Merge pull request #193 from clEsperanto/add-JIT-cache

add JIT cache on disk

CMakeLists.txt

@@ -71,6 +71,3 @@ if(BUILD_TESTING)
     include(CTest)
     add_subdirectory(${PROJECT_SOURCE_DIR}/tests)
 endif(BUILD_TESTING)
-
-
-

clic/include/backend.hpp

@@ -1,6 +1,7 @@
 #ifndef __INCLUDE_BACKEND_HPP
 #define __INCLUDE_BACKEND_HPP
 
+#include "cache.hpp"
 #include "clic.hpp"
 #include "device.hpp"
 #include "utils.hpp"
@@ -128,10 +129,7 @@ class Backend
               const std::string &     kernel_source,
               const std::string &     kernel_name,
               void *                  kernel) const -> void = 0;
-  virtual auto
-  loadProgramFromCache(const Device::Pointer & device, const std::string & hash, void * program) const -> void = 0;
-  virtual auto
-  saveProgramToCache(const Device::Pointer & device, const std::string & hash, void * program) const -> void = 0;
+
   virtual auto
   executeKernel(const Device::Pointer &       device,
                 const std::string &           kernel_source,
@@ -296,10 +294,6 @@ class CUDABackend : public Backend
             const dType &                 dtype,
             const float &                 value) -> void;
 
-  auto
-  loadProgramFromCache(const Device::Pointer & device, const std::string & hash, void * program) const -> void override;
-  auto
-  saveProgramToCache(const Device::Pointer & device, const std::string & hash, void * program) const -> void override;
   auto
   buildKernel(const Device::Pointer & device,
               const std::string &     kernel_source,
@@ -476,10 +470,6 @@ class OpenCLBackend : public Backend
             const dType &                 dtype,
             const float &                 value) -> void;
 
-  auto
-  loadProgramFromCache(const Device::Pointer & device, const std::string & hash, void * program) const -> void override;
-  auto
-  saveProgramToCache(const Device::Pointer & device, const std::string & hash, void * program) const -> void override;
   auto
   buildKernel(const Device::Pointer & device,
               const std::string &     kernel_source,

clic/include/cache.hpp

@@ -0,0 +1,69 @@
+#ifndef __INCLUDE_CACHE_HPP
+#define __INCLUDE_CACHE_HPP
+
+#include <cstdlib>
+#include <filesystem>
+#include <iostream>
+#include <string>
+
+#include "device.hpp"
+
+#if defined(WIN32) || defined(_WIN32) || defined(__WIN32) && !defined(__CYGWIN__)
+#  define NOMINMAX
+#  include <ShlObj.h>
+#  include <windows.h>
+#endif
+
+namespace cle
+{
+
+static const std::string CACHE_FOLDER = "clesperanto";
+static const std::string CACHE_DIR_WIN = "AppData\\Local";
+static const std::string CACHE_DIR_UNIX = ".cache";
+
+static auto
+get_path_with_cache_folder(const std::filesystem::path & base_path) -> std::filesystem::path
+{
+  return base_path / std::filesystem::path(CACHE_FOLDER);
+}
+
+#if defined(WIN32) || defined(_WIN32) || defined(__WIN32) && !defined(__CYGWIN__)
+static auto
+get_cache_directory_path() -> std::filesystem::path
+{
+  TCHAR path[MAX_PATH];
+  if (FAILED(SHGetFolderPath(NULL, CSIDL_LOCAL_APPDATA, NULL, 0, path)))
+  {
+    std::cerr << "Failed to get " << CACHE_DIR_WIN << " directory\n";
+    return get_path_with_cache_folder(std::filesystem::current_path());
+  }
+  return get_path_with_cache_folder(std::filesystem::path(path));
+}
+#else
+static auto
+get_cache_directory_path() -> std::filesystem::path
+{
+  char * home_dir = std::getenv("HOME");
+  if (home_dir != nullptr)
+  {
+    return get_path_with_cache_folder(std::filesystem::path(home_dir) / std::filesystem::path(CACHE_DIR_UNIX));
+  }
+  std::cerr << "Failed to get user home directory\n";
+  return get_path_with_cache_folder(std::filesystem::current_path() / std::filesystem::path(CACHE_DIR_UNIX));
+}
+#endif
+
+static auto
+is_cache_enabled() -> bool
+{
+  const char * env_var = std::getenv("CLESPERANTO_NO_CACHE");
+  bool         cache_disabled = (env_var != nullptr);
+  return !cache_disabled;
+}
+
+
+static const auto CACHE_FOLDER_PATH = get_cache_directory_path();
+
+} // namespace cle
+
+#endif // __INCLUDE_CACHE_HPP

clic/include/device.hpp

@@ -58,9 +58,9 @@ class Device
   }
 
   friend auto
-  operator<<(std::ostream & out, const Device & device) -> std::ostream &
+  operator<<(std::ostream & out, const Device::Pointer & device) -> std::ostream &
   {
-    out << "(" << device.getType() << ") " << device.getName();
+    out << "(" << device->getType() << ") " << device->getName();
     return out;
   }
 };
@@ -69,8 +69,6 @@ class Device
 class OpenCLDevice : public Device
 {
 public:
-  using CacheType = std::unordered_map<std::string, cl_program>;
-
   OpenCLDevice(const cl_platform_id & platform, const cl_device_id & device);
   ~OpenCLDevice() override;
 
@@ -99,15 +97,12 @@ class OpenCLDevice : public Device
   getName() const -> std::string override;
   [[nodiscard]] auto
   getInfo() const -> std::string override;
-  [[nodiscard]] auto
-  getCache() -> CacheType &;
 
 private:
   cl_device_id     clDevice;
   cl_platform_id   clPlatform;
   cl_context       clContext;
   cl_command_queue clCommandQueue;
-  CacheType        cache;
   bool             initialized = false;
   bool             waitFinish = false;
 };
@@ -117,8 +112,6 @@ class OpenCLDevice : public Device
 class CUDADevice : public Device
 {
 public:
-  using CacheType = std::unordered_map<std::string, CUmodule>;
-
   explicit CUDADevice(int deviceIndex);
   ~CUDADevice() override;
 
@@ -149,8 +142,6 @@ class CUDADevice : public Device
   getInfo() const -> std::string override;
   [[nodiscard]] auto
   getArch() const -> std::string;
-  [[nodiscard]] auto
-  getCache() -> CacheType &;
 
 private:
   int       cudaDeviceIndex;
@@ -159,7 +150,6 @@ class CUDADevice : public Device
   CUstream  cudaStream;
   bool      initialized = false;
   bool      waitFinish = false;
-  CacheType cache;
 };
 #endif // USE_CUDA