Fixed generic OpenCL code for AMD Navi

src/backend/opencl/cl/rx/randomx_vm.cl

@@ -316,7 +316,7 @@ __kernel void init_vm(__global const void* entropy_data, __global void* vm_state
 		uint64_t registerLatencyFP = 0;
 		uint64_t registerReadCycleFP = 0;
 		uint32_t ScratchpadHighLatency = 0;
-		uint32_t ScratchpadLatency = 0;
+		volatile uint32_t ScratchpadLatency = 0;
 
 		int32_t first_available_slot = 0;
 		int32_t first_allowed_slot_cfround = 0;
@@ -1425,8 +1425,7 @@ double fma_soft(double a, double b, double c, uint32_t rounding_mode)
 	}
 
 	const uint64_t mantissa_size = 52;
-	const uint64_t mantissa_mask = (1UL << mantissa_size) - 1;
+	const uint64_t mantissa_mask = (1UL << 52) - 1;
-	const uint64_t mantissa_high_bit = 1UL << mantissa_size;
 
 	const uint64_t exponent_size = 11;
 	const uint64_t exponent_mask = (1 << exponent_size) - 1;
@@ -1441,9 +1440,13 @@ double fma_soft(double a, double b, double c, uint32_t rounding_mode)
 		return as_double(inf);
 	}
 
-	const uint64_t mantissa_a = (as_ulong(a) & mantissa_mask) | mantissa_high_bit;
+	uint64_t mantissa_a = (as_ulong(a) & mantissa_mask);
-	const uint64_t mantissa_b = (as_ulong(b) & mantissa_mask) | mantissa_high_bit;
+	uint64_t mantissa_b = (as_ulong(b) & mantissa_mask);
-	const uint64_t mantissa_c = (as_ulong(c) & mantissa_mask) | mantissa_high_bit;
+	uint64_t mantissa_c = (as_ulong(c) & mantissa_mask);
+
+	((uint2*)&mantissa_a)->y |= 1U << 20;
+	((uint2*)&mantissa_b)->y |= 1U << 20;
+	((uint2*)&mantissa_c)->y |= 1U << 20;
 
 	const uint32_t sign_a = as_uint2(a).y >> 31;
 	const uint32_t sign_b = as_uint2(b).y >> 31;
@@ -1585,7 +1588,7 @@ double fma_soft(double a, double b, double c, uint32_t rounding_mode)
 	if (rounding_mode + sign_fma_result == 2)
 	{
 		fma_result[1] += round_up;
-		if (fma_result[1] == mantissa_high_bit)
+		if (fma_result[1] == (1UL << mantissa_size))
 		{
 			fma_result[1] = 0;
 			++exponent_fma_result;

src/backend/opencl/generators/ocl_generic_rx_generator.cpp

@@ -49,6 +49,7 @@ bool ocl_generic_rx_generator(const OclDevice &device, const Algorithm &algorith
     const size_t mem = device.globalMemSize();
     auto config      = RxAlgo::base(algorithm);
     bool gcnAsm      = false;
+    bool isNavi      = false;
 
     switch (device.type()) {
     case OclDevice::Baffin:
@@ -59,6 +60,12 @@ bool ocl_generic_rx_generator(const OclDevice &device, const Algorithm &algorith
         gcnAsm = true;
         break;
 
+    case OclDevice::Navi_10:
+    case OclDevice::Navi_12:
+    case OclDevice::Navi_14:
+        isNavi = true;
+        break;
+
     default:
         break;
     }
@@ -75,8 +82,9 @@ bool ocl_generic_rx_generator(const OclDevice &device, const Algorithm &algorith
     uint32_t intensity = static_cast<uint32_t>((mem - (datasetHost ? 0 : dataset_mem)) / per_thread_mem / 2);
 
     // Too high intensity makes hashrate worse
-    if (intensity > device.computeUnits() * 16) {
+    const uint32_t intensityCoeff = isNavi ? 64 : 16;
-        intensity = device.computeUnits() * 16;
+    if (intensity > device.computeUnits() * intensityCoeff) {
+        intensity = device.computeUnits() * intensityCoeff;
     }
 
     intensity -= intensity % 64;

src/backend/opencl/wrappers/OclDevice.cpp

@@ -107,6 +107,14 @@ static OclDevice::Type getType(const String &name)
         return OclDevice::Navi_10;
     }
 
+    if (name == "gfx1011") {
+        return OclDevice::Navi_12;
+    }
+
+    if (name == "gfx1012") {
+        return OclDevice::Navi_14;
+    }
+
     if (name == "gfx804") {
         return OclDevice::Lexa;
     }

src/backend/opencl/wrappers/OclDevice.h

@@ -56,7 +56,9 @@ public:
         Vega_10,
         Vega_20,
         Raven,
-        Navi_10
+        Navi_10,
+        Navi_12,
+        Navi_14
     };
 
     OclDevice() = delete;