optimizer rework - #25

Closed

Human9000-bit wants to merge 25 commits into

mainfrom

onnx-optimizer-ir

Commits on Aug 7, 2026

onnx: numerical-equivalence harness for graph transformations
Human9000-bit
committed
ce5c760
onnx: def-use IR and pass driver for the optimizer
Human9000-bit
committed
2faeb6f
onnx: port activation fusion to the IR, keep the Conv_Relu op
Human9000-bit
committed
12f5dd1
onnx: port node reordering to the IR
Human9000-bit
committed
c71bbab
onnx: port Conv+BatchNormalization folding to the IR
Human9000-bit
committed
c050f0c
onnx: port Conv+Mul and Conv+Add const folding, fix the broadcast rule
Human9000-bit
committed
cbaac3b
onnx: fold constants without building a session per node
Human9000-bit
committed
7f6f42b
onnx: intern operator attribute names as an enum
Human9000-bit
committed
d926f2b
onnx: move the execution plan out of loader/ into plan/
Human9000-bit
committed
af520af
plan: split slot assignment and conv-param resolution out of the builder
Human9000-bit
committed
df908e0
onnx: read the hot-path tuning switches once, not per node
Human9000-bit
committed
9db0acd
plan: match DW+PW and PW+DW fusions by dataflow, not adjacency
Human9000-bit
committed
410e046
plan: match Conv+Add residual fusion by dataflow, not adjacency
Human9000-bit
committed
a62cede
plan: stop truncating the plan at the DW+PW back-off
Human9000-bit
committed
0977deb
plan: match the quantized chains by dataflow, not adjacency
Human9000-bit
committed
b5c0cca
plan: assert the execution plan holds one action per node
Human9000-bit
committed
7d39e13
plan: only fuse Conv+Add into the later producer of the residual
Human9000-bit
committed
855ccdc
plan: resolve the NCHWc handoff at load time
Human9000-bit
committed
39cf43b
plan: match the PW_expand+DW+PW_reduce merge by dataflow too
Human9000-bit
committed
4b6d602
plan: resolve the Conv kernel entry point at load time
Human9000-bit
committed
fa1456a
conv: dispatch on the resolved kernel, not on re-derived predicates
Human9000-bit
committed
ac2dc17
plan: stop the PW-reduce merge from dropping its residual Add
Human9000-bit
committed
d0257f9
plan: cover the four untested actions, fixing two bugs they exposed
Human9000-bit
committed
697c780
docs: correct the optimizer docs that describe the pre-IR world
Human9000-bit
committed
f041585
tests: one lock for the process environment, held for whole tests
Human9000-bit
committed
696f43f

Navigation Menu

optimizer rework - #25

optimizer rework#25
Human9000-bit wants to merge 25 commits into
mainfrom
onnx-optimizer-ir

Commits on Aug 7, 2026

onnx: numerical-equivalence harness for graph transformations

onnx: def-use IR and pass driver for the optimizer

onnx: port activation fusion to the IR, keep the Conv_Relu op

onnx: port node reordering to the IR

onnx: port Conv+BatchNormalization folding to the IR

onnx: port Conv+Mul and Conv+Add const folding, fix the broadcast rule

onnx: fold constants without building a session per node

onnx: intern operator attribute names as an enum

onnx: move the execution plan out of loader/ into plan/

plan: split slot assignment and conv-param resolution out of the builder

onnx: read the hot-path tuning switches once, not per node

plan: match DW+PW and PW+DW fusions by dataflow, not adjacency

plan: match Conv+Add residual fusion by dataflow, not adjacency

plan: stop truncating the plan at the DW+PW back-off

plan: match the quantized chains by dataflow, not adjacency

plan: assert the execution plan holds one action per node

plan: only fuse Conv+Add into the later producer of the residual

plan: resolve the NCHWc handoff at load time

plan: match the PW_expand+DW+PW_reduce merge by dataflow too

plan: resolve the Conv kernel entry point at load time

conv: dispatch on the resolved kernel, not on re-derived predicates

plan: stop the PW-reduce merge from dropping its residual Add

plan: cover the four untested actions, fixing two bugs they exposed

docs: correct the optimizer docs that describe the pre-IR world

tests: one lock for the process environment, held for whole tests