Fixed CUPTI on Windows (#830)

Summary: Fixed 2 compiling errors on Windows. Changed callbackLock_ from static to member of the class since on Windows Release in Pytorch CuptiCallbackApi::initCallbackApi gets called before the static initialisation of the mutex so it crashes. Pull Request resolved: #830 Reviewed By: aaronenyeshi Differential Revision: D51314940 Pulled By: davidberard98 fbshipit-source-id: 4de09332320bb15063859be2e43a9eb1e308c6fb
pytorch · Nov 15, 2023 · 33fdaad · 33fdaad
1 parent b885c64
commit 33fdaad
Show file tree

Hide file tree

Showing 4 changed files with 15 additions and 17 deletions.
diff --git a/libkineto/src/CuptiCallbackApi.cpp b/libkineto/src/CuptiCallbackApi.cpp
@@ -15,7 +15,6 @@
 #include <chrono>
 #include <algorithm>
 #include <mutex>
-#include <shared_mutex>
 
 #ifdef HAS_CUPTI
 #include "cupti_call.h"
@@ -32,13 +31,6 @@ constexpr size_t MAX_CB_FNS_PER_CB = 8;
 // is enabled, not a specific cbid.
 constexpr uint32_t MAX_CUPTI_CALLBACK_ID_ALL = 0xffffffff;
 
-// Reader Writer lock types
-using ReaderWriterLock = std::shared_timed_mutex;
-using ReaderLockGuard = std::shared_lock<ReaderWriterLock>;
-using WriteLockGuard = std::unique_lock<ReaderWriterLock>;
-
-static ReaderWriterLock callbackLock_;
-
 /* Callback Table :
  *  Overall goal of the design is to optimize the lookup of function
  *  pointers. The table is structured at two levels and the leaf

diff --git a/libkineto/src/CuptiCallbackApi.h b/libkineto/src/CuptiCallbackApi.h
@@ -16,6 +16,7 @@
 #include <list>
 #include <memory>
 #include <mutex>
+#include <shared_mutex>
 #include <set>
 
 // TODO(T90238193)
@@ -143,6 +144,12 @@ class CuptiCallbackApi {
   // As an implementation detail, cbid == 0xffffffff means enable the domain.
   std::set<std::pair<CUpti_CallbackDomain, CUpti_CallbackId>> enabledCallbacks_;
 
+
+  // Reader Writer lock types
+  using ReaderWriterLock = std::shared_timed_mutex;
+  using ReaderLockGuard = std::shared_lock<ReaderWriterLock>;
+  using WriteLockGuard = std::unique_lock<ReaderWriterLock>;
+  ReaderWriterLock callbackLock_;
 #ifdef HAS_CUPTI
   CUptiResult lastCuptiStatus_;
   CUpti_SubscriberHandle subscriber_ {0};

diff --git a/libkineto/src/CuptiRangeProfiler.cpp b/libkineto/src/CuptiRangeProfiler.cpp
@@ -109,14 +109,14 @@ CuptiRangeProfilerSession::CuptiRangeProfilerSession(
     LOG(INFO) << "\t" << m;
   }
 
-  CuptiRangeProfilerOptions opts{
-    .metricNames = cupti_metrics,
-    .deviceId = 0,
-    .maxRanges = max_ranges,
-    .numNestingLevels = 1,
-    .cuContext = nullptr,
-    .unitTest = false};
-
+  CuptiRangeProfilerOptions opts;
+  opts.metricNames = cupti_metrics;
+  opts.deviceId = 0;
+  opts.maxRanges = max_ranges;
+  opts.numNestingLevels = 1;
+  opts.cuContext = nullptr;
+  opts.unitTest = false;
+  
   for (auto device_id : CuptiRBProfilerSession::getActiveDevices()) {
     LOG(INFO) << "Init CUPTI range profiler on gpu = " << device_id
               << " max ranges = " << max_ranges;

diff --git a/libkineto/src/CuptiRangeProfilerConfig.cpp b/libkineto/src/CuptiRangeProfilerConfig.cpp
@@ -10,7 +10,6 @@
 #include <Logger.h>
 
 #include <stdlib.h>
-#include <unistd.h>
 
 #include <fmt/format.h>
 #include <ostream>