Repository navigation
Bring the hackathon branch up to date with main - #29
Merged
Merged
Conversation
armclang rejects alignas after a GNU attribute ("'alignas' attribute
cannot be applied to types"); the attribute after the declarator is
accepted by AC6, GCC and Clang.
The FVP loaded the runner's libpython3.11 and died at start-up because that build's standard library is not at its baked prefix; PYTHONHOME on the FVP step points it at a complete Python 3.11. The flatbuffers pin check used grep -q under pipefail, so grep closing the pipe early failed the step on Linux and macOS.
--update-rte is no longer needed on a fresh checkout, the manual FVP run uses the same shim as the Run and Debug buttons, the wheel pin that has to match the pack is named, the project layout lists the task drop-in and the shim, and the MLOps flow page names the six components the generated layer selects. The Docker contrast is gone with the flow it referred to.
The region header claimed pack defaults that were this example's own values; every Default line now states the SSE_320_BSP value. The layer advertised a 768 kB heap while the scatter file reserves the 96 kB from regions_SSE-320.h. The FVP configuration no longer refers to scripts and tools this repository does not have.
…comments The arm_embedded_module headers pointed at a LICENSE in the repository root, which is Apache-2.0; src/LICENSE-ExecuTorch now carries the BSD text. The generator's docstrings name every component it selects and why the tensor extension is not among them, and the generated layer's Re-run line names the argument the script needs.
CI no longer passes --update-rte (the RTE configuration is committed) and keys the pack cache on the csolution so a pack bump refreshes it. The .gitignore loses patterns for tools and directories this repository never had and gains the FVP log the README and CI write. The shim's header shows the arguments the launch config really passes and the Windows advice that the README gives.
ExecuTorch 1.4.1 is on PyPI together with torch 2.13.0 and torchao 0.18.0, so the nightly index and requirements-executorch.txt go; every pin lives in requirements.txt. The pack pin follows. The generated layer now names the pack version cbuild setup resolved, and the generator reads that version by the csolution's exact pin: the cbuild-pack.yml lock keeps earlier resolutions, and an unversioned selector in the layer left the previous pack listed next to the pinned one. Same 8832-byte program and logits as with 1.4.0, verified on the FVP.
The Corstone-320 job becomes a toolchain matrix and runs the FVP for each image; the LLVM embedded toolchain joins vcpkg-configuration.json. Only AC6 was exercised before, which is how a component that does not build with the LLVM toolchain could go unnoticed.
The Linux x86 build of ATfE 21.1.1 crashes in its register coalescer on the quantize and dequantize kernels; 22.1.0 compiles them.
CMSIS-Toolbox 2.14.1 passes -mfpu=fpv5-sp-d16 to Clang for the Cortex-M85 (and M55), a single-precision FPU on cores with a double-precision FPU and MVE; LLVM 21.1.1 and 22.1.0 crash in the register coalescer on the quantize and dequantize kernels with that combination. A later -mfpu on the command line wins, so the csolution adds -mfpu=fp-armv8-fullfp16-d16 for Clang; newer toolboxes already leave the FPU to -mcpu.
create_ai_layer.py re-executes in .venv whenever the running interpreter is not that venv; testing whether torch imports kept the export on a host interpreter that happens to have torch. The Vela ini goes into the compile spec as a relative path so the exported program does not embed the checkout location. The pack-build recipe carries the CMake configure, all generator arguments, the structural validator and a non-interactive install; its checks now fail instead of echoing. Comments and the README no longer attribute stdout to a UART, the flatbuffers pin to the executorch wheel, or shutdown to uart0, and the adaptation section names the runner's fixed input shape.
RTE/ is no longer partly ignored: the per-context RTE_Components.h and Pre_Include_Global.h the toolbox writes are committed next to the layers' configuration files.
… wheels ExecuTorch 1.4.1 allows Python <3.15, but the tosa-tools 2026.5.0 it pins for the Arm backend only ships wheels for CPython 3.10 to 3.13. On 3.14 pip fails with "No matching distribution found for tosa-tools==2026.5.0" long after the venv was created. Check the bound up front and document it.
setup_venv.py and setup_venv.sh refuse 3.14 since tosa-tools 2026.5.0 has no wheels for it; the Windows wrapper's default range for uv and the prompt of the uv setup task still said <3.15.
create_ai_layer.py picked --accelerator-config, --system-config and --memory-mode out of vela.options of the cbuild-mlops.yml with a regular expression and dropped the rest, so `vela: misc:` in the csolution had no effect and nothing said so. The options are now tokenized; the three that are arguments of EthosUCompileSpec go there and every other one reaches Vela as an extra flag. An option given twice, and --config, --output-format and --output-dir, which the export sets itself, end the script with a message. Without misc: the exported program is byte for byte the one in ai_layer/.
The PyTorch::ExecuTorch pack 1.5.1 is in the public pack index and executorch 1.5.1 is on PyPI. It targets torch 2.14.0 (torch_pin.py at the v1.5.1 tag) and keeps torchao 0.18, ethos-u-vela 5.1.0 and tosa-tools 2026.5.0, so requirements-arm-tosa.txt is unchanged. The pack pins in the csolution and the cproject follow, and the generated layer and the RTE context headers name the new pack version. The exported program is 8784 bytes instead of 8832.
TinyCNN took its image as NCHW. The Ethos-U computes on NHWC feature maps, so the command stream began with an operation that only reorders the input. The model now takes NHWC, the layout of a camera frame, and permutes it to NCHW itself; Vela folds that into the first convolution. The layout conversion is gone from the command stream, the program is 8720 bytes instead of 8784, and the Ethos-U scratch 5376 bytes instead of 8704. The runner declares its input as 1x16x16x3; its ramp now fills the image in that order, so the logits differ from the ones before.
….0 (#27) The target-set's model: is FVP_Corstone_SSE-320 again, so Run and Debug work on Windows and Linux with the launch.json the extension generates; the Docker shim and its launch configuration (now .vscode/launch.json.mac) are the macOS setup, described in the README. CMSIS-Toolbox 2.15.0 no longer passes -mfpu=fpv5-sp-d16 to Clang for the Cortex-M85, so the CLANG misc: workaround goes; vcpkg pins the toolbox at 2.15.0.
From the preview branch, now that CMSIS-Toolbox 2.15.0 is released: the toolbox detects the simulator target from the Arm-FVP target-set, so simulator: goes, and passes extra keys under model: through to *.cbuild-mlops.yml. The NHWC input shape and the number of calibration samples move there; model_params() in create_ai_layer.py hands them to model/model.py. The exported program is unchanged.
MD schema compliance fix.
…WC input Brings the hackathon branch up to date with main (#21, #23, #25, #26, #27, #28): ExecuTorch 1.5.1 with torch 2.14.0, the NHWC model input, every Vela option of the mlops: node passed to Vela, the .venv python in the Create AI layer task, CMSIS-Toolbox 2.15.0 without the CLANG -mfpu workaround, and the FVP run directly with fvp.sh plus launch.json.mac as the macOS setup. The hackathon parts stay: the DevKit-E8 target-type, the Ensemble packs, the Ensemble Vela system config, the Alif tasks, the DevKit-E8 CI job and the one-page README. CMSIS-Toolbox 2.15 detects both the hardware and the simulator target from the target-sets, so neither is named in mlops: any more; input-shape and calibration-samples move under model:. The regenerated AI layer is 8736 bytes and prints the same logits as main on the FVP; the README, example.md and mlops-flow.md follow.
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
4 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.
Merges
mainintohackathon(a merge commit, so please merge this PR with Create a merge commit, not squash). It brings in #21, #23, #25, #26, #27 and #28:mlops:node forwarded to Vela.venvpython in the Create AI layer task-mfpuworkaroundFVP_Corstone_SSE-320, withfvp.sh+.vscode/launch.json.macas the macOS setupKept from
hackathon: the DevKit-E8 target-type, the Ensemble packs and Vela system config, the Alif tasks, the DevKit-E8 CI job and the one-page README. CMSIS-Toolbox 2.15 detects the hardware (DevKit-E8) and simulator (SSE-320-U85) targets from the target-sets, somlops:no longer names them.input-shapeandcalibration-samplesare undermodel:.Tested locally (toolbox 2.15.1, AC6): the regenerated AI layer is 8736 bytes, and both targets build. On the FVP the program prints the same logits as
mainandTest_result: PASS. No DevKit-E8 board was attached, so the hardware run is untested. CI covers both targets with AC6, GCC and CLANG.After this is merged,
hackathon-pico-faces(#22) gets the same update.