Updated for node index type stabalization.

examples/benchmark_pointer.cpp

@@ -7,6 +7,7 @@
 
 #include <pando-lib-galois/containers/host_local_storage.hpp>
 #include <pando-lib-galois/containers/pod_local_storage.hpp>
+#include <pando-rt/memory/address_translation.hpp>
 #include <pando-rt/memory/memory_guard.hpp>
 #include <pando-rt/pando-rt.hpp>
 
@@ -15,59 +16,175 @@
  * This benchmarks various types of pointer dereferences
  */
 
+std::uint64_t currentLocation = 100;
+std::uint64_t nodeDims = 100;
+
+template <typename T>
+struct CacheRef {
+  pando::GlobalPtr<T> globalPtr;
+  std::uint64_t cacheLoc;
+  T* cachePtr;
+
+private:
+  bool tryToCache() {
+    if (cacheLoc == nodeDims) {
+      std::uint64_t nodeIndex = pando::extractNodeIndex(globalPtr.address).id;
+      if (currentLocation == nodeIndex) {
+        cacheLoc = nodeIndex;
+        cachePtr = globalPtr.operator->();
+        return true;
+      }
+    }
+    return false;
+  }
+
+public:
+  operator T() const {
+    if (cacheLoc == currentLocation || tryToCache()) {
+      return *cachePtr;
+    } else {
+      return *globalPtr;
+    }
+  }
+
+  CacheRef<T> operator=(const T& value) {
+    if (cacheLoc == currentLocation || tryToCache()) {
+      *cachePtr = value;
+    } else {
+      *globalPtr = value;
+    }
+    return this;
+  }
+  template <typename U>
+  CacheRef<T> operator+=(const U& y) {
+    if (cacheLoc == currentLocation || tryToCache()) {
+      *cachePtr += y;
+    } else {
+      *globalPtr += y;
+    }
+    return *this;
+  }
+};
+
+template <typename T>
+struct CachePtr {
+  CacheRef<T> ref;
+
+  CacheRef<T> operator*() {
+    return ref;
+  }
+};
+
 void printUsageExit(char* argv0) {
-  std::cerr << "Usage: " << argv0 << " -n numAccesses" << std::endl;
+  std::cerr << "Usage: " << argv0 << " -n numAccesses [-r] [-g] [-c]" << std::endl;
   std::exit(EXIT_FAILURE);
 }
 
+enum POINTER : std::uint64_t { REGULAR = 0x1 << 0, GLOBALPTR = 0x1 << 1, CACHEPTR = 0x1 << 2 };
+
 int pandoMain(int argc, char** argv) {
   auto thisPlace = pando::getCurrentPlace();
+  currentLocation = thisPlace.node.id;
+  nodeDims = pando::getNodeDims().id;
   if (thisPlace.node.id == 0) {
     galois::HostLocalStorageHeap::HeapInit();
     galois::PodLocalStorageHeap::HeapInit();
     optind = 0;
 
     std::uint64_t numAccesses = 0;
+    std::uint64_t ptrTypes = 0;
 
     int opt;
-    while ((opt = getopt(argc, argv, "n:s:f:dm")) != -1) {
+    while ((opt = getopt(argc, argv, "n:rgc")) != -1) {
       switch (opt) {
         case 'n':
           numAccesses = strtoull(optarg, nullptr, 10);
           break;
+        case 'r':
+          ptrTypes |= REGULAR;
+          break;
+        case 'g':
+          ptrTypes |= GLOBALPTR;
+          break;
+        case 'c':
+          ptrTypes |= CACHEPTR;
+          break;
         default:
           printUsageExit(argv[0]);
       }
     }
 
+    // Checks
+    if (numAccesses == 0) {
+      printUsageExit(argv[0]);
+    }
+    if (ptrTypes == 0) {
+      printUsageExit(argv[0]);
+    }
+
+    // Initialize Regular
     std::uint64_t* ptr = (std::uint64_t*)malloc(sizeof(std::uint64_t));
+
+    // Initialize GlobalPtr
     pando::GlobalPtr<std::uint64_t> gptr;
     pando::LocalStorageGuard<std::uint64_t> gptrGuard(gptr, 1);
 
+    // Initialize CachePtr
+    CachePtr<std::uint64_t> cptr;
+    pando::LocalStorageGuard<std::uint64_t> cptrGuard(cptr.ref.globalPtr, 1);
+    cptr.ref.cacheLoc = nodeDims;
+
     std::array<std::uint64_t, 16> simpleArr;
     for (std::uint64_t i = 0; i < simpleArr.size(); i++) {
       simpleArr[i] = i;
     }
 
-    *ptr = 0;
-    auto normalBegin = std::chrono::high_resolution_clock::now();
-    for (std::uint64_t i = 0; i < numAccesses; i++) {
-      *ptr += simpleArr[i % simpleArr.size()];
+    std::chrono::time_point<std::chrono::high_resolution_clock> normalBegin, normalEnd;
+    std::chrono::time_point<std::chrono::high_resolution_clock> gptrBegin, gptrEnd;
+    std::chrono::time_point<std::chrono::high_resolution_clock> cptrBegin, cptrEnd;
+    if (ptrTypes & REGULAR) {
+      *ptr = 0;
+      normalBegin = std::chrono::high_resolution_clock::now();
+      for (std::uint64_t i = 0; i < numAccesses; i++) {
+        *ptr += simpleArr[i % simpleArr.size()];
+      }
+      normalEnd = std::chrono::high_resolution_clock::now();
     }
-    auto normalEnd = std::chrono::high_resolution_clock::now();
-    *gptr = 0;
-    auto gptrBegin = std::chrono::high_resolution_clock::now();
-    for (std::uint64_t i = 0; i < numAccesses; i++) {
-      *gptr += simpleArr[i % simpleArr.size()];
+
+    if (ptrTypes & GLOBALPTR) {
+      *gptr = 0;
+      gptrBegin = std::chrono::high_resolution_clock::now();
+      for (std::uint64_t i = 0; i < numAccesses; i++) {
+        *gptr += simpleArr[i % simpleArr.size()];
+      }
+      gptrEnd = std::chrono::high_resolution_clock::now();
     }
 
-    auto gptrEnd = std::chrono::high_resolution_clock::now();
+    if (ptrTypes & CACHEPTR) {
+      *gptr = 0;
+      cptrBegin = std::chrono::high_resolution_clock::now();
+      for (std::uint64_t i = 0; i < numAccesses; i++) {
+        *cptr += simpleArr[i % simpleArr.size()];
+      }
+      cptrEnd = std::chrono::high_resolution_clock::now();
+    }
 
-    std::cout << "Normal Pointer took: "
-              << std::chrono::duration_cast<std::chrono::nanoseconds>(normalEnd - normalBegin)
-              << "\n Global Pointer took: "
-              << std::chrono::duration_cast<std::chrono::nanoseconds>(gptrEnd - gptrBegin)
-              << std::endl;
+    if (ptrTypes & REGULAR) {
+      std::cout << "Normal Pointer took: "
+                << std::chrono::duration_cast<std::chrono::nanoseconds>(normalEnd - normalBegin)
+                << "\n";
+    }
+    if (ptrTypes & GLOBALPTR) {
+      std::cout << "Global Pointer took: "
+                << std::chrono::duration_cast<std::chrono::nanoseconds>(gptrEnd - gptrBegin)
+                << "\n";
+    }
+    if (ptrTypes & CACHEPTR) {
+      std::cout << "Cache Pointer took: "
+                << std::chrono::duration_cast<std::chrono::nanoseconds>(cptrEnd - cptrBegin)
+                << "\n";
+    }
+    std::cout << std::flush;
   }
   pando::waitAll();
   return 0;

pando-rt/examples/gups.cpp

@@ -38,7 +38,7 @@ int pandoMain(int argc, char** argv) {
 
   const auto thisPlace = pando::getCurrentPlace();
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   const auto tableByteCount = tableSize * sizeof(std::int64_t);

pando-rt/examples/gups_no_tasks.cpp

@@ -39,7 +39,7 @@ int pandoMain(int argc, char** argv) {
   std::printf("Table size per node: %lu, updates / thread: %lu\n", tableSize, threadUpdates);
 
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   const auto tableByteCount = tableSize * sizeof(std::int64_t);

pando-rt/examples/helloworld.cpp

@@ -15,7 +15,7 @@ void greetings(int level) {
     PANDO_CHECK(pando::executeOn(otherPlace, &greetings, level + 1));
   }
 
-  std::printf("%s/%i: Hello from node %i, pod x=%i,y=%i, core x=%i,y=%i\n", __func__, level,
+  std::printf("%s/%i: Hello from node %li, pod x=%i,y=%i, core x=%i,y=%i\n", __func__, level,
               thisPlace.node.id, thisPlace.pod.x, thisPlace.pod.y, thisPlace.core.x,
               thisPlace.core.y);
 }
@@ -30,14 +30,14 @@ void nodeGreetings(int level) {
     PANDO_CHECK(
         pando::executeOn(pando::Place{rightNode, {}, pando::anyCore}, &nodeGreetings, level + 1));
   }
-  std::printf("%s/%i: Hello from node %i, pod x=%i,y=%i, core x=%i,y=%i\n", __func__, level,
+  std::printf("%s/%i: Hello from node %li, pod x=%i,y=%i, core x=%i,y=%i\n", __func__, level,
               thisPlace.node.id, thisPlace.pod.x, thisPlace.pod.y, thisPlace.core.x,
               thisPlace.core.y);
 }
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   const auto thisPlace = pando::getCurrentPlace();

pando-rt/examples/memory_resource.cpp

@@ -18,7 +18,7 @@ void allocate() {
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   const auto thisPlace = pando::getCurrentPlace();

pando-rt/examples/multi_pxns/allcores_reduce_values.cpp

@@ -186,7 +186,7 @@ void correctnessCheck(pando::GlobalPtr<pando::GlobalPtr<std::int64_t>> output,
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i).\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i).\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/multi_pxns/allcores_reducebroadcast_values.cpp

@@ -233,7 +233,7 @@ void correctnessCheck(pando::GlobalPtr<pando::GlobalPtr<std::int64_t>> output,
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i).\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i).\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/multi_pxns/onecore_perpxn_reduce_values.cpp

@@ -144,7 +144,7 @@ void correctnessCheck(pando::GlobalPtr<pando::GlobalPtr<std::int64_t>> output,
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i).\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i).\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/multi_pxns/onecore_perpxn_reducebroadcast_values.cpp

@@ -169,7 +169,7 @@ void correctnessCheck(pando::GlobalPtr<pando::GlobalPtr<std::int64_t>> output,
     for (std::int32_t c = 0; c < numCoresPerNode; ++c) {
       std::int16_t ownerPXN = (c % placeDims.node.id);
       if (localArray[c] != std::int64_t{ownerPXN * (placeDims.node.id - 1) + 1}) {
-        std::printf("%i %i %li\n", pando::getCurrentPlace().node.id, c, (int64_t)localArray[c]);
+        std::printf("%li %i %li\n", pando::getCurrentPlace().node.id, c, (int64_t)localArray[c]);
         check_correctness = false;
       }
     }
@@ -183,7 +183,7 @@ void correctnessCheck(pando::GlobalPtr<pando::GlobalPtr<std::int64_t>> output,
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i).\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i).\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/multi_pxns/onecore_scatter_diffdim_arrays_values.cpp

@@ -90,7 +90,7 @@ void waitUntil(pando::GlobalPtr<bool> dones, std::int16_t numNodes) {
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i).\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i).\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/multi_pxns/onecore_scatter_samedim_arrays_values.cpp

@@ -89,7 +89,7 @@ void waitUntil(pando::GlobalPtr<bool> dones, std::int16_t numNodes) {
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i).\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i).\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/single_pxn/onecore_increment_value.cpp

@@ -21,7 +21,7 @@ void increase(std::int64_t v) {
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
   const auto thisPlace = pando::getCurrentPlace();
 

pando-rt/examples/single_pxn/onecore_scatter_values.cpp

@@ -74,7 +74,7 @@ void correctnessCheck(pando::GlobalPtr<std::int64_t> output) {
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/single_pxn/pass_value_in_allcores.cpp

@@ -19,7 +19,7 @@ constexpr std::int64_t solution{5};
 void readValue(pando::GlobalPtr<std::int64_t> sharedValue, pando::GlobalPtr<bool> checkSolution) {
   if (*sharedValue == solution) {
     auto thisPlace = pando::getCurrentPlace();
-    std::printf("[node %i, pod x=%i,y=%i, core x=%i,y=%i] read value: %li\n", thisPlace.node.id,
+    std::printf("[node %li, pod x=%i,y=%i, core x=%i,y=%i] read value: %li\n", thisPlace.node.id,
                 thisPlace.pod.x, thisPlace.pod.y, thisPlace.core.x, thisPlace.core.y,
                 static_cast<std::int64_t>(*sharedValue));
     (*checkSolution) = true;
@@ -49,7 +49,7 @@ void incrementValue(pando::GlobalPtr<std::int64_t> sharedValue,
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/single_pxn/pass_value_in_twocores.cpp

@@ -18,7 +18,7 @@ constexpr std::int64_t solution{5};
 void readValue(pando::GlobalPtr<std::int64_t> sharedValue) {
   if (*sharedValue == solution) {
     auto thisPlace = pando::getCurrentPlace();
-    std::printf("[node %i, pod x=%i,y=%i, core x=%i,y=%i] read value: %li\n", thisPlace.node.id,
+    std::printf("[node %li, pod x=%i,y=%i, core x=%i,y=%i] read value: %li\n", thisPlace.node.id,
                 thisPlace.pod.x, thisPlace.pod.y, thisPlace.core.x, thisPlace.core.y,
                 static_cast<std::int64_t>(*sharedValue));
   }
@@ -46,7 +46,7 @@ void incrementValue(pando::GlobalPtr<std::int64_t> sharedValue) {
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/single_pxn/serial_tasks.cpp

@@ -20,7 +20,7 @@ void incrementValue(pando::GlobalPtr<std::int16_t> sharedValue,
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x == 0 || placeDims.core.y == 0) {

pando-rt/examples/single_pxn/twocores_scatter_values.cpp

@@ -97,7 +97,7 @@ void correctnessCheck(pando::GlobalPtr<std::int64_t> output) {
 
 int pandoMain(int, char**) {
   const auto placeDims = pando::getPlaceDims();
-  std::printf("Configuration (nodes, pods, cores): (%i), (%i,%i), (%i,%i)\n", placeDims.node.id,
+  std::printf("Configuration (nodes, pods, cores): (%li), (%i,%i), (%i,%i)\n", placeDims.node.id,
               placeDims.pod.x, placeDims.pod.y, placeDims.core.x, placeDims.core.y);
 
   if (placeDims.core.x * placeDims.core.y < 2) {