📄 datatype_unpack.c
字号:
pData, pConv->count ); DO_DEBUG( opal_output( 0, "4. unpack dest %p src %p length %lu\n", user_memory, packed_buffer, (unsigned long)remaining ); ); MEMCPY_CSUM( user_memory, packed_buffer, remaining, pConv ); user_memory += remaining; } } pConv->bConverted += bConverted; } *out_size = iov_count; *max_data = (pConv->bConverted - initial_bytes_converted); if( pConv->bConverted == pConv->local_size ) { pConv->flags |= CONVERTOR_COMPLETED; return 1; } return 0;}/** * This function handle partial types. Depending on the send operation it might happens * that we receive only a partial type (always predefined type). In fact the outcome is * that the unpack has to be done in 2 steps. As there is no way to know if the other * part of the datatype is already received, we need to use a trick to handle this special * case. The trick is to fill the missing part with some well known value, unpack the data * as if it was completely received, and then move into the user memory only the bytes * that don't match th wekk known value. This approach work as long as there is no need * for more than structural changes. They will not work for cases where we will have to * change the content of the data (as in all conversions that require changing the size * of the exponent or mantissa). */static inline uint32_tompi_unpack_partial_datatype( ompi_convertor_t* pConvertor, dt_elem_desc_t* pElem, char* partial_data, ptrdiff_t start_position, ptrdiff_t end_position, char** user_buffer ){ char unused_byte = 0x7F, saved_data[16]; char temporary[16], *temporary_buffer = temporary; char* real_data = *user_buffer + pElem->elem.disp; uint32_t i, length, count_desc = 1; size_t data_length = ompi_ddt_basicDatatypes[pElem->elem.common.type]->size; DO_DEBUG( opal_output( 0, "unpack partial data start %d end %d data_length %lu user %p\n" "\tbConverted %lu total_length %lu count %d\n", start_position, end_position, (unsigned long)data_length, *user_buffer, (unsigned long)pConvertor->bConverted, (unsigned long)pConvertor->local_size, pConvertor->count ); ); /* Find a byte that is not used in the partial buffer */ find_unused_byte: length = (uint32_t)(end_position - start_position); for( i = 0; i < length; i++ ) { if( unused_byte == partial_data[i] ) { unused_byte--; goto find_unused_byte; } } /* Copy and fill the rest of the buffer with the unused byte */ memset( temporary, unused_byte, data_length ); MEMCPY( temporary + start_position, partial_data, (end_position - start_position) ); /* Save the content of the user memory */ MEMCPY( saved_data, real_data, data_length ); /* Then unpack the data into the user memory */ UNPACK_PREDEFINED_DATATYPE( pConvertor, pElem, count_desc, temporary_buffer, *user_buffer, data_length ); /* reload the length as it is reset by the macro */ data_length = ompi_ddt_basicDatatypes[pElem->elem.common.type]->size; /* For every occurence of the unused byte move data from the saved * buffer back into the user memory. */ for( i = 0; i < data_length; i++ ) { if( unused_byte == real_data[i] ) real_data[i] = saved_data[i]; } return 0;}/* The pack/unpack functions need a cleanup. I have to create a proper interface to access * all basic functionalities, hence using them as basic blocks for all conversion functions. * * But first let's make some global assumptions: * - a datatype (with the flag DT_DATA set) will have the contiguous flags set if and only if * the data is really contiguous (extent equal with size) * - for the DT_LOOP type the DT_CONTIGUOUS flag set means that the content of the loop is * contiguous but with a gap in the begining or at the end. * - the DT_CONTIGUOUS flag for the type DT_END_LOOP is meaningless. */int32_tompi_generic_simple_unpack_function( ompi_convertor_t* pConvertor, struct iovec* iov, uint32_t* out_size, size_t* max_data ){ dt_stack_t* pStack; /* pointer to the position on the stack */ uint32_t pos_desc; /* actual position in the description of the derived datatype */ uint32_t count_desc; /* the number of items already done in the actual pos_desc */ uint16_t type = DT_MAX_PREDEFINED; /* type at current position */ size_t total_unpacked = 0; /* total size unpacked this time */ dt_elem_desc_t* description; dt_elem_desc_t* pElem; const ompi_datatype_t *pData = pConvertor->pDesc; char *user_memory_base, *packed_buffer; size_t iov_len_local; uint32_t iov_count; DO_DEBUG( opal_output( 0, "ompi_convertor_generic_simple_unpack( %p, {%p, %lu}, %u )\n", (void*)pConvertor, iov[0].iov_base, (unsigned long)iov[0].iov_len, *out_size ); ); description = pConvertor->use_desc->desc; /* For the first step we have to add both displacement to the source. After in the * main while loop we will set back the source_base to the correct value. This is * due to the fact that the convertor can stop in the middle of a data with a count */ pStack = pConvertor->pStack + pConvertor->stack_pos; pos_desc = pStack->index; user_memory_base = pConvertor->pBaseBuf + pStack->disp; count_desc = (uint32_t)pStack->count; pStack--; pConvertor->stack_pos--; pElem = &(description[pos_desc]); user_memory_base += pStack->disp; DO_DEBUG( opal_output( 0, "unpack start pos_desc %d count_desc %d disp %ld\n" "stack_pos %d pos_desc %d count_desc %d disp %ld\n", pos_desc, count_desc, (long)(user_memory_base - pConvertor->pBaseBuf), pConvertor->stack_pos, pStack->index, (int)pStack->count, (long)(pStack->disp) ); ); for( iov_count = 0; iov_count < (*out_size); iov_count++ ) { packed_buffer = iov[iov_count].iov_base; iov_len_local = iov[iov_count].iov_len; if( 0 != pConvertor->partial_length ) { size_t element_length = ompi_ddt_basicDatatypes[pElem->elem.common.type]->size; size_t missing_length = element_length - pConvertor->partial_length; assert( pElem->elem.common.flags & DT_FLAG_DATA ); COMPUTE_CSUM( packed_buffer, missing_length, pConvertor ); ompi_unpack_partial_datatype( pConvertor, pElem, packed_buffer, pConvertor->partial_length, element_length, &user_memory_base ); --count_desc; if( 0 == count_desc ) { user_memory_base = pConvertor->pBaseBuf + pStack->disp; pos_desc++; /* advance to the next data */ UPDATE_INTERNAL_COUNTERS( description, pos_desc, pElem, count_desc ); } packed_buffer += missing_length; iov_len_local -= missing_length; pConvertor->partial_length = 0; /* nothing more inside */ } while( 1 ) { while( pElem->elem.common.flags & DT_FLAG_DATA ) { /* now here we have a basic datatype */ UNPACK_PREDEFINED_DATATYPE( pConvertor, pElem, count_desc, packed_buffer, user_memory_base, iov_len_local ); if( 0 == count_desc ) { /* completed */ user_memory_base = pConvertor->pBaseBuf + pStack->disp; pos_desc++; /* advance to the next data */ UPDATE_INTERNAL_COUNTERS( description, pos_desc, pElem, count_desc ); continue; } type = pElem->elem.common.type; assert( type < DT_MAX_PREDEFINED ); if( 0 != iov_len_local ) { char* temp = user_memory_base; /* We have some partial data here. Let's copy it into the convertor * and keep it hot until the next round. */ assert( iov_len_local < ompi_ddt_basicDatatypes[type]->size ); COMPUTE_CSUM( packed_buffer, iov_len_local, pConvertor ); ompi_unpack_partial_datatype( pConvertor, pElem, packed_buffer, 0, iov_len_local, &temp ); pConvertor->partial_length = (uint32_t)iov_len_local; iov_len_local = 0; } goto complete_loop; } if( DT_END_LOOP == pElem->elem.common.type ) { /* end of the current loop */ DO_DEBUG( opal_output( 0, "unpack end_loop count %d stack_pos %d pos_desc %d disp %ld space %lu\n", (int)pStack->count, pConvertor->stack_pos, pos_desc, (long)pStack->disp, (unsigned long)iov_len_local ); ); if( --(pStack->count) == 0 ) { /* end of loop */ if( pConvertor->stack_pos == 0 ) { /* Force the conversion to stop by lowering the number of iovecs. */ *out_size = iov_count; goto complete_loop; /* completed */ } pConvertor->stack_pos--; pStack--; pos_desc++; } else { pos_desc = pStack->index + 1; if( pStack->index == -1 ) { pStack->disp += (pData->ub - pData->lb); } else { assert( DT_LOOP == description[pStack->index].loop.common.type ); pStack->disp += description[pStack->index].loop.extent; } } user_memory_base = pConvertor->pBaseBuf + pStack->disp; UPDATE_INTERNAL_COUNTERS( description, pos_desc, pElem, count_desc ); DO_DEBUG( opal_output( 0, "unpack new_loop count %d stack_pos %d pos_desc %d disp %ld space %lu\n", (int)pStack->count, pConvertor->stack_pos, pos_desc, (long)pStack->disp, (unsigned long)iov_len_local ); ); } if( DT_LOOP == pElem->elem.common.type ) { ptrdiff_t local_disp = (ptrdiff_t)user_memory_base; if( pElem->loop.common.flags & DT_FLAG_CONTIGUOUS ) { UNPACK_CONTIGUOUS_LOOP( pConvertor, pElem, count_desc, packed_buffer, user_memory_base, iov_len_local ); if( 0 == count_desc ) { /* completed */ pos_desc += pElem->loop.items + 1; goto update_loop_description; } /* Save the stack with the correct last_count value. */ } local_disp = (ptrdiff_t)user_memory_base - local_disp; PUSH_STACK( pStack, pConvertor->stack_pos, pos_desc, DT_LOOP, count_desc, pStack->disp + local_disp); pos_desc++; update_loop_description: /* update the current state */ user_memory_base = pConvertor->pBaseBuf + pStack->disp; UPDATE_INTERNAL_COUNTERS( description, pos_desc, pElem, count_desc ); DDT_DUMP_STACK( pConvertor->pStack, pConvertor->stack_pos, pElem, "advance loop" ); continue; } } complete_loop: iov[iov_count].iov_len -= iov_len_local; /* update the amount of valid data */ total_unpacked += iov[iov_count].iov_len; } *max_data = total_unpacked; pConvertor->bConverted += total_unpacked; /* update the already converted bytes */ *out_size = iov_count; if( pConvertor->bConverted == pConvertor->remote_size ) { pConvertor->flags |= CONVERTOR_COMPLETED; return 1; } /* I complete an element, next step I should go to the next one */ PUSH_STACK( pStack, pConvertor->stack_pos, pos_desc, DT_BYTE, count_desc, user_memory_base - pStack->disp - pConvertor->pBaseBuf ); DO_DEBUG( opal_output( 0, "unpack save stack stack_pos %d pos_desc %d count_desc %d disp %ld\n", pConvertor->stack_pos, pStack->index, (int)pStack->count, (long)pStack->disp ); ); return 0;}
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -