Enable use of IBM XL compilers with openMP Offload support

lanl · Apr 7, 2022 · 12d4b86 · 12d4b86
1 parent 986ac31
commit 12d4b86
Show file tree

Hide file tree

Showing 7 changed files with 100 additions and 46 deletions.
diff --git a/src/C-interface/ellblock/bml_add_ellblock_typed.c b/src/C-interface/ellblock/bml_add_ellblock_typed.c
@@ -45,7 +45,6 @@ void TYPED_FUNC(
 
     int NB = A->NB;
     int MB = A->MB;
-    int ix[NB], jx[NB];
 
     int *A_nnzb = A->nnzb;
     int *A_indexb = A->indexb;
@@ -55,12 +54,16 @@ void TYPED_FUNC(
 
     int *bsize = A->bsize;
 
-    REAL_T *x_ptr[NB];
     REAL_T **A_ptr_value = (REAL_T **) A->ptr_value;
     REAL_T **B_ptr_value = (REAL_T **) B->ptr_value;
 
+#if !(defined(__IBMC__) || defined(__ibmxl__))
+    int ix[NB], jx[NB];
+    REAL_T *x_ptr[NB];
+
     memset(ix, 0, NB * sizeof(int));
     memset(jx, 0, NB * sizeof(int));
+#endif
 
     int maxbsize = 0;
     for (int ib = 0; ib < NB; ib++)
@@ -74,13 +77,29 @@ void TYPED_FUNC(
     REAL_T *x_ptr_storage = calloc(maxbsize2 * NB * nthreads, sizeof(REAL_T));
 
     char xptrset = 0;
+#if defined(__IBMC__) || defined(__ibmxl__)
+#pragma omp parallel for \
+    shared(A_indexb, A_ptr_value, A_nnzb)       \
+    shared(B_indexb, B_ptr_value, B_nnzb)       \
+    shared(x_ptr_storage) \
+    firstprivate(xptrset)
+#else
 #pragma omp parallel for                  \
     shared(A_indexb, A_ptr_value, A_nnzb)       \
     shared(B_indexb, B_ptr_value, B_nnzb)       \
     shared(x_ptr_storage) \
     firstprivate(ix, jx, x_ptr, xptrset)
+#endif
     for (int ib = 0; ib < NB; ib++)
     {
+
+#if defined(__IBMC__) || defined(__ibmxl__)
+        int ix[NB], jx[NB];
+        REAL_T *x_ptr[NB];
+
+        memset(ix, 0, NB * sizeof(int));
+#endif
+
         if (!xptrset)
         {
 #ifdef _OPENMP

diff --git a/src/C-interface/ellblock/bml_multiply_ellblock_typed.c b/src/C-interface/ellblock/bml_multiply_ellblock_typed.c
@@ -255,18 +255,20 @@ void *TYPED_FUNC(
     int *X2_indexb = X2->indexb;
     int *X2_nnzb = X2->nnzb;
 
-    int ix[NB], jx[NB];
-    REAL_T *x_ptr[NB];
-
     REAL_T traceX = 0.0;
     REAL_T traceX2 = 0.0;
     REAL_T **X_ptr_value = (REAL_T **) X->ptr_value;
     REAL_T **X2_ptr_value = (REAL_T **) X2->ptr_value;
 
     double *trace = bml_allocate_memory(sizeof(double) * 2);
 
+#if !(defined(__IBMC__) || defined(__ibmxl__))
+    int ix[NB], jx[NB];
+    REAL_T *x_ptr[NB];
+
     memset(ix, 0, NB * sizeof(int));
     memset(jx, 0, NB * sizeof(int));
+#endif
 
     int maxbsize = 0;
     for (int ib = 0; ib < NB; ib++)
@@ -297,14 +299,26 @@ void *TYPED_FUNC(
             TYPED_FUNC(bml_multiply_block4),
             TYPED_FUNC(bml_multiply_block5), TYPED_FUNC(bml_multiply_block6)};
 
-
+#if defined(__IBMC__) || defined(__ibmxl__)
+#pragma omp parallel for                           \
+    firstprivate(xptrset)            \
+    reduction(+: traceX, traceX2)
+#else
 #pragma omp parallel for                           \
     firstprivate(ix,jx, x_ptr, xptrset)            \
     reduction(+: traceX, traceX2)
-
+#endif
     //loop over row blocks
     for (int ib = 0; ib < NB; ib++)
     {
+
+#if defined(__IBMC__) || defined(__ibmxl__)
+        int ix[NB], jx[NB];
+        REAL_T *x_ptr[NB];
+
+        memset(ix, 0, NB * sizeof(int));
+#endif
+
         int lb = 0;
         if (!xptrset)
         {
@@ -456,15 +470,17 @@ void TYPED_FUNC(
     int *C_nnzb = C->nnzb;
     int *C_indexb = C->indexb;
 
-    int ix[NB], jx[NB];
-    REAL_T *x_ptr[NB];
-
     REAL_T **A_ptr_value = (REAL_T **) A->ptr_value;
     REAL_T **B_ptr_value = (REAL_T **) B->ptr_value;
     REAL_T **C_ptr_value = (REAL_T **) C->ptr_value;
 
+#if !(defined(__IBMC__) || defined(__ibmxl__))
+    int ix[NB], jx[NB];
+    REAL_T *x_ptr[NB];
+
     memset(ix, 0, NB * sizeof(int));
     memset(jx, 0, NB * sizeof(int));
+#endif
 
     int maxbsize = 0;
     for (int ib = 0; ib < NB; ib++)
@@ -495,11 +511,24 @@ void TYPED_FUNC(
             TYPED_FUNC(bml_multiply_block5), TYPED_FUNC(bml_multiply_block6)};
 
     //loop over row blocks
+#if defined(__IBMC__) || defined(__ibmxl__)
+#pragma omp parallel for                       \
+    firstprivate( xptrset)
+#else
 #pragma omp parallel for                       \
     firstprivate(ix, jx, x_ptr, xptrset)
+#endif
 
     for (int ib = 0; ib < NB; ib++)
     {
+
+#if defined(__IBMC__) || defined(__ibmxl__)
+        int ix[NB], jx[NB];
+        REAL_T *x_ptr[NB];
+
+        memset(ix, 0, NB * sizeof(int));
+#endif
+
         int lb = 0;
         if (!xptrset)
         {

diff --git a/src/C-interface/ellpack/bml_add_ellpack_typed.c b/src/C-interface/ellpack/bml_add_ellpack_typed.c
@@ -66,7 +66,7 @@ void TYPED_FUNC(
     memset(x, 0.0, N * sizeof(REAL_T));
 #endif
 
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
     int num_chunks = MIN(OFFLOAD_NUM_CHUNKS, rowMax - rowMin + 1);
 
     int all_ix[N * num_chunks], all_jx[N * num_chunks];
@@ -81,7 +81,7 @@ void TYPED_FUNC(
 #endif
 
 #if defined (USE_OMP_OFFLOAD)
-#if defined(INTEL_SDK) || defined(CRAY_SDK)
+#if defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__)
 #pragma omp teams distribute parallel for \
     shared(rowMin, rowMax)                \
     shared(A_index, A_value, A_nnz)       \
@@ -116,19 +116,20 @@ void TYPED_FUNC(
     firstprivate(ix, jx, x)
 #endif
 #endif
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
     for (int i = rowMin + chunk; i < rowMax; i = i + num_chunks)
+    {
 #else
     for (int i = rowMin; i < rowMax; i++)
-#endif
     {
+
 #if defined(__IBMC__) || defined(__ibmxl__)
         int ix[N], jx[N];
         REAL_T x[N];
 
         memset(ix, 0, N * sizeof(int));
 #endif
-
+#endif
         int l = 0;
         if (alpha > (double) 0.0 || alpha < (double) 0.0)
             for (int jp = 0; jp < A_nnz[i]; jp++)
@@ -175,7 +176,7 @@ void TYPED_FUNC(
         }
         A_nnz[i] = ll;
     }
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
 }
 #endif
 }
@@ -234,7 +235,7 @@ double TYPED_FUNC(
     memset(y, 0.0, N * sizeof(REAL_T));
 #endif
 
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
     int num_chunks = MIN(OFFLOAD_NUM_CHUNKS, rowMax - rowMin + 1);
 
     int all_ix[N * num_chunks], all_jx[N * num_chunks];
@@ -250,7 +251,7 @@ double TYPED_FUNC(
 #endif
 
 #if defined (USE_OMP_OFFLOAD)
-#if defined(INTEL_SDK) || defined(CRAY_SDK)
+#if defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__)
 #pragma omp teams distribute parallel for \
     shared(rowMin, rowMax)                \
     shared(A_index, A_value, A_nnz)       \
@@ -291,11 +292,11 @@ double TYPED_FUNC(
     reduction(+:trnorm)
 #endif
 #endif
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
     for (int i = rowMin + chunk; i < rowMax; i = i + num_chunks)
+    {
 #else
     for (int i = rowMin; i < rowMax; i++)
-#endif
     {
 
 #if defined(__IBMC__) || defined(__ibmxl__)
@@ -305,7 +306,7 @@ double TYPED_FUNC(
 
         memset(ix, 0, N * sizeof(int));
 #endif
-
+#endif
         int l = 0;
         for (int jp = 0; jp < A_nnz[i]; jp++)
         {
@@ -359,7 +360,7 @@ double TYPED_FUNC(
         }
         A_nnz[i] = ll;
     }
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
 }
 #endif
 
@@ -397,7 +398,7 @@ void TYPED_FUNC(
     memset(x, 0.0, A_M * sizeof(REAL_T));
 #endif
 
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
     int num_chunks = MIN(OFFLOAD_NUM_CHUNKS, N);
 
     int all_jx[N * num_chunks];
@@ -411,7 +412,7 @@ void TYPED_FUNC(
 #endif
 
 #if defined (USE_OMP_OFFLOAD)
-#if defined(INTEL_SDK) || defined(CRAY_SDK)
+#if defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__)
 #pragma omp teams distribute parallel for \
     shared(N, A_M)                \
     shared(A_index, A_value, A_nnz)
@@ -441,18 +442,18 @@ void TYPED_FUNC(
     firstprivate(jx, x)
 #endif
 #endif
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
     for (int i = chunk; i < N; i = i + num_chunks)
+    {
 #else
     for (int i = 0; i < N; i++)
-#endif
     {
 
 #if defined(__IBMC__) || defined(__ibmxl__)
         int jx[A_M];
         REAL_T x[A_M];
 #endif
-
+#endif
         int l = 0;
         int diag = -1;
 
@@ -495,7 +496,7 @@ void TYPED_FUNC(
         }
         A_nnz[i] = ll;
     }
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
 }
 #endif
 }

diff --git a/src/C-interface/ellpack/bml_element_multiply_ellpack_typed.c b/src/C-interface/ellpack/bml_element_multiply_ellpack_typed.c
@@ -74,7 +74,7 @@ void TYPED_FUNC(
     memset(x, 0.0, C->N * sizeof(REAL_T));
 #endif
 
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
     int num_chunks = MIN(OFFLOAD_NUM_CHUNKS, rowMax - rowMin + 1);
 
     int all_ix[C_N * num_chunks], all_jx[C_N * num_chunks];
@@ -89,7 +89,7 @@ void TYPED_FUNC(
 #endif
 
 #if defined (USE_OMP_OFFLOAD)
-#if defined(INTEL_SDK) || defined(CRAY_SDK)
+#if defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__)
 #pragma omp teams distribute parallel for \
     shared(A_N, A_M, A_nnz, A_index, A_value)  \
     shared(A_localRowMin, A_localRowMax)       \
@@ -129,19 +129,19 @@ void TYPED_FUNC(
 #endif
 #endif
     //for (int i = 0; i < A_N; i++)
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
     for (int i = rowMin + chunk; i < rowMax; i = i + num_chunks)
+    {
 #else
     for (int i = rowMin; i < rowMax; i++)
-#endif
     {
 #if defined(__IBMC__) || defined(__ibmxl__)
         int ix[C_N], jx[C_N];
         REAL_T x[C_N];
 
         memset(ix, 0, C_N * sizeof(int));
 #endif
-
+#endif
         int l = 0;
         for (int jp = 0; jp < A_nnz[i]; jp++)
         {
@@ -198,7 +198,7 @@ void TYPED_FUNC(
         }
         C_nnz[i] = ll;
     }
-#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK))
+#if defined(USE_OMP_OFFLOAD) && (defined(INTEL_SDK) || defined(CRAY_SDK) || defined(__IBMC__) || defined(__ibmxl__))
 }
 #endif
 }