space-time-stack: allow demangled names

CMakeLists.txt

@@ -116,6 +116,9 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}/profiling/all)
 set(COMMON_HEADERS_PATH ${CMAKE_CURRENT_BINARY_DIR}/common)
 include_directories(${COMMON_HEADERS_PATH})
 
+# Allow all tools to include any file.
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+
 set(SINGLELIB_PROFILERS "" CACHE STRING "" FORCE)
 
 # Export settings

common/utils/demangle.hpp

@@ -0,0 +1,37 @@
+#ifndef KOKKOSTOOLS_COMMON_UTILS_DEMANGLE_HPP
+#define KOKKOSTOOLS_COMMON_UTILS_DEMANGLE_HPP
+
+#include <memory>
+#include <string>
+
+#if defined(__GXX_ABI_VERSION)
+#define HAVE_GCC_ABI_DEMANGLE
+#endif
+
+#if defined(HAVE_GCC_ABI_DEMANGLE)
+#include <cxxabi.h>
+#endif  // HAVE_GCC_ABI_DEMANGLE
+
+namespace KokkosTools
+{
+
+//! Demangme @p mangled_name.
+inline std::string demangleName(const std::string& mangled_name)
+{
+#if defined(HAVE_GCC_ABI_DEMANGLE)
+    int status = 0;
+
+    std::unique_ptr<char, decltype(&std::free)> demangled_name(
+      abi::__cxa_demangle(mangled_name.c_str(), nullptr, nullptr, &status),
+      &std::free
+    );
+
+    if(status == 0) return std::string(demangled_name.get());
+    else demangled_name.reset();
+#endif
+  return mangled_name;
+}
+
+} // namespace KokkosTools
+
+#endif // KOKKOSTOOLS_COMMON_UTILS_DEMANGLE_HPP

profiling/simple-kernel-timer/kp_json_writer.cpp

@@ -57,20 +57,6 @@ inline void write_json(std::ostream& os, KernelPerformanceInfo const& kp,
 }
 // clang-format off
 
-int find_index(std::vector<KernelPerformanceInfo*>& kernels,
-	const char* kernelName) {
-
-	for(unsigned int i = 0; i < kernels.size(); i++) {
-		KernelPerformanceInfo* nextKernel = kernels[i];
-
-		if(strcmp(nextKernel->getName(), kernelName) == 0) {
-			return i;
-		}
-	}
-
-	return -1;
-}
-
 int main(int argc, char* argv[]) {
 
 	if(argc == 1) {
@@ -100,8 +86,8 @@ int main(int argc, char* argv[]) {
 		while(! feof(the_file)) {
 			KernelPerformanceInfo* new_kernel = new KernelPerformanceInfo("", PARALLEL_FOR);
 			if(new_kernel->readFromFile(the_file)) {
-			   if(strlen(new_kernel->getName()) > 0) {
-				int kernelIndex = find_index(kernelInfo, new_kernel->getName());
+			   if( ! new_kernel->getName().empty()) {
+				int kernelIndex = find_index(kernelInfo, new_kernel->getName().c_str());
 
 				if(kernelIndex > -1) {
 					kernelInfo[kernelIndex]->addTime(new_kernel->getTime());

profiling/simple-kernel-timer/kp_kernel_info.h

@@ -22,29 +22,10 @@
 #include <string>
 #include <cstring>
 
-#if defined(__GXX_ABI_VERSION)
-#define HAVE_GCC_ABI_DEMANGLE
-#endif
-
-#if defined(HAVE_GCC_ABI_DEMANGLE)
-#include <cxxabi.h>
-#endif  // HAVE_GCC_ABI_DEMANGLE
+#include "common/utils/demangle.hpp"
 
 namespace KokkosTools::KernelTimer {
 
-inline char* demangleName(char* kernelName) {
-#if defined(HAVE_GCC_ABI_DEMANGLE)
-  int status = -1;
-  char* demangledKernelName =
-      abi::__cxa_demangle(kernelName, NULL, NULL, &status);
-  if (status == 0) {
-    free(kernelName);
-    kernelName = demangledKernelName;
-  }
-#endif  // HAVE_GCC_ABI_DEMANGLE
-  return kernelName;
-}
-
 inline double seconds() {
   struct timeval now;
   gettimeofday(&now, NULL);
@@ -62,15 +43,7 @@ enum KernelExecutionType {
 class KernelPerformanceInfo {
  public:
   KernelPerformanceInfo(std::string kName, KernelExecutionType kernelType)
-      : kType(kernelType) {
-    kernelName = (char*)malloc(sizeof(char) * (kName.size() + 1));
-    strcpy(kernelName, kName.c_str());
-
-    callCount = 0;
-    time      = 0;
-  }
-
-  ~KernelPerformanceInfo() { free(kernelName); }
+      : kernelName(std::move(kName)), kType(kernelType) {}
 
   KernelExecutionType getKernelType() const { return kType; }
 
@@ -95,7 +68,7 @@ class KernelPerformanceInfo {
 
   double getTimeSq() { return timeSq; }
 
-  char* getName() const { return kernelName; }
+  const std::string& getName() const { return kernelName; }
 
   void addCallCount(const uint64_t newCalls) { callCount += newCalls; }
 
@@ -112,13 +85,7 @@ class KernelPerformanceInfo {
     copy((char*)&kernelNameLength, &entry[nextIndex], sizeof(kernelNameLength));
     nextIndex += sizeof(kernelNameLength);
 
-    if (strlen(kernelName) > 0) {
-      free(kernelName);
-    }
-
-    kernelName = (char*)malloc(sizeof(char) * (kernelNameLength + 1));
-    copy(kernelName, &entry[nextIndex], kernelNameLength);
-    kernelName[kernelNameLength] = '\0';
+    this->kernelName = std::string(&entry[nextIndex], kernelNameLength);
 
     kernelName = demangleName(kernelName);
 
@@ -152,7 +119,7 @@ class KernelPerformanceInfo {
   }
 
   void writeToBinaryFile(FILE* output) {
-    const uint32_t kernelNameLen = (uint32_t)strlen(kernelName);
+    const uint32_t kernelNameLen = kernelName.size();
     const uint32_t recordLen = sizeof(uint32_t) + sizeof(char) * kernelNameLen +
                                sizeof(uint64_t) + sizeof(double) +
                                sizeof(double) + sizeof(uint32_t);
@@ -163,7 +130,7 @@ class KernelPerformanceInfo {
     copy(&entry[nextIndex], (char*)&kernelNameLen, sizeof(kernelNameLen));
     nextIndex += sizeof(kernelNameLen);
 
-    copy(&entry[nextIndex], kernelName, kernelNameLen);
+    copy(&entry[nextIndex], kernelName.c_str(), kernelNameLen);
     nextIndex += kernelNameLen;
 
     copy(&entry[nextIndex], (char*)&callCount, sizeof(callCount));
@@ -191,7 +158,7 @@ class KernelPerformanceInfo {
     snprintf(indentBuffer, 256, "%s    ", indent);
 
     fprintf(output, "%s\"kernel-name\"    : \"%s\",\n", indentBuffer,
-            kernelName);
+            kernelName.c_str());
     // fprintf(output, "%s\"region\"         : \"%s\",\n", indentBuffer,
     // regionName);
     fprintf(output, "%s\"call-count\"     : %llu,\n", indentBuffer,
@@ -216,12 +183,12 @@ class KernelPerformanceInfo {
     }
   }
 
-  char* kernelName;
+  std::string kernelName;
   // const char* regionName;
-  uint64_t callCount;
-  double time;
-  double timeSq;
-  double startTime;
+  uint64_t callCount = 0;
+  double time = 0;
+  double timeSq = 0;
+  double startTime = 0;
   KernelExecutionType kType;
 };
 

profiling/simple-kernel-timer/kp_reader.cpp

@@ -25,19 +25,6 @@
 
 using namespace KokkosTools::KernelTimer;
 
-int find_index(std::vector<KernelPerformanceInfo*>& kernels,
-               const char* kernelName) {
-  for (unsigned int i = 0; i < kernels.size(); i++) {
-    KernelPerformanceInfo* nextKernel = kernels[i];
-
-    if (strcmp(nextKernel->getName(), kernelName) == 0) {
-      return i;
-    }
-  }
-
-  return -1;
-}
-
 int main(int argc, char* argv[]) {
   if (argc == 1) {
     fprintf(stderr, "Did you specify any data files on the command line!\n");
@@ -77,8 +64,9 @@ int main(int argc, char* argv[]) {
       KernelPerformanceInfo* new_kernel =
           new KernelPerformanceInfo("", PARALLEL_FOR);
       if (new_kernel->readFromFile(the_file)) {
-        if (strlen(new_kernel->getName()) > 0) {
-          int kernelIndex = find_index(kernelInfo, new_kernel->getName());
+        if (!new_kernel->getName().empty()) {
+          int kernelIndex =
+              find_index(kernelInfo, new_kernel->getName().c_str());
 
           if (kernelIndex > -1) {
             kernelInfo[kernelIndex]->addTime(new_kernel->getTime());
@@ -110,7 +98,7 @@ int main(int argc, char* argv[]) {
     if (kernelInfo[i]->getKernelType() != REGION) continue;
     if (fixed_width)
       printf("- %100s\n%11s%c%15.5f%c%12" PRIu64 "%c%15.5f%c%7.3f%c%7.3f\n",
-             kernelInfo[i]->getName(),
+             kernelInfo[i]->getName().c_str(),
              (kernelInfo[i]->getKernelType() == PARALLEL_FOR)
                  ? (" (ParFor)  ")
                  : ((kernelInfo[i]->getKernelType() == PARALLEL_REDUCE)
@@ -125,7 +113,7 @@ int main(int argc, char* argv[]) {
              (kernelInfo[i]->getTime() / totalExecuteTime) * 100.0);
     else
       printf("- %s\n%s%c%f%c%" PRIu64 "%c%f%c%f%c%f\n",
-             kernelInfo[i]->getName(),
+             kernelInfo[i]->getName().c_str(),
              (kernelInfo[i]->getKernelType() == PARALLEL_FOR)
                  ? (" (ParFor)  ")
                  : ((kernelInfo[i]->getKernelType() == PARALLEL_REDUCE)
@@ -152,7 +140,7 @@ int main(int argc, char* argv[]) {
     if (kernelInfo[i]->getKernelType() == REGION) continue;
     if (fixed_width)
       printf("- %100s\n%11s%c%15.5f%c%12" PRIu64 "%c%15.5f%c%7.3f%c%7.3f\n",
-             kernelInfo[i]->getName(),
+             kernelInfo[i]->getName().c_str(),
              (kernelInfo[i]->getKernelType() == PARALLEL_FOR)
                  ? (" (ParFor)  ")
                  : ((kernelInfo[i]->getKernelType() == PARALLEL_REDUCE)
@@ -167,7 +155,7 @@ int main(int argc, char* argv[]) {
              (kernelInfo[i]->getTime() / totalExecuteTime) * 100.0);
     else
       printf("- %s\n%s%c%f%c%" PRIu64 "%c%f%c%f%c%f\n",
-             kernelInfo[i]->getName(),
+             kernelInfo[i]->getName().c_str(),
              (kernelInfo[i]->getKernelType() == PARALLEL_FOR)
                  ? (" (ParFor)  ")
                  : ((kernelInfo[i]->getKernelType() == PARALLEL_REDUCE)

profiling/simple-kernel-timer/kp_shared.h

@@ -17,8 +17,11 @@
 #ifndef _H_KOKKOSP_KERNEL_SHARED
 #define _H_KOKKOSP_KERNEL_SHARED
 
+#include <algorithm>
 #include <map>
 #include <memory>
+#include <vector>
+
 #include "kp_kernel_info.h"
 
 namespace KokkosTools::KernelTimer {
@@ -39,6 +42,14 @@ inline bool compareKernelPerformanceInfo(KernelPerformanceInfo* left,
   return left->getTime() > right->getTime();
 };
 
+inline int find_index(std::vector<KernelPerformanceInfo*>& kernels,
+                      const char* kernelName) {
+  const auto it = std::find_if(
+      kernels.begin(), kernels.end(),
+      [&](const auto& kernel) { return kernel->getName() == kernelName; });
+  return it == kernels.cend() ? -1 : std::distance(kernels.begin(), it);
+}
+
 }  // namespace KokkosTools::KernelTimer
 
 #endif  // _H_KOKKOSP_KERNEL_SHARED

profiling/space-time-stack/kp_space_time_stack.cpp

@@ -31,6 +31,8 @@
 #include <algorithm>
 #include <cstring>
 
+#include "common/utils/demangle.hpp"
+
 #include "kp_core.hpp"
 
 #if USE_MPI
@@ -741,7 +743,7 @@ struct State {
   }
 
   void begin_frame(const char* name, StackKind kind) {
-    std::string name_str(name);
+    std::string name_str(demangleName(name));
     stack_frame = stack_frame->get_child(std::move(name_str), kind);
     stack_frame->begin();
   }