- ZYNQ HLS圖像處理加速總結(二)

將vdma的基址映射到虛擬地址空間上，在linux系統(tǒng)下就可以直接通過指針訪問vdma的各個寄存器
handle->vdmaVirtualAddress = (unsigned int*)mmap(NULL, 65535, PROT_READ | PROT_WRITE, MAP_SHARED, handle->vdmaHandler, (off_t)handle->baseAddr);

這個函數的返回值就是申請到的vdma基址的虛擬地址，這個地址加上寄存器offset就可以用來配置各個寄存器了，在使用之前要校驗下申請的地址是不是有效的
if (handle->vdmaVirtualAddress == MAP_FAILED) {
perror("vdmaVirtualAddress mapping for absolute memory access failed.\n");
return -1;
}

然后還要申請圖像存放的地址，這里根據vdma的運行方式可以配置多個地址存放多幅圖像數據。

下面是我的主程序，僅供參考：
int main() {

//variable start
int j, i;
Vec2b pix;
struct timeval tstart, tend, hls_start, hls_end;
float timeuse;
Mat src_rgb = imread(INPUT_IMAGE, 1);
Mat src_yuv(src_rgb.rows, src_rgb.cols, CV_8UC2);
Mat dst_yuv(src_rgb.rows, src_rgb.cols, CV_8UC2);
Mat dst_rgb(src_rgb.rows, src_rgb.cols, CV_8UC3);
//convert to yuv format
cvtcolor_rgb2yuv422(src_rgb, src_yuv);
IplImage src = src_yuv;
IplImage dst = dst_yuv;
//variable end

#if SW_GENERATE
printf("opencv software processing\n");
//calculate software used time
gettimeofday(&tstart, NULL);
opencv_sobel_init();
opencv_sobel(&src, &dst);
gettimeofday(&tend, NULL);
timeuse = 1000000 * (tend.tv_sec - tstart.tv_sec) + (tend.tv_usec - tstart.tv_usec);
timeuse /= 1000000;
printf("soft used time is %f\n", timeuse);
cvtColor(dst_yuv, dst_rgb, CV_YUV2BGR_YUYV);
imwrite(OUTPUT_IMAGE_GOLDEN, dst_rgb);
#endif

if (!(init_filter() == XST_SUCCESS))
{
printf("filter init faild!");
}
set_reg_filter();

// Setup VDMA handle and memory-mapped ranges
vdma_setup(&handle, 0x43000000, 640, 480, 2, 0x1f400000, 0x1f800000, 0x1fc00000);

gettimeofday(&tstart, NULL);
memcpy(handle.fb1VirtualAddress, (uchar *)src.imageData, 640 * 480 * 2);

#if MEMCPY_CHECK
printf("memcpy checking \n");
u32 memcpy_error_flag = 0;
for (i = 0; i < src_yuv.rows; i++) //row 480
{
for (j = 0; j < src_yuv.cols; j++) //col 640*2
{
pix = src_yuv.at(i, j);
if ((handle.fb1VirtualAddress[j * 2 + i * 640 * 2] != pix.val[0]) || (
handle.fb1VirtualAddress[j * 2 + i * 640 * 2 + 1] != pix.val[1]))
{
memcpy_error_flag = 1;
}
}
}
if (memcpy_error_flag == 1)
{
printf("img copy error");
return 0;
}
//memset(handle.fb1VirtualAddress, 0, handle.width * handle.height * handle.pixelLength);
printf("memcpy check result FB2:(ORI)\n");
for (j = 512; j < 512 + 20; j++) printf(" %02x", handle.fb2VirtualAddress[j]); printf("\n");

#endif

gettimeofday(&hls_start, NULL);
vdma_start_triple_buffering(&handle);
//printf("hahahaha\n");

wait_done_filter();
gettimeofday(&hls_end, NULL);

#if RESULT_CHECK
printf("RESULT CHECK FB2:(NOW)\n");
for (j = 635 * 2; j < 635 * 2 + 20; j++) printf(" %02x", handle.fb2VirtualAddress[j]); printf("\n");
//}
#endif
memcpy((uchar *)dst.imageData, handle.fb2VirtualAddress, 640 * 480 * 2);

gettimeofday(&tend, NULL);
timeuse = 1000000 * (tend.tv_sec - tstart.tv_sec) + (tend.tv_usec - tstart.tv_usec);
timeuse /= 1000000;
printf("hard total used time is %f\n", timeuse);

timeuse = 1000000 * (hls_end.tv_sec - hls_start.tv_sec) + (hls_end.tv_usec - hls_start.tv_usec);
timeuse /= 1000000;
printf("hard hls used time is %f\n", timeuse);
print_vdma_register_status();
cvtColor(dst_yuv, dst_rgb, CV_YUV2BGR_YUYV);
imwrite(OUTPUT_IMAGE, dst_rgb);
stop_filter();
// Halt VDMA and unmap memory ranges
vdma_halt(&handle);
return image_compare(OUTPUT_IMAGE, OUTPUT_IMAGE_GOLDEN);
}

vdma寄存器配置參考pg020_axi_vdma文檔

軟件編譯選項

這里單獨列出來是因為感覺在這種處理器性能不是很好的硬件平臺下進行大計算量算法的實施的情況下，一定要讓軟件以最高效率運行（盡力最高效率吧)下面是辛辛苦苦寫的Makefile，用的通配符@梅神，稍微改改就可以用在新工程上
CC=g++
CFLAGS= -g -O2 -mcpu=cortex-a9 -mfpu=neon -ftree-vectorize -mvectorize-with-neon-quad #-mfloat-abi=softfp -ffast-math
CFLAGS+=`pkg-config --cflags opencv`
LDFLAGS+=`pkg-config --libs opencv`

OBJS = $(patsubst %.c,%.o,$(wildcard *.c))
OBJS += $(patsubst %.cpp,%.o,$(wildcard *.cpp))
all: vdma_test

%.o: %.cpp
$(CC) -c $(CFLAGS) -o $@ $<
%.o: %.c
$(CC) -c $(CFLAGS) -o $@ $<
vdma_test: $(OBJS)
$(CC) -o $@ $(OBJS) $(LDFLAGS)
clean:
rm vdma_test $(OBJS)

cflag的優(yōu)化配置參考的xapp1206-boost-sw-performance-zynq7soc-w-neon文檔。