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:  * 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 "AVX 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  _avx  /* for renaming */
Chris@82: #define VL DS(2, 4)        /* SIMD complex vector length */
Chris@82: #define SIMD_VSTRIDE_OKA(x) ((x) == 2) 
Chris@82: #define SIMD_STRIDE_OKPAIR SIMD_STRIDE_OK
Chris@82: 
Chris@82: #if defined(__GNUC__) && !defined(__AVX__) /* sanity check */
Chris@82: #error "compiling simd-avx.h without -mavx"
Chris@82: #endif
Chris@82: 
Chris@82: #ifdef _MSC_VER
Chris@82: #ifndef inline
Chris@82: #define inline __inline
Chris@82: #endif
Chris@82: #endif
Chris@82: 
Chris@82: #include <immintrin.h>
Chris@82: 
Chris@82: typedef DS(__m256d, __m256) V;
Chris@82: #define VADD SUFF(_mm256_add_p)
Chris@82: #define VSUB SUFF(_mm256_sub_p)
Chris@82: #define VMUL SUFF(_mm256_mul_p)
Chris@82: #define VXOR SUFF(_mm256_xor_p)
Chris@82: #define VSHUF SUFF(_mm256_shuffle_p)
Chris@82: 
Chris@82: #define SHUFVALD(fp0,fp1) \
Chris@82:    (((fp1) << 3) | ((fp0) << 2) | ((fp1) << 1) | ((fp0)))
Chris@82: #define SHUFVALS(fp0,fp1,fp2,fp3) \
Chris@82:    (((fp3) << 6) | ((fp2) << 4) | ((fp1) << 2) | ((fp0)))
Chris@82: 
Chris@82: #define VDUPL(x) DS(_mm256_unpacklo_pd(x, x), VSHUF(x, x, SHUFVALS(0, 0, 2, 2)))
Chris@82: #define VDUPH(x) DS(_mm256_unpackhi_pd(x, x), VSHUF(x, x, SHUFVALS(1, 1, 3, 3)))
Chris@82: 
Chris@82: #define VLIT(x0, x1) DS(_mm256_set_pd(x0, x1, x0, x1), _mm256_set_ps(x0, x1, x0, x1, x0, x1, x0, x1))
Chris@82: #define DVK(var, val) V var = VLIT(val, val)
Chris@82: #define LDK(x) x
Chris@82: 
Chris@82: static inline V LDA(const R *x, INT ivs, const R *aligned_like)
Chris@82: {
Chris@82:      (void)aligned_like; /* UNUSED */
Chris@82:      (void)ivs; /* UNUSED */
Chris@82:      return SUFF(_mm256_loadu_p)(x);
Chris@82: }
Chris@82: 
Chris@82: static inline void STA(R *x, V v, INT ovs, const R *aligned_like)
Chris@82: {
Chris@82:      (void)aligned_like; /* UNUSED */
Chris@82:      (void)ovs; /* UNUSED */
Chris@82:      SUFF(_mm256_storeu_p)(x, v);
Chris@82: }
Chris@82: 
Chris@82: #if FFTW_SINGLE
Chris@82: 
Chris@82: #  ifdef _MSC_VER
Chris@82:      /* Temporarily disable the warning "uninitialized local variable
Chris@82: 	'name' used" and runtime checks for using a variable before it is
Chris@82: 	defined which is erroneously triggered by the LOADL0 / LOADH macros
Chris@82: 	as they only modify VAL partly each. */
Chris@82: #    ifndef __INTEL_COMPILER
Chris@82: #      pragma warning(disable : 4700)
Chris@82: #      pragma runtime_checks("u", off)
Chris@82: #    endif
Chris@82: #  endif
Chris@82: #  ifdef __INTEL_COMPILER
Chris@82: #    pragma warning(disable : 592)
Chris@82: #  endif
Chris@82: 
Chris@82: #define LOADH(addr, val) _mm_loadh_pi(val, (const __m64 *)(addr))
Chris@82: #define LOADL(addr, val) _mm_loadl_pi(val, (const __m64 *)(addr))
Chris@82: #define STOREH(addr, val) _mm_storeh_pi((__m64 *)(addr), val)
Chris@82: #define STOREL(addr, val) _mm_storel_pi((__m64 *)(addr), val)
Chris@82: 
Chris@82: /* it seems like the only AVX way to store 4 complex floats is to
Chris@82:    extract two pairs of complex floats into two __m128 registers, and
Chris@82:    then use SSE-like half-stores.  Similarly, to load 4 complex
Chris@82:    floats, we load two pairs of complex floats into two __m128
Chris@82:    registers, and then pack the two __m128 registers into one __m256
Chris@82:    value. */
Chris@82: static inline V LD(const R *x, INT ivs, const R *aligned_like)
Chris@82: {
Chris@82:      __m128 l, h;
Chris@82:      V v;
Chris@82:      (void)aligned_like; /* UNUSED */
Chris@82:      l = LOADL(x, l);
Chris@82:      l = LOADH(x + ivs, l);
Chris@82:      h = LOADL(x + 2*ivs, h);
Chris@82:      h = LOADH(x + 3*ivs, h);
Chris@82:      v = _mm256_castps128_ps256(l);
Chris@82:      v = _mm256_insertf128_ps(v, h, 1);
Chris@82:      return v;
Chris@82: }
Chris@82: 
Chris@82: #  ifdef _MSC_VER
Chris@82: #    ifndef __INTEL_COMPILER
Chris@82: #      pragma warning(default : 4700)
Chris@82: #      pragma runtime_checks("u", restore)
Chris@82: #    endif
Chris@82: #  endif
Chris@82: #  ifdef __INTEL_COMPILER
Chris@82: #    pragma warning(default : 592)
Chris@82: #  endif
Chris@82: 
Chris@82: static inline void ST(R *x, V v, INT ovs, const R *aligned_like)
Chris@82: {
Chris@82:      __m128 h = _mm256_extractf128_ps(v, 1);
Chris@82:      __m128 l = _mm256_castps256_ps128(v);
Chris@82:      (void)aligned_like; /* UNUSED */
Chris@82:      /* WARNING: the extra_iter hack depends upon STOREL occurring
Chris@82: 	after STOREH */
Chris@82:      STOREH(x + 3*ovs, h);
Chris@82:      STOREL(x + 2*ovs, h);
Chris@82:      STOREH(x + ovs, l);
Chris@82:      STOREL(x, l);
Chris@82: }
Chris@82: 
Chris@82: #define STM2(x, v, ovs, aligned_like) /* no-op */
Chris@82: static inline void STN2(R *x, V v0, V v1, INT ovs)
Chris@82: {
Chris@82:     V x0 = VSHUF(v0, v1, SHUFVALS(0, 1, 0, 1));
Chris@82:     V x1 = VSHUF(v0, v1, SHUFVALS(2, 3, 2, 3));
Chris@82:     __m128 h0 = _mm256_extractf128_ps(x0, 1);
Chris@82:     __m128 l0 = _mm256_castps256_ps128(x0);
Chris@82:     __m128 h1 = _mm256_extractf128_ps(x1, 1);
Chris@82:     __m128 l1 = _mm256_castps256_ps128(x1);
Chris@82: 
Chris@82:     *(__m128 *)(x + 3*ovs) = h1;
Chris@82:     *(__m128 *)(x + 2*ovs) = h0;
Chris@82:     *(__m128 *)(x + 1*ovs) = l1;
Chris@82:     *(__m128 *)(x + 0*ovs) = l0;
Chris@82: }
Chris@82: 
Chris@82: #define STM4(x, v, ovs, aligned_like) /* no-op */
Chris@82: #define STN4(x, v0, v1, v2, v3, ovs)				\
Chris@82: {								\
Chris@82:      V xxx0, xxx1, xxx2, xxx3;					\
Chris@82:      V yyy0, yyy1, yyy2, yyy3;					\
Chris@82:      xxx0 = _mm256_unpacklo_ps(v0, v2);				\
Chris@82:      xxx1 = _mm256_unpackhi_ps(v0, v2);				\
Chris@82:      xxx2 = _mm256_unpacklo_ps(v1, v3);				\
Chris@82:      xxx3 = _mm256_unpackhi_ps(v1, v3);				\
Chris@82:      yyy0 = _mm256_unpacklo_ps(xxx0, xxx2);			\
Chris@82:      yyy1 = _mm256_unpackhi_ps(xxx0, xxx2);			\
Chris@82:      yyy2 = _mm256_unpacklo_ps(xxx1, xxx3);			\
Chris@82:      yyy3 = _mm256_unpackhi_ps(xxx1, xxx3);			\
Chris@82:      *(__m128 *)(x + 0 * ovs) = _mm256_castps256_ps128(yyy0);	\
Chris@82:      *(__m128 *)(x + 4 * ovs) = _mm256_extractf128_ps(yyy0, 1);	\
Chris@82:      *(__m128 *)(x + 1 * ovs) = _mm256_castps256_ps128(yyy1);	\
Chris@82:      *(__m128 *)(x + 5 * ovs) = _mm256_extractf128_ps(yyy1, 1);	\
Chris@82:      *(__m128 *)(x + 2 * ovs) = _mm256_castps256_ps128(yyy2);	\
Chris@82:      *(__m128 *)(x + 6 * ovs) = _mm256_extractf128_ps(yyy2, 1);	\
Chris@82:      *(__m128 *)(x + 3 * ovs) = _mm256_castps256_ps128(yyy3);	\
Chris@82:      *(__m128 *)(x + 7 * ovs) = _mm256_extractf128_ps(yyy3, 1);	\
Chris@82: }
Chris@82: 
Chris@82: #else
Chris@82: static inline __m128d VMOVAPD_LD(const R *x)
Chris@82: {
Chris@82:      /* gcc-4.6 miscompiles the combination _mm256_castpd128_pd256(VMOVAPD_LD(x))
Chris@82: 	into a 256-bit vmovapd, which requires 32-byte aligment instead of
Chris@82: 	16-byte alignment.
Chris@82: 
Chris@82: 	Force the use of vmovapd via asm until compilers stabilize.
Chris@82:      */
Chris@82: #if defined(__GNUC__)
Chris@82:      __m128d var;
Chris@82:      __asm__("vmovapd %1, %0\n" : "=x"(var) : "m"(x[0]));
Chris@82:      return var;
Chris@82: #else
Chris@82:      return *(const __m128d *)x;
Chris@82: #endif
Chris@82: }
Chris@82: 
Chris@82: static inline V LD(const R *x, INT ivs, const R *aligned_like)
Chris@82: {
Chris@82:      V var;
Chris@82:      (void)aligned_like; /* UNUSED */
Chris@82:      var = _mm256_castpd128_pd256(VMOVAPD_LD(x));
Chris@82:      var = _mm256_insertf128_pd(var, *(const __m128d *)(x+ivs), 1);
Chris@82:      return var;
Chris@82: }
Chris@82: 
Chris@82: static inline void ST(R *x, V v, INT ovs, const R *aligned_like)
Chris@82: {
Chris@82:      (void)aligned_like; /* UNUSED */
Chris@82:      /* WARNING: the extra_iter hack depends upon the store of the low
Chris@82: 	part occurring after the store of the high part */
Chris@82:      *(__m128d *)(x + ovs) = _mm256_extractf128_pd(v, 1);
Chris@82:      *(__m128d *)x = _mm256_castpd256_pd128(v);
Chris@82: }
Chris@82: 
Chris@82: 
Chris@82: #define STM2 ST
Chris@82: #define STN2(x, v0, v1, ovs) /* nop */
Chris@82: #define STM4(x, v, ovs, aligned_like) /* no-op */
Chris@82: 
Chris@82: /* STN4 is a macro, not a function, thanks to Visual C++ developers
Chris@82:    deciding "it would be infrequent that people would want to pass more
Chris@82:    than 3 [__m128 parameters] by value."  Even though the comment
Chris@82:    was made about __m128 parameters, it appears to apply to __m256
Chris@82:    parameters as well. */
Chris@82: #define STN4(x, v0, v1, v2, v3, ovs)					\
Chris@82: {									\
Chris@82:      V xxx0, xxx1, xxx2, xxx3;						\
Chris@82:      xxx0 = _mm256_unpacklo_pd(v0, v1);					\
Chris@82:      xxx1 = _mm256_unpackhi_pd(v0, v1);					\
Chris@82:      xxx2 = _mm256_unpacklo_pd(v2, v3);					\
Chris@82:      xxx3 = _mm256_unpackhi_pd(v2, v3);					\
Chris@82:      STA(x,           _mm256_permute2f128_pd(xxx0, xxx2, 0x20), 0, 0); \
Chris@82:      STA(x +     ovs, _mm256_permute2f128_pd(xxx1, xxx3, 0x20), 0, 0); \
Chris@82:      STA(x + 2 * ovs, _mm256_permute2f128_pd(xxx0, xxx2, 0x31), 0, 0); \
Chris@82:      STA(x + 3 * ovs, _mm256_permute2f128_pd(xxx1, xxx3, 0x31), 0, 0); \
Chris@82: }
Chris@82: #endif
Chris@82: 
Chris@82: static inline V FLIP_RI(V x)
Chris@82: {
Chris@82:      return VSHUF(x, x,
Chris@82: 		  DS(SHUFVALD(1, 0), 
Chris@82: 		     SHUFVALS(1, 0, 3, 2)));
Chris@82: }
Chris@82: 
Chris@82: static inline V VCONJ(V x)
Chris@82: {
Chris@82:      /* Produce a SIMD vector[VL] of (0 + -0i). 
Chris@82: 
Chris@82:         We really want to write this:
Chris@82: 
Chris@82:            V pmpm = VLIT(-0.0, 0.0);
Chris@82: 
Chris@82:         but historically some compilers have ignored the distiction
Chris@82:         between +0 and -0.  It looks like 'gcc-8 -fast-math' treats -0
Chris@82:         as 0 too.
Chris@82:       */
Chris@82:      union uvec {
Chris@82:           unsigned u[8];
Chris@82:           V v;
Chris@82:      };
Chris@82:      static const union uvec pmpm = {
Chris@82: #ifdef FFTW_SINGLE
Chris@82:           { 0x00000000, 0x80000000, 0x00000000, 0x80000000,
Chris@82:             0x00000000, 0x80000000, 0x00000000, 0x80000000 }
Chris@82: #else
Chris@82:           { 0x00000000, 0x00000000, 0x00000000, 0x80000000,
Chris@82:             0x00000000, 0x00000000, 0x00000000, 0x80000000 }
Chris@82: #endif
Chris@82:      };
Chris@82:      return VXOR(pmpm.v, x);
Chris@82: }
Chris@82: 
Chris@82: static inline V VBYI(V x)
Chris@82: {
Chris@82:      return FLIP_RI(VCONJ(x));
Chris@82: }
Chris@82: 
Chris@82: /* FMA support */
Chris@82: #define VFMA(a, b, c) VADD(c, VMUL(a, b))
Chris@82: #define VFNMS(a, b, c) VSUB(c, VMUL(a, b))
Chris@82: #define VFMS(a, b, c) VSUB(VMUL(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) {TW_CEXP, v, x}, {TW_CEXP, v+1, x}, {TW_CEXP, v+2, x}, {TW_CEXP, v+3, x}
Chris@82: #else
Chris@82: # define VTW1(v,x) {TW_CEXP, v, x}, {TW_CEXP, v+1, x}
Chris@82: #endif
Chris@82: #define TWVL1 (VL)
Chris@82: 
Chris@82: static inline V BYTW1(const R *t, V sr)
Chris@82: {
Chris@82:      return VZMUL(LDA(t, 2, t), sr);
Chris@82: }
Chris@82: 
Chris@82: static inline V BYTWJ1(const R *t, V sr)
Chris@82: {
Chris@82:      return VZMULJ(LDA(t, 2, t), sr);
Chris@82: }
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_COS, v+2, x}, {TW_COS, v+2, x}, {TW_COS, v+3, x}, {TW_COS, v+3, x}, \
Chris@82:    {TW_SIN, v, -x}, {TW_SIN, v, x}, {TW_SIN, v+1, -x}, {TW_SIN, v+1, x}, \
Chris@82:    {TW_SIN, v+2, -x}, {TW_SIN, v+2, x}, {TW_SIN, v+3, -x}, {TW_SIN, v+3, x}
Chris@82: #else
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: #endif
Chris@82: #define TWVL2 (2 * VL)
Chris@82: 
Chris@82: static inline V BYTW2(const R *t, V sr)
Chris@82: {
Chris@82:      const V *twp = (const V *)t;
Chris@82:      V si = FLIP_RI(sr);
Chris@82:      V tr = twp[0], ti = twp[1];
Chris@82:      return VFMA(tr, sr, VMUL(ti, si));
Chris@82: }
Chris@82: 
Chris@82: static inline V BYTWJ2(const R *t, V sr)
Chris@82: {
Chris@82:      const V *twp = (const V *)t;
Chris@82:      V si = FLIP_RI(sr);
Chris@82:      V tr = twp[0], ti = twp[1];
Chris@82:      return VFNMS(ti, si, VMUL(tr, sr));
Chris@82: }
Chris@82: 
Chris@82: /* twiddle storage #3 */
Chris@82: #define VTW3 VTW1
Chris@82: #define TWVL3 TWVL1
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_COS, v+4, x}, {TW_COS, v+5, x}, {TW_COS, v+6, x}, {TW_COS, v+7, x}, \
Chris@82:   {TW_SIN, v, x}, {TW_SIN, v+1, x}, {TW_SIN, v+2, x}, {TW_SIN, v+3, x},	\
Chris@82:   {TW_SIN, v+4, x}, {TW_SIN, v+5, x}, {TW_SIN, v+6, x}, {TW_SIN, v+7, x}
Chris@82: #else
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: #endif
Chris@82: #define TWVLS (2 * VL)
Chris@82: 
Chris@82: 
Chris@82: /* Use VZEROUPPER to avoid the penalty of switching from AVX to SSE.
Chris@82:    See Intel Optimization Manual (April 2011, version 248966), Section
Chris@82:    11.3 */
Chris@82: #define VLEAVE _mm256_zeroupper
Chris@82: 
Chris@82: #include "simd-common.h"