fixes from reference version

Author: Michael Boulton

data.f90

@@ -23,7 +23,7 @@ MODULE data_module
 
    IMPLICIT NONE
 
-   REAL(KIND=8), PARAMETER :: g_version=2.0
+   REAL(KIND=8), PARAMETER :: g_version=1.2
 
    INTEGER,      PARAMETER :: g_ibig=640000
 

ocl_common.hpp

@@ -7,7 +7,7 @@
 #include <cstdlib>
 #include <map>
 
-#define JACOBI_BLOCK_SIZE 4
+#define JACOBI_BLOCK_SIZE 1
 
 // 2 dimensional arrays - use a 2D tile for local group
 const static size_t LOCAL_Y = JACOBI_BLOCK_SIZE;

start.f90

@@ -51,11 +51,11 @@ SUBROUTINE start
 
   CALL tea_get_num_chunks(number_of_chunks)
 
-  ALLOCATE(chunks(1:number_of_chunks))
-  ALLOCATE(left(1:number_of_chunks))
-  ALLOCATE(right(1:number_of_chunks))
-  ALLOCATE(bottom(1:number_of_chunks))
-  ALLOCATE(top(1:number_of_chunks))
+  ALLOCATE(chunks(1:chunks_per_task))
+  ALLOCATE(left(1:chunks_per_task))
+  ALLOCATE(right(1:chunks_per_task))
+  ALLOCATE(bottom(1:chunks_per_task))
+  ALLOCATE(top(1:chunks_per_task))
 
   CALL tea_decompose(grid%x_cells,grid%y_cells,left,right,bottom,top)
 

tea.f90

@@ -237,14 +237,15 @@ SUBROUTINE tea_exchange(fields,depth)
     LOGICAL :: test_complete
 
     ! Assuming 1 patch per task, this will be changed
+    chunk = 1
+
+    IF (ALL(chunks(chunk)%chunk_neighbours .eq. external_face)) return
 
     request_lr=0
     message_count_lr=0
     request_ud = 0
     message_count_ud = 0
 
-    chunk = 1
-
     end_pack_index_left_right=0
     end_pack_index_bottom_top=0
     left_right_offset = 0

tea.in

@@ -19,7 +19,7 @@ tl_preconditioner_type=jac_block
 tl_ppcg_inner_steps=249
 tl_check_result
 tl_eps=1.0e-12
-profiler_on
+!profiler_on
 
 !use_fortran_kernels
 

tea_solve.f90

@@ -325,6 +325,13 @@ SUBROUTINE tea_leaf()
             IF (cheby_calc_steps .EQ. 0) THEN
               cheby_calc_steps = 1
 
+              fields=0
+              fields(FIELD_U) = 1
+              IF (profiler_on) halo_time=timer()
+              CALL update_halo(fields,1)
+              !IF (profiler_on) profiler%halo_exchange = profiler%halo_exchange + (timer() - halo_time)
+              IF (profiler_on) init_time=init_time+(timer()-halo_time)
+
               IF(use_opencl_kernels) THEN
                 call tea_leaf_kernel_ppcg_init_ocl(ch_alphas, ch_betas, &
                     theta, tl_ppcg_inner_steps)
@@ -344,12 +351,6 @@ SUBROUTINE tea_leaf()
               ELSEIF(use_opencl_kernels) THEN
                 CALL tea_leaf_calc_residual_ocl()
               ENDIF
-
-              IF (profiler_on) halo_time = timer()
-              ! update p
-              CALL update_halo(fields,1)
-              !IF (profiler_on) profiler%halo_exchange = profiler%halo_exchange + (timer() - halo_time)
-              IF (profiler_on) solve_time = solve_time + (timer()-halo_time)
             ENDIF
 
             IF(use_fortran_kernels) THEN
@@ -556,6 +557,12 @@ SUBROUTINE tea_leaf()
       ENDDO
 
       IF (tl_check_result) THEN
+        fields = 0
+        fields(FIELD_U) = 1
+        IF (profiler_on) halo_time = timer()
+        CALL update_halo(fields,1)
+        !IF (profiler_on) profiler%halo_exchange = profiler%halo_exchange + (timer() - halo_time)
+        IF (profiler_on) solve_time = solve_time + (timer()-halo_time)
         IF(use_fortran_kernels) THEN
           CALL tea_leaf_calc_residual(chunks(c)%field%x_min,&
               chunks(c)%field%x_max,                        &