Merge branch 'opencl-shared-data' into evo

src/backend/opencl/OclBackend.cpp

@@ -35,6 +35,7 @@
 #include "backend/opencl/OclConfig.h"
 #include "backend/opencl/OclLaunchData.h"
 #include "backend/opencl/OclWorker.h"
+#include "backend/opencl/runners/tools/OclSharedState.h"
 #include "backend/opencl/wrappers/OclContext.h"
 #include "backend/opencl/wrappers/OclLib.h"
 #include "base/io/log/Log.h"
@@ -164,7 +165,7 @@ public:
     }
 
 
-    inline void start()
+    inline void start(const Job &job)
     {
         LOG_INFO("%s use profile " BLUE_BG(WHITE_BOLD_S " %s ") WHITE_BOLD_S " (" CYAN_BOLD("%zu") WHITE_BOLD(" threads)") " scratchpad " CYAN_BOLD("%zu KB"),
                  tag,
@@ -194,6 +195,8 @@ public:
                     i++;
         }
 
+        OclSharedState::start(threads, job);
+
         status.start(threads.size());
         workers.start(threads);
     }
@@ -329,7 +332,7 @@ void xmrig::OclBackend::setJob(const Job &job)
         return stop();
     }
 
-    if (!d_ptr->context.init(d_ptr->devices, threads, job)) {
+    if (!d_ptr->context.init(d_ptr->devices, threads)) {
         LOG_WARN("%s " RED_BOLD("disabled") YELLOW(" (OpenCL context unavailable)"), tag);
 
         return stop();
@@ -338,7 +341,7 @@ void xmrig::OclBackend::setJob(const Job &job)
     stop();
 
     d_ptr->threads = std::move(threads);
-    d_ptr->start();
+    d_ptr->start(job);
 }
 
 
@@ -371,6 +374,8 @@ void xmrig::OclBackend::stop()
     d_ptr->workers.stop();
     d_ptr->threads.clear();
 
+    OclSharedState::release();
+
     LOG_INFO("%s" YELLOW(" stopped") BLACK_BOLD(" (%" PRIu64 " ms)"), tag, Chrono::steadyMSecs() - ts);
 }
 

src/backend/opencl/OclConfig.cpp

@@ -193,32 +193,13 @@ std::vector<xmrig::OclLaunchData> xmrig::OclConfig::get(const Miner *miner, cons
             continue;
         }
 
-#       ifdef XMRIG_ALGO_RANDOMX
-        auto dataset = algorithm.family() == Algorithm::RANDOM_X ? std::make_shared<OclRxDataset>() : nullptr;
-#       endif
-
         if (thread.threads().size() > 1) {
-            auto interleave = std::make_shared<OclInterleave>(thread.threads().size());
-
             for (int64_t affinity : thread.threads()) {
-                OclLaunchData data(miner, algorithm, *this, platform, thread, devices[thread.index()], affinity);
-                data.interleave = interleave;
-
-#               ifdef XMRIG_ALGO_RANDOMX
-                data.dataset = dataset;
-#               endif
-
-                out.emplace_back(std::move(data));
+                out.emplace_back(miner, algorithm, *this, platform, thread, devices[thread.index()], affinity);
             }
         }
         else {
-            OclLaunchData data(miner, algorithm, *this, platform, thread, devices[thread.index()], thread.threads().front());
-
-#           ifdef XMRIG_ALGO_RANDOMX
-            data.dataset = dataset;
-#           endif
-
-            out.emplace_back(std::move(data));
+            out.emplace_back(miner, algorithm, *this, platform, thread, devices[thread.index()], thread.threads().front());
         }
     }
 

src/backend/opencl/OclLaunchData.h

@@ -27,19 +27,14 @@
 #define XMRIG_OCLLAUNCHDATA_H
 
 
-#include "backend/opencl/OclInterleave.h"
 #include "backend/opencl/OclThread.h"
+#include "backend/opencl/runners/tools/OclSharedData.h"
 #include "backend/opencl/wrappers/OclDevice.h"
 #include "backend/opencl/wrappers/OclPlatform.h"
 #include "crypto/common/Algorithm.h"
 #include "crypto/common/Nonce.h"
 
 
-#ifdef XMRIG_ALGO_RANDOMX
-#   include "backend/opencl/runners/tools/OclRxDataset.h"
-#endif
-
-
 using cl_context = struct _cl_context *;
 
 
@@ -72,11 +67,6 @@ public:
     const OclDevice device;
     const OclPlatform platform;
     const OclThread thread;
-    OclInterleavePtr interleave;
-
-#   ifdef XMRIG_ALGO_RANDOMX
-    OclRxDatasetPtr dataset;
-#   endif
 };
 
 

src/backend/opencl/OclWorker.cpp

@@ -28,6 +28,8 @@
 
 #include "backend/common/Tags.h"
 #include "backend/opencl/runners/OclCnRunner.h"
+#include "backend/opencl/runners/tools/OclSharedData.h"
+#include "backend/opencl/runners/tools/OclSharedState.h"
 #include "base/io/log/Log.h"
 #include "base/tools/Chrono.h"
 #include "core/Miner.h"
@@ -75,7 +77,7 @@ xmrig::OclWorker::OclWorker(size_t id, const OclLaunchData &data) :
     m_algorithm(data.algorithm),
     m_miner(data.miner),
     m_intensity(data.thread.intensity()),
-    m_interleave(data.interleave)
+    m_sharedData(OclSharedState::get(data.device.index()))
 {
     switch (m_algorithm.family()) {
     case Algorithm::RANDOM_X:
@@ -149,9 +151,7 @@ void xmrig::OclWorker::start()
 
     while (Nonce::sequence(Nonce::OPENCL) > 0) {
         if (!isReady()) {
-            if (m_interleave) {
-                m_interleave->setResumeCounter(0);
-            }
+            m_sharedData.setResumeCounter(0);
 
             do {
                 std::this_thread::sleep_for(std::chrono::milliseconds(200));
@@ -162,9 +162,7 @@ void xmrig::OclWorker::start()
                 break;
             }
 
-            if (m_interleave) {
-                m_interleave->resumeDelay(m_id);
-            }
+            m_sharedData.resumeDelay(m_id);
 
             if (!consumeJob()) {
                 return;
@@ -172,9 +170,7 @@ void xmrig::OclWorker::start()
         }
 
         while (!Nonce::isOutdated(Nonce::OPENCL, m_job.sequence())) {
-            if (m_interleave) {
-                m_interleave->adjustDelay(m_id);
-            }
+            m_sharedData.adjustDelay(m_id);
 
             const uint64_t t = Chrono::steadyMSecs();
 
@@ -233,9 +229,7 @@ void xmrig::OclWorker::storeStats(uint64_t t)
 
     m_count += m_intensity;
 
-    if (m_interleave) {
-        m_interleave->setRunTime(Chrono::steadyMSecs() - t);
-    }
+    m_sharedData.setRunTime(Chrono::steadyMSecs() - t);
 
     Worker::storeStats();
 }

src/backend/opencl/OclWorker.h

@@ -67,7 +67,7 @@ private:
     const Miner *m_miner;
     const uint32_t m_intensity;
     IOclRunner *m_runner = nullptr;
-    OclInterleavePtr m_interleave;
+    OclSharedData &m_sharedData;
     WorkerJob<1> m_job;
 };
 

src/backend/opencl/opencl.cmake

@@ -14,7 +14,6 @@ if (WITH_OPENCL)
         src/backend/opencl/OclCache.h
         src/backend/opencl/OclConfig.h
         src/backend/opencl/OclGenerator.h
-        src/backend/opencl/OclInterleave.h
         src/backend/opencl/OclLaunchData.h
         src/backend/opencl/OclThread.h
         src/backend/opencl/OclThreads.h
@@ -22,6 +21,8 @@ if (WITH_OPENCL)
         src/backend/opencl/runners/OclBaseRunner.h
         src/backend/opencl/runners/OclCnRunner.h
         src/backend/opencl/runners/tools/OclCnR.h
+        src/backend/opencl/runners/tools/OclSharedData.h
+        src/backend/opencl/runners/tools/OclSharedState.h
         src/backend/opencl/wrappers/OclContext.h
         src/backend/opencl/wrappers/OclDevice.h
         src/backend/opencl/wrappers/OclError.h
@@ -42,7 +43,6 @@ if (WITH_OPENCL)
         src/backend/opencl/OclBackend.cpp
         src/backend/opencl/OclCache.cpp
         src/backend/opencl/OclConfig.cpp
-        src/backend/opencl/OclInterleave.cpp
         src/backend/opencl/OclLaunchData.cpp
         src/backend/opencl/OclThread.cpp
         src/backend/opencl/OclThreads.cpp
@@ -50,6 +50,8 @@ if (WITH_OPENCL)
         src/backend/opencl/runners/OclBaseRunner.cpp
         src/backend/opencl/runners/OclCnRunner.cpp
         src/backend/opencl/runners/tools/OclCnR.cpp
+        src/backend/opencl/runners/tools/OclSharedData.cpp
+        src/backend/opencl/runners/tools/OclSharedState.cpp
         src/backend/opencl/wrappers/OclContext.cpp
         src/backend/opencl/wrappers/OclDevice.cpp
         src/backend/opencl/wrappers/OclError.cpp
@@ -78,7 +80,6 @@ if (WITH_OPENCL)
            src/backend/opencl/runners/OclRxBaseRunner.h
            src/backend/opencl/runners/OclRxJitRunner.h
            src/backend/opencl/runners/OclRxVmRunner.h
-           src/backend/opencl/runners/tools/OclRxDataset.h
            )
 
        list(APPEND SOURCES_BACKEND_OPENCL
@@ -95,7 +96,6 @@ if (WITH_OPENCL)
            src/backend/opencl/runners/OclRxBaseRunner.cpp
            src/backend/opencl/runners/OclRxJitRunner.cpp
            src/backend/opencl/runners/OclRxVmRunner.cpp
-           src/backend/opencl/runners/tools/OclRxDataset.cpp
            )
    endif()
 

src/backend/opencl/runners/OclBaseRunner.cpp

@@ -23,10 +23,11 @@
  */
 
 
+#include "backend/opencl/runners/OclBaseRunner.h"
 #include "backend/opencl/cl/OclSource.h"
 #include "backend/opencl/OclCache.h"
 #include "backend/opencl/OclLaunchData.h"
-#include "backend/opencl/runners/OclBaseRunner.h"
+#include "backend/opencl/runners/tools/OclSharedState.h"
 #include "backend/opencl/wrappers/OclError.h"
 #include "backend/opencl/wrappers/OclLib.h"
 #include "base/io/log/Log.h"
@@ -34,6 +35,9 @@
 #include "crypto/common/VirtualMemory.h"
 
 
+constexpr size_t oneGiB = 1024 * 1024 * 1024;
+
+
 xmrig::OclBaseRunner::OclBaseRunner(size_t id, const OclLaunchData &data) :
     m_algorithm(data.algorithm),
     m_ctx(data.ctx),
@@ -95,14 +99,15 @@ void xmrig::OclBaseRunner::init()
 {
     m_queue = OclLib::createCommandQueue(m_ctx, data().device.id());
 
-    constexpr size_t oneGiB = 1024 * 1024 * 1024;
-    size_t size             = bufferSize();
+    size_t size = align(bufferSize());
 
     if (size < oneGiB && data().device.vendorId() == OCL_VENDOR_AMD && data().device.freeMemSize() >= oneGiB) {
-        size = oneGiB;
+        m_buffer = OclSharedState::get(data().device.index()).createBuffer(m_ctx, size, m_offset);
+    }
+    else {
+        m_buffer = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, size);
     }
 
-    m_buffer = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, size);
     m_input  = createSubBuffer(CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY, Job::kMaxBlobSize);
     m_output = createSubBuffer(CL_MEM_READ_WRITE, sizeof(cl_uint) * 0x100);
 }

src/backend/opencl/runners/OclRxBaseRunner.cpp

@@ -30,6 +30,7 @@
 #include "backend/opencl/kernels/rx/FindSharesKernel.h"
 #include "backend/opencl/kernels/rx/HashAesKernel.h"
 #include "backend/opencl/OclLaunchData.h"
+#include "backend/opencl/runners/tools/OclSharedState.h"
 #include "backend/opencl/wrappers/OclLib.h"
 #include "base/net/stratum/Job.h"
 #include "crypto/rx/Rx.h"
@@ -75,6 +76,7 @@ xmrig::OclRxBaseRunner::~OclRxBaseRunner()
     OclLib::release(m_hashes);
     OclLib::release(m_rounding);
     OclLib::release(m_scratchpads);
+    OclLib::release(m_dataset);
 }
 
 
@@ -120,7 +122,7 @@ void xmrig::OclRxBaseRunner::set(const Job &job, uint8_t *blob)
         m_seed = job.seed();
 
         auto dataset = Rx::dataset(job, 0);
-        enqueueWriteBuffer(data().dataset->get(), CL_TRUE, 0, dataset->size(), dataset->raw());
+        enqueueWriteBuffer(m_dataset, CL_TRUE, 0, dataset->size(), dataset->raw());
     }
 
     if (job.size() < Job::kMaxBlobSize) {
@@ -177,4 +179,5 @@ void xmrig::OclRxBaseRunner::init()
     m_hashes      = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, 64 * m_intensity);
     m_entropy     = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, (128 + 2560) * m_intensity);
     m_rounding    = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, sizeof(uint32_t) * m_intensity);
+    m_dataset     = OclSharedState::get(data().device.index()).dataset();
 }

src/backend/opencl/runners/OclRxBaseRunner.h

@@ -62,6 +62,7 @@ protected:
     Blake2bHashRegistersKernel *m_blake2b_hash_registers_64 = nullptr;
     Blake2bInitialHashKernel *m_blake2b_initial_hash        = nullptr;
     Buffer m_seed;
+    cl_mem m_dataset                                        = nullptr;
     cl_mem m_entropy                                        = nullptr;
     cl_mem m_hashes                                         = nullptr;
     cl_mem m_rounding                                       = nullptr;

src/backend/opencl/runners/OclRxJitRunner.cpp

@@ -74,7 +74,7 @@ void xmrig::OclRxJitRunner::build()
     }
 
     m_randomx_run = new RxRunKernel(m_asmProgram);
-    m_randomx_run->setArgs(data().dataset->get(), m_scratchpads, m_registers, m_rounding, m_programs, m_intensity, m_algorithm);
+    m_randomx_run->setArgs(m_dataset, m_scratchpads, m_registers, m_rounding, m_programs, m_intensity, m_algorithm);
 }
 
 

src/backend/opencl/runners/OclRxVmRunner.cpp

@@ -70,7 +70,7 @@ void xmrig::OclRxVmRunner::build()
     m_init_vm->setArgs(m_entropy, m_vm_states, m_rounding);
 
     m_execute_vm = new ExecuteVmKernel(m_program);
-    m_execute_vm->setArgs(m_vm_states, m_rounding, m_scratchpads, data().dataset->get(), m_intensity);
+    m_execute_vm->setArgs(m_vm_states, m_rounding, m_scratchpads, m_dataset, m_intensity);
 }
 
 

src/backend/opencl/runners/tools/OclSharedData.cpp

@@ -23,17 +23,44 @@
  */
 
 
-#include "backend/opencl/OclInterleave.h"
+#include "backend/opencl/runners/tools/OclSharedData.h"
+#include "backend/opencl/wrappers/OclLib.h"
 #include "base/io/log/Log.h"
 #include "base/tools/Chrono.h"
+#include "crypto/rx/Rx.h"
+#include "crypto/rx/RxDataset.h"
 
 
+#include <algorithm>
 #include <cinttypes>
+#include <stdexcept>
 #include <thread>
 
 
-uint64_t xmrig::OclInterleave::adjustDelay(size_t id)
+constexpr size_t oneGiB = 1024 * 1024 * 1024;
+
+
+cl_mem xmrig::OclSharedData::createBuffer(cl_context context, size_t size, size_t &offset)
 {
+    std::lock_guard<std::mutex> lock(m_mutex);
+
+    offset += size * m_offset++;
+    size   = std::max(size * m_threads, oneGiB);
+
+    if (!m_buffer) {
+        m_buffer = OclLib::createBuffer(context, CL_MEM_READ_WRITE, size);
+    }
+
+    return OclLib::retain(m_buffer);
+}
+
+
+uint64_t xmrig::OclSharedData::adjustDelay(size_t id)
+{
+    if (m_threads < 2) {
+        return 0;
+    }
+
     const uint64_t t0 = Chrono::steadyMSecs();
     uint64_t delay    = 0;
 
@@ -69,8 +96,12 @@ uint64_t xmrig::OclInterleave::adjustDelay(size_t id)
 }
 
 
-uint64_t xmrig::OclInterleave::resumeDelay(size_t id)
+uint64_t xmrig::OclSharedData::resumeDelay(size_t id)
 {
+    if (m_threads < 2) {
+        return 0;
+    }
+
     uint64_t delay = 0;
 
     {
@@ -99,14 +130,28 @@ uint64_t xmrig::OclInterleave::resumeDelay(size_t id)
 }
 
 
-void xmrig::OclInterleave::setResumeCounter(uint32_t value)
+void xmrig::OclSharedData::release()
 {
+    OclLib::release(m_buffer);
+
+#   ifdef XMRIG_ALGO_RANDOMX
+    OclLib::release(m_dataset);
+#   endif
+}
+
+
+void xmrig::OclSharedData::setResumeCounter(uint32_t value)
+{
+    if (m_threads < 2) {
+        return;
+    }
+
     std::lock_guard<std::mutex> lock(m_mutex);
     m_resumeCounter = value;
 }
 
 
-void xmrig::OclInterleave::setRunTime(uint64_t time)
+void xmrig::OclSharedData::setRunTime(uint64_t time)
 {
     // averagingBias = 1.0 - only the last delta time is taken into account
     // averagingBias = 0.5 - the last delta time has the same weight as all the previous ones combined
@@ -116,3 +161,34 @@ void xmrig::OclInterleave::setRunTime(uint64_t time)
     std::lock_guard<std::mutex> lock(m_mutex);
     m_averageRunTime = m_averageRunTime * (1.0 - averagingBias) + time * averagingBias;
 }
+
+
+#ifdef XMRIG_ALGO_RANDOMX
+cl_mem xmrig::OclSharedData::dataset() const
+{
+    if (!m_dataset) {
+        throw std::runtime_error("RandomX dataset is not available");
+    }
+
+    return OclLib::retain(m_dataset);
+}
+
+
+void xmrig::OclSharedData::createDataset(cl_context ctx, const Job &job, bool host)
+{
+    if (m_dataset) {
+        return;
+    }
+
+    cl_int ret;
+
+    if (host) {
+        auto dataset = Rx::dataset(job, 0);
+
+        m_dataset = OclLib::createBuffer(ctx, CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR, RxDataset::maxSize(), dataset->raw(), &ret);
+    }
+    else {
+        m_dataset = OclLib::createBuffer(ctx, CL_MEM_READ_ONLY, RxDataset::maxSize(), nullptr, &ret);
+    }
+}
+#endif

src/backend/opencl/runners/tools/OclSharedData.h

@@ -22,42 +22,62 @@
  *   along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#ifndef XMRIG_OCLINTERLEAVE_H
-#define XMRIG_OCLINTERLEAVE_H
+#ifndef XMRIG_OCLSHAREDDATA_H
+#define XMRIG_OCLSHAREDDATA_H
 
 
 #include <memory>
 #include <mutex>
 
 
+using cl_context = struct _cl_context *;
+using cl_mem     = struct _cl_mem *;
+
+
 namespace xmrig {
 
 
-class OclInterleave
+class Job;
+
+
+class OclSharedData
 {
 public:
-    OclInterleave() = delete;
-    inline OclInterleave(size_t threads) : m_threads(threads) {}
+    OclSharedData() = default;
 
+    cl_mem createBuffer(cl_context context, size_t size, size_t &offset);
     uint64_t adjustDelay(size_t id);
     uint64_t resumeDelay(size_t id);
+    void release();
     void setResumeCounter(uint32_t value);
     void setRunTime(uint64_t time);
 
+    inline size_t threads() const       { return m_threads; }
+
+    inline OclSharedData &operator++()  { ++m_threads; return *this; }
+
+#   ifdef XMRIG_ALGO_RANDOMX
+    cl_mem dataset() const;
+    void createDataset(cl_context ctx, const Job &job, bool host);
+#   endif
+
 private:
-    const size_t m_threads;
+    cl_mem m_buffer           = nullptr;
     double m_averageRunTime   = 0.0;
     double m_threshold        = 0.95;
+    size_t m_offset           = 0;
+    size_t m_threads          = 0;
     std::mutex m_mutex;
     uint32_t m_resumeCounter  = 0;
     uint64_t m_timestamp      = 0;
+
+#   ifdef XMRIG_ALGO_RANDOMX
+    cl_mem m_dataset          = nullptr;
+#   endif
 };
 
 
-using OclInterleavePtr = std::shared_ptr<OclInterleave>;
-
-
 } /* namespace xmrig */
 
 
-#endif /* XMRIG_OCLINTERLEAVE_H */
+#endif /* XMRIG_OCLSHAREDDATA_H */

src/backend/opencl/runners/tools/OclSharedState.cpp

@@ -23,32 +23,52 @@
  */
 
 
-#include "backend/opencl/runners/tools/OclRxDataset.h"
-#include "backend/opencl/wrappers/OclLib.h"
-#include "crypto/rx/Rx.h"
-#include "crypto/rx/RxDataset.h"
+#include "backend/opencl/runners/tools/OclSharedState.h"
+#include "backend/opencl/runners/tools/OclSharedData.h"
 
 
-void xmrig::OclRxDataset::createBuffer(cl_context ctx, const Job &job, bool host)
+#include <cassert>
+#include <map>
+
+
+namespace xmrig {
+
+
+static std::map<uint32_t, OclSharedData> map;
+
+
+} // namespace xmrig
+
+
+xmrig::OclSharedData &xmrig::OclSharedState::get(uint32_t index)
 {
-    if (m_dataset) {
-        return;
-    }
-
-    cl_int ret;
-
-    if (host) {
-        auto dataset = Rx::dataset(job, 0);
-
-        m_dataset = OclLib::createBuffer(ctx, CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR, RxDataset::maxSize(), dataset->raw(), &ret);
-    }
-    else {
-        m_dataset = OclLib::createBuffer(ctx, CL_MEM_READ_ONLY, RxDataset::maxSize(), nullptr, &ret);
-    }
+    return map[index];
 }
 
 
-xmrig::OclRxDataset::~OclRxDataset()
+void xmrig::OclSharedState::release()
 {
-    OclLib::release(m_dataset);
+    for (auto &kv : map) {
+        kv.second.release();
+    }
+
+    map.clear();
+}
+
+
+void xmrig::OclSharedState::start(const std::vector<OclLaunchData> &threads, const Job &job)
+{
+    assert(map.empty());
+
+    for (const auto &data : threads) {
+        auto &sharedData = map[data.device.index()];
+
+        ++sharedData;
+
+#       ifdef XMRIG_ALGO_RANDOMX
+        if (data.algorithm.family() == Algorithm::RANDOM_X) {
+            sharedData.createDataset(data.ctx, job, data.thread.isDatasetHost());
+        }
+#       endif
+    }
 }

src/backend/opencl/runners/tools/OclSharedState.h

@@ -22,47 +22,26 @@
  *   along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#ifndef XMRIG_OCLRXDATASET_H
-#define XMRIG_OCLRXDATASET_H
+#ifndef XMRIG_OCLSHAREDSTATE_H
+#define XMRIG_OCLSHAREDSTATE_H
 
 
-#include "base/tools/Object.h"
-
-
-#include <memory>
-
-
-using cl_context = struct _cl_context *;
-using cl_mem     = struct _cl_mem *;
+#include "backend/opencl/OclLaunchData.h"
 
 
 namespace xmrig {
 
 
-class Job;
-
-
-class OclRxDataset
+class OclSharedState
 {
 public:
-    XMRIG_DISABLE_COPY_MOVE(OclRxDataset)
-
-    OclRxDataset() = default;
-    ~OclRxDataset();
-
-    inline cl_mem get() const { return m_dataset; }
-
-    void createBuffer(cl_context ctx,const Job &job, bool host);
-
-private:
-    cl_mem m_dataset = nullptr;
+    static OclSharedData &get(uint32_t index);
+    static void release();
+    static void start(const std::vector<OclLaunchData> &threads, const Job &job);
 };
 
 
-using OclRxDatasetPtr = std::shared_ptr<OclRxDataset>;
-
-
 } /* namespace xmrig */
 
 
-#endif /* XMRIG_OCLINTERLEAVE_H */
+#endif /* XMRIG_OCLSHAREDSTATE_H */

src/backend/opencl/wrappers/OclContext.cpp

@@ -24,6 +24,7 @@
 
 
 #include "backend/opencl/wrappers/OclContext.h"
+#include "backend/opencl/runners/tools/OclSharedState.h"
 #include "backend/opencl/wrappers/OclLib.h"
 
 
@@ -42,7 +43,7 @@ xmrig::OclContext::~OclContext()
 }
 
 
-bool xmrig::OclContext::init(const std::vector<OclDevice> &devices, std::vector<OclLaunchData> &threads, const Job &job)
+bool xmrig::OclContext::init(const std::vector<OclDevice> &devices, std::vector<OclLaunchData> &threads)
 {
     if (!m_ctx) {
         std::vector<cl_device_id> ids(devices.size());
@@ -59,12 +60,6 @@ bool xmrig::OclContext::init(const std::vector<OclDevice> &devices, std::vector<
 
     for (OclLaunchData &data : threads) {
         data.ctx = m_ctx;
-
-#       ifdef XMRIG_ALGO_RANDOMX
-        if (data.algorithm.family() == Algorithm::RANDOM_X) {
-            data.dataset->createBuffer(m_ctx, job, data.thread.isDatasetHost());
-        }
-#       endif
     }
 
     return true;

src/backend/opencl/wrappers/OclContext.h

@@ -49,7 +49,7 @@ public:
     OclContext(const OclDevice &device);
     ~OclContext();
 
-    bool init(const std::vector<OclDevice> &devices, std::vector<OclLaunchData> &threads, const Job &job);
+    bool init(const std::vector<OclDevice> &devices, std::vector<OclLaunchData> &threads);
 
     inline bool isValid() const     { return m_ctx != nullptr; }
     inline cl_context ctx() const   { return m_ctx; }

src/backend/opencl/wrappers/OclLib.cpp

@@ -73,6 +73,7 @@ static const char *kReleaseDevice                    = "clReleaseDevice";
 static const char *kReleaseKernel                    = "clReleaseKernel";
 static const char *kReleaseMemObject                 = "clReleaseMemObject";
 static const char *kReleaseProgram                   = "clReleaseProgram";
+static const char *kRetainMemObject                  = "clRetainMemObject";
 static const char *kRetainProgram                    = "clRetainProgram";
 static const char *kSetKernelArg                     = "clSetKernelArg";
 static const char *kSetMemObjectDestructorCallback   = "clSetMemObjectDestructorCallback";
@@ -106,6 +107,7 @@ typedef cl_int (CL_API_CALL *releaseDevice_t)(cl_device_id device);
 typedef cl_int (CL_API_CALL *releaseKernel_t)(cl_kernel);
 typedef cl_int (CL_API_CALL *releaseMemObject_t)(cl_mem);
 typedef cl_int (CL_API_CALL *releaseProgram_t)(cl_program);
+typedef cl_int (CL_API_CALL *retainMemObject_t)(cl_mem);
 typedef cl_int (CL_API_CALL *retainProgram_t)(cl_program);
 typedef cl_int (CL_API_CALL *setKernelArg_t)(cl_kernel, cl_uint, size_t, const void *);
 typedef cl_int (CL_API_CALL *setMemObjectDestructorCallback_t)(cl_mem, void (CL_CALLBACK *)(cl_mem, void *), void *);
@@ -148,6 +150,7 @@ static releaseDevice_t pReleaseDevice                                       = nu
 static releaseKernel_t pReleaseKernel                                       = nullptr;
 static releaseMemObject_t pReleaseMemObject                                 = nullptr;
 static releaseProgram_t pReleaseProgram                                     = nullptr;
+static retainMemObject_t pRetainMemObject                                   = nullptr;
 static retainProgram_t pRetainProgram                                       = nullptr;
 static setKernelArg_t pSetKernelArg                                         = nullptr;
 static setMemObjectDestructorCallback_t pSetMemObjectDestructorCallback     = nullptr;
@@ -239,6 +242,7 @@ bool xmrig::OclLib::load()
     DLSYM(SetMemObjectDestructorCallback);
     DLSYM(CreateSubBuffer);
     DLSYM(RetainProgram);
+    DLSYM(RetainMemObject);
 
 #   if defined(CL_VERSION_2_0)
     uv_dlsym(&oclLib, kCreateCommandQueueWithProperties, reinterpret_cast<void**>(&pCreateCommandQueueWithProperties));
@@ -670,6 +674,18 @@ cl_mem xmrig::OclLib::createSubBuffer(cl_mem buffer, cl_mem_flags flags, size_t
 }
 
 
+cl_mem xmrig::OclLib::retain(cl_mem memobj) noexcept
+{
+    assert(pRetainMemObject != nullptr);
+
+    if (memobj != nullptr) {
+        pRetainMemObject(memobj);
+    }
+
+    return memobj;
+}
+
+
 cl_program xmrig::OclLib::createProgramWithBinary(cl_context context, cl_uint num_devices, const cl_device_id *device_list, const size_t *lengths, const unsigned char **binaries, cl_int *binary_status, cl_int *errcode_ret) noexcept
 {
     assert(pCreateProgramWithBinary != nullptr);

src/backend/opencl/wrappers/OclLib.h

@@ -79,6 +79,7 @@ public:
     static cl_mem createBuffer(cl_context context, cl_mem_flags flags, size_t size, void *host_ptr, cl_int *errcode_ret) noexcept;
     static cl_mem createSubBuffer(cl_mem buffer, cl_mem_flags flags, size_t offset, size_t size, cl_int *errcode_ret) noexcept;
     static cl_mem createSubBuffer(cl_mem buffer, cl_mem_flags flags, size_t offset, size_t size);
+    static cl_mem retain(cl_mem memobj) noexcept;
     static cl_program createProgramWithBinary(cl_context context, cl_uint num_devices, const cl_device_id *device_list, const size_t *lengths, const unsigned char **binaries, cl_int *binary_status, cl_int *errcode_ret) noexcept;
     static cl_program createProgramWithSource(cl_context context, cl_uint count, const char **strings, const size_t *lengths, cl_int *errcode_ret) noexcept;
     static cl_program retain(cl_program program) noexcept;