Add --nvidia option

include/lo2s/config.hpp

@@ -94,6 +94,8 @@ struct Config
     std::vector<int64_t> syscall_filter;
     // NEC SX-Aurora Tsubasa
     bool use_nec;
+    // Nvidia CUPTI
+    bool use_nvidia;
     std::chrono::microseconds nec_read_interval;
     std::chrono::milliseconds nec_check_interval;
 };

src/config.cpp

@@ -161,6 +161,7 @@ void parse_program_options(int argc, const char** argv)
     auto& sensors_options = parser.group("sensors options");
     auto& io_options = parser.group("I/O recording options");
     auto& nec_options = parser.group("NEC SX-Aurora Tsubasa recording options");
+    auto& nvidia_options = parser.group("Nvidia CUDA instrumentation options");
 
     lo2s::Config config;
 
@@ -344,6 +345,8 @@ void parse_program_options(int argc, const char** argv)
         .metavar("MSEC")
         .default_value("100");
 
+    nvidia_options.toggle("nvidia", "Enable Nvidia CUPTI instrumentation");
+
     nitro::options::arguments arguments;
     try
     {
@@ -370,6 +373,7 @@ void parse_program_options(int argc, const char** argv)
     config.use_sensors = arguments.given("sensors");
     config.use_block_io = arguments.given("block-io");
     config.use_nec = arguments.given("nec");
+    config.use_nvidia = arguments.given("nvidia");
     config.command = arguments.positionals();
 
     if (arguments.given("help"))

src/cupti/lib.cpp

@@ -32,7 +32,7 @@
 #include <cupti.h>
 extern "C"
 {
-  #include <time.h>
+#include <time.h>
 #include <unistd.h>
 }
 
@@ -52,7 +52,7 @@ static void CUPTIAPI bufferRequested(uint8_t** buffer, size_t* size, size_t* max
 
     if (*buffer == NULL)
     {
-      std::cerr << "Error: Out of memory.\n";
+        std::cerr << "Error: Out of memory.\n";
         exit(-1);
     }
 }
@@ -94,7 +94,7 @@ static void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t*
     cuptiActivityGetNumDroppedRecords(ctx, streamId, &dropped);
     if (dropped != 0)
     {
-      std::cerr << "Dropped %u activity records.\n", (unsigned int)dropped;
+        std::cerr << "Dropped %u activity records.\n", (unsigned int)dropped;
     }
 
     free(buffer);
@@ -159,10 +159,10 @@ void CUPTIAPI callbackHandler(void* userdata, CUpti_CallbackDomain domain, CUpti
 
 uint64_t timestampfunc()
 {
-  struct timespec ts;
-  clock_gettime(CLOCK_MONOTONIC_RAW,&ts);
-  uint64_t res = ts.tv_sec * 1000000000 + ts.tv_nsec;
-  return res;
+    struct timespec ts;
+    clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
+    uint64_t res = ts.tv_sec * 1000000000 + ts.tv_nsec;
+    return res;
 }
 
 extern "C" int InitializeInjection(void)

src/monitor/scope_monitor.cpp

@@ -70,7 +70,7 @@ ScopeMonitor::ScopeMonitor(ExecutionScope scope, MainMonitor& parent, bool enabl
         add_fd(userspace_counter_writer_->fd());
     }
 
-    if (is_process)
+    if (config().use_nvidia && is_process)
     {
         cupti_reader_ = std::make_unique<cupti::Reader>(parent.trace(), scope.as_process());
         add_fd(cupti_reader_->fd());