honeycrisp/acpu/bench/baseline.toml

toml
ฯ€ 0.0% 463 words ยท 3 min 
# Benchmark baseline โ€” M1Pro (8P+2E), 2026-04-04
#
# Purpose: track expected runtimes, detect broken benchmarks,
# set time budgets. If a bench exceeds max_seconds, it's a bug.
#
# Run all: for f in bench/*.rs; do cargo run --release --example "bench_$(basename $f .rs)" 2>&1; done
# (skip common.rs and baseline.toml)

[meta]
chip = "M1Pro"
date = "2026-04-04"
total_time_seconds = 45  # all benches combined should finish in ~45s
                         # (excluding sgemm which takes ~90s due to Apple first-pass)

[bench.memory]
file = "bench/memory.rs"
max_seconds = 10
expected_seconds = 5
notes = "pointer chase + STREAM + prefetch"

[bench.cores]
file = "bench/cores.rs"
max_seconds = 15
expected_seconds = 8
notes = "P/E core pinning + cross-core latency (ping-pong 50K rounds)"

[bench.elementwise]
file = "bench/elementwise.rs"
max_seconds = 30
expected_seconds = 15
comparison = "Apple Accelerate"
notes = "13 ops vs vvexpf/vDSP/cblas"

[bench.sgemm]
file = "bench/sgemm.rs"
max_seconds = 120
expected_seconds = 90
comparison = "Apple cblas_sgemm"
notes = "Apple runs ALL sizes first (avoids deadlock). 500 iters for small sizes."
known_issue = "takes long because Apple Accelerate 4096x4096 at 8 iters is slow"

[bench.integer]
file = "bench/integer.rs"
max_seconds = 20
expected_seconds = 10
comparison = "scalar Rust (auto-vectorized)"
notes = "NEON vs compiler. fused ops (SAD, SSD) show real wins."

[bench.numeric]
file = "bench/numeric.rs"
max_seconds = 10
expected_seconds = 5
notes = "complex_mul_acc, bf16/i8 conversions, RoPE"

[bench.amx]
file = "bench/amx.rs"
max_seconds = 10
expected_seconds = 3
notes = "AMX instruction bandwidth + compute throughput"

[bench.zk]
file = "bench/zk.rs"
max_seconds = 60
expected_seconds = 40
notes = "Goldilocks field, Poseidon2, NTT 2^16, Merkle 4096 leaves"

[bench.crypto]
file = "bench/crypto.rs"
max_seconds = 15
expected_seconds = 8
notes = "AES-128, SHA-256, PMULL via ARM crypto extensions"

[bench.render]
file = "bench/render.rs"
max_seconds = 10
expected_seconds = 5
comparison = "scalar Rust"
notes = "rsqrt/recip via FRSQRTE/FRECPE, clamp, lerp"

[bench.ai]
file = "bench/ai.rs"
max_seconds = 30
expected_seconds = 15
notes = "inference pipeline: sgemm 4K, quant, activations, softmax"

# โ”€โ”€ key results (reference numbers, not hard requirements) โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€

[results.elementwise]
exp_ns = 1458
log_ns = 1250
tanh_ns = 2250
sigmoid_ns = 2000
gelu_ns = 3354
silu_ns = 3000
softmax_ns = 4292
normalize_ns = 1375
note = "vs Apple Accelerate: 10 wins, 3 ties, 0 losses"

[results.sgemm]
sgemm_32_gf = 524
sgemm_256_gf = 1200
sgemm_1024_gf = 2260
sgemm_4096_gf = 1470
note = "vs Apple cblas_sgemm"

[results.zk]
field_mul_batch_ns_per = 1.1
field_inv_ns = 297
poseidon2_perm_ns = 5500
poseidon2_hashes_per_sec = 182000
merkle_4096_ns = 24000000
ntt_65536_fwd_ns = 2800000
note = "Goldilocks p = 2^64 - 2^32 + 1"

[results.crypto]
aes_128_gb_s = 10.0
sha256_gb_s = 1.7
pmull_mops = 1700

[results.render]
rsqrt_ns = 916
recip_ns = 792
clamp_ns = 625
lerp_ns = 333
note = "vs scalar Rust, NOT vs Apple"

[results.integer]
dot_i8_ns = 83
sad_u8_ns = 125
scale_acc_i16_ns = 500
note = "vs scalar Rust, NOT vs Apple"

Neighbours