not inserting constexpr int loop unroll factor when 0

Author: benvanwerkhoven

kernel_tuner/util.py

@@ -606,9 +606,10 @@ def prepare_kernel_string(kernel_name, kernel_string, params, grid, threads, blo
             # pragma unroll loop_unroll_factor, loop_unroll_factor should be a constant integer expression
             # in OpenCL this isn't the case and we can just insert "#define loop_unroll_factor N"
             # using 0 to disable specifying a loop unrolling factor for this loop
-            kernel_prefix += f"constexpr int {k} = {v};\n"
             if v == "0":
                 kernel_string = re.sub(r"\n\s*#pragma\s+unroll\s+" + k, "\n", kernel_string)    # + r"[^\S]*"
+            else:
+                kernel_prefix += f"constexpr int {k} = {v};\n"
         else:
             kernel_prefix += f"#define {k} {v}\n"
 

test/test_util_functions.py

@@ -198,6 +198,28 @@ def test_prepare_kernel_string():
         prepare_kernel_string("this", kernel, params, grid, threads, block_size_names, "", invalid_defines)
 
 
+def test_prepare_kernel_string_partial_loop_unrolling():
+
+    kernel = """this is a weird kernel(what * language, is this, anyway* C) {
+        #pragma unroll loop_unroll_factor_monkey
+        for monkey in the forest {
+            eat(banana);
+        }
+    }"""
+    threads = (1, 2, 3)
+    grid = (4, 5, 6)
+    params = dict()
+    params["loop_unroll_factor_monkey"] = 8
+
+    _, output = prepare_kernel_string("this", kernel, params, grid, threads, block_size_names, "CUDA", None)
+    assert "constexpr int loop_unroll_factor_monkey = 8;" in output
+
+    params["loop_unroll_factor_monkey"] = 0
+    _, output = prepare_kernel_string("this", kernel, params, grid, threads, block_size_names, "CUDA", None)
+    assert not "constexpr int loop_unroll_factor_monkey" in output
+    assert not "#pragma unroll loop_unroll_factor_monkey" in output
+
+
 
 def test_replace_param_occurrences():
     kernel = "this is a weird kernel"