media: vicodec: process all 4 components for RGB32 formats
Only ARGB32-type pixelformat were assumed to have 4 components, which is wrong since RGB32-type pixelformats may have an alpha channel, so they should also assume 4 color components. The XRGB32-type pixelformats really have only 3 color components, but this complicated matters since that creates strides that are sometimes width * 3 and sometimes width * 4, and in fact this can result in buffer overflows. Keep things simple by just always processing all 4 color components. In the future we might want to optimize this again for the XRGB32-type pixelformats, but for now keep it simple and robust. Signed-off-by: Hans Verkuil <hverkuil-cisco@xs4all.nl> Cc: <stable@vger.kernel.org> # for v5.4 and up Signed-off-by: Mauro Carvalho Chehab <mchehab+huawei@kernel.org>
This commit is contained in:
Родитель
aa9eda7612
Коммит
49a56266f9
|
@ -27,17 +27,17 @@ static const struct v4l2_fwht_pixfmt_info v4l2_fwht_pixfmts[] = {
|
|||
{ V4L2_PIX_FMT_BGR24, 3, 3, 1, 3, 3, 1, 1, 3, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_RGB24, 3, 3, 1, 3, 3, 1, 1, 3, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_HSV24, 3, 3, 1, 3, 3, 1, 1, 3, 1, FWHT_FL_PIXENC_HSV},
|
||||
{ V4L2_PIX_FMT_BGR32, 4, 4, 1, 4, 4, 1, 1, 3, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_XBGR32, 4, 4, 1, 4, 4, 1, 1, 3, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_BGR32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_XBGR32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_ABGR32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_RGB32, 4, 4, 1, 4, 4, 1, 1, 3, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_XRGB32, 4, 4, 1, 4, 4, 1, 1, 3, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_RGB32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_XRGB32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_ARGB32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_BGRX32, 4, 4, 1, 4, 4, 1, 1, 3, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_BGRX32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_BGRA32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_RGBX32, 4, 4, 1, 4, 4, 1, 1, 3, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_RGBX32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_RGBA32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_RGB},
|
||||
{ V4L2_PIX_FMT_HSV32, 4, 4, 1, 4, 4, 1, 1, 3, 1, FWHT_FL_PIXENC_HSV},
|
||||
{ V4L2_PIX_FMT_HSV32, 4, 4, 1, 4, 4, 1, 1, 4, 1, FWHT_FL_PIXENC_HSV},
|
||||
{ V4L2_PIX_FMT_GREY, 1, 1, 1, 1, 0, 1, 1, 1, 1, FWHT_FL_PIXENC_RGB},
|
||||
};
|
||||
|
||||
|
@ -175,22 +175,14 @@ static int prepare_raw_frame(struct fwht_raw_frame *rf,
|
|||
case V4L2_PIX_FMT_RGB32:
|
||||
case V4L2_PIX_FMT_XRGB32:
|
||||
case V4L2_PIX_FMT_HSV32:
|
||||
rf->cr = rf->luma + 1;
|
||||
rf->cb = rf->cr + 2;
|
||||
rf->luma += 2;
|
||||
break;
|
||||
case V4L2_PIX_FMT_BGR32:
|
||||
case V4L2_PIX_FMT_XBGR32:
|
||||
rf->cb = rf->luma;
|
||||
rf->cr = rf->cb + 2;
|
||||
rf->luma++;
|
||||
break;
|
||||
case V4L2_PIX_FMT_ARGB32:
|
||||
rf->alpha = rf->luma;
|
||||
rf->cr = rf->luma + 1;
|
||||
rf->cb = rf->cr + 2;
|
||||
rf->luma += 2;
|
||||
break;
|
||||
case V4L2_PIX_FMT_BGR32:
|
||||
case V4L2_PIX_FMT_XBGR32:
|
||||
case V4L2_PIX_FMT_ABGR32:
|
||||
rf->cb = rf->luma;
|
||||
rf->cr = rf->cb + 2;
|
||||
|
@ -198,10 +190,6 @@ static int prepare_raw_frame(struct fwht_raw_frame *rf,
|
|||
rf->alpha = rf->cr + 1;
|
||||
break;
|
||||
case V4L2_PIX_FMT_BGRX32:
|
||||
rf->cb = rf->luma + 1;
|
||||
rf->cr = rf->cb + 2;
|
||||
rf->luma += 2;
|
||||
break;
|
||||
case V4L2_PIX_FMT_BGRA32:
|
||||
rf->alpha = rf->luma;
|
||||
rf->cb = rf->luma + 1;
|
||||
|
@ -209,10 +197,6 @@ static int prepare_raw_frame(struct fwht_raw_frame *rf,
|
|||
rf->luma += 2;
|
||||
break;
|
||||
case V4L2_PIX_FMT_RGBX32:
|
||||
rf->cr = rf->luma;
|
||||
rf->cb = rf->cr + 2;
|
||||
rf->luma++;
|
||||
break;
|
||||
case V4L2_PIX_FMT_RGBA32:
|
||||
rf->alpha = rf->luma + 3;
|
||||
rf->cr = rf->luma;
|
||||
|
|
Загрузка…
Ссылка в новой задаче