[Mesa-dev] [PATCH] [rfc] spirv/nir: handle casting in OpSampledImage

[Jason Ekstrand]

On Mon, Sep 25, 2017 at 7:36 PM, Dave Airlie <airlied at gmail.com> wrote:

> From: Dave Airlie <airlied at redhat.com>
>
> %9903 = OpImageSampleDrefExplicitLod %float %14616 %14315 %16081 Lod
> %float_0
> %14616 = OpSampledImage %510 %8499 %13137
>
> %278 = OpTypeImage %float 2D 1 0 0 1 Unknown
> %510 = OpTypeSampledImage %278
>
> %8499 = OpLoad %150 %4159
> %150 = OpTypeImage %float 2D 0 0 0 1 Unknown
>
> Is being generated by a hlsl->spirv convertor.
>
> So it appears that the 510 return type from sampledimage is a shadow
> sampler,
> however the descriptor it's loading is for a 2D non-shadow image, which
> makes it seems like OpSampledImage should cast appropriately.
>
> Now I'm not sure enough to know if this is valid spir-v in the first place,
> and I don't think this patch is good enough to fix it even if it is, but
> let's use it to open discussions.
>

Yeah, that's not a thing.... From the SPIR-V spec for OpLoad:

    *Pointer* is the pointer to load through. Its type must be an
*OpTypePointer* whose *Type* operand is the same as *Result Type*.

You're not allowed to do an implicit cast in an OpLoad.  My guess is that
HLSL only makes a distinction between shadow and non-shadow sampling at the
sample instruction and not at the interface point.  GL, however, has lots
of history encoded in these things...

Please file a bug against GLSLang.  If it's too hard to fix there, I'm
happy for John to kick it back to the SPIR-V working group and try to come
up with a way to do this properly.  I'm reluctant to just let it through
because I have no idea what will happen in our hardware if you try to do a
depth comparison on an RGBA8 image.

--Jason

---
>  src/compiler/spirv/spirv_to_nir.c | 6 +++++-
>  src/compiler/spirv/vtn_private.h  | 1 +
>  2 files changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/src/compiler/spirv/spirv_to_nir.c
> b/src/compiler/spirv/spirv_to_nir.c
> index 6ce9d1a..d962e93 100644
> --- a/src/compiler/spirv/spirv_to_nir.c
> +++ b/src/compiler/spirv/spirv_to_nir.c
> @@ -1490,6 +1490,8 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp
> opcode,
>        struct vtn_value *val =
>           vtn_push_value(b, w[2], vtn_value_type_sampled_image);
>        val->sampled_image = ralloc(b, struct vtn_sampled_image);
> +      val->sampled_image->ret_type =
> +         vtn_value(b, w[1], vtn_value_type_type)->type;
>        val->sampled_image->image =
>           vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
>        val->sampled_image->sampler =
> @@ -1521,7 +1523,9 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp
> opcode,
>     }
>
>     const struct glsl_type *image_type;
> -   if (sampled.image) {
> +   if (sampled.ret_type) {
> +      image_type = sampled.ret_type->type;
> +   } else if (sampled.image) {
>        image_type = sampled.image->var->var->interface_type;
>     } else {
>        image_type = sampled.sampler->var->var->interface_type;
> diff --git a/src/compiler/spirv/vtn_private.h b/src/compiler/spirv/vtn_
> private.h
> index 8458462..ce26326 100644
> --- a/src/compiler/spirv/vtn_private.h
> +++ b/src/compiler/spirv/vtn_private.h
> @@ -413,6 +413,7 @@ struct vtn_image_pointer {
>  struct vtn_sampled_image {
>     struct vtn_pointer *image; /* Image or array of images */
>     struct vtn_pointer *sampler; /* Sampler */
> +   struct vtn_type *ret_type;
>  };
>
>  struct vtn_value {
> --
> 2.9.5
>
> _______________________________________________
> 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/20170926/b071f982/attachment-0001.html>