pva-opgen
Generate, change, or repair a complete vectorized NVIDIA PVA operator from a validated pva-spec handoff. Deliver the contract's API, host and VPU code, independent CPU reference, sample, build integration, tests, and documentation. Verify with pva-build-test and use pva-debug within the configured attempt limit.
Try your first task
A source-grounded starter task has not been verified for this skill yet.
SKILL.md
name: pva-opgendescription: >- Generate, change, or repair a complete vectorized NVIDIA PVA operator from a validated pva-spec handoff. Deliver the contract's API, host and VPU code, independent CPU reference, sample, build integration, tests, and documentation. Verify with pva-build-test and use pva-debug within the configured attempt limit.metadata: author: "Venkat Balan <venkatramanb@nvidia.com>" tags: [pva, code-generation, build-test]Generate a PVA operator
Implement the approved Spec contract. Do not redefine it.
Resolve paths
SKILL_ROOT is the absolute directory containing this skill's SKILL.md,
after resolving symlinks. Bundled files are relative to that directory. Resolve
supplied relative paths against the invocation directory before changing
directories. Set OPGEN_INPUTS to the absolute
handoff path and WORK_DIR to the caller's report work directory. For a handoff
at <work-directory>/configure/opgen-inputs.json, default WORK_DIR to that
<work-directory>; otherwise require an explicit report work directory.
Use only the validated handoff's workspace as the project root.
Inputs
Before changing source, require installed pva-build-test and, when
PVA_OPERATOR_MAX_DEBUG_ATTEMPTS is greater than zero, pva-debug.
Resolve them by name through the agent's skill discovery; their directories
need not be adjacent. Return BLOCKED if a required skill is unavailable.
Run these skills using the agent's available skill mechanism. If no callable
skill tool exists, load the named skill's SKILL.md and execute its complete
workflow as the Build-Test or Debug phase. Save that phase's reports, then
resume OpGen with its verdict. Reading instructions alone does not execute a
phase.
Require:
- Spec's
opgen-inputs.json. - The recorded operator specification, test matrix, and evidence manifest.
- A Doctor report that is ready for
opgen, or accepted limitations. - Direct-Tegra inputs that Build-Test needs.
- A writable report directory.
| Script | Command | Inputs | Outputs | Exit codes |
|---|---|---|---|---|
| finalize_opgen_inputs.py | python3 "$SKILL_ROOT/scripts/finalize_opgen_inputs.py" --check "$OPGEN_INPUTS" | Spec handoff and its recorded inputs. | JSON validation summary; read-only check. | 0: valid; 1: failed; 2: invalid command arguments. |
Run this command before you inspect or change the workspace:
python3 "$SKILL_ROOT/scripts/finalize_opgen_inputs.py" \
--check "$OPGEN_INPUTS"
Return BLOCKED if the command reports an invalid structure, unknown
requirement link, unsafe path, missing evidence, changed input, stale hash or
size, or mismatched count. Do not repair Spec-owned files or handoff metadata.
Use the project workspace recorded in the validated handoff.
Verify the recorded hashes before reading project evidence files. Use API, CPU-reference, test, workload, KPI, benchmark, and build files only as contracts. Do not copy CPU-reference or test code into the implementation.
Read the technical basis
Use the Doctor report recorded in the validated handoff. Require its
public-reference-corpus check to be PASS. Set SDK_REFERENCES to that report's
absolute sdk_reference_root. Require a readable index.md and
reference-lock.json, a lock SHA-256 matching the report's
sdk_reference_lock_sha256, and the reported SDK version matching this task.
If these fields are absent or no longer match, return BLOCKED for a fresh
Doctor report and regenerated Spec handoff. Do not edit a finalized report or
fall back to an installed package or CWD path. Read the
pva-opgen route in $SDK_REFERENCES/index.md.
Before reading Markdown, check its SHA-256 against the lock
outputs entry; return BLOCKED on a mismatch. Resolve links relative to each
generated file.
Use rg to select only the programming, data-movement, build, tutorial, and
API Markdown that applies. Read the relevant Orin GEN2 or Thor GEN3 guide
chapters for the core, VMEM, address generation, vector operations, and
programming examples.
Before use of an intrinsic or mnemonic, find its exact type and operation in the generated guide pages. Record the SDK documentation root, local Markdown path, packaged source path, source SHA-256, guide revision, generation, and section or PDF page.
Use supplied or public implementations within the user's authorized scope and their license terms.
Inspect the project
Record:
- Public API and ownership conventions.
- Host and device target patterns.
- Registration, sample, test, and documentation patterns.
- Installed SDK version and PVA generation.
- Files that OpGen will add or change.
Do not require Git metadata or a fixed repository layout.
Design the implementation
Use the behavior and test contracts from Spec.
- Select an algorithm that gives the specified result.
- Separate parallel work from reductions and dependencies.
- Select the PVA dataflow and transfer plan.
- Select a tile and VMEM layout that fit the target generation.
- Define vector-tail, alignment, border, and scratch handling.
- Estimate compute and transfer work when the required facts exist.
- Record unavailable facts as limitations. Do not invent a Speed-of-Light value.
Use RasterDataFlow for suitable raster tiles. Use another documented dataflow only when the access pattern needs it.
Create the deliverables
Create or update all items that the Spec contract requires:
| Deliverable | Requirement |
|---|---|
| Public API | Use project naming, ownership, validation, and error conventions. Preserve asynchronous behavior. |
| Host code | Create resources, load VPU code, configure dataflows, submit work, check status, synchronize as required, and release owned resources. |
| VPU code | Use documented device APIs and vector intrinsics. Keep the main data path vectorized. Use a correct predicated or scalar tail only where required. |
| CPU reference | Preserve an approved independent reference. For a new operator, create independent scalar logic that implements the contract. |
| Build wiring | Add host, device, sample, test, and optional performance targets. Add a complete install target or deployment manifest. |
| Sample | Show one valid lifecycle, check the result, and release resources. |
| Tests | Implement the approved Spec test matrix without weakening expected results or tolerances. |
| Documentation | Document the API, formats, shapes, parameters, arithmetic, borders, limits, build steps, and sample. |
Do not reuse a VMEM buffer before its transfer or compute use is complete. Do not replace a documented vector operation with scalar work in the main data path. Record each necessary scalar region and its reason.
Do not derive an expected result from PVA output.
Verify with Build-Test
Run pva-build-test using the child-skill mechanism defined above.
Pass:
- Caller identity
OpGen,WORK_DIR, and the Debug calls used and maximum. - The validated Doctor report and Spec handoff.
- The active project root.
- The complete correctness matrix and EMU policy.
SKIPPED_OPTIONALevidence when optional EMU support is unavailable.- L4T build, install, target, and direct-SSH inputs.
build-test-opgenas the report label.- An absolute result directory under
$WORK_DIRfor each Build-Test call.
Use correctness-only when the matrix has no performance case. Use
correctness-and-performance when it has one or more performance cases. Pass
performance targets, clock inputs, filters, and sample policy only in the
second scope.
Build-Test owns all build, EMU, hardware, profiling, evidence, and verdict rules. Follow its instructions without skipping or weakening checks.
After every Build-Test call, resume OpGen and inspect its returned verdict and report. A child skill's final response does not complete OpGen. Follow the repair rules below, then write the final OpGen report before ending the run.
Repair a failed implementation
Start with zero Debug calls used and read the maximum from
PVA_OPERATOR_MAX_DEBUG_ATTEMPTS. Check the budget before every Debug call;
0 means no automatic repair. The next attempt number is calls used plus one.
Count each execution of the Debug workflow as one Debug call, regardless of
the invocation mechanism, including one that returns without a repair.
If Build-Test returns FAILED and an attempt remains, run pva-debug using
the same child-skill mechanism with the complete
failure report, failing source identity, command, logs, case, expected result,
tolerance, next attempt number, and maximum attempts.
Pass the same WORK_DIR for Debug reports.
After Debug returns, resume OpGen. For each REPAIRED result, invoke Build-Test
again on the changed source using a new result directory and the complete
original matrix. Choose the next action from the returned result:
| Result | OpGen action |
|---|---|
Build-Test PASSED | Write the final PASSED report. |
Build-Test FAILED, calls used below maximum | Invoke Debug, then resume OpGen. |
Build-Test FAILED, zero or exhausted budget | Write the final FAILED report with the failed stage, reason, evidence, and calls used. Do not attempt a repair. |
Build-Test or Debug BLOCKED | Write the final BLOCKED report with the exact unavailable prerequisite. |
Debug NOT_REPRODUCED | Write the final FAILED report, retaining the original failure and Debug's reproduction result. |
Record one attempt entry per Debug invocation, with its result and source identity. Do not reset or discount calls used when a failure remains unchanged.
Do not invoke Debug for an external BLOCKED result. Do not invoke Doctor or
Spec.
Output Format
Write $WORK_DIR/reports/opgen-report.md with
the result template.
Write it for every terminal outcome, including zero or exhausted debug budget,
before returning the final response. Report zero calls explicitly when Debug
was not invoked. If the report directory is unavailable, identify the exact
path and write error in the final response.
Return one verdict:
PASSED: All deliverables exist, and Build-Test passed the current source.FAILED: An executed implementation or verification stage produced a negative result that the bounded repair loop did not repair.BLOCKED: A required input, dependency, permission, corpus, or target is unavailable.
Report the changed files, API, supported matrix, vectorization, dataflow, VMEM plan, build targets, deployment method, Build-Test result, Debug attempts, evidence paths, hashes, and limitations.
Do not claim hardware verification from EMU execution. Stop after the OpGen result.