mirrors/tinygrad
2
0
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-06-24 02:14:17 +00:00
Code Issues Projects Releases Packages Wiki Activity

IMAGE hand_coded_optimizations update

Browse source
This commit is contained in:
Chen-Yu Yang 2026-06-23 21:34:51 -04:00
commit adeeb33dc4
2 changed files with 7 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
.github/workflows/test.yml vendored
View file

@ -327,7 +327,7 @@ jobs:
llvm: 'true'
- name: Test openpilot model kernel count and gate usage
run: |
ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1468 ALLOWED_GATED_READ_IMAGE=4 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=55 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
- name: Test openpilot CL compile fp32 (test correctness)
run: |
DEV=CL IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx

6
tinygrad/codegen/opt/heuristic.py
View file

@ -101,6 +101,12 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
# for Schedule, we check if the range is used in INDEX gates or WHERE gates
is_masked = k.rngs[axis] in where_gate_rngs
if k.full_shape[axis] <= 7 and is_masked and prod(k.full_shape[j] for j in to_upcast) * k.full_shape[axis] <= 7 * 7:
# upcasting a masked global axis moves that range out of the launch grid into each work-item
# under IMAGE, skip the upcast unless enough global work-items remain after it to hide memory latency
if IMAGE and k.axis_types[axis] is AxisType.GLOBAL:
global_upcast = prod(k.full_shape[i] for i in to_upcast if k.axis_types[i] is AxisType.GLOBAL) * k.full_shape[axis]
global_items_after = prod(k.full_shape[i] for i in k.axes_of(AxisType.GLOBAL)) // global_upcast
if resolve(global_items_after < getenv("OCCUPANCY_FLOOR", 4096), False): continue
if DEBUG >= 4: print(f"upcasting masked axis : {axis}")
to_upcast.append(axis)
for axis in to_upcast[::-1]: k.apply_opt(Opt(OptOps.UPCAST, axis, 0))