Merge pull request #246 from vortexgpgpu/bug_fixes

extending divergence test coverage

sim/common/util.cpp

@@ -12,6 +12,8 @@
 // limitations under the License.
 
 #include "util.h"
+#include <fstream>
+#include <sstream>
 #include <string.h>
 
 // return file extension
@@ -36,4 +38,21 @@ void aligned_free(void *ptr) {
   // retreive the stored unaligned address and use it to free the allocation
   void* unaligned_addr = ((void**)ptr)[-1];
   free(unaligned_addr);
+}
+
+std::string vortex::resolve_file_path(const std::string& filename, const std::string& searchPaths) {
+  std::ifstream ifs(filename);
+  if (!ifs) {
+    std::stringstream ss(searchPaths);
+    std::string path;
+    while (std::getline(ss, path, ',')) {
+      if (!path.empty()) {
+        std::string filePath = path + "/" + filename;
+        std::ifstream ifs(filePath);
+        if (ifs)
+          return filePath;
+      }
+    }
+  }
+  return filename;
 }

sim/common/util.h

@@ -17,6 +17,7 @@
 #include <algorithm>
 #include <assert.h>
 #include <bitmanip.h>
+#include <string>
 
 template <typename... Args>
 void unused(Args&&...) {}
@@ -94,4 +95,6 @@ public:
   }
 };
 
+std::string resolve_file_path(const std::string& filename, const std::string& searchPaths);
+
 }

tests/regression/diverge/kernel.cpp

@@ -5,6 +5,14 @@
 
 // Parallel Selection sort
 
+struct key_t {
+	uint32_t user = 0;
+};
+
+static __attribute__((noinline)) void hacker(key_t* key, uint32_t task_id) {
+	key->user = task_id;
+}
+
 void kernel_body(kernel_arg_t* __UNIFORM__ arg) {
 	int32_t* src_ptr = (int32_t*)arg->src_addr;
 	int32_t* dst_ptr = (int32_t*)arg->dst_addr;
@@ -13,6 +21,15 @@ void kernel_body(kernel_arg_t* __UNIFORM__ arg) {
 
 	int value = src_ptr[task_id];
 
+	key_t key;
+	uint32_t samples = arg->num_points;
+  while (samples--) {
+		hacker(&key, task_id);
+		if ((key.user & 0x1) == 0) {
+			value += 1;
+		}
+	}
+
 	// none taken
 	if (task_id >= 0x7fffffff) {
 		value = 0;

tests/regression/diverge/main.cpp

@@ -78,6 +78,14 @@ void gen_ref_data(std::vector<int>& ref_data, const std::vector<int>& src_data,
   for (int i = 0; i < (int)size; ++i) {
     int value = src_data.at(i);
 
+    key_t key;
+    uint32_t samples = size;
+    while (samples--) {
+      if ((i & 0x1) == 0) {
+        value += 1;
+      }
+    }
+
     // none taken
     if (i >= 0x7fffffff) {
       value = 0;