1
1

Cleanup the code and make it a little faster.

This commit was SVN r18038.
Этот коммит содержится в:
George Bosilca 2008-03-31 17:12:03 +00:00
родитель 493677426d
Коммит 5adaa88241

Просмотреть файл

@ -150,7 +150,7 @@ int32_t ompi_ddt_copy_content_same_ddt( const ompi_datatype_t* datatype, int32_t
/* If we have to copy a contiguous datatype then simply /* If we have to copy a contiguous datatype then simply
* do a memcpy. * do a memcpy.
*/ */
if( (datatype->flags & DT_FLAG_CONTIGUOUS) == DT_FLAG_CONTIGUOUS ) { if( datatype->flags & DT_FLAG_CONTIGUOUS ) {
ptrdiff_t extent = (datatype->ub - datatype->lb); ptrdiff_t extent = (datatype->ub - datatype->lb);
/* Now that we know the datatype is contiguous, we should move the 2 pointers /* Now that we know the datatype is contiguous, we should move the 2 pointers
* source and destination to the correct displacement. * source and destination to the correct displacement.
@ -158,7 +158,7 @@ int32_t ompi_ddt_copy_content_same_ddt( const ompi_datatype_t* datatype, int32_t
destination += datatype->lb; destination += datatype->lb;
source += datatype->lb; source += datatype->lb;
if( (ptrdiff_t)datatype->size == extent ) { /* all contiguous == no gaps around */ if( (ptrdiff_t)datatype->size == extent ) { /* all contiguous == no gaps around */
size_t total_length = datatype->size * count; size_t total_length = iov_len_local;
size_t memcpy_chunk = ompi_datatype_memcpy_block_size; size_t memcpy_chunk = ompi_datatype_memcpy_block_size;
while( total_length > 0 ) { while( total_length > 0 ) {
if( memcpy_chunk > total_length ) memcpy_chunk = total_length; if( memcpy_chunk > total_length ) memcpy_chunk = total_length;
@ -167,14 +167,14 @@ int32_t ompi_ddt_copy_content_same_ddt( const ompi_datatype_t* datatype, int32_t
OMPI_DDT_SAFEGUARD_POINTER( source, memcpy_chunk, OMPI_DDT_SAFEGUARD_POINTER( source, memcpy_chunk,
(unsigned char*)source_base, datatype, count ); (unsigned char*)source_base, datatype, count );
DO_DEBUG( opal_output( 0, "copy c1. memcpy( %p, %p, %lu ) => space %lu\n", DO_DEBUG( opal_output( 0, "copy c1. memcpy( %p, %p, %lu ) => space %lu\n",
destination, source, (unsigned long)memcpy_chunk, (unsigned long)iov_len_local ); ); destination, source, (unsigned long)memcpy_chunk, (unsigned long)total_length ); );
MEMCPY( destination, source, memcpy_chunk ); MEMCPY( destination, source, memcpy_chunk );
destination += memcpy_chunk; destination += memcpy_chunk;
source += memcpy_chunk; source += memcpy_chunk;
total_length -= memcpy_chunk; total_length -= memcpy_chunk;
iov_len_local -= memcpy_chunk;
} }
} else { return 0; /* completed */
}
for( pos_desc = 0; (int32_t)pos_desc < count; pos_desc++ ) { for( pos_desc = 0; (int32_t)pos_desc < count; pos_desc++ ) {
OMPI_DDT_SAFEGUARD_POINTER( destination, datatype->size, OMPI_DDT_SAFEGUARD_POINTER( destination, datatype->size,
(unsigned char*)destination_base, datatype, count ); (unsigned char*)destination_base, datatype, count );
@ -187,8 +187,7 @@ int32_t ompi_ddt_copy_content_same_ddt( const ompi_datatype_t* datatype, int32_t
destination += extent; destination += extent;
source += extent; source += extent;
} }
} return 0; /* completed */
return 0;
} }
pStack = (dt_stack_t*)alloca( sizeof(dt_stack_t) * (datatype->btypes[DT_LOOP] + 1) ); pStack = (dt_stack_t*)alloca( sizeof(dt_stack_t) * (datatype->btypes[DT_LOOP] + 1) );