Merge release-5-0 into master

[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sparc64_hpc_ace_double / nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_sparc64_hpc_ace_double.c

diff --git a/src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_sparc64_hpc_ace_double.c b/src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_sparc64_hpc_ace_double.c
index 7f3df8cfb4eab4da5119256506c0508b58cfbcc3..5dabbd9a3045d1c47be11df312e44b24013278a8 100644 (file)
--- a/src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_sparc64_hpc_ace_double.c
+++ b/src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_sparc64_hpc_ace_double.c
@@ -342,11 +342,11 @@ nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_VF_sparc64_hpc_ace_double
             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
-            exponent         = gmx_simd_exp_d(-ewcljrsq);
+            exponent         = gmx_simd_exp_d(ewcljrsq);
             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
            poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
-            vvdw6            = _fjsp_mul_v2r8(_fjsp_madd_v2r8(-c6grid_00,_fjsp_sub_v2r8(one,poly),c6_00),rinvsix);
+            vvdw6            = _fjsp_mul_v2r8(_fjsp_madd_v2r8(c6grid_00,_fjsp_sub_v2r8(poly,one),c6_00),rinvsix);
             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
                                _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw6,_fjsp_madd_v2r8(c6grid_00,sh_lj_ewald,_fjsp_mul_v2r8(c6_00,sh_vdw_invrcut6))),one_sixth));
@@ -923,11 +923,11 @@ nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_VF_sparc64_hpc_ace_double
             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
-            exponent         = gmx_simd_exp_d(-ewcljrsq);
+            exponent         = gmx_simd_exp_d(ewcljrsq);
             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
            poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
-            vvdw6            = _fjsp_mul_v2r8(_fjsp_madd_v2r8(-c6grid_00,_fjsp_sub_v2r8(one,poly),c6_00),rinvsix);
+            vvdw6            = _fjsp_mul_v2r8(_fjsp_madd_v2r8(c6grid_00,_fjsp_sub_v2r8(poly,one),c6_00),rinvsix);
             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
             vvdw             = _fjsp_msub_v2r8(_fjsp_nmsub_v2r8(c12_00,_fjsp_mul_v2r8(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
                                _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw6,_fjsp_madd_v2r8(c6grid_00,sh_lj_ewald,_fjsp_mul_v2r8(c6_00,sh_vdw_invrcut6))),one_sixth));
@@ -1744,11 +1744,11 @@ nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_F_sparc64_hpc_ace_double
             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
-            exponent         = gmx_simd_exp_d(-ewcljrsq);
+            exponent         = gmx_simd_exp_d(ewcljrsq);
             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
            poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
             /* f6A = 6 * C6grid * (1 - poly) */
-            f6A              = _fjsp_mul_v2r8(c6grid_00,_fjsp_msub_v2r8(one,poly));
+            f6A              = _fjsp_mul_v2r8(c6grid_00,_fjsp_sub_v2r8(one,poly));
             /* f6B = C6grid * exponent * beta^6 */
             f6B              = _fjsp_mul_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6));
             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
@@ -2230,11 +2230,11 @@ nb_kernel_ElecEwSh_VdwLJEwSh_GeomW4W4_F_sparc64_hpc_ace_double
             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
             ewcljrsq         = _fjsp_mul_v2r8(ewclj2,rsq00);
             ewclj6           = _fjsp_mul_v2r8(ewclj2,_fjsp_mul_v2r8(ewclj2,ewclj2));
-            exponent         = gmx_simd_exp_d(-ewcljrsq);
+            exponent         = gmx_simd_exp_d(ewcljrsq);
             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
            poly             = _fjsp_mul_v2r8(exponent,_fjsp_madd_v2r8(_fjsp_mul_v2r8(ewcljrsq,ewcljrsq),one_half,_fjsp_sub_v2r8(one,ewcljrsq)));
             /* f6A = 6 * C6grid * (1 - poly) */
-            f6A              = _fjsp_mul_v2r8(c6grid_00,_fjsp_msub_v2r8(one,poly));
+            f6A              = _fjsp_mul_v2r8(c6grid_00,_fjsp_sub_v2r8(one,poly));
             /* f6B = C6grid * exponent * beta^6 */
             f6B              = _fjsp_mul_v2r8(_fjsp_mul_v2r8(c6grid_00,one_sixth),_fjsp_mul_v2r8(exponent,ewclj6));
             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */