Skip to content

Releases

v1.3.0

Choose a tag to compare

@mxreyer mxreyer released this 05 Nov 19:13
53d679c

Adds pt_ocl-fixes-only.so (./build.sh fixes): pytorch_ocl with only the build and correctness fixes (dlprimitives 00–02, pytorch_dlprim 01). It’s the baseline every speed-up in these repos is measured against, so those comparisons can be rerun. pt_ocl.so is unchanged from v1.2.1.

sha256:
pt_ocl.so 283410d408e43ade29864b5cd3bb59024d1f78d9367489ce3a8b8c03437d89ad
pt_ocl-fixes-only.so fffe0d76cd4d108865799d727367b14962e7a87099cbff435429d4b176c1f4cf

v1.2.1

Choose a tag to compare

@mxreyer mxreyer released this 05 Nov 05:42
50739b8

No change in behaviour from v1.2.0. The pt_ocl.so hash differs because a
comment changed in a kernel source, and kernel sources are compiled in.

Figures: the README, patch descriptions and upstream reports now quote a
full rerun on Linux 7.2.7 and Mesa 26.2.3. ResNet-9 training at batch 128:

build img/s
fixes only (00–02) 916
fp32, all patches 2,436
fp16 (DLPRIM_CONV_FP16=1) 3,147
T4 on Colab 1,930

Some earlier claims are corrected. fp16 lands within 0.9% of fp32, typically
about 0.5%, and is 1.29x shipped fp32 rather than 1.45x.

Tools

  • profile-step.py: PROFILE_MODE=infer times inference, and
    PROFILE_DEVICE selects the device, so the T4 can run the same step.
  • wino-repro.py: an optional second argument sets the relative tolerance.

Reports: the fp16 inner loop now has its own upstream report.

sha256: 283410d408e43ade29864b5cd3bb59024d1f78d9367489ce3a8b8c03437d89ad

v1.2.0

Choose a tag to compare

@mxreyer mxreyer released this 21 Sep 19:28
Insert 01-reduce-barrier ahead of the performance series

The CUSTOM_REDUCE fallback in reduce.h has no barrier after every work-item
reads my_reduce[0], so a kernel that reduces twice in a row (softmax: max,
then sum) lets work-item 0 overwrite the first result before slower
work-items have read it. The dlprimitives softmax and log_softmax tests
fail on the large-channel shapes without the fix.

It goes in at 01, next to the build fix that exposes it, so 00-01 are
"build and correct" and 02-11 the performance series; the old 01-10 move
up by one and every reference in the README, HANDOFF, OPENCL-PERF and the
upstream reports follows. New report: upstream/dlprimitives-reduce-barrier.md.

Also folds in the 2026-09-21 HANDOFF notes on the Mesa 26.2.2 re-check and
on pinning the clock for short measurements.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

v1.1.0

Choose a tag to compare

@mxreyer mxreyer released this 19 Sep 19:18

Added a patch, re-ordered patches.

v1.0.0

Choose a tag to compare

@mxreyer mxreyer released this 14 Sep 17:26
build.sh: make pt_ocl.so reproducible

-ffile-prefix-map maps scratch/ (source and the venv's torch headers, which
__FILE__ embedded ~200 times) to /scratch, CMAKE_SKIP_RPATH drops the RPATH
naming this venv's torch/lib (never needed: import torch has already loaded
those sonames), and pybind11 is pinned. Two consecutive builds now give the
same sha256, so a release asset can be verified by rebuilding anywhere.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W9wRcEhBBBQPFzPWvCrgf1