VDR Portal

Source code

typedef struct {
	uint8_t* yuv[3];
	int height[3];
	int width[3];
	int pitch[3]; // stored width
	int yuvsize;
} vaapi_yadif_img;

static void nv12ToUV(uint8_t *dstU, uint8_t *dstV, uint8_t *src, int width)
{
	int i;
	width>>=1;
	for(i=0;i<width;i++) {
		dstU[i]=src[i*2];
		dstV[i]=src[i*2+1];
	}
}

static void UVTonv12(uint8_t *dst, uint8_t *srcU, uint8_t *srcV, int width)
{
	int i;
	width>>=1;
	for(i=0;i<width;i++) {
		dst[i*2]=srcU[i];
		dst[i*2+1]=srcV[i];
	}
}

#define YADIF_PITCH(w,a) (w%a==0) ? w : w+a-(w%a)

static void CopyVAImg2YadifImg_alloc(VADisplay *display, VAImage *src, vaapi_yadif_img *dst)
{
	uint8_t *srcp, *srcuv, *dstU, *dstV;
	int i;
	vaMapBuffer(display, src->buf, (void**)&srcp);
	dst->width[0] = src->width;
	dst->width[1] = dst->width[2] = dst->width[0]>>1;
	dst->height[0] = src->height;
	dst->height[1] = dst->height[2] = dst->height[0]>>1;
	if(src->format.fourcc == VA_FOURCC_NV12) {
		dst->pitch[0] = src->pitches[0];
		dst->pitch[2] = dst->pitch[1] = YADIF_PITCH(dst->width[1],16);
		dst->yuvsize = dst->pitch[0]*dst->height[0]+dst->pitch[1]*dst->height[1]+dst->pitch[2]*dst->height[2];
		dst->yuv[0] = malloc(dst->yuvsize);
		dst->yuv[1] = dst->yuv[0]+dst->pitch[0]*dst->height[0];
		dst->yuv[2] = dst->yuv[1]+dst->pitch[1]*dst->height[1];
		memcpy(dst->yuv[0], srcp+src->offsets[0], src->height*src->pitches[0]);
		srcuv=srcp+src->offsets[1];
		dstU=dst->yuv[1];
		dstV=dst->yuv[2];
		for(i=0;i<dst->height[1];i++) {
			nv12ToUV(dstU,dstV,srcuv,src->width);
			srcuv+=src->pitches[1];
			dstU+=dst->pitch[1];
			dstV+=dst->pitch[2];
		}
	}
	if (vaUnmapBuffer(display, src->buf) != VA_STATUS_SUCCESS) {
		Error(_("video/vaapi: can't unmap image buffer\n"));
    }
}

static void CopyVAImg2YadifImg(VADisplay *display, VAImage *src, vaapi_yadif_img *dst)
{
	uint8_t *srcp, *srcuv, *dstU, *dstV;
	int i;
	vaMapBuffer(display, src->buf, (void**)&srcp);
	if(src->format.fourcc == VA_FOURCC_NV12) {
		memcpy(dst->yuv[0], srcp+src->offsets[0], src->height*src->pitches[0]);
		srcuv=srcp+src->offsets[1];
		dstU=dst->yuv[1];
		dstV=dst->yuv[2];
		for(i=0;i<dst->height[1];i++) {
			nv12ToUV(dstU,dstV,srcuv,src->width);
			srcuv+=src->pitches[1];
			dstU+=dst->pitch[1];
			dstV+=dst->pitch[2];
		}
	}
	if (vaUnmapBuffer(display, src->buf) != VA_STATUS_SUCCESS) {
		Error(_("video/vaapi: can't unmap image buffer\n"));
    }
}

static void CopyYadifImg2VAImg(VADisplay *display, vaapi_yadif_img *src, VAImage *dst)
{
	uint8_t *dstp, *dstuv, *srcU, *srcV;
	int i;
	vaMapBuffer(display, dst->buf, (void**)&dstp);
	if(dst->format.fourcc == VA_FOURCC_NV12) {
		memcpy(dstp+dst->offsets[0], src->yuv[0], src->height[0]*src->pitch[0]);
		dstuv=dstp+dst->offsets[1];
		srcU=src->yuv[1];
		srcV=src->yuv[2];
		for(i=0;i<src->height[1];i++) {
			UVTonv12(dstuv,srcU,srcV,dst->width);
			dstuv+=dst->pitches[1];
			srcU+=src->pitch[1];
			srcV+=src->pitch[2];
		}
	}
	if (vaUnmapBuffer(display, dst->buf) != VA_STATUS_SUCCESS) {
		Error(_("video/vaapi: can't unmap image buffer\n"));
    }
}

static void InitVaapiYadif(VADisplay *display, VAImage *src, vaapi_yadif_img *buf_imgs[])
{
	buf_imgs[0]=malloc(5*sizeof(vaapi_yadif_img));
	buf_imgs[1]=buf_imgs[0]+sizeof(vaapi_yadif_img);
	buf_imgs[2]=buf_imgs[1]+sizeof(vaapi_yadif_img);
	buf_imgs[3]=buf_imgs[2]+sizeof(vaapi_yadif_img);
	buf_imgs[4]=buf_imgs[3]+sizeof(vaapi_yadif_img);

	CopyVAImg2YadifImg_alloc(display,src,buf_imgs[0]);

	memcpy(buf_imgs[1],buf_imgs[0],sizeof(vaapi_yadif_img));
	memcpy(buf_imgs[2],buf_imgs[0],sizeof(vaapi_yadif_img));
	buf_imgs[1]->yuv[0] = malloc(buf_imgs[1]->yuvsize);
	buf_imgs[2]->yuv[0] = malloc(buf_imgs[2]->yuvsize);
	memcpy(buf_imgs[1]->yuv[0],buf_imgs[0]->yuv[0],buf_imgs[1]->yuvsize);
	memcpy(buf_imgs[2]->yuv[0],buf_imgs[0]->yuv[0],buf_imgs[2]->yuvsize);
	buf_imgs[1]->yuv[1] = buf_imgs[1]->yuv[0]+buf_imgs[1]->pitch[0]*buf_imgs[1]->height[0];
	buf_imgs[1]->yuv[2] = buf_imgs[1]->yuv[1]+buf_imgs[1]->pitch[1]*buf_imgs[1]->height[1];
	buf_imgs[2]->yuv[1] = buf_imgs[2]->yuv[0]+buf_imgs[2]->pitch[0]*buf_imgs[2]->height[0];
	buf_imgs[2]->yuv[2] = buf_imgs[2]->yuv[1]+buf_imgs[2]->pitch[1]*buf_imgs[2]->height[1];
	memcpy(buf_imgs[3],buf_imgs[0],sizeof(vaapi_yadif_img));
	memcpy(buf_imgs[4],buf_imgs[0],sizeof(vaapi_yadif_img));
	buf_imgs[3]->yuv[0] = malloc(buf_imgs[3]->yuvsize);
	buf_imgs[4]->yuv[0] = malloc(buf_imgs[4]->yuvsize);
	buf_imgs[3]->yuv[1] = buf_imgs[3]->yuv[0]+buf_imgs[3]->pitch[0]*buf_imgs[3]->height[0];
	buf_imgs[3]->yuv[2] = buf_imgs[3]->yuv[1]+buf_imgs[3]->pitch[1]*buf_imgs[3]->height[1];
	buf_imgs[4]->yuv[1] = buf_imgs[4]->yuv[0]+buf_imgs[4]->pitch[0]*buf_imgs[4]->height[0];
	buf_imgs[4]->yuv[2] = buf_imgs[4]->yuv[1]+buf_imgs[4]->pitch[1]*buf_imgs[4]->height[1];
}


void VaapiYadif (VADisplay *display, VAImage *src, VAImage *dst1,
    VAImage *dst2, vaapi_yadif_img *buf_imgs[], int tff, int skip_spatial)
{
	vaapi_yadif_img *tmp;

	if(buf_imgs[2]==NULL) { InitVaapiYadif(display, src, buf_imgs); }
	else { 	CopyVAImg2YadifImg(display, src, buf_imgs[2]); }
	
	skip_spatial<<=1;
	int y, i;

	for (i = 0; i < 3; i++) {
		int w = buf_imgs[3]->width[i];
		int h = buf_imgs[3]->height[i];
		int refs = buf_imgs[3]->pitch[i];

		for (y = 0; y < h; y++) {
			if ((y ^ tff) & 1) {
				uint8_t *prev = &buf_imgs[0]->yuv[i][y*refs];
				uint8_t *cur  = &buf_imgs[1]->yuv[i][y*refs];
				uint8_t *next = &buf_imgs[2]->yuv[i][y*refs];
				// vdr_plg_shddev_yadif_filter_line_ssse3
				filter_line_c(&buf_imgs[3]->yuv[i][y*refs], prev, cur, next, w, y+1<h ? refs : -refs, y ? -refs : refs, 0, (y==1 || y+2==h) ? 2 : skip_spatial);
				memcpy(&buf_imgs[4]->yuv[i][y*refs], &buf_imgs[1] ->yuv[i][y*refs], w);
			} else {
				uint8_t *prev = &buf_imgs[0]->yuv[i][y*refs];
				uint8_t *cur  = &buf_imgs[1] ->yuv[i][y*refs];
				uint8_t *next = &buf_imgs[2]->yuv[i][y*refs];
				filter_line_c(&buf_imgs[4]->yuv[i][y*refs], prev, cur, next, w, y+1<h ? refs : -refs, y ? -refs : refs, 1, (y==1 || y+2==h) ? 2 : skip_spatial);
				memcpy(&buf_imgs[3]->yuv[i][y*refs], &buf_imgs[1] ->yuv[i][y*refs], w);
			}
		}
	}

	CopyYadifImg2VAImg(display,buf_imgs[3],dst1);
	CopyYadifImg2VAImg(display,buf_imgs[4],dst2);

	tmp = buf_imgs[0];
	buf_imgs[0] = buf_imgs[1];
	buf_imgs[1] = buf_imgs[2];
	buf_imgs[2] = tmp;
}

Source code

im decoder struct:
...
    VAImage DeintImages[5];		///< deinterlace image buffers
    vaapi_yadif_img* yadif_images[5];
    int GetPutImage;			///< flag get/put image can be used
...
in VaapiNewHwDecoder:
...
    decoder->Image->image_id = VA_INVALID_ID;
    memset(decoder->yadif_images,0,5*sizeof(void*));
    for (i = 0; i < CODEC_SURFACES_MAX; ++i) {
...
in VaapiCpuDerive:
...
	case VideoDeinterlaceSoftSpatial:
	    VaapiSpatial(decoder, image, dest1, dest2);
	    break;
	case VideoDeinterlaceSoftTemporal:
	    VaapiYadif(decoder->VaDisplay, image, dest1, dest2, decoder->yadif_images, decoder->TopFieldFirst, 1);
	    break;
	case VideoDeinterlaceSoftTemporalSpatial:
	    VaapiYadif(decoder->VaDisplay, image, dest1, dest2, decoder->yadif_images, decoder->TopFieldFirst, 0);
	    break;
    }
...

Source code

libva: VA-API version 0.33.0
libva: va_getDriverName() returns 0
libva: Trying to open /usr/lib/x86_64-linux-gnu/dri/i965_drv_video.so
libva: Found init function __vaDriverInit_0_33
libva: va_openDriver() returns 0
video/vaapi: libva 0.33 (Intel i965 driver - 1.0.18) initialized
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
0 ms / frame
done

Source code

(EE) intel(0): [DRI2] DRI2SwapComplete: bad drawable
(EE) intel(0): [DRI2] DRI2SwapComplete: bad drawable

Source code

-Audiosynchronisation: Das Bild hängt wie erwartet um 40ms, da muss der Ton angepasst werden
-Speicher Clean-Up: Bei Senderwechsel (zumindest bei Formatwechsel) muss der Speicher wieder freigegeben und die threads beendet werden,
                    damit yadif sich beim nächsten mal neu initialisiert und nicht abstürzt
-Makefile: yasm integrieren
-Komplett ungetestet auf 32 Bit
-Eine config.h / config.asm erstellen, welche die nötigen infos enthält (32/64 Bit, sse2/ssse3 vorhanden)
-Skip Chroma ist noch nicht implementiert
-Das erste Bild scheint fehlerhaft zu sein