TDA4VH-Q1: GStreamer using v4l2h264enc, input NV12 with row width not aligned 64bytes

PL Tanguay

Part Number: TDA4VH-Q1
Other Parts Discussed in Thread: TDA4VH

Tool/software:

Hello,

I am trying to encode some NV12 input data using GStreamer and encoder v4l2h264enc on TDA4VH. I am doing a "per-frame" encoding using appsrc and appsink. The NV12 data is wrapped into a GstBuffer, and then sent to the GStreamer pipeline.

Pipeline string:

"appsrc name=mysource format=time ! "
"video/x-raw,format=NV21 ! " // NV21 to reverse channels
"v4l2h264enc name=v4l2h264enc extra-controls=\"enc,prepend_sps_and_pps_to_idr=1,video_gop_size=1\" ! "
"video/x-h264 ! "
"appsink name=mysink emit-signals=false sync=false drop=false",

It works well except that I need to explicitly restructure the input data as a continuous block memory to have the right image at decoding (decoding the encoded data). My NV12 image as width of 3848, with stride of 3904 (64 byte aligned). When doing some manipulations to provide to the pipeline NV12 data with stride of 3848 (width==stride), I have the right image output at decoded.

----

My goal is to avoid doing those manipulations, before sending to the pipeline (giving the raw NV12 data). Is there a way to properly encode (v4l2h264enc) an NV12 image with width != stride ?

5 months ago

0 Fabiana Jaimes 4 months ago

TI__Mastermind 19680 points

Hello,

Could you please provide more information about your application? Are you using any edge AI GStreamer plugins (TIOVX) in your pipeline?

Thank you,

Fabiana

0 PL Tanguay 4 months ago in reply to Fabiana Jaimes

Prodigy 110 points

Hello,

We are using the GStreamer api directly, not using the edge AI plugins from TIOVX

0 Fabiana Jaimes 4 months ago in reply to PL Tanguay

TI__Mastermind 19680 points

Hi PL Tanguay,

Are you streaming from a camera? If so, which camera are you using and does it require ISP tuning binaries?

Thank you,

Fabiana

0 PL Tanguay 4 months ago in reply to Fabiana Jaimes

Prodigy 110 points

Hello,

I am doing a "per-frame" encoding. I'm basically taking a data pointer (example cv::Mat nv12.data) that contains my image data to encode, then allocating it in a GstBuffer, and sending it in the Gst pipeline. I am not streaming from a camera.

Reminder: It works if I do : cv::Mat nv12Cloned = nv12.clone() . When doing clone(), it allocates to nv12Cloned a continuous memory block with stride == width. i'm looking at how I can give to the pipeline cv::Mat nv12 without doing a clone().

FYI, I also have access to the vx_image of the nv12 image to encode.

0 Fabiana Jaimes 4 months ago in reply to PL Tanguay

TI__Mastermind 19680 points

Hello,

Could you try making the following change and let me know if this resolves your issue?

diff --git a/kernels/hwa/vpac_viss/vx_vpac_viss_target.c b/kernels/hwa/vpac_viss/vx_vpac_viss_target.c
index 01c4bd6..6d577e3 100755
--- a/kernels/hwa/vpac_viss/vx_vpac_viss_target.c
+++ b/kernels/hwa/vpac_viss/vx_vpac_viss_target.c
@@ -1688,7 +1688,7 @@ static void tivxVpacVissSetInputParams(tivxVpacVissObj *vissObj,
     /* Set the Input Format */
     fmt->width = raw_img_desc->params.width;
     fmt->height = raw_img_desc->params.height;
-    fmt->pitch[0] = (uint32_t)raw_img_desc->imagepatch_addr[0U].stride_y;
+    fmt->pitch[0] = fmt->width * 2;
     fmt->dataFormat = FVID2_DF_RAW;

     switch (raw_img_desc->params.format[0U].pixel_container)

Thank you,

Fabiana

0 PL Tanguay 4 months ago in reply to Fabiana Jaimes

Prodigy 110 points

I haven't tried this procedure yet, but I believe making modifications to the VPAC VISS kernel is not an option for us right now. It seems a bit too "low level". Is there anything I can do around GStreamer specifically ?

0 Praveen Rao 4 months ago in reply to PL Tanguay

TI__Mastermind 48173 points

The expert( Fabiana) assigned to this thread is out of the office today. Please expect a delay in response to this query. Thanks.

0 Fabiana Jaimes 4 months ago in reply to PL Tanguay

TI__Mastermind 19680 points

Hi PL Tanguay,

I will let my colleague Rahul Ravikumar comment on this.

Thank you,

Fabiana

0 Jared McArthur 4 months ago in reply to Fabiana Jaimes

TI__Genius 17615 points

Hi,

Can you try adding tiovxmemalloc to the pipeline (before the encoder)?

Best,
Jared

0 Rahul Ravikumar 3 months ago in reply to Jared McArthur

TI__Expert 4208 points

Hi Tanguay,

Can you please try below pipeline

"appsrc name=mysource format=time ! "
"video/x-raw,format=NV21, stride-y-align=64 ! tiovxdlcolorconvert ! video/x-raw, format=NV12, stride-y-align=64 ! " // NV21 to reverse channels
"v4l2h264enc name=v4l2h264enc output-io-mode=5 extra-controls=\"enc,prepend_sps_and_pps_to_idr=1,video_gop_size=1\" ! "
"video/x-h264 ! "
"appsink name=mysink emit-signals=false sync=false drop=false",

Regards
Rahul T R

Processors

Processors forum

TDA4VH-Q1: GStreamer using v4l2h264enc, input NV12 with row width not aligned 64bytes