From c0e82939acbb0ea23afa340e0fdbe1bd36881516 Mon Sep 17 00:00:00 2001
From: Anthony Scemama <scemama@irsamc.ups-tlse.fr>
Date: Tue, 14 Jun 2022 22:30:33 +0200
Subject: [PATCH] Switch for asm in AOs

---
 org/qmckl_ao.org | 33 ++++++++++++++++++++++++++-------
 1 file changed, 26 insertions(+), 7 deletions(-)

diff --git a/org/qmckl_ao.org b/org/qmckl_ao.org
index b1734de..1e10ccf 100644
--- a/org/qmckl_ao.org
+++ b/org/qmckl_ao.org
@@ -6521,8 +6521,25 @@ qmckl_compute_ao_vgl_hpc_gaussian (
           exp_mat[iprim][4] = f * (3.0 - 2.0 * ar2[iprim]);
         }
 
-        
+
 /* --- */
+      switch (8) {
+      case(5):
+
+        for (int i=0 ; i<nucleus_shell_num[inucl] ; ++i) {
+          for (int j=0 ; j<5 ; ++j) {
+            ce_mat[i][j] = 0.;
+          }
+          for (int k=0 ; k<nidx; ++k) {
+            for (int j=0 ; j<5 ; ++j) {
+              ce_mat[i][j] += coef_mat[inucl][i][k] * exp_mat[k][j];
+            }
+          }
+       }
+
+       break;
+      case(8):
+
         for (int i=0 ; i<nucleus_shell_num[inucl] ; ++i) {
 #ifdef HAVE_OPENMP
 #pragma omp simd simdlen(8)
@@ -6540,9 +6557,11 @@ qmckl_compute_ao_vgl_hpc_gaussian (
           }
        }
 
-/*
-        for (int i=0 ; i<nucleus_shell_num[inucl] ; ++i) {
+       break;
+       case(256):
+
 // Following loop is the assembly version AVX2
+        for (int i=0 ; i<nucleus_shell_num[inucl] ; ++i) {
           __asm__ volatile (
             "mov       %[k],%%RSI" "\n\t"     // &(nidx)
             "mov       %[a],%%RAX" "\n\t"     // &(coef_mat[inucl][i][k])
@@ -6577,10 +6596,10 @@ qmckl_compute_ao_vgl_hpc_gaussian (
             "ymm0", "ymm1", "ymm2", "memory" );
 
         }
-      }
-,*/
 
-/*
+       break;
+       case(512):
+
 // Following loop is the assembly version AVX512
         for (int i=0 ; i<nucleus_shell_num[inucl] ; ++i) {
           __asm__ volatile (
@@ -6608,7 +6627,7 @@ qmckl_compute_ao_vgl_hpc_gaussian (
           : : [k] "m"(nidx), [c] "m"(&(ce_mat[i][0])), [a] "m"(&(coef_mat[inucl][i][0])), [b] "m"(&(exp_mat[0][0])) : "rax", "rbx", "rcx", "rsi", "zmm0", "zmm1", "memory" );
 
         }
-,*/
+       }
 
         const int64_t ishell_start = nucleus_index[inucl];
         const int64_t ishell_end   = nucleus_index[inucl] + nucleus_shell_num[inucl];