[Mesa-dev] [PATCH 1/4] nir: add support for min/max/median of 3 srcs

[Ian Romanick]

This patch is

Reviewed-by: Ian Romanick <ian.d.romanick at intel.com>

I think we should also add a flag for lowering this instruction as is
done in GLSL IR.  You should then add some patterns to
nir_opt_algebraic.py that recognize open-code versions.  It would also
be interesting to detect clamp() patterns and convert them to med3.

I don't have any expectation that you would implement the actual lowering.

On 03/08/2018 02:14 PM, Daniel Schürmann wrote:
> From: Dave Airlie <airlied at redhat.com>
> 
> These are needed for SPV_AMD_shader_trinary_minmax,
> the AMD HW supports these.
> 
> Co-authored-by: Daniel Schürmann <daniel.schuermann at campus.tu-berlin.de>
> Signed-off-by: Dave Airlie <airlied at redhat.com>
> ---
>  src/compiler/nir/nir_opcodes.py | 14 ++++++++++++++
>  1 file changed, 14 insertions(+)
> 
> diff --git a/src/compiler/nir/nir_opcodes.py b/src/compiler/nir/nir_opcodes.py
> index 65d1320062..fdf90ef3f5 100644
> --- a/src/compiler/nir/nir_opcodes.py
> +++ b/src/compiler/nir/nir_opcodes.py
> @@ -658,6 +658,20 @@ triop("flrp", tfloat, "src0 * (1 - src2) + src1 * src2")
>  
>  
>  triop("fcsel", tfloat32, "(src0 != 0.0f) ? src1 : src2")
> +
> +# 3 way min/max/med
> +triop("fmin3", tfloat, "fminf(src0, fminf(src1, src2))")
> +triop("imin3", tint, "MIN2(src0, MIN2(src1, src2))")
> +triop("umin3", tuint, "MIN2(src0, MIN2(src1, src2))")
> +
> +triop("fmax3", tfloat, "fmaxf(src0, fmaxf(src1, src2))")
> +triop("imax3", tint, "MAX2(src0, MAX2(src1, src2))")
> +triop("umax3", tuint, "MAX2(src0, MAX2(src1, src2))")
> +
> +triop("fmed3", tfloat, "fmaxf(fminf(fmaxf(src0, src1), src2), fminf(src0, src1))")
> +triop("imed3", tint, "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
> +triop("umed3", tuint, "MAX2(MIN2(MAX2(src0, src1), src2), MIN2(src0, src1))")
> +
>  opcode("bcsel", 0, tuint, [0, 0, 0],
>        [tbool, tuint, tuint], "", "src0 ? src1 : src2")
>  
>