[Mesa-dev] [PATCH] nir: Handle double-precision in fabs, frsq, fsqrt, fexp2 and flog2

[Matt Turner]

On Thu, May 19, 2016 at 3:36 AM, Iago Toral Quiroga <itoral at igalia.com> wrote:
> We agreed in the list that it would be better to have these if they were
> easy to implement.
> ---
>  src/compiler/nir/nir_opcodes.py | 10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/src/compiler/nir/nir_opcodes.py b/src/compiler/nir/nir_opcodes.py
> index 8a3a80f..6dc0c90 100644
> --- a/src/compiler/nir/nir_opcodes.py
> +++ b/src/compiler/nir/nir_opcodes.py
> @@ -153,13 +153,13 @@ unop("fnot", tfloat, "(src0 == 0.0f) ? 1.0f : 0.0f")
>  unop("fsign", tfloat, "(src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f)")
>  unop("isign", tint, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)")
>  unop("iabs", tint, "(src0 < 0) ? -src0 : src0")
> -unop("fabs", tfloat, "fabsf(src0)")
> +unop("fabs", tfloat, "bit_size == 64 ? fabs(src0) : fabsf(src0)")

Seems like we should have had this anyway, since the extension
requires abs(double)?

As idr said in that thread, fnot, fsign, and fsat should have explicit
double support as well.

>  unop("fsat", tfloat, "(src0 > 1.0f) ? 1.0f : ((src0 <= 0.0f) ? 0.0f : src0)")
>  unop("frcp", tfloat, "1.0f / src0")
> -unop("frsq", tfloat, "1.0f / sqrtf(src0)")
> -unop("fsqrt", tfloat, "sqrtf(src0)")
> -unop("fexp2", tfloat, "exp2f(src0)")
> -unop("flog2", tfloat, "log2f(src0)")
> +unop("frsq", tfloat, "bit_size == 64 ? 1.0 / sqrt(src0) : 1.0f / sqrtf(src0)")
> +unop("fsqrt", tfloat, "bit_size == 64 ? sqrt(src0) : sqrtf(src0)")

And these?

> +unop("fexp2", tfloat, "bit_size == 64 ? exp2(src0) : exp2f(src0)")
> +unop("flog2", tfloat, "bit_size == 64 ? log2(src0) : log2f(src0)")

I don't know why we'd add these. I think this is just confusing the
issue. I'd separate functions that are actually required by the spec
into a separate patch at least.