static T get_something(cl_platform_id platform, cl_device_id device,
T Slot::*member, thread_scoped_lock &slot_locker)
{
assert(platform != NULL);
OpenCLCache &self = global_instance();
thread_scoped_lock cache_lock(self.cache_lock);
pair<CacheMap::iterator,bool> ins = self.cache.insert(
CacheMap::value_type(PlatformDevicePair(platform, device), Slot()));
Slot &slot = ins.first->second;
/* create slot lock only while holding cache lock */
if(!slot.mutex)
slot.mutex = new thread_mutex;
/* need to unlock cache before locking slot, to allow store to complete */
cache_lock.unlock();
/* lock the slot */
slot_locker = thread_scoped_lock(*slot.mutex);
/* If the thing isn't cached */
if(slot.*member == NULL) {
/* return with the caller's lock holder holding the slot lock */
return NULL;
}
/* the item was already cached, release the slot lock */
slot_locker.unlock();
return slot.*member;
}
void OpenCLCache::store_context(cl_platform_id platform,
cl_device_id device,
cl_context context,
thread_scoped_lock& slot_locker)
{
assert(platform != NULL);
assert(device != NULL);
assert(context != NULL);
OpenCLCache &self = global_instance();
thread_scoped_lock cache_lock(self.cache_lock);
CacheMap::iterator i = self.cache.find(PlatformDevicePair(platform, device));
cache_lock.unlock();
Slot &slot = i->second;
/* sanity check */
assert(i != self.cache.end());
assert(slot.context == NULL);
slot.context = context;
/* unlock the slot */
slot_locker.unlock();
/* increment reference count in OpenCL.
* The caller is going to release the object when done with it. */
cl_int ciErr = clRetainContext(context);
assert(ciErr == CL_SUCCESS);
(void)ciErr;
}
cl_program OpenCLCache::get_program(cl_platform_id platform,
cl_device_id device,
ustring key,
thread_scoped_lock& slot_locker)
{
assert(platform != NULL);
OpenCLCache& self = global_instance();
thread_scoped_lock cache_lock(self.cache_lock);
pair<CacheMap::iterator,bool> ins = self.cache.insert(
CacheMap::value_type(PlatformDevicePair(platform, device), Slot()));
Slot &slot = ins.first->second;
pair<Slot::EntryMap::iterator,bool> ins2 = slot.programs.insert(
Slot::EntryMap::value_type(key, Slot::ProgramEntry()));
Slot::ProgramEntry &entry = ins2.first->second;
/* create slot lock only while holding cache lock */
if(!entry.mutex)
entry.mutex = new thread_mutex;
/* need to unlock cache before locking slot, to allow store to complete */
cache_lock.unlock();
/* lock the slot */
slot_locker = thread_scoped_lock(*entry.mutex);
/* If the thing isn't cached */
if(entry.program == NULL) {
/* return with the caller's lock holder holding the slot lock */
return NULL;
}
/* the item was already cached, release the slot lock */
slot_locker.unlock();
cl_int ciErr = clRetainProgram(entry.program);
assert(ciErr == CL_SUCCESS);
(void)ciErr;
return entry.program;
}
void OpenCLCache::store_program(cl_platform_id platform,
cl_device_id device,
cl_program program,
ustring key,
thread_scoped_lock& slot_locker)
{
assert(platform != NULL);
assert(device != NULL);
assert(program != NULL);
OpenCLCache &self = global_instance();
thread_scoped_lock cache_lock(self.cache_lock);
CacheMap::iterator i = self.cache.find(PlatformDevicePair(platform, device));
assert(i != self.cache.end());
Slot &slot = i->second;
Slot::EntryMap::iterator i2 = slot.programs.find(key);
assert(i2 != slot.programs.end());
Slot::ProgramEntry &entry = i2->second;
assert(entry.program == NULL);
cache_lock.unlock();
entry.program = program;
/* unlock the slot */
slot_locker.unlock();
/* Increment reference count in OpenCL.
* The caller is going to release the object when done with it.
*/
cl_int ciErr = clRetainProgram(program);
assert(ciErr == CL_SUCCESS);
(void)ciErr;
}
static void store_something(cl_platform_id platform, cl_device_id device, T thing,
T Slot::*member, thread_scoped_lock &slot_locker)
{
assert(platform != NULL);
assert(device != NULL);
assert(thing != NULL);
OpenCLCache &self = global_instance();
thread_scoped_lock cache_lock(self.cache_lock);
CacheMap::iterator i = self.cache.find(PlatformDevicePair(platform, device));
cache_lock.unlock();
Slot &slot = i->second;
/* sanity check */
assert(i != self.cache.end());
assert(slot.*member == NULL);
slot.*member = thing;
/* unlock the slot */
slot_locker.unlock();
}
/* note that the lock must be already acquired upon entry.
* This is necessary because the caller often peeks at
* the header and delegates control to here when it doesn't
* specifically handle the current RPC.
* The lock must be unlocked before returning */
void process(RPCReceive& rcv, thread_scoped_lock &lock)
{
if(rcv.name == "mem_alloc") {
MemoryType type;
network_device_memory mem;
device_ptr client_pointer;
rcv.read(mem);
rcv.read(type);
lock.unlock();
client_pointer = mem.device_pointer;
/* create a memory buffer for the device buffer */
size_t data_size = mem.memory_size();
DataVector &data_v = data_vector_insert(client_pointer, data_size);
if(data_size)
mem.data_pointer = (device_ptr)&(data_v[0]);
else
mem.data_pointer = 0;
/* perform the allocation on the actual device */
device->mem_alloc(mem, type);
/* store a mapping to/from client_pointer and real device pointer */
pointer_mapping_insert(client_pointer, mem.device_pointer);
}
else if(rcv.name == "mem_copy_to") {
network_device_memory mem;
rcv.read(mem);
lock.unlock();
device_ptr client_pointer = mem.device_pointer;
DataVector &data_v = data_vector_find(client_pointer);
size_t data_size = mem.memory_size();
/* get pointer to memory buffer for device buffer */
mem.data_pointer = (device_ptr)&data_v[0];
/* copy data from network into memory buffer */
rcv.read_buffer((uint8_t*)mem.data_pointer, data_size);
/* translate the client pointer to a real device pointer */
mem.device_pointer = device_ptr_from_client_pointer(client_pointer);
/* copy the data from the memory buffer to the device buffer */
device->mem_copy_to(mem);
}
else if(rcv.name == "mem_copy_from") {
network_device_memory mem;
int y, w, h, elem;
rcv.read(mem);
rcv.read(y);
rcv.read(w);
rcv.read(h);
rcv.read(elem);
device_ptr client_pointer = mem.device_pointer;
mem.device_pointer = device_ptr_from_client_pointer(client_pointer);
DataVector &data_v = data_vector_find(client_pointer);
mem.data_pointer = (device_ptr)&(data_v[0]);
device->mem_copy_from(mem, y, w, h, elem);
size_t data_size = mem.memory_size();
RPCSend snd(socket, &error_func, "mem_copy_from");
snd.write();
snd.write_buffer((uint8_t*)mem.data_pointer, data_size);
lock.unlock();
}
else if(rcv.name == "mem_zero") {
network_device_memory mem;
rcv.read(mem);
lock.unlock();
device_ptr client_pointer = mem.device_pointer;
mem.device_pointer = device_ptr_from_client_pointer(client_pointer);
DataVector &data_v = data_vector_find(client_pointer);
mem.data_pointer = (device_ptr)&(data_v[0]);
device->mem_zero(mem);
}
else if(rcv.name == "mem_free") {
network_device_memory mem;
device_ptr client_pointer;
rcv.read(mem);
lock.unlock();
client_pointer = mem.device_pointer;
mem.device_pointer = device_ptr_from_client_pointer_erase(client_pointer);
device->mem_free(mem);
}
else if(rcv.name == "const_copy_to") {
string name_string;
size_t size;
rcv.read(name_string);
rcv.read(size);
vector<char> host_vector(size);
rcv.read_buffer(&host_vector[0], size);
lock.unlock();
device->const_copy_to(name_string.c_str(), &host_vector[0], size);
}
else if(rcv.name == "tex_alloc") {
network_device_memory mem;
string name;
InterpolationType interpolation;
bool periodic;
device_ptr client_pointer;
rcv.read(name);
rcv.read(mem);
rcv.read(interpolation);
rcv.read(periodic);
lock.unlock();
client_pointer = mem.device_pointer;
size_t data_size = mem.memory_size();
DataVector &data_v = data_vector_insert(client_pointer, data_size);
if(data_size)
mem.data_pointer = (device_ptr)&(data_v[0]);
else
mem.data_pointer = 0;
rcv.read_buffer((uint8_t*)mem.data_pointer, data_size);
device->tex_alloc(name.c_str(), mem, interpolation, periodic);
pointer_mapping_insert(client_pointer, mem.device_pointer);
}
else if(rcv.name == "tex_free") {
network_device_memory mem;
device_ptr client_pointer;
rcv.read(mem);
lock.unlock();
client_pointer = mem.device_pointer;
mem.device_pointer = device_ptr_from_client_pointer_erase(client_pointer);
device->tex_free(mem);
}
else if(rcv.name == "load_kernels") {
bool experimental;
rcv.read(experimental);
bool result;
result = device->load_kernels(experimental);
RPCSend snd(socket, &error_func, "load_kernels");
snd.add(result);
snd.write();
lock.unlock();
}
else if(rcv.name == "task_add") {
DeviceTask task;
rcv.read(task);
lock.unlock();
if(task.buffer)
task.buffer = device_ptr_from_client_pointer(task.buffer);
if(task.rgba_half)
task.rgba_half = device_ptr_from_client_pointer(task.rgba_half);
if(task.rgba_byte)
task.rgba_byte = device_ptr_from_client_pointer(task.rgba_byte);
if(task.shader_input)
task.shader_input = device_ptr_from_client_pointer(task.shader_input);
if(task.shader_output)
task.shader_output = device_ptr_from_client_pointer(task.shader_output);
task.acquire_tile = function_bind(&DeviceServer::task_acquire_tile, this, _1, _2);
task.release_tile = function_bind(&DeviceServer::task_release_tile, this, _1);
task.update_progress_sample = function_bind(&DeviceServer::task_update_progress_sample, this);
task.update_tile_sample = function_bind(&DeviceServer::task_update_tile_sample, this, _1);
task.get_cancel = function_bind(&DeviceServer::task_get_cancel, this);
device->task_add(task);
}
else if(rcv.name == "task_wait") {
lock.unlock();
blocked_waiting = true;
device->task_wait();
blocked_waiting = false;
lock.lock();
RPCSend snd(socket, &error_func, "task_wait_done");
snd.write();
lock.unlock();
}
else if(rcv.name == "task_cancel") {
lock.unlock();
device->task_cancel();
}
else if(rcv.name == "acquire_tile") {
AcquireEntry entry;
entry.name = rcv.name;
rcv.read(entry.tile);
acquire_queue.push_back(entry);
lock.unlock();
}
else if(rcv.name == "acquire_tile_none") {
AcquireEntry entry;
entry.name = rcv.name;
acquire_queue.push_back(entry);
lock.unlock();
}
else if(rcv.name == "release_tile") {
AcquireEntry entry;
entry.name = rcv.name;
acquire_queue.push_back(entry);
lock.unlock();
}
else {
cout << "Error: unexpected RPC receive call \"" + rcv.name + "\"\n";
lock.unlock();
}
}