[PATCH vkd3d 5/7] vkd3d-shader/hlsl: Write SM4 casts to float.

[Zebediah Figura (she/her)]

On 9/9/21 6:41 PM, Matteo Bruni wrote:
> On Fri, Sep 10, 2021 at 1:32 AM Zebediah Figura (she/her)
> <zfigura at codeweavers.com> wrote:
>>
>> On 9/9/21 6:16 PM, Matteo Bruni wrote:
>>> On Thu, Sep 9, 2021 at 6:02 AM Zebediah Figura <zfigura at codeweavers.com> wrote:
>>>>
>>>> Signed-off-by: Zebediah Figura <zfigura at codeweavers.com>
>>>> ---
>>>>    libs/vkd3d-shader/hlsl_sm4.c | 34 ++++++++++++++++++++++++++++++++++
>>>>    1 file changed, 34 insertions(+)
>>>>
>>>> diff --git a/libs/vkd3d-shader/hlsl_sm4.c b/libs/vkd3d-shader/hlsl_sm4.c
>>>> index 160da47fd..cac71ef7e 100644
>>>> --- a/libs/vkd3d-shader/hlsl_sm4.c
>>>> +++ b/libs/vkd3d-shader/hlsl_sm4.c
>>>> @@ -981,6 +981,40 @@ static void write_sm4_expr(struct hlsl_ctx *ctx,
>>>>            {
>>>>                switch (expr->op)
>>>>                {
>>>> +                case HLSL_OP1_CAST:
>>>> +                {
>>>> +                    const struct hlsl_type *src_type = arg1->data_type;
>>>> +
>>>> +                    /* Narrowing casts need to be lowered. */
>>>> +                    if (src_type->dimx != expr->node.data_type->dimx)
>>>> +                        hlsl_fixme(ctx, expr->node.loc, "Narrowing cast.\n");
>>>> +
>>>> +                    switch (src_type->base_type)
>>>> +                    {
>>>> +                        case HLSL_TYPE_HALF:
>>>> +                        case HLSL_TYPE_FLOAT:
>>>> +                            write_sm4_unary_op(buffer, VKD3D_SM4_OP_MOV, &expr->node, arg1, 0);
>>>> +                            break;
>>>> +
>>>> +                        case HLSL_TYPE_BOOL:
>>>> +                        case HLSL_TYPE_INT:
>>>> +                            write_sm4_unary_op(buffer, VKD3D_SM4_OP_ITOF, &expr->node, arg1, 0);
>>>> +                            break;
>>>
>>> As far as I'm aware, booleans in SM4+ are represented in quite a
>>> peculiar way: false is represented as 0, true is canonically
>>> 0xffffffff. I think the idea is to be able to use the result of a
>>> conditional expression directly as a mask and such, but don't quote me
>>> on that. For sure it's different from SM1-3, which use the traditional
>>> 0 and 1 values for false and true.
>>>
>>> That's probably the reason why native uses MOVC to convert from bool
>>> (to float or anything else) and, most importantly, NE for the opposite
>>> direction (with 0 as the other comparand).
>>>
>>
>> "(int)(bool)true" evaluates to 1 in anything I've been able to test.
>> Under what circumstances do you have it evaluating to ~0?
> 
> Yeah, the bool -> int conversion explicitly makes sure to give you 1
> (by means of the MOVC instruction). You need to look at the generated
> shader bytecode.
> Try with this shader:
> 
> bool4 main(uniform int i, uniform uint u, uniform bool b, uniform
> float f) : sv_target
> {
>      return bool4(((float)(bool)i) + 1.5, ((float)(bool)u) - 2.5,
> ((float)(bool)b) / 2, f);
> }
> 

Ah, sure, I guess it would only show up for RTV or UAV writes.

I'll leave the bool case out of this patch, then.