enable coalesced access in pre & postFDM and bug fix in fusedFDM (#121)

Nek5000 · Jul 25, 2020 · 04f17fd · 04f17fd
1 parent 1462cd1
commit 04f17fd
Showing 1 changed file with 13 additions and 17 deletions.
diff --git a/src/libP/solvers/elliptic/okl/ellipticSchwarzSolverHex3D.okl b/src/libP/solvers/elliptic/okl/ellipticSchwarzSolverHex3D.okl
@@ -26,11 +26,10 @@
       }
     }
     @barrier("local");
-    for(int k = 0; k < p_Nq_e; ++k; @inner) {
+    for(int k = 0; k < p_Nq; ++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_e; ++i; @inner) {
+          if(i < p_Nq && j < p_Nq) {
             const dlong elem_offset = elem * p_Nq * p_Nq * p_Nq;
             const dlong idx = i + j * p_Nq + k * p_Nq * p_Nq + elem_offset;
             sWork1[k + 1][j + 1][i + 1] = static_cast < pfloat > (u[idx]);
@@ -66,10 +65,9 @@
           sWork1[i][j][p_Nq_e - l1 - 1] = sWork1[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 = p_Nq_e * p_Nq_e * p_Nq_e * elem;
           const dlong idx = i + j * p_Nq_e + k * p_Nq_e * p_Nq_e + elem_offset;
           work1[idx] = sWork1[k][j][i];
@@ -86,10 +84,9 @@
   for (dlong elem = 0; elem < Nelements; ++elem; @outer) {
     @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 = elem * 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] = my_work2[idx];
@@ -158,11 +155,10 @@
                                          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; ++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_e; ++i; @inner) {
+          if(i < p_Nq && j < p_Nq) {
             const dlong elem_offset = elem * 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];
@@ -363,10 +359,10 @@
     }
 #else  /* if (!p_restrict) */
     @barrier("local");
-    for(int k = 0; k < p_Nq_e; ++k) {
+    for(int k = 0; k < p_Nq; ++k) {
       for(int j = 0; j < p_Nq_e; ++j; @inner) {
-        for(int i = 0; i < p_Nq; ++i; @inner) {
-          if(k < p_Nq && j < p_Nq) {
+        for(int i = 0; i < p_Nq_e; ++i; @inner) {
+          if(i < 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];