📄 m1_32.c
字号:
T2j = T2f - T2i; TC = Ty + TB; TJ = TF + TI; T5A = TC - TJ; { E T4E, T4F, T27, T2c; T5B = T4C + T4D; T5C = T4I + T4J; T5D = T5B - T5C; T4E = T4C - T4D; T4F = TI - TF; T4G = T4E - T4F; T5g = T4F + T4E; T27 = T23 - T26; T2c = T28 + T2b; T2d = KP707106781 * (T27 - T2c); T3F = KP707106781 * (T2c + T27); { E T4H, T4K, T2k, T2l; T4H = Ty - TB; T4K = T4I - T4J; T4L = T4H - T4K; T5h = T4H + T4K; T2k = T2b - T28; T2l = T23 + T26; T2m = KP707106781 * (T2k - T2l); T3I = KP707106781 * (T2k + T2l); } } } { E T4B, T57, T5a, T5c, T4Y, T56, T55, T5b; { E T4t, T4A, T58, T59; T4t = T4r - T4s; T4A = KP707106781 * (T4w - T4z); T4B = T4t + T4A; T57 = T4t - T4A; T58 = FNMS(KP923879532, T4L, KP382683432 * T4G); T59 = FMA(KP382683432, T4W, KP923879532 * T4R); T5a = T58 - T59; T5c = T58 + T59; } { E T4M, T4X, T51, T54; T4M = FMA(KP923879532, T4G, KP382683432 * T4L); T4X = FNMS(KP923879532, T4W, KP382683432 * T4R); T4Y = T4M + T4X; T56 = T4X - T4M; T51 = T4Z - T50; T54 = KP707106781 * (T52 - T53); T55 = T51 - T54; T5b = T51 + T54; } ro[WS(os, 22)] = T4B - T4Y; io[WS(os, 22)] = T5b - T5c; ro[WS(os, 6)] = T4B + T4Y; io[WS(os, 6)] = T5b + T5c; io[WS(os, 30)] = T55 - T56; ro[WS(os, 30)] = T57 - T5a; io[WS(os, 14)] = T55 + T56; ro[WS(os, 14)] = T57 + T5a; } { E T5f, T5r, T5u, T5w, T5m, T5q, T5p, T5v; { E T5d, T5e, T5s, T5t; T5d = T4r + T4s; T5e = KP707106781 * (T53 + T52); T5f = T5d + T5e; T5r = T5d - T5e; T5s = FNMS(KP382683432, T5h, KP923879532 * T5g); T5t = FMA(KP923879532, T5k, KP382683432 * T5j); T5u = T5s - T5t; T5w = T5s + T5t; } { E T5i, T5l, T5n, T5o; T5i = FMA(KP382683432, T5g, KP923879532 * T5h); T5l = FNMS(KP382683432, T5k, KP923879532 * T5j); T5m = T5i + T5l; T5q = T5l - T5i; T5n = T50 + T4Z; T5o = KP707106781 * (T4w + T4z); T5p = T5n - T5o; T5v = T5n + T5o; } ro[WS(os, 18)] = T5f - T5m; io[WS(os, 18)] = T5v - T5w; ro[WS(os, 2)] = T5f + T5m; io[WS(os, 2)] = T5v + T5w; io[WS(os, 26)] = T5p - T5q; ro[WS(os, 26)] = T5r - T5u; io[WS(os, 10)] = T5p + T5q; ro[WS(os, 10)] = T5r + T5u; } { E T5z, T5P, T5S, T5U, T5K, T5O, T5N, T5T; { E T5x, T5y, T5Q, T5R; T5x = T7 - Te; T5y = T1n - T1u; T5z = T5x + T5y; T5P = T5x - T5y; T5Q = T5D - T5A; T5R = T5F + T5I; T5S = KP707106781 * (T5Q - T5R); T5U = KP707106781 * (T5Q + T5R); } { E T5E, T5J, T5L, T5M; T5E = T5A + T5D; T5J = T5F - T5I; T5K = KP707106781 * (T5E + T5J); T5O = KP707106781 * (T5J - T5E); T5L = T18 - T1f; T5M = Tt - Tm; T5N = T5L - T5M; T5T = T5M + T5L; } ro[WS(os, 20)] = T5z - T5K; io[WS(os, 20)] = T5T - T5U; ro[WS(os, 4)] = T5z + T5K; io[WS(os, 4)] = T5T + T5U; io[WS(os, 28)] = T5N - T5O; ro[WS(os, 28)] = T5P - T5S; io[WS(os, 12)] = T5N + T5O; ro[WS(os, 12)] = T5P + T5S; } { E Tv, T5V, T5Y, T60, T10, T11, T1w, T5Z; { E Tf, Tu, T5W, T5X; Tf = T7 + Te; Tu = Tm + Tt; Tv = Tf + Tu; T5V = Tf - Tu; T5W = T5B + T5C; T5X = T5G + T5H; T5Y = T5W - T5X; T60 = T5W + T5X; } { E TK, TZ, T1g, T1v; TK = TC + TJ; TZ = TR + TY; T10 = TK + TZ; T11 = TZ - TK; T1g = T18 + T1f; T1v = T1n + T1u; T1w = T1g - T1v; T5Z = T1g + T1v; } ro[WS(os, 16)] = Tv - T10; io[WS(os, 16)] = T5Z - T60; ro[0] = Tv + T10; io[0] = T5Z + T60; io[WS(os, 8)] = T11 + T1w; ro[WS(os, 8)] = T5V + T5Y; io[WS(os, 24)] = T1w - T11; ro[WS(os, 24)] = T5V - T5Y; } { E T1X, T33, T31, T37, T2o, T34, T2P, T35; { E T1H, T1W, T2X, T30; T1H = T1z - T1G; T1W = T1O - T1V; T1X = T1H + T1W; T33 = T1H - T1W; T2X = T2T - T2W; T30 = T2Y - T2Z; T31 = T2X - T30; T37 = T2X + T30; } { E T2e, T2n, T2F, T2O; T2e = T22 - T2d; T2n = T2j - T2m; T2o = FMA(KP980785280, T2e, KP195090322 * T2n); T34 = FNMS(KP980785280, T2n, KP195090322 * T2e); T2F = T2t - T2E; T2O = T2K - T2N; T2P = FNMS(KP980785280, T2O, KP195090322 * T2F); T35 = FMA(KP195090322, T2O, KP980785280 * T2F); } { E T2Q, T38, T32, T36; T2Q = T2o + T2P; ro[WS(os, 23)] = T1X - T2Q; ro[WS(os, 7)] = T1X + T2Q; T38 = T34 + T35; io[WS(os, 23)] = T37 - T38; io[WS(os, 7)] = T37 + T38; T32 = T2P - T2o; io[WS(os, 31)] = T31 - T32; io[WS(os, 15)] = T31 + T32; T36 = T34 - T35; ro[WS(os, 31)] = T33 - T36; ro[WS(os, 15)] = T33 + T36; } } { E T3D, T41, T3Z, T45, T3K, T42, T3R, T43; { E T3v, T3C, T3V, T3Y; T3v = T3t - T3u; T3C = T3y - T3B; T3D = T3v + T3C; T41 = T3v - T3C; T3V = T3T - T3U; T3Y = T3W - T3X; T3Z = T3V - T3Y; T45 = T3V + T3Y; } { E T3G, T3J, T3N, T3Q; T3G = T3E - T3F; T3J = T3H - T3I; T3K = FMA(KP555570233, T3G, KP831469612 * T3J); T42 = FNMS(KP831469612, T3G, KP555570233 * T3J); T3N = T3L - T3M; T3Q = T3O - T3P; T3R = FNMS(KP831469612, T3Q, KP555570233 * T3N); T43 = FMA(KP831469612, T3N, KP555570233 * T3Q); } { E T3S, T46, T40, T44; T3S = T3K + T3R; ro[WS(os, 21)] = T3D - T3S; ro[WS(os, 5)] = T3D + T3S; T46 = T42 + T43; io[WS(os, 21)] = T45 - T46; io[WS(os, 5)] = T45 + T46; T40 = T3R - T3K; io[WS(os, 29)] = T3Z - T40; io[WS(os, 13)] = T3Z + T40; T44 = T42 - T43; ro[WS(os, 29)] = T41 - T44; ro[WS(os, 13)] = T41 + T44; } } { E T49, T4l, T4j, T4p, T4c, T4m, T4f, T4n; { E T47, T48, T4h, T4i; T47 = T3t + T3u; T48 = T3X + T3W; T49 = T47 + T48; T4l = T47 - T48; T4h = T3T + T3U; T4i = T3y + T3B; T4j = T4h - T4i; T4p = T4h + T4i; } { E T4a, T4b, T4d, T4e; T4a = T3E + T3F; T4b = T3H + T3I; T4c = FMA(KP980785280, T4a, KP195090322 * T4b); T4m = FNMS(KP195090322, T4a, KP980785280 * T4b); T4d = T3L + T3M; T4e = T3O + T3P; T4f = FNMS(KP195090322, T4e, KP980785280 * T4d); T4n = FMA(KP195090322, T4d, KP980785280 * T4e); } { E T4g, T4q, T4k, T4o; T4g = T4c + T4f; ro[WS(os, 17)] = T49 - T4g; ro[WS(os, 1)] = T49 + T4g; T4q = T4m + T4n; io[WS(os, 17)] = T4p - T4q; io[WS(os, 1)] = T4p + T4q; T4k = T4f - T4c; io[WS(os, 25)] = T4j - T4k; io[WS(os, 9)] = T4j + T4k; T4o = T4m - T4n; ro[WS(os, 25)] = T4l - T4o; ro[WS(os, 9)] = T4l + T4o; } } { E T3b, T3n, T3l, T3r, T3e, T3o, T3h, T3p; { E T39, T3a, T3j, T3k; T39 = T1z + T1G; T3a = T2Z + T2Y; T3b = T39 + T3a; T3n = T39 - T3a; T3j = T2T + T2W; T3k = T1O + T1V; T3l = T3j - T3k; T3r = T3j + T3k; } { E T3c, T3d, T3f, T3g; T3c = T22 + T2d; T3d = T2j + T2m; T3e = FMA(KP555570233, T3c, KP831469612 * T3d); T3o = FNMS(KP555570233, T3d, KP831469612 * T3c); T3f = T2t + T2E; T3g = T2K + T2N; T3h = FNMS(KP555570233, T3g, KP831469612 * T3f); T3p = FMA(KP831469612, T3g, KP555570233 * T3f); } { E T3i, T3s, T3m, T3q; T3i = T3e + T3h; ro[WS(os, 19)] = T3b - T3i; ro[WS(os, 3)] = T3b + T3i; T3s = T3o + T3p; io[WS(os, 19)] = T3r - T3s; io[WS(os, 3)] = T3r + T3s; T3m = T3h - T3e; io[WS(os, 27)] = T3l - T3m; io[WS(os, 11)] = T3l + T3m; T3q = T3o - T3p; ro[WS(os, 27)] = T3n - T3q; ro[WS(os, 11)] = T3n + T3q; } } }}static void m1_32(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_32_0(ri, ii, ro, io, is, os); ri += ivs; ii += ivs; ro += ovs; io += ovs; }}static const kdft_desc desc = { 32, "m1_32", {340, 52, 32, 0}, &GENUS, 0, 0, 0, 0 };void X(codelet_m1_32) (planner *p) { X(kdft_register) (p, m1_32, &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 + -