[Mesa-dev] [PATCH 3/3] i965/fs: emit DIM instruction to load 64-bit immediates in HSW

[Matt Turner]

On Wed, Jul 13, 2016 at 10:52 PM, Samuel Iglesias Gonsálvez
<siglesias at igalia.com> wrote:
>
>
> On 14/07/16 03:46, Matt Turner wrote:
>> On Wed, Jul 13, 2016 at 5:06 PM, Matt Turner <mattst88 at gmail.com> wrote:
>>> On Tue, Jul 12, 2016 at 11:42 PM, Samuel Iglesias Gonsálvez
>>> <siglesias at igalia.com> wrote:
>>>> Signed-off-by: Samuel Iglesias Gonsálvez <siglesias at igalia.com>
>>>> ---
>>>>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 12 ++++++++++++
>>>>  1 file changed, 12 insertions(+)
>>>>
>>>> diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
>>>> index a65c273..bf32dfd 100644
>>>> --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
>>>> +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
>>>> @@ -4558,6 +4558,18 @@ setup_imm_df(const fs_builder &bld, double v)
>>>>     if (devinfo->gen >= 8)
>>>>        return brw_imm_df(v);
>>>>
>>>> +   /* gen7.5 does not support DF immediates straighforward but the DIM
>>>> +    * instruction allows to set the 64-bit immediate value.
>>>> +    */
>>>> +   if (devinfo->is_haswell) {
>>>> +      const fs_builder ubld = bld.exec_all();
>>>> +      fs_reg dst = ubld.vgrf(BRW_REGISTER_TYPE_DF, 1);
>>>> +      struct brw_reg imm = brw_imm_reg(BRW_REGISTER_TYPE_F);
>>>> +      imm.df = v;
>>>> +      ubld.DIM(dst, imm);
>>>
>>> I know the hardware is strange and requires that src0's type is F, but
>>> I don't think we need to model that in the IR. I think that using a DF
>>> type in the IR
>>> would require otherwise unnecessary changes to dump_instructions().
>>>
>>> With the above three lines changed to just
>>>
>>>    ubld.DIM(dst, brw_imm_df(v));
>>>
>>> this patch is:
>>>
>>> Reviewed-by: Matt Turner <mattst88 at gmail.com>
>>>
>>> Patch 1 I sent comments on. With those addressed it is also
>>>
>>> Reviewed-by: Matt Turner <mattst88 at gmail.com>
>>>
>>> I believe with my comments addressed on 1/3 and 3/3 that 2/3 is unecessary.
>>
>> Actually, I guess 2/3 is necessary since the type is changed before
>> brw_eu_emit? Sorry, not thinking very clearly.
>>
>
> Exactly, patch 2/3 is needed because the type has been changed before
> brw_eu_emit.
>
> Does patch 2/3 get your R-b then?

Yes. Thank you!