Implemented fp16 source code generation method

llvm/projects/hpvm-tensor-rt/code_autogenerators/source_code_autogenerator.py

@@ -110,16 +110,49 @@ def get_new_function_calls(complete_line, knob_config):
         new_line.append("%s%s, %s)" % (knob_config.modified_func_name, old_func_call[:-1], ', '.join(knob_config.params)))
         orig_func_ind = complete_line.find(knob_config.orig_func_name, line_start_ind)
 	new_line.append(complete_line[line_start_ind : ])
-    #print(new_line)
     return ''.join(new_line)
 
-def generate_fp32_source(knob_config, new_file, source_file):
-    pass
 
-def generate_fp16_source(knob_config, new_file, source_file):
-    pass
-
-def generate_approx_source(knob_config, new_file, source_file, filename_dir):
+def generate_fp32_source(new_file, source_file):
+    # Copy the source code over 
+    new_file.write(source_file.read())
+
+
+def generate_fp16_source(knob_config, new_file, source_file, orig_source_dir):
+    file_contents = source_file.read()
+
+    # Fix all local paths 
+    last_include_ind = file_contents.rfind("#include")
+    last_include_newline_ind = file_contents.find("\n", last_include_ind)
+    include_lines = file_contents[ : last_include_newline_ind].split("\n")
+
+    new_file_contents = []
+    for line in include_lines:
+        if line.startswith("#"):
+            include_file = line.split()[1]
+            if include_file.startswith("\""):
+                new_include_path = get_new_path(include_file.replace("\"", ""), orig_source_dir.replace("\"", ""))
+                new_file_contents.append("#include \"%s\"\n" % new_include_path)
+            else:
+                new_file_contents.append(line)
+    new_file_contents.append(file_contents[last_include_newline_ind : ])
+    new_file_contents = '\n'.join(new_file_contents)
+
+    # Replace all tensorOperation calls with tensorHalfOperation calls
+    # Derived from ../bin/replace_half_calls.py 
+    # NOTE: Not very portable but I don't see another way of ONLY replacing tensorOperation FUNCTION calls
+    new_file_contents = new_file_contents.replace("tensorConvolution", "tensorHalfConvolution")
+    new_file_contents = new_file_contents.replace("tensorAdd", "tensorHalfAdd")
+    new_file_contents = new_file_contents.replace("tensorRelu", "tensorHalfRelu")
+    new_file_contents = new_file_contents.replace("tensorRelu2", "tensorHalfRelu2")
+    new_file_contents = new_file_contents.replace("tensorTanh", "tensorHalfTanh")
+    new_file_contents = new_file_contents.replace("tensorPooling", "tensorHalfPooling")
+    new_file_contents = new_file_contents.replace("tensorGemmGPU", "tensorHalfGemmGPU")
+   
+    new_file.write(new_file_contents)
+
+
+def generate_approx_source(knob_config, new_file, source_file, orig_source_dir):
 	new_file_contents = []
 
 	# Store complete line to handle cases where one line of code is split into two lines
@@ -130,7 +163,7 @@ def generate_approx_source(knob_config, new_file, source_file, filename_dir):
 		if line.startswith("#"):
 			include_file = line.split()[1]
 			if include_file.startswith("\""):
-				new_include_path = get_new_path(include_file.replace("\"", ""), filename_dir.replace("\"", ""))
+				new_include_path = get_new_path(include_file.replace("\"", ""), orig_source_dir.replace("\"", ""))
 				new_file_contents.append("#include \"%s\"\n" % new_include_path)
 			else:
 				new_file_contents.append(line)
@@ -165,19 +198,18 @@ def generate_source_code(table, dir_name, filename, source_name):
         source_name: Filename without the file extension (ex: foo/blah.cc --> blah)
     '''
     source_file = open(filename, "r") 
-    filename_dir = os.path.dirname(filename)
+    orig_source_dir = os.path.dirname(filename)
 
     for knob_config in table:
         source_file.seek(0, 0)
         new_filename = os.path.join(dir_name, "%s_%s.cc" % (source_name, knob_config.id))
         new_file = open(new_filename, "w")
-
         if knob_config.approx == Approx.FP16:
-			pass
+            generate_fp16_source(knob_config, new_file, source_file, orig_source_dir)
         elif knob_config.approx == Approx.FP32:
-			pass
+            generate_fp32_source(new_file, source_file)
         else:
-			generate_approx_source(knob_config, new_file, source_file, filename_dir)
+			generate_approx_source(knob_config, new_file, source_file, orig_source_dir)
 
         new_file.close()
         print("Generated source code as %s" % new_filename)