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:  * AVX-512 support implemented by Romain Dolbeau.
Chris@42:  * Romain Dolbeau hereby places his modifications in the public domain.
Chris@42:  *
Chris@42:  * Permission is hereby granted, free of charge, to any person obtaining a copy
Chris@42:  * of this software and associated documentation files (the "Software"), to deal
Chris@42:  * in the Software without restriction, including without limitation the rights
Chris@42:  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
Chris@42:  * copies of the Software, and to permit persons to whom the Software is
Chris@42:  * furnished to do so, subject to the following conditions:
Chris@42:  *
Chris@42:  * The above copyright notice and this permission notice shall be included in
Chris@42:  * all copies or substantial portions of the Software.
Chris@42:  *
Chris@42:  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
Chris@42:  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
Chris@42:  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
Chris@42:  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
Chris@42:  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
Chris@42:  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
Chris@42:  * THE SOFTWARE.
Chris@42:  *
Chris@42:  */
Chris@42: 
Chris@42: #if defined(FFTW_LDOUBLE) || defined(FFTW_QUAD)
Chris@42: #error "AVX-512 vector instructions 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 ## _ps
Chris@42: #  define SCAL(x) x ## f
Chris@42: #else /* !FFTW_SINGLE */
Chris@42: #  define DS(d,s) d /* double-precision option */
Chris@42: #  define SUFF(name) name ## _pd
Chris@42: #  define SCAL(x) x
Chris@42: #endif /* FFTW_SINGLE */
Chris@42: 
Chris@42: #define SIMD_SUFFIX  _avx512  /* for renaming */
Chris@42: #define VL DS(4, 8)        /* SIMD complex vector length */
Chris@42: #define SIMD_VSTRIDE_OKA(x) ((x) == 2) 
Chris@42: #define SIMD_STRIDE_OKPAIR SIMD_STRIDE_OK
Chris@42: 
Chris@42: #if defined(__GNUC__) && !defined(__AVX512F__) /* sanity check */
Chris@42: #error "compiling simd-avx512.h without avx-512f support"
Chris@42: #endif
Chris@42: 
Chris@42: #if !defined(HAVE_FMA)
Chris@42: #warning "You should probably enable FMAs with --enable-fma for AVX-512"
Chris@42: #endif
Chris@42: 
Chris@42: #if !defined(HAVE_AVX2)
Chris@42: #warning "You should probably enable AVX2 with --enable-avx2 for AVX-512"
Chris@42: #endif
Chris@42: 
Chris@42: #include <immintrin.h>
Chris@42: 
Chris@42: typedef DS(__m512d, __m512) V;
Chris@42: 
Chris@42: #define VLIT(re, im) DS(SUFF(_mm512_setr)(im, re, im, re, im, re, im, re),SUFF(_mm512_setr)(im, re, im, re, im, re, im, re, im, re, im, re, im, re, im, re))
Chris@42: #define VLIT1(val) SUFF(_mm512_set1)(val)
Chris@42: #define LDK(x) x
Chris@42: #define DVK(var, val) V var = VLIT1(val)
Chris@42: #define VZERO SUFF(_mm512_setzero)()
Chris@42: 
Chris@42: #define VDUPL(x) DS(_mm512_movedup_pd(x),_mm512_moveldup_ps(x))
Chris@42: #define VDUPH(x) DS(_mm512_unpackhi_pd(x, x),_mm512_movehdup_ps(x))
Chris@42: #define FLIP_RI(x) SUFF(_mm512_shuffle)(x, x, DS(0x55,0xB1))
Chris@42: #define VCONJ(x) SUFF(_mm512_fmsubadd)(VZERO, VZERO, x)
Chris@42: static inline V VBYI(V x)
Chris@42: {
Chris@42:      return FLIP_RI(VCONJ(x));
Chris@42: }
Chris@42: 
Chris@42: #define VADD(a,b) SUFF(_mm512_add)(a,b)
Chris@42: #define VSUB(a,b) SUFF(_mm512_sub)(a,b)
Chris@42: #define VMUL(a,b) SUFF(_mm512_mul)(a,b)
Chris@42: #define VFMA(a, b, c)  SUFF(_mm512_fmadd)(a, b, c)
Chris@42: #define VFMS(a, b, c)  SUFF(_mm512_fmsub)(a, b, c)
Chris@42: #define VFNMS(a, b, c) SUFF(_mm512_fnmadd)(a, b, c)
Chris@42: #define VFMAI(b, c)    SUFF(_mm512_fmaddsub)(VLIT1(1.), c, FLIP_RI(b))
Chris@42: #define VFNMSI(b, c)   SUFF(_mm512_fmsubadd)(VLIT1(1.), c, FLIP_RI(b))
Chris@42: #define VFMACONJ(b,c)  SUFF(_mm512_fmsubadd)(VLIT1(1.), c, b)
Chris@42: #define VFMSCONJ(b,c)  SUFF(_mm512_fmsubadd)(VLIT1(-1.), c, b)
Chris@42: #define VFNMSCONJ(b,c) SUFF(_mm512_fmaddsub)(VLIT1(1.), c, b)
Chris@42: 
Chris@42: static inline V LDA(const R *x, INT ivs, const R *aligned_like) {
Chris@42:   (void)aligned_like; /* UNUSED */
Chris@42:   (void)ivs; /* UNUSED */
Chris@42:   return SUFF(_mm512_load)(x);
Chris@42: }
Chris@42: static inline void STA(R *x, V v, INT ovs, const R *aligned_like) {
Chris@42:   (void)aligned_like; /* UNUSED */
Chris@42:   (void)ovs; /* UNUSED */
Chris@42:   SUFF(_mm512_store)(x, v);
Chris@42: }
Chris@42: 
Chris@42: #if FFTW_SINGLE
Chris@42: 
Chris@42: static inline V LDu(const R *x, INT ivs, const R *aligned_like)
Chris@42: {
Chris@42:   (void)aligned_like; /* UNUSED */
Chris@42:   __m512i index = _mm512_set_epi32(7 * ivs + 1, 7 * ivs,
Chris@42:                                    6 * ivs + 1, 6 * ivs,
Chris@42:                                    5 * ivs + 1, 5 * ivs,
Chris@42:                                    4 * ivs + 1, 4 * ivs,
Chris@42:                                    3 * ivs + 1, 3 * ivs,
Chris@42:                                    2 * ivs + 1, 2 * ivs,
Chris@42:                                    1 * ivs + 1, 1 * ivs,
Chris@42:                                    0 * ivs + 1, 0 * ivs);
Chris@42:   
Chris@42:   return _mm512_i32gather_ps(index, x, 4);
Chris@42: }
Chris@42: 
Chris@42: static inline void STu(R *x, V v, INT ovs, const R *aligned_like)
Chris@42: {
Chris@42:   (void)aligned_like; /* UNUSED */
Chris@42:   __m512i index = _mm512_set_epi32(7 * ovs + 1, 7 * ovs,
Chris@42:                                    6 * ovs + 1, 6 * ovs,
Chris@42:                                    5 * ovs + 1, 5 * ovs,
Chris@42:                                    4 * ovs + 1, 4 * ovs,
Chris@42:                                    3 * ovs + 1, 3 * ovs,
Chris@42:                                    2 * ovs + 1, 2 * ovs,
Chris@42:                                    1 * ovs + 1, 1 * ovs,
Chris@42:                                    0 * ovs + 1, 0 * ovs);
Chris@42:   
Chris@42:   _mm512_i32scatter_ps(x, index, v, 4);
Chris@42: }
Chris@42: 
Chris@42: #else /* !FFTW_SINGLE */
Chris@42: 
Chris@42: static inline V LDu(const R *x, INT ivs, const R *aligned_like)
Chris@42: {
Chris@42:   (void)aligned_like; /* UNUSED */
Chris@42:   __m256i index = _mm256_set_epi32(3 * ivs + 1, 3 * ivs,
Chris@42:                                    2 * ivs + 1, 2 * ivs,
Chris@42:                                    1 * ivs + 1, 1 * ivs,
Chris@42:                                    0 * ivs + 1, 0 * ivs);
Chris@42:   
Chris@42:   return _mm512_i32gather_pd(index, x, 8);
Chris@42: }
Chris@42: 
Chris@42: static inline void STu(R *x, V v, INT ovs, const R *aligned_like)
Chris@42: {
Chris@42:   (void)aligned_like; /* UNUSED */
Chris@42:   __m256i index = _mm256_set_epi32(3 * ovs + 1, 3 * ovs,
Chris@42:                                    2 * ovs + 1, 2 * ovs,
Chris@42:                                    1 * ovs + 1, 1 * ovs,
Chris@42:                                    0 * ovs + 1, 0 * ovs);
Chris@42:   
Chris@42:   _mm512_i32scatter_pd(x, index, v, 8);
Chris@42: }
Chris@42: 
Chris@42: #endif /* FFTW_SINGLE */
Chris@42: 
Chris@42: #define LD LDu
Chris@42: #define ST STu
Chris@42: 
Chris@42: #ifdef FFTW_SINGLE
Chris@42: #define STM2(x, v, ovs, a) ST(x, v, ovs, a)
Chris@42: #define STN2(x, v0, v1, ovs) /* nop */
Chris@42: 
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:   __m512i index = _mm512_set_epi32(15 * ovs, 14 * ovs,
Chris@42:                                    13 * ovs, 12 * ovs,
Chris@42:                                    11 * ovs, 10 * ovs,
Chris@42:                                    9 * ovs, 8 * ovs,
Chris@42:                                    7 * ovs, 6 * ovs,
Chris@42:                                    5 * ovs, 4 * ovs,
Chris@42:                                    3 * ovs, 2 * ovs,
Chris@42:                                    1 * ovs, 0 * ovs);
Chris@42:   
Chris@42:   _mm512_i32scatter_ps(x, index, v, 4);
Chris@42: }
Chris@42: #define STN4(x, v0, v1, v2, v3, ovs)  /* no-op */
Chris@42: #else /* !FFTW_SINGLE */
Chris@42: #define STM2(x, v, ovs, a) ST(x, v, ovs, a)
Chris@42: #define STN2(x, v0, v1, ovs) /* nop */
Chris@42: 
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:   __m256i index = _mm256_set_epi32(7 * ovs, 6 * ovs,
Chris@42:                                    5 * ovs, 4 * ovs,
Chris@42:                                    3 * ovs, 2 * ovs,
Chris@42:                                    1 * ovs, 0 * ovs);
Chris@42:   
Chris@42:   _mm512_i32scatter_pd(x, index, v, 8);
Chris@42: }
Chris@42: #define STN4(x, v0, v1, v2, v3, ovs)  /* no-op */
Chris@42: #endif /* FFTW_SINGLE */
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:      return SUFF(_mm512_fmaddsub)(sr, VDUPL(tx), VMUL(FLIP_RI(sr), VDUPH(tx)));
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:      return SUFF(_mm512_fmsubadd)(sr, VDUPL(tx), VMUL(FLIP_RI(sr), VDUPH(tx)));
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:   /* return SUFF(_mm512_addsub)(SUFF(_mm512_fnmadd)(sr, VDUPH(tx), VZERO), VMUL(FLIP_RI(sr), VDUPL(tx))); */
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:      return SUFF(_mm512_fmaddsub)(sr, VDUPH(tx), VMUL(FLIP_RI(sr), VDUPL(tx)));
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}, {TW_CEXP, v+4, x}, {TW_CEXP, v+5, x}, {TW_CEXP, v+6, x}, {TW_CEXP, v+7, x}
Chris@42: #else /* !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: #endif /* FFTW_SINGLE */
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_COS, v+4,  x}, {TW_COS, v+4, x}, {TW_COS, v+5,  x}, {TW_COS, v+5, x}, \
Chris@42:    {TW_COS, v+6,  x}, {TW_COS, v+6, x}, {TW_COS, v+7,  x}, {TW_COS, v+7, 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:    {TW_SIN, v+4, -x}, {TW_SIN, v+4, x}, {TW_SIN, v+5, -x}, {TW_SIN, v+5, x}, \
Chris@42:    {TW_SIN, v+6, -x}, {TW_SIN, v+6, x}, {TW_SIN, v+7, -x}, {TW_SIN, v+7, x}
Chris@42: #else /* !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: #endif /* FFTW_SINGLE */
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: /*      V tr = LD(t, 2, t), ti = LD(t + VL, 2, t + VL); */
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: /*      V tr = LD(t, 2, t), ti = LD(t + VL, 2, t + VL); */
Chris@42:      return VFNMS(ti, si, VMUL(tr, sr));
Chris@42: }
Chris@42: 
Chris@42: /* twiddle storage #3 */
Chris@42: #define VTW3(v,x) VTW1(v,x)
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_COS, v+8 , x}, {TW_COS, v+9 , x}, {TW_COS, v+10, x}, {TW_COS, v+11, x}, \
Chris@42:   {TW_COS, v+12, x}, {TW_COS, v+13, x}, {TW_COS, v+14, x}, {TW_COS, v+15, 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:   {TW_SIN, v+8 , x}, {TW_SIN, v+9 , x}, {TW_SIN, v+10, x}, {TW_SIN, v+11, x}, \
Chris@42:   {TW_SIN, v+12, x}, {TW_SIN, v+13, x}, {TW_SIN, v+14, x}, {TW_SIN, v+15, x}
Chris@42: #else /* !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: #endif /* FFTW_SINGLE */
Chris@42: #define TWVLS (2 * VL)
Chris@42: 
Chris@42: #define VLEAVE() /* nothing */
Chris@42: 
Chris@42: #include "simd-common.h"