donna_sse.h 3.0 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586
  1. // donna_sse.h - written and placed in public domain by Jeffrey Walton
  2. // Crypto++ specific implementation wrapped around Andrew
  3. // Moon's public domain curve25519-donna and ed25519-donna,
  4. // https://github.com/floodyberry/curve25519-donna and
  5. // https://github.com/floodyberry/ed25519-donna.
  6. // This source file multiplexes two different repos using namespaces. This
  7. // was a little easier from a project management standpoint. We only need
  8. // two files per architecture at the expense of namespaces and bloat.
  9. #ifndef CRYPTOPP_DONNA_SSE_H
  10. #define CRYPTOPP_DONNA_SSE_H
  11. #ifndef CRYPTOPP_DOXYGEN_PROCESSING
  12. #include "config.h"
  13. #include <emmintrin.h>
  14. NAMESPACE_BEGIN(CryptoPP)
  15. NAMESPACE_BEGIN(Donna)
  16. NAMESPACE_BEGIN(ArchSSE)
  17. using CryptoPP::byte;
  18. using CryptoPP::word32;
  19. typedef __m128i xmmi;
  20. #define ALIGN(n) CRYPTOPP_ALIGN_DATA(n)
  21. typedef union packedelem8_t {
  22. byte u[16];
  23. xmmi v;
  24. } packedelem8;
  25. typedef union packedelem32_t {
  26. word32 u[4];
  27. xmmi v;
  28. } packedelem32;
  29. typedef union packedelem64_t {
  30. word64 u[2];
  31. xmmi v;
  32. } packedelem64;
  33. /* 10 elements + an extra 2 to fit in 3 xmm registers */
  34. typedef word32 bignum25519[12];
  35. typedef packedelem32 packed32bignum25519[5];
  36. typedef packedelem64 packed64bignum25519[10];
  37. const word32 reduce_mask_26 = (1 << 26) - 1;
  38. const word32 reduce_mask_25 = (1 << 25) - 1;
  39. const packedelem32 sse2_bot32bitmask = {{0xffffffff, 0x00000000, 0xffffffff, 0x00000000}};
  40. const packedelem32 sse2_top32bitmask = {{0x00000000, 0xffffffff, 0x00000000, 0xffffffff}};
  41. const packedelem32 sse2_top64bitmask = {{0x00000000, 0x00000000, 0xffffffff, 0xffffffff}};
  42. const packedelem32 sse2_bot64bitmask = {{0xffffffff, 0xffffffff, 0x00000000, 0x00000000}};
  43. /* reduction masks */
  44. const packedelem64 packedmask26 = {{0x03ffffff, 0x03ffffff}};
  45. const packedelem64 packedmask25 = {{0x01ffffff, 0x01ffffff}};
  46. const packedelem32 packedmask2625 = {{0x3ffffff,0,0x1ffffff,0}};
  47. const packedelem32 packedmask26262626 = {{0x03ffffff, 0x03ffffff, 0x03ffffff, 0x03ffffff}};
  48. const packedelem32 packedmask25252525 = {{0x01ffffff, 0x01ffffff, 0x01ffffff, 0x01ffffff}};
  49. /* multipliers */
  50. const packedelem64 packednineteen = {{19, 19}};
  51. const packedelem64 packednineteenone = {{19, 1}};
  52. const packedelem64 packedthirtyeight = {{38, 38}};
  53. const packedelem64 packed3819 = {{19*2,19}};
  54. const packedelem64 packed9638 = {{19*4,19*2}};
  55. /* 121666,121665 */
  56. const packedelem64 packed121666121665 = {{121666, 121665}};
  57. /* 2*(2^255 - 19) = 0 mod p */
  58. const packedelem32 packed2p0 = {{0x7ffffda,0x3fffffe,0x7fffffe,0x3fffffe}};
  59. const packedelem32 packed2p1 = {{0x7fffffe,0x3fffffe,0x7fffffe,0x3fffffe}};
  60. const packedelem32 packed2p2 = {{0x7fffffe,0x3fffffe,0x0000000,0x0000000}};
  61. const packedelem32 packed32zeromodp0 = {{0x7ffffda,0x7ffffda,0x3fffffe,0x3fffffe}};
  62. const packedelem32 packed32zeromodp1 = {{0x7fffffe,0x7fffffe,0x3fffffe,0x3fffffe}};
  63. NAMESPACE_END // ArchSSE
  64. NAMESPACE_END // Donna
  65. NAMESPACE_END // CryptoPP
  66. #endif // CRYPTOPP_DOXYGEN_PROCESSING
  67. #endif // CRYPTOPP_DONNA_SSE_H