Improve workspace release in getrf*

icl-utk-edu · Dec 1, 2023 · 9aecfd2 · 9aecfd2
1 parent 686d38d
commit 9aecfd2
Show file tree

Hide file tree

Showing 3 changed files with 24 additions and 12 deletions.
diff --git a/src/getrf.cc b/src/getrf.cc
@@ -222,17 +222,21 @@ int64_t getrf(
             }
             #pragma omp task depend(inout:column[k])
             {
-                auto panel = A.sub( k, A_nt-1, k, k );
+                auto left_panel = A.sub( k, A_mt-1, k, k );
+                auto top_panel = A.sub( k, k, k+1, A_nt-1 );
 
                 // Erase remote tiles on all devices including host
-                panel.releaseRemoteWorkspace();
+                left_panel.releaseRemoteWorkspace();
+                top_panel.releaseRemoteWorkspace();
 
                 // Update the origin tiles before their
                 // workspace copies on devices are erased.
-                panel.tileUpdateAllOrigin();
+                left_panel.tileUpdateAllOrigin();
+                top_panel.tileUpdateAllOrigin();
 
                 // Erase local workspace on devices.
-                panel.releaseLocalWorkspace();
+                left_panel.releaseLocalWorkspace();
+                top_panel.releaseLocalWorkspace();
             }
             kk += A.tileNb( k );
         }

diff --git a/src/getrf_nopiv.cc b/src/getrf_nopiv.cc
@@ -206,17 +206,21 @@ int64_t getrf_nopiv(
             }
             #pragma omp task depend(inout:column[k])
             {
-                auto panel = A.sub(k, A_nt-1, k, k);
+                auto left_panel = A.sub( k, A_mt-1, k, k );
+                auto top_panel = A.sub( k, k, k+1, A_nt-1 );
 
                 // Erase remote tiles on all devices including host
-                panel.releaseRemoteWorkspace();
+                left_panel.releaseRemoteWorkspace();
+                top_panel.releaseRemoteWorkspace();
 
                 // Update the origin tiles before their
                 // workspace copies on devices are erased.
-                panel.tileUpdateAllOrigin();
+                left_panel.tileUpdateAllOrigin();
+                top_panel.tileUpdateAllOrigin();
 
                 // Erase local workspace on devices.
-                panel.releaseLocalWorkspace();
+                left_panel.releaseLocalWorkspace();
+                top_panel.releaseLocalWorkspace();
             }
             kk += A.tileNb( k );
         }

diff --git a/src/getrf_tntpiv.cc b/src/getrf_tntpiv.cc
@@ -334,17 +334,21 @@ int64_t getrf_tntpiv(
             }
             #pragma omp task depend(inout:column[k])
             {
-                auto panel = A.sub(k, A_nt-1, k, k);
+                auto left_panel = A.sub( k, A_mt-1, k, k );
+                auto top_panel = A.sub( k, k, k+1, A_nt-1 );
 
                 // Erase remote tiles on all devices including host
-                panel.releaseRemoteWorkspace();
+                left_panel.releaseRemoteWorkspace();
+                top_panel.releaseRemoteWorkspace();
 
                 // Update the origin tiles before their
                 // workspace copies on devices are erased.
-                panel.tileUpdateAllOrigin();
+                left_panel.tileUpdateAllOrigin();
+                top_panel.tileUpdateAllOrigin();
 
                 // Erase local workspace on devices.
-                panel.releaseLocalWorkspace();
+                left_panel.releaseLocalWorkspace();
+                top_panel.releaseLocalWorkspace();
             }
             kk += A.tileNb( k );
         }