[incubator-mxnet] branch master updated: Resolve the conflict with PR#20499 (#20887)

[zh...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

zhasheng pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git


The following commit(s) were added to refs/heads/master by this push:
     new f03fb23  Resolve the conflict with PR#20499 (#20887)
f03fb23 is described below

commit f03fb23f1d103fec9541b5ae59ee06b1734a51d9
Author: Guanxin Li <38...@users.noreply.github.com>
AuthorDate: Thu Feb 10 17:16:28 2022 -0800

    Resolve the conflict with PR#20499 (#20887)
---
 src/profiler/storage_profiler.cc | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)

diff --git a/src/profiler/storage_profiler.cc b/src/profiler/storage_profiler.cc
index a8846a2..237cdd9 100644
--- a/src/profiler/storage_profiler.cc
+++ b/src/profiler/storage_profiler.cc
@@ -26,6 +26,7 @@
 #include <regex>
 #include <unordered_map>
 #include <vector>
+#include <type_traits>
 #include "./profiler.h"
 #include "../common/utils.h"
 #include "../common/cuda/utils.h"
@@ -45,6 +46,19 @@ GpuDeviceStorageProfiler* GpuDeviceStorageProfiler::Get() {
   return gpu_dev_storage_profiler.get();
 }
 
+#if MXNET_USE_NVML
+// Deduce the possibly versioned variant of nvmlProcessInfo_t* expected
+// as the 3rd arg of nvmlDeviceGetComputeRunningProcesses().
+template <typename F>
+struct GetArgType;
+template <typename R, typename T1, typename T2, typename T3>
+struct GetArgType<R (*)(T1, T2, T3)> {
+  typedef T3 arg3_t;
+};
+using NvmlProcessInfoPtr = GetArgType<decltype(&nvmlDeviceGetComputeRunningProcesses)>::arg3_t;
+using NvmlProcessInfo    = std::remove_pointer_t<NvmlProcessInfoPtr>;
+#endif
+
 void GpuDeviceStorageProfiler::DumpProfile() const {
   size_t current_pid = common::current_process_id();
   std::ofstream fout((filename_prefix_ + "-pid_" + std::to_string(current_pid) + ".csv").c_str());
@@ -97,23 +111,17 @@ void GpuDeviceStorageProfiler::DumpProfile() const {
   // If NVML has been enabled, add amend term to the GPU memory profile.
   nvmlDevice_t nvml_device;
 
-#if NVML_API_VERSION < 11
-  typedef std::vector<nvmlProcessInfo_t> ProcessInfoVector;
-#else
-  typedef std::vector<nvmlProcessInfo_v1_t> ProcessInfoVector;
-#endif
-
   NVML_CALL(nvmlInit());
   for (std::pair<const int, size_t>& dev_id_total_alloc_pair : gpu_dev_id_total_alloc_map) {
     unsigned info_count = 0;
-    ProcessInfoVector infos(info_count);
+    std::vector<NvmlProcessInfo> infos(info_count);
 
     NVML_CALL(nvmlDeviceGetHandleByIndex(dev_id_total_alloc_pair.first, &nvml_device));
     // The first call to `nvmlDeviceGetComputeRunningProcesses` is to set the
     // size of info. Since `NVML_ERROR_INSUFFICIENT_SIZE` will always be
     // returned, we do not wrap the function call with `NVML_CALL`.
     nvmlDeviceGetComputeRunningProcesses(nvml_device, &info_count, infos.data());
-    infos = ProcessInfoVector(info_count);
+    infos.resize(info_count);
     NVML_CALL(nvmlDeviceGetComputeRunningProcesses(nvml_device, &info_count, infos.data()));
 
     bool amend_made = false;