add 2 phase ccsd_t2_7_3 and broken cublas of the same

jeffhammond · Oct 25, 2024 · 9d4040a · 9d4040a
1 parent 239ce8b
commit 9d4040a
Show file tree

Hide file tree

Showing 3 changed files with 546 additions and 9 deletions.
diff --git a/src/tce/ccsd/ccsd_t2.F b/src/tce/ccsd/ccsd_t2.F
@@ -302,26 +302,26 @@ SUBROUTINE ccsd_t2(rtdb,d_f1,d_i0,d_t1,d_t2,d_v2,
       CALL OFFSET_ccsd_t2_7_1(l_i1_offset,k_i1_offset,size_i1)
       CALL TCE_FILENAME('ccsd_t2_7_1_i1',filename)
       CALL CREATEFILE(filename,d_i1,size_i1)
+      CALL ccsd_t2_7_1(d_v2,k_v2_offset,d_i1,k_i1_offset)
       if (ccsd_t2_7_alg.eq.1) then
-        CALL ccsd_t2_7_1(d_v2,k_v2_offset,d_i1,k_i1_offset)
         CALL ccsd_t2_7_2(d_t1,k_t1_offset,d_v2,k_v2_offset,
      &                   d_i1,k_i1_offset)
         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)
+     &                   d_i1,k_i1_offset)
         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.4) then
+        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_y(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)