Multi-GPU and genproclimit for maximizing GPU usage support

src/libstratum/ZcashStratum.cpp

@@ -298,9 +298,100 @@ void ZcashMiner::start()
     }
 
     minerThreads = new boost::thread_group();
-    for (int i = 0; i < nThreads; i++) {
-        minerThreads->create_thread(boost::bind(&ZcashMinerThread, this, nThreads, i, conf));
-    }
+
+	// If using GPU
+  	if(conf.useGPU) {
+
+	  conf.currentPlatform = 0;
+	  conf.currentDevice = conf.selGPU;
+
+	  std::vector<cl::Platform> platforms = cl_zogminer::getPlatforms();
+
+	  // use all available GPUs
+	  if(conf.allGPU) {
+
+	        int devicesFound = 0;
+	        unsigned numPlatforms = platforms.size();
+
+	        for(unsigned platform = 0; platform < numPlatforms; ++platform) {
+
+	          std::vector<cl::Device> devices = cl_zogminer::getDevices(platforms, platform);
+	          unsigned noDevices = devices.size();
+	          devicesFound += noDevices;
+	          for(unsigned device = 0; device < noDevices; ++device) {
+
+	              conf.currentPlatform = platform;
+	              conf.currentDevice = device;
+
+	              cl_ulong result;
+	              devices[device].getInfo(CL_DEVICE_GLOBAL_MEM_SIZE, &result);
+
+				  int maxThreads = nThreads;
+				  if (!conf.forceGenProcLimit) {
+				    if (result > 7500000000) {
+				      maxThreads = std::min(4, nThreads);
+				    } else if (result > 5500000000) {
+				      maxThreads = std::min(3, nThreads);
+				    } else if (result > 3500000000) {
+				      maxThreads = std::min(2, nThreads);
+				    } else {
+				      maxThreads = std::min(1, nThreads);
+				    }
+				  }
+
+	              LogPrintf("ZcashMiner GPU[%d][%d] MemLimit: %s nThreads: %d\n", platform, device, std::to_string(result), maxThreads);
+
+	            for (int i = 0; i < maxThreads; i++)
+					minerThreads->create_thread(boost::bind(&ZcashMinerThread, this, nThreads, i, conf));
+
+	          }
+	        }
+
+	        if (devicesFound <= 0) {
+	           LogPrintf("ZcashMiner ERROR, No OpenCL devices found!\n");
+	        }
+
+	    } else {
+
+	      // mine on specified GPU device
+	      std::vector<cl::Device> devices = cl_zogminer::getDevices(platforms, conf.currentPlatform);
+
+	      if (devices.size() > conf.currentDevice) {
+
+	        cl_ulong result;
+	        devices[conf.currentDevice].getInfo(CL_DEVICE_GLOBAL_MEM_SIZE, &result);
+
+	        int maxThreads = nThreads;
+	        if (!conf.forceGenProcLimit) {
+	          if (result > 7500000000) {
+	            maxThreads = std::min(4, nThreads);
+	          } else if (result > 5500000000) {
+	            maxThreads = std::min(3, nThreads);
+	          } else if (result > 3500000000) {
+	            maxThreads = std::min(2, nThreads);
+	          } else {
+	            maxThreads = std::min(1, nThreads);
+	          }
+	        }
+
+	        LogPrintf("ZcashMiner GPU[%d][%d] MemLimit: %s nThreads: %d\n", conf.currentPlatform, conf.currentDevice, std::to_string(result), maxThreads);
+
+	        for (int i = 0; i < maxThreads; i++)
+	          	minerThreads->create_thread(boost::bind(&ZcashMinerThread, this, nThreads, i, conf));
+
+	      } else {
+	         LogPrintf("ZcashMiner ERROR, No OpenCL devices found!\n");
+	      }
+
+	    }
+
+	  } else {
+
+		 for (int i = 0; i < nThreads; i++)
+			minerThreads->create_thread(boost::bind(&ZcashMinerThread, this, nThreads, i, conf));
+
+	  }
+
 }
 
 void ZcashMiner::stop()

src/libzogminer/cl_zogminer.h

@@ -80,6 +80,8 @@ class cl_zogminer
 	static unsigned getNumDevices(unsigned _platformId = 0);
 	static std::string platform_info(unsigned _platformId = 0, unsigned _deviceId = 0);
 	static void listDevices();
+	static std::vector<cl::Device> getDevices(std::vector<cl::Platform> const& _platforms, unsigned _platformId);
+	static std::vector<cl::Platform> getPlatforms();
 
 	// Currently just prints memory of the GPU
 	static bool configureGPU(
@@ -112,8 +114,6 @@ class cl_zogminer
   static const size_t z_collision_bit_length = z_n / (z_k + 1);
   static const eh_index z_N = 1 << (z_collision_bit_length + 1);
 
-	static std::vector<cl::Device> getDevices(std::vector<cl::Platform> const& _platforms, unsigned _platformId);
-	static std::vector<cl::Platform> getPlatforms();
 	int compare_indices32(uint32_t* a, uint32_t* b, size_t n_current_indices) {
 		for(size_t i = 0; i < n_current_indices; ++i, ++a, ++b) {
 		    if(*a < *b) {

src/libzogminer/gpuconfig.h

@@ -31,7 +31,11 @@ class GPUConfig {
 	//~GPUConfig();
 
 	bool useGPU;
-	int64_t selGPU;
+	unsigned selGPU;
+	bool allGPU;
+    bool forceGenProcLimit;
+	unsigned currentPlatform;
+	unsigned currentDevice;
 	unsigned globalWorkSize;
 	unsigned workgroupSize;
 

src/standaloneminer.cpp

@@ -244,8 +244,11 @@ int main(int argc, char* argv[])
 
 	GPUConfig conf;
 
-	conf.useGPU = GetBoolArg("-G", false);
+	conf.useGPU = GetBoolArg("-G", false) || GetBoolArg("-GPU", false);
 	conf.selGPU = GetArg("-S", 0);
+	conf.allGPU = GetBoolArg("-allgpu", false);
+	conf.forceGenProcLimit = GetBoolArg("-forcenolimit", false);
+	int nThreads = GetArg("-genproclimit", 1);
 	//std::cout << GPU << " " << selGPU << std::endl;
 
     // Zcash debugging
@@ -290,7 +293,7 @@ int main(int argc, char* argv[])
             return false;
         }
 
-        ZcashMiner miner(GetArg("-genproclimit", 1), conf);
+		ZcashMiner miner(GetArg("-genproclimit", 1), conf);
         ZcashStratumClient sc {
             &miner, host, port,
             GetArg("-user", "x"),
@@ -303,7 +306,7 @@ int main(int argc, char* argv[])
         });
 
         scSig = &sc;
-        signal(SIGINT, stratum_sigint_handler);
+        signal(SIGINT, stratum_sigint_handler);	
 
         while(sc.isRunning()) {
             MilliSleep(1000);