enable coalesced access in fusedFDM (#119)

Nek5000 · Jul 24, 2020 · ef4a051 · ef4a051
1 parent ec92df4
commit ef4a051
Showing 1 changed file with 12 additions and 17 deletions.
diff --git a/src/libP/solvers/elliptic/okl/ellipticSchwarzSolverHex3D.okl b/src/libP/solvers/elliptic/okl/ellipticSchwarzSolverHex3D.okl
@@ -189,10 +189,9 @@
     @shared pfloat S_z_eT[p_Nq_e][p_Nq_e];
     @shared pfloat work1[p_Nq_e][p_Nq_e][p_Nq_e];
     @shared pfloat work2[p_Nq_e][p_Nq_e][p_Nq_e];
-    for(int k = 0; k < p_Nq_e; ++k; @inner) {
+    for(int k = 0; k < p_Nq_e; ++k) {
       for(int j = 0; j < p_Nq_e; ++j; @inner) {
-#pragma unroll
-        for(int i = 0; i < p_Nq_e; ++i) {
+        for(int i = 0; i < p_Nq_e; ++i; @inner) {
           const dlong elem_offset = e * p_Nq_e * p_Nq_e * p_Nq_e;
           const dlong idx = i + j * p_Nq_e + k * p_Nq_e * p_Nq_e + elem_offset;
           work1[k][j][i] = u[idx];
@@ -268,10 +267,9 @@
       }
     }
     @barrier("local");
-    for (int k = 0; k < p_Nq_e; k++; @inner) {
+    for (int k = 0; k < p_Nq_e; k++) {
       for (int j = 0; j < p_Nq_e; j++; @inner) {
-#pragma unroll
-        for (int i = 0; i < p_Nq_e; i++) {
+        for (int i = 0; i < p_Nq_e; i++; @inner) {
           const int v = i + j * p_Nq_e + k * p_Nq_e * p_Nq_e;
           pfloat value = 0.0;
 #pragma unroll
@@ -308,10 +306,9 @@
       }
     }
     @barrier("local");
-    for (int k = 0; k < p_Nq_e; k++; @inner) {
+    for (int k = 0; k < p_Nq_e; k++) {
       for (int j = 0; j < p_Nq_e; j++; @inner) {
-#pragma unroll
-        for (int i = 0; i < p_Nq_e; i++) {
+        for (int i = 0; i < p_Nq_e; i++; @inner) {
           pfloat value = 0.0;
 #pragma unroll
           for (int l = 0; l < p_Nq_e; l++)
@@ -355,10 +352,9 @@
           work2[i][j][p_Nq_e - l1 - 1] = work1[i][j][p_Nq_e - l2 - 1];
         }
     @barrier("local");
-    for(int k = 0; k < p_Nq_e; ++k; @inner) {
+    for(int k = 0; k < p_Nq_e; ++k) {
       for(int j = 0; j < p_Nq_e; ++j; @inner) {
-#pragma unroll
-        for(int i = 0; i < p_Nq_e; ++i) {
+        for(int i = 0; i < p_Nq_e; ++i; @inner) {
           const dlong elem_offset = e * p_Nq_e * p_Nq_e * p_Nq_e;
           const dlong idx = i + j * p_Nq_e + k * p_Nq_e * p_Nq_e + elem_offset;
           u[idx] = work2[k][j][i];
@@ -367,11 +363,10 @@
     }
 #else  /* if (!p_restrict) */
     @barrier("local");
-    for(int k = 0; k < p_Nq_e; ++k; @inner) {
+    for(int k = 0; k < p_Nq_e; ++k) {
       for(int j = 0; j < p_Nq_e; ++j; @inner) {
-        if(k < p_Nq && j < p_Nq) {
-#pragma unroll
-          for(int i = 0; i < p_Nq; ++i) {
+        for(int i = 0; i < p_Nq; ++i; @inner) {
+          if(k < p_Nq && j < p_Nq) {
             const dlong elem_offset = e * p_Nq * p_Nq * p_Nq;
             const dlong idx = i + j * p_Nq + k * p_Nq * p_Nq + elem_offset;
             Su[idx] = work1[k + 1][j + 1][i + 1];
@@ -381,4 +376,4 @@
     }
 #endif
   }
-}
+}