OpenCL

PoCL

full decoder kernel endless loop with new loop vectorizer. Minimal reproducer:

__kernel void decode(uchar flags)
{
  const ushort N = 1;

  for (ushort i = 0; i < N; i++) {
    if (i != N - 1 || flags) {
      work_group_reduce_add(0);
    }
  }
}

PoCL devs already have a patch set that also fixes this issue.

New issue: Crash when compiling stream compaction kernel to machine code in clusterizer. (TODO: reproducer)

Original issues in ZS decoder still persist even with new compiler (Maybe caused by O2 after all?)

New GPU Servers

Wanted to buy new GPU machines for a) development and b) CI Tests that run GPU code.

Both machines should have AMD + Nvidia GPU.

Ideally, just buy two new server -> not possible due to budget constraints.

Dev machine -> Buy parts and build server scratch

CI machine -> Add GPUs to existing machine

Two candidates for CI: alibi and alinsure

alibi

already has GPUs, but is occupied by long MC jobs. -> Unsuited to run CI jobs in parallel.

alinsure

Ancient machine (mainboard + CPU from 2013), unclear if BIOS even supports GPUs with >4GB VRAM.
Haven't checked if PSU is strong enough yet (base model only has 460W)

How far can we get if we build both machines from scratch instead?

Peter mentioned remaining budget of 8000CHF + 4000CHF over budget should be ok -> Assume 12000 CHF

CI Machine

Can have smaller CPU and GPU, only needs to compile GPU standalone:

Minimal configuration:

Threadripper 7000 (24 cores)
128GB RAM
RTX 5060 TI (16GB)
2TB NVMe

Total price inc basic components (Mainboard, case, etc): ~5000 CHF

GPU Machine

Max configuration possible with remaining bugdet:

Threadripper 9000 (32 cores)
128GB RAM
RTX 5080 (16GB)
4 TB NVMe
2x 20TB HDD (for RAID 0)

Total price inc basic components: ~7000 CHF

Updgrades possible if we don't need to buy a CI Machine:

CPU: Threadripper 7000 (64 cores) (maybe 9000 series)
Increase RAM to 256 GB