Merge pull request #602 from jithinjosepkl/pr/pml-cm-opt

Optimizations to PML-CM

Author: jsquyres

ompi/mca/mtl/base/mtl_base_datatype.h

@@ -38,6 +38,17 @@ ompi_mtl_datatype_pack(struct opal_convertor_t *convertor,
     struct iovec iov;
     uint32_t iov_count = 1;
 
+#if !(OPAL_ENABLE_HETEROGENEOUS_SUPPORT)
+    if (convertor->pDesc && 
+	opal_datatype_is_contiguous_memory_layout(convertor->pDesc,
+						  convertor->count)) {
+	    *freeAfter = false;
+	    *buffer = convertor->pBaseBuf;
+	    *buffer_len = convertor->local_size;
+	    return OPAL_SUCCESS;
+    }
+#endif
+
     opal_convertor_get_packed_size(convertor, buffer_len);
     *freeAfter  = false;
     if( 0 == *buffer_len ) {

ompi/mca/mtl/mtl.h

@@ -39,7 +39,7 @@
 #include "ompi_config.h"
 #include "mpi.h" /* needed for MPI_ANY_TAG */
 #include "ompi/mca/mca.h"
-#include "ompi/mca/pml/pml.h" /* for send_mode enum */
+#include "ompi/mca/pml/pml_constants.h" /* for send_mode enum */
 #include "ompi/request/request.h"
 
 BEGIN_C_DECLS
@@ -425,22 +425,24 @@ typedef struct mca_mtl_base_module_t mca_mtl_base_module_t;
 #define MCA_MTL_BASE_VERSION_2_0_0 \
     OMPI_MCA_BASE_VERSION_2_1_0("mtl", 2, 0, 0)
 
+OMPI_DECLSPEC extern mca_mtl_base_module_t *ompi_mtl;
+
 /*
  * macro for doing direct call / call through struct
  */
 #if MCA_ompi_mtl_DIRECT_CALL
 
-#include MCA_ompi_mtl_DIRECT_CALL_HEADER
 
 #define OMPI_MTL_CALL_STAMP(a, b) ompi_mtl_ ## a ## _ ## b
 #define OMPI_MTL_CALL_EXPANDER(a, b) OMPI_MTL_CALL_STAMP(a,b)
 #define OMPI_MTL_CALL(a) OMPI_MTL_CALL_EXPANDER(MCA_ompi_mtl_DIRECT_CALL_COMPONENT, a)
 
+#include MCA_ompi_mtl_DIRECT_CALL_HEADER
+
 #else
 #define OMPI_MTL_CALL(a) ompi_mtl->mtl_ ## a
 #endif
 
-OMPI_DECLSPEC extern mca_mtl_base_module_t *ompi_mtl;
 
 END_C_DECLS
 #endif

ompi/mca/mtl/mxm/mtl_mxm_send.c

@@ -57,6 +57,17 @@ static inline __opal_attribute_always_inline__ int
 
     size_t *buffer_len = &mxm_send_req->base.data.buffer.length;
 
+#if !(OPAL_ENABLE_HETEROGENEOUS_SUPPORT)
+    if (convertor->pDesc && 
+	opal_datatype_is_contiguous_memory_layout(convertor->pDesc,
+						  convertor->count)) {
+	    mxm_send_req->base.data.buffer.ptr = convertor->pBaseBuf;
+	    mxm_send_req->base.data.buffer.length = convertor->local_size;
+	    mxm_send_req->base.data_type = MXM_REQ_DATA_BUFFER;
+	    return OMPI_SUCCESS;
+    }
+#endif
+
     opal_convertor_get_packed_size(convertor, buffer_len);
     if (0 == *buffer_len) {
         mxm_send_req->base.data.buffer.ptr = NULL;

ompi/mca/pml/cm/Makefile.am

@@ -31,13 +31,10 @@ local_sources = \
 	pml_cm_cancel.c \
 	pml_cm_component.c \
 	pml_cm_component.h \
-	pml_cm_probe.c \
-	pml_cm_recv.c \
 	pml_cm_recvreq.h \
 	pml_cm_recvreq.c \
 	pml_cm_request.h \
 	pml_cm_request.c \
-	pml_cm_send.c \
 	pml_cm_sendreq.h \
 	pml_cm_sendreq.c \
 	pml_cm_start.c