瑞芯微rk3399使用rga图形加速处理后拷贝到普通内存耗时大的解决方案

rga,mpp,drm,un-cached

rga图形加速引擎处理图形很快：rk3399开发板自带rga（图形加速引擎）和mpp编解码。在MPP解码->图像后处理->显示整个流程中，RGA都是直接访问内存物理地址，整个过程零内存拷贝，效率很高。（都是在drm内存上完成，drm内存默认non-cache）。

从rga的drm内存拷贝图片到普通内存很慢：其他任何软件以cpu访问该段内存都会因为non-cached导致不必要的损失

解决方案：

将rga-API替换

老版rga-API:

extern "C" 
{
    #include "rockchip_rga/rockchip_rga.h"
}

dec_rga->ops->setSrcFormat(dec_rga, V4L2_PIX_FMT_NV12,   width, ver_stride);
dec_rga->ops->setDstFormat(dec_rga, V4L2_PIX_FMT_RGB24,  width, height);
dec_rga->ops->setSrcCrop(dec_rga, 0, 0, width, height);
dec_rga->ops->setSrcBufferPtr(dec_rga, (unsignedchar*)mpp_buffer_get_ptr(buffer));
dec_rga->ops->setDstBufferPtr(dec_rga, (unsigned char*)rgb_buffer);
dec_rga->ops->go(dec_rga);//rga处理（resize,cvtcolor）很快

img_rgb = cv::Mat(height, width, CV_8UC3, (uchar*)(rgb_buffer), cv::Mat::AUTO_STEP).clone();//从rga-drm memory拷贝到普通内存很慢

新rga-API:

#include "RgaUtils.h"
#include "im2d_api/im2d.hpp"
#include "rga.h"

 src_buffer = wrapbuffer_virtualaddr((char*)mpp_buffer_get_ptr(buff), width, ver_stride, src_format);
 mid_buffer = wrapbuffer_virtualaddr((char*)malloc(width * height * get_bpp_from_format(src_format)), width,height,  src_format);
 
 im_rect src_rect;
 src_rect.x      = 0;
 src_rect.y      = 0;
 src_rect.width  = width;
 src_rect.height = height;
  
 //对齐后的图片crop
 status = imcrop(src_buffer, mid_buffer, src_rect);
 //cvtcolor+resize
  status = imcvtcolor(mid_buffer, out_buffer , mid_buffer.format, out_buffer.format);
  if(mid_buffer.vir_addr != NULL) {
      free(mid_buffer.vir_addr);
  }//rga处理（resize,cvtcolor）很快
                 
  img_rgb=cv::Mat(dst_height, dst_width, CV_8UC3, out_buffer.vir_addr).clone();//从rga-cache memory拷贝到普通内存很快

```cpp