Fixes for multidevice/gcc.

- Allow devices to be cleared during benchmark definition.
- Fix various demangling bugs.

nvbench/benchmark_base.cu

@@ -30,7 +30,7 @@ std::unique_ptr<benchmark_base> benchmark_base::clone() const
   return std::move(result);
 }
 
-void benchmark_base::set_devices(std::vector<int> device_ids)
+benchmark_base &benchmark_base::set_devices(std::vector<int> device_ids)
 {
   std::vector<device_info> devices;
   devices.reserve(device_ids.size());
@@ -38,12 +38,12 @@ void benchmark_base::set_devices(std::vector<int> device_ids)
   {
     devices.emplace_back(dev_id);
   }
-  this->set_devices(std::move(devices));
+  return this->set_devices(std::move(devices));
 }
 
-void benchmark_base::add_device(int device_id)
+benchmark_base &benchmark_base::add_device(int device_id)
 {
-  this->add_device(device_info{device_id});
+  return this->add_device(device_info{device_id});
 }
 
 } // namespace nvbench

nvbench/benchmark_base.cuh

@@ -81,18 +81,26 @@ struct benchmark_base
     return *this;
   }
 
-  void set_devices(std::vector<int> device_ids);
+  benchmark_base &set_devices(std::vector<int> device_ids);
 
-  void set_devices(std::vector<nvbench::device_info> devices)
+  benchmark_base &set_devices(std::vector<nvbench::device_info> devices)
   {
     m_devices = std::move(devices);
+    return *this;
   }
 
-  void add_device(int device_id);
+  benchmark_base &clear_devices()
+  {
+    m_devices.clear();
+    return *this;
+  }
 
-  void add_device(nvbench::device_info device)
+  benchmark_base &add_device(int device_id);
+
+  benchmark_base &add_device(nvbench::device_info device)
   {
     m_devices.push_back(std::move(device));
+    return *this;
   }
 
   [[nodiscard]] const std::vector<nvbench::device_info> &get_devices() const

nvbench/runner.cuh

@@ -1,5 +1,7 @@
 #pragma once
 
+#include <nvbench/benchmark_base.cuh>
+
 #include <nvbench/detail/state_generator.cuh>
 
 #include <stdexcept>

nvbench/type_strings.cu

@@ -1,7 +1,5 @@
 #include <nvbench/type_strings.cuh>
 
-#include <cstdlib>
-#include <memory>
 #include <string>
 
 #if defined(__GNUC__) || defined(__clang__)
@@ -10,16 +8,28 @@
 
 #ifdef NVBENCH_CXXABI_DEMANGLE
 #include <cxxabi.h>
-#endif
+
+#include <cstdlib>
+#include <memory>
+
+namespace
+{
+struct free_wrapper
+{
+  void operator()(void* ptr) { std::free(ptr); }
+};
+} // end namespace
+
+#endif // NVBENCH_CXXABI_DEMANGLE
 
 namespace nvbench::detail
 {
 
-std::string nvbench::detail::demangle(const std::string &str)
+std::string demangle(const std::string &str)
 {
 #ifdef NVBENCH_CXXABI_DEMANGLE
-  std::unique_ptr<char, std::free> demangled =
-    abi::__cxx_demangle(str.c_str(), nullptr, nullptr, nullptr);
+  std::unique_ptr<char, free_wrapper> demangled{
+    abi::__cxa_demangle(str.c_str(), nullptr, nullptr, nullptr)};
   return std::string(demangled.get());
 #else
   return str;

testing/benchmark.cu

@@ -222,6 +222,7 @@ void test_run()
   // More exhaustive testing is in runner.cu. This just tests that the
   // runner is called.
   no_types_bench bench;
+  bench.set_devices(std::vector<int>{});
   ASSERT(bench.get_states().empty());
   bench.run();
   ASSERT(bench.get_states().size() == 1);
@@ -230,6 +231,7 @@ void test_run()
 void test_clone()
 {
   lots_of_types_bench bench;
+  bench.set_devices(std::vector<int>{});
   bench.set_type_axes_names({"Integer", "Float", "Other"});
   bench.add_string_axis("Strings", {"string a", "string b", "string c"});
   bench.add_int64_power_of_two_axis("I64 POT Axis", {10, 20});

testing/option_parser.cu

@@ -10,7 +10,7 @@
 //==============================================================================
 // Declare a couple benchmarks for testing:
 void DummyBench(nvbench::state &state) { state.skip("Skipping for testing."); }
-NVBENCH_CREATE(DummyBench);
+NVBENCH_CREATE(DummyBench).clear_devices();
 
 using Ts = nvbench::type_list<void, nvbench::int8_t, nvbench::uint8_t>;
 using Us = nvbench::type_list<bool, nvbench::float32_t, nvbench::float64_t>;
@@ -25,7 +25,8 @@ NVBENCH_CREATE_TEMPLATE(TestBench, NVBENCH_TYPE_AXES(Ts, Us))
   .add_int64_axis("Ints", {42})
   .add_int64_power_of_two_axis("PO2s", {3})
   .add_float64_axis("Floats", {3.14})
-  .add_string_axis("Strings", {"S1"});
+  .add_string_axis("Strings", {"S1"})
+  .clear_devices();
 //==============================================================================
 
 namespace

testing/runner.cu

@@ -69,6 +69,7 @@ void test_empty()
   using runner_type    = nvbench::runner<benchmark_type>;
 
   benchmark_type bench;
+  bench.set_devices(std::vector<int>{});
   runner_type runner{bench};
 
   runner.generate_states();
@@ -85,6 +86,7 @@ void test_non_types()
   using runner_type    = nvbench::runner<benchmark_type>;
 
   benchmark_type bench;
+  bench.set_devices(std::vector<int>{});
   bench.add_int64_axis("Int", {1, 2, 3});
   bench.add_float64_axis("Float", {11.0, 12.0, 13.0});
   bench.add_string_axis("String", {"One", "Two", "Three"});