[Mesa-dev] [PATCH 1/5] mesa: Add infrastructure for GL_ARB_clear_buffer_object

[Marek Olšák]

On Fri, Dec 20, 2013 at 9:21 PM, Ian Romanick <idr at freedesktop.org> wrote:
>
> I think there are two ways a driver could implement this function, and I
> suspect that drivers for UMA systems may dynamically switch between the
> two.  I'm pretty sure the i965 driver will, anyway.
>
> A driver could implement this using a blitter-based or CPU-based memcpy
> function.  In that case, having a pointer to raw data in the destination
> format is what you would want.
>
> A driver could also implement this using a blitter-based or render-based
> clear function.  That is, they'd wrap the buffer object in a fake
> surface and do whatever is done for glClearBuffer*.  I had originally
> thought you'd want the internalformat in that case, but after discussing
> a number of implementation strategies with Ken, I don't think it's
> actually needed.
>
> I also support Marek's suggestion.
>
> For the record, the possible implementation strategies Ken and I discussed:
>
>  - Repeating memcpy on the CPU.
>
>  - Using the blit engine with clever source strides.
>
>  - Using the render engine with some sort of uint outputs of raw bits.
>
>  - Using transform feedback with some sort of uint outputs of raw bits.

We use DMA (memset on the GPU) for clearValueSize <= 4 and transform
feedback in the other cases. The latter sets the clear value as a
vertex with a stride of zero.

Marek