|
| 1 | +//! `simd_profile_probe` — boot-on-silicon diagnostic for the dispatch matrix. |
| 2 | +//! |
| 3 | +//! Step 1 of the TEST-promotion checklist from |
| 4 | +//! `.claude/knowledge/td-simd-cpu-dispatch-matrix.md` § "TEST verification |
| 5 | +//! checklist": *"Boot the binary on the silicon and confirm `simd_profile()` |
| 6 | +//! returns the expected variant."* |
| 7 | +//! |
| 8 | +//! Prints every CPUID-derived capability bit plus the resolved `SimdProfile` |
| 9 | +//! variant. Used to verify silicon → profile mapping when promoting DOC |
| 10 | +//! cells in the dispatch matrix to TEST. |
| 11 | +//! |
| 12 | +//! Usage: |
| 13 | +//! ```sh |
| 14 | +//! # Runtime detection (default — same binary on any silicon): |
| 15 | +//! cargo run --example simd_profile_probe --release |
| 16 | +//! |
| 17 | +//! # Compile-time pinned (the LazyLock is not linked in): |
| 18 | +//! cargo run --example simd_profile_probe --release --features cpu-spr |
| 19 | +//! ``` |
| 20 | +
|
| 21 | +use ndarray::hpc::simd_caps::{simd_caps, ArmProfile, SimdCaps}; |
| 22 | +use ndarray::hpc::simd_profile::{is_pinned, pinned_profile, simd_profile, SimdProfile}; |
| 23 | + |
| 24 | +fn main() { |
| 25 | + let caps = simd_caps(); |
| 26 | + let profile = simd_profile(); |
| 27 | + |
| 28 | + println!("ndarray simd-profile probe"); |
| 29 | + println!("=========================="); |
| 30 | + println!(); |
| 31 | + |
| 32 | + // ── Dispatch identity ─────────────────────────────────────────── |
| 33 | + println!("Resolved profile: {}", profile.name()); |
| 34 | + println!(" is_x86: {}", profile.is_x86()); |
| 35 | + println!(" is_aarch64: {}", profile.is_aarch64()); |
| 36 | + println!(" has_avx512: {}", profile.has_avx512()); |
| 37 | + println!(" has_amx: {}", profile.has_amx()); |
| 38 | + println!(); |
| 39 | + |
| 40 | + // ── Pinning status ───────────────────────────────────────────── |
| 41 | + println!("Compile-time pinning: {}", if is_pinned() { "ACTIVE" } else { "off (runtime detection)" }); |
| 42 | + if let Some(p) = pinned_profile() { |
| 43 | + println!(" Pinned variant: {}", p.name()); |
| 44 | + } |
| 45 | + println!(); |
| 46 | + |
| 47 | + // ── Raw capability bits ──────────────────────────────────────── |
| 48 | + println!("SimdCaps (raw bits):"); |
| 49 | + print_caps(&caps); |
| 50 | + println!(); |
| 51 | + |
| 52 | + // ── ARM-specific sub-profile (heuristic; deployment-pragmatic) ── |
| 53 | + let arm = caps.arm_profile(); |
| 54 | + if !matches!(arm, ArmProfile::NotArm) { |
| 55 | + println!("ARM profile (heuristic): {}", arm.name()); |
| 56 | + println!(" est. tok/sec: {}", arm.estimated_tok_per_sec()); |
| 57 | + println!(" eff. f32 lanes:{}", arm.effective_f32_lanes()); |
| 58 | + println!(); |
| 59 | + } |
| 60 | + |
| 61 | + // ── Build configuration ───────────────────────────────────────── |
| 62 | + println!("Build:"); |
| 63 | + println!(" target_arch: {}", std::env::consts::ARCH); |
| 64 | + println!(" target_os: {}", std::env::consts::OS); |
| 65 | + #[cfg(target_feature = "avx512f")] |
| 66 | + println!(" -Ctarget-feature avx512f: yes (compile-time)"); |
| 67 | + #[cfg(not(target_feature = "avx512f"))] |
| 68 | + println!(" -Ctarget-feature avx512f: no (compile-time)"); |
| 69 | + #[cfg(target_feature = "avx2")] |
| 70 | + println!(" -Ctarget-feature avx2: yes (compile-time)"); |
| 71 | + #[cfg(not(target_feature = "avx2"))] |
| 72 | + println!(" -Ctarget-feature avx2: no (compile-time)"); |
| 73 | + println!(); |
| 74 | + |
| 75 | + // ── TEST promotion guidance ──────────────────────────────────── |
| 76 | + println!("Matrix-doc cells affected by this CPU:"); |
| 77 | + matrix_cell_summary(profile); |
| 78 | + |
| 79 | + // Sanity invariant: simd_profile() and pinned_profile() must agree |
| 80 | + // when pinning is active. This is the same check that |
| 81 | + // `pinning_consistency` runs as a unit test; we re-run it here so a |
| 82 | + // probe binary deployed on real silicon flags any future regression |
| 83 | + // in the cfg cascade. |
| 84 | + if let Some(p) = pinned_profile() { |
| 85 | + assert_eq!( |
| 86 | + profile, p, |
| 87 | + "INVARIANT VIOLATION: pinned_profile()={:?} disagrees with simd_profile()={:?}", |
| 88 | + p, profile |
| 89 | + ); |
| 90 | + } |
| 91 | +} |
| 92 | + |
| 93 | +fn print_caps(c: &SimdCaps) { |
| 94 | + let bits: &[(&str, bool)] = &[ |
| 95 | + ("avx2", c.avx2), |
| 96 | + ("avx512f", c.avx512f), |
| 97 | + ("avx512bw", c.avx512bw), |
| 98 | + ("avx512vl", c.avx512vl), |
| 99 | + ("avx512vnni", c.avx512vnni), |
| 100 | + ("avx512vbmi", c.avx512vbmi), |
| 101 | + ("avx512vpopcntdq", c.avx512vpopcntdq), |
| 102 | + ("avx512bf16", c.avx512bf16), |
| 103 | + ("avx512fp16", c.avx512fp16), |
| 104 | + ("avx512vp2intersect", c.avx512vp2intersect), |
| 105 | + ("avxvnniint8", c.avxvnniint8), |
| 106 | + ("amx_tile", c.amx_tile), |
| 107 | + ("amx_int8", c.amx_int8), |
| 108 | + ("amx_bf16", c.amx_bf16), |
| 109 | + ("amx_fp16", c.amx_fp16), |
| 110 | + ("fma", c.fma), |
| 111 | + ("sse41", c.sse41), |
| 112 | + ("sse2", c.sse2), |
| 113 | + ("neon", c.neon), |
| 114 | + ("asimd_dotprod", c.asimd_dotprod), |
| 115 | + ("fp16 (arm)", c.fp16), |
| 116 | + ("aes", c.aes), |
| 117 | + ("sha2", c.sha2), |
| 118 | + ("crc32", c.crc32), |
| 119 | + ]; |
| 120 | + for (name, present) in bits { |
| 121 | + println!(" [{}] {}", if *present { "x" } else { " " }, name); |
| 122 | + } |
| 123 | +} |
| 124 | + |
| 125 | +fn matrix_cell_summary(p: SimdProfile) { |
| 126 | + // Lifted from `td-simd-cpu-dispatch-matrix.md` § "Master matrix" |
| 127 | + // for each x86 profile. The summary is intentionally terse — the |
| 128 | + // matrix doc is the source of truth and should be consulted before |
| 129 | + // promoting any DOC cell to TEST. |
| 130 | + let summary: &[&str] = match p { |
| 131 | + SimdProfile::GraniteRapids => &[ |
| 132 | + "F+CD+VL+DQ+BW+IFMA+VBMI+VBMI2+VNNI+BF16+FP16", |
| 133 | + "VPOPCNTDQ+BITALG+GFNI+VAES+VPCLMUL", |
| 134 | + "AMX-TILE+INT8+BF16+FP16 (FP16 is the GNR discriminator)", |
| 135 | + ], |
| 136 | + SimdProfile::SapphireRapids => &[ |
| 137 | + "F+CD+VL+DQ+BW+IFMA+VBMI+VBMI2+VNNI+BF16+FP16", |
| 138 | + "VPOPCNTDQ+BITALG+GFNI+VAES+VPCLMUL", |
| 139 | + "AMX-TILE+INT8+BF16 (no AMX-FP16 — that's GNR)", |
| 140 | + ], |
| 141 | + SimdProfile::Zen4Avx512 => &[ |
| 142 | + "F+CD+VL+DQ+BW+IFMA+VBMI+VBMI2+VNNI+BF16+FP16", |
| 143 | + "No AMX of any kind; 256-bit FPU double-pumped on Zen4, native 512-bit on Zen5", |
| 144 | + ], |
| 145 | + SimdProfile::CooperLake => &[ |
| 146 | + "F+CD+VL+DQ+BW+VNNI+BF16", |
| 147 | + "No VBMI, no FP16, no AMX — unique 'BF16 without VBMI'", |
| 148 | + ], |
| 149 | + SimdProfile::TigerLakeU => &[ |
| 150 | + "F+CD+VL+DQ+BW+IFMA+VBMI+VBMI2+VNNI+VP2INTERSECT", |
| 151 | + "VP2INTERSECT is the sole discriminator vs IceLakeSp", |
| 152 | + ], |
| 153 | + SimdProfile::IceLakeSp => &[ |
| 154 | + "F+CD+VL+DQ+BW+IFMA+VBMI+VBMI2+VNNI", |
| 155 | + "No BF16, no FP16, no AMX, no VP2INTERSECT", |
| 156 | + ], |
| 157 | + SimdProfile::CascadeLake => &["F+CD+VL+DQ+BW+VNNI", "First Xeon with VNNI; no VBMI/BF16/FP16/AMX"], |
| 158 | + SimdProfile::SkylakeX => &["F+CD+VL+DQ+BW", "Founding AVX-512 baseline; everything since adds on top"], |
| 159 | + SimdProfile::ArrowLake => &[ |
| 160 | + "No AVX-512 (hybrid CPU design)", |
| 161 | + "AVX-VNNI-INT8 + AVX-IFMA + AVX-NE-CONVERT (256-bit / VEX forms)", |
| 162 | + ], |
| 163 | + SimdProfile::HaswellAvx2 => &["AVX2 + FMA + F16C + BMI1/2", "Haswell..Coffee Lake / Zen 1-3"], |
| 164 | + SimdProfile::A76DotProd => &[ |
| 165 | + "NEON + dotprod + fp16 + bf16+ + i8mm", |
| 166 | + "Pi 5 (BCM2712), Orange Pi 5 (RK3588), Apple M1+", |
| 167 | + ], |
| 168 | + SimdProfile::A72Fast => &[ |
| 169 | + "NEON 128-bit + crypto (AES/SHA-2/CRC32)", |
| 170 | + "Pi 4 (BCM2711), Pi 3-with-crypto, Orange Pi 4 — HWCAP cannot distinguish A72 from A53-with-crypto", |
| 171 | + ], |
| 172 | + SimdProfile::A53Baseline => &[ |
| 173 | + "NEON 128-bit baseline", |
| 174 | + "Rare in the wild — QEMU / minimal aarch64 without crypto", |
| 175 | + ], |
| 176 | + SimdProfile::Scalar => &["No SIMD ISA recognised", "Fallback: scalar reference kernels"], |
| 177 | + }; |
| 178 | + for line in summary { |
| 179 | + println!(" - {}", line); |
| 180 | + } |
| 181 | +} |
0 commit comments