_h = 0, out_idx = 0; out_h < output_y; in_h += stride_y, ++out_h)
    {
        for (int32_t in_w = -pad_x, out_w = 0, ker_h_start = MAX(0, -in_h); out_w < output_x; in_w += stride_x, ++out_w)
        {
            for (int32_t in_ch = 0, out_ch = 0, ker_w_start = MAX(0, -in_w); out_ch < output_ch;
                 ++in_ch, out_ch += ch_mult)
            {
                for (int mult_tile = 0; mult_tile < ch_mult; mult_tile += 4)
                {
                    int32_t out_buff[4] = {0, 0, 0, 0};
                    if (bias)
                    {
                        out_buff[0] = bias[out_ch + 0 + mult_tile];
                        out_buff[1] = bias[out_ch + 1 + mult_tile];
                        out_buff[2] = bias[out_ch + 2 + mult_tile];
                        out_buff[3] = bias[out_ch + 3 + mult_tile];
                    }

                    for (int32_t ker_h = ker_h_start; ker_h < MIN(kernel_y, input_y - in_h); ++ker_h)
                    {
                        int32_t ker_idx = ker_h * (output_ch * kernel_x) + ker_w_start * output_ch + out_ch;
                        int32_t in_idx = (in_h + ker_h) * (input_ch * input_x) + in_w * input_ch + in_ch;
#if defined(__ARMCC_VERSION) && (__ARMCC_VERSION >= 6010050)
#pragma clang loop unroll(disable)
#endif
                        for (int32_t ker_w = ker_w_start; ker_w < MIN(kernel_x, input_x - in_w);
                             ++ker_w, ker_idx += output_ch)
                        {
                            int32_t in_val = input[in_idx + ker_w * input_ch] + input_offset;
                            out_buff[0] += in_val * kernel[ker_idx + 0 + mult_tile];
                            out_buff[1] += in_val * kernel[ker_idx + 1 + mult_tile];
                            out_buff[2] += in_val * kernel[ker_idx + 2 + mult_tile];
                            out_buff[3] += in_val * kernel[ker_idx + 3 + mult_tile];
                        }
                    }
#if defined(ARM_MATH_MVEI)
                    (void)out_idx;
                    int32x4_t res = vldrwq_s32(out_buff);
                    res = arm_requantize_mve_32x4(res,
                                                  vldrwq_s32(&output_mult[out_ch + mult_tile]),
                                                  vldrwq_s32(&output_shift[out_ch + mult_tile]));
                    res = vaddq_n_s32(res, output_offset);

                    res = vmaxq_s32(res, vdupq_n_s32(output_activation_min));
                    res = vminq_s32(res, vdupq_n_s32(output_activation_max));
                    vstrbq_s32(output, res);
                    output += 4;
#else
                    out_buff[0] = arm_nn_requantize(
                        out_buff[0], output_mult[out_ch + 0 + mult_tile], output_shift[out_ch + 0 + mult_tile]);
                    out_buff[1] = arm_nn_requantize(
                        out_buff[1], output_mult[out_ch + 1 + mult_tile], output_shift[out_ch + 1 + mult_tile]);
                    out_buff[2] = arm_nn_requantize(
                        out_buff[2], output_mult[out_ch + 2 + mult_tile], output_shift[out_ch + 2 + mult_tile]);
                    out_buff[3] = arm_nn_requantize(
                        out_buff[3], output_mult[out_ch + 3 + mult_tile], output_shift[out_ch + 3 + mult_tile]);

                    out_buff[0] += output_offset;
                    out_buff[1] += output_offset;
                    out_buff[2] += output_offset;
                    out_buff[3] += output_offset;

                    out_buff[0] = MIN(MAX(out_buff[0], output_activation_min), output_activation_max);
                    out_buff[1] = MIN(MAX(out_buff[1], output_activation_min), output_activation_max);
                    out_buff[2] = MIN(MAX(out_buff[2], output_activation_min), output_activation_max);
                    out_buff[3] = MIN(MAX(out_buff[3], output_activation_min), output_activation_max);

                    output[out_idx++] = (int8_t)out_buff[0];
                    output[out_idx++] = (int8_t)out_buff[1];
                    output[out_idx++] = (int8_t)out_buff[2];
                    output[out_idx++] = (int8_t)out_buff[3];

#endif
                }
            }
        }
    }
}

static void depthwise_conv_s8_generic(const q7_t *input,
                                      const uint16_t input_batches,
                                      const uint16_t input_x,
                                      const uint16_t input_y,
                                      const uint16_t input_ch,
                                      const q7_t *kernel,
                                      const uint16_t output_ch,
                                      const uint16_t ch_mult,
                                      const uint16_t kernel_x,
                                      const uint16_t kernel_y,
                                      const uint16_t pad_x,
                                      const uint16_t pad_y,
                                      const uint16_t stride_x,
                                      const uint16_t stride_y,
                                      const int32_t *bias,
                                      q7_t *output,
                                      const int32_t *output_shift,
                                      const int32_t *output_mult,
                                      const uint16_t output_x,
                                      const uint16_t output_y,
                                      const int32_t output_offset,
                                      const int32_t input_offset,
                                      const int32_t output_activation_min,
                                      const int32_t output_activation_max,
                                      const uint16_t dilation_x,
                                      const uint16_t dilation_y)

{
    (void)output_ch;
    int i_out = 0;
    int i_batch;

    for (i_batch = 0; i_batch < input_batches; i_batch++)
    {
        for (int i_out_y = 0; i_out_y < output_y; i_out_y++)
        {
            const int16_t base_idx_y = (i_out_y * stride_y) - pad_y;
            for (int i_out_x = 0; i_out_x < output_x; i_out_x++)
            {
                const int16_t base_idx_x = (i_out_x * stride_x) - pad_x;
                for (int i_input_ch = 0; i_input_ch < input_ch; i_input_ch++)
                {
                    for (int i_ch_mult = 0; i_ch_mult < ch_mult; i_ch_mult++)
                    {
                        const int idx_out_ch = i_ch_mult + i_input_ch * ch_mult;
                        int32_t acc_0 = 0;

                        int ker_y_start;
                        int ker_x_start;
                        int ker_y_end;
                        int ker_x_end;

                        if (dilation_x > 1)
                        {
                            const int32_t start_x_max = (-base_idx_x + dilation_x - 1) / dilation_x;
                            ker_x_start = MAX(0, start_x_max);
                            const int32_t end_min_x = (input_x - base_idx_x + dilation_x - 1) / dilation_x;
                            ker_x_end = MIN(kernel_x, end_min_x);
                        }
                        else
                        {
                            ker_x_start = MAX(0, -base_idx_x);
                            ker_x_end = MIN(kernel_x, input_x - base_idx_x);
                        }

                        if (dilation_y > 1)
                        {
                            const int32_t start_y_max = (-base_idx_y + dilation_y - 1) / dilation_y;
                            ker_y_start = MAX(0, start_y_max);
                            const int32_t end_min_y = (input_y - base_idx_y + dilation_y - 1) / dilation_y;
                            ker_y_end = MIN(kernel_y, end_min_y);
                        }
                        else
                        {
                            ker_y_start = MAX(0, -base_idx_y);
                            ker_y_end = MIN(kernel_y, input_y - base_idx_y);
                        }

                        if (bias)
                        {
                            acc_0 = bias[idx_out_ch];
                        }

                        for (int i_ker_y = ker_y_start; i_ker_y < ker_y_end; i_ker_y++)
                        {
                            const int32_t idx_y = base_idx_y + dilation_y * i_ker_y;
                            for (int i_ker_x = ker_x_start; i_ker_x < ker_x_end; i_ker_x++)
                            {
                                const int32_t idx_x = base_idx_x + dilation_x * i_ker_x;
                                int32_t idx_0 = (idx_y * input_x + idx_x) * input_ch + i_input_ch;
                                int32_t ker_idx_0 = (i_ker_y * kernel_x + i_ker_x) * (input_ch * ch_mult) + idx_out_ch;

                                acc_0 += (input[idx_0] + input_offset) * kernel[ker_idx_0];
                            }
                        }

                        /* Requantize and clamp output to provided range */
                        acc_0 = arm_nn_requantize(acc_0, output_mult[idx_out_ch], output_shift[idx_out_ch]);
                        acc_0 += output_offset;
                        acc_0 = MAX(acc_0, output_activation_min);
                        acc_0 = MIN(acc_0, output_activation_max);

                        output[i_out++] = acc_0;
                    }
                }
            }
        }
        /* Advance to the next batch */
        input += (input_x * input_y * input_ch);
    }
}

/*
 *  Basic s8 depthwise convolution function.
 *
 *  Refer header file for details.
 *  Optimization using DSP extension is not available for the generic case where channel multiplier is > 1.
 *
 */
arm_status arm_depthwise_conv_s8(const cmsis_nn_context *ctx,
                                 const cmsis_nn_dw_conv_params *dw_conv_params,
                                 const cmsis_nn_per_channel_quant_params *quant_params,
                                 const cmsis_nn_dims *input_dims,
                                 const q7_t *input,
                                 const cmsis_nn_dims *filter_dims,
                                 const q7_t *kernel,
                                 const cmsis_nn_dims *bias_dims,
                                 const int32_t *bias,
                                 const cmsis_nn_dims *output_dims,
                                 q7_t *output)
{
    const uint16_t dilation_x = dw_conv_params->dilation.w;
    const uint16_t dilation_y = dw_conv_params->dilation.h;

    (void)dw_conv_params->dilation;
    (void)bias_dims;
    (void)ctx;

    if (dw_conv_params->ch_mult % 4 == 0 && input_dims->n == 1 && dw_conv_params->dilation.w == 1 &&
        dw_conv_params->dilation.h == 1)
    {
        depthwise_conv_s8_mult_4(input,
                                 input_dims->w,
                                 input_dims->h,
                                 input_dims->c,
                                 kernel,
                                 output_dims->c,
                                 dw_conv_params->ch_mult,
                                 filter_dims->w,
                                 filter_dims->h,
                                 dw_conv_params->padding.w,
                                 dw_conv_params->padding.h,
                                 dw_conv_params->stride.w,
                                 dw_conv_params->stride.h,
                                 bias,
                                 output,
                                 quant_params->shift,
                                 quant_params->multiplier,
                                 output_dims->w,
                                 output_dims->h,
                                 dw_conv_params->output_offset,
                                 dw_conv_params->input_offset,
                                 dw_conv_params->activation.min,
                                 dw_conv_params->activation.max);
    }
    else
    {
        depthwise_conv_s8_generic(input,
                                  input_dims->n,
                                  input_dims->w,
                                  input_dims->h,
                                  input_dims->c,
                                  kernel,
                                  output_dims->c,
                                  dw_conv_params->ch_mult,
                                  filter_dims->w,
                                  filter_dims->h,
                                  dw_conv_params->padding.w,
                                  dw_conv_params->padding.h,
                                  dw_conv_params->stride.w,
                                  dw_conv_params->stride.h,
                                  bias,
                                  output,
                                  quant_params->shift,
                                  quant_params->multiplier,
                                  output_dims->w,
                                  output_dims->h,
                                  dw_conv_params->output_offset,
                                  dw_conv_params->input_offset,
                                  dw_conv_params->activation.min,
                                  dw_conv_params->activation.max,
                                  dilation_x,
                                  dilation_y);
    }

    /* Return to application */
    return ARM_MATH_SUCCESS;
}

/**
 * @} end of NNConv group
 */