📄 mr2hc_128.c
字号:
TeY = TeW - TeX; Tfi = TeX + TeW; TeN = TeL - TeM; TeQ = TeO - TeP; TeR = FNMS(KP471396736, TeQ, KP881921264 * TeN); Tf3 = FMA(KP881921264, TeQ, KP471396736 * TeN); } { E TeD, TeS, Tf1, Tf4; TeD = Tev + TeC; TeS = TeK + TeR; ro[WS(ros, 54)] = TeD - TeS; ro[WS(ros, 10)] = TeD + TeS; Tf1 = TeY + TeV; Tf4 = Tf2 + Tf3; io[WS(ios, 10)] = Tf1 + Tf4; io[WS(ios, 54)] = Tf4 - Tf1; } { E TeZ, Tf0, Tf5, Tf6; TeZ = TeV - TeY; Tf0 = TeR - TeK; io[WS(ios, 22)] = TeZ + Tf0; io[WS(ios, 42)] = Tf0 - TeZ; Tf5 = Tev - TeC; Tf6 = Tf3 - Tf2; ro[WS(ros, 42)] = Tf5 - Tf6; ro[WS(ros, 22)] = Tf5 + Tf6; } { E Tf9, Tfg, Tfl, Tfo; Tf9 = Tf7 + Tf8; Tfg = Tfc + Tff; ro[WS(ros, 58)] = Tf9 - Tfg; ro[WS(ros, 6)] = Tf9 + Tfg; Tfl = Tfi + Tfh; Tfo = Tfm + Tfn; io[WS(ios, 6)] = Tfl + Tfo; io[WS(ios, 58)] = Tfo - Tfl; } { E Tfj, Tfk, Tfp, Tfq; Tfj = Tfh - Tfi; Tfk = Tff - Tfc; io[WS(ios, 26)] = Tfj + Tfk; io[WS(ios, 38)] = Tfk - Tfj; Tfp = Tf7 - Tf8; Tfq = Tfn - Tfm; ro[WS(ros, 38)] = Tfp - Tfq; ro[WS(ros, 26)] = Tfp + Tfq; } } { E TcP, Te9, Tee, Teo, Teh, Tep, Td8, Tej, Tdu, Te4, TdT, Tea, Te0, Tek, TdP; E Te5; { E TcH, TcO, Tec, Ted; TcH = TcD + TcG; TcO = TcK + TcN; TcP = TcH + TcO; Te9 = TcH - TcO; Tec = Tde - Tdl; Ted = Tds - Tdp; Tee = FMA(KP773010453, Tec, KP634393284 * Ted); Teo = FNMS(KP634393284, Tec, KP773010453 * Ted); } { E Tef, Teg, TcY, Td7; Tef = Tdz - TdG; Teg = TdN - TdK; Teh = FNMS(KP634393284, Teg, KP773010453 * Tef); Tep = FMA(KP634393284, Tef, KP773010453 * Teg); TcY = FMA(KP980785280, TcU, KP195090322 * TcX); Td7 = FNMS(KP195090322, Td6, KP980785280 * Td3); Td8 = TcY + Td7; Tej = Td7 - TcY; } { E Tdm, Tdt, TdR, TdS; Tdm = Tde + Tdl; Tdt = Tdp + Tds; Tdu = FMA(KP995184726, Tdm, KP098017140 * Tdt); Te4 = FNMS(KP098017140, Tdm, KP995184726 * Tdt); TdR = FNMS(KP195090322, TcU, KP980785280 * TcX); TdS = FMA(KP195090322, Td3, KP980785280 * Td6); TdT = TdR + TdS; Tea = TdS - TdR; } { E TdW, TdZ, TdH, TdO; TdW = TdU + TdV; TdZ = TdX + TdY; Te0 = TdW + TdZ; Tek = TdZ - TdW; TdH = Tdz + TdG; TdO = TdK + TdN; TdP = FNMS(KP098017140, TdO, KP995184726 * TdH); Te5 = FMA(KP098017140, TdH, KP995184726 * TdO); } { E Td9, TdQ, Te3, Te6; Td9 = TcP + Td8; TdQ = Tdu + TdP; ro[WS(ros, 62)] = Td9 - TdQ; ro[WS(ros, 2)] = Td9 + TdQ; Te3 = Te0 + TdT; Te6 = Te4 + Te5; io[WS(ios, 2)] = Te3 + Te6; io[WS(ios, 62)] = Te6 - Te3; } { E Te1, Te2, Te7, Te8; Te1 = TdT - Te0; Te2 = TdP - Tdu; io[WS(ios, 30)] = Te1 + Te2; io[WS(ios, 34)] = Te2 - Te1; Te7 = TcP - Td8; Te8 = Te5 - Te4; ro[WS(ros, 34)] = Te7 - Te8; ro[WS(ros, 30)] = Te7 + Te8; } { E Teb, Tei, Ten, Teq; Teb = Te9 + Tea; Tei = Tee + Teh; ro[WS(ros, 50)] = Teb - Tei; ro[WS(ros, 14)] = Teb + Tei; Ten = Tek + Tej; Teq = Teo + Tep; io[WS(ios, 14)] = Ten + Teq; io[WS(ios, 50)] = Teq - Ten; } { E Tel, Tem, Ter, Tes; Tel = Tej - Tek; Tem = Teh - Tee; io[WS(ios, 18)] = Tel + Tem; io[WS(ios, 46)] = Tem - Tel; Ter = Te9 - Tea; Tes = Tep - Teo; ro[WS(ros, 46)] = Ter - Tes; ro[WS(ros, 18)] = Ter + Tes; } } { E T6v, T77, T6C, T7h, T6Y, T7i, T6V, T78, T6R, T7n, T73, T7f, T6K, T7m, T72; E T7c; { E T6t, T6u, T6T, T6U; T6t = T27 - T2e; T6u = T5Y - T5X; T6v = T6t - T6u; T77 = T6t + T6u; { E T6y, T6B, T6W, T6X; T6y = FMA(KP773010453, T6w, KP634393284 * T6x); T6B = FNMS(KP634393284, T6A, KP773010453 * T6z); T6C = T6y - T6B; T7h = T6B + T6y; T6W = T2x - T2o; T6X = T5V - T5S; T6Y = T6W - T6X; T7i = T6X + T6W; } T6T = FNMS(KP634393284, T6w, KP773010453 * T6x); T6U = FMA(KP634393284, T6z, KP773010453 * T6A); T6V = T6T - T6U; T78 = T6U + T6T; { E T6N, T7d, T6Q, T7e, T6M, T6O; T6M = T5I - T5H; T6N = T6L - T6M; T7d = T6L + T6M; T6O = T5v - T5e; T6Q = T6O - T6P; T7e = T6P + T6O; T6R = FNMS(KP427555093, T6Q, KP903989293 * T6N); T7n = FMA(KP941544065, T7e, KP336889853 * T7d); T73 = FMA(KP903989293, T6Q, KP427555093 * T6N); T7f = FNMS(KP336889853, T7e, KP941544065 * T7d); } { E T6G, T7a, T6J, T7b, T6E, T6I; T6E = T4g - T3Z; T6G = T6E - T6F; T7a = T6F + T6E; T6I = T4t - T4s; T6J = T6H - T6I; T7b = T6H + T6I; T6K = FMA(KP427555093, T6G, KP903989293 * T6J); T7m = FNMS(KP336889853, T7b, KP941544065 * T7a); T72 = FNMS(KP427555093, T6J, KP903989293 * T6G); T7c = FMA(KP336889853, T7a, KP941544065 * T7b); } } { E T6D, T6S, T71, T74; T6D = T6v + T6C; T6S = T6K + T6R; ro[WS(ros, 55)] = T6D - T6S; ro[WS(ros, 9)] = T6D + T6S; T71 = T6Y + T6V; T74 = T72 + T73; io[WS(ios, 9)] = T71 + T74; io[WS(ios, 55)] = T74 - T71; } { E T6Z, T70, T75, T76; T6Z = T6V - T6Y; T70 = T6R - T6K; io[WS(ios, 23)] = T6Z + T70; io[WS(ios, 41)] = T70 - T6Z; T75 = T6v - T6C; T76 = T73 - T72; ro[WS(ros, 41)] = T75 - T76; ro[WS(ros, 23)] = T75 + T76; } { E T79, T7g, T7l, T7o; T79 = T77 + T78; T7g = T7c + T7f; ro[WS(ros, 57)] = T79 - T7g; ro[WS(ros, 7)] = T79 + T7g; T7l = T7i + T7h; T7o = T7m + T7n; io[WS(ios, 7)] = T7l + T7o; io[WS(ios, 57)] = T7o - T7l; } { E T7j, T7k, T7p, T7q; T7j = T7h - T7i; T7k = T7f - T7c; io[WS(ios, 25)] = T7j + T7k; io[WS(ios, 39)] = T7k - T7j; T7p = T77 - T78; T7q = T7n - T7m; ro[WS(ros, 39)] = T7p - T7q; ro[WS(ros, 25)] = T7p + T7q; } } { E T99, T9L, T9g, T9V, T9C, T9W, T9z, T9M, T9v, Ta1, T9H, T9T, T9o, Ta0, T9G; E T9Q; { E T97, T98, T9x, T9y; T97 = T7r - T7s; T98 = T8C - T8B; T99 = T97 - T98; T9L = T97 + T98; { E T9c, T9f, T9A, T9B; T9c = FMA(KP471396736, T9a, KP881921264 * T9b); T9f = FNMS(KP471396736, T9e, KP881921264 * T9d); T9g = T9c - T9f; T9V = T9f + T9c; T9A = T7z - T7w; T9B = T8z - T8y; T9C = T9A - T9B; T9W = T9B + T9A; } T9x = FNMS(KP471396736, T9b, KP881921264 * T9a); T9y = FMA(KP881921264, T9e, KP471396736 * T9d); T9z = T9x - T9y; T9M = T9y + T9x; { E T9r, T9R, T9u, T9S, T9q, T9s; T9q = T8q - T8p; T9r = T9p - T9q; T9R = T9p + T9q; T9s = T8j - T8g; T9u = T9s - T9t; T9S = T9t + T9s; T9v = FNMS(KP514102744, T9u, KP857728610 * T9r); Ta1 = FMA(KP970031253, T9S, KP242980179 * T9R); T9H = FMA(KP857728610, T9u, KP514102744 * T9r); T9T = FNMS(KP242980179, T9S, KP970031253 * T9R); } { E T9k, T9O, T9n, T9P, T9i, T9m; T9i = T80 - T7X; T9k = T9i - T9j; T9O = T9j + T9i; T9m = T87 - T86; T9n = T9l - T9m; T9P = T9l + T9m; T9o = FMA(KP514102744, T9k, KP857728610 * T9n); Ta0 = FNMS(KP242980179, T9P, KP970031253 * T9O); T9G = FNMS(KP514102744, T9n, KP857728610 * T9k); T9Q = FMA(KP242980179, T9O, KP970031253 * T9P); } } { E T9h, T9w, T9F, T9I; T9h = T99 + T9g; T9w = T9o + T9v; ro[WS(ros, 53)] = T9h - T9w; ro[WS(ros, 11)] = T9h + T9w; T9F = T9C + T9z; T9I = T9G + T9H; io[WS(ios, 11)] = T9F + T9I; io[WS(ios, 53)] = T9I - T9F; } { E T9D, T9E, T9J, T9K; T9D = T9z - T9C; T9E = T9v - T9o; io[WS(ios, 21)] = T9D + T9E; io[WS(ios, 43)] = T9E - T9D; T9J = T99 - T9g; T9K = T9H - T9G; ro[WS(ros, 43)] = T9J - T9K; ro[WS(ros, 21)] = T9J + T9K; } { E T9N, T9U, T9Z, Ta2; T9N = T9L + T9M; T9U = T9Q + T9T; ro[WS(ros, 59)] = T9N - T9U; ro[WS(ros, 5)] = T9N + T9U; T9Z = T9W + T9V; Ta2 = Ta0 + Ta1; io[WS(ios, 5)] = T9Z + Ta2; io[WS(ios, 59)] = Ta2 - T9Z; } { E T9X, T9Y, Ta3, Ta4; T9X = T9V - T9W; T9Y = T9T - T9Q; io[WS(ios, 27)] = T9X + T9Y; io[WS(ios, 37)] = T9Y - T9X; Ta3 = T9L - T9M; Ta4 = Ta1 - Ta0; ro[WS(ros, 37)] = Ta3 - Ta4; ro[WS(ros, 27)] = Ta3 + Ta4; } } { E T2z, T69, T3g, T6j, T60, T6k, T5P, T6a, T5L, T6p, T65, T6h, T4w, T6o, T64; E T6e; { E T2f, T2y, T5N, T5O; T2f = T27 + T2e; T2y = T2o + T2x; T2z = T2f + T2y; T69 = T2f - T2y; { E T2U, T3f, T5W, T5Z; T2U = FMA(KP098017140, T2M, KP995184726 * T2T); T3f = FNMS(KP098017140, T3e, KP995184726 * T37); T3g = T2U + T3f; T6j = T3f - T2U; T5W = T5S + T5V; T5Z = T5X + T5Y; T60 = T5W + T5Z; T6k = T5Z - T5W; } T5N = FNMS(KP098017140, T2T, KP995184726 * T2M); T5O = FMA(KP995184726, T3e, KP098017140 * T37); T5P = T5N + T5O; T6a = T5O - T5N; { E T5x, T6f, T5K, T6g, T5w, T5J; T5w = T5e + T5v; T5x = T4X + T5w; T6f = T4X - T5w; T5J = T5H + T5I; T5K = T5G + T5J; T6g = T5J - T5G; T5L = FNMS(KP049067674, T5K, KP998795456 * T5x); T6p = FMA(KP671558954, T6f, KP740951125 * T6g); T65 = FMA(KP049067674, T5x, KP998795456 * T5K); T6h = FNMS(KP671558954, T6g, KP740951125 * T6f); } { E T4i, T6c, T4v, T6d, T4h, T4u; T4h = T3Z + T4g; T4i = T3I + T4h; T6c = T3I - T4h; T4u = T4s + T4t; T4v = T4r + T4u; T6d = T4u - T4r; T4w = FMA(KP998795456, T4i, KP049067674 * T4v); T6o = FNMS(KP671558954, T6c, KP740951125 * T6d); T64 = FNMS(KP049067674, T4i, KP998795456 * T4v); T6e = FMA(KP740951125, T6c, KP671558954 * T6d); } } { E T3h, T5M, T63, T66; T3h = T2z + T3g; T5M = T4w + T5L; ro[WS(ros, 63)] = T3h - T5M; ro[WS(ros, 1)] = T3h + T5M; T63 = T60 + T5P; T66 = T64 + T65; io[WS(ios, 1)] = T63 + T66; io[WS(ios, 63)] = T66 - T63; } { E T61, T62, T67, T68; T61 = T5P - T60; T62 = T5L - T4w; io[WS(ios, 31)] = T61 + T62; io[WS(ios, 33)] = T62 - T61; T67 = T2z - T3g; T68 = T65 - T64; ro[WS(ros, 33)] = T67 - T68; ro[WS(ros, 31)] = T67 + T68; } { E T6b, T6i, T6n, T6q; T6b = T69 + T6a; T6i = T6e + T6h; ro[WS(ros, 49)] = T6b - T6i; ro[WS(ros, 15)] = T6b + T6i; T6n = T6k + T6j; T6q = T6o + T6p; io[WS(ios, 15)] = T6n + T6q; io[WS(ios, 49)] = T6q - T6n; } { E T6l, T6m, T6r, T6s; T6l = T6j - T6k; T6m = T6h - T6e; io[WS(ios, 17)] = T6l + T6m; io[WS(ios, 47)] = T6m - T6l; T6r = T69 - T6a; T6s = T6p - T6o; ro[WS(ros, 47)] = T6r - T6s; ro[WS(ros, 17)] = T6r + T6s; } } { E T7B, T8N, T7Q, T8X, T8E, T8Y, T8x, T8O, T8t, T93, T8J, T8V, T8a, T92, T8I; E T8S; { E T7t, T7A, T8v, T8w; T7t = T7r + T7s; T7A = T7w + T7z; T7B = T7t + T7A; T8N = T7t - T7A; { E T7I, T7P, T8A, T8D; T7I = FMA(KP956940335, T7E, KP290284677 * T7H); T7P = FNMS(KP290284677, T7O, KP956940335 * T7L); T7Q = T7I + T7P; T8X = T7P - T7I; T8A = T8y + T8z; T8D = T8B + T8C; T8E = T8A + T8D; T8Y = T8D - T8A; } T8v = FNMS(KP290284677, T7E, KP956940335 * T7H); T8w = FMA(KP290284677, T7L, KP956940335 * T7O); T8x = T8v + T8w; T8O = T8w - T8v; { E T8l, T8T, T8s, T8U, T8k, T8r; T8k = T8g + T8j; T8l = T8d + T8k; T8T = T8d - T8k; T8r = T8p + T8q; T8s = T8o + T8r; T8U = T8r - T8o; T8t = FNMS(KP146730474, T8s, KP989176509 * T8l); T93 = FMA(KP595699304, T8T, KP803207531 * T8U); T8J = FMA(KP146730474, T8l, KP989176509 * T8s); T8V = FNMS(KP595699304, T8U, KP803207531 * T8T); } { E T82, T8Q, T89, T8R, T81, T88; T81 = T7X + T80; T82 = T7U + T81; T8Q = T7U - T81; T88 = T86 + T87; T89 = T85 + T88; T8R = T88 - T85; T8a = FMA(KP989176509, T82, KP146730474 * T89); T92 = FNMS(KP595699304, T8Q, KP803207531 * T8R); T8I = FNMS(KP146730474, T82, KP989176509 * T89); T8S = FMA(KP803207531, T8Q, KP595699304 * T8R); } } { E T7R, T8u, T8H, T8K; T7R = T7B + T7Q; T8u = T8a + T8t; ro[WS(ros, 61)] = T7R - T8u; ro[WS(ros, 3)] = T7R + T8u; T8H = T8E + T8x; T8K = T8I + T8J; io[WS(ios, 3)] = T8H + T8K; io[WS(ios, 61)] = T8K - T8H; } { E T8F, T8G, T8L, T8M; T8F = T8x - T8E; T8G = T8t - T8a; io[WS(ios, 29)] = T8F + T8G; io[WS(ios, 35)] = T8G - T8F; T8L = T7B - T7Q; T8M = T8J - T8I; ro[WS(ros, 35)] = T8L - T8M; ro[WS(ros, 29)] = T8L + T8M; } { E T8P, T8W, T91, T94; T8P = T8N + T8O; T8W = T8S + T8V; ro[WS(ros, 51)] = T8P - T8W; ro[WS(ros, 13)] = T8P + T8W; T91 = T8Y + T8X; T94 = T92 + T93; io[WS(ios, 13)] = T91 + T94; io[WS(ios, 51)] = T94 - T91; } { E T8Z, T90, T95, T96; T8Z = T8X - T8Y; T90 = T8V - T8S; io[WS(ios, 19)] = T8Z + T90; io[WS(ios, 45)] = T90 - T8Z; T95 = T8N - T8O; T96 = T93 - T92; ro[WS(ros, 45)] = T95 - T96; ro[WS(ros, 19)] = T95 + T96; } } }}static void mr2hc_128(const R *I, R *ro, R *io, stride is, stride ros, stride ios, int v, int ivs, int ovs){ int i; for (i = v; i > 0; --i) { mr2hc_128_0(I, ro, io, is, ros, ios); I += ivs; ro += ovs; io += ovs; }}static const kr2hc_desc desc = { 128, "mr2hc_128", {812, 186, 144, 0}, &GENUS, 0, 0, 0, 0, 0 };void X(codelet_mr2hc_128) (planner *p) { X(kr2hc_register) (p, mr2hc_128, &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 + -