Add benchmarking script (#20188)#20188
Merged
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20188
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 9 Pending, 1 Unrelated FailureAs of commit ea24622 with merge base 4519036 ( NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@metascroy has exported this pull request. If you are a Meta employee, you can view the originating Diff in D108048181. |
This PR needs a
|
meta-codesync
Bot
changed the title
metascroy
added a commit
to metascroy/executorch
that referenced
this pull request
Jun 10, 2026
Summary:
Adds a standalone microbenchmark for the ImageProcessor reuse APIs and a companion script to diff two runs, so kernel/pipeline changes (e.g. the NEON deinterleave switch) can be measured reproducibly.
New directory xplat/executorch/extension/image/benchmark/:
* image_processor_benchmark.cpp (cxx_binary) — times process_into (BGRA/RGBA) and process_yuv_into (NV12/NV21) over a sweep of common input sizes × target sizes. Per cell it runs variants covering execution path (CPU / GPU / size-default), resize mode (stretch / letterbox), orientation (upright + 90°), cropped ROI, and the allocating process() vs process_into(). Each row reports mean/median/p95/stddev over 100 iters (10 warmup) on a synthetic gradient input; a row that fails is reported as ERROR rather than timed.
* Flags (all optional): --format=bgra|rgba|nv12|nv21, --unit=cpu|gpu|default (both default to all), --out=PATH (writes a clean results table; the input-size sweep and rotation always run). Output is grouped under === API-section banners with a column legend, and --- per-cell separators.
* compare_benchmarks.py (python_binary, stdlib-only) — matches rows by (API section, input→target cell, variant) and prints per-row base / new speedup plus a summary bucketed by execution path (CPU / GPU / default).
* README.md — usage, the build-mode caveat, and the capture→compare workflow.
* BUCK / TARGETS / targets.bzl — build defs.
Note: benchmark only with an optimized build (-c cxx.extra_cxxflags=-Os); the default buck2 run is -O0 and unrepresentative.
Differential Revision: D108048181
metascroy
force-pushed
the
export-D108048181
branch
from
257df08 to
9f42724
Compare
metascroy
added a commit
to metascroy/executorch
that referenced
this pull request
Jun 10, 2026
Summary:
Adds a standalone microbenchmark for the ImageProcessor reuse APIs and a companion script to diff two runs, so kernel/pipeline changes (e.g. the NEON deinterleave switch) can be measured reproducibly.
New directory xplat/executorch/extension/image/benchmark/:
* image_processor_benchmark.cpp (cxx_binary) — times process_into (BGRA/RGBA) and process_yuv_into (NV12/NV21) over a sweep of common input sizes × target sizes. Per cell it runs variants covering execution path (CPU / GPU / size-default), resize mode (stretch / letterbox), orientation (upright + 90°), cropped ROI, and the allocating process() vs process_into(). Each row reports mean/median/p95/stddev over 100 iters (10 warmup) on a synthetic gradient input; a row that fails is reported as ERROR rather than timed.
* Flags (all optional): --format=bgra|rgba|nv12|nv21, --unit=cpu|gpu|default (both default to all), --out=PATH (writes a clean results table; the input-size sweep and rotation always run). Output is grouped under === API-section banners with a column legend, and --- per-cell separators.
* compare_benchmarks.py (python_binary, stdlib-only) — matches rows by (API section, input→target cell, variant) and prints per-row base / new speedup plus a summary bucketed by execution path (CPU / GPU / default).
* README.md — usage, the build-mode caveat, and the capture→compare workflow.
* BUCK / TARGETS / targets.bzl — build defs.
Note: benchmark only with an optimized build (-c cxx.extra_cxxflags=-Os); the default buck2 run is -O0 and unrepresentative.
Differential Revision: D108048181
metascroy
force-pushed
the
export-D108048181
branch
from
9f42724 to
803f6d3
Compare
Summary:
Adds a standalone microbenchmark for the ImageProcessor reuse APIs and a companion script to diff two runs, so kernel/pipeline changes (e.g. the NEON deinterleave switch) can be measured reproducibly.
New directory xplat/executorch/extension/image/benchmark/:
* image_processor_benchmark.cpp (cxx_binary) — times process_into (BGRA/RGBA) and process_yuv_into (NV12/NV21) over a sweep of common input sizes × target sizes. Per cell it runs variants covering execution path (CPU / GPU / size-default), resize mode (stretch / letterbox), orientation (upright + 90°), cropped ROI, and the allocating process() vs process_into(). Each row reports mean/median/p95/stddev over 100 iters (10 warmup) on a synthetic gradient input; a row that fails is reported as ERROR rather than timed.
* Flags (all optional): --format=bgra|rgba|nv12|nv21, --unit=cpu|gpu|default (both default to all), --out=PATH (writes a clean results table; the input-size sweep and rotation always run). Output is grouped under === API-section banners with a column legend, and --- per-cell separators.
* compare_benchmarks.py (python_binary, stdlib-only) — matches rows by (API section, input→target cell, variant) and prints per-row base / new speedup plus a summary bucketed by execution path (CPU / GPU / default).
* README.md — usage, the build-mode caveat, and the capture→compare workflow.
* BUCK / TARGETS / targets.bzl — build defs.
Note: benchmark only with an optimized build (-c cxx.extra_cxxflags=-Os); the default buck2 run is -O0 and unrepresentative.
Differential Revision: D108048181
metascroy
force-pushed
the
export-D108048181
branch
2 times, most recently
from
803f6d3 to
ea24622
Compare
psiddh
approved these changes
Jun 10, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Adds a standalone microbenchmark for the ImageProcessor reuse APIs and a companion script to diff two runs, so kernel/pipeline changes (e.g. the NEON deinterleave switch) can be measured reproducibly.
New directory xplat/executorch/extension/image/benchmark/:
Note: benchmark only with an optimized build (-c cxx.extra_cxxflags=-Os); the default buck2 run is -O0 and unrepresentative.
Differential Revision: D108048181