option to hand out a built context

README.org

@@ -7,3 +7,8 @@ This is a mini-example of distributed gemm for starpu mpi, both in Fortran and C
   make -j
   make install
 #+end_src
+
+
+* TODO
+
+- add warmup run

fstarpu_example_dgemm.f90

@@ -39,6 +39,8 @@ program fstarpu_example_dgemm
   logical                         :: super_prune = .false.
   logical                         :: prune_handles = .false.
   logical                         :: delay = .false.
+  logical                         :: provide_context = .false.
+  logical                         :: warmup = .true.
   integer(c_int)                  :: comm_size, comm_rank
   integer(c_int), target          :: comm_world
 
@@ -58,6 +60,11 @@ program fstarpu_example_dgemm
   integer                         :: te, ts, tr
   real                            :: tf, gflops
 
+  integer(c_int), dimension(:), allocatable :: procs
+  integer(c_int)                            :: ctx
+  integer(c_int),target                     :: arg_ctx
+  character(kind=c_char,len=4), target      :: ctx_policy = C_CHAR_"lws"//C_NULL_CHAR
+
   write(*,*) "initializing starpu ..."
   ret = fstarpu_init(C_NULL_PTR)
   if (ret == -19) then
@@ -117,6 +124,11 @@ program fstarpu_example_dgemm
         prune_handles = .true.
       case('-d')
         delay = .true.
+      case('-c')
+        provide_context = .true.
+      case('-now')
+        warmup = .false.
+      ! keep -e as an empty argument for debug purpose
     end select
   end do
   
@@ -133,12 +145,14 @@ program fstarpu_example_dgemm
      write(*,'("mbxnbxkb = ",i5,"x",i5,"x",i5)') mb, nb, kb
      write(*,'("B        = ",i5)') bs
      write(*,'("PxQ      = ",i3,"x",i3)') p,q
-     if (trace)         write(*,*) "(T)racing enabled"
-     if (lflush)        write(*,*) "(F)lushing enabled"
-     if (super_prune)   write(*,*) "(S)uper-pruning enabled"
-     if (prune)         write(*,*) "(P)runing enabled"
-     if (prune_handles) write(*,*) "(H)andles pruning enabled"
-     if (delay)         write(*,*) "(D)elayed handle registration enabled"
+     if (trace)           write(*,*) "(T)racing enabled"
+     if (lflush)          write(*,*) "(F)lushing enabled"
+     if (super_prune)     write(*,*) "(S)uper-pruning enabled"
+     if (prune)           write(*,*) "(P)runing enabled"
+     if (prune_handles)   write(*,*) "(H)andles pruning enabled"
+     if (delay)           write(*,*) "(D)elayed handle registration enabled"
+     if (provide_context) write(*,*) "(C)ontext provided at submission"
+     if (.not.warmup)       write(*,*) "(W)armup disabled"
      write(*,'("========================================")')
   end if
   ret = fstarpu_mpi_barrier(comm_world)
@@ -148,6 +162,14 @@ program fstarpu_example_dgemm
   alpha = 0.42
   beta = 3.14
 
+  if (provide_context) then
+    allocate(procs(ncpu))
+    err = fstarpu_worker_get_ids_by_type(FSTARPU_CPU_WORKER, procs, ncpu)
+    ctx = fstarpu_sched_ctx_create(procs, ncpu, C_CHAR_"stdalone"//C_NULL_CHAR,&
+        (/ FSTARPU_SCHED_CTX_POLICY_NAME, c_loc(ctx_policy), c_null_ptr /) )
+  end if
+
+  if (warmup) t = t + 1
   do trial=1,t
      ! allocate matrices
      call initialize_matrix(A,mb,kb,"A",.true. ,.false.)
@@ -186,13 +208,25 @@ program fstarpu_example_dgemm
                 call block_register(B,l,j)
                 call block_register(C,i,j)
              end if
-             call fstarpu_mpi_task_insert((/ c_loc(comm_world), cl_mm, &
+             if (provide_context) then
+                 arg_ctx = ctx
+                 call fstarpu_mpi_task_insert((/ c_loc(comm_world), cl_mm, &
+                    FSTARPU_VALUE, c_loc(alpha), FSTARPU_SZ_REAL8,       &
+                    FSTARPU_VALUE, c_loc(zbeta), FSTARPU_SZ_REAL8,       &
+                    FSTARPU_R,  A%blocks(i,l)%h,                         &
+                    FSTARPU_R,  B%blocks(l,j)%h,                         &
+                    FSTARPU_RW, C%blocks(i,j)%h,                         &
+                    FSTARPU_SCHED_CTX, c_loc(arg_ctx),            &
+                    c_null_ptr /))
+             else 
+               call fstarpu_mpi_task_insert((/ c_loc(comm_world), cl_mm, &
                   FSTARPU_VALUE, c_loc(alpha), FSTARPU_SZ_REAL8,       &
                   FSTARPU_VALUE, c_loc(zbeta), FSTARPU_SZ_REAL8,       &
                   FSTARPU_R,  A%blocks(i,l)%h,                         &
                   FSTARPU_R,  B%blocks(l,j)%h,                         &
                   FSTARPU_RW, C%blocks(i,j)%h,                         &
                   c_null_ptr /))
+             end if
            else
               !could write something 
            end if
@@ -211,7 +245,8 @@ program fstarpu_example_dgemm
      call system_clock(te,tr)
      tf = max(real(te-ts)/real(tr),1e-20)
      gflops = 2.0*m*n*k/(tf*10**9)
-     if (comm_rank.eq.0) write(*,'("RANK ",i3," -> took ",e15.8," s | ", e15.8," Gflop/s")') &
+     if (comm_rank.eq.0.and.(.not.warmup.or.trial.gt.1)) &
+        write(*,'("RANK ",i3," -> took ",e15.8," s | ", e15.8," Gflop/s")') &
              comm_rank, tf, gflops
      
      ! unregister matrices
@@ -220,6 +255,10 @@ program fstarpu_example_dgemm
      call unregister_matrix(C,mb,nb)
   end do
   
+  if (provide_context) then
+     call fstarpu_sched_ctx_delete(ctx)
+     deallocate(procs)
+  endif
   
   call fstarpu_codelet_free(cl_mm)
   call fstarpu_codelet_free(cl_fill)

starpu_example_dgemm.c

@@ -273,7 +273,7 @@ static void cpu_gemm(void *handles[], void *args)
 //	if (verbose) printf("gemm_task\n");
 //	printf("DATA %d | ld A %d B %d C %d | alpha %f beta %f \n", datatype, ld_A, ld_B, ld_C, clargs->alpha, clargs->beta);
   double start = starpu_timing_now();
-  cblas_dgemm( CblasColMajor, CblasNoTrans, CblasNoTrans, // 2
+  cblas_dgemm( CblasRowMajor, CblasNoTrans, CblasNoTrans, // 2
                n_row_C, n_col_C, n_col_A, clargs->alpha, block_A, ld_A, block_B, // 9
                ld_B, clargs->beta, block_C, ld_C ); // 13
   double stop = starpu_timing_now();