Merge pull request #1823 from SChernykh/dev

RandomX: added parameter for scratchpad prefetch mode

src/config.json

@@ -21,7 +21,8 @@
         "rdmsr": true,
         "wrmsr": true,
         "cache_qos": false,
-        "numa": true
+        "numa": true,
+        "scratchpad_prefetch_mode": 1
     },
     "cpu": {
         "enabled": true,

src/core/config/Config_default.h

@@ -51,7 +51,12 @@ R"===(
     "randomx": {
         "init": -1,
         "mode": "auto",
-        "numa": true
+        "1gb-pages": false,
+        "rdmsr": true,
+        "wrmsr": true,
+        "cache_qos": false,
+        "numa": true,
+        "scratchpad_prefetch_mode": 1
     },
     "cpu": {
         "enabled": true,

src/crypto/randomx/randomx.cpp

@@ -211,6 +211,13 @@ RandomX_ConfigurationBase::RandomX_ConfigurationBase()
 static uint32_t Log2(size_t value) { return (value > 1) ? (Log2(value / 2) + 1) : 0; }
 #endif
 
+static int scratchpadPrefetchMode = 1;
+
+void randomx_set_scratchpad_prefetch_mode(int mode)
+{
+	scratchpadPrefetchMode = mode;
+}
+
 void RandomX_ConfigurationBase::Apply()
 {
 	const uint32_t ScratchpadL1Mask_Calculated = (ScratchpadL1_Size / sizeof(uint64_t) - 1) * 8;
@@ -240,6 +247,36 @@ void RandomX_ConfigurationBase::Apply()
 	*(uint32_t*)(codePrefetchScratchpadTweaked + 4) = ScratchpadL3Mask64_Calculated;
 	*(uint32_t*)(codePrefetchScratchpadTweaked + 18) = ScratchpadL3Mask64_Calculated;
 
+	// Apply scratchpad prefetch mode
+	{
+		uint32_t* a = (uint32_t*)(codePrefetchScratchpadTweaked + 8);
+		uint32_t* b = (uint32_t*)(codePrefetchScratchpadTweaked + 22);
+
+		switch (scratchpadPrefetchMode)
+		{
+		case 0:
+			*a = 0x00401F0FUL; // 4-byte nop
+			*b = 0x00401F0FUL; // 4-byte nop
+			break;
+
+		case 1:
+		default:
+			*a = 0x060C180FUL; // prefetcht0 [rsi+rax]
+			*b = 0x160C180FUL; // prefetcht0 [rsi+rdx]
+			break;
+
+		case 2:
+			*a = 0x0604180FUL; // prefetchnta [rsi+rax]
+			*b = 0x1604180FUL; // prefetchnta [rsi+rdx]
+			break;
+
+		case 3:
+			*a = 0x060C8B48UL; // mov rcx, [rsi+rax]
+			*b = 0x160C8B48UL; // mov rcx, [rsi+rdx]
+			break;
+		}
+	}
+
 #define JIT_HANDLE(x, prev) randomx::JitCompilerX86::engine[k] = &randomx::JitCompilerX86::h_##x
 
 #elif defined(XMRIG_ARMv8)

src/crypto/randomx/randomx.h

@@ -200,6 +200,8 @@ void randomx_apply_config(const T& config)
 	RandomX_CurrentConfig.Apply();
 }
 
+void randomx_set_scratchpad_prefetch_mode(int mode);
+
 #if defined(__cplusplus)
 extern "C" {
 #endif

src/crypto/rx/Rx.cpp

@@ -32,6 +32,7 @@
 #include "base/io/log/Log.h"
 #include "crypto/rx/RxConfig.h"
 #include "crypto/rx/RxQueue.h"
+#include "crypto/randomx/randomx.h"
 
 
 namespace xmrig {
@@ -99,6 +100,8 @@ bool xmrig::Rx::init(const T &seed, const RxConfig &config, const CpuConfig &cpu
         return true;
     }
 
+    randomx_set_scratchpad_prefetch_mode(config.scratchpadPrefetchMode());
+
     if (isReady(seed)) {
         return true;
     }

src/crypto/rx/RxConfig.cpp

@@ -57,6 +57,8 @@ static const char *kCacheQoS    = "cache_qos";
 static const char *kNUMA        = "numa";
 #endif
 
+static const char *kScratchpadPrefetchMode = "scratchpad_prefetch_mode";
+
 static const std::array<const char *, RxConfig::ModeMax> modeNames = { "auto", "fast", "light" };
 
 
@@ -118,6 +120,11 @@ bool xmrig::RxConfig::read(const rapidjson::Value &value)
         }
 #       endif
 
+        const int mode = Json::getInt(value, kScratchpadPrefetchMode, static_cast<int>(m_scratchpadPrefetchMode));
+        if ((mode >= ScratchpadPrefetchOff) && (mode < ScratchpadPrefetchMax)) {
+            m_scratchpadPrefetchMode = static_cast<ScratchpadPrefetchMode>(mode);
+        }
+
         return true;
     }
 
@@ -171,6 +178,8 @@ rapidjson::Value xmrig::RxConfig::toJSON(rapidjson::Document &doc) const
     }
 #   endif
 
+    obj.AddMember(StringRef(kScratchpadPrefetchMode), static_cast<int>(m_scratchpadPrefetchMode), allocator);
+
     return obj;
 }
 

src/crypto/rx/RxConfig.h

@@ -50,6 +50,14 @@ public:
         ModeMax
     };
 
+    enum ScratchpadPrefetchMode : uint32_t {
+        ScratchpadPrefetchOff,
+        ScratchpadPrefetchT0,
+        ScratchpadPrefetchNTA,
+        ScratchpadPrefetchMov,
+        ScratchpadPrefetchMax,
+    };
+
     bool read(const rapidjson::Value &value);
     rapidjson::Value toJSON(rapidjson::Document &doc) const;
 
@@ -68,6 +76,8 @@ public:
     inline bool cacheQoS() const        { return m_cacheQoS; }
     inline Mode mode() const            { return m_mode; }
 
+    inline ScratchpadPrefetchMode scratchpadPrefetchMode() const { return m_scratchpadPrefetchMode; }
+
 #   ifdef XMRIG_FEATURE_MSR
     const char *msrPresetName() const;
     const MsrItems &msrPreset() const;
@@ -94,6 +104,8 @@ private:
     int m_threads       = -1;
     Mode m_mode         = AutoMode;
 
+    ScratchpadPrefetchMode m_scratchpadPrefetchMode = ScratchpadPrefetchT0;
+
 #   ifdef XMRIG_FEATURE_HWLOC
     std::vector<uint32_t> m_nodeset;
 #   endif