// Enqueues a kernel, waits for completion, and checks for errors
void RunKernel(Kernel &kernel, Queue &queue, const Device &device,
std::vector<size_t> global, const std::vector<size_t> &local,
EventPointer event, const std::vector<Event> &waitForEvents) {
if (!local.empty()) {
// Tests for validity of the local thread sizes
if (local.size() > device.MaxWorkItemDimensions()) {
throw RuntimeErrorCode(StatusCode::kInvalidLocalNumDimensions);
}
const auto max_work_item_sizes = device.MaxWorkItemSizes();
for (auto i=size_t{0}; i<local.size(); ++i) {
if (local[i] > max_work_item_sizes[i]) {
throw RuntimeErrorCode(StatusCode::kInvalidLocalThreadsDim);
}
}
auto local_size = size_t{1};
for (auto &item: local) { local_size *= item; }
if (local_size > device.MaxWorkGroupSize()) {
throw RuntimeErrorCode(StatusCode::kInvalidLocalThreadsTotal);
}
// Make sure the global thread sizes are at least equal to the local sizes
for (auto i=size_t{0}; i<global.size(); ++i) {
if (global[i] < local[i]) { global[i] = local[i]; }
}
}
// Tests for local memory usage
const auto local_mem_usage = kernel.LocalMemUsage(device);
if (!device.IsLocalMemoryValid(local_mem_usage)) {
throw RuntimeErrorCode(StatusCode::kInvalidLocalMemUsage);
}
// Prints the name of the kernel to launch in case of debugging in verbose mode
#ifdef VERBOSE
queue.Finish();
printf("[DEBUG] Running kernel '%s'\n", kernel.GetFunctionName().c_str());
const auto start_time = std::chrono::steady_clock::now();
#endif
// Launches the kernel (and checks for launch errors)
kernel.Launch(queue, global, local, event, waitForEvents);
// Prints the elapsed execution time in case of debugging in verbose mode
#ifdef VERBOSE
queue.Finish();
const auto elapsed_time = std::chrono::steady_clock::now() - start_time;
const auto timing = std::chrono::duration<double,std::milli>(elapsed_time).count();
printf("[DEBUG] Completed kernel in %.2lf ms\n", timing);
#endif
}
// Retrieves a parameter from the database
size_t operator[](const std::string &key) const {
for (const auto &kernel_name : kernel_names_) {
const auto &kernel_db = databases_.find(kernel_name)->second;
if (kernel_db.exists(key)) { return kernel_db[key]; }
}
throw RuntimeErrorCode(StatusCode::kDatabaseError);
}
void Routine::InitProgram(std::initializer_list<const char *> source) {
// Determines the identifier for this particular routine call
auto routine_info = routine_name_;
for (const auto &kernel_name : kernel_names_) {
routine_info += "_" + kernel_name + db_(kernel_name).GetValuesString();
}
log_debug(routine_info);
// Queries the cache to see whether or not the program (context-specific) is already there
bool has_program;
program_ = ProgramCache::Instance().Get(ProgramKeyRef{ context_(), device_(), precision_, routine_info },
&has_program);
if (has_program) { return; }
// Sets the build options from an environmental variable (if set)
auto options = std::vector<std::string>();
const auto environment_variable = std::getenv("CLBLAST_BUILD_OPTIONS");
if (environment_variable != nullptr) {
options.push_back(std::string(environment_variable));
}
// Queries the cache to see whether or not the binary (device-specific) is already there. If it
// is, a program is created and stored in the cache
const auto device_name = GetDeviceName(device_);
const auto platform_id = device_.PlatformID();
bool has_binary;
auto binary = BinaryCache::Instance().Get(BinaryKeyRef{platform_id, precision_, routine_info, device_name },
&has_binary);
if (has_binary) {
program_ = std::make_shared<Program>(device_, context_, binary);
program_->Build(device_, options);
ProgramCache::Instance().Store(ProgramKey{ context_(), device_(), precision_, routine_info },
std::shared_ptr<Program>{program_});
return;
}
// Otherwise, the kernel will be compiled and program will be built. Both the binary and the
// program will be added to the cache.
// Inspects whether or not FP64 is supported in case of double precision
if ((precision_ == Precision::kDouble && !PrecisionSupported<double>(device_)) ||
(precision_ == Precision::kComplexDouble && !PrecisionSupported<double2>(device_))) {
throw RuntimeErrorCode(StatusCode::kNoDoublePrecision);
}
// As above, but for FP16 (half precision)
if (precision_ == Precision::kHalf && !PrecisionSupported<half>(device_)) {
throw RuntimeErrorCode(StatusCode::kNoHalfPrecision);
}
// Collects the parameters for this device in the form of defines
auto source_string = std::string{""};
for (const auto &kernel_name : kernel_names_) {
source_string += db_(kernel_name).GetDefines();
}
// Adds routine-specific code to the constructed source string
for (const char *s: source) {
source_string += s;
}
// Completes the source and compiles the kernel
program_ = CompileFromSource(source_string, precision_, routine_name_,
device_, context_, options, 0);
// Store the compiled binary and program in the cache
BinaryCache::Instance().Store(BinaryKey{platform_id, precision_, routine_info, device_name},
program_->GetIR());
ProgramCache::Instance().Store(ProgramKey{context_(), device_(), precision_, routine_info},
std::shared_ptr<Program>{program_});
}
