argon2_avx2.cpp

Go to the documentation of this file.

/**
* (C) 2023 Jack Lloyd
*
* Botan is released under the Simplified BSD License (see license.txt)
*/
 
#include <botan/argon2.h>
#include <immintrin.h>
 
namespace Botan {
 
namespace {
 
class SIMD_4x64 final {
   public:
      SIMD_4x64& operator=(const SIMD_4x64& other) = default;
      SIMD_4x64(const SIMD_4x64& other) = default;
 
      SIMD_4x64& operator=(SIMD_4x64&& other) = default;
      SIMD_4x64(SIMD_4x64&& other) = default;
 
      ~SIMD_4x64() = default;
 
      // zero initialized
      BOTAN_FUNC_ISA("avx2") SIMD_4x64() { m_simd = _mm256_setzero_si256(); }
 
      // Load two halves at different addresses
      static BOTAN_FUNC_ISA("avx2") SIMD_4x64 load_le2(const void* inl, const void* inh) {
         return SIMD_4x64(
            _mm256_loadu2_m128i(reinterpret_cast<const __m128i*>(inl), reinterpret_cast<const __m128i*>(inh)));
      }
 
      static BOTAN_FUNC_ISA("avx2") SIMD_4x64 load_le(const void* in) {
         return SIMD_4x64(_mm256_loadu_si256(reinterpret_cast<const __m256i*>(in)));
      }
 
      void store_le(uint64_t out[4]) const { this->store_le(reinterpret_cast<uint8_t*>(out)); }
 
      BOTAN_FUNC_ISA("avx2") void store_le(uint8_t out[]) const {
         _mm256_storeu_si256(reinterpret_cast<__m256i*>(out), m_simd);
      }
 
      BOTAN_FUNC_ISA("avx2") void store_le2(void* outh, void* outl) {
         _mm256_storeu2_m128i(reinterpret_cast<__m128i*>(outh), reinterpret_cast<__m128i*>(outl), m_simd);
      }
 
      SIMD_4x64 operator+(const SIMD_4x64& other) const {
         SIMD_4x64 retval(*this);
         retval += other;
         return retval;
      }
 
      SIMD_4x64 operator^(const SIMD_4x64& other) const {
         SIMD_4x64 retval(*this);
         retval ^= other;
         return retval;
      }
 
      BOTAN_FUNC_ISA("avx2") void operator+=(const SIMD_4x64& other) {
         m_simd = _mm256_add_epi64(m_simd, other.m_simd);
      }
 
      BOTAN_FUNC_ISA("avx2") void operator^=(const SIMD_4x64& other) {
         m_simd = _mm256_xor_si256(m_simd, other.m_simd);
      }
 
      template <size_t ROT>
      BOTAN_FUNC_ISA("avx2")
      SIMD_4x64 rotr() const
         requires(ROT > 0 && ROT < 64)
      {
         if constexpr(ROT == 16) {
            auto shuf_rot_16 =
               _mm256_set_epi64x(0x09080f0e0d0c0b0a, 0x0100070605040302, 0x09080f0e0d0c0b0a, 0x0100070605040302);
 
            return SIMD_4x64(_mm256_shuffle_epi8(m_simd, shuf_rot_16));
         } else if constexpr(ROT == 24) {
            auto shuf_rot_24 =
               _mm256_set_epi64x(0x0a09080f0e0d0c0b, 0x0201000706050403, 0x0a09080f0e0d0c0b, 0x0201000706050403);
 
            return SIMD_4x64(_mm256_shuffle_epi8(m_simd, shuf_rot_24));
         } else if constexpr(ROT == 32) {
            auto shuf_rot_32 =
               _mm256_set_epi64x(0x0b0a09080f0e0d0c, 0x0302010007060504, 0x0b0a09080f0e0d0c, 0x0302010007060504);
 
            return SIMD_4x64(_mm256_shuffle_epi8(m_simd, shuf_rot_32));
         } else {
            return SIMD_4x64(_mm256_or_si256(_mm256_srli_epi64(m_simd, static_cast<int>(ROT)),
                                             _mm256_slli_epi64(m_simd, static_cast<int>(64 - ROT))));
         }
      }
 
      template <size_t ROT>
      SIMD_4x64 rotl() const {
         return this->rotr<64 - ROT>();
      }
 
      // Argon2 specific operation
      static BOTAN_FUNC_ISA("avx2") SIMD_4x64 mul2_32(SIMD_4x64 x, SIMD_4x64 y) {
         const __m256i m = _mm256_mul_epu32(x.m_simd, y.m_simd);
         return SIMD_4x64(_mm256_add_epi64(m, m));
      }
 
      template <uint8_t CTRL>
      static BOTAN_FUNC_ISA("avx2") SIMD_4x64 permute_4x64(SIMD_4x64 x) {
         return SIMD_4x64(_mm256_permute4x64_epi64(x.m_simd, CTRL));
      }
 
      // Argon2 specific
      static void twist(SIMD_4x64& B, SIMD_4x64& C, SIMD_4x64& D) {
         B = SIMD_4x64::permute_4x64<0b00'11'10'01>(B);
         C = SIMD_4x64::permute_4x64<0b01'00'11'10>(C);
         D = SIMD_4x64::permute_4x64<0b10'01'00'11>(D);
      }
 
      // Argon2 specific
      static void untwist(SIMD_4x64& B, SIMD_4x64& C, SIMD_4x64& D) {
         B = SIMD_4x64::permute_4x64<0b10'01'00'11>(B);
         C = SIMD_4x64::permute_4x64<0b01'00'11'10>(C);
         D = SIMD_4x64::permute_4x64<0b00'11'10'01>(D);
      }
 
      explicit BOTAN_FUNC_ISA("avx2") SIMD_4x64(__m256i x) : m_simd(x) {}
 
   private:
      __m256i m_simd;
};
 
BOTAN_FORCE_INLINE void blamka_G(SIMD_4x64& A, SIMD_4x64& B, SIMD_4x64& C, SIMD_4x64& D) {
   A += B + SIMD_4x64::mul2_32(A, B);
   D ^= A;
   D = D.rotr<32>();
 
   C += D + SIMD_4x64::mul2_32(C, D);
   B ^= C;
   B = B.rotr<24>();
 
   A += B + SIMD_4x64::mul2_32(A, B);
   D ^= A;
   D = D.rotr<16>();
 
   C += D + SIMD_4x64::mul2_32(C, D);
   B ^= C;
   B = B.rotr<63>();
}
 
BOTAN_FORCE_INLINE void blamka_R(SIMD_4x64& A, SIMD_4x64& B, SIMD_4x64& C, SIMD_4x64& D) {
   blamka_G(A, B, C, D);
 
   SIMD_4x64::twist(B, C, D);
   blamka_G(A, B, C, D);
   SIMD_4x64::untwist(B, C, D);
}
 
}  // namespace
 
BOTAN_FUNC_ISA("avx2") void Argon2::blamka_avx2(uint64_t N[128], uint64_t T[128]) {
   for(size_t i = 0; i != 8; ++i) {
      SIMD_4x64 A = SIMD_4x64::load_le(&N[16 * i + 4 * 0]);
      SIMD_4x64 B = SIMD_4x64::load_le(&N[16 * i + 4 * 1]);
      SIMD_4x64 C = SIMD_4x64::load_le(&N[16 * i + 4 * 2]);
      SIMD_4x64 D = SIMD_4x64::load_le(&N[16 * i + 4 * 3]);
 
      blamka_R(A, B, C, D);
 
      A.store_le(&T[16 * i + 4 * 0]);
      B.store_le(&T[16 * i + 4 * 1]);
      C.store_le(&T[16 * i + 4 * 2]);
      D.store_le(&T[16 * i + 4 * 3]);
   }
 
   for(size_t i = 0; i != 8; ++i) {
      SIMD_4x64 A = SIMD_4x64::load_le2(&T[2 * i + 32 * 0], &T[2 * i + 32 * 0 + 16]);
      SIMD_4x64 B = SIMD_4x64::load_le2(&T[2 * i + 32 * 1], &T[2 * i + 32 * 1 + 16]);
      SIMD_4x64 C = SIMD_4x64::load_le2(&T[2 * i + 32 * 2], &T[2 * i + 32 * 2 + 16]);
      SIMD_4x64 D = SIMD_4x64::load_le2(&T[2 * i + 32 * 3], &T[2 * i + 32 * 3 + 16]);
 
      blamka_R(A, B, C, D);
 
      A.store_le2(&T[2 * i + 32 * 0], &T[2 * i + 32 * 0 + 16]);
      B.store_le2(&T[2 * i + 32 * 1], &T[2 * i + 32 * 1 + 16]);
      C.store_le2(&T[2 * i + 32 * 2], &T[2 * i + 32 * 2 + 16]);
      D.store_le2(&T[2 * i + 32 * 3], &T[2 * i + 32 * 3 + 16]);
   }
 
   for(size_t i = 0; i != 128 / 8; ++i) {
      SIMD_4x64 n0 = SIMD_4x64::load_le(&N[8 * i]);
      SIMD_4x64 n1 = SIMD_4x64::load_le(&N[8 * i + 4]);
      SIMD_4x64 t0 = SIMD_4x64::load_le(&T[8 * i]);
      SIMD_4x64 t1 = SIMD_4x64::load_le(&T[8 * i + 4]);
 
      n0 ^= t0;
      n1 ^= t1;
      n0.store_le(&N[8 * i]);
      n1.store_le(&N[8 * i + 4]);
   }
}
 
}  // namespace Botan