07:01 <tomeu> austriancoder: btw, just stumbled upon this in galcore:

07:01 <tomeu> gctUINT32 threadAllocation = gcmALIGN_NP2(groupsizeX * groupsizeY, db->NumShaderCores * 4)/db->NumShaderCores * 4;

07:01 <tomeu> which matches my observation

07:01 <tomeu> my current problem is that anything above 0x4 will hang the GPU

07:02 <tomeu> but that doesn't happen to galcore with the same cmdstream

07:02 <tomeu> so I'm seeing if I need to initialize the HW differently in order to "activate" those shaders

07:08 <austriancoder> tomeu: -ENOTIME .. but I have seen something like this on GC2000: groupsizeX * groupsizeY * groupsizeZ / (NumShaderCores * 4)

07:08 <tomeu> yeah, unfortauntely that was already known :(

07:09 <tomeu> but we even have some bits about images in _ProgramPPUCommand

07:09 <tomeu> but just confirm things we already knew

07:10 <austriancoder> I think I can spend more time to help out in 2-3 weeks

07:10 <austriancoder> maybe you get it working by then :)

07:10 <tomeu> don't worry, no rush

07:10 <tomeu> I surely hope!

07:10 <tomeu> or at least work around it

07:11 <tomeu> I'm really close to run the whole pipeline in tensorflow lite, I think this is the last operation

07:11 <austriancoder> nice

07:38 <tomeu> yay! finally got to run inference correctly with etnaviv :)

07:38 <tomeu> worked around that cores limitation by reducing the number of workgroups on the tflite side

11:16 <tomeu> does anybody know what is this flop reset think in galcore?

12:13 <lynxeye> tomeu: My _guess_ is that it's kind of a internal state reset, so you can't leak information from one context to another.

12:13 <tomeu> hmm, that sounds pausible

12:15 <tomeu> lynxeye: do you have any idea of what the galcore ko may be doing so that when userspace submits THREAD_ALLOCATION values higher than 0x4, the HW doesn't hang?

12:15 <tomeu> for identical cmdstream, ioctl dumps and shaders, I get a hang with etnaviv but not with galcore

12:17 <lynxeye> Nope, sorry. Haven't really looked at this part of the HW/blob.

12:21 <tomeu> np

13:50 <tomeu> lynxeye: austriancoder : btw, don't know if this was already known, but I have just stumbled upon a new source of info on the ISA in galcore: https://github.com/ArcturusNetworks/uCLS1012A-kernel/blob/fb781a15024e347bfedb40ec2233c031ce6c9f3c/drivers/mxc/gpu-viv/hal/kernel/arch/gc_hal_kernel_hardware_func_ppu.h#L1095

13:51 <tomeu> they are showing even how the dot product EVIS instruction is encoded!

13:53 <austriancoder> tomeu: that's new to me.. nice