portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity
c72db6c6f6
Grid/lib/simd/Grid_vComplexD.h
Peter Boyle 25d523c0f4 Shaken out stencil to the point where I think wilson dslash is correct. 
Need to audit code carefully, consolidate between stencil and cshift,
and then benchmark and optimise.
2015-04-28 08:11:59 +01:00

388 lines
13 KiB
C++
Raw Blame History

#ifndef GRID_VCOMPLEXD_H
#define GRID_VCOMPLEXD_H
namespace Grid {
class vComplexD {
public:
zvec v;
public:
typedef zvec vector_type;
typedef ComplexD scalar_type;
vComplexD & operator = ( Zero & z){
vzero(*this);
return (*this);
}
vComplexD(){};
vComplexD(ComplexD a){
vsplat(*this,a);
};
vComplexD(double a){
vsplat(*this,ComplexD(a));
};
///////////////////////////////////////////////
// mac, mult, sub, add, adj
// Should do an AVX2 version with mac.
///////////////////////////////////////////////
friend inline void mac (vComplexD * __restrict__ y,const vComplexD * __restrict__ a,const vComplexD *__restrict__ x) {*y = (*a)*(*x)+(*y);};
friend inline void mult(vComplexD * __restrict__ y,const vComplexD * __restrict__ l,const vComplexD *__restrict__ r) {*y = (*l) * (*r);}
friend inline void sub (vComplexD * __restrict__ y,const vComplexD * __restrict__ l,const vComplexD *__restrict__ r) {*y = (*l) - (*r);}
friend inline void add (vComplexD * __restrict__ y,const vComplexD * __restrict__ l,const vComplexD *__restrict__ r) {*y = (*l) + (*r);}
friend inline vComplexD adj(const vComplexD &in){ return conj(in); }
//////////////////////////////////
// Initialise to 1,0,i
//////////////////////////////////
friend inline void vone (vComplexD &ret){ vsplat(ret,1.0,0.0);}
friend inline void vzero (vComplexD &ret){ vsplat(ret,0.0,0.0);}
friend inline void vcomplex_i(vComplexD &ret){ vsplat(ret,0.0,1.0);}
////////////////////////////////////
// Arithmetic operator overloads +,-,*
////////////////////////////////////
friend inline vComplexD operator + (vComplexD a, vComplexD b)
{
vComplexD ret;
#if defined (AVX1)|| defined (AVX2)
ret.v = _mm256_add_pd(a.v,b.v);
#endif
#ifdef SSE4
ret.v = _mm_add_pd(a.v,b.v);
#endif
#ifdef AVX512
ret.v = _mm512_add_pd(a.v,b.v);
#endif
#ifdef QPX
ret.v = vec_add(a.v,b.v);
#endif
return ret;
};
friend inline vComplexD operator - (vComplexD a, vComplexD b)
{
vComplexD ret;
#if defined (AVX1)|| defined (AVX2)
ret.v = _mm256_sub_pd(a.v,b.v);
#endif
#ifdef SSE4
ret.v = _mm_sub_pd(a.v,b.v);
#endif
#ifdef AVX512
ret.v = _mm512_sub_pd(a.v,b.v);
#endif
#ifdef QPX
ret.v = vec_sub(a.v,b.v);
#endif
return ret;
};
friend inline vComplexD operator * (vComplexD a, vComplexD b)
{
vComplexD ret;
//Multiplicationof (ak+ibk)*(ck+idk)
// a + i b can be stored as a data structure
//From intel optimisation reference guide
/*
movsldup xmm0, Src1; load real parts into the destination,
; a1, a1, a0, a0
movaps xmm1, src2; load the 2nd pair of complex values, ; i.e. d1, c1, d0, c0
mulps xmm0, xmm1; temporary results, a1d1, a1c1, a0d0, ; a0c0
shufps xmm1, xmm1, b1; reorder the real and imaginary ; parts, c1, d1, c0, d0
movshdup xmm2, Src1; load the imaginary parts into the ; destination, b1, b1, b0, b0
mulps xmm2, xmm1; temporary results, b1c1, b1d1, b0c0, ; b0d0
addsubps xmm0, xmm2; b1c1+a1d1, a1c1 -b1d1, b0c0+a0d
VSHUFPD (VEX.256 encoded version)
IF IMM0[0] = 0
THEN DEST[63:0]=SRC1[63:0] ELSE DEST[63:0]=SRC1[127:64] FI;
IF IMM0[1] = 0
THEN DEST[127:64]=SRC2[63:0] ELSE DEST[127:64]=SRC2[127:64] FI;
IF IMM0[2] = 0
THEN DEST[191:128]=SRC1[191:128] ELSE DEST[191:128]=SRC1[255:192] FI;
IF IMM0[3] = 0
THEN DEST[255:192]=SRC2[191:128] ELSE DEST[255:192]=SRC2[255:192] FI; // Ox5 r<->i ; 0xC unchanged
*/
#if defined (AVX1)|| defined (AVX2)
zvec ymm0,ymm1,ymm2;
ymm0 = _mm256_shuffle_pd(a.v,a.v,0x0); // ymm0 <- ar ar, ar,ar b'00,00
ymm0 = _mm256_mul_pd(ymm0,b.v); // ymm0 <- ar bi, ar br
ymm1 = _mm256_shuffle_pd(b.v,b.v,0x5); // ymm1 <- br,bi b'01,01
ymm2 = _mm256_shuffle_pd(a.v,a.v,0xF); // ymm2 <- ai,ai b'11,11
ymm1 = _mm256_mul_pd(ymm1,ymm2); // ymm1 <- br ai, ai bi
ret.v= _mm256_addsub_pd(ymm0,ymm1);
#endif
#ifdef SSE4
zvec ymm0,ymm1,ymm2;
ymm0 = _mm_shuffle_pd(a.v,a.v,0x0); // ymm0 <- ar ar,
ymm0 = _mm_mul_pd(ymm0,b.v); // ymm0 <- ar bi, ar br
ymm1 = _mm_shuffle_pd(b.v,b.v,0x1); // ymm1 <- br,bi b01
ymm2 = _mm_shuffle_pd(a.v,a.v,0x3); // ymm2 <- ai,ai b11
ymm1 = _mm_mul_pd(ymm1,ymm2); // ymm1 <- br ai, ai bi
ret.v= _mm_addsub_pd(ymm0,ymm1);
#endif
#ifdef AVX512
/* This is from
* Automatic SIMD Vectorization of Fast Fourier Transforms for the Larrabee and AVX Instruction Sets
* @inproceedings{McFarlin:2011:ASV:1995896.1995938,
* author = {McFarlin, Daniel S. and Arbatov, Volodymyr and Franchetti, Franz and P\"{u}schel, Markus},
* title = {Automatic SIMD Vectorization of Fast Fourier Transforms for the Larrabee and AVX Instruction Sets},
* booktitle = {Proceedings of the International Conference on Supercomputing},
* series = {ICS '11},
* year = {2011},
* isbn = {978-1-4503-0102-2},
* location = {Tucson, Arizona, USA},
* pages = {265--274},
* numpages = {10},
* url = {http://doi.acm.org/10.1145/1995896.1995938},
* doi = {10.1145/1995896.1995938},
* acmid = {1995938},
* publisher = {ACM},
* address = {New York, NY, USA},
* keywords = {autovectorization, fourier transform, program generation, simd, super-optimization},
* }
*/
zvec vzero,ymm0,ymm1,real,imag;
vzero = _mm512_setzero();
ymm0 = _mm512_swizzle_pd(a.v, _MM_SWIZ_REG_CDAB); //
real = _mm512_mask_or_epi64(a.v, 0xAAAA,vzero, ymm0);
imag = _mm512_mask_sub_pd(a.v, 0x5555,vzero, ymm0);
ymm1 = _mm512_mul_pd(real, b.v);
ymm0 = _mm512_swizzle_pd(b.v, _MM_SWIZ_REG_CDAB); // OK
ret.v= _mm512_fmadd_pd(ymm0,imag,ymm1);
/* Imag OK */
#endif
#ifdef QPX
ret.v = vec_mul(a.v,b.v);
#endif
return ret;
};
////////////////////////////////////////////////////////////////////
// General permute; assumes vector length is same across
// all subtypes; may not be a good assumption, but could
// add the vector width as a template param for BG/Q for example
////////////////////////////////////////////////////////////////////
/*
friend inline void permute(vComplexD &y,vComplexD b,int perm)
{
Gpermute<vComplexD>(y,b,perm);
}
friend inline void merge(vComplexD &y,std::vector<ComplexD *> &extracted)
{
Gmerge<vComplexD,ComplexD >(y,extracted);
}
friend inline void extract(const vComplexD &y,std::vector<ComplexD *> &extracted)
{
Gextract<vComplexD,ComplexD>(y,extracted);
}
friend inline void merge(vComplexD &y,std::vector<ComplexD > &extracted)
{
Gmerge<vComplexD,ComplexD >(y,extracted);
}
friend inline void extract(const vComplexD &y,std::vector<ComplexD > &extracted)
{
Gextract<vComplexD,ComplexD>(y,extracted);
}
*/
///////////////////////
// Splat
///////////////////////
friend inline void vsplat(vComplexD &ret,ComplexD c){
float a= real(c);
float b= imag(c);
vsplat(ret,a,b);
}
friend inline void vsplat(vComplexD &ret,double rl,double ig){
#if defined (AVX1)|| defined (AVX2)
ret.v = _mm256_set_pd(ig,rl,ig,rl);
#endif
#ifdef SSE4
ret.v = _mm_set_pd(ig,rl);
#endif
#ifdef AVX512
ret.v = _mm512_set_pd(ig,rl,ig,rl,ig,rl,ig,rl);
#endif
#ifdef QPX
ret.v = {ig,rl,ig,rl};
#endif
}
friend inline void vset(vComplexD &ret,ComplexD *a){
#if defined (AVX1)|| defined (AVX2)
ret.v = _mm256_set_pd(a[1].imag(),a[1].real(),a[0].imag(),a[0].real());
#endif
#ifdef SSE4
ret.v = _mm_set_pd(a[0].imag(),a[0].real());
#endif
#ifdef AVX512
ret.v = _mm512_set_pd(a[3].imag(),a[3].real(),a[2].imag(),a[2].real(),a[1].imag(),a[1].real(),a[0].imag(),a[0].real());
// Note v has a0 a1 a2 a3
#endif
#ifdef QPX
ret.v = {a[0].real(),a[0].imag(),a[1].real(),a[3].imag()};
#endif
}
friend inline void vstore(const vComplexD &ret, ComplexD *a){
#if defined (AVX1)|| defined (AVX2)
_mm256_store_pd((double *)a,ret.v);
#endif
#ifdef SSE4
_mm_store_pd((double *)a,ret.v);
#endif
#ifdef AVX512
_mm512_store_pd((double *)a,ret.v);
//Note v has a3 a2 a1 a0
#endif
#ifdef QPX
assert(0);
#endif
}
friend inline void vprefetch(const vComplexD &v)
{
_mm_prefetch((const char*)&v.v,_MM_HINT_T0);
}
////////////////////////
// Conjugate
////////////////////////
friend inline vComplexD conj(const vComplexD &in){
vComplexD ret ; vzero(ret);
#if defined (AVX1)|| defined (AVX2)
// addsubps 0, inv=>0+in.v[3] 0-in.v[2], 0+in.v[1], 0-in.v[0], ...
zvec tmp = _mm256_addsub_pd(ret.v,_mm256_shuffle_pd(in.v,in.v,0x5));
ret.v =_mm256_shuffle_pd(tmp,tmp,0x5);
#endif
#ifdef SSE4
zvec tmp = _mm_addsub_pd(ret.v,_mm_shuffle_pd(in.v,in.v,0x1));
ret.v = _mm_shuffle_pd(tmp,tmp,0x1);
#endif
#ifdef AVX512
ret.v = _mm512_mask_sub_pd(in.v, 0xaaaa,ret.v, in.v);
#endif
#ifdef QPX
assert(0);
#endif
return ret;
}
friend inline vComplexD timesMinusI(const vComplexD &in){
vComplexD ret; vzero(ret);
vComplexD tmp;
#if defined (AVX1)|| defined (AVX2)
tmp.v =_mm256_addsub_pd(ret.v,in.v); // r,-i
ret.v =_mm256_shuffle_pd(tmp.v,tmp.v,0x5);
#endif
#ifdef SSE4
tmp.v =_mm_addsub_pd(ret.v,in.v); // r,-i
ret.v =_mm_shuffle_pd(tmp.v,tmp.v,0x1);
#endif
#ifdef AVX512
ret.v = _mm512_mask_sub_pd(in.v,0xaaaa,ret.v,in.v); // real -imag
ret.v = _mm512_swizzle_pd(ret.v, _MM_SWIZ_REG_CDAB);// OK
#endif
#ifdef QPX
assert(0);
#endif
return ret;
}
friend inline vComplexD timesI(const vComplexD &in){
vComplexD ret; vzero(ret);
vComplexD tmp;
#if defined (AVX1)|| defined (AVX2)
tmp.v =_mm256_shuffle_pd(in.v,in.v,0x5);
ret.v =_mm256_addsub_pd(ret.v,tmp.v); // i,-r
#endif
#ifdef SSE4
tmp.v =_mm_shuffle_pd(in.v,in.v,0x1);
ret.v =_mm_addsub_pd(ret.v,tmp.v); // r,-i
#endif
#ifdef AVX512
tmp.v = _mm512_swizzle_pd(in.v, _MM_SWIZ_REG_CDAB);// OK
ret.v = _mm512_mask_sub_pd(tmp.v,0xaaaa,ret.v,tmp.v); // real -imag
#endif
#ifdef QPX
assert(0);
#endif
return ret;
}
// REDUCE FIXME must be a cleaner implementation
friend inline ComplexD Reduce(const vComplexD & in)
{
#if defined (AVX1) || defined(AVX2)
// return std::complex<double>(_mm256_mask_reduce_add_pd(0x55, in.v),_mm256_mask_reduce_add_pd(0xAA, in.v));
__attribute__ ((aligned(32))) double c_[4];
_mm256_store_pd(c_,in.v);
return ComplexD(c_[0]+c_[2],c_[1]+c_[3]);
#endif
#ifdef AVX512
return ComplexD(_mm512_mask_reduce_add_pd(0x5555, in.v),_mm512_mask_reduce_add_pd(0xAAAA, in.v));
#endif
#ifdef QPX
#endif
}
// Unary negation
friend inline vComplexD operator -(const vComplexD &r) {
vComplexD ret;
vzero(ret);
ret = ret - r;
return ret;
}
// *=,+=,-= operators
inline vComplexD &operator *=(const vComplexD &r) {
*this = (*this)*r;
return *this;
}
inline vComplexD &operator +=(const vComplexD &r) {
*this = *this+r;
return *this;
}
inline vComplexD &operator -=(const vComplexD &r) {
*this = *this-r;
return *this;
}
public:
static int Nsimd(void) { return sizeof(zvec)/sizeof(double)/2;}
};
inline vComplexD innerProduct(const vComplexD & l, const vComplexD & r) { return conj(l)*r; }
typedef vComplexD vDComplex;
inline void zeroit(vComplexD &z){ vzero(z);}
inline vComplexD outerProduct(const vComplexD &l, const vComplexD& r)
{
return l*r;
}
inline vComplexD trace(const vComplexD &arg){
return arg;
}
/////////////////////////////////////////////////////////////////////////
//// Generic routine to promote object<complex> -> object<vcomplex>
//// Supports the array reordering transformation that gives me SIMD utilisation
///////////////////////////////////////////////////////////////////////////
/*
template<template<class> class object>
inline object<vComplex> splat(object<Complex >s){
object<vComplex> ret;
vComplex * v_ptr = (vComplex *)& ret;
Complex * s_ptr = (Complex *) &s;
for(int i=0;i<sizeof(ret);i+=sizeof(vComplex)){
vsplat(*(v_ptr++),*(s_ptr++));
}
return ret;
}
*/
}
#endif
View Git Blame Copy Permalink