Merge release-4-6 into master

[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse4_1_single / nb_kernel_ElecRF_VdwNone_GeomW4P1_sse4_1_single.c

diff --git a/src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRF_VdwNone_GeomW4P1_sse4_1_single.c b/src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRF_VdwNone_GeomW4P1_sse4_1_single.c
index 69c2ca0851a3ea4400a62d19a70b71891db70c11..41beb98b8949ba31d5221c8acbee5100fe061e4f 100644 (file)
--- a/src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRF_VdwNone_GeomW4P1_sse4_1_single.c
+++ b/src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRF_VdwNone_GeomW4P1_sse4_1_single.c
@@ -199,6 +199,10 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_sse4_1_single
             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
                                                               charge+jnrC+0,charge+jnrD+0);
 
+            fjx0             = _mm_setzero_ps();
+            fjy0             = _mm_setzero_ps();
+            fjz0             = _mm_setzero_ps();
+
             /**************************
              * CALCULATE INTERACTIONS *
              **************************/
@@ -225,11 +229,9 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_sse4_1_single
             fiy1             = _mm_add_ps(fiy1,ty);
             fiz1             = _mm_add_ps(fiz1,tz);
 
-            fjptrA             = f+j_coord_offsetA;
-            fjptrB             = f+j_coord_offsetB;
-            fjptrC             = f+j_coord_offsetC;
-            fjptrD             = f+j_coord_offsetD;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
 
             /**************************
              * CALCULATE INTERACTIONS *
@@ -257,11 +259,9 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_sse4_1_single
             fiy2             = _mm_add_ps(fiy2,ty);
             fiz2             = _mm_add_ps(fiz2,tz);
 
-            fjptrA             = f+j_coord_offsetA;
-            fjptrB             = f+j_coord_offsetB;
-            fjptrC             = f+j_coord_offsetC;
-            fjptrD             = f+j_coord_offsetD;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
 
             /**************************
              * CALCULATE INTERACTIONS *
@@ -289,11 +289,16 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_sse4_1_single
             fiy3             = _mm_add_ps(fiy3,ty);
             fiz3             = _mm_add_ps(fiz3,tz);
 
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
+
             fjptrA             = f+j_coord_offsetA;
             fjptrB             = f+j_coord_offsetB;
             fjptrC             = f+j_coord_offsetC;
             fjptrD             = f+j_coord_offsetD;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+
+           gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 
             /* Inner loop uses 96 flops */
         }
@@ -353,6 +358,10 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_sse4_1_single
             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
                                                               charge+jnrC+0,charge+jnrD+0);
 
+            fjx0             = _mm_setzero_ps();
+            fjy0             = _mm_setzero_ps();
+            fjz0             = _mm_setzero_ps();
+
             /**************************
              * CALCULATE INTERACTIONS *
              **************************/
@@ -382,11 +391,9 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_sse4_1_single
             fiy1             = _mm_add_ps(fiy1,ty);
             fiz1             = _mm_add_ps(fiz1,tz);
 
-            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;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
 
             /**************************
              * CALCULATE INTERACTIONS *
@@ -417,11 +424,9 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_sse4_1_single
             fiy2             = _mm_add_ps(fiy2,ty);
             fiz2             = _mm_add_ps(fiz2,tz);
 
-            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;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
 
             /**************************
              * CALCULATE INTERACTIONS *
@@ -452,11 +457,16 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_sse4_1_single
             fiy3             = _mm_add_ps(fiy3,ty);
             fiz3             = _mm_add_ps(fiz3,tz);
 
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
+
             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;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+
+           gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 
             /* Inner loop uses 96 flops */
         }
@@ -646,6 +656,10 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_F_sse4_1_single
             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
                                                               charge+jnrC+0,charge+jnrD+0);
 
+            fjx0             = _mm_setzero_ps();
+            fjy0             = _mm_setzero_ps();
+            fjz0             = _mm_setzero_ps();
+
             /**************************
              * CALCULATE INTERACTIONS *
              **************************/
@@ -668,11 +682,9 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_F_sse4_1_single
             fiy1             = _mm_add_ps(fiy1,ty);
             fiz1             = _mm_add_ps(fiz1,tz);
 
-            fjptrA             = f+j_coord_offsetA;
-            fjptrB             = f+j_coord_offsetB;
-            fjptrC             = f+j_coord_offsetC;
-            fjptrD             = f+j_coord_offsetD;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
 
             /**************************
              * CALCULATE INTERACTIONS *
@@ -696,11 +708,9 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_F_sse4_1_single
             fiy2             = _mm_add_ps(fiy2,ty);
             fiz2             = _mm_add_ps(fiz2,tz);
 
-            fjptrA             = f+j_coord_offsetA;
-            fjptrB             = f+j_coord_offsetB;
-            fjptrC             = f+j_coord_offsetC;
-            fjptrD             = f+j_coord_offsetD;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
 
             /**************************
              * CALCULATE INTERACTIONS *
@@ -724,11 +734,16 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_F_sse4_1_single
             fiy3             = _mm_add_ps(fiy3,ty);
             fiz3             = _mm_add_ps(fiz3,tz);
 
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
+
             fjptrA             = f+j_coord_offsetA;
             fjptrB             = f+j_coord_offsetB;
             fjptrC             = f+j_coord_offsetC;
             fjptrD             = f+j_coord_offsetD;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+
+           gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 
             /* Inner loop uses 81 flops */
         }
@@ -788,6 +803,10 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_F_sse4_1_single
             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
                                                               charge+jnrC+0,charge+jnrD+0);
 
+            fjx0             = _mm_setzero_ps();
+            fjy0             = _mm_setzero_ps();
+            fjz0             = _mm_setzero_ps();
+
             /**************************
              * CALCULATE INTERACTIONS *
              **************************/
@@ -812,11 +831,9 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_F_sse4_1_single
             fiy1             = _mm_add_ps(fiy1,ty);
             fiz1             = _mm_add_ps(fiz1,tz);
 
-            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;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
 
             /**************************
              * CALCULATE INTERACTIONS *
@@ -842,11 +859,9 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_F_sse4_1_single
             fiy2             = _mm_add_ps(fiy2,ty);
             fiz2             = _mm_add_ps(fiz2,tz);
 
-            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;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
 
             /**************************
              * CALCULATE INTERACTIONS *
@@ -872,11 +887,16 @@ nb_kernel_ElecRF_VdwNone_GeomW4P1_F_sse4_1_single
             fiy3             = _mm_add_ps(fiy3,ty);
             fiz3             = _mm_add_ps(fiz3,tz);
 
+            fjx0             = _mm_add_ps(fjx0,tx);
+            fjy0             = _mm_add_ps(fjy0,ty);
+            fjz0             = _mm_add_ps(fjz0,tz);
+
             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;
-            gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
+
+           gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 
             /* Inner loop uses 81 flops */
         }