[Mesa-dev] [PATCH] r600g: fix and optimize tgsi_cmp when using ABS and NEG modifier

[Patrick Rudolph]

Are there optimizations done on TGSI ?
I can't find any file in src/gallium/auxiliary/tgsi that does so.

Kind Regards,
Patrick

On 2016-03-30 06:59 PM, Ian Romanick wrote:
> I'd recommend doing this at a higher level.  We've been doing this in
> NIR (in nir_opt_algebraic.py) for quite some time, and we've found that
> having x==0 instead of -abs(x)>=0 in higher-level IR helps other
> optimizations make progress.  Maybe do this when translating out of TGSI?
> 
> On 03/30/2016 03:30 AM, Patrick Rudolph wrote:
>> It affects the game "RIFT" and fixes slightly wrong colors on
>> the player model, when using the nine state-tracker.
>>
>> Kind Regards,
>> Patrick
>>
>> On 2016-03-29 11:09 PM, Dave Airlie wrote:
>>> On 28 March 2016 at 19:52,  <siro at das-labor.org> wrote:
>>>> From: Patrick Rudolph <siro at das-labor.org>
>>>>
>>>> Some apps set NEG and ABS on the source param to test for zero.
>>>> Use ALU_OP3_CNDE insted of ALU_OP3_CNDGE and unset both modifiers.
>>>>
>>>> It also removes the need for a MOV instruction, as ABS isn't
>>>> supported on op3.
>>>>
>>>> Tested on AMD CAYMAN and AMD RV770.
>>>>
>>>> Signed-off-by: Patrick Rudolph <siro at das-labor.org>
>>>
>>> Seems sane to me, does this help in any app in particular?
>>>
>>> Reviewed-by: Dave Airlie <airlied at redhat.com>
>>>
>>> I just have to remember to push this tomorrow.
>>>
>>> Dave.
>>>
>>>> ---
>>>>  src/gallium/drivers/r600/r600_shader.c | 11 ++++++++++-
>>>>  1 file changed, 10 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/src/gallium/drivers/r600/r600_shader.c b/src/gallium/drivers/r600/r600_shader.c
>>>> index 77658f5..e5e2e3f 100644
>>>> --- a/src/gallium/drivers/r600/r600_shader.c
>>>> +++ b/src/gallium/drivers/r600/r600_shader.c
>>>> @@ -7707,6 +7707,15 @@ static int tgsi_cmp(struct r600_shader_ctx *ctx)
>>>>         int i, r, j;
>>>>         int lasti = tgsi_last_instruction(inst->Dst[0].Register.WriteMask);
>>>>         int temp_regs[3];
>>>> +       unsigned op;
>>>> +
>>>> +       if (ctx->src[0].abs && ctx->src[0].neg) {
>>>> +               op = ALU_OP3_CNDE;
>>>> +               ctx->src[0].abs = 0;
>>>> +               ctx->src[0].neg = 0;
>>>> +       } else {
>>>> +               op = ALU_OP3_CNDGE;
>>>> +       }
>>>>
>>>>         for (j = 0; j < inst->Instruction.NumSrcRegs; j++) {
>>>>                 temp_regs[j] = 0;
>>>> @@ -7719,7 +7728,7 @@ static int tgsi_cmp(struct r600_shader_ctx *ctx)
>>>>                         continue;
>>>>
>>>>                 memset(&alu, 0, sizeof(struct r600_bytecode_alu));
>>>> -               alu.op = ALU_OP3_CNDGE;
>>>> +               alu.op = op;
>>>>                 r = tgsi_make_src_for_op3(ctx, temp_regs[0], i, &alu.src[0], &ctx->src[0]);
>>>>                 if (r)
>>>>                         return r;
>>>> --
>>>> 2.5.0
>>>>
>>>> _______________________________________________
>>>> mesa-dev mailing list
>>>> mesa-dev at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>>