Upload kernels/noesa24_avx2.cpp with huggingface_hub
Browse files- kernels/noesa24_avx2.cpp +77 -0
kernels/noesa24_avx2.cpp
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
// noesa24_avx2.cpp — IKNN-Rl1-A1 — NoeSA-24 AVX2 Kernel (Ryzen5 5650U)
|
| 2 |
+
// Version: v1.0
|
| 3 |
+
// Created: 2026-09-03T17:10:00+07:00
|
| 4 |
+
// Last Updated: 2026-09-03T17:10:00+07:00
|
| 5 |
+
// Status: PUBLISHABLE — EN ONLY — M1 Kernel Validation
|
| 6 |
+
// Repo: IKNN-Rl1-A1 — Integrated Knowledge-phase Neural Network — Recursive Language Iteration 1 — Architecture 1
|
| 7 |
+
// Hardware: Ryzen5 5650U — 6C/12T Zen3, AVX2 (NO AVX-512), DDR4 35-45GB/s
|
| 8 |
+
// Description: Same as AVX-512 but AVX2 version — LUT576 decode 12 cycles per 13 units
|
| 9 |
+
|
| 10 |
+
#include "noesa24_common.h"
|
| 11 |
+
#include <cstdint>
|
| 12 |
+
#include <immintrin.h>
|
| 13 |
+
|
| 14 |
+
namespace iknn {
|
| 15 |
+
namespace noesa24 {
|
| 16 |
+
namespace avx2 {
|
| 17 |
+
|
| 18 |
+
// AVX2 version — processes 8 activations at once (256-bit / 32-bit = 8)
|
| 19 |
+
inline __m256i compute_noesa24_block_avx2(
|
| 20 |
+
__m256i activations, // 8x INT32
|
| 21 |
+
const uint8_t states[8],
|
| 22 |
+
int32_t bias_0b = 1)
|
| 23 |
+
{
|
| 24 |
+
alignas(32) int32_t act_array[8];
|
| 25 |
+
_mm256_store_si256(reinterpret_cast<__m256i*>(act_array), activations);
|
| 26 |
+
|
| 27 |
+
alignas(32) int32_t result[8];
|
| 28 |
+
for (int i = 0; i < 8; ++i) {
|
| 29 |
+
result[i] = compute_noesa24_single(static_cast<int8_t>(act_array[i]), states[i], bias_0b);
|
| 30 |
+
}
|
| 31 |
+
return _mm256_load_si256(reinterpret_cast<const __m256i*>(result));
|
| 32 |
+
}
|
| 33 |
+
|
| 34 |
+
inline void decode_60bit_avx2(uint64_t packed60, uint8_t out_digits[13]) {
|
| 35 |
+
static LUT576 lut[576];
|
| 36 |
+
static bool lut_built = false;
|
| 37 |
+
if (!lut_built) {
|
| 38 |
+
build_lut576(lut);
|
| 39 |
+
lut_built = true;
|
| 40 |
+
}
|
| 41 |
+
unpack_13x24_lut(packed60, out_digits, lut);
|
| 42 |
+
}
|
| 43 |
+
|
| 44 |
+
} // namespace avx2
|
| 45 |
+
} // namespace noesa24
|
| 46 |
+
} // namespace iknn
|
| 47 |
+
|
| 48 |
+
#ifdef NOESA24_AVX2_TEST
|
| 49 |
+
#include <random>
|
| 50 |
+
#include <iostream>
|
| 51 |
+
|
| 52 |
+
int main() {
|
| 53 |
+
using namespace iknn::noesa24;
|
| 54 |
+
using namespace iknn::noesa24::avx2;
|
| 55 |
+
|
| 56 |
+
std::cout << "[NoeSA-24 AVX2 Test] Ryzen5 5650U" << std::endl;
|
| 57 |
+
|
| 58 |
+
std::mt19937 rng(42);
|
| 59 |
+
uint8_t digits[13];
|
| 60 |
+
for (int i = 0; i < 13; ++i) digits[i] = rng() % 24;
|
| 61 |
+
|
| 62 |
+
uint64_t packed = pack_13x24(digits);
|
| 63 |
+
uint8_t unpacked[13];
|
| 64 |
+
unpack_13x24(packed, unpacked);
|
| 65 |
+
|
| 66 |
+
bool ok = true;
|
| 67 |
+
for (int i = 0; i < 13; ++i) if (digits[i] != unpacked[i]) ok = false;
|
| 68 |
+
std::cout << "Pack/unpack: " << (ok ? "[PASS]" : "[FAIL]") << std::endl;
|
| 69 |
+
|
| 70 |
+
int8_t act = 10;
|
| 71 |
+
uint8_t state = encode_state(Operator::POS_A, Scale::S2);
|
| 72 |
+
int32_t res = compute_noesa24_single(act, state);
|
| 73 |
+
std::cout << "Compute +1a S2: " << res << " expected 40 " << (res==40 ? "[PASS]" : "[FAIL]") << std::endl;
|
| 74 |
+
|
| 75 |
+
return 0;
|
| 76 |
+
}
|
| 77 |
+
#endif
|