OpenMP tasks

Makefile

@@ -1,6 +1,7 @@
 IRPF90 = irpf90/bin/irpf90 --codelet=factor_een:2 --align=4096 # -s nelec_8:504 -s nnuc:100 -s ncord:5 #-a -d
 #FC     = ifort -xCORE-AVX512 -g -mkl=sequential  -qopt-zmm-usage=high 
 FC     = ifort -xCORE-AVX2 -g
+CC     = gcc -fopenmp
 FCFLAGS= -O3 -I .
 NINJA  = ninja
 ARCHIVE = ar crs
@@ -8,7 +9,7 @@ RANLIB = ranlib
 
 SRC= IRPF90_temp/qmckl_blas_f.f90 IRPF90_temp/qmckl_dgemm.c
 OBJ= IRPF90_temp/qmckl_blas_f.o   IRPF90_temp/qmckl_dgemm.o
-LIB= -mkl=sequential
+LIB= -mkl=sequential -lgomp
 
 -include irpf90.make
 export

qmckl_dgemm.c

@@ -27,10 +27,13 @@ static void qmckl_dgemm_rec(struct dgemm_args args) {
 //  printf("%5d %5d\n", args.m, args.n);
 
   if ( (args.m <= MIN_SIZE) || (args.n <= MIN_SIZE)) {
+    #pragma omp task
+    {
       cblas_dgemm(CblasColMajor, args.transa, args.transb,
                   args.m, args.n, args.k, args.alpha,
                   args.A, args.lda, args.B, args.ldb,
                   args.beta, args.C, args.ldc);
+    }
   } else {
 
     int m1 = args.m / 2;
@@ -38,11 +41,16 @@ static void qmckl_dgemm_rec(struct dgemm_args args) {
     int n1 = args.n / 2;
     int n2 = args.n - n1;
       
+    #pragma omp task
+    {
       struct dgemm_args args_1 = args;
       args_1.m = m1;
       args_1.n = n1;
       qmckl_dgemm_rec(args_1);
+    }
     
+    #pragma omp task
+    {
       // TODO: assuming 'N', 'N' here
       struct dgemm_args args_2 = args;
       args_2.B = args.B + args.ldb*n1;
@@ -50,14 +58,20 @@ static void qmckl_dgemm_rec(struct dgemm_args args) {
       args_2.m = m1;
       args_2.n = n2;
       qmckl_dgemm_rec(args_2);
+    }
     
+    #pragma omp task
+    {
       struct dgemm_args args_3 = args;
       args_3.A = args.A + m1;
       args_3.C = args.C + m1;
       args_3.m = m2;
       args_3.n = n1;
       qmckl_dgemm_rec(args_3);
+    }
     
+    #pragma omp task
+    {
       struct dgemm_args args_4 = args;
       args_4.A = args.A + m1;
       args_4.B = args.B + args.ldb*n1;
@@ -66,6 +80,7 @@ static void qmckl_dgemm_rec(struct dgemm_args args) {
       args_4.n = n2;
       qmckl_dgemm_rec(args_4);
     }
+  }
 
 }
 
@@ -104,5 +119,12 @@ void qmckl_dgemm(char transa, char transb,
     args.transb = CblasNoTrans;
   }
 
+  #pragma omp parallel
+  {
+    #pragma omp single
+    {
       qmckl_dgemm_rec(args);
     }
+    #pragma omp taskwait
+  }
+}

qmckl_dgemm.org

@@ -92,8 +92,15 @@ void qmckl_dgemm(char transa, char transb,
     args.transb = CblasNoTrans;
   }
 
+  #pragma omp parallel
+  {
+    #pragma omp single
+    {
       qmckl_dgemm_rec(args);
     }
+    #pragma omp taskwait
+  }
+}
 
   #+END_SRC
 
@@ -108,10 +115,13 @@ static void qmckl_dgemm_rec(struct dgemm_args args) {
 //  printf("%5d %5d\n", args.m, args.n);
 
   if ( (args.m <= MIN_SIZE) || (args.n <= MIN_SIZE)) {
+    #pragma omp task
+    {
       cblas_dgemm(CblasColMajor, args.transa, args.transb,
                   args.m, args.n, args.k, args.alpha,
                   args.A, args.lda, args.B, args.ldb,
                   args.beta, args.C, args.ldc);
+    }
   } else {
 
     int m1 = args.m / 2;
@@ -119,11 +129,16 @@ static void qmckl_dgemm_rec(struct dgemm_args args) {
     int n1 = args.n / 2;
     int n2 = args.n - n1;
       
+    #pragma omp task
+    {
       struct dgemm_args args_1 = args;
       args_1.m = m1;
       args_1.n = n1;
       qmckl_dgemm_rec(args_1);
+    }
     
+    #pragma omp task
+    {
       // TODO: assuming 'N', 'N' here
       struct dgemm_args args_2 = args;
       args_2.B = args.B + args.ldb*n1;
@@ -131,14 +146,20 @@ static void qmckl_dgemm_rec(struct dgemm_args args) {
       args_2.m = m1;
       args_2.n = n2;
       qmckl_dgemm_rec(args_2);
+    }
     
+    #pragma omp task
+    {
       struct dgemm_args args_3 = args;
       args_3.A = args.A + m1;
       args_3.C = args.C + m1;
       args_3.m = m2;
       args_3.n = n1;
       qmckl_dgemm_rec(args_3);
+    }
     
+    #pragma omp task
+    {
       struct dgemm_args args_4 = args;
       args_4.A = args.A + m1;
       args_4.B = args.B + args.ldb*n1;
@@ -147,6 +168,7 @@ static void qmckl_dgemm_rec(struct dgemm_args args) {
       args_4.n = n2;
       qmckl_dgemm_rec(args_4);
     }
+  }
 
 }
   #+END_SRC