portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-03 21:44:33 +00:00
Code Releases Activity

Big updates with progress towards wilson matrix

Browse Source
This commit is contained in:
Peter Boyle
2015-04-26 15:51:09 +01:00
parent 51f0da7b93
commit 94f728bee4
27 changed files with 1008 additions and 355 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
lib/simd/Grid_vComplexD.h
View File

@@ -251,13 +251,13 @@ friend inline void vstore(const vComplexD &ret, ComplexD *a){
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], ...
// __m256d tmp = _mm256_addsub_pd(ret.v,_mm256_shuffle_pd(in.v,in.v,0x5));
// ret.v=_mm256_shuffle_pd(tmp,tmp,0x5);
ret.v = _mm256_addsub_pd(ret.v,in.v);
// 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
ret.v = _mm_addsub_pd(ret.v,in.v);
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);
@@ -268,48 +268,41 @@ friend inline void vstore(const vComplexD &ret, ComplexD *a){
return ret;
}
friend inline vComplexD timesI(const vComplexD &in){
friend inline vComplexD timesMinusI(const vComplexD &in){
vComplexD ret; vzero(ret);
vComplexD tmp;
#if defined (AVX1)|| defined (AVX2)
cvec tmp =_mm256_addsub_ps(ret.v,in.v); // r,-i
/*
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;
*/
ret.v =_mm256_shuffle_ps(tmp,tmp,0x5);
tmp.v =_mm256_addsub_pd(ret.v,in.v); // r,-i
ret.v =_mm256_shuffle_pd(tmp.v,tmp.v,0x5);
#endif
#ifdef SSE4
cvec tmp =_mm_addsub_ps(ret.v,in.v); // r,-i
ret.v =_mm_shuffle_ps(tmp,tmp,0x5);
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_ps(in.v,0xaaaa,ret.v,in.v); // real -imag
ret.v = _mm512_swizzle_ps(ret.v, _MM_SWIZ_REG_CDAB);// OK
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 timesMinusI(const vComplexD &in){
friend inline vComplexD timesI(const vComplexD &in){
vComplexD ret; vzero(ret);
vComplexD tmp;
#if defined (AVX1)|| defined (AVX2)
cvec tmp =_mm256_shuffle_ps(in.v,in.v,0x5);
ret.v =_mm256_addsub_ps(ret.v,tmp); // i,-r
tmp.v =_mm256_shuffle_pd(in.v,in.v,0x5);
ret.v =_mm256_addsub_pd(ret.v,tmp.v); // i,-r
#endif
#ifdef SSE4
cvec tmp =_mm_shuffle_ps(in.v,in.v,0x5);
ret.v =_mm_addsub_ps(ret.v,tmp); // r,-i
tmp.v =_mm_shuffle_pd(in.v,in.v,0x1);
ret.v =_mm_addsub_pd(ret.v,tmp.v); // r,-i
#endif
#ifdef AVX512
cvec tmp = _mm512_swizzle_ps(in.v, _MM_SWIZ_REG_CDAB);// OK
ret.v = _mm512_mask_sub_ps(tmp,0xaaaa,ret.v,tmp); // real -imag
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);

36
lib/simd/Grid_vComplexF.h
View File

@@ -214,10 +214,10 @@ friend inline void vstore(const vComplexF &ret, ComplexF *a){
{
#ifdef SSE4
union {
__m128 v1; // SSE 4 x float vector
cvec v1; // SSE 4 x float vector
float f[4]; // scalar array of 4 floats
} u128;
u128.v1= _mm_add_ps(v, _mm_shuffle_ps(v, v, 0b01001110)); // FIXME Prefer to use _MM_SHUFFLE macros
u128.v1= _mm_add_ps(in.v, _mm_shuffle_ps(in.v,in.v, 0b01001110)); // FIXME Prefer to use _MM_SHUFFLE macros
return ComplexF(u128.f[0], u128.f[1]);
#endif
#ifdef AVX1
@@ -329,13 +329,15 @@ friend inline void vstore(const vComplexF &ret, ComplexF *a){
friend inline vComplexF conj(const vComplexF &in){
vComplexF ret ; vzero(ret);
#if defined (AVX1)|| defined (AVX2)
// cvec tmp;
// tmp = _mm256_addsub_ps(ret.v,_mm256_shuffle_ps(in.v,in.v,_MM_SHUFFLE(2,3,0,1))); // ymm1 <- br,bi
// ret.v=_mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(2,3,0,1));
ret.v = _mm256_addsub_ps(ret.v,in.v);
cvec tmp;
tmp = _mm256_addsub_ps(ret.v,_mm256_shuffle_ps(in.v,in.v,_MM_SHUFFLE(2,3,0,1))); // ymm1 <- br,bi
ret.v=_mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(2,3,0,1));
#endif
#ifdef SSE4
ret.v = _mm_addsub_ps(ret.v,in.v);
cvec tmp;
tmp = _mm_addsub_ps(ret.v,_mm_shuffle_ps(in.v,in.v,_MM_SHUFFLE(2,3,0,1))); // ymm1 <- br,bi
ret.v=_mm_shuffle_ps(tmp,tmp,_MM_SHUFFLE(2,3,0,1));
#endif
#ifdef AVX512
ret.v = _mm512_mask_sub_ps(in.v,0xaaaa,ret.v,in.v); // Zero out 0+real 0-imag
@@ -345,15 +347,16 @@ friend inline void vstore(const vComplexF &ret, ComplexF *a){
#endif
return ret;
}
friend inline vComplexF timesI(const vComplexF &in){
vComplexF ret; vzero(ret);
friend inline vComplexF timesMinusI(const vComplexF &in){
vComplexF ret;
vzero(ret);
#if defined (AVX1)|| defined (AVX2)
cvec tmp =_mm256_addsub_ps(ret.v,in.v); // r,-i
ret.v = _mm256_shuffle_ps(tmp,tmp,0x5);
ret.v = _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(2,3,0,1)); //-i,r
#endif
#ifdef SSE4
cvec tmp =_mm_addsub_ps(ret.v,in.v); // r,-i
ret.v = _mm_shuffle_ps(tmp,tmp,0x5);
ret.v = _mm_shuffle_ps(tmp,tmp,_MM_SHUFFLE(2,3,0,1));
#endif
#ifdef AVX512
ret.v = _mm512_mask_sub_ps(in.v,0xaaaa,ret.v,in.v); // real -imag
@@ -364,14 +367,14 @@ friend inline void vstore(const vComplexF &ret, ComplexF *a){
#endif
return ret;
}
friend inline vComplexF timesMinusI(const vComplexF &in){
friend inline vComplexF timesI(const vComplexF &in){
vComplexF ret; vzero(ret);
#if defined (AVX1)|| defined (AVX2)
cvec tmp =_mm256_shuffle_ps(in.v,in.v,0x5);
ret.v = _mm256_addsub_ps(ret.v,tmp); // i,-r
cvec tmp =_mm256_shuffle_ps(in.v,in.v,_MM_SHUFFLE(2,3,0,1));//i,r
ret.v =_mm256_addsub_ps(ret.v,tmp); //i,-r
#endif
#ifdef SSE4
cvec tmp =_mm_shuffle_ps(in.v,in.v,0x5);
cvec tmp =_mm_shuffle_ps(in.v,in.v,_MM_SHUFFLE(2,3,0,1));
ret.v = _mm_addsub_ps(ret.v,tmp); // r,-i
#endif
#ifdef AVX512
@@ -443,5 +446,8 @@ friend inline void vstore(const vComplexF &ret, ComplexF *a){
inline vComplexF trace(const vComplexF &arg){
return arg;
}
}
#endif

11
lib/simd/Grid_vRealD.h
View File

@@ -146,7 +146,7 @@ namespace Grid {
ret.v = _mm256_set_pd(a[3],a[2],a[1],a[0]);
#endif
#ifdef SSE4
ret.v = _mm_set_pd(a[0],a[1]);
ret.v = _mm_set_pd(a[1],a[0]);
#endif
#ifdef AVX512
ret.v = _mm512_set_pd(a[7],a[6],a[5],a[4],a[3],a[2],a[1],a[0]);
@@ -186,6 +186,15 @@ namespace Grid {
friend inline RealD Reduce(const vRealD & in)
{
#if defined (SSE4)
// FIXME Hack
const RealD * ptr =(const RealD *) &in;
RealD ret = 0;
for(int i=0;i<vRealD::Nsimd();i++){
ret = ret+ptr[i];
}
return ret;
#endif
#if defined (AVX1) || defined(AVX2)
typedef union {
uint64_t l;

11
lib/simd/Grid_vRealF.h
View File

@@ -175,7 +175,7 @@ namespace Grid {
ret.v = _mm256_set_ps(a[7],a[6],a[5],a[4],a[3],a[2],a[1],a[0]);
#endif
#ifdef SSE4
ret.v = _mm_set_ps(a[0],a[1],a[2],a[3]);
ret.v = _mm_set_ps(a[3],a[2],a[1],a[0]);
#endif
#ifdef AVX512
ret.v = _mm512_set_ps( a[15],a[14],a[13],a[12],a[11],a[10],a[9],a[8],
@@ -220,6 +220,15 @@ friend inline void vstore(const vRealF &ret, float *a){
}
friend inline RealF Reduce(const vRealF & in)
{
#if defined (SSE4)
// FIXME Hack
const RealF * ptr = (const RealF *) &in;
RealF ret = 0;
for(int i=0;i<vRealF::Nsimd();i++){
ret = ret+ptr[i];
}
return ret;
#endif
#if defined (AVX1) || defined(AVX2)
__attribute__ ((aligned(32))) float c_[16];
__m256 tmp = _mm256_permute2f128_ps(in.v,in.v,0x01);