Added first (untested) version of a CUDA API

3 files changed, 47 insertions, 15 deletions

diff --git a/scripts/generator/generator.py b/scripts/generator/generator.py
index 0d34d7fe..520e3fc8 100755
--- a/scripts/generator/generator.py
+++ b/scripts/generator/generator.py
@@ -12,6 +12,8 @@
 #    clblast.cpp
 #    clblast_c.h
 #    clblast_c.cpp
+#    clblast_cuda.h
+#    clblast_cuda.cpp
 #    clblast_netlib_c.h
 #    clblast_netlib_c.cpp
 #    wrapper_clblas.h
@@ -41,9 +43,11 @@ FILES = [
     "/test/wrapper_cublas.hpp",
     "/include/clblast_netlib_c.h",
     "/src/clblast_netlib_c.cpp",
+    "/include/clblast_cuda.h",
+    "/src/clblast_cuda.cpp",
 ]
-HEADER_LINES = [122, 21, 126, 24, 29, 41, 29, 65, 32]
-FOOTER_LINES = [25, 3, 27, 38, 6, 6, 6, 9, 2]
+HEADER_LINES = [122, 21, 126, 24, 29, 41, 29, 65, 32, 94, 21]
+FOOTER_LINES = [25, 3, 27, 38, 6, 6, 6, 9, 2, 25, 3]
 HEADER_LINES_DOC = 0
 FOOTER_LINES_DOC = 63
 
@@ -224,6 +228,10 @@ def main(argv):
                     if i == 8:
                         if not routine.batched:
                             body += cpp.clblast_netlib_c_cc(routine)
+                    if i == 9:
+                        body += cpp.clblast_h(routine, cuda=True)
+                    if i == 10:
+                        body += cpp.clblast_cc(routine, cuda=True)
             f.write("".join(file_header))
             f.write(body)
             f.write("".join(file_footer))
diff --git a/scripts/generator/generator/cpp.py b/scripts/generator/generator/cpp.py
index 5fef3083..f1ee1959 100644
--- a/scripts/generator/generator/cpp.py
+++ b/scripts/generator/generator/cpp.py
@@ -36,19 +36,19 @@ HEADER = NL + SEPARATOR + """
 """ + SEPARATOR + NL
 
 
-def clblast_h(routine):
+def clblast_h(routine, cuda=False):
     """The C++ API header (.h)"""
     result = NL + "// " + routine.description + ": " + routine.short_names() + NL
-    result += routine.routine_header_cpp(12, " = nullptr") + ";" + NL
+    result += routine.routine_header_cpp(12, " = nullptr", cuda) + ";" + NL
     return result
 
 
-def clblast_cc(routine):
+def clblast_cc(routine, cuda=False):
     """The C++ API implementation (.cpp)"""
     indent1 = " " * (15 + routine.length())
     result = NL + "// " + routine.description + ": " + routine.short_names() + NL
     if routine.implemented:
-        result += routine.routine_header_cpp(12, "") + " {" + NL
+        result += routine.routine_header_cpp(12, "", cuda) + " {" + NL
         result += "  try {" + NL
         result += "    auto queue_cpp = Queue(*queue);" + NL
         result += "    auto routine = X" + routine.plain_name() + "<" + routine.template.template + ">(queue_cpp, event);" + NL
@@ -60,14 +60,22 @@ def clblast_cc(routine):
         result += "    return StatusCode::kSuccess;" + NL
         result += "  } catch (...) { return DispatchException(); }" + NL
     else:
-        result += routine.routine_header_type_cpp(12) + " {" + NL
+        result += routine.routine_header_type_cpp(12, cuda) + " {" + NL
         result += "  return StatusCode::kNotImplemented;" + NL
     result += "}" + NL
     for flavour in routine.flavours:
         indent2 = " " * (34 + routine.length() + len(flavour.template))
         result += "template StatusCode PUBLIC_API " + routine.capitalized_name() + "<" + flavour.template + ">("
-        result += ("," + NL + indent2).join([a for a in routine.arguments_type(flavour)])
-        result += "," + NL + indent2 + "cl_command_queue*, cl_event*);" + NL
+        arguments = routine.arguments_type(flavour)
+        if cuda:
+            arguments = [a.replace("cl_mem", "CUdeviceptr") for a in arguments]
+        result += ("," + NL + indent2).join([a for a in arguments])
+        result += "," + NL + indent2
+        if cuda:
+            result += "CUstream*"
+        else:
+            result += "cl_command_queue*, cl_event*"
+        result += ");" + NL
     return result
 
 
diff --git a/scripts/generator/generator/routine.py b/scripts/generator/generator/routine.py
index cef7db87..c3c1f775 100644
--- a/scripts/generator/generator/routine.py
+++ b/scripts/generator/generator/routine.py
@@ -802,22 +802,38 @@ class Routine:
         """Retrieves a list of routine requirements for documentation"""
         return self.requirements
 
-    def routine_header_cpp(self, spaces, default_event):
+    def routine_header_cpp(self, spaces, default_event, cuda=False):
         """Retrieves the C++ templated definition for a routine"""
         indent = " " * (spaces + self.length())
+        arguments = self.arguments_def(self.template)
+        if cuda:
+            arguments = [a.replace("cl_mem", "CUdeviceptr") for a in arguments]
         result = "template <" + self.template.name + ">\n"
         result += "StatusCode " + self.capitalized_name() + "("
-        result += (",\n" + indent).join([a for a in self.arguments_def(self.template)])
-        result += ",\n" + indent + "cl_command_queue* queue, cl_event* event" + default_event + ")"
+        result += (",\n" + indent).join([a for a in arguments])
+        result += ",\n" + indent
+        if cuda:
+            result += "CUstream* stream"
+        else:
+            result += "cl_command_queue* queue, cl_event* event" + default_event
+        result += ")"
         return result
 
-    def routine_header_type_cpp(self, spaces):
+    def routine_header_type_cpp(self, spaces, cuda=False):
         """As above, but now without variable names"""
         indent = " " * (spaces + self.length())
+        arguments = self.arguments_type(self.template)
+        if cuda:
+            arguments = [a.replace("cl_mem", "CUdeviceptr") for a in arguments]
         result = "template <" + self.template.name + ">\n"
         result += "StatusCode " + self.capitalized_name() + "("
-        result += (",\n" + indent).join([a for a in self.arguments_type(self.template)])
-        result += ",\n" + indent + "cl_command_queue*, cl_event*)"
+        result += (",\n" + indent).join([a for a in arguments])
+        result += ",\n" + indent
+        if cuda:
+            result += "CUstream* stream"
+        else:
+            result += "cl_command_queue*, cl_event*"
+        result += ")"
         return result
 
     def routine_header_c(self, flavour, spaces, extra_qualifier):