rsqrt.cc

#include <cstdint>
#include <limits>
#include <iostream>
#include <immintrin.h>
#include <openssl/sha.h>

#pragma GCC target "avx2"

int main()
{
    constexpr std::uint32_t min_value = 0x00800000;
    constexpr std::uint32_t stop_value = 0x7F800000;
    auto value = _mm256_castsi256_ps(_mm256_setr_epi32(
        min_value, min_value + 1, min_value + 2, min_value + 3,
        min_value + 4, min_value + 5, min_value + 7, min_value + 8));
    const auto increment = _mm256_set1_epi32(8);
    const auto final_value = _mm256_set1_epi32(stop_value);
    const auto test_mask = _mm256_set1_epi8(-1);
    SHA512_CTX sha_ctx;
    SHA384_Init(&sha_ctx);
    do {
        alignas(32) float buffer[32];
        auto it = buffer;
        _mm256_store_ps(it, _mm256_rsqrt_ps(value));
        it += 8;
        value = _mm256_castsi256_ps(
            _mm256_add_epi32(_mm256_castps_si256(value), increment));
        _mm256_store_ps(it, _mm256_rsqrt_ps(value));
        it += 8;
        value = _mm256_castsi256_ps(
            _mm256_add_epi32(_mm256_castps_si256(value), increment));
        _mm256_store_ps(it, _mm256_rsqrt_ps(value));
        it += 8;
        value = _mm256_castsi256_ps(
            _mm256_add_epi32(_mm256_castps_si256(value), increment));
        _mm256_store_ps(it, _mm256_rsqrt_ps(value));
        it += 8;
        value = _mm256_castsi256_ps(
            _mm256_add_epi32(_mm256_castps_si256(value), increment));
        SHA384_Update(&sha_ctx, buffer, sizeof(buffer));
    } while(_mm256_testz_si256(
        _mm256_cmpeq_epi32(_mm256_castps_si256(value), final_value),
        test_mask));
    unsigned char output[SHA384_DIGEST_LENGTH];
    SHA384_Final(output, &sha_ctx);
    std::cout.setf(std::ios_base::hex, std::ios_base::basefield);
    std::cout.fill('0');
    for(const auto& ch: output) {
        std::cout.width(2);
        std::cout << static_cast<unsigned>(ch);
    }
}