Merge branch 'dev'

CHANGELOG.md

@@ -1,3 +1,9 @@
+# v6.3.0
+- [#1771](https://github.com/xmrig/xmrig/pull/1771) Adopted new SSE2NEON and reduced ARM-specific changes.
+- [#1774](https://github.com/xmrig/xmrig/pull/1774) RandomX: Added new option `cache_qos` in `randomx` object for cache QoS support.
+- [#1777](https://github.com/xmrig/xmrig/pull/1777) Added support for upcoming Haven offshore fork.
+  - [#1780](https://github.com/xmrig/xmrig/pull/1780) CryptoNight OpenCL: fix for long input data.
+
 # v6.2.3
 - [#1745](https://github.com/xmrig/xmrig/pull/1745) AstroBWT: fixed OpenCL compilation on some systems.
 - [#1749](https://github.com/xmrig/xmrig/pull/1749) KawPow: optimized CPU share verification.

src/backend/cpu/interfaces/ICpuInfo.h

@@ -62,6 +62,7 @@ public:
         FLAG_SSSE3,
         FLAG_XOP,
         FLAG_POPCNT,
+        FLAG_CAT_L3,
         FLAG_MAX
     };
 
@@ -79,6 +80,7 @@ public:
     virtual bool hasAVX2() const                                                    = 0;
     virtual bool hasBMI2() const                                                    = 0;
     virtual bool hasOneGbPages() const                                              = 0;
+    virtual bool hasCatL3() const                                                   = 0;
     virtual const char *backend() const                                             = 0;
     virtual const char *brand() const                                               = 0;
     virtual CpuThreads threads(const Algorithm &algorithm, uint32_t limit) const    = 0;

src/backend/cpu/platform/BasicCpuInfo.cpp

@@ -57,7 +57,7 @@
 namespace xmrig {
 
 
-static const std::array<const char *, ICpuInfo::FLAG_MAX> flagNames     = { "aes", "avx2", "avx512f", "bmi2", "osxsave", "pdpe1gb", "sse2", "ssse3", "xop", "popcnt" };
+static const std::array<const char *, ICpuInfo::FLAG_MAX> flagNames     = { "aes", "avx2", "avx512f", "bmi2", "osxsave", "pdpe1gb", "sse2", "ssse3", "xop", "popcnt", "cat_l3" };
 static const std::array<const char *, ICpuInfo::MSR_MOD_MAX> msrNames   = { "none", "ryzen", "intel", "custom" };
 
 
@@ -66,7 +66,7 @@ static inline void cpuid(uint32_t level, int32_t output[4])
     memset(output, 0, sizeof(int32_t) * 4);
 
 #   ifdef _MSC_VER
-    __cpuid(output, static_cast<int>(level));
+    __cpuidex(output, static_cast<int>(level), 0);
 #   else
     __cpuid_count(level, 0, output[0], output[1], output[2], output[3]);
 #   endif
@@ -143,6 +143,7 @@ static inline bool has_sse2()       { return has_feature(PROCESSOR_INFO,
 static inline bool has_ssse3()      { return has_feature(PROCESSOR_INFO,        ECX_Reg, 1 << 9); }
 static inline bool has_xop()        { return has_feature(0x80000001,            ECX_Reg, 1 << 11); }
 static inline bool has_popcnt()     { return has_feature(PROCESSOR_INFO,        ECX_Reg, 1 << 23); }
+static inline bool has_cat_l3()     { return has_feature(EXTENDED_FEATURES,     EBX_Reg, 1 << 15) && has_feature(0x10, EBX_Reg, 1 << 1); }
 
 
 } // namespace xmrig
@@ -178,6 +179,7 @@ xmrig::BasicCpuInfo::BasicCpuInfo() :
     m_flags.set(FLAG_SSSE3,   has_ssse3());
     m_flags.set(FLAG_XOP,     has_xop());
     m_flags.set(FLAG_POPCNT,  has_popcnt());
+    m_flags.set(FLAG_CAT_L3,  has_cat_l3());
 
 #   ifdef XMRIG_FEATURE_ASM
     if (hasAES()) {

src/backend/cpu/platform/BasicCpuInfo.h

@@ -51,6 +51,7 @@ protected:
     inline bool hasAVX2() const override            { return has(FLAG_AVX2); }
     inline bool hasBMI2() const override            { return has(FLAG_BMI2); }
     inline bool hasOneGbPages() const override      { return has(FLAG_PDPE1GB); }
+    inline bool hasCatL3() const override           { return has(FLAG_CAT_L3); }
     inline const char *brand() const override       { return m_brand; }
     inline MsrMod msrMod() const override           { return m_msrMod; }
     inline size_t cores() const override            { return 0; }

src/backend/opencl/cl/cn/cryptonight.cl

@@ -71,7 +71,7 @@ inline ulong getIdx()
 
 
 __attribute__((reqd_work_group_size(8, 8, 1)))
-__kernel void cn0(__global ulong *input, __global uint4 *Scratchpad, __global ulong *states, uint Threads)
+__kernel void cn0(__global ulong *input, int inlen, __global uint4 *Scratchpad, __global ulong *states, uint Threads)
 {
     uint ExpandedKey1[40];
     __local uint AES0[256], AES1[256], AES2[256], AES3[256];
@@ -109,34 +109,25 @@ __kernel void cn0(__global ulong *input, __global uint4 *Scratchpad, __global ul
         if (get_local_id(1) == 0) {
             __local ulong* State = State_buf + get_local_id(0) * 25;
 
-            ((__local ulong8 *)State)[0] = vload8(0, input);
+            #pragma unroll
-            State[8]  = input[8];
+            for (int i = 0; i < 25; ++i) {
-            State[9]  = input[9];
+                State[i] = 0;
-            State[10] = input[10];
+            }
-            State[11] = input[11];
-            State[12] = input[12];
-            State[13] = input[13];
-            State[14] = input[14];
-            State[15] = input[15];
 
+            // Input length must be a multiple of 136 and padded on the host side
+            for (int i = 0; inlen > 0; i += 17, inlen -= 136) {
+                #pragma unroll
+                for (int j = 0; j < 17; ++j) {
+                    State[j] ^= input[i + j];
+                }
+                if (i == 0) {
                     ((__local uint *)State)[9]  &= 0x00FFFFFFU;
                     ((__local uint *)State)[9]  |= (((uint)get_global_id(0)) & 0xFF) << 24;
                     ((__local uint *)State)[10] &= 0xFF000000U;
-            /* explicit cast to `uint` is required because some OpenCL implementations (e.g. NVIDIA)
-             * handle get_global_id and get_global_offset as signed long long int and add
-             * 0xFFFFFFFF... to `get_global_id` if we set on host side a 32bit offset where the first bit is `1`
-             * (even if it is correct casted to unsigned on the host)
-             */
                     ((__local uint *)State)[10] |= (((uint)get_global_id(0) >> 8));
-
-            // Last bit of padding
-            State[16] = 0x8000000000000000UL;
-
-            for (int i = 17; i < 25; ++i) {
-                State[i] = 0x00UL;
                 }
-
                 keccakf1600_2(State);
+            }
 
             #pragma unroll 1
             for (int i = 0; i < 25; ++i) {

src/backend/opencl/kernels/Cn0Kernel.cpp

@@ -38,10 +38,11 @@ void xmrig::Cn0Kernel::enqueue(cl_command_queue queue, uint32_t nonce, size_t th
 
 
 // __kernel void cn0(__global ulong *input, __global uint4 *Scratchpad, __global ulong *states, uint Threads)
-void xmrig::Cn0Kernel::setArgs(cl_mem input, cl_mem scratchpads, cl_mem states, uint32_t threads)
+void xmrig::Cn0Kernel::setArgs(cl_mem input, int inlen, cl_mem scratchpads, cl_mem states, uint32_t threads)
 {
     setArg(0, sizeof(cl_mem), &input);
-    setArg(1, sizeof(cl_mem), &scratchpads);
+    setArg(1, sizeof(int), &inlen);
-    setArg(2, sizeof(cl_mem), &states);
+    setArg(2, sizeof(cl_mem), &scratchpads);
-    setArg(3, sizeof(uint32_t), &threads);
+    setArg(3, sizeof(cl_mem), &states);
+    setArg(4, sizeof(uint32_t), &threads);
 }

src/backend/opencl/kernels/Cn0Kernel.h

@@ -38,7 +38,7 @@ public:
     inline Cn0Kernel(cl_program program) : OclKernel(program, "cn0") {}
 
     void enqueue(cl_command_queue queue, uint32_t nonce, size_t threads);
-    void setArgs(cl_mem input, cl_mem scratchpads, cl_mem states, uint32_t threads);
+    void setArgs(cl_mem input, int inlen, cl_mem scratchpads, cl_mem states, uint32_t threads);
 };
 
 

src/backend/opencl/runners/OclCnRunner.cpp

@@ -122,10 +122,16 @@ void xmrig::OclCnRunner::set(const Job &job, uint8_t *blob)
         throw std::length_error("job size too big");
     }
 
-    blob[job.size()] = 0x01;
+    const int inlen = static_cast<int>(job.size() + 136 - (job.size() % 136));
-    memset(blob + job.size() + 1, 0, Job::kMaxBlobSize - job.size() - 1);
 
-    enqueueWriteBuffer(m_input, CL_TRUE, 0, Job::kMaxBlobSize, blob);
+    blob[job.size()] = 0x01;
+    memset(blob + job.size() + 1, 0, inlen - job.size() - 1);
+
+    blob[inlen - 1] |= 0x80;
+
+    enqueueWriteBuffer(m_input, CL_TRUE, 0, inlen, blob);
+
+    m_cn0->setArg(1, sizeof(int), &inlen);
 
     if (m_algorithm == Algorithm::CN_R && m_height != job.height()) {
         delete m_cn1;
@@ -152,7 +158,7 @@ void xmrig::OclCnRunner::build()
     OclBaseRunner::build();
 
     m_cn0 = new Cn0Kernel(m_program);
-    m_cn0->setArgs(m_input, m_scratchpads, m_states, m_intensity);
+    m_cn0->setArgs(m_input, 0, m_scratchpads, m_states, m_intensity);
 
     m_cn2 = new Cn2Kernel(m_program);
     m_cn2->setArgs(m_scratchpads, m_states, m_branches, m_intensity);

src/base/io/json/JsonChain.cpp

@@ -36,9 +36,7 @@ static const rapidjson::Value kNullValue;
 }
 
 
-xmrig::JsonChain::JsonChain()
+xmrig::JsonChain::JsonChain() = default;
-{
-}
 
 
 bool xmrig::JsonChain::add(rapidjson::Document &&doc)
@@ -66,8 +64,10 @@ bool xmrig::JsonChain::addFile(const char *fileName)
     if (doc.HasParseError()) {
         const size_t offset = doc.GetErrorOffset();
 
-        size_t line, pos;
+        size_t line;
+        size_t pos;
         std::vector<std::string> s;
+
         if (Json::convertOffset(fileName, offset, line, pos, s)) {
             for (const auto& t : s) {
                 LOG_ERR("%s", t.c_str());

src/base/io/log/Tags.cpp

@@ -73,6 +73,16 @@ const char *xmrig::Tags::randomx()
 #endif
 
 
+#ifdef XMRIG_PROXY_PROJECT
+const char *xmrig::Tags::proxy()
+{
+    static const char *tag = MAGENTA_BG_BOLD(WHITE_BOLD_S " proxy   ");
+
+    return tag;
+}
+#endif
+
+
 #ifdef XMRIG_FEATURE_CUDA
 const char *xmrig::Tags::nvidia()
 {

src/base/io/log/Tags.h

@@ -42,6 +42,10 @@ public:
 #   endif
 #   endif
 
+#   ifdef XMRIG_PROXY_PROJECT
+    static const char *proxy();
+#   endif
+
 #   ifdef XMRIG_FEATURE_CUDA
     static const char *nvidia();
 #   endif

src/base/net/stratum/Job.h

@@ -45,7 +45,9 @@ class Job
 public:
     // Max blob size is 84 (75 fixed + 9 variable), aligned to 96. https://github.com/xmrig/xmrig/issues/1 Thanks fireice-uk.
     // SECOR increase requirements for blob size: https://github.com/xmrig/xmrig/issues/913
-    static constexpr const size_t kMaxBlobSize = 128;
+    // Haven (XHV) offshore increases requirements by adding pricing_record struct (192 bytes) to block_header.
+    // Round it up to 408 (136*3) for a convenient keccak calculation in OpenCL
+    static constexpr const size_t kMaxBlobSize = 408;
     static constexpr const size_t kMaxSeedSize = 32;
 
     Job() = default;

src/config.json

@@ -20,6 +20,7 @@
         "1gb-pages": false,
         "rdmsr": true,
         "wrmsr": true,
+        "cache_qos": false,
         "numa": true
     },
     "cpu": {

src/crypto/cn/CryptoNight_arm.h

@@ -33,7 +33,6 @@
 #include "crypto/cn/CryptoNight_monero.h"
 #include "crypto/cn/CryptoNight.h"
 #include "crypto/cn/soft_aes.h"
-#include "crypto/common/portable/mm_malloc.h"
 
 
 extern "C"
@@ -68,34 +67,6 @@ static inline void do_skein_hash(const uint8_t *input, size_t len, uint8_t *outp
 void (* const extra_hashes[4])(const uint8_t *, size_t, uint8_t *) = {do_blake_hash, do_groestl_hash, do_jh_hash, do_skein_hash};
 
 
-static inline __attribute__((always_inline)) __m128i _mm_set_epi64x(const uint64_t a, const uint64_t b)
-{
-    return vcombine_u64(vcreate_u64(b), vcreate_u64(a));
-}
-
-
-#if __ARM_FEATURE_CRYPTO
-static inline __attribute__((always_inline)) __m128i _mm_aesenc_si128(__m128i v, __m128i rkey)
-{
-    alignas(16) const __m128i zero = { 0 };
-    return veorq_u8(vaesmcq_u8(vaeseq_u8(v, zero)), rkey );
-}
-#else
-static inline __attribute__((always_inline)) __m128i _mm_aesenc_si128(__m128i v, __m128i rkey)
-{
-    alignas(16) const __m128i zero = { 0 };
-    return zero;
-}
-#endif
-
-
-/* this one was not implemented yet so here it is */
-static inline __attribute__((always_inline)) uint64_t _mm_cvtsi128_si64(__m128i a)
-{
-    return vgetq_lane_u64(a, 0);
-}
-
-
 #if defined (__arm64__) || defined (__aarch64__)
 static inline uint64_t __umul128(uint64_t a, uint64_t b, uint64_t* hi)
 {

src/crypto/cn/soft_aes.h

@@ -28,7 +28,7 @@
 
 
 #if defined(XMRIG_ARM)
-#   include "crypto/cn/SSE2NEON.h"
+#   include "crypto/cn/sse2neon.h"
 #elif defined(__GNUC__)
 #   include <x86intrin.h>
 #else

src/crypto/rx/Rx.cpp

@@ -28,6 +28,7 @@
 #include "crypto/rx/Rx.h"
 #include "backend/common/Tags.h"
 #include "backend/cpu/CpuConfig.h"
+#include "backend/cpu/CpuThreads.h"
 #include "base/io/log/Log.h"
 #include "base/io/log/Tags.h"
 #include "crypto/rx/RxConfig.h"
@@ -78,7 +79,7 @@ bool xmrig::Rx::init(const Job &job, const RxConfig &config, const CpuConfig &cp
     }
 
     if (!msrInitialized) {
-        msrInit(config);
+        msrInit(config, cpu.threads().get(job.algorithm()).data());
         msrInitialized = true;
     }
 
@@ -130,7 +131,7 @@ void xmrig::Rx::init(IRxListener *listener)
 
 
 #ifndef XMRIG_FEATURE_MSR
-void xmrig::Rx::msrInit(const RxConfig &)
+void xmrig::Rx::msrInit(const RxConfig &, const std::vector<CpuThread> &)
 {
 }
 

src/crypto/rx/Rx.h

@@ -30,6 +30,7 @@
 
 #include <cstdint>
 #include <utility>
+#include <vector>
 
 
 #include "crypto/common/HugePagesInfo.h"
@@ -41,6 +42,7 @@ namespace xmrig
 
 class Algorithm;
 class CpuConfig;
+class CpuThread;
 class IRxListener;
 class Job;
 class RxConfig;
@@ -62,7 +64,7 @@ public:
 #   endif
 
 private:
-    static void msrInit(const RxConfig &config);
+    static void msrInit(const RxConfig &config, const std::vector<CpuThread>& threads);
     static void msrDestroy();
     static void setupMainLoopExceptionFrame();
 };

src/crypto/rx/RxConfig.cpp

@@ -51,6 +51,7 @@ static const char *kMode        = "mode";
 static const char *kOneGbPages  = "1gb-pages";
 static const char *kRdmsr       = "rdmsr";
 static const char *kWrmsr       = "wrmsr";
+static const char *kCacheQoS    = "cache_qos";
 
 #ifdef XMRIG_FEATURE_HWLOC
 static const char *kNUMA        = "numa";
@@ -89,6 +90,8 @@ bool xmrig::RxConfig::read(const rapidjson::Value &value)
         readMSR(Json::getValue(value, kWrmsr));
 #       endif
 
+        m_cacheQoS = Json::getBool(value, kCacheQoS, m_cacheQoS);
+
 #       ifdef XMRIG_OS_LINUX
         m_oneGbPages = Json::getBool(value, kOneGbPages, m_oneGbPages);
 #       endif
@@ -151,6 +154,8 @@ rapidjson::Value xmrig::RxConfig::toJSON(rapidjson::Document &doc) const
     obj.AddMember(StringRef(kWrmsr), false, allocator);
 #   endif
 
+    obj.AddMember(StringRef(kCacheQoS), m_cacheQoS, allocator);
+
 #   ifdef XMRIG_FEATURE_HWLOC
     if (!m_nodeset.empty()) {
         Value numa(kArrayType);

src/crypto/rx/RxConfig.h

@@ -65,6 +65,7 @@ public:
     inline bool isOneGbPages() const    { return m_oneGbPages; }
     inline bool rdmsr() const           { return m_rdmsr; }
     inline bool wrmsr() const           { return m_wrmsr; }
+    inline bool cacheQoS() const        { return m_cacheQoS; }
     inline Mode mode() const            { return m_mode; }
 
 #   ifdef XMRIG_FEATURE_MSR
@@ -83,6 +84,8 @@ private:
     bool m_wrmsr = false;
 #   endif
 
+    bool m_cacheQoS = false;
+
     Mode readMode(const rapidjson::Value &value) const;
 
     bool m_numa         = true;

src/crypto/rx/Rx_linux.cpp

@@ -29,6 +29,7 @@
 
 #include "crypto/rx/Rx.h"
 #include "backend/cpu/Cpu.h"
+#include "backend/cpu/CpuThread.h"
 #include "base/io/log/Log.h"
 #include "base/tools/Chrono.h"
 #include "crypto/rx/RxConfig.h"
@@ -123,14 +124,15 @@ static bool wrmsr_on_cpu(uint32_t reg, uint32_t cpu, uint64_t value, uint64_t ma
 }
 
 
-static bool wrmsr_on_all_cpus(uint32_t reg, uint64_t value, uint64_t mask)
+template<typename T>
+static bool wrmsr_on_all_cpus(uint32_t reg, uint64_t value, uint64_t mask, T&& callback)
 {
     struct dirent **namelist;
     int dir_entries = scandir("/dev/cpu", &namelist, dir_filter, 0);
     int errors      = 0;
 
     while (dir_entries--) {
-        if (!wrmsr_on_cpu(reg, strtoul(namelist[dir_entries]->d_name, nullptr, 10), value, mask)) {
+        if (!callback(reg, strtoul(namelist[dir_entries]->d_name, nullptr, 10), value, mask)) {
             ++errors;
         }
 
@@ -159,7 +161,7 @@ static bool wrmsr_modprobe()
 }
 
 
-static bool wrmsr(const MsrItems &preset, bool save)
+static bool wrmsr(const MsrItems& preset, const std::vector<CpuThread>& threads, bool cache_qos, bool save)
 {
     if (!wrmsr_modprobe()) {
         return false;
@@ -177,12 +179,66 @@ static bool wrmsr(const MsrItems &preset, bool save)
     }
 
     for (const auto &i : preset) {
-        if (!wrmsr_on_all_cpus(i.reg(), i.value(), i.mask())) {
+        if (!wrmsr_on_all_cpus(i.reg(), i.value(), i.mask(), [](uint32_t reg, uint32_t cpu, uint64_t value, uint64_t mask) { return wrmsr_on_cpu(reg, cpu, value, mask); })) {
             return false;
         }
     }
 
+    const uint32_t n = Cpu::info()->threads();
+
+    // Which CPU cores will have access to the full L3 cache
+    std::vector<bool> cacheEnabled(n, false);
+    bool cacheQoSDisabled = threads.empty();
+
+    for (const CpuThread& t : threads) {
+        // If some thread has no affinity or wrong affinity, disable cache QoS
+        if ((t.affinity() < 0) || (t.affinity() >= n)) {
+            cacheQoSDisabled = true;
+            if (cache_qos) {
+                LOG_WARN(CLEAR "%s" YELLOW_BOLD_S "Cache QoS can only be enabled when all mining threads have affinity set", tag);
+            }
+            break;
+        }
+
+        cacheEnabled[t.affinity()] = true;
+    }
+
+    if (cache_qos && !Cpu::info()->hasCatL3()) {
+        if (!threads.empty()) {
+            LOG_WARN(CLEAR "%s" YELLOW_BOLD_S "This CPU doesn't support cat_l3, cache QoS is unavailable", tag);
+        }
+        cache_qos = false;
+    }
+
+    bool result = true;
+
+    if (cache_qos) {
+        result = wrmsr_on_all_cpus(0xC8F, 0, MsrItem::kNoMask, [&cacheEnabled, cacheQoSDisabled](uint32_t, uint32_t cpu, uint64_t, uint64_t) {
+            if (cacheQoSDisabled || (cpu >= cacheEnabled.size()) || cacheEnabled[cpu]) {
+                // Assign Class Of Service 0 to current CPU core (default, full L3 cache available)
+                if (!wrmsr_on_cpu(0xC8F, cpu, 0, MsrItem::kNoMask)) {
+                    return false;
+                }
+            }
+            else {
+                // Disable L3 cache for Class Of Service 1
+                if (!wrmsr_on_cpu(0xC91, cpu, 0, MsrItem::kNoMask)) {
+                    // Some CPUs don't let set it to all zeros
+                    if (!wrmsr_on_cpu(0xC91, cpu, 1, MsrItem::kNoMask)) {
+                        return false;
+                    }
+                }
+
+                // Assign Class Of Service 1 to current CPU core
+                if (!wrmsr_on_cpu(0xC8F, cpu, 1ULL << 32, MsrItem::kNoMask)) {
+                    return false;
+                }
+            }
             return true;
+        });
+    }
+
+    return result;
 }
 
 
@@ -216,7 +272,7 @@ void Rx::setMainLoopBounds(const std::pair<const void*, const void*>& bounds)
 } // namespace xmrig
 
 
-void xmrig::Rx::msrInit(const RxConfig &config)
+void xmrig::Rx::msrInit(const RxConfig &config, const std::vector<CpuThread>& threads)
 {
     const auto &preset = config.msrPreset();
     if (preset.empty()) {
@@ -225,7 +281,7 @@ void xmrig::Rx::msrInit(const RxConfig &config)
 
     const uint64_t ts = Chrono::steadyMSecs();
 
-    if (wrmsr(preset, config.rdmsr())) {
+    if (wrmsr(preset, threads, config.cacheQoS(), config.rdmsr())) {
         LOG_NOTICE(CLEAR "%s" GREEN_BOLD_S "register values for \"%s\" preset has been set successfully" BLACK_BOLD(" (%" PRIu64 " ms)"), tag, config.msrPresetName(), Chrono::steadyMSecs() - ts);
     }
     else {
@@ -242,7 +298,7 @@ void xmrig::Rx::msrDestroy()
 
     const uint64_t ts = Chrono::steadyMSecs();
 
-    if (!wrmsr(savedState, false)) {
+    if (!wrmsr(savedState, std::vector<CpuThread>(), true, false)) {
         LOG_ERR(CLEAR "%s" RED_BOLD_S "failed to restore initial state" BLACK_BOLD(" (%" PRIu64 " ms)"), tag, Chrono::steadyMSecs() - ts);
     }
 }

src/crypto/rx/Rx_win.cpp

@@ -30,6 +30,7 @@
 
 #include "crypto/rx/Rx.h"
 #include "backend/cpu/Cpu.h"
+#include "backend/cpu/CpuThread.h"
 #include "base/io/log/Log.h"
 #include "base/kernel/Platform.h"
 #include "base/tools/Chrono.h"
@@ -256,7 +257,7 @@ static bool wrmsr(HANDLE driver, uint32_t reg, uint64_t value, uint64_t mask)
 }
 
 
-static bool wrmsr(const MsrItems &preset, bool save)
+static bool wrmsr(const MsrItems &preset, const std::vector<CpuThread>& threads, bool cache_qos, bool save)
 {
     bool success = true;
 
@@ -282,14 +283,59 @@ static bool wrmsr(const MsrItems &preset, bool save)
         }
     }
 
-    std::thread wrmsr_thread([driver, &preset, &success]() {
+    const uint32_t n = Cpu::info()->threads();
-        for (uint32_t i = 0, n = Cpu::info()->threads(); i < n; ++i) {
+
+    // Which CPU cores will have access to the full L3 cache
+    std::vector<bool> cacheEnabled(n, false);
+    bool cacheQoSDisabled = threads.empty();
+
+    for (const CpuThread& t : threads) {
+        // If some thread has no affinity or wrong affinity, disable cache QoS
+        if ((t.affinity() < 0) || (t.affinity() >= n)) {
+            cacheQoSDisabled = true;
+            if (cache_qos) {
+                LOG_WARN(CLEAR "%s" YELLOW_BOLD_S "Cache QoS can only be enabled when all mining threads have affinity set", tag);
+            }
+            break;
+        }
+
+        cacheEnabled[t.affinity()] = true;
+    }
+
+    if (cache_qos && !Cpu::info()->hasCatL3()) {
+        if (!threads.empty()) {
+            LOG_WARN(CLEAR "%s" YELLOW_BOLD_S "This CPU doesn't support cat_l3, cache QoS is unavailable", tag);
+        }
+        cache_qos = false;
+    }
+
+    std::thread wrmsr_thread([n, driver, &preset, &cacheEnabled, cache_qos, cacheQoSDisabled, &success]() {
+        for (uint32_t i = 0; i < n; ++i) {
             if (!Platform::setThreadAffinity(i)) {
                 continue;
             }
 
             for (const auto &i : preset) {
-                success = wrmsr(driver, i.reg(), i.value(), i.mask());
+                success &= wrmsr(driver, i.reg(), i.value(), i.mask());
+            }
+
+            if (cache_qos) {
+                if (cacheQoSDisabled || cacheEnabled[i]) {
+                    // Assign Class Of Service 0 to current CPU core (default, full L3 cache available)
+                    success &= wrmsr(driver, 0xC8F, 0, MsrItem::kNoMask);
+                }
+                else {
+                    // Disable L3 cache for Class Of Service 1
+                    if (!wrmsr(driver, 0xC91, 0, MsrItem::kNoMask)) {
+                        // Some CPUs don't let set it to all zeros
+                        if (!wrmsr(driver, 0xC91, 1, MsrItem::kNoMask)) {
+                            success = false;
+                        }
+                    }
+
+                    // Assign Class Of Service 1 to current CPU core
+                    success &= wrmsr(driver, 0xC8F, 1ULL << 32, MsrItem::kNoMask);
+                }
             }
 
             if (!success) {
@@ -349,7 +395,7 @@ void Rx::setMainLoopBounds(const std::pair<const void*, const void*>& bounds)
 } // namespace xmrig
 
 
-void xmrig::Rx::msrInit(const RxConfig &config)
+void xmrig::Rx::msrInit(const RxConfig &config, const std::vector<CpuThread>& threads)
 {
     const auto &preset = config.msrPreset();
     if (preset.empty()) {
@@ -358,7 +404,7 @@ void xmrig::Rx::msrInit(const RxConfig &config)
 
     const uint64_t ts = Chrono::steadyMSecs();
 
-    if (wrmsr(preset, config.rdmsr())) {
+    if (wrmsr(preset, threads, config.cacheQoS(), config.rdmsr())) {
         LOG_NOTICE(CLEAR "%s" GREEN_BOLD_S "register values for \"%s\" preset has been set successfully" BLACK_BOLD(" (%" PRIu64 " ms)"), tag, config.msrPresetName(), Chrono::steadyMSecs() - ts);
     }
     else {
@@ -375,7 +421,7 @@ void xmrig::Rx::msrDestroy()
 
     const uint64_t ts = Chrono::steadyMSecs();
 
-    if (!wrmsr(savedState, false)) {
+    if (!wrmsr(savedState, std::vector<CpuThread>(), true, false)) {
         LOG_ERR(CLEAR "%s" RED_BOLD_S "failed to restore initial state" BLACK_BOLD(" (%" PRIu64 " ms)"), tag, Chrono::steadyMSecs() - ts);
     }
 }

src/version.h

@@ -28,15 +28,15 @@
 #define APP_ID        "xmrig"
 #define APP_NAME      "XMRig"
 #define APP_DESC      "XMRig miner"
-#define APP_VERSION   "6.2.3"
+#define APP_VERSION   "6.3.0-dev"
 #define APP_DOMAIN    "xmrig.com"
 #define APP_SITE      "www.xmrig.com"
 #define APP_COPYRIGHT "Copyright (C) 2016-2020 xmrig.com"
 #define APP_KIND      "miner"
 
 #define APP_VER_MAJOR  6
-#define APP_VER_MINOR  2
+#define APP_VER_MINOR  3
-#define APP_VER_PATCH  3
+#define APP_VER_PATCH  0
 
 #ifdef _MSC_VER
 #   if (_MSC_VER >= 1920)