new-kernel-checklist

name

new-kernel-checklist

description

PROACTIVELY USE THIS SKILL when adding a new GPU kernel, NVRTC kernel, CCCL primitive wrapper, or any new GPU-dispatched operation to src/vibespatial/. This checklist ensures every registration, warmup, caching, precision, dispatch, test, benchmark, and documentation step is completed. The CUDA-focused worker sub-agent and gpu-code-review skill should reference this checklist. Trigger on: "new kernel", "add kernel", "implement kernel", "write kernel", "create kernel", "new GPU operation", "add GPU", "scaffold kernel".

New Kernel Checklist — vibeSpatial

You are adding a new GPU kernel or GPU-dispatched operation. This checklist ensures nothing is missed. Work through each section in order — every item marked [REQUIRED] must be completed before the kernel can land.

Target kernel: $ARGUMENTS

---

Phase 1: Classification and Design

1.1 Tier Classification (ADR-0033) [REQUIRED]

Run the decision tree to determine the correct tier:

Is the inner loop geometry-specific (ring traversal, winding, segment intersection)?
  -> Yes: Tier 1 (custom NVRTC kernel)
  -> No: Is it segmented (per-row/ring/group reduction, sort, scan)?
    -> Yes: Tier 3a (CCCL segmented_*)
    -> No: Is it sort/unique/search/partition/compaction/scan?
      -> Yes: Tier 3a (CCCL)
      -> No: Is it element-wise/gather/scatter/concat?
        -> Yes: Tier 2 (CuPy)
        -> No: Tier 1 (custom NVRTC kernel)

Then for CCCL (Tier 3a):

• Can input use CountingIterator/TransformIterator? -> Tier 3c iterator
• Called repeatedly with same types/ops? -> Tier 3b make_*

Record: Tier: ___ Rationale: ___

1.2 Kernel Class (ADR-0002) [REQUIRED]

Classify precision behavior using KernelClass from src/vibespatial/runtime/precision.py:

Class	When	Precision Policy (Consumer GPU)
COARSE	Bounds, filters, sort keys, spatial index	Staged fp32 + centering (but bounds stay fp64)
METRIC	Area, length, distance, centroid	Staged fp32 + Kahan compensation + centering
PREDICATE	Point-in-polygon, binary predicates	Staged fp32 coarse pass + selective fp64 refinement
CONSTRUCTIVE	Clip, overlay, buffer, union	Native fp64 (wire plan for observability only)

Record: KernelClass: ___

1.3 Geometry Families [REQUIRED]

List which geometry types this kernel handles: point, multipoint, linestring, multilinestring, polygon, multipolygon

Record: geometry_families: (___)

---

Phase 2: Kernel Source (Tier 1 only — skip for Tier 2/3/4)

2.1 Write Kernel Source String [REQUIRED for Tier 1]

Define as a module-level constant. Key rules:

• Source is a raw string constant: _KERNEL_SOURCE = r""" ... """
• Use typedef {compute_type} compute_t; for precision templating
• All compute paths use compute_t, not hardcoded double
• Storage reads stay double: compute_t lx = (compute_t)(x[i] - center_x);
• All read-only pointers use const ... * __restrict__
• Use __launch_bounds__(threads, min_blocks) on kernel functions
• Use SoA layout (separate x[], y[]), never AoS
• Float constants use type-correct suffix: (compute_t)0.0, (compute_t)1e-7
• Integer division by power-of-2 uses bitwise: >> 5 not / 32

2.2 Define Kernel Names Tuple [REQUIRED for Tier 1]

_KERNEL_NAMES = ("my_kernel_main", "my_kernel_cooperative")

Every extern "C" __global__ function in the source must appear here.

2.3 Precision Variants [REQUIRED for Tier 1]

Create fp64 and fp32 variants of the source:

_KERNEL_FP64 = _KERNEL_SOURCE_TEMPLATE.format(compute_type="double")
_KERNEL_FP32 = _KERNEL_SOURCE_TEMPLATE.format(compute_type="float")

For METRIC kernels, also include Kahan summation macros and centering macros (CX(val), CY(val)) in the fp32 variant.

---

Phase 3: Compilation and Caching

3.1 Kernel Compilation (Tier 1) [REQUIRED for Tier 1]

Use the standard compilation pattern:

from vibespatial.cuda._runtime import get_cuda_runtime, make_kernel_cache_key

def _compile_my_kernel(compute_type: str = "double"):
    source = _KERNEL_FP64 if compute_type == "double" else _KERNEL_FP32
    suffix = "fp64" if compute_type == "double" else "fp32"
    runtime = get_cuda_runtime()
    cache_key = make_kernel_cache_key(f"my-kernel-{suffix}", source)
    return runtime.compile_kernels(
        cache_key=cache_key,
        source=source,
        kernel_names=_KERNEL_NAMES,
    )

• Cache key includes precision suffix (-fp64 / -fp32)
• Uses make_kernel_cache_key() (SHA1-based)
• Uses runtime.compile_kernels() (memory + disk cached)

3.2 CCCL Primitive Wrapper (Tier 3) [REQUIRED for Tier 3]

If adding a new CCCL primitive to cccl_primitives.py:

• Wrap behind a strategy enum (e.g., MyStrategy.AUTO/CCCL/CUPY)
• Implement make_* fast path checking CCCLPrecompiler.get().get_compiled()
• Implement one-shot fallback via cccl_algorithms.*() API
• Add synchronize parameter (default True)
• Benchmark make_* vs CuPy before setting AUTO default

---

Phase 4: Warmup and Precompilation Registration (ADR-0034)

4.1 NVRTC Warmup (Tier 1) [REQUIRED for Tier 1]

At module scope (after defining source constants), register for background precompilation:

from vibespatial.cuda.nvrtc_precompile import request_nvrtc_warmup

request_nvrtc_warmup([
    ("my-kernel-fp64", _KERNEL_FP64, _KERNEL_NAMES),
    ("my-kernel-fp32", _KERNEL_FP32, _KERNEL_NAMES),
])

• Called at module scope (not inside a function)
• Registers BOTH fp64 and fp32 variants
• Prefix matches the cache key prefix used in compilation
• Kernel names tuple matches exactly

4.2 Add to NVRTC Consumer Modules List [REQUIRED for Tier 1]

Add your module to _NVRTC_CONSUMER_MODULES in src/vibespatial/cuda/cccl_precompile.py (line ~1066):

_NVRTC_CONSUMER_MODULES: tuple[str, ...] = (
    ...
    "vibespatial.my_module.my_kernel",  # <-- ADD HERE
    ...
)

This ensures precompile_all() (used in CI and post-install) discovers and compiles your kernel.

4.3 CCCL Warmup Specs (Tier 3b/3c) [REQUIRED if using make_*]

If your kernel uses CCCL make_* callables:

Step A — Add specs to SPEC_REGISTRY in src/vibespatial/cuda/cccl_precompile.py, inside _build_spec_registry():

CCCLWarmupSpec(
    name="my_primitive_i32",
    family=AlgorithmFamily.MY_PRIMITIVE,
    key_dtype=np.int32,
    value_dtype=np.int32,
    op_name="sum",
)

• One spec per unique (algorithm, dtype, op) combination
• Name follows convention: {algorithm}_{dtype} (e.g., exclusive_scan_i32)

Step B — Request warmup at module scope in your consumer module:

from vibespatial.cuda.cccl_precompile import request_warmup

request_warmup(["my_primitive_i32", "my_primitive_i64"])

• Called at module scope
• Lists only the specs THIS module actually uses
• Idempotent (safe if multiple modules request the same spec)

4.4 CCCL CUBIN Disk Cache [CONDITIONAL — if using make_*]

If your make_* family is cacheable (no embedded LTOIR / Numba closures):

• Add a _Cached{Family} wrapper class in cccl_cubin_cache.py
• Add ctypes struct definition matching the C build_result layout
• Register the family in extract_cache_entry() and reconstruct_build()
• Test with VIBESPATIAL_CCCL_CACHE=1 and verify cache hits

Note: select and segmented_sort are NOT cacheable due to embedded LTOIR pointers. If your algorithm bakes closures into the build result, it falls into this category.

---

Phase 5: Dispatch and Precision Wiring

5.1 Kernel Variant Registration [REQUIRED]

Register both GPU and CPU variants using @register_kernel_variant():

from vibespatial.runtime.kernel_registry import register_kernel_variant
from vibespatial.runtime.precision import KernelClass, PrecisionMode
from vibespatial.runtime._runtime import ExecutionMode
from vibespatial.runtime.residency import Residency

@register_kernel_variant(
    "my_kernel",
    "gpu-cuda-python",
    kernel_class=KernelClass.THE_CLASS,
    execution_modes=(ExecutionMode.GPU,),
    geometry_families=("point", "polygon"),
    supports_mixed=True,
    precision_modes=(PrecisionMode.AUTO, PrecisionMode.FP32, PrecisionMode.FP64),
    preferred_residency=Residency.DEVICE,
    tags=("cuda-python",),
)
def _my_kernel_gpu(owned, *, precision_plan=None):
    ...

@register_kernel_variant(
    "my_kernel",
    "cpu",
    kernel_class=KernelClass.THE_CLASS,
    execution_modes=(ExecutionMode.CPU,),
    geometry_families=("point", "polygon"),
    supports_mixed=True,
    tags=("shapely",),
)
def _my_kernel_cpu(owned):
    ...

• GPU variant has preferred_residency=Residency.DEVICE
• GPU variant lists all three precision_modes
• CPU variant exists as Shapely fallback
• kernel_class matches your classification from Phase 1
• geometry_families matches your list from Phase 1

5.2 Precision Plan at Public API [REQUIRED]

At the public entry point, compute and pass through the precision plan:

from vibespatial.runtime.precision import select_precision_plan, KernelClass
from vibespatial.runtime.adaptive import plan_dispatch_selection

selection = plan_dispatch_selection(
    kernel_name="my_kernel",
    kernel_class=KernelClass.THE_CLASS,
    row_count=n,
    requested_mode=dispatch_mode,
)
precision_plan = select_precision_plan(
    runtime_selection=selection,
    kernel_class=KernelClass.THE_CLASS,
    requested=precision,
    coordinate_stats=coord_stats,  # if available
)

• select_precision_plan() called before kernel launch
• Plan passed to GPU kernel implementation
• Plan included in result metadata (if applicable)
• Public API accepts precision: PrecisionMode | str = PrecisionMode.AUTO

5.3 Kernel Launch (Tier 1) [REQUIRED for Tier 1]

runtime = get_cuda_runtime()
grid, block = runtime.launch_config(kernels["my_kernel"], item_count)
runtime.launch(kernels["my_kernel"], grid=grid, block=block, params=params)

• Uses runtime.launch_config() for occupancy-based sizing (NEVER hardcode block)
• Uses correct parameter type constants (KERNEL_PARAM_PTR, KERNEL_PARAM_I32, KERNEL_PARAM_F64)
• No unnecessary runtime.synchronize() between same-stream operations
• Sync only before host reads of device data

---

Phase 6: Tests

6.1 Unit Tests [REQUIRED]

Create tests/test_{kernel_name}.py:

• Uses @compare_with_shapely(reference=ref_func) decorator
• Parametrized with dispatch_mode fixture (tests both CPU and GPU)
• Tests null/empty/mixed geometry inputs
• Uses oracle_runner fixture for Shapely comparison
• Handles NotImplementedError for scaffold stubs via pytest.xfail
• Marked with @pytest.mark.gpu for GPU-specific tests

6.2 Precision Tests [REQUIRED]

• Test with explicit PrecisionMode.FP32 — verify correct results
• Test with explicit PrecisionMode.FP64 — verify correct results
• fp32 results within acceptable tolerance of fp64 ground truth
• If PREDICATE: test that ambiguous-case refinement produces correct results

6.3 Edge Cases [REQUIRED]

• Empty geometry collections
• Single-element inputs
• Very large inputs (>1M rows if applicable)
• Degenerate geometries (zero-area polygons, coincident points, etc.)

---

Phase 7: Benchmarks

7.1 Operation Benchmark [RECOMMENDED]

Add an entry in src/vibespatial/bench/operations/{category}_ops.py:

from vibespatial.bench.catalog import benchmark_operation

@benchmark_operation(
    name="my_kernel",
    description="Description of what is benchmarked",
    category="predicate",  # or constructive, spatial, overlay, io, misc
    geometry_types=("point", "polygon"),
    default_scale="100K",
    tier=1,
)
def bench_my_kernel(scale, precision, **kwargs):
    ...

• Registered via @benchmark_operation() decorator
• Discoverable via vsbench list operations

7.2 Kernel Microbenchmark (NVBench) [RECOMMENDED for Tier 1]

Create src/vibespatial/bench/kernels/bench_{kernel_name}.py:

• Auto-discovered by nvbench_runner.py via bench_*.py naming
• Supports --scale, --precision, --output-json CLI args
• Reports bandwidth metrics if memory-bound

---

Phase 8: Manifest and Documentation

8.1 Variant Manifest [REQUIRED]

Update src/vibespatial/kernels/variant_manifest.json:

{
  "kernel": "my_kernel",
  "module": "vibespatial.kernels.predicates",
  "tier": 1,
  "geom_types": ["point", "polygon"],
  "variants": ["cpu", "gpu-cuda-python"]
}

Or run the scaffold generator which does this automatically:

uv run python scripts/generate_kernel_scaffold.py my_kernel --tier 1 --geom-types point,polygon

8.2 Kernel Inventory Doc [REQUIRED]

Ensure docs/testing/kernel-inventory.md has a row for the new kernel. The scaffold generator updates this automatically.

8.3 Package Exports [REQUIRED]

If the kernel module is in a sub-package under src/vibespatial/kernels/, ensure the __init__.py exports it:

# src/vibespatial/kernels/{subpackage}/__init__.py
from .my_kernel import my_kernel

---

Phase 9: Quality Gates

9.1 Run $cuda-optimizer [REQUIRED]

Invoke the cuda-optimizer skill on your kernel file to verify:

• No unnecessary host round-trips
• No redundant synchronizations
• Efficient memory access patterns
• Correct use of streams

9.2 Run $precision-compliance [REQUIRED]

Invoke the precision-compliance skill to verify:

• Kernel source is templated on compute_t
• Cache key includes precision
• Centering and compensation wired correctly
• Update the compliance ledger in that skill

9.3 Run $gpu-code-review [REQUIRED]

Invoke the gpu-code-review skill to catch:

• Anti-patterns (hardcoded block sizes, missing __restrict__, AoS layout)
• Memory management issues
• Synchronization bugs
• Performance regressions

9.4 Run $pre-land-review [REQUIRED before commit]

Final gate. Must pass before creating a git commit.

---

Quick Reference: Files to Touch

What	File(s)
Kernel source	src/vibespatial/{module}/{kernel}.py
NVRTC warmup call	Same file as kernel source (module scope)
Consumer modules list	src/vibespatial/cuda/cccl_precompile.py (_NVRTC_CONSUMER_MODULES)
CCCL warmup specs	src/vibespatial/cuda/cccl_precompile.py (_build_spec_registry)
CCCL warmup call	Consumer module (module scope)
CCCL disk cache	src/vibespatial/cuda/cccl_cubin_cache.py (if new family)
CCCL primitive wrapper	src/vibespatial/cuda/cccl_primitives.py (if new primitive)
Variant registration	Same file as kernel source (decorator)
Dispatch wiring	Public API module
Tests	tests/test_{kernel}.py
Precision tests	tests/test_{kernel}.py or tests/test_precision_policy.py
Operation benchmark	src/vibespatial/bench/operations/{category}_ops.py
Kernel benchmark	src/vibespatial/bench/kernels/bench_{kernel}.py
Variant manifest	src/vibespatial/kernels/variant_manifest.json
Kernel inventory	docs/testing/kernel-inventory.md
Package __init__.py	src/vibespatial/kernels/{subpackage}/__init__.py

---

Conditional Checklist Summary

Use this table to determine which phases apply based on your tier:

Phase	Tier 1 (NVRTC)	Tier 2 (CuPy)	Tier 3 (CCCL)	Tier 4 (CuPy default)
1. Classification	YES	YES	YES	YES
2. Kernel Source	YES	skip	skip	skip
3.1 NVRTC Compilation	YES	skip	skip	skip
3.2 CCCL Wrapper	skip	skip	YES	skip
4.1 NVRTC Warmup	YES	skip	skip	skip
4.2 Consumer Modules	YES	skip	skip	skip
4.3 CCCL Warmup Specs	skip	skip	if make_*	skip
4.4 CCCL Disk Cache	skip	skip	if cacheable	skip
5. Dispatch & Precision	YES	YES	YES	YES
6. Tests	YES	YES	YES	YES
7. Benchmarks	YES	optional	optional	optional
8. Manifest & Docs	YES	YES	YES	YES
9. Quality Gates	YES	YES	YES	YES