[v2 PATCH vkd3d 2/4] vkd3d-shader/hlsl: Cast round() input to float.

[Matteo Bruni]

On Thu, Jan 20, 2022 at 6:51 PM Zebediah Figura (she/her)
<zfigura at codeweavers.com> wrote:
>
> On 1/20/22 06:12, Matteo Bruni wrote:
> > On Thu, Jan 20, 2022 at 12:42 AM Zebediah Figura (she/her)
> > <zfigura at codeweavers.com> wrote:
> >>
> >> On 1/19/22 06:03, Matteo Bruni wrote:
> >>> On Mon, Dec 20, 2021 at 10:44 PM Zebediah Figura (she/her)
> >>> <zfigura at codeweavers.com> wrote:
> >>>>
> >>>> On 12/20/21 12:15, Zebediah Figura (she/her) wrote:
> >>>>> On 12/19/21 06:41, Nikolay Sivov wrote:
> >>>>>> +static struct hlsl_ir_node *intrinsic_float_convert_arg(struct
> >>>>>> hlsl_ctx *ctx,
> >>>>>> +        const struct parse_initializer *params, struct hlsl_ir_node
> >>>>>> *arg, const struct vkd3d_shader_location *loc)
> >>>>>> +{
> >>>>>> +    struct hlsl_type *type = arg->data_type;
> >>>>>> +
> >>>>>> +    if (type->base_type == HLSL_TYPE_FLOAT || type->base_type ==
> >>>>>> HLSL_TYPE_HALF)
> >>>>>> +        return arg;
> >>>>>
> >>>>> add_implicit_conversion() already does this.
> >>>>
> >>>> Sorry, turns out I'm wrong; this preserves the input type for half,
> >>>> whereas add_implicit_conversion() alone would not.
> >>>
> >>> I wonder if that's what it should do in the first place?
> >>
> >> Yes, witness the following shader:
> >>
> >> ---
> >>
> >> uint func(float x) { return 1; }
> >> uint func(half x) { return 2; }
> >>
> >> float4 main(void) : sv_target
> >> {
> >>       return float4(func(float(0)), func(half(0)), func(int(0)),
> >> func(uint(0)));
> >> }
> >>
> >> ---
> >>
> >> Native reduces this to (1, 2, 1, 1).
> >
> > Fun... But at least it means we can fix add_implicit_conversion() and
> > avoid that kind of workarounds, if I understand correctly.
>
> I... guess, but it seems janky. The main point of
> add_implicit_conversion() is for stuff like assignment RHS or return
> values. If we had add_implicit_conversion() pass through half it'd
> generate IR like
>
> 2:  half | 1.0
> 3:       | x = @2
> 4: float | x
>
>
> (i.e. "x" is a float variable). Which... works, I guess? Maybe it's not
> that bad, since we already do a somewhat similar thing for scalar vs vec1.

I think that's an improvement to the current state of things, right
now we're effectively losing information from the get-go. We can
always add a half-to-float lowering pass to simplify the IR.