Added a program cache (per-context) next to the per-device binary cache

3 files changed, 53 insertions, 16 deletions

diff --git a/include/internal/cache.h b/include/internal/cache.h
index fa33b78f..4a11b70f 100644
--- a/include/internal/cache.h
+++ b/include/internal/cache.h
@@ -7,7 +7,7 @@
 // Author(s):
 //   Cedric Nugteren <www.cedricnugteren.nl>
 //
-// This file implements the caching functionality of compiled binaries.
+// This file implements the caching functionality of compiled binaries and programs.
 //
 // =================================================================================================
 
@@ -46,18 +46,46 @@ static std::mutex binary_cache_mutex_;
 
 // =================================================================================================
 
-// Stores the compiled binary in the cache
-void StoreBinaryToCache(const std::string& binary, const std::string &device_name,
+// The cache of compiled OpenCL programs, along with some meta-data
+struct ProgramCache {
+  Program program;
+  ContextPointer context_ptr;
+  Precision precision;
+  std::string routine_name_;
+
+  // Finds out whether the properties match
+  bool MatchInCache(const ContextPointer ref_context, const Precision &ref_precision,
+                    const std::string &ref_routine) {
+    return (context_ptr == ref_context &&
+            precision == ref_precision &&
+            routine_name_ == ref_routine);
+  }
+};
+
+// The actual cache, implemented as a vector of the above data-type, and its mutex
+static std::vector<ProgramCache> program_cache_;
+static std::mutex program_cache_mutex_;
+
+// =================================================================================================
+
+// Stores the compiled binary or program in the cache
+void StoreBinaryToCache(const std::string &binary, const std::string &device_name,
                         const Precision &precision, const std::string &routine_name);
+void StoreProgramToCache(const Program &program, const Context &context,
+                         const Precision &precision, const std::string &routine_name);
 
-// Queries the cache and retrieves a matching binary. Assumes that the match is available, throws
-// otherwise.
+// Queries the cache and retrieves a matching binary or program. Assumes that the match is
+// available, throws otherwise.
 const std::string& GetBinaryFromCache(const std::string &device_name, const Precision &precision,
                                       const std::string &routine_name);
+const Program& GetProgramFromCache(const Context &context, const Precision &precision,
+                                   const std::string &routine_name);
 
 // Queries the cache to see whether or not the compiled kernel is already there
 bool BinaryIsInCache(const std::string &device_name, const Precision &precision,
                      const std::string &routine_name);
+bool ProgramIsInCache(const Context &context, const Precision &precision,
+                      const std::string &routine_name);
 
 // =================================================================================================
 
diff --git a/include/internal/clpp11.h b/include/internal/clpp11.h
index b865ab1e..e70f9000 100644
--- a/include/internal/clpp11.h
+++ b/include/internal/clpp11.h
@@ -269,10 +269,14 @@ class Context {
 
   // Accessor to the private data-member
   const cl_context& operator()() const { return *context_; }
+  cl_context* pointer() const { return &(*context_); }
  private:
   std::shared_ptr<cl_context> context_;
 };
 
+// Pointer to an OpenCL context
+using ContextPointer = cl_context*;
+
 // =================================================================================================
 
 // Enumeration of build statuses of the run-time compilation process
diff --git a/include/internal/routine.h b/include/internal/routine.h
index 32be6012..f2f236ac 100644
--- a/include/internal/routine.h
+++ b/include/internal/routine.h
@@ -84,24 +84,29 @@ class Routine {
                                     const bool upper = false, const bool lower = false,
                                     const bool diagonal_imag_zero = false);
 
-  // Stores a newly compiled binary into the cache
+  // Stores a newly compiled binary/program into the cache
   void StoreBinaryToCache(const std::string& binary) const {
-    return cache::StoreBinaryToCache(binary, device_name_, precision_, routine_name_);
+    cache::StoreBinaryToCache(binary, device_name_, precision_, routine_name_);
+  }
+  void StoreProgramToCache(const Program& program) const {
+    cache::StoreProgramToCache(program, context_, precision_, routine_name_);
   }
 
-  // Queries the cache and retrieve either a matching program or a boolean whether a match exists.
-  // The first assumes that the program is available in the cache and will throw an exception
-  // otherwise.
+  // Queries the cache and retrieve either a matching binary/program or a boolean whether a match
+  // exists. The first assumes that the binary/program is available in the cache and will throw an
+  // exception otherwise.
+  std::string GetBinaryFromCache() const {
+    return cache::GetBinaryFromCache(device_name_, precision_, routine_name_);
+  }
   Program GetProgramFromCache() const {
-    auto& binary = cache::GetBinaryFromCache(device_name_, precision_, routine_name_);
-    auto program = Program(device_, context_, binary);
-    auto options = std::vector<std::string>();
-    program.Build(device_, options);
-    return program;
+    return cache::GetProgramFromCache(context_, precision_, routine_name_);
   }
-  bool ProgramIsInCache() const {
+  bool BinaryIsInCache() const {
     return cache::BinaryIsInCache(device_name_, precision_, routine_name_);
   }
+  bool ProgramIsInCache() const {
+    return cache::ProgramIsInCache(context_, precision_, routine_name_);
+  }
 
   // Non-static variable for the precision. Note that the same variable (but static) might exist in
   // a derived class.