Uh oh!
There was an error while loading. Please reload this page.
Fix VisualWakeWords on GAP9 - #187
Merged
Merged
Conversation
`PULP2DDWConvTemplate` inherited `computeTransientBuffersSize` from the regular-conv template, which sizes the scratch as `2 * 8 * ch_in * KH * KW` — independent of `H_in`. The upstream `pulp_nn_depthwise_*` kernel actually walks a column-shaped im2col of `dim_kernel_x * (dim_in_y + pad_top + pad_bot) + dim_kernel_x` bytes per core, so for any input taller than the inherited bound the kernel writes past its scratch into the next L1 tensor. With 8 channels and a 3×3 kernel the two formulas coincide at H=45 and diverge from H=46 on (e.g. VWW PASS_1's 1×8×48×48 DW reproduces the corruption with ~74% of outputs wrong). Override `computeTransientBuffersSize` in `PULP2DDWConvTemplate` to use the kernel's actual per-core formula × NUM_CORES. Add a `Kernels/Integer/Conv/DW_2D_RQ_8x16x16` repro test exercising the previously-overflowing geometry, and wire VWW + the new kernel into the GAP9 L2 single- and double-buffer tiling rosters so CI catches any future regression. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bump pulp-nn-mixed to 1d8eeee0 (PR pulp-platform#11), which fixes the `pulp_nn_linear_*` family: when `num_o_neurons` is not a multiple of 2*NUM_CORES, the trailing odd neuron's `if (lft_neurons …)` branch called `pulp_nn_bn_quant_*(sum, *pKappa, *pLambda, …)` instead of the per-core advanced `*k1, *lambda1` pointers, so every core wrote element 0's RQS parameters into its own neuron. Any layer with an odd channel count (e.g. VWW's 2-class classifier) ended up with the wrong kappa/lambda. Rebuild the prebuilt GAP9 library against the fixed source. Split the dropped Clang-only `-Wno-incompatible-pointer-types- discards-qualifiers` on PULPOpen into the two flags GCC accepts (`-Wno-incompatible-pointer-types`, `-Wno-discards-qualifiers`) and add the same `-Wno-incompatible-pointer-types` plus `-Wno-implicit-function-declaration` to the GAP9 target so `pulp-nn-mixed` builds clean from source under the GCC GAP9 toolchain. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Xeratec
marked this pull request as ready for review
April 30, 2026 11:04
Contributor
Warning Review limit reached
Next review available in:47 minutes Limit details: You’ve used the included review currently available. You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits within each organization. For paid Pro and Pro+ reviews, CodeRabbit uses a developer's included PR review attempts over the past 7 days to set the current hourly allowance. At typical activity levels, the full plan allowance applies. Higher sustained activity can lower the allowance until earlier attempts leave the 7-day window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
📝 WalkthroughWalkthroughConverted several transient-buffer sizing helpers from static to instance methods, added depthwise-specific per-core im2col sizing in PULP DW conv templates, updated tiled test configs to include a DW kernel and VisualWakeWords model entries, tweaked compiler warning flags, and advanced a pulp-nn-mixed submodule pointer. Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Suggested reviewers
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@Deeploy/Targets/PULPOpen/Templates/ConvTemplate.py`:
- Around line 65-80: The hoisting path still calls
PULP2DConvTemplate.computeTransientBuffersSize directly so the depthwise
override in PULP2DDWConvTemplate is never used; update
PULP2DConvTemplate.hoistTransientBuffers to call the class/instance
computeTransientBuffersSize dynamically (e.g., self.computeTransientBuffersSize
or type(obj).computeTransientBuffersSize) and then pass each returned tuple into
ctxt.hoistTransientBuffer so that
PULP2DDWConvTemplate.computeTransientBuffersSize results are actually hoisted
via ctxt.hoistTransientBuffer.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: 7512f408-b7f3-4da9-8f16-434097f40bfb
📒 Files selected for processing (11)
Deeploy/Targets/PULPOpen/Templates/ConvTemplate.pyDeeployTest/Tests/Kernels/Integer/Conv/DW_2D_RQ_8x16x16/activations.npzDeeployTest/Tests/Kernels/Integer/Conv/DW_2D_RQ_8x16x16/inputs.npzDeeployTest/Tests/Kernels/Integer/Conv/DW_2D_RQ_8x16x16/network.onnxDeeployTest/Tests/Kernels/Integer/Conv/DW_2D_RQ_8x16x16/outputs.npzDeeployTest/test_gap9_tiled_config.pyDeeployTest/test_siracusa_tiled_config.pyTargetLibraries/GAP9/CMakeLists.txtTargetLibraries/GAP9/prebuilt/libpulp-nn-mixed.aTargetLibraries/PULPOpen/CMakeLists.txtTargetLibraries/third_party/pulp-nn-mixed
Uh oh!
There was an error while loading. Please reload this page.
`hoistTransientBuffers` called `computeTransientBuffersSize` via the hard-coded class name (e.g. `PULP2DConvTemplate.computeTransient...`), which bypassed any subclass override regardless of the actual `self` type. As a result the `PULP2DDWConvTemplate` override added in 65eda17 was dead code: VWW codegen kept using the regular-conv formula instead of the depthwise per-core size. Convert `computeTransientBuffersSize` from `@staticmethod` to an instance method on each affected template (PULP 1D/2D conv, PULP 2D float conv + DW, PULP iSoftmax, generic iSoftmax pre-allocated) and call it via `self.computeTransientBuffersSize(...)` so dispatch goes through the actual class. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`PULP1DDWConvTemplate` previously inherited `computeTransientBuffersSize` from the regular 1D conv template, which sizes scratch as `8 * 2 * ch_in * KY` — independent of `H_in`. The upstream `pulp_nn_depthwise_*` kernel walks a column-shaped im2col of `dim_kernel_y * (dim_in_y + pad_top + pad_bot) + dim_kernel_y` bytes per core, so for tall enough inputs the kernel overflows the inherited buffer. Add the depthwise per-core formula × NUM_CORES, mirroring the 2D fix in 65eda17. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Victor-Jung
commented
May 3, 2026
Member
NITPICK: Update the link #11 to point at the pulp-nn-mixed repo PR. |
Victor-Jung
left a comment
Member
There was a problem hiding this comment.
Changes look good to me. Thanks for the fixes! Why is the CI failing? Hasn't the Docker been updated with the new pulp-nn-mixed version? As soon as it passes, we can merge.
Victor-Jung
marked this pull request as draft
June 18, 2026 13:27
5 tasks
PULP2DDWConvTemplate.computeTransientBuffersSize sized the scratch with dim_kernel_x * (dim_im_in_y + pad_top + pad_bot) + dim_kernel_x. That matches the kernel's formula, but the call site passes the spatial axes transposed: pulp_nn_depthwise's dim_in_y receives dim_im_in_x and its dim_kernel_x receives dim_kernel_y. Both terms have to be transposed here as well. Square feature maps hide this. MLPerf KeywordSpotting does not: its depthwise layers are 20x5 with a 3x3 kernel, so the buffer was sized 8 * (3 * (5 + 1 + 1) + 3) = 192 B where the kernel writes 8 * (3 * (20 + 1 + 1) + 3) = 552 B The 360 B overrun corrupts the L2 heap and the run dies after "Input copied" with /chip/cluster/pe8/lsu Invalid access (offset: 0xa63b0045, size: 0x3) which is what the gap9-models CI job has been reporting. PULP1DDWConvTemplate is left alone: it sizes with dim_kernel_y while the 1D call site pins the kernel's dim_kernel_x to 1, so it over-allocates rather than under-allocates. Verified on GAP9 / GVSoC, 8 cores: pytest -m "gap9 and models" 1 failed -> 5 passed pytest -m "gap9 and kernels" 31 passed pytest -m "gap9_tiled and models and singlebuffer and l2" 12 passed pytest -m "gap9_tiled and models and doublebuffer and l2" 12 passed
runwangdl
marked this pull request as ready for review
August 19, 2026 16:51
runwangdl
left a comment
Contributor
There was a problem hiding this comment.
LGTM. I have fixed the bugs for CI.
runwangdl
approved these changes
Aug 19, 2026
Victor-Jung
commented
Aug 20, 2026
Member
Amazing! Thank you, Run! 🚀 |
Uh oh!
There was an error while loading. Please reload this page.
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 freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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
Two independent correctness bugs surfaced while bringing up the MLPerf VisualWakeWords (VWW) model on GAP9 with tiling + double buffering. Each is fixable on its own, and either alone is enough to corrupt VWW outputs. Both are fixed here, and VWW is wired into CI on GAP9 so we don't regress.
1. DWConv im2col scratch buffer is undersized on PULPOpen
PULP2DDWConvTemplatewas inheritingcomputeTransientBuffersSizefrom the regular-conv template, which sizes the scratch as2 * 8 * ch_in * KH * KW— independent ofH_in. The upstreampulp_nn_depthwise_*kernel actually walks a column-shaped im2col ofdim_kernel_x * (dim_in_y + pad_top + pad_bot) + dim_kernel_xbytes per core, so for any input taller than the inherited bound the kernel writes past its scratch into the next L1 tensor.With 8 channels and a 3×3 kernel the two formulas coincide at
H=45and diverge fromH=46onward — VWW PASS_1's1×8×48×48DW reproduces the corruption with ~74% of outputs wrong.Fix: override
computeTransientBuffersSizeinPULP2DDWConvTemplateto use the kernel's actual per-core formula ×NUM_CORES.Also adds a
Kernels/Integer/Conv/DW_2D_RQ_8x16x16repro test exercising the previously-overflowing geometry, so CI catches it.2.
pulp_nn_linear_*leftover-neuron RQS parameter aliasingWhen
num_o_neuronsis not a multiple of2*NUM_CORES, the trailing odd-neuronif (lft_neurons …)branch in thepulp_nn_linear_*family calledpulp_nn_bn_quant_*(sum, *pKappa, *pLambda, …)instead of the per-core advanced*k1, *lambda1pointers. Every core ended up writing element-0's RQS parameters into its own neuron, so any layer with an odd channel count (e.g. VWW's 2-class classifier) was silently quantized with the wrong kappa/lambda.Fix lives in
pulp-nn-mixed(PR #11 —1d8eeee0); this PR bumps the submodule and rebuilds the prebuilt GAP9 library against the fixed source.While in there: split the dropped Clang-only
-Wno-incompatible-pointer-types-discards-qualifierson PULPOpen into the two flags GCC accepts (-Wno-incompatible-pointer-types,-Wno-discards-qualifiers), and add the same-Wno-incompatible-pointer-typesplus-Wno-implicit-function-declarationto GAP9 sopulp-nn-mixedbuilds cleanly from source under the GCC GAP9 toolchain.PR Merge Checklist
develcommit and pointing todevel.CHANGELOG.mdfile has been updated.