"""The GUDA-role path: a fast, vectorized INT8 GEMM on real CPU silicon. This is where the *throughput* comes from -- real CPU SIMD (via numpy/torch int32 accumulation), exactly GUDA's job. It is NOT a neural net and NOT GPU-fast; its ceiling is the CPU. The neural verified op certifies that this kernel is bit-faithful to the reference (verify_kernel.py) -- turning GUDA-style "IEEE parity within tolerance" into "provably equal on this finite integer op". """ from __future__ import annotations import numpy as np #: Largest contraction length an int32 accumulator can take without overflow. #: |a*b| <= 128*128 = 16384 for int8 operands, so K*16384 must stay inside int32. #: The bound is real headroom (131071), but layer width is user-selectable from the #: SpikeWhale panel, so it is checked rather than asserted in a comment -- an #: overflow here would silently produce wrong numbers, not raise. MAX_INT32_K = (2 ** 31 - 1) // (128 * 128) def gemm_int8(A: np.ndarray, B: np.ndarray) -> np.ndarray: """Signed INT8 GEMM with exact int32 accumulation (tensor-core semantics).""" k = A.shape[-1] if k > MAX_INT32_K: raise ValueError( "contraction length %d exceeds int32 accumulator headroom %d; " "accumulate in int64 for this size" % (k, MAX_INT32_K)) a = A.astype(np.int32) b = B.astype(np.int32) return a @ b # exact integer matmul within the checked bound def random_int8(shape, rng) -> np.ndarray: return rng.integers(-128, 128, size=shape, dtype=np.int16).astype(np.int8)