File size: 3,698 Bytes
3ac63cb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 | // benchmark_satu1.cpp — IKNN-Rl1-A1 — Combined benchmark SatU1 AVX2 vs AVX-512
// Version: v1.0
// Created: 2026-09-03T17:00:00+07:00
// Status: PUBLISHABLE — EN ONLY
// Repo: IKNN-Rl1-A1
#include "satu1_common.h"
#include <iostream>
#include <random>
#include <chrono>
#include <cmath>
// Declare functions from other translation units (we include via header but need extern)
// Instead include cpp directly for simplicity in this benchmark
#include "satu1_avx2.cpp"
#include "satu1_avx512.cpp"
int main() {
using namespace iknn::satu1;
std::cout << "=== IKNN-Rl1-A1 — SatU1 Kernel Benchmark — M1 ===" << std::endl;
std::cout << "Repo: IKNN-Rl1-A1 — Integrated Knowledge-phase Neural Network — Recursive Language Iteration 1 — Architecture 1" << std::endl;
std::cout << "Hardware: Xeon AVX-512 (VPOPCNTDQ, VBMI2) + Ryzen5 5650U AVX2 reference" << std::endl;
std::cout << "Timestamp: " << __DATE__ << " " << __TIME__ << std::endl;
std::mt19937_64 rng(12345);
const int WORDS = 8; // 512 weights
uint64_t act[WORDS], w[WORDS];
for (int i = 0; i < WORDS; ++i) {
act[i] = rng();
w[i] = rng();
}
float alpha = 0.75f;
// Correctness
float naive = compute_satu1_block_naive(act, w, alpha, WORDS);
float avx2 = avx2::compute_satu1_block_avx2(act, w, alpha, WORDS);
float avx512 = avx512::compute_satu1_block_avx512(act, w, alpha, WORDS);
std::cout << "[Correctness] Naive: " << naive << " AVX2: " << avx2 << " AVX-512: " << avx512 << std::endl;
bool pass = (std::abs(naive-avx2) < 1e-3f) && (std::abs(naive-avx512) < 1e-3f);
std::cout << (pass ? "[PASS] All kernels match naive" : "[FAIL] Mismatch") << std::endl;
if (!pass) return 1;
// Benchmark AVX2
{
const int ITERS = 2000000;
auto start = std::chrono::high_resolution_clock::now();
float sum = 0;
for (int it = 0; it < ITERS; ++it) {
act[0] ^= it;
sum += avx2::compute_satu1_block_avx2(act, w, alpha, WORDS);
}
auto end = std::chrono::high_resolution_clock::now();
double ms = std::chrono::duration_cast<std::chrono::milliseconds>(end-start).count();
double giga = (double)ITERS * WORDS * 64 / 1e9;
std::cout << "[BENCH AVX2] Iters " << ITERS << " Time " << ms << "ms Giga-popcnt/s " << giga/(ms/1000.0) << " sum " << sum << std::endl;
}
// Benchmark AVX-512
{
const int ITERS = 2000000;
auto start = std::chrono::high_resolution_clock::now();
float sum = 0;
for (int it = 0; it < ITERS; ++it) {
act[0] ^= it;
sum += avx512::compute_satu1_block_avx512(act, w, alpha, WORDS);
}
auto end = std::chrono::high_resolution_clock::now();
double ms = std::chrono::duration_cast<std::chrono::milliseconds>(end-start).count();
double giga = (double)ITERS * WORDS * 64 / 1e9;
double tps_equiv = ITERS / (ms/1000.0);
std::cout << "[BENCH AVX-512] Iters " << ITERS << " Time " << ms << "ms Giga-popcnt/s " << giga/(ms/1000.0) << " TPS equiv " << tps_equiv << " sum " << sum << std::endl;
}
// Memory bandwidth estimation
// Per token: 0.70GB read (SatU1 2.66GB*16% + NoeSA 0.97GB*20%)
// DDR4 38GB/s -> 54 TPS upper, DDR5 70GB/s -> 100 TPS upper
std::cout << "[BANDWIDTH] Estimated per token read 0.70GB" << std::endl;
std::cout << "[BANDWIDTH] Ryzen5 DDR4 38GB/s -> 54 TPS upper bound" << std::endl;
std::cout << "[BANDWIDTH] Xeon DDR5 70GB/s -> 100 TPS upper bound" << std::endl;
std::cout << "[TARGET] Ryzen5 28-42 / 60-85 TPS, AVX-512 65-90 / 120-165 TPS" << std::endl;
return 0;
}
|