On Monday 30 March 2009 16:34:42 Hiremath, Vaibhav wrote:
> Hi,
>
> With reference to the mail-thread started by Sakari on Resizer driver
> interface,
>
> http://marc.info/?l=linux-omap&m=123628392325716&w=2
>
> I would like to bring some issues and propose changes to adapt such
> devices under V4L2 framework. Sorry for delayed response on this
> mail-thread, actually I was on vacation.
>
> As proposed by Sakari, I do agree with the approach of having V4L2
> interface for memory-to-memory operation of the ISP (like resizer and
> previewer), but I would like to bring some important aspects/issues here
> -
>
> - Some drawbacks of V4L2-buf layer framework for such kind of devices
> - Need for backward compatibility.
>
> Drawbacks for V4L2-Buf layer -
> -----------------------------
>
> 1} In case of resizer driver, the input buffer will always be different
> than output buffer.
>
> In case of Mmapped buffer there is no issue, since driver allocates
> maximum of input and output. User doesn't have control on this, although
> there is loss of memory from system point of view.
>
> In case of User Pointer Exchange, User would expect and may allocate
> different sizes of buffers for input and output which V4L2-buf layer
> doesn't support with single queue. And to address this, I think here we
> need to have either of following approaches -
>
> - Use two separate buffer queues, one for input and another for output.
> - Or hack the driver for v4l2_buffer, internally using different buffer
> params for input and output. [Not recommended]
>
> Please note that sometimes application receives buffers from another
> driver or from some codecs engine that's why input and output buffer will
> never be of same size.
>
> 2) V4L2-buf layer doesn't support IOMEM coming from user application.
> Just to give low level details about this -
>
> When application tries to configure for 'V4L2_MEMORY_USERPTR' with buffer
> coming from another driver (which is iomampped), then QUEUEBUF ioctl will
> fail from 'videobuf_iolock' --> videobuf_dma_init_user_locked -->
> get_user_pages.
>
> In 'get_user_pages' it checks for IOMEM flag and returns error, which is
> expected behavior from Kernel point of view. We are trying to map buffer
> which is already mapped to kernel by another driver.
>
> One thing I am not able to understand, how nobody came across such
> use-case which is very common. And to address this issue we need to add
> support for IOMEM in V4L2-buf layer.
>
> NOTE: Currently both of these issues have been addressed as a custom
> implementation for our internal use case.
>
> Backward Compatibility -
> -----------------------
>
> This is an important aspect, since similar hardware modules are available
> on Davinci as well as on OMAP and their driver interface is completely
> different.
>
> On Davinci, resizer driver is supported through plane char driver
> interface which handles all data/buffer processing and control paths. It
> maintains internal queue for priority of resizing tasks and stuff.
>
> The driver is present under /drivers/char/Davinci.
>
> Here I feel, V4L2 way is better, since all image processing drivers
> should go under "drivers/media/video/". And again we can make use of
> readily available V4L2 framework interface for data and control path to
> manage buffers. We should enhance V4L2 framework to support such devices.
>
>
> Proposed Required Changes -
> --------------------------
>
> I am putting some high level changes required to be done for supporting
> such devices -
>
> - Enhancement in V4L2-buf layer for above issues
>
> - Will be directly using sub-device frame-work and have to enhance it to
> support such devices.
>
> - Below are the parameters we need to configure for Resizer from user
> application -
>
> __s32 in_hsize; /* input frame horizontal size.*/
> __s32 in_vsize; /* input frame vertical size */
> __s32 in_pitch; /* offset between 2 rows of input frame.*/
> __s32 inptyp; /* for determining 16 bit or 8 bit data.*/
> __s32 vert_starting_pixel; /* vertical starting pixel in input.*/
> __s32 horz_starting_pixel; /* horizontal starting pixel in input.*/
> __s32 cbilin; /* filter with luma or bi-linear interpolation.*/
> __s32 pix_fmt; /* UYVY or YUYV */
> __s32 out_hsize; /* output frame horizontal size. */
> __s32 out_vsize; /* output frame vertical size.*/
> __s32 out_pitch; /* offset between 2 rows of output frame.*/
> __s32 hstph; /* for specifying horizontal starting phase.*/
> __s32 vstph; /* for specifying vertical starting phase.*/
> __u16 tap4filt_coeffs[32]; /* horizontal filtercoefficients.*/
> __u16 tap7filt_coeffs[32]; /* vertical filter coefficients. */
> struct rsz_yenh yenh_params;
>
> (Pasted from previous patches posted by Sergio)
>
>
> Putting one sample proposal using VIDIOC_S_FMT -
>
> APPROACH 1 -
> ----------
>
> Either we can add this under "struct v4l2_format" or need to enhance
> "stuct v4l2_crop" to support such device.
>
> We can use 'VIDIOC_S_FMT' ioctl to configure the resizer parameters. From
> top level it will look something like -
>
> struct v4l2_buffer buf;
> struct v4l2_format fmt;
> struct rsz_parm parm; /*Contains custom configuration specific to module
> other than input and output params*/
>
> <Fill the "rsz_parm" structure>
>
> fmt.type = V4L2_BUF_TYPE_VIDEO_CAPTURE;
> fmt.fmt.pix.width = IN_WIDTH;
> fmt.fmt.pix.height = IN_HEIGHT;
> fmt.fmt.pix.bytesperline= IN_WIDTH*2;
> fmt.fmt.pix.pixelformat= V4L2_PIX_FMT_YUYV/V4L2_PIX_FMT_UYVY
> fmt.fmt.pix.priv = &parm;
>
> ret = ioctl(rsz_fd, VIDIOC_S_FMT, &fmt);
> if(ret<0) {
> perror("Set Format failed\n");
> return -1;
> }
>
> To set output buffer we can use VIDIOC_S_FMT
>
> fmt.type = V4L2_BUF_TYPE_VIDEO_OUTPUT;
> fmt.fmt.pix.width = OUT_WIDTH;
> fmt.fmt.pix.height = OUT_HEIGHT;
> fmt.fmt.pix.bytesperline= OUT_WIDTH*2;
>
> ret = ioctl(rsz_fd, VIDIOC_S_FMT, &fmt);
> if(ret<0) {
> perror("Set Format failed\n");
> return -1;
> }
>
> In case of resizer driver we would need to call VIDIOC_S_FMT twice, one
> for input params and second for output params. And in case of Previewer
> we need to call VIDIOC_S_FMT only once, with all required params as a
> part of custom struct (priv).
>
>
> APPROACH 2 -
> ----------
>
> Instead of using "priv" variable, we can make it capability based
> interface. For this we have to create separate class of devices, called
> "V4L2_BUF_TYPE_VIDEO_RESIZE" with "struct v4l2_fmt_resize" and "struct
> v4l2_fmt_previewer".
>
> And depending on the capability application will configure params using
> VIDIOC_S_FMT.
>
> In both the cases we need to add one IOCTL to trigger the operation and
> this should be standard, should be common for all such memory-to-memory
> device operations.
>
>
> APPROACH 3 -
> ----------
>
> .....
>
> (Any other approach which I could not think of would be appreciated)
>
>
> I would prefer second approach, since this will provide standard
> interface to applications independent on underneath hardware.
>
> There may be many number of such configuration parameters required for
> different such devices, we need to work on this and come up with some
> standard capability fields covering most of available devices.
>
> Does anybody have some other opinions on this?
> Any suggestions will be helpful here,
FYI: I have very little time to look at this for the next 2-3 weeks. As you
know I'm working on the last pieces of the v4l2_subdev conversion for 2.6.30
that should be finished this week. After that I'm attending the Embedded
Linux Conference in San Francisco.
But I always thought that something like this would be just a regular video
device that can do both 'output' and 'capture'. For a resizer I would
expect that you set the 'output' size (the size of your source image) and
the 'capture' size (the size of the resized image), then just send the
frames to the device (== resizer) and get them back on the capture side.
Regards,
Hans
--
Hans Verkuil - video4linux developer - sponsored by TANDBERG
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
