vibespatial.io.gpu_parse.properties¶

GPU property extraction primitives for GeoJSON Feature objects.

Extracts numeric, boolean, and null property values directly on GPU, avoiding the per-feature orjson.loads() CPU bottleneck. String properties are intentionally NOT handled – the caller falls back to CPU for those columns.

Pipeline overview:

# Inputs: d_bytes, d_quote_parity, d_depth, feature boundaries
#
# 1. Locate property key positions via pattern_match + depth filter
# 2. Classify value types (string/bool/null/number/complex)
# 3. For numeric columns: extract floats via number_boundaries + parse
# 4. For boolean columns: NVRTC kernel reads "true"/"false"
# 5. For null columns: mark in validity mask
# 6. Schema inference on first N features (only D->H transfer)

The only host round-trip is schema inference (reading unique key names and their types from a small sample). All column extraction runs entirely on GPU.

Kernel classification (ADR-0033):

classify_property_values: Tier 1 (byte-level pattern at specific positions)
extract_booleans: Tier 1 (byte-level “true”/”false” recognition)
All other operations: Tier 2 (CuPy element-wise / gpu_parse reuse)

Precision (ADR-0002):

Integer-only byte classification kernels – no PrecisionPlan needed.
Numeric property values are parsed to fp64 via parse_ascii_floats (storage is always fp64 per ADR-0002).

Attributes¶

`cp`
`KERNEL_PARAM_I64`
`VTYPE_STRING`
`VTYPE_BOOLEAN`
`VTYPE_NULL`
`VTYPE_NUMBER`
`VTYPE_COMPLEX`

Functions¶

`infer_property_schema`(→ dict[str, int])	Infer property names and their types from a sample of features.
`extract_gpu_properties`(→ dict[str, cupy.ndarray])	Extract numeric/boolean properties on GPU.

Module Contents¶

vibespatial.io.gpu_parse.properties.cp = None¶

vibespatial.io.gpu_parse.properties.KERNEL_PARAM_I64¶

vibespatial.io.gpu_parse.properties.VTYPE_STRING: int = 0¶

vibespatial.io.gpu_parse.properties.VTYPE_BOOLEAN: int = 1¶

vibespatial.io.gpu_parse.properties.VTYPE_NULL: int = 2¶

vibespatial.io.gpu_parse.properties.VTYPE_NUMBER: int = 3¶

vibespatial.io.gpu_parse.properties.VTYPE_COMPLEX: int = 4¶

vibespatial.io.gpu_parse.properties.infer_property_schema(d_bytes: cupy.ndarray, d_quote_parity: cupy.ndarray, d_depth: cupy.ndarray, d_feature_starts: cupy.ndarray, d_feature_ends: cupy.ndarray, *, sample_size: int = 100, property_depth: int = 4) → dict[str, int]¶

Infer property names and their types from a sample of features.

Scans the first sample_size features for property keys at the specified depth, classifies each key’s value type, and returns a schema mapping property name to value type code.

This is the ONLY acceptable D->H transfer in the property extraction pipeline.

Parameters¶

d_bytescp.ndarray: Device-resident uint8 byte array.
d_quote_paritycp.ndarray: Device-resident uint8 quote parity mask.
d_depthcp.ndarray: Device-resident int32 bracket depth array.
d_feature_startscp.ndarray: Device-resident int64 array of feature start byte offsets.
d_feature_endscp.ndarray: Device-resident int64 array of feature end byte offsets.
sample_sizeint: Number of features to sample for schema detection.
property_depthint: Bracket depth of property keys (default 4 for GeoJSON).

Returns¶

dict[str, int]: Mapping of property name to value type code (VTYPE_* constants). String columns are included (VTYPE_STRING) so the caller knows which columns need CPU fallback.

vibespatial.io.gpu_parse.properties.extract_gpu_properties(d_bytes: cupy.ndarray, d_feature_starts: cupy.ndarray, d_feature_ends: cupy.ndarray, d_quote_parity: cupy.ndarray, d_depth: cupy.ndarray, *, property_depth: int = 4, sample_size: int = 100) → dict[str, cupy.ndarray]¶

Extract numeric/boolean properties on GPU.

Performs schema inference on a small sample (the only D->H transfer), then extracts each non-string property column entirely on GPU.

Parameters¶

d_bytescp.ndarray: Device-resident uint8 byte array of the full file.
d_feature_startscp.ndarray: Device-resident int64 feature start byte offsets.
d_feature_endscp.ndarray: Device-resident int64 feature end byte offsets.
d_quote_paritycp.ndarray: Device-resident uint8 quote parity mask.
d_depthcp.ndarray: Device-resident int32 bracket depth array.
property_depthint: Bracket depth of property keys (default 4 for GeoJSON).
sample_sizeint: Number of features to sample for schema inference.

Returns¶

dict[str, cp.ndarray]

Mapping of property name to device array of values.

Numeric properties: float64 array of shape (n_features,)
Boolean properties: uint8 array (1=true, 0=false) of shape (n_features,)
Null-only properties: excluded from output (all-null columns are dropped)
String properties: NOT included (caller falls back to CPU)
Complex properties: NOT included (nested objects/arrays skipped)

For columns with mixed types across features (e.g., some features have a numeric value and others have null), the output array has NaN for missing numeric values and 0 for missing boolean values.