[Mesa-dev] [PATCH 28/53] st/nine: Match REP implementation to LOOP

[Ilia Mirkin]

On Wed, Jan 7, 2015 at 5:33 PM, Axel Davy <axel.davy at ens.fr> wrote:
> Le 07/01/2015 21:13, Ilia Mirkin a écrit :
>
>> On Wed, Jan 7, 2015 at 11:36 AM, Axel Davy <axel.davy at ens.fr> wrote:
>>>
>>> Previous implementation was fine,
>>> just instead of having increasing counter,
>>> have a decreasing counter.
>>>
>>> Signed-off-by: Axel Davy <axel.davy at ens.fr>
>>> ---
>>>   src/gallium/state_trackers/nine/nine_shader.c | 41
>>> +++++++++++++++------------
>>>   1 file changed, 23 insertions(+), 18 deletions(-)
>>>
>>> diff --git a/src/gallium/state_trackers/nine/nine_shader.c
>>> b/src/gallium/state_trackers/nine/nine_shader.c
>>> index 21b06ce..88d4c07 100644
>>> --- a/src/gallium/state_trackers/nine/nine_shader.c
>>> +++ b/src/gallium/state_trackers/nine/nine_shader.c
>>> @@ -1562,9 +1562,7 @@ DECL_SPECIAL(REP)
>>>       unsigned *label;
>>>       struct ureg_src rep = tx_src_param(tx, &tx->insn.src[0]);
>>>       struct ureg_dst ctr;
>>> -    struct ureg_dst tmp = tx_scratch_scalar(tx);
>>> -    struct ureg_src imm =
>>> -        tx->native_integers ? ureg_imm1u(ureg, 0) : ureg_imm1f(ureg,
>>> 0.0f);
>>> +    struct ureg_dst tmp;
>>>
>>>       label = tx_bgnloop(tx);
>>>       ctr = tx_get_loopctr(tx, FALSE);
>>> @@ -1572,33 +1570,40 @@ DECL_SPECIAL(REP)
>>>       /* NOTE: rep must be constant, so we don't have to save the count
>>> */
>>>       assert(rep.File == TGSI_FILE_CONSTANT || rep.File ==
>>> TGSI_FILE_IMMEDIATE);
>>>
>>> -    ureg_MOV(ureg, ctr, imm);
>>> +    ureg_MOV(ureg, ureg_writemask(ctr, NINED3DSP_WRITEMASK_0), rep);
>>> +    /* in the case ctr is float, remove 0.5 to avoid precision issues
>>> for comparisons */
>>> +    if (!tx->native_integers)
>>> +        ureg_ADD(ureg, ureg_writemask(ctr, NINED3DSP_WRITEMASK_0),
>>> ureg_src(ctr), ureg_imm1f(ureg, -0.5f));
>>> +
>>>       ureg_BGNLOOP(ureg, label);
>>> -    if (tx->native_integers)
>>> -    {
>>> -        ureg_USGE(ureg, tmp, tx_src_scalar(ctr), rep);
>>> -        ureg_UIF(ureg, tx_src_scalar(tmp), tx_cond(tx));
>>> -    }
>>> -    else
>>> -    {
>>> -        ureg_SGE(ureg, tmp, tx_src_scalar(ctr), rep);
>>> +    tmp = tx_scratch_scalar(tx);
>>> +
>>> +    /* stop when crt.x <= 0 */
>>
>> ctr.x
>>
>>> +    if (!tx->native_integers) {
>>> +        ureg_SLE(ureg, tmp, ureg_scalar(ureg_src(ctr), TGSI_SWIZZLE_X),
>>> ureg_imm1f(ureg, 0.0f));
>>
>> It would be less confusing if this were the same as below, i.e. switch
>> it to SGT. Or switch the other one to ISLT. Having one be greater and
>> one less-than is weird.
>>
>>>           ureg_IF(ureg, tx_src_scalar(tmp), tx_cond(tx));
>>> +    } else {
>>> +        ureg_ISGE(ureg, tmp, ureg_imm1i(ureg, 0),
>>> ureg_scalar(ureg_src(ctr), TGSI_SWIZZLE_X));
>>> +        ureg_UIF(ureg, tx_src_scalar(tmp), tx_cond(tx));
>>>       }
>>>       ureg_BRK(ureg);
>>>       tx_endcond(tx);
>>>       ureg_ENDIF(ureg);
>>>
>>> -    if (tx->native_integers) {
>>> -        ureg_UADD(ureg, ctr, tx_src_scalar(ctr), ureg_imm1u(ureg, 1));
>>> -    } else {
>>> -        ureg_ADD(ureg, ctr, tx_src_scalar(ctr), ureg_imm1f(ureg, 1.0f));
>>> -    }
>>> -
>>>       return D3D_OK;
>>>   }
>>>
>>>   DECL_SPECIAL(ENDREP)
>>>   {
>>> +    struct ureg_program *ureg = tx->ureg;
>>> +    struct ureg_dst ctr = tx_get_loopctr(tx, FALSE);
>>> +
>>> +    if (!tx->native_integers) {
>>> +        ureg_ADD(ureg, ureg_writemask(ctr, NINED3DSP_WRITEMASK_0),
>>> ureg_src(ctr), ureg_imm1f(ureg, -1.0f));
>>> +    } else {
>>> +        ureg_UADD(ureg, ureg_writemask(ctr, NINED3DSP_WRITEMASK_0),
>>> ureg_src(ctr), ureg_imm1i(ureg, -1.0));
>>
>> -1 -- presumably this is an integer...
>
> Yep, should be -1. Strange compiler didn't warn.
>>
>>
>> But at what point does ctr get converted to an integer? At the
>> beginning, when you subtract 0.5, it appears to still be a float... I
>> think that ADD is wrong and needs to be a UADD? (But then obviously
>> -0.5 doesn't make sense).
>
> ctr is either an integer from the start, or not.
>
> When the card doesn't support integers, ctr is a float. To avoid precision
> issues when doing comparison to 0, I choosed to substract 0.5 to ctr.
>
> I realise now it makes more sense to compare ctr to 0.5 and not substract
> anything at the start.

Oh oops, I misread that add as happening if native ints, not if
!native ints. Having both paths being SGE (or SLT) would definitely
make for easier reading.

  -ilia