📄 m1_64.c
字号:
Tcz = FNMS(KP831469612, Tcy, KP555570233 * Tcv); TcL = FMA(KP831469612, Tcv, KP555570233 * Tcy); } { E Tcl, TcA, TcN, TcO; Tcl = Tcd + Tck; TcA = Tcs + Tcz; ro[WS(os, 42)] = Tcl - TcA; ro[WS(os, 10)] = Tcl + TcA; TcN = TcD + TcG; TcO = TcK + TcL; io[WS(os, 42)] = TcN - TcO; io[WS(os, 10)] = TcN + TcO; } { E TcH, TcI, TcJ, TcM; TcH = TcD - TcG; TcI = Tcz - Tcs; io[WS(os, 58)] = TcH - TcI; io[WS(os, 26)] = TcH + TcI; TcJ = Tcd - Tck; TcM = TcK - TcL; ro[WS(os, 58)] = TcJ - TcM; ro[WS(os, 26)] = TcJ + TcM; } { E TcR, TcY, Td7, Td8; TcR = TcP + TcQ; TcY = TcU + TcX; ro[WS(os, 34)] = TcR - TcY; ro[WS(os, 2)] = TcR + TcY; Td7 = TcZ + Td0; Td8 = Td4 + Td5; io[WS(os, 34)] = Td7 - Td8; io[WS(os, 2)] = Td7 + Td8; } { E Td1, Td2, Td3, Td6; Td1 = TcZ - Td0; Td2 = TcX - TcU; io[WS(os, 50)] = Td1 - Td2; io[WS(os, 18)] = Td1 + Td2; Td3 = TcP - TcQ; Td6 = Td4 - Td5; ro[WS(os, 50)] = Td3 - Td6; ro[WS(os, 18)] = Td3 + Td6; } } { E Tap, TbR, TbF, Tc1, TaE, Tc2, TbZ, Tc7, Tb6, TbM, TbI, TbS, TbW, Tc6, Tbx; E TbN, Tao, TbE; Tao = KP707106781 * (Tak - Tan); Tap = Tah - Tao; TbR = Tah + Tao; TbE = KP707106781 * (TbC - TbD); TbF = TbB - TbE; Tc1 = TbB + TbE; { E Taw, TaD, TbX, TbY; Taw = FNMS(KP923879532, Tav, KP382683432 * Tas); TaD = FMA(KP382683432, Taz, KP923879532 * TaC); TaE = Taw - TaD; Tc2 = Taw + TaD; TbX = Tbb + Tbm; TbY = Tbs + Tbv; TbZ = FNMS(KP555570233, TbY, KP831469612 * TbX); Tc7 = FMA(KP831469612, TbY, KP555570233 * TbX); } { E TaW, Tb5, TbG, TbH; TaW = TaK - TaV; Tb5 = Tb1 - Tb4; Tb6 = FMA(KP980785280, TaW, KP195090322 * Tb5); TbM = FNMS(KP980785280, Tb5, KP195090322 * TaW); TbG = FNMS(KP923879532, Taz, KP382683432 * TaC); TbH = FMA(KP923879532, Tas, KP382683432 * Tav); TbI = TbG - TbH; TbS = TbH + TbG; } { E TbU, TbV, Tbn, Tbw; TbU = TaK + TaV; TbV = Tb1 + Tb4; TbW = FMA(KP555570233, TbU, KP831469612 * TbV); Tc6 = FNMS(KP555570233, TbV, KP831469612 * TbU); Tbn = Tbb - Tbm; Tbw = Tbs - Tbv; Tbx = FNMS(KP980785280, Tbw, KP195090322 * Tbn); TbN = FMA(KP195090322, Tbw, KP980785280 * Tbn); } { E TaF, Tby, TbP, TbQ; TaF = Tap + TaE; Tby = Tb6 + Tbx; ro[WS(os, 46)] = TaF - Tby; ro[WS(os, 14)] = TaF + Tby; TbP = TbF + TbI; TbQ = TbM + TbN; io[WS(os, 46)] = TbP - TbQ; io[WS(os, 14)] = TbP + TbQ; } { E TbJ, TbK, TbL, TbO; TbJ = TbF - TbI; TbK = Tbx - Tb6; io[WS(os, 62)] = TbJ - TbK; io[WS(os, 30)] = TbJ + TbK; TbL = Tap - TaE; TbO = TbM - TbN; ro[WS(os, 62)] = TbL - TbO; ro[WS(os, 30)] = TbL + TbO; } { E TbT, Tc0, Tc9, Tca; TbT = TbR + TbS; Tc0 = TbW + TbZ; ro[WS(os, 38)] = TbT - Tc0; ro[WS(os, 6)] = TbT + Tc0; Tc9 = Tc1 + Tc2; Tca = Tc6 + Tc7; io[WS(os, 38)] = Tc9 - Tca; io[WS(os, 6)] = Tc9 + Tca; } { E Tc3, Tc4, Tc5, Tc8; Tc3 = Tc1 - Tc2; Tc4 = TbZ - TbW; io[WS(os, 54)] = Tc3 - Tc4; io[WS(os, 22)] = Tc3 + Tc4; Tc5 = TbR - TbS; Tc8 = Tc6 - Tc7; ro[WS(os, 54)] = Tc5 - Tc8; ro[WS(os, 22)] = Tc5 + Tc8; } } { E T6F, T7h, T7m, T7w, T7p, T7x, T6M, T7s, T6U, T7c, T75, T7r, T78, T7i, T71; E T7d; { E T6D, T6E, T7k, T7l; T6D = T37 + T3e; T6E = T65 + T64; T6F = T6D - T6E; T7h = T6D + T6E; T7k = T6O + T6P; T7l = T6R + T6S; T7m = FMA(KP956940335, T7k, KP290284677 * T7l); T7w = FNMS(KP290284677, T7k, KP956940335 * T7l); } { E T7n, T7o, T6I, T6L; T7n = T6V + T6W; T7o = T6Y + T6Z; T7p = FNMS(KP290284677, T7o, KP956940335 * T7n); T7x = FMA(KP290284677, T7n, KP956940335 * T7o); T6I = FNMS(KP555570233, T6H, KP831469612 * T6G); T6L = FMA(KP831469612, T6J, KP555570233 * T6K); T6M = T6I - T6L; T7s = T6I + T6L; } { E T6Q, T6T, T73, T74; T6Q = T6O - T6P; T6T = T6R - T6S; T6U = FMA(KP471396736, T6Q, KP881921264 * T6T); T7c = FNMS(KP881921264, T6Q, KP471396736 * T6T); T73 = T5Z + T62; T74 = T3m + T3t; T75 = T73 - T74; T7r = T73 + T74; } { E T76, T77, T6X, T70; T76 = FNMS(KP555570233, T6J, KP831469612 * T6K); T77 = FMA(KP555570233, T6G, KP831469612 * T6H); T78 = T76 - T77; T7i = T77 + T76; T6X = T6V - T6W; T70 = T6Y - T6Z; T71 = FNMS(KP881921264, T70, KP471396736 * T6X); T7d = FMA(KP881921264, T6X, KP471396736 * T70); } { E T6N, T72, T7f, T7g; T6N = T6F + T6M; T72 = T6U + T71; ro[WS(os, 43)] = T6N - T72; ro[WS(os, 11)] = T6N + T72; T7f = T75 + T78; T7g = T7c + T7d; io[WS(os, 43)] = T7f - T7g; io[WS(os, 11)] = T7f + T7g; } { E T79, T7a, T7b, T7e; T79 = T75 - T78; T7a = T71 - T6U; io[WS(os, 59)] = T79 - T7a; io[WS(os, 27)] = T79 + T7a; T7b = T6F - T6M; T7e = T7c - T7d; ro[WS(os, 59)] = T7b - T7e; ro[WS(os, 27)] = T7b + T7e; } { E T7j, T7q, T7z, T7A; T7j = T7h + T7i; T7q = T7m + T7p; ro[WS(os, 35)] = T7j - T7q; ro[WS(os, 3)] = T7j + T7q; T7z = T7r + T7s; T7A = T7w + T7x; io[WS(os, 35)] = T7z - T7A; io[WS(os, 3)] = T7z + T7A; } { E T7t, T7u, T7v, T7y; T7t = T7r - T7s; T7u = T7p - T7m; io[WS(os, 51)] = T7t - T7u; io[WS(os, 19)] = T7t + T7u; T7v = T7h - T7i; T7y = T7w - T7x; ro[WS(os, 51)] = T7v - T7y; ro[WS(os, 19)] = T7v + T7y; } } { E T9j, T9V, Ta0, Taa, Ta3, Tab, T9q, Ta6, T9y, T9Q, T9J, Ta5, T9M, T9W, T9F; E T9R; { E T9h, T9i, T9Y, T9Z; T9h = T7B + T7C; T9i = T8J + T8I; T9j = T9h - T9i; T9V = T9h + T9i; T9Y = T9s + T9t; T9Z = T9v + T9w; Ta0 = FMA(KP995184726, T9Y, KP098017140 * T9Z); Taa = FNMS(KP098017140, T9Y, KP995184726 * T9Z); } { E Ta1, Ta2, T9m, T9p; Ta1 = T9z + T9A; Ta2 = T9C + T9D; Ta3 = FNMS(KP098017140, Ta2, KP995184726 * Ta1); Tab = FMA(KP098017140, Ta1, KP995184726 * Ta2); T9m = FNMS(KP195090322, T9l, KP980785280 * T9k); T9p = FMA(KP195090322, T9n, KP980785280 * T9o); T9q = T9m - T9p; Ta6 = T9m + T9p; } { E T9u, T9x, T9H, T9I; T9u = T9s - T9t; T9x = T9v - T9w; T9y = FMA(KP634393284, T9u, KP773010453 * T9x); T9Q = FNMS(KP773010453, T9u, KP634393284 * T9x); T9H = T8F + T8G; T9I = T7G + T7J; T9J = T9H - T9I; Ta5 = T9H + T9I; } { E T9K, T9L, T9B, T9E; T9K = FNMS(KP195090322, T9o, KP980785280 * T9n); T9L = FMA(KP980785280, T9l, KP195090322 * T9k); T9M = T9K - T9L; T9W = T9L + T9K; T9B = T9z - T9A; T9E = T9C - T9D; T9F = FNMS(KP773010453, T9E, KP634393284 * T9B); T9R = FMA(KP773010453, T9B, KP634393284 * T9E); } { E T9r, T9G, T9T, T9U; T9r = T9j + T9q; T9G = T9y + T9F; ro[WS(os, 41)] = T9r - T9G; ro[WS(os, 9)] = T9r + T9G; T9T = T9J + T9M; T9U = T9Q + T9R; io[WS(os, 41)] = T9T - T9U; io[WS(os, 9)] = T9T + T9U; } { E T9N, T9O, T9P, T9S; T9N = T9J - T9M; T9O = T9F - T9y; io[WS(os, 57)] = T9N - T9O; io[WS(os, 25)] = T9N + T9O; T9P = T9j - T9q; T9S = T9Q - T9R; ro[WS(os, 57)] = T9P - T9S; ro[WS(os, 25)] = T9P + T9S; } { E T9X, Ta4, Tad, Tae; T9X = T9V + T9W; Ta4 = Ta0 + Ta3; ro[WS(os, 33)] = T9X - Ta4; ro[WS(os, 1)] = T9X + Ta4; Tad = Ta5 + Ta6; Tae = Taa + Tab; io[WS(os, 33)] = Tad - Tae; io[WS(os, 1)] = Tad + Tae; } { E Ta7, Ta8, Ta9, Tac; Ta7 = Ta5 - Ta6; Ta8 = Ta3 - Ta0; io[WS(os, 49)] = Ta7 - Ta8; io[WS(os, 17)] = Ta7 + Ta8; Ta9 = T9V - T9W; Tac = Taa - Tab; ro[WS(os, 49)] = Ta9 - Tac; ro[WS(os, 17)] = Ta9 + Tac; } } { E T3v, T6j, T6o, T6y, T6r, T6z, T48, T6u, T52, T6e, T67, T6t, T6a, T6k, T5V; E T6f; { E T3f, T3u, T6m, T6n; T3f = T37 - T3e; T3u = T3m - T3t; T3v = T3f - T3u; T6j = T3f + T3u; T6m = T4q + T4N; T6n = T4X + T50; T6o = FMA(KP634393284, T6m, KP773010453 * T6n); T6y = FNMS(KP634393284, T6n, KP773010453 * T6m); } { E T6p, T6q, T3O, T47; T6p = T5j + T5G; T6q = T5Q + T5T; T6r = FNMS(KP634393284, T6q, KP773010453 * T6p); T6z = FMA(KP773010453, T6q, KP634393284 * T6p); T3O = FNMS(KP980785280, T3N, KP195090322 * T3G); T47 = FMA(KP195090322, T3Z, KP980785280 * T46); T48 = T3O - T47; T6u = T3O + T47; } { E T4O, T51, T63, T66; T4O = T4q - T4N; T51 = T4X - T50; T52 = FMA(KP995184726, T4O, KP098017140 * T51); T6e = FNMS(KP995184726, T51, KP098017140 * T4O); T63 = T5Z - T62; T66 = T64 - T65; T67 = T63 - T66; T6t = T63 + T66; } { E T68, T69, T5H, T5U; T68 = FNMS(KP980785280, T3Z, KP195090322 * T46); T69 = FMA(KP980785280, T3G, KP195090322 * T3N); T6a = T68 - T69; T6k = T69 + T68; T5H = T5j - T5G; T5U = T5Q - T5T; T5V = FNMS(KP995184726, T5U, KP098017140 * T5H); T6f = FMA(KP098017140, T5U, KP995184726 * T5H); } { E T49, T5W, T6h, T6i; T49 = T3v + T48; T5W = T52 + T5V; ro[WS(os, 47)] = T49 - T5W; ro[WS(os, 15)] = T49 + T5W; T6h = T67 + T6a; T6i = T6e + T6f; io[WS(os, 47)] = T6h - T6i; io[WS(os, 15)] = T6h + T6i; } { E T6b, T6c, T6d, T6g; T6b = T67 - T6a; T6c = T5V - T52; io[WS(os, 63)] = T6b - T6c; io[WS(os, 31)] = T6b + T6c; T6d = T3v - T48; T6g = T6e - T6f; ro[WS(os, 63)] = T6d - T6g; ro[WS(os, 31)] = T6d + T6g; } { E T6l, T6s, T6B, T6C; T6l = T6j + T6k; T6s = T6o + T6r; ro[WS(os, 39)] = T6l - T6s; ro[WS(os, 7)] = T6l + T6s; T6B = T6t + T6u; T6C = T6y + T6z; io[WS(os, 39)] = T6B - T6C; io[WS(os, 7)] = T6B + T6C; } { E T6v, T6w, T6x, T6A; T6v = T6t - T6u; T6w = T6r - T6o; io[WS(os, 55)] = T6v - T6w; io[WS(os, 23)] = T6v + T6w; T6x = T6j - T6k; T6A = T6y - T6z; ro[WS(os, 55)] = T6x - T6A; ro[WS(os, 23)] = T6x + T6A; } } { E T7L, T8X, T92, T9c, T95, T9d, T80, T98, T8k, T8S, T8L, T97, T8O, T8Y, T8D; E T8T; { E T7D, T7K, T90, T91; T7D = T7B - T7C; T7K = T7G - T7J; T7L = T7D - T7K; T8X = T7D + T7K; T90 = T84 + T8b; T91 = T8f + T8i; T92 = FMA(KP471396736, T90, KP881921264 * T91); T9c = FNMS(KP471396736, T91, KP881921264 * T90); } { E T93, T94, T7S, T7Z; T93 = T8n + T8u; T94 = T8y + T8B; T95 = FNMS(KP471396736, T94, KP881921264 * T93); T9d = FMA(KP881921264, T94, KP471396736 * T93); T7S = FNMS(KP831469612, T7R, KP555570233 * T7O); T7Z = FMA(KP831469612, T7V, KP555570233 * T7Y); T80 = T7S - T7Z; T98 = T7S + T7Z; } { E T8c, T8j, T8H, T8K; T8c = T84 - T8b; T8j = T8f - T8i; T8k = FMA(KP956940335, T8c, KP290284677 * T8j); T8S = FNMS(KP956940335, T8j, KP290284677 * T8c); T8H = T8F - T8G; T8K = T8I - T8J; T8L = T8H - T8K; T97 = T8H + T8K; } { E T8M, T8N, T8v, T8C; T8M = FNMS(KP831469612, T7Y, KP555570233 * T7V); T8N = FMA(KP555570233, T7R, KP831469612 * T7O); T8O = T8M - T8N; T8Y = T8N + T8M; T8v = T8n - T8u; T8C = T8y - T8B; T8D = FNMS(KP956940335, T8C, KP290284677 * T8v); T8T = FMA(KP290284677, T8C, KP956940335 * T8v); } { E T81, T8E, T8V, T8W; T81 = T7L + T80; T8E = T8k + T8D; ro[WS(os, 45)] = T81 - T8E; ro[WS(os, 13)] = T81 + T8E; T8V = T8L + T8O; T8W = T8S + T8T; io[WS(os, 45)] = T8V - T8W; io[WS(os, 13)] = T8V + T8W; } { E T8P, T8Q, T8R, T8U; T8P = T8L - T8O; T8Q = T8D - T8k; io[WS(os, 61)] = T8P - T8Q; io[WS(os, 29)] = T8P + T8Q; T8R = T7L - T80; T8U = T8S - T8T; ro[WS(os, 61)] = T8R - T8U; ro[WS(os, 29)] = T8R + T8U; } { E T8Z, T96, T9f, T9g; T8Z = T8X + T8Y; T96 = T92 + T95; ro[WS(os, 37)] = T8Z - T96; ro[WS(os, 5)] = T8Z + T96; T9f = T97 + T98; T9g = T9c + T9d; io[WS(os, 37)] = T9f - T9g; io[WS(os, 5)] = T9f + T9g; } { E T99, T9a, T9b, T9e; T99 = T97 - T98; T9a = T95 - T92; io[WS(os, 53)] = T99 - T9a; io[WS(os, 21)] = T99 + T9a; T9b = T8X - T8Y; T9e = T9c - T9d; ro[WS(os, 53)] = T9b - T9e; ro[WS(os, 21)] = T9b + T9e; } } }}static void m1_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, int v, int ivs, int ovs){ int i; for (i = v; i > 0; i -= 1) { m1_64_0(ri, ii, ro, io, is, os); ri += ivs; ii += ivs; ro += ovs; io += ovs; }}static const kdft_desc desc = { 64, "m1_64", {808, 144, 104, 0}, &GENUS, 0, 0, 0, 0 };void X(codelet_m1_64) (planner *p) { X(kdft_register) (p, m1_64, &desc);}
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -