Added config option for AVX2 dataset init

-1 = Auto detect
0 = Always disabled
1 = Enabled if AVX2 is supported

src/config.json

@@ -16,6 +16,7 @@
     "title": true,
     "randomx": {
         "init": -1,
+        "init-avx2": -1,
         "mode": "auto",
         "1gb-pages": false,
         "rdmsr": true,

src/core/config/Config_default.h

@@ -50,6 +50,7 @@ R"===(
     "colors": true,
     "randomx": {
         "init": -1,
+        "init-avx2": -1,
         "mode": "auto",
         "1gb-pages": false,
         "rdmsr": true,

src/crypto/randomx/jit_compiler_a64.cpp

@@ -36,12 +36,18 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "crypto/randomx/virtual_memory.hpp"
 
 static bool hugePagesJIT = false;
+static int optimizedDatasetInit = -1;
 
 void randomx_set_huge_pages_jit(bool hugePages)
 {
 	hugePagesJIT = hugePages;
 }
 
+void randomx_set_optimized_dataset_init(int value)
+{
+	optimizedDatasetInit = value;
+}
+
 namespace ARMV8A {
 
 constexpr uint32_t B           = 0x14000000;
@@ -98,7 +104,7 @@ static size_t CalcDatasetItemSize()
 
 constexpr uint32_t IntRegMap[8] = { 4, 5, 6, 7, 12, 13, 14, 15 };
 
-JitCompilerA64::JitCompilerA64(bool hugePagesEnable) :
+JitCompilerA64::JitCompilerA64(bool hugePagesEnable, bool) :
 	hugePages(hugePagesJIT && hugePagesEnable),
 	literalPos(ImulRcpLiteralsEnd)
 {

src/crypto/randomx/jit_compiler_a64.hpp

@@ -47,7 +47,7 @@ namespace randomx {
 
 	class JitCompilerA64 {
 	public:
-		explicit JitCompilerA64(bool hugePagesEnable);
+		explicit JitCompilerA64(bool hugePagesEnable, bool optimizedInitDatasetEnable);
 		~JitCompilerA64();
 
 		void prepare() {}

src/crypto/randomx/jit_compiler_fallback.cpp

@@ -35,3 +35,6 @@ void randomx_set_huge_pages_jit(bool)
 {
 }
 
+void randomx_set_optimized_dataset_init(int)
+{
+}

src/crypto/randomx/jit_compiler_fallback.hpp

@@ -43,7 +43,7 @@ namespace randomx {
 
 	class JitCompilerFallback {
 	public:
-		explicit JitCompilerFallback(bool) {
+		explicit JitCompilerFallback(bool, bool) {
 			throw std::runtime_error("JIT compilation is not supported on this platform");
 		}
 		void prepare() {}

src/crypto/randomx/jit_compiler_x86.cpp

@@ -52,12 +52,18 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #endif
 
 static bool hugePagesJIT = false;
+static int optimizedDatasetInit = -1;
 
 void randomx_set_huge_pages_jit(bool hugePages)
 {
 	hugePagesJIT = hugePages;
 }
 
+void randomx_set_optimized_dataset_init(int value)
+{
+	optimizedDatasetInit = value;
+}
+
 namespace randomx {
 	/*
 
@@ -209,41 +215,52 @@ namespace randomx {
 	static std::atomic<size_t> codeOffset;
 	constexpr size_t codeOffsetIncrement = 59 * 64;
 
-	JitCompilerX86::JitCompilerX86(bool hugePagesEnable) {
+	JitCompilerX86::JitCompilerX86(bool hugePagesEnable, bool optimizedInitDatasetEnable) {
 		BranchesWithin32B = xmrig::Cpu::info()->jccErratum();
 
 		hasAVX = xmrig::Cpu::info()->hasAVX();
 		hasAVX2 = xmrig::Cpu::info()->hasAVX2();
 
-		// Set to false by default
+		// Disable by default
 		initDatasetAVX2 = false;
 
-		xmrig::ICpuInfo::Vendor vendor = xmrig::Cpu::info()->vendor();
-		xmrig::ICpuInfo::Arch arch = xmrig::Cpu::info()->arch();
-
-		if (vendor == xmrig::ICpuInfo::VENDOR_INTEL) {
-			// AVX2 init is faster on Intel CPUs without HT
-			initDatasetAVX2 = xmrig::Cpu::info()->cores() == xmrig::Cpu::info()->threads();
-		}
-		else if (vendor == xmrig::ICpuInfo::VENDOR_AMD) {
-			switch (arch) {
-			case xmrig::ICpuInfo::ARCH_ZEN:
-			case xmrig::ICpuInfo::ARCH_ZEN_PLUS:
-				// AVX2 init is slow on Zen/Zen+
-				initDatasetAVX2 = false;
-				break;
-			case xmrig::ICpuInfo::ARCH_ZEN2:
-				// AVX2 init is faster on Zen2 without SMT (mobile CPUs)
-				initDatasetAVX2 = xmrig::Cpu::info()->cores() == xmrig::Cpu::info()->threads();
-				break;
-			case xmrig::ICpuInfo::ARCH_ZEN3:
-				// AVX2 init is faster on Zen3
+		if (optimizedInitDatasetEnable) {
+			// Dataset init using AVX2:
+			// -1 = Auto detect
+			//  0 = Always disabled
+			// +1 = Always enabled
+			if (optimizedDatasetInit > 0) {
 				initDatasetAVX2 = true;
-				break;
+			}
+			else if (optimizedDatasetInit < 0) {
+				xmrig::ICpuInfo::Vendor vendor = xmrig::Cpu::info()->vendor();
+				xmrig::ICpuInfo::Arch arch = xmrig::Cpu::info()->arch();
+
+				if (vendor == xmrig::ICpuInfo::VENDOR_INTEL) {
+					// AVX2 init is faster on Intel CPUs without HT
+					initDatasetAVX2 = (xmrig::Cpu::info()->cores() == xmrig::Cpu::info()->threads());
+				}
+				else if (vendor == xmrig::ICpuInfo::VENDOR_AMD) {
+					switch (arch) {
+					case xmrig::ICpuInfo::ARCH_ZEN:
+					case xmrig::ICpuInfo::ARCH_ZEN_PLUS:
+						// AVX2 init is slower on Zen/Zen+
+						initDatasetAVX2 = false;
+						break;
+					case xmrig::ICpuInfo::ARCH_ZEN2:
+						// AVX2 init is faster on Zen2 without SMT (mobile CPUs)
+						initDatasetAVX2 = (xmrig::Cpu::info()->cores() == xmrig::Cpu::info()->threads());
+						break;
+					case xmrig::ICpuInfo::ARCH_ZEN3:
+						// AVX2 init is faster on Zen3
+						initDatasetAVX2 = true;
+						break;
+					}
+				}
 			}
 		}
 
-		// Sorry low-end Intel CPUs
+		// Sorry, low-end Intel CPUs
 		if (!hasAVX2) {
 			initDatasetAVX2 = false;
 		}

src/crypto/randomx/jit_compiler_x86.hpp

@@ -49,7 +49,7 @@ namespace randomx {
 
 	class JitCompilerX86 {
 	public:
-		explicit JitCompilerX86(bool hugePagesEnable);
+		explicit JitCompilerX86(bool hugePagesEnable, bool optimizedInitDatasetEnable);
 		~JitCompilerX86();
 		void prepare();
 		void generateProgram(Program&, ProgramConfiguration&, uint32_t);

src/crypto/randomx/randomx.cpp

@@ -381,7 +381,7 @@ extern "C" {
 					break;
 
 				case RANDOMX_FLAG_JIT:
-					cache->jit          = new randomx::JitCompiler(false);
+					cache->jit          = new randomx::JitCompiler(false, true);
 					cache->initialize   = &randomx::initCacheCompile;
 					cache->datasetInit  = nullptr;
 					cache->memory       = memory;

src/crypto/randomx/randomx.h

@@ -170,6 +170,7 @@ void randomx_apply_config(const T& config)
 
 void randomx_set_scratchpad_prefetch_mode(int mode);
 void randomx_set_huge_pages_jit(bool hugePages);
+void randomx_set_optimized_dataset_init(int value);
 
 #if defined(__cplusplus)
 extern "C" {

src/crypto/randomx/vm_compiled.hpp

@@ -59,7 +59,7 @@ namespace randomx {
 	protected:
 		void execute();
 
-		JitCompiler compiler{ true };
+		JitCompiler compiler{ true, false };
 	};
 
 	using CompiledVmDefault = CompiledVm<1>;

src/crypto/rx/Rx.cpp

@@ -96,6 +96,7 @@ bool xmrig::Rx::init(const T &seed, const RxConfig &config, const CpuConfig &cpu
 
     randomx_set_scratchpad_prefetch_mode(config.scratchpadPrefetchMode());
     randomx_set_huge_pages_jit(cpu.isHugePagesJit());
+    randomx_set_optimized_dataset_init(config.initDatasetAVX2());
 
     if (!msrInitialized) {
         msrEnabled      = msrInit(config, cpu.threads().get(seed.algorithm()).data());

src/crypto/rx/RxConfig.cpp

@@ -47,6 +47,7 @@
 namespace xmrig {
 
 const char *RxConfig::kInit                     = "init";
+const char *RxConfig::kInitAVX2                 = "init-avx2";
 const char *RxConfig::kField                    = "randomx";
 const char *RxConfig::kMode                     = "mode";
 const char *RxConfig::kOneGbPages               = "1gb-pages";
@@ -86,9 +87,10 @@ static_assert (kMsrArraySize == ICpuInfo::MSR_MOD_MAX, "kMsrArraySize and MSR_MO
 bool xmrig::RxConfig::read(const rapidjson::Value &value)
 {
     if (value.IsObject()) {
-        m_threads    = Json::getInt(value, kInit, m_threads);
-        m_mode       = readMode(Json::getValue(value, kMode));
-        m_rdmsr      = Json::getBool(value, kRdmsr, m_rdmsr);
+        m_threads         = Json::getInt(value, kInit, m_threads);
+        m_initDatasetAVX2 = Json::getInt(value, kInitAVX2, m_initDatasetAVX2);
+        m_mode            = readMode(Json::getValue(value, kMode));
+        m_rdmsr           = Json::getBool(value, kRdmsr, m_rdmsr);
 
 #       ifdef XMRIG_FEATURE_MSR
         readMSR(Json::getValue(value, kWrmsr));
@@ -141,6 +143,7 @@ rapidjson::Value xmrig::RxConfig::toJSON(rapidjson::Document &doc) const
 
     Value obj(kObjectType);
     obj.AddMember(StringRef(kInit),         m_threads, allocator);
+    obj.AddMember(StringRef(kInitAVX2),     m_initDatasetAVX2, allocator);
     obj.AddMember(StringRef(kMode),         StringRef(modeName()), allocator);
     obj.AddMember(StringRef(kOneGbPages),   m_oneGbPages, allocator);
     obj.AddMember(StringRef(kRdmsr),        m_rdmsr, allocator);

src/crypto/rx/RxConfig.h

@@ -61,6 +61,7 @@ public:
     static const char *kCacheQoS;
     static const char *kField;
     static const char *kInit;
+    static const char *kInitAVX2;
     static const char *kMode;
     static const char *kOneGbPages;
     static const char *kRdmsr;
@@ -83,6 +84,7 @@ public:
     const char *modeName() const;
     uint32_t threads(uint32_t limit = 100) const;
 
+    inline int initDatasetAVX2() const  { return m_initDatasetAVX2; }
     inline bool isOneGbPages() const    { return m_oneGbPages; }
     inline bool rdmsr() const           { return m_rdmsr; }
     inline bool wrmsr() const           { return m_wrmsr; }
@@ -111,11 +113,12 @@ private:
 
     Mode readMode(const rapidjson::Value &value) const;
 
-    bool m_numa         = true;
-    bool m_oneGbPages   = false;
-    bool m_rdmsr        = true;
-    int m_threads       = -1;
-    Mode m_mode         = AutoMode;
+    bool m_numa           = true;
+    bool m_oneGbPages     = false;
+    bool m_rdmsr          = true;
+    int m_threads         = -1;
+    int m_initDatasetAVX2 = -1;
+    Mode m_mode           = AutoMode;
 
     ScratchpadPrefetchMode m_scratchpadPrefetchMode = ScratchpadPrefetchT0;