textureGather performance

castano · May 13, 2024, 11:49pm

It appears that textureGather is natively supported on PowerVR GPUs and that each textureGather invocation is translated into a smp2d.fcnorm.replace.data instruction. However, for every one of these instructions the compiler also produces many additional branch and alu instructions.

Does anyone know what these instructions are doing? I’ve put together this gist comparing the results of using textureGather vs textureFetch:

gist.github.com

https://gist.github.com/castano/705eb6d9717003db2a6f0d9164e67b87

PVR GE8320 textureGather performance.md


I'm investigating the performance difference between textureGather and textureFetch on PVR GE8320.

The instruction set reference does not provide much information about the available functionality:

https://docs.imgtec.com/reference-manuals/powervr-instruction-set-reference/html/topics/data-access-instructions/ITRSMP.html

But compiling some simple shaders that load a 4x4 texture block results in 4 textureGather instructions vs 16 textureFetch instructions, which seems to indicate that the instruction set supports this functionality. In particular the former uses 4 `smp2d.fcnorm.replace.data` while the latter issues 16 `smp2d.fcnorm.nncoords.replace`. 

However, in the textureGather code path the compiler emits many additional branch and ALU instructions, which seems to indicate that the inputs or outputs of the `smp2d` instruction need to be processed in some way to produce the desired result.

This file has been truncated. show original

spark_test_sol64.textureFetch.GE8320.txt


Coefficient update program:

---------------------------------------------

 Address: 0x0000000000[         0] 0: 
mov ft0, ft1, c0, c0
       mov ft2, cf0
       cbs ft3, cf0
       or ft4, _, ft2, _, c0

This file has been truncated. show original

spark_test_sol64.textureGather.GE8320.txt


Coefficient update program:

---------------------------------------------

 Address: 0x0000000000[         0] 0: 
mov ft0, ft1, c0, c0
       mov ft2, cf0
       cbs ft3, cf0
       or ft4, _, ft2, _, c0

This file has been truncated. show original

AlejandroC · May 16, 2024, 1:23pm

Hi castano,

Thanks for your message.

I will need to look into it and possibly ask some internal teams to try shed light on the reasons behind this difference in the instruction count when using textureGather. I will come back with any findings.

Best regards,
Alejandro

castano · May 17, 2024, 7:10am

Thanks Alejandro! Most of my clients would like to avoid having to use vendor-specific shader variants, which puts me in the uncomfortable position of choosing what vendor to favor. I could use uniform branching and pick the best path dynamically, but if we could make the textureGather code path work well in all cases that would be much better!

AlejandroC · May 21, 2024, 2:00pm

Hi castano,

I replied to your questions through an internal message.

Best regards,
Alejandro

Topic		Replies	Views
Version 320 es. Not supported PowerVR Insider	4	1677	October 16, 2018
Bug in glGenerateMipmap() for SGX 540 ? PowerVR Insider pvrtrace	10	412	May 10, 2013
Can I use the GCMP instruction on a GLSL shader? PowerVR Insider	1	645	January 28, 2019
developing on gentoo - shader compile or link fail PowerVR Insider	5	330	May 28, 2012
glGenerateMipmap fails silently... PowerVR Insider	2	350	March 1, 2011

textureGather performance

Related topics