Resurrect GPU autodetection.

Fixes issue #1632.

Pull request #1633.

src/OpenCL.cpp

@@ -946,8 +946,7 @@ std::vector<size_t> OpenCL::get_sgemm_tuners(void) {
     return tuners;
 }
 
-void OpenCL::initialize(const int channels, const std::vector<int> & gpus,
-                        bool silent) {
+void OpenCL::initialize(const int channels, int gpu, bool silent) {
     std::vector<cl::Platform> platforms;
     try {
         cl::Platform::get(&platforms);
@@ -1022,8 +1021,7 @@ void OpenCL::initialize(const int channels, const std::vector<int> & gpus,
                 myprintf("Device score:  %d\n", this_score);
             }
 
-            bool preferred =
-                std::find(cbegin(gpus), cend(gpus), id) != cend(gpus);
+            bool preferred = (gpu == id);
 
             if ((this_score > best_score) || preferred) {
                 best_version = opencl_version;

src/OpenCL.h

@@ -176,8 +176,7 @@ class OpenCL {
     friend class OpenCL_Network;
     friend class Tuner;
 public:
-    void initialize(const int channels, const std::vector<int> & gpus,
-                    bool silent = false);
+    void initialize(const int channels, int gpu, bool silent = false);
     void ensure_context_initialized(OpenCLContext & opencl_context);
     std::string get_device_name();
 

src/OpenCLScheduler.cpp

@@ -26,8 +26,11 @@
 void OpenCLScheduler::initialize(const int channels) {
     // multi-gpu?
     auto gpus = cfg_gpus;
+
+    // an empty GPU list from the command line represents autodetect.
+    // put a minus one GPU index here.
     if (gpus.empty()) {
-        gpus = {0};
+        gpus = {-1};
     }
 
     auto silent{false};
@@ -42,7 +45,7 @@ void OpenCLScheduler::initialize(const int channels) {
     for (auto gpu : gpus) {
         auto opencl = std::make_unique<OpenCL>();
         auto net = std::make_unique<OpenCL_Network>(*opencl);
-        opencl->initialize(channels, {gpu}, silent);
+        opencl->initialize(channels, gpu, silent);
         m_opencl.push_back(std::move(opencl));
         m_networks.push_back(std::move(net));