Chris@42: /* Chris@42: * Copyright (c) 2003, 2007-11 Matteo Frigo Chris@42: * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology Chris@42: * Chris@42: * Generic256d added by Romain Dolbeau, and turned into simd-generic256.h Chris@42: * with single & double precision by Erik Lindahl. Chris@42: * Romain Dolbeau hereby places his modifications in the public domain. Chris@42: * Erik Lindahl hereby places his modifications in the public domain. Chris@42: * Chris@42: * This program is free software; you can redistribute it and/or modify Chris@42: * it under the terms of the GNU General Public License as published by Chris@42: * the Free Software Foundation; either version 2 of the License, or Chris@42: * (at your option) any later version. Chris@42: * Chris@42: * This program is distributed in the hope that it will be useful, Chris@42: * but WITHOUT ANY WARRANTY; without even the implied warranty of Chris@42: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the Chris@42: * GNU General Public License for more details. Chris@42: * Chris@42: * You should have received a copy of the GNU General Public License Chris@42: * along with this program; if not, write to the Free Software Chris@42: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA Chris@42: * Chris@42: */ Chris@42: Chris@42: #if defined(FFTW_LDOUBLE) || defined(FFTW_QUAD) Chris@42: # error "Generic simd256 only works in single or double precision" Chris@42: #endif Chris@42: Chris@42: #define SIMD_SUFFIX _generic_simd256 /* for renaming */ Chris@42: Chris@42: #ifdef FFTW_SINGLE Chris@42: # define DS(d,s) s /* single-precision option */ Chris@42: # define VDUPL(x) {x[0],x[0],x[2],x[2],x[4],x[4],x[6],x[6]} Chris@42: # define VDUPH(x) {x[1],x[1],x[3],x[3],x[5],x[5],x[7],x[7]} Chris@42: # define DVK(var, val) V var = {val,val,val,val,val,val,val,val} Chris@42: #else Chris@42: # define DS(d,s) d /* double-precision option */ Chris@42: # define VDUPL(x) {x[0],x[0],x[2],x[2]} Chris@42: # define VDUPH(x) {x[1],x[1],x[3],x[3]} Chris@42: # define DVK(var, val) V var = {val, val, val, val} Chris@42: #endif Chris@42: Chris@42: #define VL DS(2,4) /* SIMD vector length, in term of complex numbers */ Chris@42: #define SIMD_VSTRIDE_OKA(x) DS(1,((x) == 2)) Chris@42: #define SIMD_STRIDE_OKPAIR SIMD_STRIDE_OK Chris@42: Chris@42: typedef DS(double,float) V __attribute__ ((vector_size(32))); Chris@42: Chris@42: #define VADD(a,b) ((a)+(b)) Chris@42: #define VSUB(a,b) ((a)-(b)) Chris@42: #define VMUL(a,b) ((a)*(b)) Chris@42: Chris@42: #define LDK(x) x Chris@42: Chris@42: static inline V LDA(const R *x, INT ivs, const R *aligned_like) Chris@42: { Chris@42: V var; Chris@42: (void)aligned_like; /* UNUSED */ Chris@42: return *(const V *)x; Chris@42: } Chris@42: Chris@42: static inline void STA(R *x, V v, INT ovs, const R *aligned_like) Chris@42: { Chris@42: (void)aligned_like; /* UNUSED */ Chris@42: (void)ovs; /* UNUSED */ Chris@42: *(V *)x = v; Chris@42: } Chris@42: Chris@42: static inline V LD(const R *x, INT ivs, const R *aligned_like) Chris@42: { Chris@42: V var; Chris@42: (void)aligned_like; /* UNUSED */ Chris@42: var[0] = x[0]; Chris@42: var[1] = x[1]; Chris@42: var[2] = x[ivs]; Chris@42: var[3] = x[ivs+1]; Chris@42: #ifdef FFTW_SINGLE Chris@42: var[4] = x[2*ivs]; Chris@42: var[5] = x[2*ivs+1]; Chris@42: var[6] = x[3*ivs]; Chris@42: var[7] = x[3*ivs+1]; Chris@42: #endif Chris@42: return var; Chris@42: } Chris@42: Chris@42: Chris@42: /* ST has to be separate due to the storage hack requiring reverse order */ Chris@42: Chris@42: static inline void ST(R *x, V v, INT ovs, const R *aligned_like) Chris@42: { Chris@42: (void)aligned_like; /* UNUSED */ Chris@42: #ifdef FFTW_SINGLE Chris@42: *(x + 3*ovs ) = v[6]; Chris@42: *(x + 3*ovs + 1) = v[7]; Chris@42: *(x + 2*ovs ) = v[4]; Chris@42: *(x + 2*ovs + 1) = v[5]; Chris@42: *(x + ovs ) = v[2]; Chris@42: *(x + ovs + 1) = v[3]; Chris@42: *(x ) = v[0]; Chris@42: *(x + 1) = v[1]; Chris@42: #else Chris@42: *(x + ovs ) = v[2]; Chris@42: *(x + ovs + 1) = v[3]; Chris@42: *(x ) = v[0]; Chris@42: *(x + 1) = v[1]; Chris@42: #endif Chris@42: } Chris@42: Chris@42: #ifdef FFTW_SINGLE Chris@42: #define STM2(x, v, ovs, a) /* no-op */ Chris@42: static inline void STN2(R *x, V v0, V v1, INT ovs) Chris@42: { Chris@42: x[ 0] = v0[0]; Chris@42: x[ 1] = v0[1]; Chris@42: x[ 2] = v1[0]; Chris@42: x[ 3] = v1[1]; Chris@42: x[ ovs ] = v0[2]; Chris@42: x[ ovs + 1] = v0[3]; Chris@42: x[ ovs + 2] = v1[2]; Chris@42: x[ ovs + 3] = v1[3]; Chris@42: x[2*ovs ] = v0[4]; Chris@42: x[2*ovs + 1] = v0[5]; Chris@42: x[2*ovs + 2] = v1[4]; Chris@42: x[2*ovs + 3] = v1[5]; Chris@42: x[3*ovs ] = v0[6]; Chris@42: x[3*ovs + 1] = v0[7]; Chris@42: x[3*ovs + 2] = v1[6]; Chris@42: x[3*ovs + 3] = v1[7]; Chris@42: } Chris@42: Chris@42: # define STM4(x, v, ovs, aligned_like) /* no-op */ Chris@42: static inline void STN4(R *x, V v0, V v1, V v2, V v3, INT ovs) Chris@42: { Chris@42: *(x ) = v0[0]; Chris@42: *(x + 1) = v1[0]; Chris@42: *(x + 2) = v2[0]; Chris@42: *(x + 3) = v3[0]; Chris@42: *(x + ovs ) = v0[1]; Chris@42: *(x + ovs + 1) = v1[1]; Chris@42: *(x + ovs + 2) = v2[1]; Chris@42: *(x + ovs + 3) = v3[1]; Chris@42: *(x + 2 * ovs ) = v0[2]; Chris@42: *(x + 2 * ovs + 1) = v1[2]; Chris@42: *(x + 2 * ovs + 2) = v2[2]; Chris@42: *(x + 2 * ovs + 3) = v3[2]; Chris@42: *(x + 3 * ovs ) = v0[3]; Chris@42: *(x + 3 * ovs + 1) = v1[3]; Chris@42: *(x + 3 * ovs + 2) = v2[3]; Chris@42: *(x + 3 * ovs + 3) = v3[3]; Chris@42: *(x + 4 * ovs ) = v0[4]; Chris@42: *(x + 4 * ovs + 1) = v1[4]; Chris@42: *(x + 4 * ovs + 2) = v2[4]; Chris@42: *(x + 4 * ovs + 3) = v3[4]; Chris@42: *(x + 5 * ovs ) = v0[5]; Chris@42: *(x + 5 * ovs + 1) = v1[5]; Chris@42: *(x + 5 * ovs + 2) = v2[5]; Chris@42: *(x + 5 * ovs + 3) = v3[5]; Chris@42: *(x + 6 * ovs ) = v0[6]; Chris@42: *(x + 6 * ovs + 1) = v1[6]; Chris@42: *(x + 6 * ovs + 2) = v2[6]; Chris@42: *(x + 6 * ovs + 3) = v3[6]; Chris@42: *(x + 7 * ovs ) = v0[7]; Chris@42: *(x + 7 * ovs + 1) = v1[7]; Chris@42: *(x + 7 * ovs + 2) = v2[7]; Chris@42: *(x + 7 * ovs + 3) = v3[7]; Chris@42: } Chris@42: Chris@42: #else Chris@42: /* FFTW_DOUBLE */ Chris@42: Chris@42: #define STM2 ST Chris@42: #define STN2(x, v0, v1, ovs) /* nop */ Chris@42: #define STM4(x, v, ovs, aligned_like) /* no-op */ Chris@42: Chris@42: static inline void STN4(R *x, V v0, V v1, V v2, V v3, INT ovs) { Chris@42: *(x ) = v0[0]; Chris@42: *(x + 1) = v1[0]; Chris@42: *(x + 2) = v2[0]; Chris@42: *(x + 3) = v3[0]; Chris@42: *(x + ovs ) = v0[1]; Chris@42: *(x + ovs + 1) = v1[1]; Chris@42: *(x + ovs + 2) = v2[1]; Chris@42: *(x + ovs + 3) = v3[1]; Chris@42: *(x + 2 * ovs ) = v0[2]; Chris@42: *(x + 2 * ovs + 1) = v1[2]; Chris@42: *(x + 2 * ovs + 2) = v2[2]; Chris@42: *(x + 2 * ovs + 3) = v3[2]; Chris@42: *(x + 3 * ovs ) = v0[3]; Chris@42: *(x + 3 * ovs + 1) = v1[3]; Chris@42: *(x + 3 * ovs + 2) = v2[3]; Chris@42: *(x + 3 * ovs + 3) = v3[3]; Chris@42: } Chris@42: #endif Chris@42: Chris@42: static inline V FLIP_RI(V x) Chris@42: { Chris@42: #ifdef FFTW_SINGLE Chris@42: return (V){x[1],x[0],x[3],x[2],x[5],x[4],x[7],x[6]}; Chris@42: #else Chris@42: return (V){x[1],x[0],x[3],x[2]}; Chris@42: #endif Chris@42: } Chris@42: Chris@42: static inline V VCONJ(V x) Chris@42: { Chris@42: #ifdef FFTW_SINGLE Chris@42: return (x * (V){1.0,-1.0,1.0,-1.0,1.0,-1.0,1.0,-1.0}); Chris@42: #else Chris@42: return (x * (V){1.0,-1.0,1.0,-1.0}); Chris@42: #endif Chris@42: } Chris@42: Chris@42: static inline V VBYI(V x) Chris@42: { Chris@42: return FLIP_RI(VCONJ(x)); Chris@42: } Chris@42: Chris@42: /* FMA support */ Chris@42: #define VFMA(a, b, c) VADD(c, VMUL(a, b)) Chris@42: #define VFNMS(a, b, c) VSUB(c, VMUL(a, b)) Chris@42: #define VFMS(a, b, c) VSUB(VMUL(a, b), c) Chris@42: #define VFMAI(b, c) VADD(c, VBYI(b)) Chris@42: #define VFNMSI(b, c) VSUB(c, VBYI(b)) Chris@42: #define VFMACONJ(b,c) VADD(VCONJ(b),c) Chris@42: #define VFMSCONJ(b,c) VSUB(VCONJ(b),c) Chris@42: #define VFNMSCONJ(b,c) VSUB(c, VCONJ(b)) Chris@42: Chris@42: static inline V VZMUL(V tx, V sr) Chris@42: { Chris@42: V tr = VDUPL(tx); Chris@42: V ti = VDUPH(tx); Chris@42: tr = VMUL(sr, tr); Chris@42: sr = VBYI(sr); Chris@42: return VFMA(ti, sr, tr); Chris@42: } Chris@42: Chris@42: static inline V VZMULJ(V tx, V sr) Chris@42: { Chris@42: V tr = VDUPL(tx); Chris@42: V ti = VDUPH(tx); Chris@42: tr = VMUL(sr, tr); Chris@42: sr = VBYI(sr); Chris@42: return VFNMS(ti, sr, tr); Chris@42: } Chris@42: Chris@42: static inline V VZMULI(V tx, V sr) Chris@42: { Chris@42: V tr = VDUPL(tx); Chris@42: V ti = VDUPH(tx); Chris@42: ti = VMUL(ti, sr); Chris@42: sr = VBYI(sr); Chris@42: return VFMS(tr, sr, ti); Chris@42: } Chris@42: Chris@42: static inline V VZMULIJ(V tx, V sr) Chris@42: { Chris@42: V tr = VDUPL(tx); Chris@42: V ti = VDUPH(tx); Chris@42: ti = VMUL(ti, sr); Chris@42: sr = VBYI(sr); Chris@42: return VFMA(tr, sr, ti); Chris@42: } Chris@42: Chris@42: /* twiddle storage #1: compact, slower */ Chris@42: #ifdef FFTW_SINGLE Chris@42: # define VTW1(v,x) {TW_CEXP, v, x}, {TW_CEXP, v+1, x}, {TW_CEXP, v+2, x}, {TW_CEXP, v+3, x} Chris@42: #else Chris@42: # define VTW1(v,x) {TW_CEXP, v, x}, {TW_CEXP, v+1, x} Chris@42: #endif Chris@42: #define TWVL1 (VL) Chris@42: Chris@42: static inline V BYTW1(const R *t, V sr) Chris@42: { Chris@42: return VZMUL(LDA(t, 2, t), sr); Chris@42: } Chris@42: Chris@42: static inline V BYTWJ1(const R *t, V sr) Chris@42: { Chris@42: return VZMULJ(LDA(t, 2, t), sr); Chris@42: } Chris@42: Chris@42: /* twiddle storage #2: twice the space, faster (when in cache) */ Chris@42: #ifdef FFTW_SINGLE Chris@42: # define VTW2(v,x) \ Chris@42: {TW_COS, v, x}, {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_COS, v+1, x}, \ Chris@42: {TW_COS, v+2, x}, {TW_COS, v+2, x}, {TW_COS, v+3, x}, {TW_COS, v+3, x}, \ Chris@42: {TW_SIN, v, -x}, {TW_SIN, v, x}, {TW_SIN, v+1, -x}, {TW_SIN, v+1, x}, \ Chris@42: {TW_SIN, v+2, -x}, {TW_SIN, v+2, x}, {TW_SIN, v+3, -x}, {TW_SIN, v+3, x} Chris@42: #else Chris@42: # define VTW2(v,x) \ Chris@42: {TW_COS, v, x}, {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_COS, v+1, x}, \ Chris@42: {TW_SIN, v, -x}, {TW_SIN, v, x}, {TW_SIN, v+1, -x}, {TW_SIN, v+1, x} Chris@42: #endif Chris@42: #define TWVL2 (2 * VL) Chris@42: Chris@42: static inline V BYTW2(const R *t, V sr) Chris@42: { Chris@42: const V *twp = (const V *)t; Chris@42: V si = FLIP_RI(sr); Chris@42: V tr = twp[0], ti = twp[1]; Chris@42: return VFMA(tr, sr, VMUL(ti, si)); Chris@42: } Chris@42: Chris@42: static inline V BYTWJ2(const R *t, V sr) Chris@42: { Chris@42: const V *twp = (const V *)t; Chris@42: V si = FLIP_RI(sr); Chris@42: V tr = twp[0], ti = twp[1]; Chris@42: return VFNMS(ti, si, VMUL(tr, sr)); Chris@42: } Chris@42: Chris@42: /* twiddle storage #3 */ Chris@42: #define VTW3 VTW1 Chris@42: #define TWVL3 TWVL1 Chris@42: Chris@42: /* twiddle storage for split arrays */ Chris@42: #ifdef FFTW_SINGLE Chris@42: # define VTWS(v,x) \ Chris@42: {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_COS, v+2, x}, {TW_COS, v+3, x}, \ Chris@42: {TW_COS, v+4, x}, {TW_COS, v+5, x}, {TW_COS, v+6, x}, {TW_COS, v+7, x}, \ Chris@42: {TW_SIN, v, x}, {TW_SIN, v+1, x}, {TW_SIN, v+2, x}, {TW_SIN, v+3, x}, \ Chris@42: {TW_SIN, v+4, x}, {TW_SIN, v+5, x}, {TW_SIN, v+6, x}, {TW_SIN, v+7, x} Chris@42: #else Chris@42: # define VTWS(v,x) \ Chris@42: {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_COS, v+2, x}, {TW_COS, v+3, x}, \ Chris@42: {TW_SIN, v, x}, {TW_SIN, v+1, x}, {TW_SIN, v+2, x}, {TW_SIN, v+3, x} Chris@42: #endif Chris@42: #define TWVLS (2 * VL) Chris@42: Chris@42: #define VLEAVE() /* nothing */ Chris@42: Chris@42: #include "simd-common.h"