/**
* Distribute the tunecache from node 0 to all other nodes.
*/
static void broadcastTuneCache()
{
#ifdef MULTI_GPU
std::stringstream serialized;
size_t size;
if (comm_rank() == 0) {
serializeTuneCache(serialized);
size = serialized.str().length();
}
comm_broadcast(&size, sizeof(size_t));
if (size > 0) {
if (comm_rank() == 0) {
comm_broadcast(const_cast<char *>(serialized.str().c_str()), size);
} else {
char *serstr = new char[size+1];
comm_broadcast(serstr, size);
serstr[size] ='\0'; // null-terminate
serialized.str(serstr);
deserializeTuneCache(serialized);
delete[] serstr;
}
}
#endif
}
/*
* Read tunecache from disk.
*/
void loadTuneCache(QudaVerbosity verbosity)
{
char *path;
struct stat pstat;
std::string cache_path, line, token;
std::ifstream cache_file;
std::stringstream ls;
path = getenv("QUDA_RESOURCE_PATH");
if (!path) {
warningQuda("Environment variable QUDA_RESOURCE_PATH is not set.");
warningQuda("Caching of tuned parameters will be disabled.");
return;
} else if (stat(path, &pstat) || !S_ISDIR(pstat.st_mode)) {
warningQuda("The path \"%s\" specified by QUDA_RESOURCE_PATH does not exist or is not a directory.", path);
warningQuda("Caching of tuned parameters will be disabled.");
return;
} else {
resource_path = path;
}
#ifdef MULTI_GPU
if (comm_rank() == 0) {
#endif
cache_path = resource_path;
cache_path += "/tunecache.tsv";
cache_file.open(cache_path.c_str());
if (cache_file) {
if (!cache_file.good()) errorQuda("Bad format in %s", cache_path.c_str());
getline(cache_file, line);
ls.str(line);
ls >> token;
if (token.compare("tunecache")) errorQuda("Bad format in %s", cache_path.c_str());
ls >> token;
if (token.compare(quda_version)) errorQuda("Cache file %s does not match current QUDA version", cache_path.c_str());
ls >> token;
if (token.compare(quda_hash)) warningQuda("Cache file %s does not match current QUDA build", cache_path.c_str());
if (!cache_file.good()) errorQuda("Bad format in %s", cache_path.c_str());
getline(cache_file, line); // eat the blank line
if (!cache_file.good()) errorQuda("Bad format in %s", cache_path.c_str());
getline(cache_file, line); // eat the description line
deserializeTuneCache(cache_file);
cache_file.close();
initial_cache_size = tunecache.size();
if (verbosity >= QUDA_SUMMARIZE) {
printfQuda("Loaded %d sets of cached parameters from %s\n", static_cast<int>(initial_cache_size), cache_path.c_str());
}
} else {
