int i_shift_offset,i_coord_offset,outeriter,inneriter;
int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
int jnrA,jnrB,jnrC,jnrD;
+ int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
int *iinr,*jindex,*jjnr,*shiftidx,*gid;
- real shX,shY,shZ,rcutoff_scalar;
+ real rcutoff_scalar;
real *shiftvec,*fshift,*x,*f;
+ real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
+ real scratch[4*DIM];
__m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
/* #for I in PARTICLES_I */
int vdwioffset{I};
/* #endif */
/* #if 'GeneralizedBorn' in KERNEL_ELEC */
__m128i gbitab;
- __m128 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,dvdatmp;
+ __m128 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
__m128 minushalf = _mm_set1_ps(-0.5);
real *invsqrta,*dvda,*gbtab;
/* #endif */
outeriter = 0;
inneriter = 0;
+ for(iidx=0;iidx<4*DIM;iidx++)
+ {
+ scratch[iidx] = 0.0;
+ }
+
/* Start outer loop over neighborlists */
for(iidx=0; iidx<nri; iidx++)
{
/* Load shift vector for this list */
i_shift_offset = DIM*shiftidx[iidx];
- shX = shiftvec[i_shift_offset+XX];
- shY = shiftvec[i_shift_offset+YY];
- shZ = shiftvec[i_shift_offset+ZZ];
/* Load limits for loop over neighbors */
j_index_start = jindex[iidx];
i_coord_offset = DIM*inr;
/* Load i particle coords and add shift vector */
- /* ## Loop over i particles, but only include ones that we use - skip e.g. vdw-only sites for elec-only kernel */
- /* #for I in PARTICLES_I */
- ix{I} = _mm_set1_ps(shX + x[i_coord_offset+DIM*{I}+XX]);
- iy{I} = _mm_set1_ps(shY + x[i_coord_offset+DIM*{I}+YY]);
- iz{I} = _mm_set1_ps(shZ + x[i_coord_offset+DIM*{I}+ZZ]);
- /* #define OUTERFLOPS OUTERFLOPS+3 */
- /* #endfor */
-
+ /* #if GEOMETRY_I == 'Particle' */
+ gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
+ /* #elif GEOMETRY_I == 'Water3' */
+ gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
+ &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
+ /* #elif GEOMETRY_I == 'Water4' */
+ /* #if 0 in PARTICLES_I */
+ gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
+ &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
+ /* #else */
+ gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
+ &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
+ /* #endif */
+ /* #endif */
+
/* #if 'Force' in KERNEL_VF */
/* #for I in PARTICLES_I */
fix{I} = _mm_setzero_ps();
/* #define INNERFLOPS 0 */
/* Get j neighbor index, and coordinate index */
+ /* #if ROUND =='Loop' */
jnrA = jjnr[jidx];
jnrB = jjnr[jidx+1];
jnrC = jjnr[jidx+2];
jnrD = jjnr[jidx+3];
-
- /* #if ROUND =='Epilogue' */
+ /* #else */
+ jnrlistA = jjnr[jidx];
+ jnrlistB = jjnr[jidx+1];
+ jnrlistC = jjnr[jidx+2];
+ jnrlistD = jjnr[jidx+3];
/* Sign of each element will be negative for non-real atoms.
* This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
* so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
*/
dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
- jnrA = (jnrA>=0) ? jnrA : 0;
- jnrB = (jnrB>=0) ? jnrB : 0;
- jnrC = (jnrC>=0) ? jnrC : 0;
- jnrD = (jnrD>=0) ? jnrD : 0;
-
+ jnrA = (jnrlistA>=0) ? jnrlistA : 0;
+ jnrB = (jnrlistB>=0) ? jnrlistB : 0;
+ jnrC = (jnrlistC>=0) ? jnrlistC : 0;
+ jnrD = (jnrlistD>=0) ? jnrlistD : 0;
/* #endif */
j_coord_offsetA = DIM*jnrA;
j_coord_offsetB = DIM*jnrB;
* CALCULATE INTERACTIONS *
**************************/
- /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
+ /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
+ /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
/* ## We always calculate rinv/rinvsq above to enable pipelineing in compilers (performance tested on x86) */
if (gmx_mm_any_lt(rsq{I}{J},rcutoff2))
{
isaprod = _mm_mul_ps(isai{I},isaj{J});
gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq{I}{J},_mm_mul_ps(isaprod,gbinvepsdiff)));
gbscale = _mm_mul_ps(isaprod,gbtabscale);
- dvdaj = gmx_mm_load_4real_swizzle_ps(dvda+jnrA+{J},dvda+jnrB+{J},dvda+jnrC+{J},dvda+jnrD+{J});
/* #define INNERFLOPS INNERFLOPS+5 */
/* Calculate generalized born table index - this is a separate table from the normal one,
fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r{I}{J})));
dvdasum = _mm_add_ps(dvdasum,dvdatmp);
- gmx_mm_store_4real_swizzle_ps(dvda+jnrA,dvda+jnrB,dvda+jnrC,dvda+jnrD,
- _mm_mul_ps(dvdatmp,_mm_mul_ps(isaj{J},isaj{J})));
+ /* #if ROUND == 'Loop' */
+ fjptrA = dvda+jnrA;
+ fjptrB = dvda+jnrB;
+ fjptrC = dvda+jnrC;
+ fjptrD = dvda+jnrD;
+ /* #else */
+ /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
+ fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
+ fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
+ fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
+ fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
+ /* #endif */
+ gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj{J},isaj{J})));
/* #define INNERFLOPS INNERFLOPS+13 */
/* #endif */
velec = _mm_mul_ps(qq{I}{J},rinv{I}{J});
/* #endif */
/* #endif */
/* #endif */
- /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
+ /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
+ /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
cutoff_mask = _mm_cmplt_ps(rsq{I}{J},rcutoff2);
/* #define INNERFLOPS INNERFLOPS+1 */
/* #endif */
/* #endif */
/* #endif */
/* #if 'vdw' in INTERACTION_FLAGS[I][J] */
- /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
+ /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
+ /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
vvdw = _mm_and_ps(vvdw,cutoff_mask);
/* #define INNERFLOPS INNERFLOPS+1 */
/* #endif */
fscal = fvdw;
/* #endif */
- /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
+ /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
+ /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
fscal = _mm_and_ps(fscal,cutoff_mask);
- /* #define INNERFLOPS INNERFLOPS+1 */
- /* #endif */
+ /* #define INNERFLOPS INNERFLOPS+1 */
+ /* #endif */
/* #if ROUND == 'Epilogue' */
fscal = _mm_andnot_ps(dummy_mask,fscal);
/* #define INNERFLOPS INNERFLOPS+6 */
/* #if GEOMETRY_J == 'Particle' */
- gmx_mm_decrement_1rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
- f+j_coord_offsetC,f+j_coord_offsetD,
- tx,ty,tz);
+ /* #if ROUND == 'Loop' */
+ fjptrA = f+j_coord_offsetA;
+ fjptrB = f+j_coord_offsetB;
+ fjptrC = f+j_coord_offsetC;
+ fjptrD = f+j_coord_offsetD;
+ /* #else */
+ fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
+ fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
+ fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
+ fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
+ /* #endif */
+ gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
/* #define INNERFLOPS INNERFLOPS+3 */
/* #else */
fjx{J} = _mm_add_ps(fjx{J},tx);
fjz{J} = _mm_add_ps(fjz{J},tz);
/* #define INNERFLOPS INNERFLOPS+3 */
/* #endif */
-
+
/* #endif */
- /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
+ /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
+ /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
/* #if 0 ## This and next two lines is a hack to maintain indentation in template file */
{
/* #endif */
/* #endfor */
/* ## End of loop over i-j interaction pairs */
+ /* #if GEOMETRY_J != 'Particle' */
+ /* #if ROUND == 'Loop' */
+ fjptrA = f+j_coord_offsetA;
+ fjptrB = f+j_coord_offsetB;
+ fjptrC = f+j_coord_offsetC;
+ fjptrD = f+j_coord_offsetD;
+ /* #else */
+ fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
+ fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
+ fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
+ fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
+ /* #endif */
+ /* #endif */
+
/* #if GEOMETRY_J == 'Water3' */
- gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
- f+j_coord_offsetC,f+j_coord_offsetD,
+ gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
/* #define INNERFLOPS INNERFLOPS+9 */
/* #elif GEOMETRY_J == 'Water4' */
/* #if 0 in PARTICLES_J */
- gmx_mm_decrement_4rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
- f+j_coord_offsetC,f+j_coord_offsetD,
+ gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
/* #define INNERFLOPS INNERFLOPS+12 */
/* #else */
- gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
- f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
+ gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
/* #define INNERFLOPS INNERFLOPS+9 */
/* #endif */
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff