Update CN-GPU integration

CMakeLists.txt

@@ -5,8 +5,8 @@ option(WITH_HWLOC           "Enable hwloc support" ON)
 option(WITH_CN_LITE         "Enable CryptoNight-Lite algorithms family" ON)
 option(WITH_CN_HEAVY        "Enable CryptoNight-Heavy algorithms family" ON)
 option(WITH_CN_PICO         "Enable CryptoNight-Pico algorithm" ON)
-option(WITH_CN_GPU          "Enable CryptoNight-GPU algorithm" ON)
 option(WITH_CN_FEMTO        "Enable CryptoNight-UPX2 algorithm" ON)
+option(WITH_CN_GPU          "Enable CryptoNight-GPU algorithm" ON)
 option(WITH_RANDOMX         "Enable RandomX algorithms family" ON)
 option(WITH_ARGON2          "Enable Argon2 algorithms family" ON)
 option(WITH_ASTROBWT        "Enable AstroBWT algorithms family" ON)
@@ -206,6 +206,7 @@ add_definitions(-D__STDC_FORMAT_MACROS -DUNICODE -D_FILE_OFFSET_BITS=64)
 find_package(UV REQUIRED)
 
 include(cmake/flags.cmake)
+include(cmake/cn-gpu.cmake)
 include(cmake/randomx.cmake)
 include(cmake/argon2.cmake)
 include(cmake/astrobwt.cmake)
@@ -213,7 +214,6 @@ include(cmake/kawpow.cmake)
 include(cmake/ghostrider.cmake)
 include(cmake/OpenSSL.cmake)
 include(cmake/asm.cmake)
-include(cmake/cn-gpu.cmake)
 
 if (WITH_CN_LITE)
     add_definitions(/DXMRIG_ALGO_CN_LITE)

src/backend/cpu/CpuConfig_gen.h

@@ -48,16 +48,16 @@ size_t inline generate<Algorithm::CN>(Threads<CpuThreads> &threads, uint32_t lim
     size_t count = 0;
 
     count += generate(Algorithm::kCN, threads, Algorithm::CN_1, limit);
+    count += generate(Algorithm::kCN_2, threads, Algorithm::CN_2, limit);
+#   ifdef XMRIG_ALGO_CN_GPU
+    count += generate(Algorithm::kCN_GPU, threads, Algorithm::CN_GPU, limit);
+#   endif
 
     if (!threads.isExist(Algorithm::CN_0)) {
         threads.disable(Algorithm::CN_0);
         ++count;
     }
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    count += generate(Algorithm::kCN_GPU, threads, Algorithm::CN_GPU, limit);
-#   endif
-
     return count;
 }
 

src/backend/cpu/CpuWorker.cpp

@@ -181,20 +181,19 @@ bool xmrig::CpuWorker<N>::selfTest()
                         verify(Algorithm::CN_RWZ,    test_output_rwz)  &&
                         verify(Algorithm::CN_ZLS,    test_output_zls)  &&
                         verify(Algorithm::CN_CCX,    test_output_ccx)  &&
+                        verify(Algorithm::CN_DOUBLE, test_output_double)
 #                       ifdef XMRIG_ALGO_CN_GPU
-                        verify(Algorithm::CN_GPU,    test_output_gpu)  &&
+                        &&
+                        verify(Algorithm::CN_GPU,    test_output_gpu)
 #                       endif
-                        verify(Algorithm::CN_DOUBLE, test_output_double);
+                        ;
 
 #       ifdef XMRIG_ALGO_CN_GPU
-        if (!rc || N > 1) {
+        if (! (!rc || N > 1)) {
-            return rc;
+            return verify(Algorithm::CN_GPU, test_output_gpu);
-        }
+        } else
-
-        return verify(Algorithm::CN_GPU, test_output_gpu);
-#       else
-        return rc;
 #       endif
+        return rc;
     }
 
 #   ifdef XMRIG_ALGO_CN_LITE

src/backend/cpu/platform/BasicCpuInfo.cpp

@@ -340,6 +340,12 @@ xmrig::CpuThreads xmrig::BasicCpuInfo::threads(const Algorithm &algorithm, uint3
     }
 #   endif
 
+#   ifdef XMRIG_ALGO_CN_GPU
+    if (algorithm == Algorithm::CN_GPU) {
+        return count_limit;
+    }
+#   endif
+
 #   ifdef XMRIG_ALGO_RANDOMX
     if (f == Algorithm::RANDOM_X) {
         if (algorithm == Algorithm::RX_WOW) {
@@ -374,12 +380,6 @@ xmrig::CpuThreads xmrig::BasicCpuInfo::threads(const Algorithm &algorithm, uint3
     }
 #   endif
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    if (algorithm == Algorithm::CN_GPU) {
-        return count_limit;
-    }
-#   endif
-
 #   ifdef XMRIG_ALGO_GHOSTRIDER
     if (f == Algorithm::GHOSTRIDER) {
         return CpuThreads(std::max<size_t>(count / 2, 1), 8);

src/backend/cuda/CudaConfig_gen.h

@@ -52,16 +52,15 @@ size_t inline generate<Algorithm::CN>(Threads<CudaThreads> &threads, const std::
 
     count += generate(Algorithm::kCN, threads, Algorithm::CN_1, devices);
     count += generate(Algorithm::kCN_2, threads, Algorithm::CN_2, devices);
+#   ifdef XMRIG_ALGO_CN_GPU
+    count += generate(Algorithm::kCN_GPU, threads, Algorithm::CN_GPU, devices);
+#   endif
 
     if (!threads.isExist(Algorithm::CN_0)) {
         threads.disable(Algorithm::CN_0);
         count++;
     }
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    count += generate("cn/gpu", threads, Algorithm::CN_GPU, devices);
-#   endif
-
     return count;
 }
 

src/backend/opencl/OclConfig_gen.h

@@ -51,16 +51,15 @@ size_t inline generate<Algorithm::CN>(Threads<OclThreads> &threads, const std::v
 
     count += generate(Algorithm::kCN, threads, Algorithm::CN_1, devices);
     count += generate(Algorithm::kCN_2, threads, Algorithm::CN_2, devices);
+#   ifdef XMRIG_ALGO_CN_GPU
+    count += generate(Algorithm::kCN_GPU, threads, Algorithm::CN_GPU, devices);
+#   endif
 
     if (!threads.isExist(Algorithm::CN_0)) {
         threads.disable(Algorithm::CN_0);
         count++;
     }
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    count += generate("cn/gpu", threads, Algorithm::CN_GPU, devices);
-#   endif
-
     return count;
 }
 

src/backend/opencl/OclThread.h

@@ -45,6 +45,20 @@ public:
         setIntensity(intensity);
     }
 
+#   ifdef XMRIG_ALGO_CN_GPU
+    OclThread(uint32_t index, uint32_t intensity, uint32_t worksize, uint32_t threads, uint32_t unrollFactor) :
+        m_fields(0),
+        m_threads(threads, -1),
+        m_index(index),
+        m_memChunk(0),
+        m_stridedIndex(0),
+        m_unrollFactor(unrollFactor),
+        m_worksize(worksize)
+    {
+        setIntensity(intensity);
+    }
+#   endif
+
 #   ifdef XMRIG_ALGO_RANDOMX
     OclThread(uint32_t index, uint32_t intensity, uint32_t worksize, uint32_t threads, bool gcnAsm, bool datasetHost, uint32_t bfactor) :
         m_datasetHost(datasetHost),
@@ -89,20 +103,6 @@ public:
     }
 #   endif
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    OclThread(uint32_t index, uint32_t intensity, uint32_t worksize, uint32_t threads, uint32_t unrollFactor) :
-        m_fields(0),
-        m_threads(threads, -1),
-        m_index(index),
-        m_memChunk(0),
-        m_stridedIndex(0),
-        m_unrollFactor(unrollFactor),
-        m_worksize(worksize)
-    {
-        setIntensity(intensity);
-    }
-#   endif
-
     OclThread(const rapidjson::Value &value);
 
     inline bool isAsm() const                               { return m_gcnAsm; }

src/backend/opencl/OclWorker.cpp

@@ -29,6 +29,10 @@
 #include "net/JobResults.h"
 
 
+#ifdef XMRIG_ALGO_CN_GPU
+#   include "backend/opencl/runners/OclRyoRunner.h"
+#endif
+
 #ifdef XMRIG_ALGO_RANDOMX
 #   include "backend/opencl/runners/OclRxJitRunner.h"
 #   include "backend/opencl/runners/OclRxVmRunner.h"
@@ -42,10 +46,6 @@
 #   include "backend/opencl/runners/OclKawPowRunner.h"
 #endif
 
-#ifdef XMRIG_ALGO_CN_GPU
-#   include "backend/opencl/runners/OclRyoRunner.h"
-#endif
-
 #include <cassert>
 #include <thread>
 

src/backend/opencl/cl/OclSource.cpp

@@ -28,6 +28,10 @@
 #include "base/crypto/Algorithm.h"
 
 
+#ifdef XMRIG_ALGO_CN_GPU
+#   include "backend/opencl/cl/cn/cryptonight_gpu_cl.h"
+#endif
+
 #ifdef XMRIG_ALGO_RANDOMX
 #   include "backend/opencl/cl/rx/randomx_cl.h"
 #endif
@@ -41,13 +45,15 @@
 #   include "backend/opencl/cl/kawpow/kawpow_dag_cl.h"
 #endif
 
-#ifdef XMRIG_ALGO_CN_GPU
-#   include "backend/opencl/cl/cn/cryptonight_gpu_cl.h"
-#endif
-
 
 const char *xmrig::OclSource::get(const Algorithm &algorithm)
 {
+#   ifdef XMRIG_ALGO_CN_GPU
+    if (algorithm == Algorithm::CN_GPU) {
+        return cryptonight_gpu_cl;
+    }
+#   endif
+
 #   ifdef XMRIG_ALGO_RANDOMX
     if (algorithm.family() == Algorithm::RANDOM_X) {
         return randomx_cl;
@@ -66,11 +72,5 @@ const char *xmrig::OclSource::get(const Algorithm &algorithm)
     }
 #   endif
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    if (algorithm == Algorithm::CN_GPU) {
-        return cryptonight_gpu_cl;
-    }
-#   endif
-
     return cryptonight_cl;
 }

src/backend/opencl/cl/cn/algorithm.cl

@@ -18,6 +18,7 @@
 #define ALGO_CN_PICO_0      0x63120200
 #define ALGO_CN_PICO_TLO    0x63120274
 #define ALGO_CN_UPX2        0x63110200
+#define ALGO_CN_GPU         0x63150300
 #define ALGO_RX_0           0x72151200
 #define ALGO_RX_WOW         0x72141177
 #define ALGO_RX_ARQMA       0x72121061
@@ -30,7 +31,6 @@
 #define ALGO_ASTROBWT_DERO  0x41000000
 #define ALGO_KAWPOW_RVN     0x6b0f0000
 
-#define ALGO_CN_GPU         0x631500ff
 #define ALGO_RX_XLA         0x721211ff
 
 #define FAMILY_UNKNOWN      0

src/backend/opencl/cl/cn/cryptonight_cl.h

@@ -30,19 +30,19 @@ static const char cryptonight_cl[61582] = {
     0x30,0x36,0x38,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x50,0x49,0x43,0x4f,0x5f,0x30,0x20,0x30,0x78,0x36,0x33,0x31,
     0x32,0x30,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x50,0x49,0x43,0x4f,0x5f,0x54,0x4c,0x4f,0x20,0x30,
     0x78,0x36,0x33,0x31,0x32,0x30,0x32,0x37,0x34,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x55,0x50,0x58,0x32,0x20,0x30,
-    0x78,0x36,0x33,0x31,0x31,0x30,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x30,0x20,0x30,0x78,0x37,0x32,
+    0x78,0x36,0x33,0x31,0x31,0x30,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x47,0x50,0x55,0x20,0x30,0x78,
-    0x31,0x35,0x31,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x57,0x4f,0x57,0x20,0x30,0x78,0x37,0x32,0x31,
+    0x36,0x33,0x31,0x35,0x30,0x33,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x30,0x20,0x30,0x78,0x37,0x32,0x31,
-    0x34,0x31,0x31,0x37,0x37,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x41,0x52,0x51,0x4d,0x41,0x20,0x30,0x78,0x37,0x32,
+    0x35,0x31,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x57,0x4f,0x57,0x20,0x30,0x78,0x37,0x32,0x31,0x34,
-    0x31,0x32,0x31,0x30,0x36,0x31,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x53,0x46,0x58,0x20,0x30,0x78,0x37,0x32,0x31,
+    0x31,0x31,0x37,0x37,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x41,0x52,0x51,0x4d,0x41,0x20,0x30,0x78,0x37,0x32,0x31,
-    0x35,0x31,0x32,0x37,0x33,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x4b,0x45,0x56,0x41,0x20,0x30,0x78,0x37,0x32,0x31,
+    0x32,0x31,0x30,0x36,0x31,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x53,0x46,0x58,0x20,0x30,0x78,0x37,0x32,0x31,0x35,
-    0x34,0x31,0x31,0x36,0x62,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x47,0x52,0x41,0x46,0x54,0x20,0x30,0x78,0x37,0x32,
+    0x31,0x32,0x37,0x33,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x4b,0x45,0x56,0x41,0x20,0x30,0x78,0x37,0x32,0x31,0x34,
-    0x31,0x35,0x31,0x32,0x36,0x37,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x43,0x48,0x55,0x4b,0x57,0x41,0x20,0x30,
+    0x31,0x31,0x36,0x62,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x47,0x52,0x41,0x46,0x54,0x20,0x30,0x78,0x37,0x32,0x31,
-    0x78,0x36,0x31,0x31,0x33,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x43,0x48,0x55,0x4b,0x57,
+    0x35,0x31,0x32,0x36,0x37,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x43,0x48,0x55,0x4b,0x57,0x41,0x20,0x30,0x78,
-    0x41,0x5f,0x56,0x32,0x20,0x30,0x78,0x36,0x31,0x31,0x34,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,
+    0x36,0x31,0x31,0x33,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x43,0x48,0x55,0x4b,0x57,0x41,
-    0x5f,0x57,0x52,0x4b,0x5a,0x20,0x30,0x78,0x36,0x31,0x31,0x32,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x53,
+    0x5f,0x56,0x32,0x20,0x30,0x78,0x36,0x31,0x31,0x34,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,
-    0x54,0x52,0x4f,0x42,0x57,0x54,0x5f,0x44,0x45,0x52,0x4f,0x20,0x30,0x78,0x34,0x31,0x30,0x30,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,
+    0x57,0x52,0x4b,0x5a,0x20,0x30,0x78,0x36,0x31,0x31,0x32,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x53,0x54,
-    0x4c,0x47,0x4f,0x5f,0x4b,0x41,0x57,0x50,0x4f,0x57,0x5f,0x52,0x56,0x4e,0x20,0x30,0x78,0x36,0x62,0x30,0x66,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,
+    0x52,0x4f,0x42,0x57,0x54,0x5f,0x44,0x45,0x52,0x4f,0x20,0x30,0x78,0x34,0x31,0x30,0x30,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,
-    0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x47,0x50,0x55,0x20,0x30,0x78,0x36,0x33,0x31,0x35,0x30,0x30,0x66,0x66,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,
+    0x47,0x4f,0x5f,0x4b,0x41,0x57,0x50,0x4f,0x57,0x5f,0x52,0x56,0x4e,0x20,0x30,0x78,0x36,0x62,0x30,0x66,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,
     0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x58,0x4c,0x41,0x20,0x30,0x78,0x37,0x32,0x31,0x32,0x31,0x31,0x66,0x66,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,
     0x46,0x41,0x4d,0x49,0x4c,0x59,0x5f,0x55,0x4e,0x4b,0x4e,0x4f,0x57,0x4e,0x20,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x46,0x41,0x4d,0x49,0x4c,0x59,0x5f,
     0x43,0x4e,0x20,0x30,0x78,0x36,0x33,0x31,0x35,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x46,0x41,0x4d,0x49,0x4c,0x59,0x5f,0x43,0x4e,0x5f,

src/backend/opencl/cl/rx/randomx_cl.h

@@ -23,19 +23,19 @@ static const char randomx_cl[131086] = {
     0x36,0x38,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x50,0x49,0x43,0x4f,0x5f,0x30,0x20,0x30,0x78,0x36,0x33,0x31,0x32,
     0x30,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x50,0x49,0x43,0x4f,0x5f,0x54,0x4c,0x4f,0x20,0x30,0x78,
     0x36,0x33,0x31,0x32,0x30,0x32,0x37,0x34,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x55,0x50,0x58,0x32,0x20,0x30,0x78,
-    0x36,0x33,0x31,0x31,0x30,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x30,0x20,0x30,0x78,0x37,0x32,0x31,
+    0x36,0x33,0x31,0x31,0x30,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x47,0x50,0x55,0x20,0x30,0x78,0x36,
-    0x35,0x31,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x57,0x4f,0x57,0x20,0x30,0x78,0x37,0x32,0x31,0x34,
+    0x33,0x31,0x35,0x30,0x33,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x30,0x20,0x30,0x78,0x37,0x32,0x31,0x35,
-    0x31,0x31,0x37,0x37,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x41,0x52,0x51,0x4d,0x41,0x20,0x30,0x78,0x37,0x32,0x31,
+    0x31,0x32,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x57,0x4f,0x57,0x20,0x30,0x78,0x37,0x32,0x31,0x34,0x31,
-    0x32,0x31,0x30,0x36,0x31,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x53,0x46,0x58,0x20,0x30,0x78,0x37,0x32,0x31,0x35,
+    0x31,0x37,0x37,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x41,0x52,0x51,0x4d,0x41,0x20,0x30,0x78,0x37,0x32,0x31,0x32,
-    0x31,0x32,0x37,0x33,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x4b,0x45,0x56,0x41,0x20,0x30,0x78,0x37,0x32,0x31,0x34,
+    0x31,0x30,0x36,0x31,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x53,0x46,0x58,0x20,0x30,0x78,0x37,0x32,0x31,0x35,0x31,
-    0x31,0x31,0x36,0x62,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x47,0x52,0x41,0x46,0x54,0x20,0x30,0x78,0x37,0x32,0x31,
+    0x32,0x37,0x33,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x4b,0x45,0x56,0x41,0x20,0x30,0x78,0x37,0x32,0x31,0x34,0x31,
-    0x35,0x31,0x32,0x36,0x37,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x43,0x48,0x55,0x4b,0x57,0x41,0x20,0x30,0x78,
+    0x31,0x36,0x62,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x47,0x52,0x41,0x46,0x54,0x20,0x30,0x78,0x37,0x32,0x31,0x35,
-    0x36,0x31,0x31,0x33,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x43,0x48,0x55,0x4b,0x57,0x41,
+    0x31,0x32,0x36,0x37,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x43,0x48,0x55,0x4b,0x57,0x41,0x20,0x30,0x78,0x36,
-    0x5f,0x56,0x32,0x20,0x30,0x78,0x36,0x31,0x31,0x34,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,
+    0x31,0x31,0x33,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x43,0x48,0x55,0x4b,0x57,0x41,0x5f,
-    0x57,0x52,0x4b,0x5a,0x20,0x30,0x78,0x36,0x31,0x31,0x32,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x53,0x54,
+    0x56,0x32,0x20,0x30,0x78,0x36,0x31,0x31,0x34,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x52,0x32,0x5f,0x57,
-    0x52,0x4f,0x42,0x57,0x54,0x5f,0x44,0x45,0x52,0x4f,0x20,0x30,0x78,0x34,0x31,0x30,0x30,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,
+    0x52,0x4b,0x5a,0x20,0x30,0x78,0x36,0x31,0x31,0x32,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,0x4f,0x5f,0x41,0x53,0x54,0x52,
-    0x47,0x4f,0x5f,0x4b,0x41,0x57,0x50,0x4f,0x57,0x5f,0x52,0x56,0x4e,0x20,0x30,0x78,0x36,0x62,0x30,0x66,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,
+    0x4f,0x42,0x57,0x54,0x5f,0x44,0x45,0x52,0x4f,0x20,0x30,0x78,0x34,0x31,0x30,0x30,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x41,0x4c,0x47,
-    0x20,0x41,0x4c,0x47,0x4f,0x5f,0x43,0x4e,0x5f,0x47,0x50,0x55,0x20,0x30,0x78,0x36,0x33,0x31,0x35,0x30,0x30,0x66,0x66,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,
+    0x4f,0x5f,0x4b,0x41,0x57,0x50,0x4f,0x57,0x5f,0x52,0x56,0x4e,0x20,0x30,0x78,0x36,0x62,0x30,0x66,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,
     0x41,0x4c,0x47,0x4f,0x5f,0x52,0x58,0x5f,0x58,0x4c,0x41,0x20,0x30,0x78,0x37,0x32,0x31,0x32,0x31,0x31,0x66,0x66,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x46,
     0x41,0x4d,0x49,0x4c,0x59,0x5f,0x55,0x4e,0x4b,0x4e,0x4f,0x57,0x4e,0x20,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x46,0x41,0x4d,0x49,0x4c,0x59,0x5f,0x43,
     0x4e,0x20,0x30,0x78,0x36,0x33,0x31,0x35,0x30,0x30,0x30,0x30,0x0a,0x23,0x64,0x65,0x66,0x69,0x6e,0x65,0x20,0x46,0x41,0x4d,0x49,0x4c,0x59,0x5f,0x43,0x4e,0x5f,0x4c,

src/base/crypto/Algorithm.cpp

@@ -74,6 +74,10 @@ const char *Algorithm::kCN_PICO_TLO     = "cn-pico/tlo";
 const char *Algorithm::kCN_UPX2         = "cn/upx2";
 #endif
 
+#ifdef XMRIG_ALGO_CN_GPU
+const char *Algorithm::kCN_GPU          = "cn/gpu";
+#endif
+
 #ifdef XMRIG_ALGO_RANDOMX
 const char *Algorithm::kRX              = "rx";
 const char *Algorithm::kRX_0            = "rx/0";
@@ -106,10 +110,6 @@ const char* Algorithm::kGHOSTRIDER      = "ghostrider";
 const char* Algorithm::kGHOSTRIDER_RTM  = "ghostrider";
 #endif
 
-
-#ifdef XMRIG_ALGO_CN_GPU
-const char *Algorithm::kCN_GPU          = "cn/gpu";
-#endif
 #ifdef XMRIG_ALGO_RANDOMX
 const char *Algorithm::kRX_XLA          = "panthera";
 #endif
@@ -154,6 +154,10 @@ static const std::map<uint32_t, const char *> kAlgorithmNames = {
     ALGO_NAME(CN_UPX2),
 #   endif
 
+#   ifdef XMRIG_ALGO_CN_GPU
+    ALGO_NAME(CN_GPU),
+#   endif
+
 #   ifdef XMRIG_ALGO_RANDOMX
     ALGO_NAME(RX_0),
     ALGO_NAME(RX_WOW),
@@ -177,10 +181,6 @@ static const std::map<uint32_t, const char *> kAlgorithmNames = {
     ALGO_NAME(KAWPOW_RVN),
 #   endif
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    ALGO_NAME(CN_GPU),
-#   endif
-
 #   ifdef XMRIG_ALGO_RANDOMX
     ALGO_NAME(RX_XLA),
 #   endif
@@ -272,6 +272,11 @@ static const std::map<const char *, Algorithm::Id, aliasCompare> kAlgorithmAlias
                                     ALGO_ALIAS(CN_UPX2,         "cryptonight-upx/2"),
 #   endif
 
+#   ifdef XMRIG_ALGO_CN_GPU
+    ALGO_ALIAS_AUTO(CN_GPU),        ALGO_ALIAS(CN_GPU,          "cryptonight/gpu"),
+                                    ALGO_ALIAS(CN_GPU,          "cryptonight_gpu"),
+#   endif
+
 #   ifdef XMRIG_ALGO_RANDOMX
     ALGO_ALIAS_AUTO(RX_0),          ALGO_ALIAS(RX_0,            "randomx/0"),
                                     ALGO_ALIAS(RX_0,            "randomx/test"),
@@ -304,10 +309,6 @@ static const std::map<const char *, Algorithm::Id, aliasCompare> kAlgorithmAlias
     ALGO_ALIAS_AUTO(KAWPOW_RVN),    ALGO_ALIAS(KAWPOW_RVN,      "kawpow/rvn"),
 #   endif
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    ALGO_ALIAS_AUTO(CN_GPU),        ALGO_ALIAS(CN_GPU,          "cryptonight/gpu"),
-#   endif
-
 #   ifdef XMRIG_ALGO_RANDOMX
     ALGO_ALIAS_AUTO(RX_XLA),        ALGO_ALIAS(RX_XLA,          "Panthera"),
 #   endif
@@ -387,8 +388,9 @@ std::vector<xmrig::Algorithm> xmrig::Algorithm::all(const std::function<bool(con
         CN_HEAVY_0, CN_HEAVY_TUBE, CN_HEAVY_XHV,
         CN_PICO_0, CN_PICO_TLO,
         CN_UPX2,
-        CN_GPU, RX_XLA,
+        CN_GPU,
         RX_0, RX_WOW, RX_ARQ, RX_GRAFT, RX_SFX, RX_KEVA,
+        RX_XLA,
         AR2_CHUKWA, AR2_CHUKWA_V2, AR2_WRKZ,
         ASTROBWT_DERO,
         KAWPOW_RVN,

src/base/crypto/Algorithm.h

@@ -65,6 +65,7 @@ public:
         CN_PICO_0       = 0x63120200,   // "cn-pico"          CryptoNight-Pico
         CN_PICO_TLO     = 0x63120274,   // "cn-pico/tlo"      CryptoNight-Pico (TLO)
         CN_UPX2         = 0x63110200,   // "cn/upx2"          Uplexa (UPX2)
+        CN_GPU          = 0x63150300,   // "cn/gpu"           CryptoNight-GPU (Ryo).
         CN_GR_0         = 0x63130100,   // "cn/dark"          GhostRider
         CN_GR_1         = 0x63130101,   // "cn/dark-lite"     GhostRider
         CN_GR_2         = 0x63150102,   // "cn/fast"          GhostRider
@@ -84,7 +85,6 @@ public:
         ASTROBWT_DERO   = 0x41000000,   // "astrobwt"         AstroBWT (Dero)
         KAWPOW_RVN      = 0x6b0f0000,   // "kawpow/rvn"       KawPow (RVN)
 
-        CN_GPU          = 0x631500ff,   // "cn/gpu"           CryptoNight-GPU (Ryo).
         RX_XLA          = 0x721211ff,   // "panthera"         Panthera (Scala2).
     };
 
@@ -141,6 +141,10 @@ public:
     static const char *kCN_UPX2;
 #   endif
 
+#   ifdef XMRIG_ALGO_CN_GPU
+    static const char *kCN_GPU;
+#   endif
+
 #   ifdef XMRIG_ALGO_RANDOMX
     static const char *kRX;
     static const char *kRX_0;
@@ -168,10 +172,6 @@ public:
     static const char *kKAWPOW_RVN;
 #   endif
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    static const char *kCN_GPU;
-#   endif
-
 #   ifdef XMRIG_ALGO_RANDOMX
     static const char *kRX_XLA;
 #   endif

src/crypto/cn/CnAlgo.h

@@ -83,16 +83,16 @@ public:
             return CN_ITER / 8;
 #       endif
 
-#       ifdef XMRIG_ALGO_CN_GPU
-        case Algorithm::CN_GPU:
-            return 0xC000;
-#       endif
-
 #       ifdef XMRIG_ALGO_CN_FEMTO
         case Algorithm::CN_UPX2:
             return CN_ITER / 32;
 #       endif
 
+#       ifdef XMRIG_ALGO_CN_GPU
+        case Algorithm::CN_GPU:
+            return 0xC000;
+#       endif
+
         default:
             break;
         }
@@ -108,18 +108,18 @@ public:
         }
 #       endif
 
-#       ifdef XMRIG_ALGO_CN_GPU
-        if (algo == Algorithm::CN_GPU) {
-            return 0x1FFFC0;
-	}
-#       endif
-
 #       ifdef XMRIG_ALGO_CN_FEMTO
         if (algo == Algorithm::CN_UPX2) {
             return 0x1FFF0;
         }
 #       endif
 
+#       ifdef XMRIG_ALGO_CN_GPU
+        if (algo == Algorithm::CN_GPU) {
+            return 0x1FFFC0;
+        }
+#       endif
+
 #       ifdef XMRIG_ALGO_GHOSTRIDER
         if (algo == Algorithm::CN_GR_1) {
             return 0x3FFF0;
@@ -152,13 +152,13 @@ template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_ZLS>::iterations() con
 template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_PICO_0>::iterations() const       { return CN_ITER / 8; }
 template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_PICO_TLO>::iterations() const     { return CN_ITER / 8; }
 template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_CCX>::iterations() const          { return CN_ITER / 2; }
-template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_GPU>::iterations() const          { return 0xC000; }
 template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_UPX2>::iterations() const         { return CN_ITER / 32; }
+template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_GPU>::iterations() const          { return 0xC000; }
 
 
 template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_PICO_0>::mask() const             { return 0x1FFF0; }
-template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_GPU>::mask() const                { return 0x1FFFC0; }
 template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_UPX2>::mask() const               { return 0x1FFF0; }
+template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_GPU>::mask() const                { return 0x1FFFC0; }
 
 #ifdef XMRIG_ALGO_GHOSTRIDER
 template<> constexpr inline uint32_t CnAlgo<Algorithm::CN_GR_0>::iterations() const         { return CN_ITER / 4; }

src/crypto/cn/CnHash.cpp

@@ -290,6 +290,12 @@ xmrig::CnHash::CnHash()
     ADD_FN_ASM(Algorithm::CN_UPX2);
 #   endif
 
+#   ifdef XMRIG_ALGO_CN_GPU
+    m_map[Algorithm::CN_GPU] = new cn_hash_fun_array{};
+    m_map[Algorithm::CN_GPU]->data[AV_SINGLE][Assembly::NONE]      = cryptonight_single_hash_gpu<Algorithm::CN_GPU, false>;
+    m_map[Algorithm::CN_GPU]->data[AV_SINGLE_SOFT][Assembly::NONE] = cryptonight_single_hash_gpu<Algorithm::CN_GPU, true>;
+#   endif
+
 #   ifdef XMRIG_ALGO_ARGON2
     m_map[Algorithm::AR2_CHUKWA] = new cn_hash_fun_array{};
     m_map[Algorithm::AR2_CHUKWA]->data[AV_SINGLE][Assembly::NONE]         = argon2::single_hash<Algorithm::AR2_CHUKWA>;
@@ -310,12 +316,6 @@ xmrig::CnHash::CnHash()
     m_map[Algorithm::ASTROBWT_DERO]->data[AV_SINGLE_SOFT][Assembly::NONE] = astrobwt::single_hash<Algorithm::ASTROBWT_DERO>;
 #   endif
 
-#   ifdef XMRIG_ALGO_CN_GPU
-    m_map[Algorithm::CN_GPU] = new cn_hash_fun_array{};
-    m_map[Algorithm::CN_GPU]->data[AV_SINGLE][Assembly::NONE]      = cryptonight_single_hash_gpu<Algorithm::CN_GPU, false>;
-    m_map[Algorithm::CN_GPU]->data[AV_SINGLE_SOFT][Assembly::NONE] = cryptonight_single_hash_gpu<Algorithm::CN_GPU, true>;
-#   endif
-
 #   ifdef XMRIG_ALGO_GHOSTRIDER
     ADD_FN(Algorithm::CN_GR_0);
     ADD_FN(Algorithm::CN_GR_1);

src/crypto/cn/CryptoNight_test.h

@@ -370,22 +370,6 @@ const static uint8_t test_output_pico_tlo[256] = {
 #endif
 
 
-#ifdef XMRIG_ALGO_CN_GPU
-// "cn/gpu"
-const static uint8_t test_output_gpu[160] = {
-    0xE5, 0x5C, 0xB2, 0x3E, 0x51, 0x64, 0x9A, 0x59, 0xB1, 0x27, 0xB9, 0x6B, 0x51, 0x5F, 0x2B, 0xF7,
-    0xBF, 0xEA, 0x19, 0x97, 0x41, 0xA0, 0x21, 0x6C, 0xF8, 0x38, 0xDE, 0xD0, 0x6E, 0xFF, 0x82, 0xDF,
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
-};
-#endif
-
 #ifdef XMRIG_ALGO_CN_FEMTO
 // "cn/upx2"
 const static uint8_t test_output_femto_upx2[256] = {
@@ -403,6 +387,23 @@ const static uint8_t test_output_femto_upx2[256] = {
 #endif
 
 
+#ifdef XMRIG_ALGO_CN_GPU
+// "cn/gpu"
+const static uint8_t test_output_gpu[160] = {
+    0xE5, 0x5C, 0xB2, 0x3E, 0x51, 0x64, 0x9A, 0x59, 0xB1, 0x27, 0xB9, 0x6B, 0x51, 0x5F, 0x2B, 0xF7,
+    0xBF, 0xEA, 0x19, 0x97, 0x41, 0xA0, 0x21, 0x6C, 0xF8, 0x38, 0xDE, 0xD0, 0x6E, 0xFF, 0x82, 0xDF,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+    0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
+};
+#endif
+
+
 #ifdef XMRIG_ALGO_ARGON2
 // "argon2/chukwa"
 const static uint8_t argon2_chukwa_test_out[256] = {

src/crypto/cn/CryptoNight_x86.h

@@ -420,7 +420,6 @@ static NOINLINE void cn_implode_scratchpad(cryptonight_ctx *ctx)
 #   else
     constexpr bool IS_HEAVY = props.isHeavy();
 #   endif
-
     constexpr size_t N = (props.memory() / sizeof(__m128i)) / (props.half_mem() ? 2 : 1);
 
     __m128i xout0, xout1, xout2, xout3, xout4, xout5, xout6, xout7;
@@ -841,8 +840,11 @@ namespace xmrig {
 
 
 template<size_t MEM>
-void cn_explode_scratchpad_gpu(const uint8_t *input, uint8_t *output)
+static NOINLINE void cn_explode_scratchpad_gpu(cryptonight_ctx *ctx)
 {
+    const uint8_t* input = reinterpret_cast<const uint8_t*>(ctx->state);
+    uint8_t* output = reinterpret_cast<uint8_t*>(ctx->memory);
+
     constexpr size_t hash_size = 200; // 25x8 bytes
     alignas(16) uint64_t hash[25];
 
@@ -866,12 +868,12 @@ void cn_explode_scratchpad_gpu(const uint8_t *input, uint8_t *output)
 
 
 template<Algorithm::Id ALGO, bool SOFT_AES>
-inline void cryptonight_single_hash_gpu(const uint8_t *__restrict__ input, size_t size, uint8_t *__restrict__ output, cryptonight_ctx **__restrict__ ctx, uint64_t)
+inline void cryptonight_single_hash_gpu(const uint8_t *__restrict__ input, size_t size, uint8_t *__restrict__ output, cryptonight_ctx **__restrict__ ctx, uint64_t height)
 {
     constexpr CnAlgo<ALGO> props;
 
     keccak(input, size, ctx[0]->state);
-    cn_explode_scratchpad_gpu<props.memory()>(ctx[0]->state, ctx[0]->memory);
+    cn_explode_scratchpad_gpu<props.memory()>(ctx[0]);
 
 #   ifdef _MSC_VER
     _control87(RC_NEAR, MCW_RC);

src/crypto/cn/CryptoNight_x86_vaes.cpp

@@ -518,13 +518,13 @@ void (*vaes_instances[])() = {
     VAES_Instance<Algorithm::CN_PICO_0>,
     VAES_Instance<Algorithm::CN_PICO_TLO>,
     VAES_Instance<Algorithm::CN_UPX2>,
+    VAES_Instance<Algorithm::CN_GPU>,
     VAES_Instance<Algorithm::CN_GR_0>,
     VAES_Instance<Algorithm::CN_GR_1>,
     VAES_Instance<Algorithm::CN_GR_2>,
     VAES_Instance<Algorithm::CN_GR_3>,
     VAES_Instance<Algorithm::CN_GR_4>,
     VAES_Instance<Algorithm::CN_GR_5>,
-    VAES_Instance<Algorithm::CN_GPU>,
 };