void reset( ) {
DVLOG(5) << "In " << __PRETTY_FUNCTION__;
CHECK( !acquire_started_ || acquired_ ) << "inconsistent state for reset";
acquire_started_ = false;
acquired_ = false;
thread_ = NULL;
num_messages_ = 0;
response_count_ = 0;
expected_reply_payload_ = sizeof( T ) * *count_;
total_reply_payload_ = 0;
start_time_ = 0;
network_time_ = 0;
if( *count_ == 0 ) {
DVLOG(5) << "Zero-length acquire";
*pointer_ = NULL;
acquire_started_ = true;
acquired_ = true;
} else if( request_address_->is_2D() ) {
num_messages_ = 1;
if( request_address_->core() == Grappa::mycore() ) {
DVLOG(5) << "Short-circuiting to address " << request_address_->pointer();
*pointer_ = request_address_->pointer();
acquire_started_ = true;
acquired_ = true;
}
} else {
DVLOG(5) << "Straddle: block_max is " << (*request_address_ + *count_).block_max() ;
DVLOG(5) << ", request_address is " << *request_address_;
DVLOG(5) << ", sizeof(T) is " << sizeof(T);
DVLOG(5) << ", count is " << *count_;
DVLOG(5) << ", block_min is " << request_address_->block_min();
DVLOG(5) << "Straddle: address is " << *request_address_ ;
DVLOG(5) << ", address + count is " << *request_address_ + *count_;
ptrdiff_t byte_diff = ( (*request_address_ + *count_ - 1).last_byte().block_max() -
request_address_->first_byte().block_min() );
DVLOG(5) << "Straddle: address block max is " << request_address_->block_max();
DVLOG(5) << " address + count block max is " << (*request_address_ + *count_).block_max();
DVLOG(5) << " address block min " << request_address_->block_min();
DVLOG(5) << "Straddle: diff is " << byte_diff << " bs " << block_size;
num_messages_ = byte_diff / block_size;
}
if( num_messages_ > 1 ) DVLOG(5) << "****************************** MULTI BLOCK CACHE REQUEST ******************************";
DVLOG(5) << "In " << __PRETTY_FUNCTION__ << "; detecting straddle for sizeof(T):" << sizeof(T)
<< " count:" << *count_
<< " num_messages_:" << num_messages_
<< " request_address:" << *request_address_;
}
void do_release() {
size_t total_bytes = *count_ * sizeof(T);
RequestArgs args;
args.request_address = *request_address_;
DVLOG(5) << "Computing request_bytes from block_max " << request_address_->first_byte().block_max() << " and " << *request_address_;
args.reply_address = make_global( this );
size_t offset = 0;
size_t request_bytes = 0;
for( size_t i = 0;
offset < total_bytes;
offset += request_bytes, i++) {
request_bytes = args.request_address.first_byte().block_max() - args.request_address.first_byte();
if( request_bytes > total_bytes - offset ) {
request_bytes = total_bytes - offset;
}
DVLOG(5) << "sending release request with " << request_bytes
<< " of total bytes = " << *count_ * sizeof(T)
<< " to " << args.request_address;
Grappa::send_heap_message(args.request_address.core(),
[args](void * payload, size_t payload_size) {
IRMetrics::count_release_ams( payload_size );
DVLOG(5) << "Worker " << Grappa::current_worker()
<< " received release request to " << args.request_address
<< " reply to " << args.reply_address;
memcpy( args.request_address.pointer(), payload, payload_size );
auto reply_address = args.reply_address;
Grappa::send_heap_message(args.reply_address.core(), [reply_address]{
DVLOG(5) << "Worker " << Grappa::current_worker() << " received release reply to " << reply_address;
reply_address.pointer()->release_reply();
});
DVLOG(5) << "Worker " << Grappa::current_worker()
<< " sent release reply to " << reply_address;
},
(char*)(*pointer_) + offset, request_bytes
);
// TODO: change type so we don't screw with pointer like this
args.request_address = GlobalAddress<T>::Raw( args.request_address.raw_bits() + request_bytes );
}
DVLOG(5) << "release started for " << args.request_address;
// blocks here waiting for messages to be sent
}
void reset( ) {
CHECK( !release_started_ || released_ ) << "inconsistent state for reset";
release_started_ = false;
released_ = false;
thread_ = NULL;
num_messages_ = 0;
response_count_ = 0;
if( *count_ == 0 ) {
DVLOG(5) << "Zero-length release";
release_started_ = true;
released_ = true;
} else if( request_address_->is_2D() ) {
num_messages_ = 1;
if( request_address_->core() == Grappa::mycore() ) {
release_started_ = true;
released_ = true;
}
} else {
DVLOG(5) << "Straddle: block_max is " << (*request_address_ + *count_).block_max() ;
DVLOG(5) << ", request_address is " << *request_address_;
DVLOG(5) << ", sizeof(T) is " << sizeof(T);
DVLOG(5) << ", count is " << *count_;
DVLOG(5) << ", block_min is " << request_address_->block_min();
DVLOG(5) << "Straddle: address is " << *request_address_ ;
DVLOG(5) << ", address + count is " << *request_address_ + *count_;
ptrdiff_t byte_diff = ( (*request_address_ + *count_ - 1).last_byte().block_max() -
request_address_->first_byte().block_min() );
DVLOG(5) << "Straddle: address block max is " << request_address_->block_max();
DVLOG(5) << " address + count block max is " << (*request_address_ + *count_).block_max();
DVLOG(5) << " address + count -1 block max is " << (*request_address_ + *count_ - 1).block_max();
DVLOG(5) << " difference is " << ( (*request_address_ + *count_ - 1).block_max() - request_address_->block_min() );
DVLOG(5) << " multiplied difference is " << ( (*request_address_ + *count_ - 1).block_max() - request_address_->block_min() ) * sizeof(T);
DVLOG(5) << " address block min " << request_address_->block_min();
DVLOG(5) << "Straddle: diff is " << byte_diff << " bs " << block_size;
num_messages_ = byte_diff / block_size;
}
if( num_messages_ > 1 ) DVLOG(5) << "****************************** MULTI BLOCK CACHE REQUEST ******************************";
DVLOG(5) << "Detecting straddle for sizeof(T):" << sizeof(T)
<< " count:" << *count_
<< " num_messages_:" << num_messages_
<< " request_address:" << *request_address_;
}
void do_acquire() {
size_t total_bytes = *count_ * sizeof(T);
RequestArgs args;
args.request_address = *request_address_;
DVLOG(5) << "Computing request_bytes from block_max " << request_address_->first_byte().block_max() << " and " << *request_address_;
args.reply_address = make_global( this );
args.offset = 0;
for(size_t i = 0;
args.offset < total_bytes;
args.offset += args.request_bytes, i++) {
args.request_bytes = args.request_address.first_byte().block_max() - args.request_address.first_byte();
if( args.request_bytes > total_bytes - args.offset ) {
args.request_bytes = total_bytes - args.offset;
}
DVLOG(5) << "sending acquire request for " << args.request_bytes
<< " of total bytes = " << *count_ * sizeof(T)
<< " from " << args.request_address;
Grappa::send_heap_message(args.request_address.core(), [args]{
IAMetrics::count_acquire_ams( args.request_bytes );
DVLOG(5) << "Worker " << Grappa::current_worker()
<< " received acquire request to " << args.request_address
<< " size " << args.request_bytes
<< " offset " << args.offset
<< " reply to " << args.reply_address;
DVLOG(5) << "Worker " << Grappa::current_worker()
<< " sending acquire reply to " << args.reply_address
<< " offset " << args.offset
<< " request address " << args.request_address
<< " payload address " << args.request_address.pointer()
<< " payload size " << args.request_bytes;
// note: this will read the payload *later* when the message is copied into the actual send buffer,
// should be okay because we're already assuming DRF, but something to watch out for
auto reply_address = args.reply_address;
auto offset = args.offset;
Grappa::send_heap_message(args.reply_address.core(),
[reply_address, offset](void * payload, size_t payload_size) {
DVLOG(5) << "Worker " << Grappa::current_worker()
<< " received acquire reply to " << reply_address
<< " offset " << offset
<< " payload size " << payload_size;
reply_address.pointer()->acquire_reply( offset, payload, payload_size);
},
args.request_address.pointer(), args.request_bytes
);
DVLOG(5) << "Worker " << Grappa::current_worker()
<< " sent acquire reply to " << args.reply_address
<< " offset " << args.offset
<< " request address " << args.request_address
<< " payload address " << args.request_address.pointer()
<< " payload size " << args.request_bytes;
});
// TODO: change type so we don't screw with pointer like this
args.request_address = GlobalAddress<T>::Raw( args.request_address.raw_bits() + args.request_bytes );
}
DVLOG(5) << "acquire started for " << args.request_address;
}
