feat(cuda): TexTable — layered 1-D lookup-table textures with point/hardware filtering

[NicolasRouquette]

TexTable.ofFloatArray data width layers hwFilter builds an immutable layered 1-D float32 table; TexTable.fetch tbl coords layerIdx evaluates piecewise-linear interpolation at grid-space coordinates u ∈ [0, width−1] (clamp addressing; the +0.5 texel-center offset is handled internally) in one launch over the existing Buffer marshaling path. On the CUDA build the table is a layered cudaArray bound once to a cudaTextureObject_t; the default build ships a host-memory parity stub, so lake build stays green without a GPU.

Two filter modes, fixed at construction:

• point (default): two point fetches + an explicit contraction-blocked (__fadd_rn/__fmul_rn/__fsub_rn) float32 lerp — bit-identical between the CUDA kernel and the CPU stub. (The native backend compiles without --fmad=false, so FMA is blocked by intrinsics, not flags.)
• hardware: a single tex1DLayered filtered fetch — zero-ALU lerp using the texture unit's 9-bit fixed-point weight, tolerance-validated (CUDA does not specify coefficient rounding); the stub emulates the quantized weight.

Motivation

Small tabulated transfer functions — calibration curves, inverted sensor-response tables, activation LUTs — are read-only, reused across many launches, and small enough to stay texture-cache resident. Binding one as a texture object turns a gather that would otherwise be a multi-launch composed pipeline into a single fetch per element. The op is forward-only by design (no autograd node); the lerp-slope backward is well-defined future work.

Tests

NN/Tests/Runtime/Cuda/TexTable.lean (wired into the CUDA coverage suite): point mode bit-exact vs the executable texLerpSpec reference — compared by Float.toBits, not a tolerance; hardware mode within the 9-bit-weight bound of the unquantized reference; integer-node texel-center probes exact in both modes (catches ±0.5 coordinate errors); clamp-below/above, width = 1, multi-layer, and empty-coords edges. Like the rest of the suite it runs on the CPU stub (lake build) and on the GPU (-K cuda) alike.

Verification

• nvcc compiles the kernel TU; cc compiles the stub; the @[extern] bindings elaborate; both TUs export the same symbol set so the default build links.
• nn_tests_suite is green on the CPU stub and with -K cuda=true on an RTX A4500; the point-mode stub↔GPU bit-identity holds on real texture hardware, and the texel-center probes pass under both filter modes.