topic: simd
24 modules sharing the simd name family (derived from the tree's prefix discipline).
The 'simd' topic family in the Nishi sovereign ecosystem focuses on vectorized operations and performance benchmarking for SIMD (Single Instruction, Multiple Data) processing. The nx_simd module provides lane-wise vector operations and a vectorizing transform, enabling efficient data parallelism. The nx_simd_dot_bench module compares x86 integer-SIMD dot products with scalar implementations, highlighting performance gains. Meanwhile, nx_simd_census evaluates the vector backend's competitiveness through internal benchmarks, ensuring the system's reliability and efficiency.
auto-narrated by the local model from this topic's module headers; links verified against the wiki index.
narrated overview -- maintained by the narration lane, module links verified against this wiki.
| module | description | lines | funcs |
|---|---|---|---|
| nx_simd.nx | sovereign SIMD: lane-wise vector ops + a VECTORIZING TRANSFORM, the | 54 | 7 |
| nx_simd_avx2_kat_gate.nx | SOVEREIGN 256-bit AVX2 vector run-KAT (the VEX rung): proves the | 113 | 7 |
| nx_simd_bench.nx | honest perf comparison: scalar i64 add vs | 116 | 3 |
| nx_simd_census.nx | SOVEREIGN vector-backend competitive census. The TEAM measures its OWN | 156 | 10 |
| nx_simd_dot_bench.nx | x86 integer-SIMD dot (vpmaddwd via __i16x16_madd) vs scalar: verify + speedup. | 125 | 4 |
| nx_simd_i16x16_test.nx | exercise the i16x16 SIMD intrinsics. | 118 | 7 |
| nx_simd_i32x8_test.nx | exercise the i32x8 SIMD intrinsics. | 128 | 6 |
| nx_simd_i8x32_test.nx | exercise the i8x32 SIMD intrinsics. | 106 | 6 |
| nx_simd_isa.nx | a sovereign VECTOR ISA + SIMD emulator, so vectorized kernels | 50 | 2 |
| nx_simd_isa_test.nx | the vectorized FMA kernel as real vector MACHINE CODE, | 73 | 3 |
| nx_simd_lever_gate.nx | PROVE the SIMD lever that WINS the PyTorch h2h, using the REAL supported intrinsic | 49 | 6 |
| nx_simd_madd_probe.nx | test the x86 integer-SIMD primitive __i16x16_madd (vpmaddwd) on nx_cc_known_good. | 65 | 2 |
| nx_simd_minmax_test.nx | i16x16 per-lane and horizontal min/max. | 146 | 8 |
| nx_simd_probe.nx | confirm the __simd i16x16 vdot intrinsics compile + work in a nx_syscalls organ. | 30 | 2 |
| nx_simd_probe2.nx | verify the POINTER-FORM __simd_vdot_i16_x16 compiles AND computes on x86. | 36 | 2 |
| nx_simd_sat_test.nx | i16x16 signed saturating add/sub. | 148 | 6 |
| nx_simd_sse_kat_gate.nx | SOVEREIGN 128-bit SSE2 vector run-KAT: the | 127 | 7 |
| nx_simd_sse_rot_gate.nx | SOVEREIGN 128-bit SSE2 packed-shift/bitwise run-KAT | 116 | 7 |
| nx_simd_sse_shuf_gate.nx | SOVEREIGN 128-bit pshufd run-KAT (rung R2+R3: | 109 | 7 |
| nx_simd_test.nx | prove the vectorizing transform is 1:1 with the scalar kernel | 60 | 3 |
| nx_simd_usat_shift_test.nx | unsigned saturating arith + per-lane shifts (i16x16). | 159 | 7 |
| nx_simd_vdot_test.nx | widening dot product i16x16 -> i64. | 149 | 8 |
| nx_simd_x86_test.nx | AVX2-subset SIMD smoke (excludes VMUL which | 49 | 1 |
| simd.nx | SIMD vector types + intrinsics (spec + soft-fallback). | 203 | 13 |