Fix: do not use gpu order to set its id, instead use the same method as blender

src/com/sheepit/client/hardware/gpu/GPU.java

@@ -74,8 +74,8 @@ public class GPU {
 		return devices;
 	}
 	
-	public static GPUDevice getGPUDevice(String device_model) {
+	public static GPUDevice getGPUDevice(String deviceId) {
-		if (device_model == null) {
+		if (deviceId == null) {
 			return null;
 		}
 		
@@ -88,7 +88,7 @@ public class GPU {
 		}
 		
 		for (GPUDevice dev : devices) {
-			if (device_model.equals(dev.getId()) || device_model.equals(dev.getModel())) {
+			if (deviceId.equals(dev.getId()) || deviceId.equals(dev.getOldId())) {
 				return dev;
 			}
 		}

src/com/sheepit/client/hardware/gpu/GPUDevice.java

@@ -26,6 +26,8 @@ public class GPUDevice {
 	
 	private String id;
 	
+	private String oldId; // for backward compatibility
+	
 	public GPUDevice(String type, String model, long ram, String id) {
 		this.type = type;
 		this.model = model;
@@ -65,6 +67,14 @@ public class GPUDevice {
 		this.id = id;
 	}
 	
+	public String getOldId() {
+		return oldId;
+	}
+	
+	public void setOldId(String id) {
+		this.oldId = id;
+	}
+	
 	@Override
 	public String toString() {
 		return "GPUDevice [type=" + type + ", model='" + model + "', memory=" + memory + ", id=" + id + "]";

src/com/sheepit/client/hardware/gpu/nvidia/Nvidia.java

@@ -66,7 +66,7 @@ public class Nvidia implements GPULister {
 		
 		List<GPUDevice> devices = new LinkedList<GPUDevice>();
 		
-		HashMap<Integer, GPUDevice> devicesWithPciId = new HashMap<Integer, GPUDevice>(count.getValue());
+		HashMap<String, GPUDevice> devicesWithPciId = new HashMap<String, GPUDevice>(count.getValue());
 		for (int num = 0; num < count.getValue(); num++) {
 			IntByReference aDevice = new IntByReference();
 			
@@ -76,12 +76,24 @@ public class Nvidia implements GPULister {
 				continue;
 			}
 			
+			IntByReference pciDomainId = new IntByReference();
 			IntByReference pciBusId = new IntByReference();
+			IntByReference pciDeviceId = new IntByReference();
+			result =  cudalib.cuDeviceGetAttribute(pciDomainId, CUDeviceAttribute.CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, aDevice.getValue());
+			if (result != CUresult.CUDA_SUCCESS) {
+				System.out.println("Nvidia::getGpus cuDeviceGetAttribute for CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID failed (ret: " + CUresult.stringFor(result) + ")");
+				continue;
+			}
 			result =  cudalib.cuDeviceGetAttribute(pciBusId, CUDeviceAttribute.CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, aDevice.getValue());
 			if (result != CUresult.CUDA_SUCCESS) {
 				System.out.println("Nvidia::getGpus cuDeviceGetAttribute for CU_DEVICE_ATTRIBUTE_PCI_BUS_ID failed (ret: " + CUresult.stringFor(result) + ")");
 				continue;
 			}
+			result =  cudalib.cuDeviceGetAttribute(pciDeviceId, CUDeviceAttribute.CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, aDevice.getValue());
+			if (result != CUresult.CUDA_SUCCESS) {
+				System.out.println("Nvidia::getGpus cuDeviceGetAttribute for CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID failed (ret: " + CUresult.stringFor(result) + ")");
+				continue;
+			}
 			
 			byte name[] = new byte[256];
 			
@@ -105,18 +117,24 @@ public class Nvidia implements GPULister {
 				return null;
 			}
 			
-			devicesWithPciId.put(pciBusId.getValue(), new GPUDevice(TYPE, new String(name).trim(), ram.getValue(), "FAKE"));
+			String blenderId = String.format("CUDA_%s_%04x:%02x:%02x",
+					new String(name).trim(),
+					pciDomainId.getValue(),
+					pciBusId.getValue(),
+					pciDeviceId.getValue());
+			devicesWithPciId.put(Integer.toString(pciBusId.getValue()), new GPUDevice(TYPE, new String(name).trim(), ram.getValue(), blenderId));
 		}
 		
-		// generate proper cuda id
+		// for backward compatibility generate a CUDA_N id
 		// in theory a set to environment "CUDA_DEVICE_ORDER=PCI_BUS_ID" should be enough but it didn't work
 		int i = 0;
-		for (Map.Entry<Integer, GPUDevice> entry : devicesWithPciId.entrySet()){
+		for (Map.Entry<String, GPUDevice> entry : devicesWithPciId.entrySet()){
 			GPUDevice aDevice = entry.getValue();
-			aDevice.setId(TYPE + "_" + Integer.toString(i));
+			aDevice.setOldId(TYPE + "_" + Integer.toString(i));
 			devices.add(aDevice);
 			i++;
 		}
+		
 		return devices;
 	}
 	

src/com/sheepit/client/hardware/gpu/opencl/OpenCL.java

@@ -105,13 +105,12 @@ public class OpenCL implements GPULister {
 			}
 			
 			for (int j = 0; j < device_count.getValue(); j++) {
-				String platform_vendor = getInfoPlatform(lib, plateforms[i], OpenCLLib.CL_PLATFORM_VENDOR);
+				String name = getInfodeviceString(lib, devices[j], OpenCLLib.CL_DEVICE_BOARD_NAME_AMD);
-				if (platform_vendor != null && platform_vendor.toLowerCase().equals("advanced micro devices, inc.")) { // opencl is only used for amd gpus
+				long vram = getInfodeviceLong(lib, devices[j], OpenCLLib.CL_DEVICE_GLOBAL_MEM_SIZE);
-					String name = getInfodeviceString(lib, devices[j], OpenCLLib.CL_DEVICE_BOARD_NAME_AMD);
+				if (name != null && vram > 0) {
-					long vram = getInfodeviceLong(lib, devices[j], OpenCLLib.CL_DEVICE_GLOBAL_MEM_SIZE);
+					GPUDevice gpu = new GPUDevice(TYPE, new String(name).trim(), vram, getBlenderId(lib, devices[j]));
-					if (name != null && vram > 0) {
+					gpu.setOldId(TYPE + "_" + id);
-						available_devices.add(new GPUDevice(TYPE, name, vram, TYPE + "_" + id));
+					available_devices.add(gpu);
-					}
 				}
 				id++;
 			}
@@ -158,4 +157,16 @@ public class OpenCL implements GPULister {
 		
 		return new String(name).trim();
 	}
+	
+	private static String getBlenderId(OpenCLLib lib, CLDeviceId.ByReference device) {
+		byte topology[] = new byte[24];
+		
+		int status = lib.clGetDeviceInfo(device, 0x4037, 24, topology, null);
+		if (status != OpenCLLib.CL_SUCCESS) {
+			System.out.println("OpenCL::getBlenderId failed(I) status: " + status);
+			return "";
+		}
+		
+		return String.format("%02x:%02x.%01x", topology[21], topology[22], topology[23]);
+	}
 }

src/com/sheepit/client/standalone/Worker.java

@@ -154,26 +154,17 @@ public class Worker {
 		
 		if (gpu_device != null) {
 			if (gpu_device.startsWith(Nvidia.TYPE) == false && gpu_device.startsWith(OpenCL.TYPE) == false) {
-				System.err.println("CUDA_DEVICE should look like '" + Nvidia.TYPE + "_X' or '" + OpenCL.TYPE + "_X' where X is a number");
+				System.err.println("GPU_ID should look like '" + Nvidia.TYPE + "_X' or '" + OpenCL.TYPE + "_X' more info on gpus available with --show-gpu");
-				return;
+				System.exit(2);
 			}
 			String family = "";
-			try {
+			if (gpu_device.startsWith(Nvidia.TYPE) == false && gpu_device.startsWith(OpenCL.TYPE) == false) {
-				if (gpu_device.startsWith(Nvidia.TYPE)) {
+				System.err.println("GPU_ID should look like '" + Nvidia.TYPE + "_X' or '" + OpenCL.TYPE + "_X' more info on gpus available with --show-gpu");
-					family = Nvidia.TYPE;
-				}
-				else if (gpu_device.startsWith(OpenCL.TYPE)) {
-					family = OpenCL.TYPE;
-				}
-				Integer.parseInt(gpu_device.substring(family.length() + 1)); // for the _
-			}
-			catch (NumberFormatException en) {
-				System.err.println("Gpu device code should look like '" + family + "_X' where X is a number");
 				return;
 			}
 			GPUDevice gpu = GPU.getGPUDevice(gpu_device);
 			if (gpu == null) {
-				System.err.println("GPU unknown");
+				System.err.println("GPU unknown, list of available gpus can be display with --show-gpu");
 				System.exit(2);
 			}
 			config.setUseGPU(gpu);