OpenCL tweaks and cleanup

- Reorganize kernels folder
- Adjust default opencl global work size
- Always use single search iteration
- Add --cl-nobinary. Bypass binary kernel load attempt.
- Remove unused isa kernel files

Author: jean-m-cyr

ethminer/main.cpp

@@ -331,7 +331,8 @@ class MinerCLI
 
 #if ETH_ETHASHCL
 
-		app.add_option("--cl-kernel", m_clKernel,
+		int clKernel = -1;
+		app.add_option("--cl-kernel", clKernel,
 			"Ignored parameter. Kernel is auto-selected.", true)
 			->group(OpenCLGroup)
 			->check(CLI::Range(2));
@@ -345,15 +346,11 @@ class MinerCLI
 			"Select list of devices to mine on (default: use all available)")
 			->group(OpenCLGroup);
 
-		app.add_set("--cl-parallel-hash", m_openclThreadsPerHash, {1, 2, 4, 8},
+		int openclThreadsPerHash = -1;
+		app.add_set("--cl-parallel-hash", openclThreadsPerHash, {1, 2, 4, 8},
 			"ignored parameter", true)
 			->group(OpenCLGroup);
 
-		app.add_option("--cl-iterations", m_openclIterations,
-			"Number of outer iterations to perform before enqeueing on a new nonce", true)
-			->group(OpenCLGroup)
-			->check(CLI::Range(1,99999));
-
 		app.add_option("--cl-global-work", m_globalWorkSizeMultiplier,
 			"Set the global work size multipler.", true)
 			->group(OpenCLGroup);
@@ -362,6 +359,9 @@ class MinerCLI
 			"Set the local work size", true)
 			->group(OpenCLGroup);
 
+		app.add_flag("--cl-nobinary", m_noBinary,
+			"Don't attempt to load binary kernel")
+			->group(OpenCLGroup);
 #endif
 
 #if ETH_ETHASHCUDA
@@ -498,6 +498,13 @@ class MinerCLI
 			exit(-1);
     	}
 
+#if ETH_ETHASHCL
+		if (clKernel >= 0)
+			clog << "--cl-kernel ignored. Kernel is auto-selected\n";
+		if (openclThreadsPerHash >= 0)
+			clog << "--cl-parallel-hash ignored. No longer applies\n";
+#endif
+
 		if (hwmon_opt->count()) {
 			m_show_hwmonitors = true;
 			if (hwmon)
@@ -621,9 +628,6 @@ class MinerCLI
 				m_miningThreads = m_openclDeviceCount;
 			}
 
-			CLMiner::setNumberIterations(m_openclIterations);
-			CLMiner::setThreadsPerHash(m_openclThreadsPerHash);
-
 			if (!CLMiner::configureGPU(
 				m_localWorkSize,
 				m_globalWorkSizeMultiplier,
@@ -633,7 +637,7 @@ class MinerCLI
 				m_dagCreateDevice,
 				m_noEval,
 				m_exit,
-				m_clKernel
+				m_noBinary
 			)) {
 				stop_io_service();
 				exit(1);
@@ -896,13 +900,11 @@ class MinerCLI
 	unsigned m_miningThreads = UINT_MAX;
 	bool m_shouldListDevices = false;
 #if ETH_ETHASHCL
-	unsigned m_openclIterations = 1;  ///< A numeric value for the number of iterations
 	unsigned m_openclDeviceCount = 0;
 	vector<unsigned> m_openclDevices;
-	int m_openclThreadsPerHash = -1;
 	unsigned m_globalWorkSizeMultiplier = CLMiner::c_defaultGlobalWorkSizeMultiplier;
 	unsigned m_localWorkSize = CLMiner::c_defaultLocalWorkSize;
-	int m_clKernel = -1;
+	bool m_noBinary = false;
 #endif
 #if ETH_ETHASHCUDA
 	unsigned m_cudaDeviceCount = 0;

libethash-cl/CLMiner.cpp

@@ -19,8 +19,6 @@ namespace eth
 
 unsigned CLMiner::s_workgroupSize = CLMiner::c_defaultLocalWorkSize;
 unsigned CLMiner::s_initialGlobalWorkSize = CLMiner::c_defaultGlobalWorkSizeMultiplier * CLMiner::c_defaultLocalWorkSize;
-int CLMiner::s_threadsPerHash = -1;
-unsigned CLMiner::s_kernelIterations = 1;
 
 constexpr size_t c_maxSearchResults = 255;
 
@@ -256,6 +254,7 @@ std::vector<cl::Device> getDevices(std::vector<cl::Platform> const& _platforms,
 unsigned CLMiner::s_platformId = 0;
 unsigned CLMiner::s_numInstances = 0;
 vector<int> CLMiner::s_devices(MAX_MINERS, -1);
+bool CLMiner::s_noBinary = false;
 
 CLMiner::CLMiner(FarmFace& _farm, unsigned _index):
     Miner("cl-", _farm, _index)
@@ -449,19 +448,16 @@ void CLMiner::listDevices()
 
 bool CLMiner::configureGPU(unsigned _localWorkSize, unsigned _globalWorkSizeMultiplier,
     unsigned _platformId, int epoch, unsigned _dagLoadMode, unsigned _dagCreateDevice,
-	bool _noeval, bool _exit, int _kernel)
+	bool _noeval, bool _exit, bool _nobinary)
 {
 	s_noeval = _noeval;
     s_dagLoadMode = _dagLoadMode;
     s_dagCreateDevice = _dagCreateDevice;
     s_exit = _exit;
+	s_noBinary = _nobinary;
 
 	if (_noeval)
 		cwarn << "--no-eval not yet supported for AMD.";
-	if (_kernel >= 0)
-		cwarn << "--cl-kernel ignored. Auto-selecting kernel";
-	if (s_threadsPerHash >= 0)
-		cwarn << "--cl-parallel-hash ignored.";
 
     s_platformId = _platformId;
 
@@ -653,7 +649,7 @@ bool CLMiner::init(int epoch)
            the default kernel if loading fails for whatever reason */
 		bool loadedBinary = false;
 
-        {
+        if (!s_noBinary) {
             std::ifstream kernel_file;
             vector<unsigned char> bin_data;
             std::stringstream fname_strm;
@@ -750,7 +746,7 @@ bool CLMiner::init(int epoch)
         m_searchKernel.setArg(2, m_dag[0]);
         m_searchKernel.setArg(3, m_dagItems);
         m_searchKernel.setArg(6, ~0u);
-        m_searchKernel.setArg(7, uint32_t(s_kernelIterations));  // Number of iterations
+        m_searchKernel.setArg(7, 1u);  // Number of iterations
 
         // create mining buffers
         ETHCL_LOG("Creating mining buffer");

libethash-cl/CLMiner.h

@@ -50,7 +50,7 @@ class CLMiner: public Miner
 	/// Default value of the local work size. Also known as workgroup size.
 	static const unsigned c_defaultLocalWorkSize = 192;
 	/// Default value of the global work size as a multiplier of the local work size
-	static const unsigned c_defaultGlobalWorkSizeMultiplier = 18000;
+	static const unsigned c_defaultGlobalWorkSizeMultiplier = 16392;
 
 	CLMiner(FarmFace& _farm, unsigned _index);
 	~CLMiner() override;
@@ -60,17 +60,15 @@ class CLMiner: public Miner
 	static void listDevices();
     static bool configureGPU(unsigned _localWorkSize, unsigned _globalWorkSizeMultiplier,
         unsigned _platformId, int epoch, unsigned _dagLoadMode, unsigned _dagCreateDevice,
-        bool _noeval, bool _exit, int _kernel);
+        bool _noeval, bool _exit, bool _nobinary);
     static void setNumInstances(unsigned _instances) { s_numInstances = std::min<unsigned>(_instances, getNumDevices()); }
-	static void setThreadsPerHash(int _threadsPerHash){s_threadsPerHash = _threadsPerHash; }
 	static void setDevices(const vector<unsigned>& _devices, unsigned _selectedDeviceCount)
 	{
 		for (unsigned i = 0; i < _selectedDeviceCount; i++)
 		{
 			s_devices[i] = _devices[i];
 		}
 	}
-    static void setNumberIterations(unsigned _iterations) {s_kernelIterations = _iterations <= 1 ? 1 : _iterations;}
 protected:
 	void kick_miner() override;
 
@@ -94,8 +92,7 @@ class CLMiner: public Miner
 
 	static unsigned s_platformId;
 	static unsigned s_numInstances;
-	static int s_threadsPerHash;
-    static unsigned s_kernelIterations;
+	static bool s_noBinary;
 	static vector<int> s_devices;
 
 	/// The local work size for the search

libethash-cl/CMakeLists.txt

@@ -6,14 +6,14 @@
 add_custom_command(
 	OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ethash.h
 	COMMAND ${CMAKE_COMMAND} ARGS
-	-DBIN2H_SOURCE_FILE="${CMAKE_CURRENT_SOURCE_DIR}/kernels/ethash.cl"
+	-DBIN2H_SOURCE_FILE="${CMAKE_CURRENT_SOURCE_DIR}/kernels/cl/ethash.cl"
 	-DBIN2H_VARIABLE_NAME=ethash_cl
 	-DBIN2H_HEADER_FILE="${CMAKE_CURRENT_BINARY_DIR}/ethash.h"
 	-P "${CMAKE_CURRENT_SOURCE_DIR}/bin2h.cmake"
 	COMMENT "Generating OpenCL Kernel Byte Array"
-	DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ethash.cl
+	DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/kernels/cl/ethash.cl
 )
-add_custom_target(clbin2h_stable DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/ethash.h ${CMAKE_CURRENT_SOURCE_DIR}/ethash.cl)
+add_custom_target(clbin2h_stable DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/ethash.h ${CMAKE_CURRENT_SOURCE_DIR}/cl/ethash.cl)
 
 set(SOURCES
 	CLMiner.h CLMiner.cpp

libethash-cl/kernels/Makefile

@@ -1,4 +1,4 @@
-SDIR=./src
+SDIR=./isa
 BUILD=./bin
 
 ASFLAGS=-I$(SDIR)