[Mesa-dev] [PATCH 1/5] nir: Define image load, store and atomic intrinsics.

[Kenneth Graunke]

On Tuesday, May 05, 2015 11:29:52 PM Francisco Jerez wrote:
> ---
>  src/glsl/nir/nir_intrinsics.h | 27 +++++++++++++++++++++++++++
>  1 file changed, 27 insertions(+)
> 
> diff --git a/src/glsl/nir/nir_intrinsics.h b/src/glsl/nir/nir_intrinsics.h
> index 8e28765..4b13c75 100644
> --- a/src/glsl/nir/nir_intrinsics.h
> +++ b/src/glsl/nir/nir_intrinsics.h
> @@ -89,6 +89,33 @@ ATOMIC(inc, 0)
>  ATOMIC(dec, 0)
>  ATOMIC(read, NIR_INTRINSIC_CAN_ELIMINATE)
>  
> +/*
> + * Image load, store and atomic intrinsics.
> + *
> + * All image intrinsics take an image target passed as a nir_variable.  Image
> + * variables contain a number of memory and layout qualifiers that influence
> + * the semantics of the intrinsic.
> + *
> + * All image intrinsics take a four-coordinate vector and a sample index as
> + * first two sources, determining the location within the image that will be
> + * accessed by the intrinsic.  Components not applicable to the image target
> + * in use are equal to zero by convention.  Image store takes an additional
> + * four-component argument with the value to be written, and image atomic
> + * operations take either one or two additional scalar arguments with the same
> + * meaning as in the ARB_shader_image_load_store specification.
> + */
> +INTRINSIC(image_load, 2, ARR(4, 1), true, 4, 1, 0,
> +          NIR_INTRINSIC_CAN_ELIMINATE)
> +INTRINSIC(image_store, 3, ARR(4, 1, 4), false, 0, 1, 0, 0)
> +INTRINSIC(image_atomic_add, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
> +INTRINSIC(image_atomic_min, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
> +INTRINSIC(image_atomic_max, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
> +INTRINSIC(image_atomic_and, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
> +INTRINSIC(image_atomic_or, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
> +INTRINSIC(image_atomic_xor, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
> +INTRINSIC(image_atomic_exchange, 3, ARR(4, 1, 1), true, 1, 1, 0, 0)
> +INTRINSIC(image_atomic_comp_swap, 4, ARR(4, 1, 1, 1), true, 1, 1, 0, 0)
> +
>  #define SYSTEM_VALUE(name, components) \
>     INTRINSIC(load_##name, 0, ARR(), true, components, 0, 0, \
>     NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
> 

Looks great, Curro.  Expanding the coordinate to a vec4 and always
having the parameters present for e.g. sample_index does reduce the
combinatorial explosion a lot.  FWIW, I also prefer undefined.

This should work out fine for SPIR-V too, it's pretty trivial to go
from their style to this (just combine the two - it's SSA after all).

These 5 are:
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20150508/71c2baa2/attachment.sig>