Adding computation offloading support for different OP and Datatypes

pmodels · Jun 26, 2019 · 4eea475 · 4eea475
1 parent eeace26
commit 4eea475
Show file tree

Hide file tree

Showing 11 changed files with 32 additions and 24 deletions.
diff --git a/src/include/mpir_op_util.h b/src/include/mpir_op_util.h
@@ -8,6 +8,9 @@
 #ifndef MPIR_OP_UTIL_H_INCLUDED
 #define MPIR_OP_UTIL_H_INCLUDED
 
+#include "ch4_cuda_helper.h"
+#include "ch4_cuda_kernel_ops.h"
+
 /* The MPI Standard (MPI-2.1, sec 5.9.2) defines which predfined reduction
    operators are valid by groups of types:
      C integer
@@ -45,13 +48,18 @@ MPIR_OP_TYPE_GROUP(C_INTEGER)
  * emits a warning and generates invalid arithmetic code.  We could drop the
  * restrict instead, but we are more likely to get an optimization from it than
  * const.  [goodell@ 2010-12-15] */
-#define MPIR_OP_TYPE_REDUCE_CASE(mpi_type_,c_type_,op_macro_) \
-    case (mpi_type_): {                                       \
-        c_type_ * restrict a = (c_type_ *)inoutvec;           \
-        /*const*/ c_type_ * restrict b = (c_type_ *)invec;    \
-        for (i=0; i<len; i++)                               \
-            a[i] = op_macro_(a[i],b[i]);                      \
-        break;                                                \
+#define MPIR_OP_TYPE_REDUCE_CASE(mpi_type_,c_type_,type_name_,op_macro_)           \
+    case (mpi_type_): {                                                 \
+        c_type_ * restrict a = (c_type_ *)inoutvec;                     \
+        /*const*/ c_type_ * restrict b = (c_type_ *)invec;              \
+        if(is_mem_type_device(inoutvec) && is_mem_type_device(invec))   \
+        {                                                              \
+            call_##type_name_##_##op_macro_(a, b, len);            \
+        } else {                                                        \
+            for (i=0; i<len; i++)                                       \
+                a[i] = op_macro_(a[i],b[i]);                            \
+        }                                                               \
+        break;                                                          \
     }
 /* helps enforce consistent naming */
 #define MPIR_OP_TYPE_GROUP(group) MPIR_OP_TYPE_GROUP_##group

diff --git a/src/mpi/coll/op/opband.c b/src/mpi/coll/op/opband.c
@@ -22,9 +22,9 @@ void MPIR_BAND(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPIR_LBAND)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPIR_LBAND)
             /* no semicolons by necessity */
-            MPIR_OP_TYPE_GROUP(C_INTEGER)
+             MPIR_OP_TYPE_GROUP(C_INTEGER)
                 MPIR_OP_TYPE_GROUP(FORTRAN_INTEGER)
                 MPIR_OP_TYPE_GROUP(BYTE)
                 /* extra types that are not required to be supported by the MPI Standard */

diff --git a/src/mpi/coll/op/opbor.c b/src/mpi/coll/op/opbor.c
@@ -22,7 +22,7 @@ void MPIR_BOR(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPIR_LBOR)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPIR_LBOR)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
                 MPIR_OP_TYPE_GROUP(FORTRAN_INTEGER)

diff --git a/src/mpi/coll/op/opbxor.c b/src/mpi/coll/op/opbxor.c
@@ -22,7 +22,7 @@ void MPIR_BXOR(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPIR_LBXOR)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPIR_LBXOR)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
                 MPIR_OP_TYPE_GROUP(FORTRAN_INTEGER)

diff --git a/src/mpi/coll/op/opland.c b/src/mpi/coll/op/opland.c
@@ -25,7 +25,7 @@ void MPIR_LAND(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPIR_LLAND)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPIR_LLAND)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
 

diff --git a/src/mpi/coll/op/oplor.c b/src/mpi/coll/op/oplor.c
@@ -25,7 +25,7 @@ void MPIR_LOR(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPIR_LLOR)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPIR_LLOR)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
 

diff --git a/src/mpi/coll/op/oplxor.c b/src/mpi/coll/op/oplxor.c
@@ -25,7 +25,7 @@ void MPIR_LXOR(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPIR_LLXOR)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPIR_LLXOR)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
 

diff --git a/src/mpi/coll/op/opmax.c b/src/mpi/coll/op/opmax.c
@@ -19,7 +19,7 @@ void MPIR_MAXF(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPL_MAX)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPL_MAX)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
                 MPIR_OP_TYPE_GROUP(FORTRAN_INTEGER)

diff --git a/src/mpi/coll/op/opmin.c b/src/mpi/coll/op/opmin.c
@@ -18,7 +18,7 @@ void MPIR_MINF(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPL_MIN)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPL_MIN)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
                 MPIR_OP_TYPE_GROUP(FORTRAN_INTEGER)

diff --git a/src/mpi/coll/op/opprod.c b/src/mpi/coll/op/opprod.c
@@ -20,7 +20,7 @@ void MPIR_PROD(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPIR_LPROD)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPIR_LPROD)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
                 MPIR_OP_TYPE_GROUP(FORTRAN_INTEGER)
@@ -45,9 +45,9 @@ void MPIR_PROD(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
             break;                                      \
         }
 #undef MPIR_OP_C_COMPLEX_TYPE_MACRO
-#define MPIR_OP_C_COMPLEX_TYPE_MACRO(mpi_type_,c_type_,type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_,c_type_,MPIR_LPROD)
-                MPIR_OP_TYPE_GROUP(COMPLEX)
-                MPIR_OP_TYPE_GROUP(COMPLEX_EXTRA)
+#define MPIR_OP_C_COMPLEX_TYPE_MACRO(mpi_type_,c_type_,type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_,c_type_,type_name_,MPIR_LPROD)
+                // MPIR_OP_TYPE_GROUP(COMPLEX)
+                // MPIR_OP_TYPE_GROUP(COMPLEX_EXTRA)
                 /* put things back where we found them */
 #undef MPIR_OP_TYPE_MACRO
 #undef MPIR_OP_C_COMPLEX_TYPE_MACRO

diff --git a/src/mpi/coll/op/opsum.c b/src/mpi/coll/op/opsum.c
@@ -20,7 +20,7 @@ void MPIR_SUM(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
 
     switch (*type) {
 #undef MPIR_OP_TYPE_MACRO
-#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, MPIR_LSUM)
+#define MPIR_OP_TYPE_MACRO(mpi_type_, c_type_, type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_, c_type_, type_name_, MPIR_LSUM)
             /* no semicolons by necessity */
             MPIR_OP_TYPE_GROUP(C_INTEGER)
                 MPIR_OP_TYPE_GROUP(FORTRAN_INTEGER)
@@ -44,9 +44,9 @@ void MPIR_SUM(void *invec, void *inoutvec, int *Len, MPI_Datatype * type)
         }
                 /* C complex types are just simple sums */
 #undef MPIR_OP_C_COMPLEX_TYPE_MACRO
-#define MPIR_OP_C_COMPLEX_TYPE_MACRO(mpi_type_,c_type_,type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_,c_type_,MPIR_LSUM)
-                MPIR_OP_TYPE_GROUP(COMPLEX)
-                MPIR_OP_TYPE_GROUP(COMPLEX_EXTRA)
+#define MPIR_OP_C_COMPLEX_TYPE_MACRO(mpi_type_,c_type_,type_name_) MPIR_OP_TYPE_REDUCE_CASE(mpi_type_,c_type_,type_name_,MPIR_LSUM)
+                //MPIR_OP_TYPE_GROUP(COMPLEX)
+                // MPIR_OP_TYPE_GROUP(COMPLEX_EXTRA)
                 /* put things back where we found them */
 #undef MPIR_OP_TYPE_MACRO
 #undef MPIR_OP_C_COMPLEX_TYPE_MACRO