Integrated pre-processor in compilation flow, default is still disabled

4 files changed, 17 insertions, 6 deletions

diff --git a/src/routine.cpp b/src/routine.cpp
index 93882fbf..8c9317d1 100644
--- a/src/routine.cpp
+++ b/src/routine.cpp
@@ -148,7 +148,7 @@ void Routine::InitProgram(std::initializer_list<const char *> source) {
 
   // Completes the source and compiles the kernel
   program_ = CompileFromSource(source_string, precision_, routine_name_,
-                               device_, context_, options);
+                               device_, context_, options, false);
 
 
   // Store the compiled binary and program in the cache
diff --git a/src/tuning/tuning.hpp b/src/tuning/tuning.hpp
index 12826d0c..329314e5 100644
--- a/src/tuning/tuning.hpp
+++ b/src/tuning/tuning.hpp
@@ -227,7 +227,7 @@ void Tuner(int argc, char* argv[]) {
     // Compiles the kernel
     auto compiler_options = std::vector<std::string>();
     const auto program = CompileFromSource(settings.sources, args.precision, settings.kernel_name,
-                                           device, context, compiler_options);
+                                           device, context, compiler_options, false);
     auto kernel = Kernel(program, settings.kernel_name);
     C::SetArguments(kernel, args, device_buffers);
     printf("             %sOK%s |", kPrintSuccess.c_str(), kPrintEnd.c_str());
@@ -286,7 +286,7 @@ void Tuner(int argc, char* argv[]) {
       const auto start_time = std::chrono::steady_clock::now();
       auto compiler_options = std::vector<std::string>();
       const auto program = CompileFromSource(kernel_source, args.precision, settings.kernel_name,
-                                             device, context, compiler_options, true);
+                                             device, context, compiler_options, false, true);
       auto kernel = Kernel(program, settings.kernel_name);
       const auto elapsed_time = std::chrono::steady_clock::now() - start_time;
       const auto timing = std::chrono::duration<double,std::milli>(elapsed_time).count();
diff --git a/src/utilities/compile.cpp b/src/utilities/compile.cpp
index 2a55506e..4d1e8929 100644
--- a/src/utilities/compile.cpp
+++ b/src/utilities/compile.cpp
@@ -15,6 +15,7 @@
 #include <chrono>
 
 #include "routines/common.hpp"
+#include "kernel_preprocessor.hpp"
 
 namespace clblast {
 // =================================================================================================
@@ -23,7 +24,8 @@ namespace clblast {
 Program CompileFromSource(const std::string &source_string, const Precision precision,
                           const std::string &routine_name,
                           const Device& device, const Context& context,
-                          std::vector<std::string>& options, const bool silent) {
+                          std::vector<std::string>& options,
+                          const bool run_preprocessor, const bool silent) {
   auto header_string = std::string{""};
 
   header_string += "#define PRECISION " + ToString(static_cast<int>(precision)) + "\n";
@@ -73,8 +75,15 @@ Program CompileFromSource(const std::string &source_string, const Precision prec
     const auto start_time = std::chrono::steady_clock::now();
   #endif
 
+  // Runs a pre-processor to unroll loops and perform array-to-register promotion
+  auto kernel_string = header_string + source_string;
+  if (run_preprocessor) {
+    log_debug("Running built-in pre-processor");
+    kernel_string = PreprocessKernelSource(kernel_string);
+  }
+
   // Compiles the kernel
-  auto program = Program(context, header_string + source_string);
+  auto program = Program(context, kernel_string);
   try {
     program.Build(device, options);
   } catch (const CLCudaAPIBuildError &e) {
diff --git a/src/utilities/compile.hpp b/src/utilities/compile.hpp
index 0315d70c..0df2ded5 100644
--- a/src/utilities/compile.hpp
+++ b/src/utilities/compile.hpp
@@ -27,7 +27,9 @@ namespace clblast {
 Program CompileFromSource(const std::string &source_string, const Precision precision,
                           const std::string &routine_name,
                           const Device& device, const Context& context,
-                          std::vector<std::string>& options, const bool silent = false);
+                          std::vector<std::string>& options,
+                          const bool run_preprocessor,
+                          const bool silent = false);
 
 // =================================================================================================
 } // namespace clblast