Use sub buffers.

src/backend/opencl/interfaces/IOclRunner.h

@@ -63,6 +63,9 @@ public:
     virtual uint32_t deviceIndex() const                    = 0;
     virtual void build()                                    = 0;
     virtual void init()                                     = 0;
+
+protected:
+    virtual size_t bufferSize() const                       = 0;
 };
 
 

src/backend/opencl/runners/OclBaseRunner.cpp

@@ -27,10 +27,11 @@
 #include "backend/opencl/OclCache.h"
 #include "backend/opencl/OclLaunchData.h"
 #include "backend/opencl/runners/OclBaseRunner.h"
+#include "backend/opencl/wrappers/OclError.h"
 #include "backend/opencl/wrappers/OclLib.h"
 #include "base/io/log/Log.h"
 #include "base/net/stratum/Job.h"
-#include "backend/opencl/wrappers/OclError.h"
+#include "crypto/common/VirtualMemory.h"
 
 
 xmrig::OclBaseRunner::OclBaseRunner(size_t id, const OclLaunchData &data) :
@@ -38,6 +39,7 @@ xmrig::OclBaseRunner::OclBaseRunner(size_t id, const OclLaunchData &data) :
     m_ctx(data.ctx),
     m_source(OclSource::get(data.algorithm)),
     m_data(data),
+    m_align(OclLib::getUint(data.device.id(), CL_DEVICE_MEM_BASE_ADDR_ALIGN)),
     m_threadId(id)
 {
     m_deviceKey = data.device.name();
@@ -61,10 +63,17 @@ xmrig::OclBaseRunner::~OclBaseRunner()
     OclLib::release(m_program);
     OclLib::release(m_input);
     OclLib::release(m_output);
+    OclLib::release(m_buffer);
     OclLib::release(m_queue);
 }
 
 
+size_t xmrig::OclBaseRunner::bufferSize() const
+{
+    return align(Job::kMaxBlobSize) + align(sizeof(cl_uint) * 0x100);
+}
+
+
 uint32_t xmrig::OclBaseRunner::deviceIndex() const
 {
     return data().thread.index();
@@ -84,8 +93,33 @@ void xmrig::OclBaseRunner::build()
 void xmrig::OclBaseRunner::init()
 {
     m_queue  = OclLib::createCommandQueue(m_ctx, data().device.id());
-    m_input  = OclLib::createBuffer(m_ctx, CL_MEM_READ_ONLY, Job::kMaxBlobSize);
+
-    m_output = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, sizeof(cl_uint) * 0x100);
+    constexpr size_t oneGiB = 1024 * 1024 * 1024;
+    size_t size             = bufferSize();
+
+    if (size < oneGiB && data().device.freeMemSize() >= oneGiB) {
+        size = oneGiB;
+    }
+
+    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);
+}
+
+
+cl_mem xmrig::OclBaseRunner::createSubBuffer(cl_mem_flags flags, size_t size)
+{
+    auto mem = OclLib::createSubBuffer(m_buffer, flags, m_offset, size);
+
+    m_offset += align(size);
+
+    return mem;
+}
+
+
+size_t xmrig::OclBaseRunner::align(size_t size) const
+{
+    return VirtualMemory::align(size, m_align);
 }
 
 

src/backend/opencl/runners/OclBaseRunner.h

@@ -57,11 +57,14 @@ protected:
     inline const OclLaunchData &data() const override   { return m_data; }
     inline size_t threadId() const override             { return m_threadId; }
 
+    size_t bufferSize() const override;
     uint32_t deviceIndex() const override;
     void build() override;
     void init() override;
 
 protected:
+    cl_mem createSubBuffer(cl_mem_flags flags, size_t size);
+    size_t align(size_t size) const;
     void enqueueReadBuffer(cl_mem buffer, cl_bool blocking_read, size_t offset, size_t size, void *ptr);
     void enqueueWriteBuffer(cl_mem buffer, cl_bool blocking_write, size_t offset, size_t size, const void *ptr);
     void finalize(uint32_t *hashOutput);
@@ -69,12 +72,15 @@ protected:
     Algorithm m_algorithm;
     cl_command_queue m_queue    = nullptr;
     cl_context m_ctx;
+    cl_mem m_buffer             = nullptr;
     cl_mem m_input              = nullptr;
     cl_mem m_output             = nullptr;
     cl_program m_program        = nullptr;
     const char *m_source;
     const OclLaunchData &m_data;
+    const size_t m_align;
     const size_t m_threadId;
+    size_t m_offset             = 0;
     std::string m_deviceKey;
     std::string m_options;
 };

src/backend/opencl/runners/OclCnRunner.cpp

@@ -81,6 +81,17 @@ xmrig::OclCnRunner::~OclCnRunner()
 }
 
 
+size_t xmrig::OclCnRunner::bufferSize() const
+{
+    const size_t g_thd = data().thread.intensity();
+
+    return OclBaseRunner::bufferSize() +
+           align(m_algorithm.l3() * g_thd) +
+           align(200 * g_thd) +
+           (align(sizeof(cl_uint) * (g_thd + 2)) * BRANCH_MAX);
+}
+
+
 void xmrig::OclCnRunner::run(uint32_t nonce, uint32_t *hashOutput)
 {
     static const cl_uint zero = 0;
@@ -167,10 +178,10 @@ void xmrig::OclCnRunner::init()
 
     const size_t g_thd = data().thread.intensity();
 
-    m_scratchpads = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, m_algorithm.l3() * g_thd);
+    m_scratchpads = createSubBuffer(CL_MEM_READ_WRITE, m_algorithm.l3() * g_thd);
-    m_states      = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, 200 * g_thd);
+    m_states      = createSubBuffer(CL_MEM_READ_WRITE, 200 * g_thd);
 
     for (size_t i = 0; i < BRANCH_MAX; ++i) {
-        m_branches[i] = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, sizeof(cl_uint) * (g_thd + 2));
+        m_branches[i] = createSubBuffer(CL_MEM_READ_WRITE, sizeof(cl_uint) * (g_thd + 2));
     }
 }

src/backend/opencl/runners/OclCnRunner.h

@@ -47,6 +47,7 @@ public:
     ~OclCnRunner() override;
 
 protected:
+    size_t bufferSize() const override;
     void run(uint32_t nonce, uint32_t *hashOutput) override;
     void set(const Job &job, uint8_t *blob) override;
     void build() override;

src/backend/opencl/runners/OclRxBaseRunner.cpp

@@ -136,6 +136,18 @@ void xmrig::OclRxBaseRunner::set(const Job &job, uint8_t *blob)
 }
 
 
+size_t xmrig::OclRxBaseRunner::bufferSize() const
+{
+    const size_t g_thd = data().thread.intensity();
+
+    return OclBaseRunner::bufferSize() +
+           align((m_algorithm.l3() + 64) * g_thd) +
+           align(64 * g_thd) +
+           align((128 + 2560) * g_thd) +
+           align(sizeof(uint32_t) * g_thd);
+}
+
+
 void xmrig::OclRxBaseRunner::build()
 {
     OclBaseRunner::build();
@@ -168,8 +180,8 @@ void xmrig::OclRxBaseRunner::init()
 
     const size_t g_thd = data().thread.intensity();
 
-    m_scratchpads = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, (m_algorithm.l3() + 64) * g_thd);
+    m_scratchpads = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, (m_algorithm.l3() + 64) * g_thd);
-    m_hashes      = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, 64 * g_thd);
+    m_hashes      = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, 64 * g_thd);
-    m_entropy     = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, (128 + 2560) * g_thd);
+    m_entropy     = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, (128 + 2560) * g_thd);
-    m_rounding    = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, sizeof(uint32_t) * g_thd);
+    m_rounding    = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, sizeof(uint32_t) * g_thd);
 }

src/backend/opencl/runners/OclRxBaseRunner.h

@@ -49,10 +49,11 @@ public:
     ~OclRxBaseRunner() override;
 
 protected:
-    void run(uint32_t nonce, uint32_t *hashOutput) override;
+    size_t bufferSize() const override;
-    void set(const Job &job, uint8_t *blob) override;
     void build() override;
     void init() override;
+    void run(uint32_t nonce, uint32_t *hashOutput) override;
+    void set(const Job &job, uint8_t *blob) override;
 
 protected:
     virtual void execute(uint32_t iteration) = 0;

src/backend/opencl/runners/OclRxJitRunner.cpp

@@ -52,6 +52,14 @@ xmrig::OclRxJitRunner::~OclRxJitRunner()
 }
 
 
+size_t xmrig::OclRxJitRunner::bufferSize() const
+{
+    const size_t g_thd = data().thread.intensity();
+
+    return OclRxBaseRunner::bufferSize() + align(256 * g_thd) + align(5120 * g_thd) + align(10048 * g_thd);
+}
+
+
 void xmrig::OclRxJitRunner::build()
 {
     OclRxBaseRunner::build();
@@ -92,9 +100,9 @@ void xmrig::OclRxJitRunner::init()
 
     const size_t g_thd = data().thread.intensity();
 
-    m_registers             = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, 256 * g_thd, nullptr);
+    m_registers             = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, 256 * g_thd);
-    m_intermediate_programs = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, 5120 * g_thd, nullptr);
+    m_intermediate_programs = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, 5120 * g_thd);
-    m_programs              = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, 10048 * g_thd, nullptr);
+    m_programs              = createSubBuffer(CL_MEM_READ_WRITE | CL_MEM_HOST_NO_ACCESS, 10048 * g_thd);
 }
 
 

src/backend/opencl/runners/OclRxJitRunner.h

@@ -45,6 +45,7 @@ public:
     ~OclRxJitRunner() override;
 
 protected:
+    size_t bufferSize() const override;
     void build() override;
     void execute(uint32_t iteration) override;
     void init() override;

src/backend/opencl/runners/OclRxVmRunner.cpp

@@ -47,6 +47,12 @@ xmrig::OclRxVmRunner::~OclRxVmRunner()
 }
 
 
+size_t xmrig::OclRxVmRunner::bufferSize() const
+{
+    return OclRxBaseRunner::bufferSize() + (align(2560 * data().thread.intensity()));
+}
+
+
 void xmrig::OclRxVmRunner::build()
 {
     OclRxBaseRunner::build();
@@ -94,5 +100,5 @@ void xmrig::OclRxVmRunner::init()
 {
     OclRxBaseRunner::init();
 
-    m_vm_states = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, 2560 * data().thread.intensity());
+    m_vm_states = createSubBuffer(CL_MEM_READ_WRITE, 2560 * data().thread.intensity());
 }

src/backend/opencl/runners/OclRxVmRunner.h

@@ -45,6 +45,7 @@ public:
     ~OclRxVmRunner() override;
 
 protected:
+    size_t bufferSize() const override;
     void build() override;
     void execute(uint32_t iteration) override;
     void init() override;

src/backend/opencl/runners/OclRyoRunner.cpp

@@ -61,6 +61,14 @@ xmrig::OclRyoRunner::~OclRyoRunner()
 }
 
 
+size_t xmrig::OclRyoRunner::bufferSize() const
+{
+    const size_t g_thd = data().thread.intensity();
+
+    return OclBaseRunner::bufferSize() + align(data().algorithm.l3() * g_thd) + align(200 * g_thd);
+}
+
+
 void xmrig::OclRyoRunner::run(uint32_t nonce, uint32_t *hashOutput)
 {
     static const cl_uint zero = 0;
@@ -123,6 +131,6 @@ void xmrig::OclRyoRunner::init()
 
     const size_t g_thd = data().thread.intensity();
 
-    m_scratchpads = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, data().algorithm.l3() * g_thd);
+    m_scratchpads = createSubBuffer(CL_MEM_READ_WRITE, data().algorithm.l3() * g_thd);
-    m_states      = OclLib::createBuffer(m_ctx, CL_MEM_READ_WRITE, 200 * g_thd);
+    m_states      = createSubBuffer(CL_MEM_READ_WRITE, 200 * g_thd);
 }

src/backend/opencl/runners/OclRyoRunner.h

@@ -48,6 +48,7 @@ public:
     ~OclRyoRunner() override;
 
 protected:
+    size_t bufferSize() const override;
     void run(uint32_t nonce, uint32_t *hashOutput) override;
     void set(const Job &job, uint8_t *blob) override;
     void build() override;

src/backend/opencl/wrappers/OclLib.cpp

@@ -51,6 +51,7 @@ static const char *kCreateContext                    = "clCreateContext";
 static const char *kCreateKernel                     = "clCreateKernel";
 static const char *kCreateProgramWithBinary          = "clCreateProgramWithBinary";
 static const char *kCreateProgramWithSource          = "clCreateProgramWithSource";
+static const char *kCreateSubBuffer                  = "clCreateSubBuffer";
 static const char *kEnqueueNDRangeKernel             = "clEnqueueNDRangeKernel";
 static const char *kEnqueueReadBuffer                = "clEnqueueReadBuffer";
 static const char *kEnqueueWriteBuffer               = "clEnqueueWriteBuffer";
@@ -72,6 +73,8 @@ static const char *kReleaseKernel                    = "clReleaseKernel";
 static const char *kReleaseMemObject                 = "clReleaseMemObject";
 static const char *kReleaseProgram                   = "clReleaseProgram";
 static const char *kSetKernelArg                     = "clSetKernelArg";
+static const char *kSetMemObjectDestructorCallback   = "clSetMemObjectDestructorCallback";
+static const char *kUnloadPlatformCompiler           = "clUnloadPlatformCompiler";
 
 
 #if defined(CL_VERSION_2_0)
@@ -102,8 +105,11 @@ 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 *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 *);
+typedef cl_int (CL_API_CALL *unloadPlatformCompiler_t)(cl_platform_id);
 typedef cl_kernel (CL_API_CALL *createKernel_t)(cl_program, const char *, cl_int *);
 typedef cl_mem (CL_API_CALL *createBuffer_t)(cl_context, cl_mem_flags, size_t, void *, cl_int *);
+typedef cl_mem (CL_API_CALL *createSubBuffer_t)(cl_mem, cl_mem_flags, cl_buffer_create_type, const void *, cl_int *);
 typedef cl_program (CL_API_CALL *createProgramWithBinary_t)(cl_context, cl_uint, const cl_device_id *, const size_t *, const unsigned char **, cl_int *, cl_int *);
 typedef cl_program (CL_API_CALL *createProgramWithSource_t)(cl_context, cl_uint, const char **, const size_t *, cl_int *);
 
@@ -118,6 +124,7 @@ static createContext_t pCreateContext                                       = nu
 static createKernel_t pCreateKernel                                         = nullptr;
 static createProgramWithBinary_t pCreateProgramWithBinary                   = nullptr;
 static createProgramWithSource_t pCreateProgramWithSource                   = nullptr;
+static createSubBuffer_t pCreateSubBuffer                                   = nullptr;
 static enqueueNDRangeKernel_t pEnqueueNDRangeKernel                         = nullptr;
 static enqueueReadBuffer_t pEnqueueReadBuffer                               = nullptr;
 static enqueueWriteBuffer_t pEnqueueWriteBuffer                             = nullptr;
@@ -139,6 +146,8 @@ static releaseKernel_t pReleaseKernel                                       = nu
 static releaseMemObject_t pReleaseMemObject                                 = nullptr;
 static releaseProgram_t pReleaseProgram                                     = nullptr;
 static setKernelArg_t pSetKernelArg                                         = nullptr;
+static setMemObjectDestructorCallback_t pSetMemObjectDestructorCallback     = nullptr;
+static unloadPlatformCompiler_t pUnloadPlatformCompiler                     = nullptr;
 
 #define DLSYM(x) if (uv_dlsym(&oclLib, k##x, reinterpret_cast<void**>(&p##x)) == -1) { return false; }
 
@@ -222,6 +231,9 @@ bool xmrig::OclLib::load()
     DLSYM(GetMemObjectInfo);
     DLSYM(GetContextInfo);
     DLSYM(ReleaseDevice);
+    DLSYM(UnloadPlatformCompiler);
+    DLSYM(SetMemObjectDestructorCallback);
+    DLSYM(CreateSubBuffer);
 
 #   if defined(CL_VERSION_2_0)
     uv_dlsym(&oclLib, kCreateCommandQueueWithProperties, reinterpret_cast<void**>(&pCreateCommandQueueWithProperties));
@@ -563,6 +575,12 @@ cl_int xmrig::OclLib::setKernelArg(cl_kernel kernel, cl_uint arg_index, size_t a
 }
 
 
+cl_int xmrig::OclLib::unloadPlatformCompiler(cl_platform_id platform) noexcept
+{
+    return pUnloadPlatformCompiler(platform);
+}
+
+
 cl_kernel xmrig::OclLib::createKernel(cl_program program, const char *kernel_name, cl_int *errcode_ret) noexcept
 {
     assert(pCreateKernel != nullptr);
@@ -619,6 +637,34 @@ cl_mem xmrig::OclLib::createBuffer(cl_context context, cl_mem_flags flags, size_
 }
 
 
+cl_mem xmrig::OclLib::createSubBuffer(cl_mem buffer, cl_mem_flags flags, size_t offset, size_t size, cl_int *errcode_ret) noexcept
+{
+    const cl_buffer_region region = { offset, size };
+
+    auto result = pCreateSubBuffer(buffer, flags, CL_BUFFER_CREATE_TYPE_REGION, &region, errcode_ret);
+    if (*errcode_ret != CL_SUCCESS) {
+        LOG_ERR("%s" RED(" error ") RED_BOLD("%s") RED(" when calling ") RED_BOLD("%s") RED(" with offset ") RED_BOLD("%zu") RED(" and size ") RED_BOLD("%zu"),
+                ocl_tag(), OclError::toString(*errcode_ret), kCreateSubBuffer, offset, size);
+
+        return nullptr;
+    }
+
+    return result;
+}
+
+
+cl_mem xmrig::OclLib::createSubBuffer(cl_mem buffer, cl_mem_flags flags, size_t offset, size_t size)
+{
+    cl_int ret;
+    cl_mem mem = createSubBuffer(buffer, flags, offset, size, &ret);
+    if (ret != CL_SUCCESS) {
+        throw std::runtime_error(OclError::toString(ret));
+    }
+
+    return mem;
+}
+
+
 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

@@ -72,10 +72,13 @@ public:
     static cl_int release(cl_mem mem_obj) noexcept;
     static cl_int release(cl_program program) noexcept;
     static cl_int setKernelArg(cl_kernel kernel, cl_uint arg_index, size_t arg_size, const void *arg_value) noexcept;
+    static cl_int unloadPlatformCompiler(cl_platform_id platform) noexcept;
     static cl_kernel createKernel(cl_program program, const char *kernel_name, cl_int *errcode_ret) noexcept;
     static cl_kernel createKernel(cl_program program, const char *kernel_name);
     static cl_mem createBuffer(cl_context context, cl_mem_flags flags, size_t size, void *host_ptr = nullptr);
     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_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_uint getNumPlatforms() noexcept;