Chris@82: /*
Chris@82:  * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@82:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@82:  *
Chris@82:  * VSX SIMD implementation added 2015 Erik Lindahl.
Chris@82:  * Erik Lindahl places his modifications in the public domain.
Chris@82:  *
Chris@82:  * This program is free software; you can redistribute it and/or modify
Chris@82:  * it under the terms of the GNU General Public License as published by
Chris@82:  * the Free Software Foundation; either version 2 of the License, or
Chris@82:  * (at your option) any later version.
Chris@82:  *
Chris@82:  * This program is distributed in the hope that it will be useful,
Chris@82:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@82:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@82:  * GNU General Public License for more details.
Chris@82:  *
Chris@82:  * You should have received a copy of the GNU General Public License
Chris@82:  * along with this program; if not, write to the Free Software
Chris@82:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@82:  *
Chris@82:  */
Chris@82: 
Chris@82: #if defined(FFTW_LDOUBLE) || defined(FFTW_QUAD)
Chris@82: #  error "VSX only works in single or double precision"
Chris@82: #endif
Chris@82: 
Chris@82: #ifdef FFTW_SINGLE
Chris@82: #  define DS(d,s) s /* single-precision option */
Chris@82: #  define SUFF(name) name ## s
Chris@82: #else
Chris@82: #  define DS(d,s) d /* double-precision option */
Chris@82: #  define SUFF(name) name ## d
Chris@82: #endif
Chris@82: 
Chris@82: #define SIMD_SUFFIX  _vsx  /* for renaming */
Chris@82: #define VL DS(1,2)         /* SIMD vector length, in term of complex numbers */
Chris@82: #define SIMD_VSTRIDE_OKA(x) DS(1,((x) == 2))
Chris@82: #define SIMD_STRIDE_OKPAIR SIMD_STRIDE_OK
Chris@82: 
Chris@82: #include <altivec.h>
Chris@82: #include <stdio.h>
Chris@82: 
Chris@82: typedef DS(vector double,vector float) V;
Chris@82: 
Chris@82: #define VADD(a,b)   vec_add(a,b)
Chris@82: #define VSUB(a,b)   vec_sub(a,b)
Chris@82: #define VMUL(a,b)   vec_mul(a,b)
Chris@82: #define VXOR(a,b)   vec_xor(a,b)
Chris@82: #define UNPCKL(a,b) vec_mergel(a,b)
Chris@82: #define UNPCKH(a,b) vec_mergeh(a,b)
Chris@82: #ifdef FFTW_SINGLE
Chris@82: #    define VDUPL(a)    ({ const vector unsigned char perm = {0,1,2,3,0,1,2,3,8,9,10,11,8,9,10,11}; vec_perm(a,a,perm); })
Chris@82: #    define VDUPH(a)    ({ const vector unsigned char perm = {4,5,6,7,4,5,6,7,12,13,14,15,12,13,14,15}; vec_perm(a,a,perm); })
Chris@82: #else
Chris@82: #    define VDUPL(a)    ({ const vector unsigned char perm = {0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7}; vec_perm(a,a,perm); })
Chris@82: #    define VDUPH(a)    ({ const vector unsigned char perm = {8,9,10,11,12,13,14,15,8,9,10,11,12,13,14,15}; vec_perm(a,a,perm); })
Chris@82: #endif
Chris@82: 
Chris@82: static inline V LDK(R f) { return vec_splats(f); }
Chris@82: 
Chris@82: #define DVK(var, val) const R var = K(val)
Chris@82: 
Chris@82: static inline V VCONJ(V x)
Chris@82: {
Chris@82:   const V pmpm = vec_mergel(vec_splats((R)0.0),-(vec_splats((R)0.0)));
Chris@82:   return vec_xor(x, pmpm);
Chris@82: }
Chris@82: 
Chris@82: static inline V LDA(const R *x, INT ivs, const R *aligned_like)
Chris@82: {
Chris@82: #ifdef __ibmxl__
Chris@82:   return vec_xl(0,(DS(double,float) *)x);
Chris@82: #else
Chris@82:   return (*(const V *)(x));
Chris@82: #endif
Chris@82: }
Chris@82: 
Chris@82: static inline void STA(R *x, V v, INT ovs, const R *aligned_like)
Chris@82: {
Chris@82: #ifdef __ibmxl__
Chris@82:   vec_xst(v,0,x);
Chris@82: #else
Chris@82:   *(V *)x = v;
Chris@82: #endif
Chris@82: }
Chris@82: 
Chris@82: static inline V FLIP_RI(V x)
Chris@82: {
Chris@82: #ifdef FFTW_SINGLE
Chris@82:   const vector unsigned char perm = { 4,5,6,7,0,1,2,3,12,13,14,15,8,9,10,11 };
Chris@82: #else
Chris@82:   const vector unsigned char perm = { 8,9,10,11,12,13,14,15,0,1,2,3,4,5,6,7 };
Chris@82: #endif
Chris@82:   return vec_perm(x,x,perm);
Chris@82: }
Chris@82: 
Chris@82: #ifdef FFTW_SINGLE
Chris@82: 
Chris@82: static inline V LD(const R *x, INT ivs, const R *aligned_like)
Chris@82: {
Chris@82:   const vector unsigned char perm = {0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23};
Chris@82: 
Chris@82:   return vec_perm((vector float)vec_splats(*(double *)(x)),
Chris@82: 		  (vector float)vec_splats(*(double *)(x+ivs)),perm);
Chris@82: }
Chris@82: 
Chris@82: static inline void ST(R *x, V v, INT ovs, const R *aligned_like)
Chris@82: {
Chris@82:   *(double *)(x+ovs) = vec_extract( (vector double)v, 1 );
Chris@82:   *(double *)x       = vec_extract( (vector double)v, 0 );
Chris@82: }
Chris@82: #else
Chris@82: /* DOUBLE */
Chris@82: 
Chris@82: #  define LD LDA
Chris@82: #  define ST STA
Chris@82: 
Chris@82: #endif
Chris@82: 
Chris@82: #define STM2 DS(STA,ST)
Chris@82: #define STN2(x, v0, v1, ovs) /* nop */
Chris@82: 
Chris@82: #ifdef FFTW_SINGLE
Chris@82: 
Chris@82: #  define STM4(x, v, ovs, aligned_like) /* no-op */
Chris@82: static inline void STN4(R *x, V v0, V v1, V v2, V v3, int ovs)
Chris@82: {
Chris@82:     V xxx0, xxx1, xxx2, xxx3;
Chris@82:     xxx0 = vec_mergeh(v0,v1);
Chris@82:     xxx1 = vec_mergel(v0,v1);
Chris@82:     xxx2 = vec_mergeh(v2,v3);
Chris@82:     xxx3 = vec_mergel(v2,v3);
Chris@82:     *(double *)x           = vec_extract( (vector double)xxx0, 0 );
Chris@82:     *(double *)(x+ovs)     = vec_extract( (vector double)xxx0, 1 );
Chris@82:     *(double *)(x+2*ovs)   = vec_extract( (vector double)xxx1, 0 );
Chris@82:     *(double *)(x+3*ovs)   = vec_extract( (vector double)xxx1, 1 );
Chris@82:     *(double *)(x+2)       = vec_extract( (vector double)xxx2, 0 );
Chris@82:     *(double *)(x+ovs+2)   = vec_extract( (vector double)xxx2, 1 );
Chris@82:     *(double *)(x+2*ovs+2) = vec_extract( (vector double)xxx3, 0 );
Chris@82:     *(double *)(x+3*ovs+2) = vec_extract( (vector double)xxx3, 1 );
Chris@82: }
Chris@82: #else /* !FFTW_SINGLE */
Chris@82: 
Chris@82: static inline void STM4(R *x, V v, INT ovs, const R *aligned_like)
Chris@82: {
Chris@82:      (void)aligned_like; /* UNUSED */
Chris@82:      x[0]    = vec_extract(v,0);
Chris@82:      x[ovs]  = vec_extract(v,1);
Chris@82: }
Chris@82: #  define STN4(x, v0, v1, v2, v3, ovs) /* nothing */
Chris@82: #endif
Chris@82: 
Chris@82: static inline V VBYI(V x)
Chris@82: {
Chris@82:      /* FIXME [matteof 2017-09-21] It is possible to use vpermxor(),
Chris@82:         but gcc and xlc treat the permutation bits differently, and
Chris@82:         gcc-6 seems to generate incorrect code when using
Chris@82:         __builtin_crypto_vpermxor() (i.e., VBYI() works for a small
Chris@82:         test case but fails in the large).
Chris@82: 
Chris@82:         Punt on vpermxor() for now and do the simple thing.
Chris@82:      */
Chris@82:      return FLIP_RI(VCONJ(x));
Chris@82: }
Chris@82: 
Chris@82: /* FMA support */
Chris@82: #define VFMA(a, b, c)  vec_madd(a,b,c)
Chris@82: #define VFNMS(a, b, c) vec_nmsub(a,b,c)
Chris@82: #define VFMS(a, b, c)  vec_msub(a,b,c)
Chris@82: #define VFMAI(b, c)    VADD(c, VBYI(b))
Chris@82: #define VFNMSI(b, c)   VSUB(c, VBYI(b))
Chris@82: #define VFMACONJ(b,c)  VADD(VCONJ(b),c)
Chris@82: #define VFMSCONJ(b,c)  VSUB(VCONJ(b),c)
Chris@82: #define VFNMSCONJ(b,c) VSUB(c, VCONJ(b))
Chris@82: 
Chris@82: static inline V VZMUL(V tx, V sr)
Chris@82: {
Chris@82:     V tr = VDUPL(tx);
Chris@82:     V ti = VDUPH(tx);
Chris@82:     tr = VMUL(sr, tr);
Chris@82:     sr = VBYI(sr);
Chris@82:     return VFMA(ti, sr, tr);
Chris@82: }
Chris@82: 
Chris@82: static inline V VZMULJ(V tx, V sr)
Chris@82: {
Chris@82:     V tr = VDUPL(tx);
Chris@82:     V ti = VDUPH(tx);
Chris@82:     tr = VMUL(sr, tr);
Chris@82:     sr = VBYI(sr);
Chris@82:     return VFNMS(ti, sr, tr);
Chris@82: }
Chris@82: 
Chris@82: static inline V VZMULI(V tx, V sr)
Chris@82: {
Chris@82:     V tr = VDUPL(tx);
Chris@82:     V ti = VDUPH(tx);
Chris@82:     ti = VMUL(ti, sr);
Chris@82:     sr = VBYI(sr);
Chris@82:     return VFMS(tr, sr, ti);
Chris@82: }
Chris@82: 
Chris@82: static inline V VZMULIJ(V tx, V sr)
Chris@82: {
Chris@82:     V tr = VDUPL(tx);
Chris@82:     V ti = VDUPH(tx);
Chris@82:     ti = VMUL(ti, sr);
Chris@82:     sr = VBYI(sr);
Chris@82:     return VFMA(tr, sr, ti);
Chris@82: }
Chris@82: 
Chris@82: /* twiddle storage #1: compact, slower */
Chris@82: #ifdef FFTW_SINGLE
Chris@82: #  define VTW1(v,x)  \
Chris@82:   {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_SIN, v, x}, {TW_SIN, v+1, x}
Chris@82: static inline V BYTW1(const R *t, V sr)
Chris@82: {
Chris@82:      V tx = LDA(t,0,t);
Chris@82:      V tr = UNPCKH(tx, tx);
Chris@82:      V ti = UNPCKL(tx, tx);
Chris@82:      tr = VMUL(tr, sr);
Chris@82:      sr = VBYI(sr);
Chris@82:      return VFMA(ti, sr, tr);
Chris@82: }
Chris@82: static inline V BYTWJ1(const R *t, V sr)
Chris@82: {
Chris@82:      V tx = LDA(t,0,t);
Chris@82:      V tr = UNPCKH(tx, tx);
Chris@82:      V ti = UNPCKL(tx, tx);
Chris@82:      tr = VMUL(tr, sr);
Chris@82:      sr = VBYI(sr);
Chris@82:      return VFNMS(ti, sr, tr);
Chris@82: }
Chris@82: #else /* !FFTW_SINGLE */
Chris@82: #  define VTW1(v,x) {TW_CEXP, v, x}
Chris@82: static inline V BYTW1(const R *t, V sr)
Chris@82: {
Chris@82:      V tx = LD(t, 1, t);
Chris@82:      return VZMUL(tx, sr);
Chris@82: }
Chris@82: static inline V BYTWJ1(const R *t, V sr)
Chris@82: {
Chris@82:      V tx = LD(t, 1, t);
Chris@82:      return VZMULJ(tx, sr);
Chris@82: }
Chris@82: #endif
Chris@82: #define TWVL1 (VL)
Chris@82: 
Chris@82: /* twiddle storage #2: twice the space, faster (when in cache) */
Chris@82: #ifdef FFTW_SINGLE
Chris@82: #  define VTW2(v,x)							\
Chris@82:   {TW_COS, v, x}, {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_COS, v+1, x},	\
Chris@82:   {TW_SIN, v, -x}, {TW_SIN, v, x}, {TW_SIN, v+1, -x}, {TW_SIN, v+1, x}
Chris@82: #else /* !FFTW_SINGLE */
Chris@82: #  define VTW2(v,x)							\
Chris@82:   {TW_COS, v, x}, {TW_COS, v, x}, {TW_SIN, v, -x}, {TW_SIN, v, x}
Chris@82: #endif
Chris@82: #define TWVL2 (2 * VL)
Chris@82: static inline V BYTW2(const R *t, V sr)
Chris@82: {
Chris@82:      V si = FLIP_RI(sr);
Chris@82:      V ti = LDA(t+2*VL,0,t);
Chris@82:      V tt = VMUL(ti, si);
Chris@82:      V tr = LDA(t,0,t);
Chris@82:      return VFMA(tr, sr, tt);
Chris@82: }
Chris@82: static inline V BYTWJ2(const R *t, V sr)
Chris@82: {
Chris@82:      V si = FLIP_RI(sr);
Chris@82:      V tr = LDA(t,0,t);
Chris@82:      V tt = VMUL(tr, sr);
Chris@82:      V ti = LDA(t+2*VL,0,t);
Chris@82:      return VFNMS(ti, si, tt);
Chris@82: }
Chris@82: 
Chris@82: /* twiddle storage #3 */
Chris@82: #ifdef FFTW_SINGLE
Chris@82: #  define VTW3(v,x) {TW_CEXP, v, x}, {TW_CEXP, v+1, x}
Chris@82: #  define TWVL3 (VL)
Chris@82: #else
Chris@82: #  define VTW3(v,x) VTW1(v,x)
Chris@82: #  define TWVL3 TWVL1
Chris@82: #endif
Chris@82: 
Chris@82: /* twiddle storage for split arrays */
Chris@82: #ifdef FFTW_SINGLE
Chris@82: #  define VTWS(v,x)							  \
Chris@82:     {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_COS, v+2, x}, {TW_COS, v+3, x}, \
Chris@82:     {TW_SIN, v, x}, {TW_SIN, v+1, x}, {TW_SIN, v+2, x}, {TW_SIN, v+3, x}
Chris@82: #else
Chris@82: #  define VTWS(v,x)							  \
Chris@82:     {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_SIN, v, x}, {TW_SIN, v+1, x}
Chris@82: #endif
Chris@82: #define TWVLS (2 * VL)
Chris@82: 
Chris@82: #define VLEAVE() /* nothing */
Chris@82: 
Chris@82: #include "simd-common.h"