Group E exponent changed from a static value (-240) to dynamic

src/asm/program_xmm_constants.inc

@@ -1,6 +1,6 @@
 mantissaMask:
-	db 255, 255, 255, 255, 255, 255, 15, 0, 255, 255, 255, 255, 255, 255, 15, 0
+	db 255, 255, 255, 255, 255, 255, 255, 0, 255, 255, 255, 255, 255, 255, 255, 0
 exp240:
-	db 0, 0, 0, 0, 0, 0, 240, 48, 0, 0, 0, 0, 0, 0, 240, 48
+	db 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
 scaleMask:
 	db 0, 0, 0, 0, 0, 0, 240, 129, 0, 0, 0, 0, 0, 0, 240, 129

src/assembly_generator_x86.cpp

@@ -35,9 +35,9 @@ namespace randomx {
 	static const char* regA[] = { "xmm8", "xmm9", "xmm10", "xmm11" };
 
 	static const char* tempRegx = "xmm12";
-	static const char* mantissaMask = "xmm13";
+	static const char* mantissaMaskReg = "xmm13";
-	static const char* exponentMask = "xmm14";
+	static const char* exponentMaskReg = "xmm14";
-	static const char* scaleMask = "xmm15";
+	static const char* scaleMaskReg = "xmm15";
 	static const char* regIc = "rbx";
 	static const char* regIc32 = "ebx";
 	static const char* regIc8 = "bl";
@@ -328,7 +328,6 @@ namespace randomx {
 		traceint(instr);
 	}
 
-	//4 uOPs
 	void AssemblyGeneratorX86::h_IMULH_R(Instruction& instr, int i) {
 		registerUsage[instr.dst].lastUsed = i;
 		asmCode << "\tmov rax, " << regR[instr.dst] << std::endl;
@@ -489,7 +488,7 @@ namespace randomx {
 
 	void AssemblyGeneratorX86::h_FSCAL_R(Instruction& instr, int i) {
 		instr.dst %= RegisterCountFlt;
-		asmCode << "\txorps " << regF[instr.dst] << ", " << scaleMask << std::endl;
+		asmCode << "\txorps " << regF[instr.dst] << ", " << scaleMaskReg << std::endl;
 		traceflt(instr);
 	}
 
@@ -504,8 +503,8 @@ namespace randomx {
 		instr.dst %= RegisterCountFlt;
 		genAddressReg(instr);
 		asmCode << "\tcvtdq2pd " << tempRegx << ", qword ptr [" << regScratchpadAddr << "+rax]" << std::endl;
-		asmCode << "\tandps " << tempRegx << ", " << mantissaMask << std::endl;
+		asmCode << "\tandps " << tempRegx << ", " << mantissaMaskReg << std::endl;
-		asmCode << "\torps " << tempRegx << ", " << exponentMask << std::endl;
+		asmCode << "\torps " << tempRegx << ", " << exponentMaskReg << std::endl;
 		asmCode << "\tdivpd " << regE[instr.dst] << ", " << tempRegx << std::endl;
 		traceflt(instr);
 	}

src/common.hpp

@@ -122,6 +122,16 @@ namespace randomx {
 		return minIndex;
 	}
 
+	constexpr int mantissaSize = 52;
+	constexpr int exponentSize = 11;
+	constexpr uint64_t mantissaMask = (1ULL << mantissaSize) - 1;
+	constexpr uint64_t exponentMask = (1ULL << exponentSize) - 1;
+	constexpr int exponentBias = 1023;
+	constexpr int dynamicExponentBits = 4;
+	constexpr int staticExponentBits = 4;
+	constexpr uint64_t constExponentBits = 0x300;
+	constexpr uint64_t dynamicMantissaMask = (1ULL << (mantissaSize + dynamicExponentBits)) - 1;
+
 	struct MemoryRegisters {
 		addr_t mx, ma;
 		uint8_t* memory = nullptr;

src/intrin_portable.h

@@ -312,12 +312,6 @@ inline __m128d load_cvt_i32x2(const void* addr) {
 	return _mm_cvtepi32_pd(ix);
 }
 
-template<int E>
-constexpr uint64_t ieee_get_exponent_mask() {
-	static_assert(E > -1023, "Invalid exponent value");
-	return (uint64_t)(E + 1023U) << 52;
-}
-
 double loadDoublePortable(const void* addr);
 uint64_t mulh(uint64_t, uint64_t);
 int64_t smulh(int64_t, int64_t);

src/tests/benchmark.cpp

@@ -229,7 +229,7 @@ int main(int argc, char** argv) {
 		std::cout << "Calculated result: ";
 		result.print(std::cout);
 		if (noncesCount == 1000 && seedValue == 0)
-			std::cout << "Reference result:  092868e4cee629a5b3848b97a52199d8a158e5b56ab9064764cda7ff656f3741" << std::endl;
+			std::cout << "Reference result:  6d95d8d07fa3a80771f33d1b20452b61ab2d0bf21058b5e586fad38bf3e1e0ca" << std::endl;
 		if (!miningMode) {
 			std::cout << "Performance: " << 1000 * elapsed / noncesCount << " ms per hash" << std::endl;
 		}

src/virtual_machine.cpp

@@ -35,30 +35,40 @@ void randomx_vm::resetRoundingMode() {
 	initFpu();
 }
 
-constexpr int mantissaSize = 52;
+namespace randomx {
-constexpr int exponentSize = 11;
+
-constexpr uint64_t mantissaMask = (1ULL << mantissaSize) - 1;
+	static inline uint64_t getSmallPositiveFloatBits(uint64_t entropy) {
-constexpr uint64_t exponentMask = (1ULL << exponentSize) - 1;
+		auto exponent = entropy >> 59; //0..31
-constexpr int exponentBias = 1023;
+		auto mantissa = entropy & mantissaMask;
+		exponent += exponentBias;
+		exponent &= exponentMask;
+		exponent <<= mantissaSize;
+		return exponent | mantissa;
+	}
+
+	static inline uint64_t getStaticExponent(uint64_t entropy) {
+		auto exponent = constExponentBits;
+		exponent |= (entropy >> (64 - staticExponentBits)) << dynamicExponentBits;
+		exponent <<= mantissaSize;
+		return exponent;
+	}
+
+	static inline uint64_t getFloatMask(uint64_t entropy) {
+		constexpr uint64_t mask22bit = (1ULL << 22) - 1;
+		return (entropy & mask22bit) | getStaticExponent(entropy);
+	}
 
-static inline uint64_t getSmallPositiveFloatBits(uint64_t entropy) {
-	auto exponent = entropy >> 59; //0..31
-	auto mantissa = entropy & mantissaMask;
-	exponent += exponentBias;
-	exponent &= exponentMask;
-	exponent <<= mantissaSize;
-	return exponent | mantissa;
 }
 
 void randomx_vm::initialize() {
-	store64(&reg.a[0].lo, getSmallPositiveFloatBits(program.getEntropy(0)));
+	store64(&reg.a[0].lo, randomx::getSmallPositiveFloatBits(program.getEntropy(0)));
-	store64(&reg.a[0].hi, getSmallPositiveFloatBits(program.getEntropy(1)));
+	store64(&reg.a[0].hi, randomx::getSmallPositiveFloatBits(program.getEntropy(1)));
-	store64(&reg.a[1].lo, getSmallPositiveFloatBits(program.getEntropy(2)));
+	store64(&reg.a[1].lo, randomx::getSmallPositiveFloatBits(program.getEntropy(2)));
-	store64(&reg.a[1].hi, getSmallPositiveFloatBits(program.getEntropy(3)));
+	store64(&reg.a[1].hi, randomx::getSmallPositiveFloatBits(program.getEntropy(3)));
-	store64(&reg.a[2].lo, getSmallPositiveFloatBits(program.getEntropy(4)));
+	store64(&reg.a[2].lo, randomx::getSmallPositiveFloatBits(program.getEntropy(4)));
-	store64(&reg.a[2].hi, getSmallPositiveFloatBits(program.getEntropy(5)));
+	store64(&reg.a[2].hi, randomx::getSmallPositiveFloatBits(program.getEntropy(5)));
-	store64(&reg.a[3].lo, getSmallPositiveFloatBits(program.getEntropy(6)));
+	store64(&reg.a[3].lo, randomx::getSmallPositiveFloatBits(program.getEntropy(6)));
-	store64(&reg.a[3].hi, getSmallPositiveFloatBits(program.getEntropy(7)));
+	store64(&reg.a[3].hi, randomx::getSmallPositiveFloatBits(program.getEntropy(7)));
 	mem.ma = program.getEntropy(8) & randomx::CacheLineAlignMask;
 	mem.mx = program.getEntropy(10);
 	auto addressRegisters = program.getEntropy(12);
@@ -70,10 +80,8 @@ void randomx_vm::initialize() {
 	addressRegisters >>= 1;
 	config.readReg3 = 6 + (addressRegisters & 1);
 	datasetOffset = (program.getEntropy(13) & randomx::DatasetExtraItems) * randomx::CacheLineSize;
-	constexpr uint64_t mask22bit = (1ULL << 22) - 1;
+	store64(&config.eMask[0], randomx::getFloatMask(program.getEntropy(14)));
-	constexpr uint64_t maskExp240 = ieee_get_exponent_mask<-240>();
+	store64(&config.eMask[1], randomx::getFloatMask(program.getEntropy(15)));
-	store64(&config.eMask[0], (program.getEntropy(14) & mask22bit) | maskExp240);
-	store64(&config.eMask[1], (program.getEntropy(15) & mask22bit) | maskExp240);
 }
 
 namespace randomx {

src/vm_interpreted.cpp

@@ -61,11 +61,10 @@ namespace randomx {
 
 	template<class Allocator, bool softAes>
 	FORCE_INLINE __m128d InterpretedVm<Allocator, softAes>::maskRegisterExponentMantissa(__m128d x) {
-		constexpr uint64_t mantissaMask64 = (1ULL << 52) - 1;
+		const __m128d xmantissaMask = _mm_castsi128_pd(_mm_set_epi64x(dynamicMantissaMask, dynamicMantissaMask));
-		const __m128d mantissaMask = _mm_castsi128_pd(_mm_set_epi64x(mantissaMask64, mantissaMask64));
+		const __m128d xexponentMask = _mm_load_pd((const double*)&config.eMask);
-		const __m128d exponentMask = _mm_load_pd((const double*)&config.eMask);
+		x = _mm_and_pd(x, xmantissaMask);
-		x = _mm_and_pd(x, mantissaMask);
+		x = _mm_or_pd(x, xexponentMask);
-		x = _mm_or_pd(x, exponentMask);
 		return x;
 	}
 

vcxproj/benchmark.vcxproj

@@ -96,7 +96,7 @@
     <ClCompile>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
-      <SDLCheck>true</SDLCheck>
+      <SDLCheck>false</SDLCheck>
       <ConformanceMode>true</ConformanceMode>
     </ClCompile>
   </ItemDefinitionGroup>