📄 t2sv_32.c
字号:
T4W = VMUL(LDK(KP707106781), VSUB(T4Q, T4V)); T5R = VMUL(LDK(KP707106781), VADD(T4Q, T4V)); T53 = VSUB(T4R, T4U); T54 = VADD(T4P, T4O); T55 = VMUL(LDK(KP707106781), VSUB(T53, T54)); T5O = VMUL(LDK(KP707106781), VADD(T54, T53)); } } { V T2b, T7x, T7K, T7M, T3A, T7L, T7A, T7B; { V T1j, T2a, T7C, T7J; T1j = VADD(TL, T1i); T2a = VADD(T1E, T29); T2b = VADD(T1j, T2a); T7x = VSUB(T1j, T2a); T7C = VADD(T7e, T7f); T7J = VADD(T7D, T7I); T7K = VADD(T7C, T7J); T7M = VSUB(T7J, T7C); } { V T2S, T3z, T7y, T7z; T2S = VADD(T2y, T2R); T3z = VADD(T3h, T3y); T3A = VADD(T2S, T3z); T7L = VSUB(T3z, T2S); T7y = VADD(T7j, T7k); T7z = VADD(T7o, T7p); T7A = VSUB(T7y, T7z); T7B = VADD(T7y, T7z); } ST(&(ri[WS(rs, 16)]), VSUB(T2b, T3A), ms, &(ri[0])); ST(&(ii[WS(rs, 16)]), VSUB(T7K, T7B), ms, &(ii[0])); ST(&(ri[0]), VADD(T2b, T3A), ms, &(ri[0])); ST(&(ii[0]), VADD(T7B, T7K), ms, &(ii[0])); ST(&(ri[WS(rs, 24)]), VSUB(T7x, T7A), ms, &(ri[0])); ST(&(ii[WS(rs, 24)]), VSUB(T7M, T7L), ms, &(ii[0])); ST(&(ri[WS(rs, 8)]), VADD(T7x, T7A), ms, &(ri[0])); ST(&(ii[WS(rs, 8)]), VADD(T7L, T7M), ms, &(ii[0])); } { V T7h, T7t, T7Q, T7S, T7m, T7u, T7r, T7v; { V T7d, T7g, T7O, T7P; T7d = VSUB(TL, T1i); T7g = VSUB(T7e, T7f); T7h = VADD(T7d, T7g); T7t = VSUB(T7d, T7g); T7O = VSUB(T29, T1E); T7P = VSUB(T7I, T7D); T7Q = VADD(T7O, T7P); T7S = VSUB(T7P, T7O); } { V T7i, T7l, T7n, T7q; T7i = VSUB(T2y, T2R); T7l = VSUB(T7j, T7k); T7m = VADD(T7i, T7l); T7u = VSUB(T7l, T7i); T7n = VSUB(T3h, T3y); T7q = VSUB(T7o, T7p); T7r = VSUB(T7n, T7q); T7v = VADD(T7n, T7q); } { V T7s, T7N, T7w, T7R; T7s = VMUL(LDK(KP707106781), VADD(T7m, T7r)); ST(&(ri[WS(rs, 20)]), VSUB(T7h, T7s), ms, &(ri[0])); ST(&(ri[WS(rs, 4)]), VADD(T7h, T7s), ms, &(ri[0])); T7N = VMUL(LDK(KP707106781), VADD(T7u, T7v)); ST(&(ii[WS(rs, 4)]), VADD(T7N, T7Q), ms, &(ii[0])); ST(&(ii[WS(rs, 20)]), VSUB(T7Q, T7N), ms, &(ii[0])); T7w = VMUL(LDK(KP707106781), VSUB(T7u, T7v)); ST(&(ri[WS(rs, 28)]), VSUB(T7t, T7w), ms, &(ri[0])); ST(&(ri[WS(rs, 12)]), VADD(T7t, T7w), ms, &(ri[0])); T7R = VMUL(LDK(KP707106781), VSUB(T7r, T7m)); ST(&(ii[WS(rs, 12)]), VADD(T7R, T7S), ms, &(ii[0])); ST(&(ii[WS(rs, 28)]), VSUB(T7S, T7R), ms, &(ii[0])); } } { V T6j, T7X, T83, T6X, T6u, T7U, T77, T7b, T70, T82, T6G, T6U, T74, T7a, T6R; V T6V; { V T6o, T6t, T6A, T6F; T6j = VSUB(T6f, T6i); T7X = VADD(T7V, T7W); T83 = VSUB(T7W, T7V); T6X = VADD(T6f, T6i); T6o = VSUB(T6m, T6n); T6t = VADD(T6p, T6s); T6u = VMUL(LDK(KP707106781), VSUB(T6o, T6t)); T7U = VMUL(LDK(KP707106781), VADD(T6o, T6t)); { V T75, T76, T6Y, T6Z; T75 = VADD(T6H, T6K); T76 = VADD(T6O, T6P); T77 = VFNMS(LDK(KP382683432), T76, VMUL(LDK(KP923879532), T75)); T7b = VFMA(LDK(KP923879532), T76, VMUL(LDK(KP382683432), T75)); T6Y = VADD(T6n, T6m); T6Z = VSUB(T6p, T6s); T70 = VMUL(LDK(KP707106781), VADD(T6Y, T6Z)); T82 = VMUL(LDK(KP707106781), VSUB(T6Z, T6Y)); } T6A = VSUB(T6y, T6z); T6F = VSUB(T6B, T6E); T6G = VFMA(LDK(KP923879532), T6A, VMUL(LDK(KP382683432), T6F)); T6U = VFNMS(LDK(KP923879532), T6F, VMUL(LDK(KP382683432), T6A)); { V T72, T73, T6L, T6Q; T72 = VADD(T6y, T6z); T73 = VADD(T6B, T6E); T74 = VFMA(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T73)); T7a = VFNMS(LDK(KP382683432), T73, VMUL(LDK(KP923879532), T72)); T6L = VSUB(T6H, T6K); T6Q = VSUB(T6O, T6P); T6R = VFNMS(LDK(KP923879532), T6Q, VMUL(LDK(KP382683432), T6L)); T6V = VFMA(LDK(KP382683432), T6Q, VMUL(LDK(KP923879532), T6L)); } } { V T6v, T6S, T81, T84; T6v = VADD(T6j, T6u); T6S = VADD(T6G, T6R); ST(&(ri[WS(rs, 22)]), VSUB(T6v, T6S), ms, &(ri[0])); ST(&(ri[WS(rs, 6)]), VADD(T6v, T6S), ms, &(ri[0])); T81 = VADD(T6U, T6V); T84 = VADD(T82, T83); ST(&(ii[WS(rs, 6)]), VADD(T81, T84), ms, &(ii[0])); ST(&(ii[WS(rs, 22)]), VSUB(T84, T81), ms, &(ii[0])); } { V T6T, T6W, T85, T86; T6T = VSUB(T6j, T6u); T6W = VSUB(T6U, T6V); ST(&(ri[WS(rs, 30)]), VSUB(T6T, T6W), ms, &(ri[0])); ST(&(ri[WS(rs, 14)]), VADD(T6T, T6W), ms, &(ri[0])); T85 = VSUB(T6R, T6G); T86 = VSUB(T83, T82); ST(&(ii[WS(rs, 14)]), VADD(T85, T86), ms, &(ii[0])); ST(&(ii[WS(rs, 30)]), VSUB(T86, T85), ms, &(ii[0])); } { V T71, T78, T7T, T7Y; T71 = VADD(T6X, T70); T78 = VADD(T74, T77); ST(&(ri[WS(rs, 18)]), VSUB(T71, T78), ms, &(ri[0])); ST(&(ri[WS(rs, 2)]), VADD(T71, T78), ms, &(ri[0])); T7T = VADD(T7a, T7b); T7Y = VADD(T7U, T7X); ST(&(ii[WS(rs, 2)]), VADD(T7T, T7Y), ms, &(ii[0])); ST(&(ii[WS(rs, 18)]), VSUB(T7Y, T7T), ms, &(ii[0])); } { V T79, T7c, T7Z, T80; T79 = VSUB(T6X, T70); T7c = VSUB(T7a, T7b); ST(&(ri[WS(rs, 26)]), VSUB(T79, T7c), ms, &(ri[0])); ST(&(ri[WS(rs, 10)]), VADD(T79, T7c), ms, &(ri[0])); T7Z = VSUB(T77, T74); T80 = VSUB(T7X, T7U); ST(&(ii[WS(rs, 10)]), VADD(T7Z, T80), ms, &(ii[0])); ST(&(ii[WS(rs, 26)]), VSUB(T80, T7Z), ms, &(ii[0])); } } { V T3R, T5d, T8r, T8x, T4e, T8o, T5n, T5r, T4G, T5a, T5g, T8w, T5k, T5q, T57; V T5b, T3Q, T8p; T3Q = VMUL(LDK(KP707106781), VSUB(T3K, T3P)); T3R = VSUB(T3F, T3Q); T5d = VADD(T3F, T3Q); T8p = VMUL(LDK(KP707106781), VSUB(T5v, T5u)); T8r = VADD(T8p, T8q); T8x = VSUB(T8q, T8p); { V T42, T4d, T5l, T5m; T42 = VFNMS(LDK(KP923879532), T41, VMUL(LDK(KP382683432), T3W)); T4d = VFMA(LDK(KP382683432), T47, VMUL(LDK(KP923879532), T4c)); T4e = VSUB(T42, T4d); T8o = VADD(T42, T4d); T5l = VADD(T4L, T4W); T5m = VADD(T52, T55); T5n = VFNMS(LDK(KP555570233), T5m, VMUL(LDK(KP831469612), T5l)); T5r = VFMA(LDK(KP831469612), T5m, VMUL(LDK(KP555570233), T5l)); } { V T4w, T4F, T5e, T5f; T4w = VSUB(T4k, T4v); T4F = VSUB(T4B, T4E); T4G = VFMA(LDK(KP980785280), T4w, VMUL(LDK(KP195090322), T4F)); T5a = VFNMS(LDK(KP980785280), T4F, VMUL(LDK(KP195090322), T4w)); T5e = VFMA(LDK(KP923879532), T3W, VMUL(LDK(KP382683432), T41)); T5f = VFNMS(LDK(KP923879532), T47, VMUL(LDK(KP382683432), T4c)); T5g = VADD(T5e, T5f); T8w = VSUB(T5f, T5e); } { V T5i, T5j, T4X, T56; T5i = VADD(T4k, T4v); T5j = VADD(T4B, T4E); T5k = VFMA(LDK(KP555570233), T5i, VMUL(LDK(KP831469612), T5j)); T5q = VFNMS(LDK(KP555570233), T5j, VMUL(LDK(KP831469612), T5i)); T4X = VSUB(T4L, T4W); T56 = VSUB(T52, T55); T57 = VFNMS(LDK(KP980785280), T56, VMUL(LDK(KP195090322), T4X)); T5b = VFMA(LDK(KP195090322), T56, VMUL(LDK(KP980785280), T4X)); } { V T4f, T58, T8v, T8y; T4f = VADD(T3R, T4e); T58 = VADD(T4G, T57); ST(&(ri[WS(rs, 23)]), VSUB(T4f, T58), ms, &(ri[WS(rs, 1)])); ST(&(ri[WS(rs, 7)]), VADD(T4f, T58), ms, &(ri[WS(rs, 1)])); T8v = VADD(T5a, T5b); T8y = VADD(T8w, T8x); ST(&(ii[WS(rs, 7)]), VADD(T8v, T8y), ms, &(ii[WS(rs, 1)])); ST(&(ii[WS(rs, 23)]), VSUB(T8y, T8v), ms, &(ii[WS(rs, 1)])); } { V T59, T5c, T8z, T8A; T59 = VSUB(T3R, T4e); T5c = VSUB(T5a, T5b); ST(&(ri[WS(rs, 31)]), VSUB(T59, T5c), ms, &(ri[WS(rs, 1)])); ST(&(ri[WS(rs, 15)]), VADD(T59, T5c), ms, &(ri[WS(rs, 1)])); T8z = VSUB(T57, T4G); T8A = VSUB(T8x, T8w); ST(&(ii[WS(rs, 15)]), VADD(T8z, T8A), ms, &(ii[WS(rs, 1)])); ST(&(ii[WS(rs, 31)]), VSUB(T8A, T8z), ms, &(ii[WS(rs, 1)])); } { V T5h, T5o, T8n, T8s; T5h = VADD(T5d, T5g); T5o = VADD(T5k, T5n); ST(&(ri[WS(rs, 19)]), VSUB(T5h, T5o), ms, &(ri[WS(rs, 1)])); ST(&(ri[WS(rs, 3)]), VADD(T5h, T5o), ms, &(ri[WS(rs, 1)])); T8n = VADD(T5q, T5r); T8s = VADD(T8o, T8r); ST(&(ii[WS(rs, 3)]), VADD(T8n, T8s), ms, &(ii[WS(rs, 1)])); ST(&(ii[WS(rs, 19)]), VSUB(T8s, T8n), ms, &(ii[WS(rs, 1)])); } { V T5p, T5s, T8t, T8u; T5p = VSUB(T5d, T5g); T5s = VSUB(T5q, T5r); ST(&(ri[WS(rs, 27)]), VSUB(T5p, T5s), ms, &(ri[WS(rs, 1)])); ST(&(ri[WS(rs, 11)]), VADD(T5p, T5s), ms, &(ri[WS(rs, 1)])); T8t = VSUB(T5n, T5k); T8u = VSUB(T8r, T8o); ST(&(ii[WS(rs, 11)]), VADD(T8t, T8u), ms, &(ii[WS(rs, 1)])); ST(&(ii[WS(rs, 27)]), VSUB(T8u, T8t), ms, &(ii[WS(rs, 1)])); } } { V T5x, T5Z, T8d, T8j, T5E, T88, T69, T6d, T5M, T5W, T62, T8i, T66, T6c, T5T; V T5X, T5w, T89; T5w = VMUL(LDK(KP707106781), VADD(T5u, T5v)); T5x = VSUB(T5t, T5w); T5Z = VADD(T5t, T5w); T89 = VMUL(LDK(KP707106781), VADD(T3K, T3P)); T8d = VADD(T89, T8c); T8j = VSUB(T8c, T89); { V T5A, T5D, T67, T68; T5A = VFNMS(LDK(KP382683432), T5z, VMUL(LDK(KP923879532), T5y)); T5D = VFMA(LDK(KP923879532), T5B, VMUL(LDK(KP382683432), T5C)); T5E = VSUB(T5A, T5D); T88 = VADD(T5A, T5D); T67 = VADD(T5N, T5O); T68 = VADD(T5Q, T5R); T69 = VFNMS(LDK(KP195090322), T68, VMUL(LDK(KP980785280), T67)); T6d = VFMA(LDK(KP195090322), T67, VMUL(LDK(KP980785280), T68)); } { V T5I, T5L, T60, T61; T5I = VSUB(T5G, T5H); T5L = VSUB(T5J, T5K); T5M = VFMA(LDK(KP555570233), T5I, VMUL(LDK(KP831469612), T5L)); T5W = VFNMS(LDK(KP831469612), T5I, VMUL(LDK(KP555570233), T5L)); T60 = VFMA(LDK(KP382683432), T5y, VMUL(LDK(KP923879532), T5z)); T61 = VFNMS(LDK(KP382683432), T5B, VMUL(LDK(KP923879532), T5C)); T62 = VADD(T60, T61); T8i = VSUB(T61, T60); } { V T64, T65, T5P, T5S; T64 = VADD(T5G, T5H); T65 = VADD(T5J, T5K); T66 = VFMA(LDK(KP980785280), T64, VMUL(LDK(KP195090322), T65)); T6c = VFNMS(LDK(KP195090322), T64, VMUL(LDK(KP980785280), T65)); T5P = VSUB(T5N, T5O); T5S = VSUB(T5Q, T5R); T5T = VFNMS(LDK(KP831469612), T5S, VMUL(LDK(KP555570233), T5P)); T5X = VFMA(LDK(KP831469612), T5P, VMUL(LDK(KP555570233), T5S)); } { V T5F, T5U, T8h, T8k; T5F = VADD(T5x, T5E); T5U = VADD(T5M, T5T); ST(&(ri[WS(rs, 21)]), VSUB(T5F, T5U), ms, &(ri[WS(rs, 1)])); ST(&(ri[WS(rs, 5)]), VADD(T5F, T5U), ms, &(ri[WS(rs, 1)])); T8h = VADD(T5W, T5X); T8k = VADD(T8i, T8j); ST(&(ii[WS(rs, 5)]), VADD(T8h, T8k), ms, &(ii[WS(rs, 1)])); ST(&(ii[WS(rs, 21)]), VSUB(T8k, T8h), ms, &(ii[WS(rs, 1)])); } { V T5V, T5Y, T8l, T8m; T5V = VSUB(T5x, T5E); T5Y = VSUB(T5W, T5X); ST(&(ri[WS(rs, 29)]), VSUB(T5V, T5Y), ms, &(ri[WS(rs, 1)])); ST(&(ri[WS(rs, 13)]), VADD(T5V, T5Y), ms, &(ri[WS(rs, 1)])); T8l = VSUB(T5T, T5M); T8m = VSUB(T8j, T8i); ST(&(ii[WS(rs, 13)]), VADD(T8l, T8m), ms, &(ii[WS(rs, 1)])); ST(&(ii[WS(rs, 29)]), VSUB(T8m, T8l), ms, &(ii[WS(rs, 1)])); } { V T63, T6a, T87, T8e; T63 = VADD(T5Z, T62); T6a = VADD(T66, T69); ST(&(ri[WS(rs, 17)]), VSUB(T63, T6a), ms, &(ri[WS(rs, 1)])); ST(&(ri[WS(rs, 1)]), VADD(T63, T6a), ms, &(ri[WS(rs, 1)])); T87 = VADD(T6c, T6d); T8e = VADD(T88, T8d); ST(&(ii[WS(rs, 1)]), VADD(T87, T8e), ms, &(ii[WS(rs, 1)])); ST(&(ii[WS(rs, 17)]), VSUB(T8e, T87), ms, &(ii[WS(rs, 1)])); } { V T6b, T6e, T8f, T8g; T6b = VSUB(T5Z, T62); T6e = VSUB(T6c, T6d); ST(&(ri[WS(rs, 25)]), VSUB(T6b, T6e), ms, &(ri[WS(rs, 1)])); ST(&(ri[WS(rs, 9)]), VADD(T6b, T6e), ms, &(ri[WS(rs, 1)])); T8f = VSUB(T69, T66); T8g = VSUB(T8d, T88); ST(&(ii[WS(rs, 9)]), VADD(T8f, T8g), ms, &(ii[WS(rs, 1)])); ST(&(ii[WS(rs, 25)]), VSUB(T8g, T8f), ms, &(ii[WS(rs, 1)])); } } } }}static const tw_instr twinstr[] = { VTW(0, 1), VTW(0, 3), VTW(0, 9), VTW(0, 27), {TW_NEXT, (2 * VL), 0}};static const ct_desc desc = { 32, "t2sv_32", twinstr, &GENUS, {376, 168, 112, 0}, 0, 0, 0 };void X(codelet_t2sv_32) (planner *p) { X(kdft_dit_register) (p, t2sv_32, &desc);}#endif /* HAVE_FMA */
⌨️ 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 + -