edge-impulse-sdk/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_fast_s16.c

#include "edge-impulse-sdk/classifier/ei_classifier_config.h"
#if EI_CLASSIFIER_TFLITE_LOAD_CMSIS_NN_SOURCES
/*
 * Copyright (C) 2010-2022 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_convolve_fast_s16.c
 * Description:  Optimized s16 version of convolution.
 *
 * $Date:        19 April 2022
 * $Revision:    V.2.0.0
 *
 * Target Processor:  Cortex-M cores
 *
 * -------------------------------------------------------------------- */

#include "edge-impulse-sdk/CMSIS/NN/Include/arm_nnfunctions.h"
#include "edge-impulse-sdk/CMSIS/NN/Include/arm_nnsupportfunctions.h"

/**
 *  @ingroup groupNN
 */

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

/*
 * Basic s16 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_cmsis_nn_status arm_convolve_fast_s16(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 q15_t *input_data,
                                          const cmsis_nn_dims *filter_dims,
                                          const q7_t *filter_data,
                                          const cmsis_nn_dims *bias_dims,
                                          const int64_t *bias_data,
                                          const cmsis_nn_dims *output_dims,
                                          q15_t *output_data)
{
    (void)bias_dims;
    if (filter_dims->w * filter_dims->h * input_dims->c >= 512)
    {
        return ARM_CMSIS_NN_ARG_ERROR;
    }

    if (ctx->buf == NULL && arm_convolve_s8_get_buffer_size(input_dims, filter_dims) > 0)
    {
        return ARM_CMSIS_NN_ARG_ERROR;
    }
    q15_t *buffer_a = (q15_t *)ctx->buf;

    const int32_t input_batches = input_dims->n;
    const int32_t input_x = input_dims->w;
    const int32_t input_y = input_dims->h;
    const int32_t input_ch = input_dims->c;
    const int32_t kernel_x = filter_dims->w;
    const int32_t kernel_y = filter_dims->h;
    const int32_t output_x = output_dims->w;
    const int32_t output_y = output_dims->h;
    const int32_t output_ch = output_dims->c;

    const int32_t pad_x = conv_params->padding.w;
    const int32_t pad_y = conv_params->padding.h;
    const int32_t stride_x = conv_params->stride.w;
    const int32_t stride_y = conv_params->stride.h;

    const int16_t out_activation_min = conv_params->activation.min;
    const int16_t out_activation_max = conv_params->activation.max;
    int32_t *output_mult = quant_params->multiplier;
    int32_t *output_shift = quant_params->shift;

    for (int i_batch = 0; i_batch < input_batches; i_batch++)
    {
#if defined(ARM_MATH_DSP) && !defined(ARM_MATH_MVEI)
        /* Generate two columns from the input tensor a GEMM computation */
        q15_t *two_column_buf = buffer_a;
        q15_t *out = output_data;
        /* This part implements the im2col function */
        for (int32_t i_out_y = 0; i_out_y < output_y; i_out_y++)
        {
            for (int32_t i_out_x = 0; i_out_x < output_x; i_out_x++)
            {
                for (int32_t 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 (int32_t 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 */
                            arm_memset_q7((q7_t *)two_column_buf, 0, sizeof(q15_t) * input_ch);
                        }
                        else
                        {
                            arm_memcpy_q7((q7_t *)two_column_buf,
                                          (const q7_t *)(input_data + (i_ker_y * input_x + i_ker_x) * input_ch),
                                          input_ch * sizeof(q15_t));
                        }
                        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_s16(filter_data,
                                                     buffer_a,
                                                     output_ch,
                                                     output_shift,
                                                     output_mult,
                                                     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++)
            {
                /* Init the accumulator*/
                q31_t sum = 0;

                /* 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--;
                }
                if (bias_data)
                {
                    q31_t reduced_multiplier = REDUCE_MULTIPLIER(output_mult[i]);
                    q63_t acc_64 = sum + bias_data[i];
                    sum = arm_nn_requantize_s64(acc_64, reduced_multiplier, output_shift[i]);
                }
                else
                {
                    sum = arm_nn_requantize(sum, output_mult[i], output_shift[i]);
                }
                sum = MAX(sum, out_activation_min);
                sum = MIN(sum, out_activation_max);
                *out++ = (q15_t)sum;
            }
        }
#else
        (void)input_data;
        (void)output_data;
        (void)bias_data;
        (void)filter_data;
        (void)buffer_a;
        (void)kernel_x;
        (void)kernel_y;
        (void)pad_x;
        (void)pad_y;
        (void)stride_x;
        (void)stride_y;
        (void)out_activation_min;
        (void)out_activation_max;
        (void)output_mult;
        (void)output_shift;
        return ARM_CMSIS_NN_ARG_ERROR;
#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_CMSIS_NN_SUCCESS;
}

int32_t arm_convolve_fast_s16_get_buffer_size(const cmsis_nn_dims *input_dims, const cmsis_nn_dims *filter_dims)
{
#if defined(ARM_MATH_DSP) && !defined(ARM_MATH_MVEI)
    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
 */

#endif // EI_CLASSIFIER_TFLITE_LOAD_CMSIS_NN_SOURCES