Chris@42: /*
Chris@42:  * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42:  *
Chris@42:  * 128-bit AVX support by Erik Lindahl, 2015.
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 "AVX only works in single or double precision"
Chris@42: #endif
Chris@42: 
Chris@42: #ifdef FFTW_SINGLE
Chris@42: #  define DS(d,s) s /* single-precision option */
Chris@42: #  define SUFF(name) name ## s
Chris@42: #else
Chris@42: #  define DS(d,s) d /* double-precision option */
Chris@42: #  define SUFF(name) name ## d
Chris@42: #endif
Chris@42: 
Chris@42: #define SIMD_SUFFIX  _avx_128_fma  /* for renaming */
Chris@42: #define VL DS(1,2)         /* 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: #ifdef _MSC_VER
Chris@42: #ifndef inline
Chris@42: #define inline __inline
Chris@42: #endif
Chris@42: #endif
Chris@42: 
Chris@42: #include <immintrin.h>
Chris@42: #ifdef _MSC_VER
Chris@42: #    include <intrin.h>
Chris@42: #elif defined (__GNUC__)
Chris@42: #    include <x86intrin.h>
Chris@42: #endif
Chris@42: 
Chris@42: #if !(defined(__AVX__) && defined(__FMA4__)) /* sanity check */
Chris@42: #error "compiling simd-avx-128-fma.h without -mavx or -mfma4"
Chris@42: #endif
Chris@42: 
Chris@42: typedef DS(__m128d,__m128) V;
Chris@42: #define VADD SUFF(_mm_add_p)
Chris@42: #define VSUB SUFF(_mm_sub_p)
Chris@42: #define VMUL SUFF(_mm_mul_p)
Chris@42: #define VXOR SUFF(_mm_xor_p)
Chris@42: #define SHUF SUFF(_mm_shuffle_p)
Chris@42: #define VPERM1 SUFF(_mm_permute_p)
Chris@42: #define UNPCKL SUFF(_mm_unpacklo_p)
Chris@42: #define UNPCKH SUFF(_mm_unpackhi_p)
Chris@42: 
Chris@42: #define SHUFVALS(fp0,fp1,fp2,fp3) \
Chris@42:    (((fp3) << 6) | ((fp2) << 4) | ((fp1) << 2) | ((fp0)))
Chris@42: 
Chris@42: #define VDUPL(x) DS(_mm_permute_pd(x,0), _mm_moveldup_ps(x))
Chris@42: #define VDUPH(x) DS(_mm_permute_pd(x,3), _mm_movehdup_ps(x))
Chris@42: #define LOADH(addr, val) _mm_loadh_pi(val, (const __m64 *)(addr))
Chris@42: #define LOADL(addr, val) _mm_loadl_pi(val, (const __m64 *)(addr))
Chris@42: #define STOREH(a, v) DS(_mm_storeh_pd(a, v), _mm_storeh_pi((__m64 *)(a), v))
Chris@42: #define STOREL(a, v) DS(_mm_storel_pd(a, v), _mm_storel_pi((__m64 *)(a), v))
Chris@42: 
Chris@42: #define VLIT(x0, x1) DS(_mm_set_pd(x0, x1), _mm_set_ps(x0, x1, x0, x1))
Chris@42: #define DVK(var, val) V var = VLIT(val, val)
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:      (void)aligned_like; /* UNUSED */
Chris@42:      (void)ivs; /* 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: #ifdef FFTW_SINGLE
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: #if defined(__ICC) || (__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ > 8)
Chris@42:     var = LOADL(x, SUFF(_mm_undefined_p)());
Chris@42:     var = LOADH(x + ivs, var);
Chris@42: #else
Chris@42:     var = LOADL(x, var);
Chris@42:     var = LOADH(x + ivs, var);
Chris@42: #endif
Chris@42:     return var;
Chris@42: }
Chris@42: 
Chris@42: #  ifdef _MSC_VER
Chris@42: #    pragma warning(default : 4700)
Chris@42: #    pragma runtime_checks("u", restore)
Chris@42: #  endif
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:      /* WARNING: the extra_iter hack depends upon STOREL occurring
Chris@42: 	after STOREH */
Chris@42:      STOREH(x + ovs, v);
Chris@42:      STOREL(x, v);
Chris@42: }
Chris@42: 
Chris@42: #else /* ! FFTW_SINGLE */
Chris@42: #  define LD LDA
Chris@42: #  define ST STA
Chris@42: #endif
Chris@42: 
Chris@42: #define STM2 DS(STA,ST)
Chris@42: #define STN2(x, v0, v1, ovs) /* nop */
Chris@42: 
Chris@42: #ifdef FFTW_SINGLE
Chris@42: #  define STM4(x, v, ovs, aligned_like) /* no-op */
Chris@42: /* STN4 is a macro, not a function, thanks to Visual C++ developers
Chris@42:    deciding "it would be infrequent that people would want to pass more
Chris@42:    than 3 [__m128 parameters] by value."  3 parameters ought to be enough
Chris@42:    for anybody. */
Chris@42: #  define STN4(x, v0, v1, v2, v3, ovs)			\
Chris@42: {							\
Chris@42:      V xxx0, xxx1, xxx2, xxx3;				\
Chris@42:      xxx0 = UNPCKL(v0, v2);				\
Chris@42:      xxx1 = UNPCKH(v0, v2);				\
Chris@42:      xxx2 = UNPCKL(v1, v3);				\
Chris@42:      xxx3 = UNPCKH(v1, v3);				\
Chris@42:      STA(x, UNPCKL(xxx0, xxx2), 0, 0);			\
Chris@42:      STA(x + ovs, UNPCKH(xxx0, xxx2), 0, 0);		\
Chris@42:      STA(x + 2 * ovs, UNPCKL(xxx1, xxx3), 0, 0);	\
Chris@42:      STA(x + 3 * ovs, UNPCKH(xxx1, xxx3), 0, 0);	\
Chris@42: }
Chris@42: #else /* !FFTW_SINGLE */
Chris@42: static inline void STM4(R *x, V v, INT ovs, const R *aligned_like)
Chris@42: {
Chris@42:      (void)aligned_like; /* UNUSED */
Chris@42:      STOREL(x, v);
Chris@42:      STOREH(x + ovs, v);
Chris@42: }
Chris@42: #  define STN4(x, v0, v1, v2, v3, ovs) /* nothing */
Chris@42: #endif
Chris@42: 
Chris@42: static inline V FLIP_RI(V x)
Chris@42: {
Chris@42:   return VPERM1(x, DS(1, SHUFVALS(1, 0, 3, 2)));
Chris@42: }
Chris@42: 
Chris@42: 
Chris@42: static inline V VCONJ(V x)
Chris@42: {
Chris@42:   V pmpm = VLIT(-0.0, 0.0);
Chris@42:   return VXOR(pmpm, x);
Chris@42: }
Chris@42: 
Chris@42: static inline V VBYI(V x)
Chris@42: {
Chris@42:      x = VCONJ(x);
Chris@42:      x = FLIP_RI(x);
Chris@42:      return x;
Chris@42: }
Chris@42: 
Chris@42: /* FMA support */
Chris@42: #define VFMA(a, b, c)  SUFF(_mm_macc_p)(a,b,c)
Chris@42: #define VFNMS(a, b, c) SUFF(_mm_nmacc_p)(a,b,c)
Chris@42: #define VFMS(a, b, c)  SUFF(_mm_msub_p)(a,b,c)
Chris@42: #define VFMAI(b, c)  SUFF(_mm_addsub_p)(c,FLIP_RI(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) SUFF(_mm_addsub_p)(c,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(tr, sr);
Chris@42:     ti = VMUL(ti, FLIP_RI(sr));
Chris@42:     return SUFF(_mm_addsub_p)(tr,ti);
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(tr, sr);
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:     tr = VMUL(tr, FLIP_RI(sr));
Chris@42:     return SUFF(_mm_addsub_p)(ti,tr);
Chris@42: }
Chris@42: 
Chris@42: /* twiddle storage #1: compact, slower */
Chris@42: #ifdef FFTW_SINGLE
Chris@42: #  define VTW1(v,x)  \
Chris@42:   {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_SIN, v, x}, {TW_SIN, v+1, x}
Chris@42: static inline V BYTW1(const R *t, V sr)
Chris@42: {
Chris@42:     const V *twp = (const V *)t;
Chris@42:     V tx = twp[0];
Chris@42:     V tr = UNPCKL(tx, tx);
Chris@42:     V ti = UNPCKH(tx, tx);
Chris@42:     tr = VMUL(tr, sr);
Chris@42:     ti = VMUL(ti, FLIP_RI(sr));
Chris@42:     return SUFF(_mm_addsub_p)(tr,ti);
Chris@42: }
Chris@42: static inline V BYTWJ1(const R *t, V sr)
Chris@42: {
Chris@42:     const V *twp = (const V *)t;
Chris@42:     V tx = twp[0];
Chris@42:     V tr = UNPCKL(tx, tx);
Chris@42:     V ti = UNPCKH(tx, tx);
Chris@42:     tr = VMUL(tr, sr);
Chris@42:     sr = VBYI(sr);
Chris@42:     return VFNMS(ti, sr, tr);
Chris@42: }
Chris@42: #else /* !FFTW_SINGLE */
Chris@42: #  define VTW1(v,x) {TW_CEXP, v, x}
Chris@42: static inline V BYTW1(const R *t, V sr)
Chris@42: {
Chris@42:      V tx = LD(t, 1, t);
Chris@42:      return VZMUL(tx, sr);
Chris@42: }
Chris@42: static inline V BYTWJ1(const R *t, V sr)
Chris@42: {
Chris@42:      V tx = LD(t, 1, t);
Chris@42:      return VZMULJ(tx, sr);
Chris@42: }
Chris@42: #endif
Chris@42: #define TWVL1 (VL)
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_SIN, v, -x}, {TW_SIN, v, x}, {TW_SIN, v+1, -x}, {TW_SIN, v+1, x}
Chris@42: #else /* !FFTW_SINGLE */
Chris@42: #  define VTW2(v,x)							\
Chris@42:   {TW_COS, v, x}, {TW_COS, v, x}, {TW_SIN, v, -x}, {TW_SIN, v, x}
Chris@42: #endif
Chris@42: #define TWVL2 (2 * VL)
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: 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: #ifdef FFTW_SINGLE
Chris@42: #  define VTW3(v,x) {TW_CEXP, v, x}, {TW_CEXP, v+1, x}
Chris@42: #  define TWVL3 (VL)
Chris@42: #else
Chris@42: #  define VTW3(v,x) VTW1(v,x)
Chris@42: #  define TWVL3 TWVL1
Chris@42: #endif
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_SIN, v, x}, {TW_SIN, v+1, x}, {TW_SIN, v+2, x}, {TW_SIN, v+3, x}
Chris@42: #else
Chris@42: #  define VTWS(v,x)							  \
Chris@42:     {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_SIN, v, x}, {TW_SIN, v+1, 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"