[Mesa-dev] [PATCH 06/10] i965: Reduce maximum GL_ARB_fragment_program instruction count to 1024.

[Kenneth Graunke]

On 09/24/2012 01:04 PM, Roland Scheidegger wrote:
> Am 24.09.2012 21:06, schrieb Kenneth Graunke:
>> On 09/22/2012 02:04 PM, Eric Anholt wrote:
>>> I don't know of any programs that would need more than this.  The larger
>>> programs I've seen have neared 100 instructions.  This prevent excessive
>>> runtimes of automatic tests that attempt to test up to the exposed maximums
>>> (like fp-long-alu).
>>> ---
>>>  src/mesa/drivers/dri/i965/brw_context.c |    8 ++++----
>>>  1 file changed, 4 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/src/mesa/drivers/dri/i965/brw_context.c b/src/mesa/drivers/dri/i965/brw_context.c
>>> index 1083e28..a7d61bd 100644
>>> --- a/src/mesa/drivers/dri/i965/brw_context.c
>>> +++ b/src/mesa/drivers/dri/i965/brw_context.c
>>> @@ -250,10 +250,10 @@ brwCreateContext(int api,
>>>        MIN2(ctx->Const.VertexProgram.MaxNativeParameters,
>>>  	   ctx->Const.VertexProgram.MaxEnvParams);
>>>  
>>> -   ctx->Const.FragmentProgram.MaxNativeInstructions = (16 * 1024);
>>> -   ctx->Const.FragmentProgram.MaxNativeAluInstructions = (16 * 1024);
>>> -   ctx->Const.FragmentProgram.MaxNativeTexInstructions = (16 * 1024);
>>> -   ctx->Const.FragmentProgram.MaxNativeTexIndirections = (16 * 1024);
>>> +   ctx->Const.FragmentProgram.MaxNativeInstructions = (1 * 1024);
>>> +   ctx->Const.FragmentProgram.MaxNativeAluInstructions = (1 * 1024);
>>> +   ctx->Const.FragmentProgram.MaxNativeTexInstructions = (1 * 1024);
>>> +   ctx->Const.FragmentProgram.MaxNativeTexIndirections = (1 * 1024);
>>>     ctx->Const.FragmentProgram.MaxNativeAttribs = 12;
>>>     ctx->Const.FragmentProgram.MaxNativeTemps = 256;
>>>     ctx->Const.FragmentProgram.MaxNativeAddressRegs = 0;
>>
>> I've definitely seen 600 instruction shaders generated by Cg.  (Some
>> Crystal Space based games use shaders written in Cg, which translates to
>> stellar ARBfp code...)  I'm afraid there may be even bigger ones.
> 
> I'm not sure there are really larger ones out in the wild.
> I would guess that the arbfp1 target was mostly used for targetting
> ati cards (since geforce cards had their own target profile), and these were
> limited to 512 (hw) instructions (for r4xx, much less for r3xx).
> I guess the stellar arbfp code is mostly due to the necessary loop unrolling,
> and if you were targetting cards which actually could do loops you'd be
> using the glsl target of cg (of course you could use glsl target for those
> older cards too but arbfp1 target is probably older?).
> And I bet you could cut down the MaxNativeTexInstructions and especially
> the number of MaxNativeTexIndirections a lot more without hurting any
> real apps.
> But of course I don't actually know for sure if there isn't some stupid
> app out there trying to compile some nice small simple cg shader but with loops
> into some arb_fp mess even if it should really use glsl target.
> 
> Roland

For some reason using Cg's glsl target always crashed and burned for me.
 But it's been a while since I tried and I never did figure out whose
bug it was.

The game I tested also recently switched to using GLSL directly, so it
may be a moot point.

I suppose we could always lower it and then re-raise it if something
broke.  It is really hard to imagine >= 1024 instruction ARB fp programs.