[Mesa-dev] [PATCH 05/22] anv/image: Append CCS/MCS with a clear value buffer

[Jason Ekstrand]

On Thu, Apr 27, 2017 at 11:32 AM, Nanley Chery <nanleychery at gmail.com>
wrote:

> Signed-off-by: Nanley Chery <nanley.g.chery at intel.com>
> ---
>  src/intel/vulkan/anv_image.c | 67 ++++++++++++++++++++++++++++++
> ++++++++++++++
>  1 file changed, 67 insertions(+)
>
> diff --git a/src/intel/vulkan/anv_image.c b/src/intel/vulkan/anv_image.c
> index cf34dbe3b0..8d946e8e93 100644
> --- a/src/intel/vulkan/anv_image.c
> +++ b/src/intel/vulkan/anv_image.c
> @@ -114,6 +114,71 @@ add_surface(struct anv_image *image, struct
> anv_surface *surf)
>     image->alignment = MAX2(image->alignment, surf->isl.alignment);
>  }
>
> +
> +/**
> + * For color buffers that have an auxiliary buffer enabled, there is an
> + * additional buffer that keeps track of fast clear values. Using this
> + * buffer allows us to access image subresources while being aware of
> their
> + * fast clear values in non-trivial cases.
> + *
> + * The clear values in this buffer are updated when a fast clear is
> performed
> + * on a subresource. Two synchronization operations can be performed in
> order
> + * for the next memory access to use the fast-clear value:
> + *   a. Copy the value from the buffer to the surface state object used
> for
> + *      reading (if different).
> + *   b. Do (a) onto a surface state object and use it to resolve the
> + *      subresource.
> + *
> + * The synchronization approach currently taken is as follows. We place
> + * resolves in the hands of the user via layout transitions. We also
> sometimes
> + * skip resolves when a clear value predetermined to be the default in
> other
> + * surface state objects is used in a fast clear.
> + *
> + * We fast-clear (and so cause things to get temporarily unsynchronized)
> + * whenever the hardware allows except for two cases: when in the GENERAL
> + * layout and when clearing a proper subset of a layered image. The
> exception
> + * is made for the former because a layout transition isn't required
> before
> + * sampling from an image in that layout. The exception is made for the
> latter
> + * because a layout transition isn't required between a render pass that
> + * renders to a proper subset of a layered image and another render pass
> that
> + * renders to every layer in the layered image.
> + */
> +static void
> +add_clear_value_buffer(struct anv_image * const image,
> +                       const struct anv_device * const device)
> +{
> +   assert(image && device);
> +
> +   /* Only color auxiliary buffers have use for this. */
> +   assert(anv_image_has_color_aux(image));
> +
> +   /* The offset to the buffer of clear values must be dword-aligned for
> GPU
> +    * memcpy operations. It is located immediately after the auxiliary
> surface.
> +    */
> +
> +   /* Tiled images are guaranteed to be 4K aligned, so the image alignment
> +    * should also be dword-aligned.
> +    */
> +   assert(image->alignment % 4 == 0);
> +
> +   /* Auxiliary buffers should be a multiple of 4K, so the start of the
> clear
> +    * values buffer should already be dword-aligned.
> +    */
> +   assert(image->aux_surface.isl.size % 4 == 0);
> +
> +   /* This surface should be at the very end of the image. */
> +   assert(image->size ==
> +          image->aux_surface.offset + image->aux_surface.isl.size);
> +
> +   /* Entire surface state objects are stored instead of just clear
> colors for
> +    * two reasons:
> +    *   1. Pre-SKL, we must compute some state fields that lie in the same
> +    *      dword as the clear value.
> +    *   2. Storing either set of objects requires less than a page of
> memory.
>

In the not-too-distant future, we're going to start sharing fast-cleared
images with the display server and with scanout.  (We can do this today on
SKL for the non-scanout case.  The scanout case will have to wait for new
hardware.)  When we do, there will be a clear color in a "metadata page"
which will just be 4 dwords.  We don't necessarily need to change anything
now to accommodate but be advised that we will probably want to drop to
just 1 or 4 dwords in the future.


> +    */
> +   image->size += device->isl_dev.ss.size * anv_color_aux_levels(image);
> +}
> +
>  /**
>   * Initialize the anv_image::*_surface selected by \a aspect. Then update
> the
>   * image's memory requirements (that is, the image's size and alignment).
> @@ -213,6 +278,7 @@ make_surface(const struct anv_device *dev,
>                                      &image->aux_surface.isl);
>           if (ok) {
>              add_surface(image, &image->aux_surface);
> +            add_clear_value_buffer(image, dev);
>
>              /* For images created without MUTABLE_FORMAT_BIT set, we know
> that
>               * they will always be used with the original format.  In
> @@ -236,6 +302,7 @@ make_surface(const struct anv_device *dev,
>                                   &image->aux_surface.isl);
>        if (ok) {
>           add_surface(image, &image->aux_surface);
> +         add_clear_value_buffer(image, dev);
>           image->aux_usage = ISL_AUX_USAGE_MCS;
>        }
>     }
> --
> 2.12.2
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20170502/146eb61e/attachment-0001.html>