Changeset 3765 for LMDZ6/branches/Optimisation_LMDZ

Timestamp:

Jul 15, 2020, 10:14:37 PM (4 years ago)

Author:

adurocher

Message:

Use cvl_comp_threshold to control whether to use compression or not

cvl_comp_threshold now defaults to 0 (never compress)

Location:

LMDZ6/branches/Optimisation_LMDZ/libf/phylmd

Files:

• conf_phys_m.F90 (modified) (4 diffs)
• cv3a_driver.f90 (modified) (3 diffs)
• cva_driver.F90 (modified) (1 diff)

LMDZ6/branches/Optimisation_LMDZ/libf/phylmd/conf_phys_m.F90

@@ -291 +291 @@
     !
     !Config Key  = ok_all_xml 
-    !Config Desc = utiliser les xml pourles définitions des champs pour xios
+    !Config Desc = utiliser les xml pourles d�finitions des champs pour xios
     !Config Def  = .FALSE.
     !Config Help = 
@@ -836 +836 @@
     !Config  Desc = Flag de fisrtilp
     !Config  Def  = 0
-    !Config  Help = Flag  pour la pluie grande-échelle les options suivantes existent :
+    !Config  Help = Flag  pour la pluie grande-�chelle les options suivantes existent :
     !Config         >1 nb iterations pour converger dans le calcul de qsat
     iflag_fisrtilp_qsat_omp = 0
@@ -844 +844 @@
     !Config  Desc = Flag de fisrtilp
     !Config  Def  = 0
-    !Config  Help = Flag  pour la pluie grande-échelle les options suivantes existent :
+    !Config  Help = Flag  pour la pluie grande-�chelle les options suivantes existent :
     !Config         0 pas d effet Bergeron
     !Config         1 effet Bergeron pour T<0
@@ -860 +860 @@
     !Config key  = cvl_comp_threshold
     !Config Desc = maximum fraction of convective points enabling compression
-    !Config Def  = 1.00
+    !Config Def  = 0.
     !Config Help = fields are compressed when less than a fraction cvl_comp_threshold
     !Config Help = of the points is convective.
-    cvl_comp_threshold_omp = 1.00
+    cvl_comp_threshold_omp = 0.
     CALL getin('cvl_comp_threshold', cvl_comp_threshold_omp)
 

LMDZ6/branches/Optimisation_LMDZ/libf/phylmd/cv3a_driver.f90

@@ -141 +141 @@
   SUBROUTINE cv3a_driver(len, nd, ndp1, ntra, nloc, k_upper, &
                          iflag_con, iflag_mix, iflag_ice_thermo, iflag_clos, ok_conserv_q, &
-                         delt, &
+                         delt, comp_threshold, &
                          t1, q1, qs1, t1_wake, q1_wake, qs1_wake, s1_wake, &
                          u1, v1, &
@@ -181 +181 @@
     REAL, INTENT(IN)                                  :: coefw_cld_cv ! coefficient for updraft velocity in convection
     REAL, INTENT(IN)                                  :: delt ! time step
+    REAL, INTENT (IN)                                 :: comp_threshold
     REAL, DIMENSION(len, nd), INTENT(IN)             :: t1 ! temperature (sat draught envt)
     REAL, DIMENSION(len, nd), INTENT(IN)             :: q1 ! specific hum (sat draught envt)
@@ -406 +407 @@
 ! ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-    compress = .false.
+    ! Compression has 3 possible modes:
+    ! 1) Compress = true :
+    compress_mode = COMPRESS_MODE_COMPRESS ! 1.1) Copy convective cells in contiguous array
+    !compress_mode = COMPRESS_MODE_COPY ! 1.2) Copy all cells and also compute on non-convective cells
+    ! 2) Compress = false : Don't copy and use original arrays, compute on non-convective cells
+    ! Never compress when comp_threshold = 0, always compress when comp_threshold = 1
+    ! comp_threshold = 0 (default) offers best performance when using many CPUs
+    compress = (comp_threshold == 1) .or. ( get_compress_size(len, (iflag1(:) == 0)) < len*comp_threshold ) 
+
     if (compress) then
-      compress_mode = COMPRESS_MODE_COMPRESS
-      !compress_mode = COMPRESS_MODE_COPY
-
       nloc = get_compress_size(len, (iflag1(:) == 0))
 

LMDZ6/branches/Optimisation_LMDZ/libf/phylmd/cva_driver.F90

@@ -127 +127 @@
     call cv3a_driver(len, nd, ndp1, ntra, nloc, k_upper, &
                 iflag_con, iflag_mix, iflag_ice_thermo, iflag_clos, ok_conserv_q, &
-                delt, &
+                delt, comp_threshold, &
                 t1, q1, qs1, t1_wake, q1_wake, qs1_wake, s1_wake, &
                 u1, v1, &