Nat512.cs 14 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363
  1. #if !BESTHTTP_DISABLE_ALTERNATE_SSL && (!UNITY_WEBGL || UNITY_EDITOR)
  2. #pragma warning disable
  3. using System;
  4. #if NETCOREAPP3_0_OR_GREATER
  5. using System.Runtime.CompilerServices;
  6. using System.Runtime.InteropServices;
  7. using System.Runtime.Intrinsics;
  8. using System.Runtime.Intrinsics.X86;
  9. #endif
  10. namespace Best.HTTP.SecureProtocol.Org.BouncyCastle.Math.Raw
  11. {
  12. internal abstract class Nat512
  13. {
  14. public static void Mul(uint[] x, uint[] y, uint[] zz)
  15. {
  16. Nat256.Mul(x, y, zz);
  17. Nat256.Mul(x, 8, y, 8, zz, 16);
  18. uint c24 = Nat256.AddToEachOther(zz, 8, zz, 16);
  19. uint c16 = c24 + Nat256.AddTo(zz, 0, zz, 8, 0);
  20. c24 += Nat256.AddTo(zz, 24, zz, 16, c16);
  21. uint[] dx = Nat256.Create(), dy = Nat256.Create();
  22. bool neg = Nat256.Diff(x, 8, x, 0, dx, 0) != Nat256.Diff(y, 8, y, 0, dy, 0);
  23. uint[] tt = Nat256.CreateExt();
  24. Nat256.Mul(dx, dy, tt);
  25. c24 += neg ? Nat.AddTo(16, tt, 0, zz, 8) : (uint)Nat.SubFrom(16, tt, 0, zz, 8);
  26. Nat.AddWordAt(32, c24, zz, 24);
  27. }
  28. public static void Square(uint[] x, uint[] zz)
  29. {
  30. Nat256.Square(x, zz);
  31. Nat256.Square(x, 8, zz, 16);
  32. uint c24 = Nat256.AddToEachOther(zz, 8, zz, 16);
  33. uint c16 = c24 + Nat256.AddTo(zz, 0, zz, 8, 0);
  34. c24 += Nat256.AddTo(zz, 24, zz, 16, c16);
  35. uint[] dx = Nat256.Create();
  36. Nat256.Diff(x, 8, x, 0, dx, 0);
  37. uint[] m = Nat256.CreateExt();
  38. Nat256.Square(dx, m);
  39. c24 += (uint)Nat.SubFrom(16, m, 0, zz, 8);
  40. Nat.AddWordAt(32, c24, zz, 24);
  41. }
  42. public static void Xor(uint[] x, int xOff, uint[] y, int yOff, uint[] z, int zOff)
  43. {
  44. #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER || UNITY_2021_2_OR_NEWER
  45. Xor(x.AsSpan(xOff), y.AsSpan(yOff), z.AsSpan(zOff));
  46. #else
  47. for (int i = 0; i < 16; i += 4)
  48. {
  49. z[zOff + i + 0] = x[xOff + i + 0] ^ y[yOff + i + 0];
  50. z[zOff + i + 1] = x[xOff + i + 1] ^ y[yOff + i + 1];
  51. z[zOff + i + 2] = x[xOff + i + 2] ^ y[yOff + i + 2];
  52. z[zOff + i + 3] = x[xOff + i + 3] ^ y[yOff + i + 3];
  53. }
  54. #endif
  55. }
  56. #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER || UNITY_2021_2_OR_NEWER
  57. public static void Xor(ReadOnlySpan<uint> x, ReadOnlySpan<uint> y, Span<uint> z)
  58. {
  59. #if NETCOREAPP3_0_OR_GREATER
  60. if (Avx2.IsSupported && Unsafe.SizeOf<Vector256<byte>>() == 32)
  61. {
  62. var X = MemoryMarshal.AsBytes(x[..16]);
  63. var Y = MemoryMarshal.AsBytes(y[..16]);
  64. var Z = MemoryMarshal.AsBytes(z[..16]);
  65. var X0 = MemoryMarshal.Read<Vector256<byte>>(X[0x00..0x20]);
  66. var X1 = MemoryMarshal.Read<Vector256<byte>>(X[0x20..0x40]);
  67. var Y0 = MemoryMarshal.Read<Vector256<byte>>(Y[0x00..0x20]);
  68. var Y1 = MemoryMarshal.Read<Vector256<byte>>(Y[0x20..0x40]);
  69. var Z0 = Avx2.Xor(X0, Y0);
  70. var Z1 = Avx2.Xor(X1, Y1);
  71. MemoryMarshal.Write(Z[0x00..0x20], ref Z0);
  72. MemoryMarshal.Write(Z[0x20..0x40], ref Z1);
  73. return;
  74. }
  75. if (Sse2.IsSupported && Unsafe.SizeOf<Vector128<byte>>() == 16)
  76. {
  77. var X = MemoryMarshal.AsBytes(x[..16]);
  78. var Y = MemoryMarshal.AsBytes(y[..16]);
  79. var Z = MemoryMarshal.AsBytes(z[..16]);
  80. var X0 = MemoryMarshal.Read<Vector128<byte>>(X[0x00..0x10]);
  81. var X1 = MemoryMarshal.Read<Vector128<byte>>(X[0x10..0x20]);
  82. var X2 = MemoryMarshal.Read<Vector128<byte>>(X[0x20..0x30]);
  83. var X3 = MemoryMarshal.Read<Vector128<byte>>(X[0x30..0x40]);
  84. var Y0 = MemoryMarshal.Read<Vector128<byte>>(Y[0x00..0x10]);
  85. var Y1 = MemoryMarshal.Read<Vector128<byte>>(Y[0x10..0x20]);
  86. var Y2 = MemoryMarshal.Read<Vector128<byte>>(Y[0x20..0x30]);
  87. var Y3 = MemoryMarshal.Read<Vector128<byte>>(Y[0x30..0x40]);
  88. var Z0 = Sse2.Xor(X0, Y0);
  89. var Z1 = Sse2.Xor(X1, Y1);
  90. var Z2 = Sse2.Xor(X2, Y2);
  91. var Z3 = Sse2.Xor(X3, Y3);
  92. MemoryMarshal.Write(Z[0x00..0x10], ref Z0);
  93. MemoryMarshal.Write(Z[0x10..0x20], ref Z1);
  94. MemoryMarshal.Write(Z[0x20..0x30], ref Z2);
  95. MemoryMarshal.Write(Z[0x30..0x40], ref Z3);
  96. return;
  97. }
  98. #endif
  99. for (int i = 0; i < 16; i += 4)
  100. {
  101. z[i + 0] = x[i + 0] ^ y[i + 0];
  102. z[i + 1] = x[i + 1] ^ y[i + 1];
  103. z[i + 2] = x[i + 2] ^ y[i + 2];
  104. z[i + 3] = x[i + 3] ^ y[i + 3];
  105. }
  106. }
  107. #endif
  108. public static void XorTo(uint[] x, int xOff, uint[] z, int zOff)
  109. {
  110. #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER || UNITY_2021_2_OR_NEWER
  111. XorTo(x.AsSpan(xOff), z.AsSpan(zOff));
  112. #else
  113. for (int i = 0; i < 16; i += 4)
  114. {
  115. z[zOff + i + 0] ^= x[xOff + i + 0];
  116. z[zOff + i + 1] ^= x[xOff + i + 1];
  117. z[zOff + i + 2] ^= x[xOff + i + 2];
  118. z[zOff + i + 3] ^= x[xOff + i + 3];
  119. }
  120. #endif
  121. }
  122. #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER || UNITY_2021_2_OR_NEWER
  123. public static void XorTo(ReadOnlySpan<uint> x, Span<uint> z)
  124. {
  125. #if NETCOREAPP3_0_OR_GREATER
  126. if (Avx2.IsSupported && Unsafe.SizeOf<Vector256<byte>>() == 32)
  127. {
  128. var X = MemoryMarshal.AsBytes(x[..16]);
  129. var Z = MemoryMarshal.AsBytes(z[..16]);
  130. var X0 = MemoryMarshal.Read<Vector256<byte>>(X[0x00..0x20]);
  131. var X1 = MemoryMarshal.Read<Vector256<byte>>(X[0x20..0x40]);
  132. var Y0 = MemoryMarshal.Read<Vector256<byte>>(Z[0x00..0x20]);
  133. var Y1 = MemoryMarshal.Read<Vector256<byte>>(Z[0x20..0x40]);
  134. var Z0 = Avx2.Xor(X0, Y0);
  135. var Z1 = Avx2.Xor(X1, Y1);
  136. MemoryMarshal.Write(Z[0x00..0x20], ref Z0);
  137. MemoryMarshal.Write(Z[0x20..0x40], ref Z1);
  138. return;
  139. }
  140. if (Sse2.IsSupported && Unsafe.SizeOf<Vector128<byte>>() == 16)
  141. {
  142. var X = MemoryMarshal.AsBytes(x[..16]);
  143. var Z = MemoryMarshal.AsBytes(z[..16]);
  144. var X0 = MemoryMarshal.Read<Vector128<byte>>(X[0x00..0x10]);
  145. var X1 = MemoryMarshal.Read<Vector128<byte>>(X[0x10..0x20]);
  146. var X2 = MemoryMarshal.Read<Vector128<byte>>(X[0x20..0x30]);
  147. var X3 = MemoryMarshal.Read<Vector128<byte>>(X[0x30..0x40]);
  148. var Y0 = MemoryMarshal.Read<Vector128<byte>>(Z[0x00..0x10]);
  149. var Y1 = MemoryMarshal.Read<Vector128<byte>>(Z[0x10..0x20]);
  150. var Y2 = MemoryMarshal.Read<Vector128<byte>>(Z[0x20..0x30]);
  151. var Y3 = MemoryMarshal.Read<Vector128<byte>>(Z[0x30..0x40]);
  152. var Z0 = Sse2.Xor(X0, Y0);
  153. var Z1 = Sse2.Xor(X1, Y1);
  154. var Z2 = Sse2.Xor(X2, Y2);
  155. var Z3 = Sse2.Xor(X3, Y3);
  156. MemoryMarshal.Write(Z[0x00..0x10], ref Z0);
  157. MemoryMarshal.Write(Z[0x10..0x20], ref Z1);
  158. MemoryMarshal.Write(Z[0x20..0x30], ref Z2);
  159. MemoryMarshal.Write(Z[0x30..0x40], ref Z3);
  160. return;
  161. }
  162. #endif
  163. for (int i = 0; i < 16; i += 4)
  164. {
  165. z[i + 0] ^= x[i + 0];
  166. z[i + 1] ^= x[i + 1];
  167. z[i + 2] ^= x[i + 2];
  168. z[i + 3] ^= x[i + 3];
  169. }
  170. }
  171. #endif
  172. public static void Xor64(ulong[] x, int xOff, ulong[] y, int yOff, ulong[] z, int zOff)
  173. {
  174. #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER || UNITY_2021_2_OR_NEWER
  175. Xor64(x.AsSpan(xOff), y.AsSpan(yOff), z.AsSpan(zOff));
  176. #else
  177. for (int i = 0; i < 8; i += 4)
  178. {
  179. z[zOff + i + 0] = x[xOff + i + 0] ^ y[yOff + i + 0];
  180. z[zOff + i + 1] = x[xOff + i + 1] ^ y[yOff + i + 1];
  181. z[zOff + i + 2] = x[xOff + i + 2] ^ y[yOff + i + 2];
  182. z[zOff + i + 3] = x[xOff + i + 3] ^ y[yOff + i + 3];
  183. }
  184. #endif
  185. }
  186. #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER || UNITY_2021_2_OR_NEWER
  187. public static void Xor64(ReadOnlySpan<ulong> x, ReadOnlySpan<ulong> y, Span<ulong> z)
  188. {
  189. #if NETCOREAPP3_0_OR_GREATER
  190. if (Avx2.IsSupported && Unsafe.SizeOf<Vector256<byte>>() == 32)
  191. {
  192. var X = MemoryMarshal.AsBytes(x[..8]);
  193. var Y = MemoryMarshal.AsBytes(y[..8]);
  194. var Z = MemoryMarshal.AsBytes(z[..8]);
  195. var X0 = MemoryMarshal.Read<Vector256<byte>>(X[0x00..0x20]);
  196. var X1 = MemoryMarshal.Read<Vector256<byte>>(X[0x20..0x40]);
  197. var Y0 = MemoryMarshal.Read<Vector256<byte>>(Y[0x00..0x20]);
  198. var Y1 = MemoryMarshal.Read<Vector256<byte>>(Y[0x20..0x40]);
  199. var Z0 = Avx2.Xor(X0, Y0);
  200. var Z1 = Avx2.Xor(X1, Y1);
  201. MemoryMarshal.Write(Z[0x00..0x20], ref Z0);
  202. MemoryMarshal.Write(Z[0x20..0x40], ref Z1);
  203. return;
  204. }
  205. if (Sse2.IsSupported && Unsafe.SizeOf<Vector128<byte>>() == 16)
  206. {
  207. var X = MemoryMarshal.AsBytes(x[..8]);
  208. var Y = MemoryMarshal.AsBytes(y[..8]);
  209. var Z = MemoryMarshal.AsBytes(z[..8]);
  210. var X0 = MemoryMarshal.Read<Vector128<byte>>(X[0x00..0x10]);
  211. var X1 = MemoryMarshal.Read<Vector128<byte>>(X[0x10..0x20]);
  212. var X2 = MemoryMarshal.Read<Vector128<byte>>(X[0x20..0x30]);
  213. var X3 = MemoryMarshal.Read<Vector128<byte>>(X[0x30..0x40]);
  214. var Y0 = MemoryMarshal.Read<Vector128<byte>>(Y[0x00..0x10]);
  215. var Y1 = MemoryMarshal.Read<Vector128<byte>>(Y[0x10..0x20]);
  216. var Y2 = MemoryMarshal.Read<Vector128<byte>>(Y[0x20..0x30]);
  217. var Y3 = MemoryMarshal.Read<Vector128<byte>>(Y[0x30..0x40]);
  218. var Z0 = Sse2.Xor(X0, Y0);
  219. var Z1 = Sse2.Xor(X1, Y1);
  220. var Z2 = Sse2.Xor(X2, Y2);
  221. var Z3 = Sse2.Xor(X3, Y3);
  222. MemoryMarshal.Write(Z[0x00..0x10], ref Z0);
  223. MemoryMarshal.Write(Z[0x10..0x20], ref Z1);
  224. MemoryMarshal.Write(Z[0x20..0x30], ref Z2);
  225. MemoryMarshal.Write(Z[0x30..0x40], ref Z3);
  226. return;
  227. }
  228. #endif
  229. for (int i = 0; i < 8; i += 4)
  230. {
  231. z[i + 0] = x[i + 0] ^ y[i + 0];
  232. z[i + 1] = x[i + 1] ^ y[i + 1];
  233. z[i + 2] = x[i + 2] ^ y[i + 2];
  234. z[i + 3] = x[i + 3] ^ y[i + 3];
  235. }
  236. }
  237. #endif
  238. public static void XorTo64(ulong[] x, int xOff, ulong[] z, int zOff)
  239. {
  240. #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER || UNITY_2021_2_OR_NEWER
  241. XorTo64(x.AsSpan(xOff), z.AsSpan(zOff));
  242. #else
  243. for (int i = 0; i < 8; i += 4)
  244. {
  245. z[zOff + i + 0] ^= x[xOff + i + 0];
  246. z[zOff + i + 1] ^= x[xOff + i + 1];
  247. z[zOff + i + 2] ^= x[xOff + i + 2];
  248. z[zOff + i + 3] ^= x[xOff + i + 3];
  249. }
  250. #endif
  251. }
  252. #if NETCOREAPP2_1_OR_GREATER || NETSTANDARD2_1_OR_GREATER || UNITY_2021_2_OR_NEWER
  253. public static void XorTo64(ReadOnlySpan<ulong> x, Span<ulong> z)
  254. {
  255. #if NETCOREAPP3_0_OR_GREATER
  256. if (Avx2.IsSupported && Unsafe.SizeOf<Vector256<byte>>() == 32)
  257. {
  258. var X = MemoryMarshal.AsBytes(x[..8]);
  259. var Z = MemoryMarshal.AsBytes(z[..8]);
  260. var X0 = MemoryMarshal.Read<Vector256<byte>>(X[0x00..0x20]);
  261. var X1 = MemoryMarshal.Read<Vector256<byte>>(X[0x20..0x40]);
  262. var Y0 = MemoryMarshal.Read<Vector256<byte>>(Z[0x00..0x20]);
  263. var Y1 = MemoryMarshal.Read<Vector256<byte>>(Z[0x20..0x40]);
  264. var Z0 = Avx2.Xor(X0, Y0);
  265. var Z1 = Avx2.Xor(X1, Y1);
  266. MemoryMarshal.Write(Z[0x00..0x20], ref Z0);
  267. MemoryMarshal.Write(Z[0x20..0x40], ref Z1);
  268. return;
  269. }
  270. if (Sse2.IsSupported && Unsafe.SizeOf<Vector128<byte>>() == 16)
  271. {
  272. var X = MemoryMarshal.AsBytes(x[..8]);
  273. var Z = MemoryMarshal.AsBytes(z[..8]);
  274. var X0 = MemoryMarshal.Read<Vector128<byte>>(X[0x00..0x10]);
  275. var X1 = MemoryMarshal.Read<Vector128<byte>>(X[0x10..0x20]);
  276. var X2 = MemoryMarshal.Read<Vector128<byte>>(X[0x20..0x30]);
  277. var X3 = MemoryMarshal.Read<Vector128<byte>>(X[0x30..0x40]);
  278. var Y0 = MemoryMarshal.Read<Vector128<byte>>(Z[0x00..0x10]);
  279. var Y1 = MemoryMarshal.Read<Vector128<byte>>(Z[0x10..0x20]);
  280. var Y2 = MemoryMarshal.Read<Vector128<byte>>(Z[0x20..0x30]);
  281. var Y3 = MemoryMarshal.Read<Vector128<byte>>(Z[0x30..0x40]);
  282. var Z0 = Sse2.Xor(X0, Y0);
  283. var Z1 = Sse2.Xor(X1, Y1);
  284. var Z2 = Sse2.Xor(X2, Y2);
  285. var Z3 = Sse2.Xor(X3, Y3);
  286. MemoryMarshal.Write(Z[0x00..0x10], ref Z0);
  287. MemoryMarshal.Write(Z[0x10..0x20], ref Z1);
  288. MemoryMarshal.Write(Z[0x20..0x30], ref Z2);
  289. MemoryMarshal.Write(Z[0x30..0x40], ref Z3);
  290. return;
  291. }
  292. #endif
  293. for (int i = 0; i < 8; i += 4)
  294. {
  295. z[i + 0] ^= x[i + 0];
  296. z[i + 1] ^= x[i + 1];
  297. z[i + 2] ^= x[i + 2];
  298. z[i + 3] ^= x[i + 3];
  299. }
  300. }
  301. #endif
  302. }
  303. }
  304. #pragma warning restore
  305. #endif