io/ompio and common/ompio: add initial support for cuda buffers in ompio

this commit adds the initial support for cuda buffers in ompio, for blocking
and non-blocking individual read and write operations.

Signed-off-by: Edgar Gabriel <egabriel@central.uh.edu>

ompi/mca/common/ompio/Makefile.am

@@ -9,7 +9,7 @@
 #                         University of Stuttgart.  All rights reserved.
 # Copyright (c) 2004-2005 The Regents of the University of California.
 #                         All rights reserved.
-# Copyright (c) 2008-2016 University of Houston. All rights reserved.
+# Copyright (c) 2008-2018 University of Houston. All rights reserved.
 # Copyright (c) 2016      IBM Corporation.  All rights reserved.
 # Copyright (c) 2017-2018 Research Organization for Information Science
 #                         and Technology (RIST). All rights reserved.
@@ -74,6 +74,11 @@ else
 ompidir = $(includedir)
 endif
 
+if OPAL_cuda_support
+headers += common_ompio_cuda.h
+sources += common_ompio_cuda.c
+endif
+
 # These two rules will sym link the "noinst" libtool library filename
 # to the installable libtool library filename in the case where we are
 # compiling this component statically (case 2), described above).

ompi/mca/common/ompio/common_ompio_cuda.c

@@ -0,0 +1,69 @@
+/*
+ *  Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
+ *                          University Research and Technology
+ *                          Corporation.  All rights reserved.
+ *  Copyright (c) 2004-2016 The University of Tennessee and The University
+ *                          of Tennessee Research Foundation.  All rights
+ *                          reserved.
+ *  Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
+ *                          University of Stuttgart.  All rights reserved.
+ *  Copyright (c) 2004-2005 The Regents of the University of California.
+ *                          All rights reserved.
+ *  Copyright (c) 2008-2018 University of Houston. All rights reserved.
+ *  $COPYRIGHT$
+ *
+ *  Additional copyrights may follow
+ *
+ *  $HEADER$
+ */
+
+#include "ompi_config.h"
+
+#include "opal/datatype/opal_convertor.h"
+#include "opal/datatype/opal_datatype_cuda.h"
+#include "opal/mca/common/cuda/common_cuda.h"
+
+#include "ompi/mca/io/ompio/io_ompio.h"
+#include "common_ompio.h"
+#include "common_ompio_cuda.h"
+
+void mca_common_ompio_check_gpu_buf ( ompio_file_t *fh, const void *buf, int *is_gpu, 
+				      int *is_managed)
+{
+    opal_convertor_t    convertor;  
+    
+    *is_gpu=0;
+    *is_managed=0;
+    
+    convertor.flags=0;
+    if ( opal_cuda_check_one_buf ( (char *)buf, &convertor ) ) {
+        *is_gpu = 1;
+        if ( convertor.flags & CONVERTOR_CUDA_UNIFIED ){
+            *is_managed =1;
+        }
+    } 
+    
+    return;
+}
+
+
+void mca_common_ompio_register_buf ( ompio_file_t *fh, const void *buf, 
+                                    size_t bufsize )
+{
+    mca_common_cuda_register ( ( char *)buf, bufsize,  (char *) fh->f_filename );
+    return;
+}
+
+void mca_common_ompio_unregister_buf ( ompio_file_t *fh, void *buf )
+{
+    if ( NULL != fh ) {
+        mca_common_cuda_unregister ( (char *)buf, (char *)fh->f_filename);
+    }
+    else {
+        char dummy_filename[]="dummy_ompio_filename";
+        mca_common_cuda_unregister ( (char *)buf, (char *)dummy_filename);
+    }
+    free (buf);
+    return;
+}
+

ompi/mca/common/ompio/common_ompio_cuda.h

@@ -0,0 +1,51 @@
+/* -*- Mode: C; c-basic-offset:4 ; -*- */
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
+ *                         University Research and Technology
+ *                         Corporation.  All rights reserved.
+ * Copyright (c) 2004-2007 The University of Tennessee and The University
+ *                         of Tennessee Research Foundation.  All rights
+ *                         reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
+ *                         University of Stuttgart.  All rights reserved.
+ * Copyright (c) 2004-2005 The Regents of the University of California.
+ *                         All rights reserved.
+ * Copyright (c) 2008-2018 University of Houston. All rights reserved.
+ * $COPYRIGHT$
+ *
+ * Additional copyrights may follow
+ *
+ * $HEADER$
+ */
+
+#ifndef MCA_COMMON_OMPIO_CUDA_H
+#define MCA_COMMON_OMPIO_CUDA_H
+
+
+#define OMPIO_CUDA_PREPARE_BUF(_fh,_buf,_count,_datatype,_tbuf,_convertor,_max_data,_decoded_iov,_iov_count){ \
+    opal_convertor_clone ( _fh->f_convertor, _convertor, 0);                          \
+    opal_convertor_prepare_for_send ( _convertor, &(_datatype->super), _count, _buf );\
+    opal_convertor_get_packed_size( _convertor, &_max_data );           \
+    _tbuf = (char *) malloc ( _max_data );                              \
+    if ( NULL == _tbuf ) {                                              \
+        opal_output(1, "common_ompio: could not allocate memory.\n");   \
+        return OMPI_ERR_OUT_OF_RESOURCE;                                \
+    }                                                                   \
+    mca_common_ompio_register_buf (_fh, _tbuf, _max_data);              \
+    _decoded_iov = (struct iovec *) malloc ( sizeof ( struct iovec ));  \
+    if ( NULL == _decoded_iov ) {                                       \
+        opal_output(1, "common_ompio: could not allocate memory.\n");   \
+        return OMPI_ERR_OUT_OF_RESOURCE;                                \
+    }                                                                   \
+    _decoded_iov->iov_base = _tbuf;                                     \
+    _decoded_iov->iov_len  = _max_data;                                 \
+    _iov_count=1;}
+
+
+void mca_common_ompio_check_gpu_buf ( ompio_file_t *fh, const void *buf, 
+				      int *is_gpu, int *is_managed);
+void mca_common_ompio_register_buf ( ompio_file_t *fh,  const void *buf, 
+				    size_t bufsize);
+void mca_common_ompio_unregister_buf ( ompio_file_t *fh,  void *buf );
+
+#endif

ompi/mca/common/ompio/common_ompio_file_read.c

@@ -36,6 +36,10 @@
 #include "math.h"
 #include <unistd.h>
 
+#if OPAL_CUDA_SUPPORT
+#include "common_ompio_cuda.h"
+#endif
+
 /* Read and write routines are split into two interfaces.
 **   The
 **   mca_io_ompio_file_read/write[_at]
@@ -86,6 +90,17 @@ int mca_common_ompio_file_read (ompio_file_t *fh,
       return ret;
     }
 
+#if OPAL_CUDA_SUPPORT
+    int is_gpu=0, is_managed=0;
+    opal_convertor_t convertor;
+    mca_common_ompio_check_gpu_buf ( fh, buf, &is_gpu, &is_managed);
+    if ( is_gpu && !is_managed ) {
+        char *tbuf=NULL;
+
+        OMPIO_CUDA_PREPARE_BUF(fh,buf,count,datatype,tbuf,&convertor,max_data,decoded_iov,iov_count);        
+        
+    }
+    else {
         mca_common_ompio_decode_datatype (fh,
                                           datatype,
                                           count,
@@ -93,8 +108,18 @@ int mca_common_ompio_file_read (ompio_file_t *fh,
                                           &max_data,
                                           &decoded_iov,
                                           &iov_count);
+    }
+#else
+    mca_common_ompio_decode_datatype (fh,
+                                      datatype,
+                                      count,
+                                      buf,
+                                      &max_data,
+                                      &decoded_iov,
+                                      &iov_count);
+#endif
 
-    if ( -1 == OMPIO_MCA_GET(fh, cycle_buffer_size) ) {
+    if ( -1 == OMPIO_MCA_GET(fh, cycle_buffer_size )) {
         bytes_per_cycle = max_data;
     }
     else {
@@ -136,6 +161,15 @@ int mca_common_ompio_file_read (ompio_file_t *fh,
         }
     }
 
+#if OPAL_CUDA_SUPPORT
+    if ( is_gpu && !is_managed ) {
+        size_t pos=0;
+
+        opal_convertor_unpack (&convertor, decoded_iov, &iov_count, &pos );
+        opal_convertor_cleanup (&convertor);
+        mca_common_ompio_unregister_buf (fh, decoded_iov->iov_base);
+    }
+#endif
     if (NULL != decoded_iov) {
         free (decoded_iov);
         decoded_iov = NULL;
@@ -208,6 +242,18 @@ int mca_common_ompio_file_iread (ompio_file_t *fh,
         int i = 0; /* index into the decoded iovec of the buffer */
         int j = 0; /* index into the file vie iovec */
         
+#if OPAL_CUDA_SUPPORT
+        int is_gpu=0, is_managed=0;
+        mca_common_ompio_check_gpu_buf ( fh, buf, &is_gpu, &is_managed);
+        if ( is_gpu && !is_managed ) {
+            char *tbuf=NULL;
+            
+            OMPIO_CUDA_PREPARE_BUF(fh,buf,count,datatype,tbuf,&ompio_req->req_convertor,max_data,decoded_iov,iov_count);        
+            
+            ompio_req->req_tbuf = tbuf;
+            ompio_req->req_size = max_data;
+        }
+        else {
             mca_common_ompio_decode_datatype (fh,
                                               datatype,
                                               count,
@@ -215,7 +261,16 @@ int mca_common_ompio_file_iread (ompio_file_t *fh,
                                               &max_data,
                                               &decoded_iov,
                                               &iov_count);
-
+        }
+#else
+        mca_common_ompio_decode_datatype (fh,
+                                          datatype,
+                                          count,
+                                          buf,
+                                          &max_data,
+                                          &decoded_iov,
+                                          &iov_count);
+#endif
         // Non-blocking operations have to occur in a single cycle
         j = fh->f_index_in_file_view;
         

ompi/mca/common/ompio/common_ompio_file_write.c

@@ -34,6 +34,10 @@
 #include "math.h"
 #include <unistd.h>
 
+#if OPAL_CUDA_SUPPORT
+#include "common_ompio_cuda.h"
+#endif
+
 int mca_common_ompio_file_write (ompio_file_t *fh,
 			       const void *buf,
 			       int count,
@@ -61,6 +65,20 @@ int mca_common_ompio_file_write (ompio_file_t *fh,
         return ret;
     }
 
+#if OPAL_CUDA_SUPPORT
+    int is_gpu=0, is_managed=0;
+    mca_common_ompio_check_gpu_buf ( fh, buf, &is_gpu, &is_managed);
+    if ( is_gpu && !is_managed ) {
+        size_t pos=0;
+        char *tbuf=NULL;
+        opal_convertor_t convertor;
+        
+        OMPIO_CUDA_PREPARE_BUF(fh,buf,count,datatype,tbuf,&convertor,max_data,decoded_iov,iov_count);        
+        
+        opal_convertor_pack (&convertor, decoded_iov, &iov_count, &pos );
+        opal_convertor_cleanup ( &convertor);
+    }
+    else {
         mca_common_ompio_decode_datatype (fh,
                                           datatype,
                                           count,
@@ -68,8 +86,17 @@ int mca_common_ompio_file_write (ompio_file_t *fh,
                                           &max_data,
                                           &decoded_iov,
                                           &iov_count);
-
+    }
-    if ( -1 == OMPIO_MCA_GET(fh, cycle_buffer_size) ) {
+#else
+    mca_common_ompio_decode_datatype (fh,
+                                      datatype,
+                                      count,
+                                      buf,
+                                      &max_data,
+                                      &decoded_iov,
+                                      &iov_count);
+#endif
+    if ( -1 == OMPIO_MCA_GET(fh, cycle_buffer_size )) {
         bytes_per_cycle = max_data;
     }
     else {
@@ -108,6 +135,11 @@ int mca_common_ompio_file_write (ompio_file_t *fh,
             fh->f_io_array = NULL;
         }
     }
+#if OPAL_CUDA_SUPPORT
+    if ( is_gpu && !is_managed ) {
+        mca_common_ompio_unregister_buf (fh, decoded_iov->iov_base);
+    }
+#endif
 
     if (NULL != decoded_iov) {
         free (decoded_iov);
@@ -176,6 +208,23 @@ int mca_common_ompio_file_iwrite (ompio_file_t *fh,
         int i = 0; /* index into the decoded iovec of the buffer */
         int j = 0; /* index into the file vie iovec */
 
+#if OPAL_CUDA_SUPPORT
+        int is_gpu=0, is_managed=0;
+        mca_common_ompio_check_gpu_buf ( fh, buf, &is_gpu, &is_managed);
+        if ( is_gpu && !is_managed ) {
+            size_t pos=0;
+            char *tbuf=NULL;
+            opal_convertor_t convertor;
+
+            OMPIO_CUDA_PREPARE_BUF(fh,buf,count,datatype,tbuf,&convertor,max_data,decoded_iov,iov_count);        
+            
+            opal_convertor_pack (&convertor, decoded_iov, &iov_count, &pos );
+            opal_convertor_cleanup (&convertor);
+
+            ompio_req->req_tbuf = tbuf;
+            ompio_req->req_size = max_data;
+        }
+        else {
             mca_common_ompio_decode_datatype (fh,
                                               datatype,
                                               count,
@@ -183,6 +232,16 @@ int mca_common_ompio_file_iwrite (ompio_file_t *fh,
                                               &max_data,
                                               &decoded_iov,
                                               &iov_count);
+        }
+#else
+        mca_common_ompio_decode_datatype (fh,
+                                          datatype,
+                                          count,
+                                          buf,
+                                          &max_data,
+                                          &decoded_iov,
+                                          &iov_count);
+#endif
         j = fh->f_index_in_file_view;
 
         /* Non blocking operations have to occur in a single cycle */

ompi/mca/common/ompio/common_ompio_request.c

@@ -19,6 +19,9 @@
  */
 
 #include "common_ompio_request.h"
+#if OPAL_CUDA_SUPPORT
+#include "common_ompio_cuda.h"
+#endif
 
 static void mca_common_ompio_request_construct(mca_ompio_request_t* req);
 static void mca_common_ompio_request_destruct(mca_ompio_request_t *req);
@@ -34,6 +37,20 @@ opal_list_t mca_common_ompio_pending_requests = {{0}};
 static int mca_common_ompio_request_free ( struct ompi_request_t **req)
 {
     mca_ompio_request_t *ompio_req = ( mca_ompio_request_t *)*req;
+#if OPAL_CUDA_SUPPORT
+    if ( NULL != ompio_req->req_tbuf ) {
+        if ( MCA_OMPIO_REQUEST_READ == ompio_req->req_type ){
+            struct iovec decoded_iov;
+            uint32_t iov_count=1;
+            size_t pos=0;
+
+            decoded_iov.iov_base = ompio_req->req_tbuf;
+            decoded_iov.iov_len  = ompio_req->req_size;
+            opal_convertor_unpack (&ompio_req->req_convertor, &decoded_iov, &iov_count, &pos );
+        }
+        mca_common_ompio_unregister_buf ( NULL, ompio_req->req_tbuf );
+    }
+#endif
     if ( NULL != ompio_req->req_free_fn ) {
         ompio_req->req_free_fn (ompio_req );
     }
@@ -60,6 +77,10 @@ void mca_common_ompio_request_construct(mca_ompio_request_t* req)
     req->req_ompi.req_cancel = mca_common_ompio_request_cancel;
     req->req_ompi.req_type   = OMPI_REQUEST_IO;
     req->req_data            = NULL;
+#if OPAL_CUDA_SUPPORT
+    req->req_tbuf            = NULL;
+    req->req_size            = 0;
+#endif
     req->req_progress_fn     = NULL;
     req->req_free_fn         = NULL;
 

ompi/mca/common/ompio/common_ompio_request.h

@@ -52,6 +52,11 @@ struct mca_ompio_request_t {
     mca_ompio_request_type_t                       req_type;
     void                                          *req_data;
     opal_list_item_t                               req_item;
+#if OPAL_CUDA_SUPPORT
+    void                                          *req_tbuf;
+    size_t                                         req_size;
+    opal_convertor_t                          req_convertor;
+#endif
     mca_fbtl_base_module_progress_fn_t      req_progress_fn;
     mca_fbtl_base_module_request_free_fn_t      req_free_fn;
 };