Re: [PATCH 1/2] media: qcom: iris: venc: keep the crop rectangle in visible resolution space

From: Bryan O'Donoghue

Date: Fri Nov 13 2026 - 11:38:08 EST


On 12/11/2026 16:15, Loic Poulain wrote:
The encoder keeps two source resolutions. inst->fmt_src holds the padded
buffer geometry handed to userspace, with width aligned to 128 and height
to 32. enc_raw_width and enc_raw_height hold the visible resolution sent
to the firmware as HFI_PROP_RAW_RESOLUTION.

inst->crop must use the visible resolution, because
iris_hfi_gen2_set_crop_offsets() computes

bottom_offset = inst->enc_raw_height - inst->crop.height;

iris_venc_s_fmt_input() seeded inst->crop from the padded fmt_src instead,
so any height that is not a multiple of 32 underflows.

I think this explains the bug enough TBH.

<snip>For 1280x720,
crop.height becomes 736 and bottom_offset becomes 0xfffffff0. The two
offsets share one payload word, so the right offset is corrupted too.

iris_venc_s_fmt_output() compared the requested coded size against
fmt_src, so a plain 1280x720 session looked like a 736 -> 720 downscale.
Seed fmt_dst from the visible size. The encoder does not use fmt_dst for
stride programming.

Seeding fmt_dst from the visible size also means the codec alignment has
to be applied on every path and not only when scaling, so that the
CAPTURE queue keeps reporting the coded size. For 1920x1080 that is 1088,
and the firmware signals the visible 1080 with SPS frame cropping.
</snip>

Report CROP_BOUNDS and CROP_DEFAULT from enc_raw_*, so the bounds match
the real source frame.

G_FMT is left as is. For contiguous NV12 pix_mp.height is the only way to
report the vertical luma padding, and the hardware is programmed with
scanline_y = ALIGN(h, 32).

v4l2-ctl -d /dev/video1 \
--set-fmt-video-out=width=1280,height=720,pixelformat=NV12 \
--get-selection-output=target=crop

Before:

Selection Video Output: crop, Left 0, Top 0, Width 1280, Height 736

After:

Selection Video Output: crop, Left 0, Top 0, Width 1280, Height 720

Nice.

Fixes: d9967fa37ca9 ("media: qcom: iris: Improve format alignment for encoder")
Signed-off-by: Loic Poulain <loic.poulain@xxxxxxxxxxxxxxxx>
---
drivers/media/platform/qcom/iris/iris_venc.c | 20 ++++++++++++--------
drivers/media/platform/qcom/iris/iris_vidc.c | 4 ++--
2 files changed, 14 insertions(+), 10 deletions(-)

diff --git a/drivers/media/platform/qcom/iris/iris_venc.c b/drivers/media/platform/qcom/iris/iris_venc.c
index 2f2c56bf9122c73e10e86815b1aa5fad99b0fb42..b367367ac2ccf874a00606b8b1352fed37b37a57 100644
--- a/drivers/media/platform/qcom/iris/iris_venc.c
+++ b/drivers/media/platform/qcom/iris/iris_venc.c
@@ -240,14 +240,18 @@ static int iris_venc_s_fmt_output(struct iris_inst *inst, struct v4l2_format *f)
* If output format size != input format size,
* it is considered a scaling case,
* and the scaled size needs to be saved.
+ * Compare against the visible source resolution.
*/
- if (f->fmt.pix_mp.width != inst->fmt_src->fmt.pix_mp.width ||
- f->fmt.pix_mp.height != inst->fmt_src->fmt.pix_mp.height) {
+ if (f->fmt.pix_mp.width != inst->enc_raw_width ||
+ f->fmt.pix_mp.height != inst->enc_raw_height) {
inst->enc_scale_width = f->fmt.pix_mp.width;
inst->enc_scale_height = f->fmt.pix_mp.height;
- fmt->fmt.pix_mp.width = ALIGN(f->fmt.pix_mp.width, codec_align);
- fmt->fmt.pix_mp.height = ALIGN(f->fmt.pix_mp.height, codec_align);
}
+
+ /* The CAPTURE queue reports the coded size, in whole macroblocks. */
+ fmt->fmt.pix_mp.width = ALIGN(f->fmt.pix_mp.width, codec_align);
+ fmt->fmt.pix_mp.height = ALIGN(f->fmt.pix_mp.height, codec_align);
+
fmt->fmt.pix_mp.num_planes = 1;
fmt->fmt.pix_mp.plane_fmt[0].bytesperline = 0;
fmt->fmt.pix_mp.plane_fmt[0].sizeimage = iris_get_buffer_size(inst, BUF_OUTPUT);
@@ -293,8 +297,8 @@ static int iris_venc_s_fmt_input(struct iris_inst *inst, struct v4l2_format *f)
fmt->fmt.pix_mp.quantization = f->fmt.pix_mp.quantization;
output_fmt = inst->fmt_dst;
- output_fmt->fmt.pix_mp.width = fmt->fmt.pix_mp.width;
- output_fmt->fmt.pix_mp.height = fmt->fmt.pix_mp.height;
+ output_fmt->fmt.pix_mp.width = f->fmt.pix_mp.width;
+ output_fmt->fmt.pix_mp.height = f->fmt.pix_mp.height;
output_fmt->fmt.pix_mp.colorspace = fmt->fmt.pix_mp.colorspace;
output_fmt->fmt.pix_mp.xfer_func = fmt->fmt.pix_mp.xfer_func;
output_fmt->fmt.pix_mp.ycbcr_enc = fmt->fmt.pix_mp.ycbcr_enc;
@@ -312,8 +316,8 @@ static int iris_venc_s_fmt_input(struct iris_inst *inst, struct v4l2_format *f)
f->fmt.pix_mp.height != inst->crop.height) {
inst->crop.top = 0;
inst->crop.left = 0;
- inst->crop.width = fmt->fmt.pix_mp.width;
- inst->crop.height = fmt->fmt.pix_mp.height;
+ inst->crop.width = f->fmt.pix_mp.width;
+ inst->crop.height = f->fmt.pix_mp.height;
iris_venc_s_fmt_output(inst, output_fmt);
}
diff --git a/drivers/media/platform/qcom/iris/iris_vidc.c b/drivers/media/platform/qcom/iris/iris_vidc.c
index 05aab6e3005d5c3788f5dfe5857f2d6b782e7a83..e3eaec1b7c2e30c065d8419ac3278756f70a8a97 100644
--- a/drivers/media/platform/qcom/iris/iris_vidc.c
+++ b/drivers/media/platform/qcom/iris/iris_vidc.c
@@ -522,8 +522,8 @@ static int iris_g_selection(struct file *filp, void *fh, struct v4l2_selection *
switch (s->target) {
case V4L2_SEL_TGT_CROP_BOUNDS:
case V4L2_SEL_TGT_CROP_DEFAULT:
- s->r.width = inst->fmt_src->fmt.pix_mp.width;
- s->r.height = inst->fmt_src->fmt.pix_mp.height;
+ s->r.width = inst->enc_raw_width;
+ s->r.height = inst->enc_raw_height;
break;
case V4L2_SEL_TGT_CROP:
s->r.width = inst->crop.width;


Could crop and scaling go into separate patches ? Feels like it could be made more granular.

---
bod