不断执行推理，最终🧨 E RKNN: failed to allocate fd, ret: -1, errno: 12, errstr: Cannot allocate memory

• ezreal

  Posts: 15

  Joined: 2024-03-10 6:01

  2024-06-19 2:56

  不断对同一张图片进行推理（大概运行了几百次后），最终会显示内存不够。
  这是输出的日志：

  rknn_init fail! ret=-1
  ret=-1 model_path=/data/RKNN/model/model.rknn
  E RKNN: failed to allocate fd, ret: -1, errno: 12, errstr: Cannot allocate memory
  E RKNN: failed to allocate model memory!, size: 8239104, flags: #2
  rknn_init fail! ret=-1

  一些观察：
  观察一：使用MobaXterm_Personal软件，在进行重复推理过程中，会看到Used RAM从一开始的25MB慢慢上升到27MB。在27MB附近，最终会开始报错。
  观察二：在报错后，我关闭重复推理程序，这时Used RAM会恢复到25MB，但是这时候再执行推理，会同样报错（已经无法再进行推理）

  rknn_init fail! ret=-1
  ret=-1 model_path=/data/RKNN/model/model.rknn
  E RKNN: failed to allocate fd, ret: -1, errno: 12, errstr: Cannot allocate memory
  E RKNN: failed to allocate model memory!, size: 8239104, flags: #2
  rknn_init fail! ret=-1

  观察三：重新拔插之后，可以恢复推理功能。
  不知道是哪个地方出问题了？
  
  下面提供推理代码：
  main.cc

  Code: Select all

  
  /*-------------------------------------------
                  Includes
  -------------------------------------------*/
  #include <stdint.h>
  #include <stdio.h>
  #include <stdlib.h>
  #include <string.h>
  
  #include "yolov5.h"
  #include "image_utils.h"
  #include "file_utils.h"
  
  #if defined(RV1106_1103)
      #include "dma_alloc.cpp"
  #endif
  
  void print_output_results(output_result *results, int num_results) {
      printf("Predictions:\n");
      for (int i = 0; i < num_results; i++) {
          printf("Class: %d, Score: %.4f\n", results[i].cls, results[i].score);
      }
      // 打开文件
      FILE *file = fopen("/tmp/output", "w");
      if (file == NULL) {
          printf("无法打开文件。\n");
      }
  
      // 将整数写入文件
      fprintf(file, "%d", results[0].cls);
  
      // 关闭文件
      fclose(file);
  
      printf("The prediction of %d has been save into /tmp/output.\n", results[0].cls);
  }
  
  /*-------------------------------------------
                    Main Function
  -------------------------------------------*/
  int main(int argc, char **argv)
  {
      if (argc != 3 && argc != 4)
      {
          printf("%s <model_path> <image_path>\n", argv[0]);
          return -1;
      }
  
      const char *model_path = argv[1];
      const char *image_path = argv[2];
  
      int ret;
      rknn_app_context_t rknn_app_ctx;
      memset(&rknn_app_ctx, 0, sizeof(rknn_app_context_t));
  
      int topk = 2;
      output_result result[topk];
  
      ret = init_yolov5_model(model_path, &rknn_app_ctx);
      if (ret != 0)
      {
          printf("init_yolov5_model fail! ret=%d model_path=%s\n", ret, model_path);
          goto out;
      }
  
      image_buffer_t src_image;
      memset(&src_image, 0, sizeof(image_buffer_t));
      ret = read_image(image_path, &src_image);
      if (ret != 0) {
          printf("read image fail! ret=%d image_path=%s\n", ret, image_path);
          return -1;
      }
  
  #if defined(RV1106_1103)
      //RV1106 rga requires that input and output bufs are memory allocated by dma
      ret = dma_buf_alloc(RV1106_CMA_HEAP_PATH, src_image.size, &rknn_app_ctx.img_dma_buf.dma_buf_fd, 
                         (void **) & (rknn_app_ctx.img_dma_buf.dma_buf_virt_addr));
      memcpy(rknn_app_ctx.img_dma_buf.dma_buf_virt_addr, src_image.virt_addr, src_image.size);
      dma_sync_cpu_to_device(rknn_app_ctx.img_dma_buf.dma_buf_fd);
      free(src_image.virt_addr);
      src_image.virt_addr = (unsigned char *)rknn_app_ctx.img_dma_buf.dma_buf_virt_addr;
      src_image.fd = rknn_app_ctx.img_dma_buf.dma_buf_fd;
      rknn_app_ctx.img_dma_buf.size = src_image.size;
  #endif
  
      printf(">>> going to inference_yolov5_model\n");
  
      ret = inference_yolov5_model(&rknn_app_ctx, &src_image, result, topk);
      if (ret != 0) {
          printf("inference_yolov5_model fail! ret=%d\n", ret);
          goto out;
      }
  
      print_output_results(result, topk);
  
  out:
      printf(">>> going to release_yolov5_model\n");
      ret = release_yolov5_model(&rknn_app_ctx);
      if (ret != 0)
      {
          printf("release_yolov5_model fail! ret=%d\n", ret);
      }
  
      if (src_image.virt_addr != NULL) {
  #if defined(RV1106_1103)
          dma_buf_free(rknn_app_ctx.img_dma_buf.size, &rknn_app_ctx.img_dma_buf.dma_buf_fd,
                  rknn_app_ctx.img_dma_buf.dma_buf_virt_addr);
  #else
          free(src_image.virt_addr);
  #endif
      }
  
      return 0;
  }
  

  Code: Select all

  
  int inference_yolov5_model(rknn_app_context_t *app_ctx, image_buffer_t *src_img, output_result *out_result, int topk)
  {
      int ret;
      image_buffer_t img;
      memset(&img, 0, sizeof(image_buffer_t));
  
      // Pre Process
      img.width = app_ctx->model_width;
      img.height = app_ctx->model_height;
      img.format = IMAGE_FORMAT_RGB888;
      img.size = get_image_size(&img);
      img.virt_addr = (unsigned char *)app_ctx->input_mems[0]->virt_addr;
      if (img.virt_addr == NULL) {
          printf("malloc buffer size:%d fail!\n", img.size);
          return -1;
      }
  
      ret = convert_image(src_img, &img, NULL, NULL, 0);
  //    ret = convert_image_gray(src_img, &img, NULL, NULL, 0);
      if (ret < 0) {
          printf("convert_image fail! ret=%d\n", ret);
          return -1;
      }
  
      // Run
      printf("rknn_run...\n");
      ret = rknn_run(app_ctx->rknn_ctx, nullptr);
      if (ret < 0) {
          printf("rknn_run fail! ret=%d\n", ret);
          return -1;
      }
  
      float outputs_float[app_ctx->output_attrs[0].size_with_stride];
      if (app_ctx->output_attrs[0].fmt == RKNN_TENSOR_NC1HWC2) {
          printf("RKNN_TENSOR_NC1HWC2\n");
          int size = app_ctx->output_attrs[0].size_with_stride;
          int zp = app_ctx->output_attrs[0].zp;
          float scale = app_ctx->output_attrs[0].scale;
          int channel = app_ctx->output_attrs[0].dims[1];
          int h = app_ctx->output_attrs[0].n_dims > 2 ? app_ctx->output_attrs[0].dims[2] : 1;
          int w = app_ctx->output_attrs[0].n_dims > 3 ? app_ctx->output_attrs[0].dims[3] : 1;
          if (app_ctx->output_attrs[0].type == RKNN_TENSOR_INT8) {
              NC1HWC2_int8_to_NCHW_float((int8_t *)app_ctx->output_mems[0]->virt_addr, outputs_float,
                                         (int *)app_ctx->output_attrs[0].dims, channel, h, w, zp, scale);
          } else {
              printf("dtype: %s cannot convert!", get_type_string(app_ctx->output_attrs[0].type));
          }
      } else {
          if (app_ctx->output_attrs[0].type == RKNN_TENSOR_INT8) {
              printf("RKNN_TENSOR_INT8\n");
              int8_t *src = (int8_t *)app_ctx->output_mems[0]->virt_addr;
              for (int index = 0; index < app_ctx->output_attrs[0].n_elems; index++) {
                  outputs_float[index] = ((float)src[index] - app_ctx->output_attrs[0].zp) * app_ctx->output_attrs[0].scale;
              }
              // 输出转换前的数组
              printf("Original array:\n");
              for (int index = 0; index < app_ctx->output_attrs[0].n_elems; index++) {
                  printf("%d ", src[index]);
              }
              printf("\n");
          } else if(app_ctx->output_attrs[0].type == RKNN_TENSOR_UINT8) {
              printf("RKNN_TENSOR_UINT8\n");
              uint8_t *src = (uint8_t *)app_ctx->output_mems[0]->virt_addr;
              for (int index = 0; index < app_ctx->output_attrs[0].n_elems; index++) {
                  outputs_float[index] = ((float)src[index] - app_ctx->output_attrs[0].zp) * app_ctx->output_attrs[0].scale;
              }
              // 输出转换前的数组
              printf("Original array:\n");
              for (int index = 0; index < app_ctx->output_attrs[0].n_elems; index++) {
                  printf("%d ", src[index]);
              }
              printf("\n");
          }
      }
  
      // Post Process
      printf("softmax with num_classes %d...\n", app_ctx->output_attrs[0].n_elems);
      softmax(outputs_float, app_ctx->output_attrs[0].n_elems);
  
      printf("get_topk_with_indices...\n");
      get_topk_with_indices(outputs_float, app_ctx->output_attrs[0].n_elems, topk, out_result);
      printf("inference done...\n");
  
  out:
      return ret;
  }
  

  • Quote