Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Caching compiled binaries #148

Merged
merged 2 commits into from
Oct 16, 2023
Merged

Caching compiled binaries #148

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
290eee8
Caching compiled binaries
chhwang Oct 13, 2023
fee2c96
Merge branch 'main' into chhwang/bin-caching
chhwang Oct 16, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 24 additions & 23 deletions ark/gpu/gpu_compile.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,10 +14,12 @@
#include <cstring>
#include <fstream>
#include <functional>
#include <iomanip>
#include <mutex>

#include "cpu_timer.h"
#include "env.h"
#include "file_io.h"
#include "gpu/gpu_logging.h"
#include "include/ark.h"
#include "random.h"
Expand Down Expand Up @@ -169,6 +171,20 @@ static void para_exec(std::vector<ItemType> &items, int max_num_threads,
}
}

// TODO: use a stronger hash function
static std::string fnv1a_hash(const std::string &str) {
const uint64_t FNV_prime = 1099511628211u;
const uint64_t FNV_offset_basis = 14695981039346656037u;
uint64_t hash = FNV_offset_basis;
for (const auto &c : str) {
hash ^= static_cast<uint64_t>(c);
hash *= FNV_prime;
}
std::stringstream ss;
ss << std::hex << std::setfill('0') << std::setw(16) << hash;
return ss.str();
}

const string gpu_compile(const vector<string> &codes,
const GpuArchType &arch_type,
unsigned int max_reg_cnt) {
Expand Down Expand Up @@ -200,27 +216,18 @@ const string gpu_compile(const vector<string> &codes,
items.reserve(codes.size());
srand();
for (auto &code : codes) {
string rand_str;
for (;;) {
rand_str = rand_anum(16);
bool retry = false;
for (auto &p : items) {
if (p.second == rand_str) {
retry = true;
break;
}
}
if (!retry) {
break;
}
}
// TODO: retry if the file name already exists.
items.emplace_back(code, "/tmp/ark_" + rand_str);
string hash_str = fnv1a_hash(code);
items.emplace_back(code, "/tmp/ark_" + hash_str);
}
assert(items.size() == 1);
para_exec<pair<string, string> >(
items, 20, [&arch, &ark_root, max_reg_cnt](pair<string, string> &item) {
string cu_file_path = item.second + ".cu";
string cubin_file_path = item.second + ".cubin";
if (is_exist(cu_file_path) && is_exist(cubin_file_path)) {
LOG(INFO, "Reusing cached binary for ", cu_file_path);
return;
}
// Write CUDA code file.
{
ofstream cu_file(cu_file_path, ios::out | ios::trunc);
Expand Down Expand Up @@ -269,14 +276,8 @@ const string gpu_compile(const vector<string> &codes,
LOG(INFO, "Compile succeed: ", cu_file_path, " (",
cpu_timer() - start, " seconds)");
});
string cu_file_path = items[0].second + ".cu";
string cubin_file_path = items[0].second + ".cubin";
ifstream cubin_file(cubin_file_path);
stringstream ss;
ss << cubin_file.rdbuf();
// remove(cu_file_path.c_str());
remove(cubin_file_path.c_str());
return ss.str();
return read_file(cubin_file_path);
#endif // (ARK_USE_NVRTC)
}

Expand Down
Loading