Merge pull request #506 from EoD/restore_old_cl_kernel

Restore old cl kernel v4

ethminer/MinerAux.h

@@ -292,6 +292,18 @@ class MinerCLI
 				BOOST_THROW_EXCEPTION(BadArgument());
 			}
 		}
+		else if (arg == "--cl-kernel" && i + 1 < argc)
+		{
+			try
+			{
+				m_openclSelectedKernel = stol(argv[++i]);
+			}
+			catch (...)
+			{
+				cerr << "Bad " << arg << " option: " << argv[i] << endl;
+				BOOST_THROW_EXCEPTION(BadArgument());
+			}
+		}
 #endif
 #if ETH_ETHASHCL || ETH_ETHASHCUDA
 		else if ((arg == "--cl-global-work" || arg == "--cuda-grid-size")  && i + 1 < argc)
@@ -501,7 +513,9 @@ class MinerCLI
 				m_miningThreads = m_openclDeviceCount;
 			}
 
+			CLMiner::setCLKernel(m_openclSelectedKernel);
 			CLMiner::setThreadsPerHash(m_openclThreadsPerHash);
+
 			if (!CLMiner::configureGPU(
 					m_localWorkSize,
 					m_globalWorkSizeMultiplier,
@@ -601,11 +615,17 @@ class MinerCLI
 			<< "        sequential  - load DAG on GPUs one after another. Use this when the miner crashes during DAG generation" << endl
 			<< "        single <n>  - generate DAG on device n, then copy to other devices" << endl
 #if ETH_ETHASHCL
+			<< " OpenCL configuration:" << endl
+			<< "    --cl-kernel <n>  Use a different OpenCL kernel (default: use stable kernel)" << endl
+			<< "        0: stable kernel" << endl
+			<< "        1: unstable kernel" << endl
+//			<< "        2: experimental kernel" << endl
 			<< "    --cl-local-work Set the OpenCL local work size. Default is " << CLMiner::c_defaultLocalWorkSize << endl
 			<< "    --cl-global-work Set the OpenCL global work size as a multiple of the local work size. Default is " << CLMiner::c_defaultGlobalWorkSizeMultiplier << " * " << CLMiner::c_defaultLocalWorkSize << endl
 			<< "    --cl-parallel-hash <1 2 ..8> Define how many threads to associate per hash. Default=8" << endl
 #endif
 #if ETH_ETHASHCUDA
+			<< " CUDA configuration:" << endl
 			<< "    --cuda-block-size Set the CUDA block work size. Default is " << toString(ethash_cuda_miner::c_defaultBlockSize) << endl
 			<< "    --cuda-grid-size Set the CUDA grid size. Default is " << toString(ethash_cuda_miner::c_defaultGridSize) << endl
 			<< "    --cuda-streams Set the number of CUDA streams. Default is " << toString(ethash_cuda_miner::c_defaultNumStreams) << endl
@@ -618,6 +638,7 @@ class MinerCLI
 			<< "    --cuda-parallel-hash <1 2 ..8> Define how many hashes to calculate in a kernel, can be scaled to achieve better performance. Default=4" << endl
 #endif
 #if API_CORE
+			<< " API core configuration:" << endl
 			<< "    --api-port Set the api port, the miner should listen to. Use 0 to disable. Default=0, use negative numbers to run in readonly mode. for example -3333." << endl
 #endif
 			;
@@ -1048,6 +1069,7 @@ class MinerCLI
 	unsigned m_miningThreads = UINT_MAX;
 	bool m_shouldListDevices = false;
 #if ETH_ETHASHCL
+	unsigned m_openclSelectedKernel = 0;  ///< A numeric value for the selected OpenCL kernel
 	unsigned m_openclDeviceCount = 0;
 	unsigned m_openclDevices[16];
 	unsigned m_openclThreadsPerHash = 8;

libethash-cl/CLMiner.cpp

@@ -5,7 +5,8 @@
 
 #include "CLMiner.h"
 #include <libethash/internal.h>
-#include "CLMiner_kernel.h"
+#include "CLMiner_kernel_stable.h"
+#include "CLMiner_kernel_unstable.h"
 
 using namespace dev;
 using namespace eth;
@@ -18,6 +19,7 @@ namespace eth
 unsigned CLMiner::s_workgroupSize = CLMiner::c_defaultLocalWorkSize;
 unsigned CLMiner::s_initialGlobalWorkSize = CLMiner::c_defaultGlobalWorkSizeMultiplier * CLMiner::c_defaultLocalWorkSize;
 unsigned CLMiner::s_threadsPerHash = 8;
+CLKernelName CLMiner::s_clKernelName = CLMiner::c_defaultKernelName;
 
 constexpr size_t c_maxSearchResults = 1;
 
@@ -632,10 +634,26 @@ bool CLMiner::init(const h256& seed)
 		uint32_t lightSize64 = (unsigned)(light->data().size() / sizeof(node));
 
 		// patch source code
-		// note: CLMiner_kernel is simply ethash_cl_miner_kernel.cl compiled
+		// note: The kernels here are simply compiled version of the respective .cl kernels
 		// into a byte array by bin2h.cmake. There is no need to load the file by hand in runtime
+		// See libethash-cl/CMakeLists.txt: add_custom_command()
 		// TODO: Just use C++ raw string literal.
-		string code(CLMiner_kernel, CLMiner_kernel + sizeof(CLMiner_kernel));
+		string code;
+
+		if ( s_clKernelName == CLKernelName::Unstable ) {
+			cllog << "OpenCL kernel: Unstable kernel";
+			code = string(CLMiner_kernel_unstable, CLMiner_kernel_unstable + sizeof(CLMiner_kernel_unstable));
+		}
+		else { //if(s_clKernelName == CLKernelName::Stable)
+			cllog << "OpenCL kernel: Stable kernel";
+
+			//CLMiner_kernel_stable.cl will do a #undef THREADS_PER_HASH
+			if(s_threadsPerHash != 8) {
+				cwarn << "The current stable OpenCL kernel only supports exactly 8 threads. Thread parameter will be ignored.";
+			}
+
+			code = string(CLMiner_kernel_stable, CLMiner_kernel_stable + sizeof(CLMiner_kernel_stable));
+		}
 		addDefinition(code, "GROUP_SIZE", m_workgroupSize);
 		addDefinition(code, "DAG_SIZE", dagSize128);
 		addDefinition(code, "LIGHT_SIZE", lightSize64);

libethash-cl/CLMiner.h

@@ -41,6 +41,11 @@ namespace dev
 namespace eth
 {
 
+enum CLKernelName {
+	Stable,
+	Unstable,
+};
+
 class CLMiner: public Miner
 {
 public:
@@ -50,6 +55,9 @@ class CLMiner: public Miner
 	/// Default value of the global work size as a multiplier of the local work size
 	static const unsigned c_defaultGlobalWorkSizeMultiplier = 8192;
 
+	/// Default value of the kernel is the original one
+	static const CLKernelName c_defaultKernelName = CLKernelName::Stable;
+
 	CLMiner(FarmFace& _farm, unsigned _index);
 	~CLMiner();
 
@@ -73,6 +81,7 @@ class CLMiner: public Miner
 			s_devices[i] = _devices[i];
 		}
 	}
+	static void setCLKernel(unsigned _clKernel) { s_clKernelName = _clKernel == 1 ? CLKernelName::Unstable : CLKernelName::Stable; }
 	HwMonitor hwmon() override;
 protected:
 	void kickOff() override;
@@ -98,6 +107,7 @@ class CLMiner: public Miner
 	static unsigned s_platformId;
 	static unsigned s_numInstances;
 	static unsigned s_threadsPerHash;
+	static CLKernelName s_clKernelName;
 	static int s_devices[16];
 
 	/// The local work size for the search