check point

This commit was SVN r17826.

ompi/mca/coll/sm2/coll_sm2.h

@@ -44,8 +44,8 @@ BEGIN_C_DECLS
 /*
  * Memory Management
  * - All memory allocation will be done on a per-communictor basis
- * - The two banks of memory will be used
- * - Each bank of memory will have M buffers
+ * - At least two banks of memory will be used
+ * - Each bank of memory will have M buffers (or segments)
  * - These buffers will be used in a cirucular buffer order
  * - Each buffer will be contigous in virtual memory, and will have page-aligned
  *     regions belonging to each process in the communicator
@@ -210,6 +210,34 @@ BEGIN_C_DECLS
     /* forward declartion */
     struct mca_coll_sm2_module_t;
 
+    /*
+     * shared memory region descriptor
+     */
+    struct sm_memory_region_desc_t {
+
+        /* pointer to control structures */
+        volatile mca_coll_sm2_nb_request_process_shared_mem_t *control_region;
+
+        /* pointer to data segment, and lower half of data segment */
+        volatile char *data_segment;
+
+    };
+    typedef struct sm_memory_region_desc_t sm_memory_region_desc_t;
+
+    /*
+     * Shared memory buffer management strcucture
+     */
+    struct sm_work_buffer_t {
+        /* pointer to segment base */
+        volatile char * base_segment_address;
+
+        /* description of how the memory segment is mapped on
+         *   a per process basis
+         */
+        sm_memory_region_desc_t *proc_memory;
+    };
+    typedef struct sm_work_buffer_t sm_work_buffer_t;
+
     /* process private barrier request object */
     struct mca_coll_sm2_nb_request_process_private_mem_t {
         struct ompi_request_t super;
@@ -251,6 +279,12 @@ BEGIN_C_DECLS
         /* Pointer to the collective buffers */
         char *collective_buffer_region;
 
+        /* description of allocated temp buffers - one struct per
+         * buffer.  Each buffer has space "owned" by each process
+         * in the group.
+         */
+        sm_work_buffer_t *sm_buffer_descriptor;
+
         /* size of memory region, per process, for memory bank management */
         size_t sm2_size_management_region_per_proc;
 
@@ -373,7 +407,7 @@ BEGIN_C_DECLS
             struct mca_coll_base_module_1_1_0_t *module);
 
     /* allocate working buffer */
-    char *alloc_sm2_shared_buffer(mca_coll_sm2_module_t *module);
+    sm_work_buffer_t *alloc_sm2_shared_buffer(mca_coll_sm2_module_t *module);
 
     /* free working buffer - it is assumed that buffers are released in
      * the order they are allocated.  We can assume this because each

ompi/mca/coll/sm2/coll_sm2_allreduce.c

@@ -52,6 +52,7 @@ int mca_coll_sm2_allreduce_intra_fanin_fanout(void *sbuf, void *rbuf, int count,
     volatile mca_coll_sm2_nb_request_process_shared_mem_t * parent_ctl_pointer;
     mca_coll_sm2_module_t *sm_module;
     tree_node_t *my_reduction_node, *my_fanout_read_tree;
+    sm_work_buffer_t *sm_buffer_desc;
 
 
     sm_module=(mca_coll_sm2_module_t *) module;
@@ -98,7 +99,9 @@ int mca_coll_sm2_allreduce_intra_fanin_fanout(void *sbuf, void *rbuf, int count,
     /* get a pointer to the shared-memory working buffer */
     /* NOTE: starting with a rather synchronous approach */
     for( stripe_number=0 ; stripe_number < n_data_segments ; stripe_number++ ) {
-        sm_buffer=alloc_sm2_shared_buffer(sm_module);
+    
+        sm_buffer_desc=alloc_sm2_shared_buffer(sm_module);
+        sm_buffer=sm_buffer_desc->base_segment_address;
         if( NULL == sm_buffer) {
             rc=OMPI_ERR_OUT_OF_RESOURCE;
             goto Error;
@@ -350,6 +353,7 @@ int mca_coll_sm2_allreduce_intra_recursive_doubling(void *sbuf, void *rbuf,
     int my_rank,count_processed,count_this_stripe;
     size_t message_extent,dt_extent,ctl_size,len_data_buffer;
     long long tag, base_tag;
+    sm_work_buffer_t *sm_buffer_desc;
     volatile char * sm_buffer;
     volatile char * my_tmp_data_buffer[2];
     volatile char * my_write_pointer;
@@ -412,7 +416,8 @@ int mca_coll_sm2_allreduce_intra_recursive_doubling(void *sbuf, void *rbuf,
             /* debug */
             t0=opal_sys_timer_get_cycles();
             /* end debug */
-        sm_buffer=alloc_sm2_shared_buffer(sm_module);
+        sm_buffer_desc=alloc_sm2_shared_buffer(sm_module);
+        sm_buffer=sm_buffer_desc->base_segment_address;
         if( NULL == sm_buffer) {
             rc=OMPI_ERR_OUT_OF_RESOURCE;
             goto Error;
@@ -558,9 +563,9 @@ int mca_coll_sm2_allreduce_intra_recursive_doubling(void *sbuf, void *rbuf,
            
                {
                    int ii,n_ints;
-                   int *my_read=(int *)my_read_pointer;
-                   int *my_write=(int *)my_write_pointer;
-                   int *exchange_read=(int *)partner_read_pointer;
+                   int * restrict my_read=(int *)my_read_pointer;
+                   int * restrict my_write=(int *)my_write_pointer;
+                   int * restrict exchange_read=(int *)partner_read_pointer;
                    n_ints=count_this_stripe;
                    for(ii=0 ; ii < n_ints ; ii++ ) {
                        my_write[ii]=my_read[ii]+exchange_read[ii];

ompi/mca/coll/sm2/coll_sm2_module.c

@@ -459,7 +459,7 @@ mca_coll_sm2_comm_query(struct ompi_communicator_t *comm, int *priority)
 {
     /* local variables */
     mca_coll_sm2_module_t *sm_module;
-    int group_size,ret;
+    int i,j,group_size,ret;
     size_t alignment,size;
     ssize_t size_tot_per_proc_per_seg;
     size_t tot_size_per_bank,size_tot_per_segment;
@@ -542,7 +542,25 @@ mca_coll_sm2_comm_query(struct ompi_communicator_t *comm, int *priority)
         mca_coll_sm2_component.sm2_num_mem_banks;
     sm_module->sm2_module_num_regions_per_bank=
         mca_coll_sm2_component.sm2_num_regions_per_bank;
+    sm_module->sm2_module_num_buffers=
+        mca_coll_sm2_component.sm2_num_regions_per_bank *
+        mca_coll_sm2_component.sm2_num_mem_banks;
 
+
+    /* allocate the array of memory descriptors used to describe the
+     *   shared memory buffers.  This structure resides in process
+     *   private memory, but describes the shared memory.
+     */
+    sm_module->sm_buffer_descriptor=(sm_work_buffer_t *)malloc(
+            sizeof(sm_work_buffer_t)*sm_module->sm2_module_num_buffers);
+    if( NULL == sm_module->sm_buffer_descriptor ) {
+        goto CLEANUP;
+    }
+
+    /*
+     *  Now figure out how much memory to allocate for use as
+     *  working memory for the shared memory collectives.
+     */
     /* 
      * get control region size 
      */ 
@@ -611,11 +629,6 @@ mca_coll_sm2_comm_query(struct ompi_communicator_t *comm, int *priority)
     /* total memory management required */
     mem_management_total=mem_management_per_proc * group_size;
 
-    /* set the total number of working buffers */
-    sm_module->sm2_module_num_buffers=
-        mca_coll_sm2_component.sm2_num_regions_per_bank *
-        mca_coll_sm2_component.sm2_num_mem_banks;
-
     /* total size of backing file - this assumes the mmap allocation
      *   occurs on page boundaries, and that all segments are paged
      *   aligned 
@@ -694,7 +707,36 @@ mca_coll_sm2_comm_query(struct ompi_communicator_t *comm, int *priority)
         sm_module->sm2_first_buffer_index_next_bank=0;
     }
 
-    /* set pointers */
+    /* setup shared memory memory descriptors */
+    for( i=0 ; i < sm_module->sm2_module_num_buffers ; i++ ) {
+
+        char *base_buffer;
+        volatile mca_coll_sm2_nb_request_process_shared_mem_t *ctl_ptr;
+
+        /* set the base address for this working buffer */
+        base_buffer= sm_module->collective_buffer_region+
+            i*sm_module->segment_size;
+        sm_module->sm_buffer_descriptor[i].base_segment_address=base_buffer;
+
+        /* allocate array to keep data on each segment in the buffer.
+         *   One segment per process in the group.
+         */
+        sm_module->sm_buffer_descriptor[i].proc_memory=
+            (sm_memory_region_desc_t *)malloc(sizeof(sm_memory_region_desc_t)*
+                                            group_size);
+        if( NULL == sm_module->sm_buffer_descriptor[i].proc_memory ) {
+            goto CLEANUP;
+        }
+        for(j=0 ; j < group_size ; j++ ) {
+            ctl_ptr=(volatile mca_coll_sm2_nb_request_process_shared_mem_t *)
+                base_buffer+j* sm_module->segement_size_per_process;
+            sm_module->sm_buffer_descriptor[i].proc_memory[j].control_region=
+                ctl_ptr;
+            sm_module->sm_buffer_descriptor[i].proc_memory[j].data_segment=
+                (char *)ctl_ptr+sm_module->ctl_memory_per_proc_per_segment;
+        }
+
+    }
 
     /* touch pages to apply memory affinity - Note: do we really need this or will
      * the algorithms do this */
@@ -716,6 +758,17 @@ CLEANUP:
         sm_module->coll_sm2_file_name=NULL;
     }
 
+    if( NULL != sm_module->sm_buffer_descriptor ) {
+        for(i=0 ; i < group_size ; i++ ) {
+            if(NULL != sm_module->sm_buffer_descriptor[i].proc_memory) {
+                free(sm_module->sm_buffer_descriptor[i].proc_memory);
+                sm_module->sm_buffer_descriptor[i].proc_memory=NULL;
+            }
+        }
+        free(sm_module->sm_buffer_descriptor);
+        sm_module->sm_buffer_descriptor=NULL;
+    }
+
     OBJ_RELEASE(sm_module);
 
     return NULL;
@@ -742,7 +795,7 @@ sm2_module_enable(struct mca_coll_base_module_1_1_0_t *module,
 }
 
 /* allocate working buffer */
-char *alloc_sm2_shared_buffer(mca_coll_sm2_module_t *module)
+sm_work_buffer_t *alloc_sm2_shared_buffer(mca_coll_sm2_module_t *module)
 {
     /* local variables */
     int rc,buffer_index, memory_bank_index;
@@ -836,11 +889,7 @@ char *alloc_sm2_shared_buffer(mca_coll_sm2_module_t *module)
 
     buffer_index=module->sm2_allocated_buffer_index;
 
-    /* get base address of return buffer */
-    return_buffer=module->collective_buffer_region+
-        buffer_index*module->segment_size;
-
-    return return_buffer;
+    return &(module->sm_buffer_descriptor[buffer_index]);
 
 }