SK-AM62P-LP: How to obtain GPU rendered buffer?

Hi,

Can you share an example of TDA2/TDA4 that you used in the past? 

Regards,
Krunal

Some code snippets are as follows：

        pSystemVideoFrameBuffer->bufAddr[0] = OSA_memAllocSR(
                                           OSA_HEAPID_DDR_CACHED_SR1,
                                           (maxHeight*(pFrameCopyObj->pitch)*2),
                                           SYSTEM_BUFFER_ALIGNMENT);

        /*
         * Carving out memory pointer for chroma which will get used in case of
         * SYSTEM_DF_YUV422SP_UV
         */
        pSystemVideoFrameBuffer->bufAddr[1] = (void*)(
            (UInt32) pSystemVideoFrameBuffer->bufAddr[0] +
            (UInt32)(maxHeight*(pFrameCopyObj->pitch))
            );

        OSA_assert(pSystemVideoFrameBuffer->bufAddr[0] != NULL);

        System_exportDmaBuf(pSystemVideoFrameBuffer->bufAddr[0],
                                (UInt32)(maxHeight*(pFrameCopyObj->pitch)*2),
                                &pSystemVideoFrameBuffer->dmaFd[0]);

Among them, the System_dexportDmaBuf function is implemented as follows：

Int32 System_exportDmaBuf(void * vAddr, UInt32 size, UInt32 *fdBuf)
{
    Int32 retVal = -1;
#ifndef QNX_BUILD
    struct dmabuf_vmem_export exp;
    exp.vaddr = (unsigned long)vAddr;
    exp.size = size;

    if(gSystem_objCommon.devBufFD > 0)
    {
        /* Export as DMAbuf handle */
        retVal = ioctl(gSystem_objCommon.devBufFD, DBUFIOC_EXPORT_VIRTMEM, &exp);
        if(retVal == SYSTEM_LINK_STATUS_SOK)
        {
            *fdBuf = exp.fd;
        }
        else
        {
            Vps_printf(" exportDmaBuf failed \n ");
        }
    }
#else
    retVal = SYSTEM_LINK_STATUS_SOK;
#endif
return retVal;
}

Hi,

Are you referring to the following: https://www.ti.com/lit/an/spracg2/spracg2.pdf?ts=1725553055154&ref_url=https%253A%252F%252Fwww.google.com%252F ? If yes, what's your usecase?

Regards,
Krunal

I'm sorry, I'm not very familiar with use cases. Chapter 3.1 in the link you provided is basically the same as what I saw using it. We use the name lvds-vip_multi_cam_3d_srv  on TDA2

I am not sure I follow, what do you plan on doing with the above code and what are you trying to do with the GPU? The app note is very old and we don't even support it on AM62P. 

Regards,
Krunal

We need to send the image data processed by GPU through RTSP, which requires obtaining GPU rendered buffer. The above method can easily obtain GPU rendered buffer with no memcopy

Hi,

I recommend you to look at something more modern like: https://gitlab.freedesktop.org/wayland/weston/-/blob/master/clients/simple-dmabuf-egl.c. In the example, you can do something like the following:

    void* mapped_ptr;
    void *bsdata;
    mapped_ptr = gbm_bo_map(buffer->bo,0,0,buffer->width,buffer->height,GBM_BO_TRANSFER_READ_WRITE,buffer->strides,&bsdata);
    if (mapped_ptr) {
        //printf("I got pointer\n");
        memset(mapped_ptr, 0xFFFFFFFF, 4 * buffer->height/2 * buffer->width/2);
    }
    else {
        printf("None");
    }
    gbm_bo_unmap(buffer->bo, bsdata);

More about GBM_BO_MAP API: https://github.com/anholt/mesa/blob/master/src/gbm/main/gbm.c#L526. In summary, GPU is reading and writing to a memory location and using the gbm_bo_map function, we mapped the GPU buffer and passed the mapped address to CPU. CPU can perform any operation on the buffer and for the above code, I used memset to draw a white rectangle. Lastly, ensure you unmap the buffer to free the memory.

Regards,
Krunal

Hi,

Actually, we have already done this. Our pipeline is GPU render buffer -> GStreamer encoding -> RTSP streaming. There are two questions: Can I directly send this address to the GStreamer pipeline for encoding? The other question is, if the subsequent link takes too long and the address is not released, won't it affect the GPU rendering frame rate? Based on these questions, I want to use the method of allocating multiple dma_buf as the GPU render buffer queue to improve the frame rate.

Regards

Hi,

I would recommend you to review the kmscube example: https://gitlab.freedesktop.org/mesa/kmscube. In that example, there a zero copy buffer between decoder and GPU. You can apply the concepts from GPU to encoder.

Regards,
Krunal

Hi, Krunal:

Are you referring to the method of creating an EGLImage? I will try that.

Yes, they use eglCreateImageKHR.

Regards,
Krunal

Hi, Krunal:

Before calling eglCreateImageKHR to create an EGLImage, configure the egl properties section. There is still a logical decision to determine whether the output gstbuf from decode is a dmabuf. If not, a bo object will be created through gbm for value copying. Therefore, in any case, the dma_buf is required. It seems that our own application of gstbuf is not a dma_buf

if (is_dmabuf_mem) {
	dmabuf_fd = dup(gst_dmabuf_memory_get_fd(mem));
} else {
	GstMapInfo map_info;
	gst_buffer_map(buf, &map_info, GST_MAP_READ);
	dmabuf_fd = buf_to_fd(dec->gbm, map_info.size, map_info.data);
	gst_buffer_unmap(buf, &map_info);
}

static int
buf_to_fd(const struct gbm *gbm, int size, void *ptr)
{
	struct gbm_bo *bo;
	void *map, *map_data = NULL;
	uint32_t stride;
	int fd;

	/* NOTE: do not actually use GBM_BO_USE_WRITE since that gets us a dumb buffer: */
	bo = gbm_bo_create(gbm->dev, size, 1, GBM_FORMAT_R8, GBM_BO_USE_LINEAR);

	map = gbm_bo_map(bo, 0, 0, size, 1, GBM_BO_TRANSFER_WRITE, &stride, &map_data);

	memcpy(map, ptr, size);

	gbm_bo_unmap(bo, map_data);

	fd = gbm_bo_get_fd(bo);

	/* we have the fd now, no longer need the bo: */
	gbm_bo_destroy(bo);

	return fd;
}

Hi,

When you try running the above example on AM62P, does it not show zero-copy equals true?

Regards,
Krunal

Hi Krunal,

From the verification, both branches of the example are executed. The key point of my problem is that I cannot ensure that the gst_buffer generated by image_data_to_gst_buffer has the same dma_buf attribute. If it is not a dma_buf, I cannot obtain its address.

Hello Quan,

Krunal is on vacation this week. Feel free to ping the thread if you do not get a response by September 25th.

Regards,

Nick

Hi,

My recommendation would be to work with a 3P like KDAB(https://www.kdab.com/) who are more acquainted with programming OpenGLES applications. They have worked with TI on AM62P platform and will be able to assist you with userspace development.

Regards,
Krunal