[Mesa-dev] [PATCH] anv/cmd_buffer: Re-emit the pipeline at every subpass

[Iago Toral]

On Fri, 2018-01-26 at 16:23 -0800, Jason Ekstrand wrote:
> If we ever hit this edge-case, it can theoretically cause problem for
> CNL because we could end up changing render targets without re-
> emitting
> 3DSTATE_MULTISAMPLE which is part of the pipeline.  Just get rid of
> the
> edge case.
> 
> Cc: mesa-stable at lists.freedesktop.org
> ---
>  src/intel/vulkan/genX_cmd_buffer.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/src/intel/vulkan/genX_cmd_buffer.c
> b/src/intel/vulkan/genX_cmd_buffer.c
> index e8d44d0..bf6e267 100644
> --- a/src/intel/vulkan/genX_cmd_buffer.c
> +++ b/src/intel/vulkan/genX_cmd_buffer.c
> @@ -3200,6 +3200,17 @@ genX(cmd_buffer_set_subpass)(struct
> anv_cmd_buffer *cmd_buffer,
>     if (GEN_GEN == 7)
>        cmd_buffer->state.gfx.vb_dirty |= ~0;
>  
> +   /* It is possible to start a render pass with an old
> pipeline.  Because the
> +    * render pass and subpass index are both baked into the
> pipeline, this is
> +    * highly unlikely.  In order to do so, it requires that you have
> a render
> +    * pass with a single subpass and that you use that render pass
> twice
> +    * back-to-back and use the same pipeline at the start of the
> second render
> +    * pass as at the end of the first.  In order to avoid
> unpredictable issues
> +    * with this edge case, we just dirty the pipeline at the start
> of every
> +    * subpass.
> +    */
> +   cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_PIPELINE;
> +

I was going to say that since this can only happen when there is a
single subpass in the renderpass we could just do this at
CmdBeginRenderPass instead, but I guess it doesn't really matter since
apps need to bind a new pipeline at each subpass anyway and that will
dirty this state. It would be redundant for most cases but I guess it
does't hurt.

Reviewed-by: Iago Toral Quiroga <itoral at igalia.com>

Iago

>     /* Perform transitions to the subpass layout before any writes
> have
>      * occurred.
>      */