add alg that does O(N^5) on CPU but other on GPU, which is slower

nwchemgit · Oct 16, 2024 · f3b5539 · f3b5539
1 parent c562fc8
commit f3b5539
Showing 1 changed file with 9 additions and 1 deletion.
diff --git a/src/tce/ccsd/ccsd_t2.F b/src/tce/ccsd/ccsd_t2.F
@@ -309,11 +309,19 @@ SUBROUTINE ccsd_t2(rtdb,d_f1,d_i0,d_t1,d_t2,d_v2,
         CALL ccsd_t2_7_3(d_t2,k_t2_offset,d_v2,k_v2_offset,
      &                   d_i1,k_i1_offset)
       else if (ccsd_t2_7_alg.eq.2) then
+        ! no GPU in 7_1_x yet - it just does transpose
         CALL ccsd_t2_7_1_x(d_v2,k_v2_offset,
      &                     d_i1,k_i1_offset,maxh,maxp)
         CALL ccsd_t2_7_2_x(d_t1,k_t1_offset,d_v2,k_v2_offset,
      &                   d_i1,k_i1_offset,maxh,maxp)
         CALL ccsd_t2_7_3_x(d_t2,k_t2_offset,d_v2,k_v2_offset,
+     &                   d_i1,k_i1_offset,maxh,maxp)
+      else if (ccsd_t2_7_alg.eq.3) then
+        CALL ccsd_t2_7_1(d_v2,k_v2_offset,
+     &                     d_i1,k_i1_offset,maxh,maxp)
+        CALL ccsd_t2_7_2(d_t1,k_t1_offset,d_v2,k_v2_offset,
+     &                   d_i1,k_i1_offset,maxh,maxp)
+        CALL ccsd_t2_7_3_x(d_t2,k_t2_offset,d_v2,k_v2_offset,
      &                   d_i1,k_i1_offset,maxh,maxp)
       else
         CALL errquit('ccsd_t2_7_alg',ccsd_t2_7_alg,INPUT_ERR)
@@ -322,7 +330,7 @@ SUBROUTINE ccsd_t2(rtdb,d_f1,d_i0,d_t1,d_t2,d_v2,
       if (ccsd_t2_7_alg.eq.1) then
         CALL ccsd_t2_7(d_t2,k_t2_offset,d_i1,k_i1_offset,
      &                 d_i0,k_i0_offset)
-      else if (ccsd_t2_7_alg.eq.2) then
+      else if (ccsd_t2_7_alg.gt.1) then
         CALL ccsd_t2_7_x(d_t2,k_t2_offset,d_i1,k_i1_offset,
      &                 d_i0,k_i0_offset,maxh,maxp)
       else