RV1106 YOLOV5

最近研究瑞星微的rv1106的yolov5检测算法，但是官方给的例子真是让人欲哭无泪，首先读取时局用的是stb的图像库，前处理加上resize需要250ms，后处理只能输出nc1hwc2,里面的举证转换需要250ms，后续自己做了很多工作，发这个只是为了记录代码。

// Copyright (c) 2021 by Rockchip Electronics Co., Ltd. All Rights Reserved.
//
// 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
//
//     http://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.

/*-------------------------------------------
                Includes
-------------------------------------------*/
#include "rknn_api.h"

#include <float.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#include <vector>
#include "fstream"
#include "iostream"
#define STB_IMAGE_IMPLEMENTATION
#include "stb/stb_image.h"
#define STB_IMAGE_RESIZE_IMPLEMENTATION
#include <stb/stb_image_resize.h>
#include "opencv2/core/core.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "postprocess.h"

#define PERF_WITH_POST 1

/*-------------------------------------------
                  Functions
-------------------------------------------*/
static inline int64_t getCurrentTimeUs()
{
   
  struct timeval tv;
  gettimeofday(&tv, NULL);
  return tv.tv_sec * 1000000 + tv.tv_usec;
}

static void dump_tensor_attr(rknn_tensor_attr *attr)
{
   
  char dims[128] = {
   0};
  for (int i = 0; i < attr->n_dims; ++i)
  {
   
    int idx = strlen(dims);
    sprintf(&dims[idx], "%d%s", attr->dims[i], (i == attr->n_dims - 1) ? "" : ", ");
  }
  printf("  index=%d, name=%s, n_dims=%d, dims=[%s], n_elems=%d, size=%d, fmt=%s, type=%s, qnt_type=%s, "
         "zp=%d, scale=%f\n",
         attr->index, attr->name, attr->n_dims, dims, attr->n_elems, attr->size, get_format_string(attr->fmt),
         get_type_string(attr->type), get_qnt_type_string(attr->qnt_type), attr->zp, attr->scale);
}

static void *load_file(const char *file_path, size_t *file_size)
{
   
  FILE *fp = fopen(file_path, "r");
  if (fp == NULL)
  {
   
    printf("failed to open file: %s\n", file_path);
    return NULL;
  }

  fseek(fp, 0, SEEK_END);
  size_t size = (size_t)ftell(fp);
  fseek(fp, 0, SEEK_SET);

  void *file_data = malloc(size);
  if (file_data == NULL)
  {
   
    fclose(fp);
    printf("failed allocate file size: %zu\n", size);
    return NULL;
  }

  if (fread(file_data, 1, size, fp) != size)
  {
   
    fclose(fp);
    free(file_data);
    printf("failed to read file data!\n");
    return NULL;
  }

  fclose(fp);

  *file_size = size;

  return file_data;
}

static unsigned char *load_image(const char *image_path, rknn_tensor_attr *input_attr, int *img_height, int *img_width)
{
   
  int req_height = 0;
  int req_width = 0;
  int req_channel = 0;

  switch (input_attr->fmt)
  {
   
  case RKNN_TENSOR_NHWC:
    req_height = input_attr->dims[1];
    req_width = input_attr->dims[2];
    req_channel = input_attr->dims[3];
    break;
  case RKNN_TENSOR_NCHW:
    req_height = input_attr->dims[2];
    req_width = input_attr->dims[3];
    req_channel = input_attr->dims[1];
    break;
  default:
    printf("meet unsupported layout\n");
    return NULL;
  }

  int channel = 0;

  unsigned char *image_data = stbi_load(image_path, img_width, img_height, &channel, req_channel);
  if (image_data == NULL)
  {
   
    printf("load image failed!\n");
    return NULL;
  }

  if (*img_width != req_width || *img_height != req_height)
  {
   




    unsigned char *image_resized = (unsigned char *)STBI_MALLOC(req_width * req_height * req_channel);
    if (!image_resized)
    {
   
      printf("malloc image failed!\n");
      STBI_FREE(image_data);
      return NULL;
    }       
     int64_t start_us = getCurrentTimeUs();
    if (stbir_resize_uint8(image_data, *img_width, *img_height, 0, image_resized, req_width, req_height, 0, channel) != 1)
    {
   
      printf("resize image failed!\n");
      STBI_FREE(image_data);
      return NULL;
    }
        int64_t elapse_us = getCurrentTimeUs() - start_us;

    printf(" resize Time = %.2fms,\n",  elapse_us / 1000.f);
    STBI_FREE(image_data);
    image_data = image_resized;
  }

  return image_data;
}

// 量化模型的npu输出结果为int8数据类型，后处理要按照int8数据类型处理
// 如下提供了int8排布的NC1HWC2转换成int8的nchw转换代码
int NC1HWC2_int8_to_NCHW_int8(const int8_t *src, int8_t *dst, int *dims, int channel, int h, int w)
{
   
  int batch = dims[0];
  int C1 = dims[1];
  int C2 = dims[4];
  int hw_src = dims[2] * dims[3];
  std::cout <<batch <<" " <<C1 <<" " <<C2 <<" " <<hw_src <<" "  <<std::endl;
  int hw_dst = h * w;
  for (int i = 0; i < batch; i++)
  {
   
    src = src + i * C1 * hw_src * C2;
    dst = dst + i * channel * hw_dst;
    for (int c = 0; c < channel; ++c)
    {
   
      int plane = c / C2;
      const int8_t *src_c = plane * hw_src * C2 + src;
      int offset = c % C2;
      for (int cur_h = 0; cur_h < h; ++cur_h)
        for (int cur_w = 0; cur_w < w; ++cur_w)
        {
   
          int cur_hw = cur_h * w + cur_w;
          dst[c * hw_dst + cur_h * w + cur_w] = src_c[C2 * cur_hw + offset];
        }
    }
  }

  return 0;
}

// 量化模型的npu输出结果为int8数据类型，后处理要按照int8数据类型处理
// 如下提供了int8排布的NC1HWC2转换成float的nchw转换代码
int NC1HWC2_int8_to_NCHW_float(const int8_t *src, float *dst, int *dims, int channel, int h, int w, int zp, float scale)
{
   
  int batch = dims[0];
  int C1 = dims[1];
  int C2 = dims[4];
  int hw_src = dims[2] * dims[3];
  int hw_dst = h * w;
  for (int i = 0; i < batch; i++)
  {
   
    src = src + i * C1 * hw_src * C2;
    dst = dst + i * channel * hw_dst;
    for (int c = 0; c < channel; ++c)
    {
   
      int plane = c / C2;
      const int8_t *src_c = plane * hw_src * C2 + src;
      int offset = c % C2;
      for (int cur_h = 0; cur_h < h; ++cur_h)
        for (int cur_w = 0; cur_w < w; ++cur_w)
        {
   
          int cur_hw = cur_h * w + cur_w;
          dst[c * hw_dst + cur_h * w + cur_w] = (src_c[C2 * cur_hw + offset] - zp) * scale; // int8-->float
        }
    }
  }

  return 0;
}

/*-------------------------------------------
                  Main Functions
-------------------------------------------*/
int main(int argc, char *argv[])
{
   
  if (argc < 3)
  {
   
    printf("Usage:%s model_path input_path [loop_count]\n", argv[0]);
    return -1;
  }

  char *model_path = argv[1];
  char *input_path = argv[2];

  int loop_count = 10;
  if (argc > 3)
  {
   
    loop_count = atoi(argv[3]);
  }

  const float nms_threshold = NMS_THRESH;
  const float box_conf_threshold = BOX_THRESH;

  int img_width = 0;
  int img_height = 0;

  rknn_context ctx = 0;

  // Load RKNN Model
#if 1
  // Init rknn from model path
  int ret = rknn_init(&ctx, model_path, 0, 0, NULL);
#else
  // Init rknn from model data