Chris@10: /*
Chris@10:  * Copyright (c) 2003, 2007-11 Matteo Frigo
Chris@10:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
Chris@10:  *
Chris@10:  * This program is free software; you can redistribute it and/or modify
Chris@10:  * it under the terms of the GNU General Public License as published by
Chris@10:  * the Free Software Foundation; either version 2 of the License, or
Chris@10:  * (at your option) any later version.
Chris@10:  *
Chris@10:  * This program is distributed in the hope that it will be useful,
Chris@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@10:  * GNU General Public License for more details.
Chris@10:  *
Chris@10:  * You should have received a copy of the GNU General Public License
Chris@10:  * along with this program; if not, write to the Free Software
Chris@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@10:  *
Chris@10:  */
Chris@10: 
Chris@10: #ifndef FFTW_SINGLE
Chris@10: #error "NEON only works in single precision"
Chris@10: #endif
Chris@10: 
Chris@10: /* define these unconditionally, because they are used by
Chris@10:    taint.c which is compiled without neon */
Chris@10: #define SIMD_SUFFIX _neon	/* for renaming */
Chris@10: #define VL 2            /* SIMD complex vector length */
Chris@10: #define SIMD_VSTRIDE_OKA(x) ((x) == 2)
Chris@10: #define SIMD_STRIDE_OKPAIR SIMD_STRIDE_OK
Chris@10: 
Chris@10: #if defined(__GNUC__) && !defined(__ARM_NEON__)
Chris@10: #error "compiling simd-neon.h requires -mfpu=neon or equivalent"
Chris@10: #endif
Chris@10: 
Chris@10: #include <arm_neon.h>
Chris@10: 
Chris@10: /* FIXME: I am not sure whether this code assumes little-endian
Chris@10:    ordering.  VLIT may or may not be wrong for big-endian systems. */
Chris@10: typedef float32x4_t V;
Chris@10: 
Chris@10: #define VLIT(x0, x1, x2, x3) {x0, x1, x2, x3}
Chris@10: #define LDK(x) x
Chris@10: #define DVK(var, val) const V var = VLIT(val, val, val, val)
Chris@10: 
Chris@10: /* NEON has FMA, but a three-operand FMA is not too useful
Chris@10:    for FFT purposes.  We normally compute
Chris@10: 
Chris@10:       t0=a+b*c
Chris@10:       t1=a-b*c
Chris@10: 
Chris@10:    In a three-operand instruction set this translates into
Chris@10: 
Chris@10:       t0=a
Chris@10:       t0+=b*c
Chris@10:       t1=a
Chris@10:       t1-=b*c
Chris@10: 
Chris@10:    At least one move must be implemented, negating the advantage of
Chris@10:    the FMA in the first place.  At least some versions of gcc generate
Chris@10:    both moves.  So we are better off generating t=b*c;t0=a+t;t1=a-t;*/
Chris@10: #if HAVE_FMA
Chris@10: #warning "--enable-fma on NEON is probably a bad idea (see source code)"
Chris@10: #endif
Chris@10: 
Chris@10: #define VADD(a, b) vaddq_f32(a, b)
Chris@10: #define VSUB(a, b) vsubq_f32(a, b)
Chris@10: #define VMUL(a, b) vmulq_f32(a, b)
Chris@10: #define VFMA(a, b, c) vmlaq_f32(c, a, b)	        /* a*b+c */
Chris@10: #define VFNMS(a, b, c) vmlsq_f32(c, a, b)	/* FNMS=-(a*b-c) in powerpc terminology; MLS=c-a*b
Chris@10: 						   in ARM terminology */
Chris@10: #define VFMS(a, b, c) VSUB(VMUL(a, b), c)	/* FMS=a*b-c in powerpc terminology; no equivalent
Chris@10: 						   arm instruction (?) */
Chris@10: 
Chris@10: static inline V LDA(const R *x, INT ivs, const R *aligned_like)
Chris@10: {
Chris@10:      (void) aligned_like;	/* UNUSED */
Chris@10:      return vld1q_f32((const float32_t *)x);
Chris@10: }
Chris@10: 
Chris@10: static inline V LD(const R *x, INT ivs, const R *aligned_like)
Chris@10: {
Chris@10:      (void) aligned_like;	/* UNUSED */
Chris@10:      return vcombine_f32(vld1_f32((float32_t *)x), vld1_f32((float32_t *)(x + ivs)));
Chris@10: }
Chris@10: 
Chris@10: static inline void STA(R *x, V v, INT ovs, const R *aligned_like)
Chris@10: {
Chris@10:      (void) aligned_like;	/* UNUSED */
Chris@10:      vst1q_f32((float32_t *)x, v);
Chris@10: }
Chris@10: 
Chris@10: static inline void ST(R *x, V v, INT ovs, const R *aligned_like)
Chris@10: {
Chris@10:      (void) aligned_like;	/* UNUSED */
Chris@10:      /* WARNING: the extra_iter hack depends upon store-low occurring
Chris@10: 	after store-high */
Chris@10:      vst1_f32((float32_t *)(x + ovs), vget_high_f32(v));
Chris@10:      vst1_f32((float32_t *)x, vget_low_f32(v));
Chris@10: }
Chris@10: 
Chris@10: /* 2x2 complex transpose and store */
Chris@10: #define STM2 ST
Chris@10: #define STN2(x, v0, v1, ovs) /* nop */
Chris@10: 
Chris@10: /* store and 4x4 real transpose */
Chris@10: static inline void STM4(R *x, V v, INT ovs, const R *aligned_like)
Chris@10: {
Chris@10:      (void) aligned_like;	/* UNUSED */
Chris@10:      vst1_lane_f32((float32_t *)(x)      , vget_low_f32(v), 0);
Chris@10:      vst1_lane_f32((float32_t *)(x + ovs), vget_low_f32(v), 1);
Chris@10:      vst1_lane_f32((float32_t *)(x + 2 * ovs), vget_high_f32(v), 0);
Chris@10:      vst1_lane_f32((float32_t *)(x + 3 * ovs), vget_high_f32(v), 1);
Chris@10: }
Chris@10: #define STN4(x, v0, v1, v2, v3, ovs)	/* use STM4 */
Chris@10: 
Chris@10: #define FLIP_RI(x) vrev64q_f32(x)
Chris@10: 
Chris@10: static inline V VCONJ(V x)
Chris@10: {
Chris@10: #if 1
Chris@10:      static const uint32x4_t pm = {0, 0x80000000u, 0, 0x80000000u};
Chris@10:      return vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(x), pm));
Chris@10: #else
Chris@10:      const V pm = VLIT(1.0, -1.0, 1.0, -1.0);
Chris@10:      return VMUL(x, pm);
Chris@10: #endif
Chris@10: }
Chris@10: 
Chris@10: static inline V VBYI(V x)
Chris@10: {
Chris@10:      return FLIP_RI(VCONJ(x));
Chris@10: }
Chris@10: 
Chris@10: static inline V VFMAI(V b, V c)
Chris@10: {
Chris@10:      const V mp = VLIT(-1.0, 1.0, -1.0, 1.0);
Chris@10:      return VFMA(FLIP_RI(b), mp, c);
Chris@10: }
Chris@10: 
Chris@10: static inline V VFNMSI(V b, V c)
Chris@10: {
Chris@10:      const V mp = VLIT(-1.0, 1.0, -1.0, 1.0);
Chris@10:      return VFNMS(FLIP_RI(b), mp, c);
Chris@10: }
Chris@10: 
Chris@10: static inline V VFMACONJ(V b, V c)
Chris@10: {
Chris@10:      const V pm = VLIT(1.0, -1.0, 1.0, -1.0);
Chris@10:      return VFMA(b, pm, c);
Chris@10: }
Chris@10: 
Chris@10: static inline V VFNMSCONJ(V b, V c)
Chris@10: {
Chris@10:      const V pm = VLIT(1.0, -1.0, 1.0, -1.0);
Chris@10:      return VFNMS(b, pm, c);
Chris@10: }
Chris@10: 
Chris@10: static inline V VFMSCONJ(V b, V c)
Chris@10: {
Chris@10:      return VSUB(VCONJ(b), c);
Chris@10: }
Chris@10: 
Chris@10: #if 1
Chris@10: #define VEXTRACT_REIM(tr, ti, tx)                               \
Chris@10: {                                                               \
Chris@10:      tr = vcombine_f32(vdup_lane_f32(vget_low_f32(tx), 0),      \
Chris@10:                        vdup_lane_f32(vget_high_f32(tx), 0));    \
Chris@10:      ti = vcombine_f32(vdup_lane_f32(vget_low_f32(tx), 1),      \
Chris@10:                        vdup_lane_f32(vget_high_f32(tx), 1));    \
Chris@10: }
Chris@10: #else
Chris@10: /* this alternative might be faster in an ideal world, but gcc likes
Chris@10:    to spill VVV onto the stack */
Chris@10: #define VEXTRACT_REIM(tr, ti, tx)               \
Chris@10: {                                               \
Chris@10:      float32x4x2_t vvv = vtrnq_f32(tx, tx);     \
Chris@10:      tr = vvv.val[0];                           \
Chris@10:      ti = vvv.val[1];                           \
Chris@10: }
Chris@10: #endif
Chris@10: 
Chris@10: static inline V VZMUL(V tx, V sr)
Chris@10: {
Chris@10:      V tr, ti;
Chris@10:      VEXTRACT_REIM(tr, ti, tx);
Chris@10:      tr = VMUL(sr, tr);
Chris@10:      sr = VBYI(sr);
Chris@10:      return VFMA(ti, sr, tr);
Chris@10: }
Chris@10: 
Chris@10: static inline V VZMULJ(V tx, V sr)
Chris@10: {
Chris@10:      V tr, ti;
Chris@10:      VEXTRACT_REIM(tr, ti, tx);
Chris@10:      tr = VMUL(sr, tr);
Chris@10:      sr = VBYI(sr);
Chris@10:      return VFNMS(ti, sr, tr);
Chris@10: }
Chris@10: 
Chris@10: static inline V VZMULI(V tx, V sr)
Chris@10: {
Chris@10:      V tr, ti;
Chris@10:      VEXTRACT_REIM(tr, ti, tx);
Chris@10:      ti = VMUL(ti, sr);
Chris@10:      sr = VBYI(sr);
Chris@10:      return VFMS(tr, sr, ti);
Chris@10: }
Chris@10: 
Chris@10: static inline V VZMULIJ(V tx, V sr)
Chris@10: {
Chris@10:      V tr, ti;
Chris@10:      VEXTRACT_REIM(tr, ti, tx);
Chris@10:      ti = VMUL(ti, sr);
Chris@10:      sr = VBYI(sr);
Chris@10:      return VFMA(tr, sr, ti);
Chris@10: }
Chris@10: 
Chris@10: /* twiddle storage #1: compact, slower */
Chris@10: #define VTW1(v,x) {TW_CEXP, v, x}, {TW_CEXP, v+1, x}
Chris@10: #define TWVL1 VL
Chris@10: static inline V BYTW1(const R *t, V sr)
Chris@10: {
Chris@10:      V tx = LDA(t, 2, 0);
Chris@10:      return VZMUL(tx, sr);
Chris@10: }
Chris@10: 
Chris@10: static inline V BYTWJ1(const R *t, V sr)
Chris@10: {
Chris@10:      V tx = LDA(t, 2, 0);
Chris@10:      return VZMULJ(tx, sr);
Chris@10: }
Chris@10: 
Chris@10: /* twiddle storage #2: twice the space, faster (when in cache) */
Chris@10: #  define VTW2(v,x)							\
Chris@10:   {TW_COS, v, x}, {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_COS, v+1, x},	\
Chris@10:   {TW_SIN, v, -x}, {TW_SIN, v, x}, {TW_SIN, v+1, -x}, {TW_SIN, v+1, x}
Chris@10: #define TWVL2 (2 * VL)
Chris@10: 
Chris@10: static inline V BYTW2(const R *t, V sr)
Chris@10: {
Chris@10:      V si = FLIP_RI(sr);
Chris@10:      V tr = LDA(t, 2, 0), ti = LDA(t+2*VL, 2, 0);
Chris@10:      return VFMA(ti, si, VMUL(tr, sr));
Chris@10: }
Chris@10: 
Chris@10: static inline V BYTWJ2(const R *t, V sr)
Chris@10: {
Chris@10:      V si = FLIP_RI(sr);
Chris@10:      V tr = LDA(t, 2, 0), ti = LDA(t+2*VL, 2, 0);
Chris@10:      return VFNMS(ti, si, VMUL(tr, sr));
Chris@10: }
Chris@10: 
Chris@10: /* twiddle storage #3 */
Chris@10: #  define VTW3(v,x) {TW_CEXP, v, x}, {TW_CEXP, v+1, x}
Chris@10: #  define TWVL3 (VL)
Chris@10: 
Chris@10: /* twiddle storage for split arrays */
Chris@10: #  define VTWS(v,x)							  \
Chris@10:     {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_COS, v+2, x}, {TW_COS, v+3, x}, \
Chris@10:     {TW_SIN, v, x}, {TW_SIN, v+1, x}, {TW_SIN, v+2, x}, {TW_SIN, v+3, x}
Chris@10: #define TWVLS (2 * VL)
Chris@10: 
Chris@10: #define VLEAVE()		/* nothing */
Chris@10: 
Chris@10: #include "simd-common.h"