fzorb/REDACTED-rig
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

KawPow: fixed OpenCL memory leak

Browse source
This commit is contained in:
SChernykh 2020-09-25 10:53:24 +02:00
parent e4929d7c06
commit bef9031b03
4 changed files with 30 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

5
src/backend/opencl/runners/OclKawPowRunner.cpp
View file

@ -69,8 +69,6 @@ OclKawPowRunner::~OclKawPowRunner()
delete m_calculateDagKernel; delete m_calculateDagKernel;
OclLib::release(m_searchKernel);
OclLib::release(m_controlQueue); OclLib::release(m_controlQueue);
OclLib::release(m_stop); OclLib::release(m_stop);
@ -120,8 +118,7 @@ void OclKawPowRunner::run(uint32_t nonce, uint32_t *hashOutput)
void OclKawPowRunner::set(const Job &job, uint8_t *blob) void OclKawPowRunner::set(const Job &job, uint8_t *blob)
{ {
m_blockHeight = static_cast<uint32_t>(job.height()); m_blockHeight = static_cast<uint32_t>(job.height());
m_searchProgram = OclKawPow::get(*this, m_blockHeight, m_workGroupSize); m_searchKernel = OclKawPow::get(*this, m_blockHeight, m_workGroupSize);
m_searchKernel = OclLib::createKernel(m_searchProgram, "progpow_search");
const uint32_t epoch = m_blockHeight / KPHash::EPOCH_LENGTH; const uint32_t epoch = m_blockHeight / KPHash::EPOCH_LENGTH;

1
src/backend/opencl/runners/OclKawPowRunner.h
View file

@ -69,7 +69,6 @@ private:
KawPow_CalculateDAGKernel* m_calculateDagKernel = nullptr; KawPow_CalculateDAGKernel* m_calculateDagKernel = nullptr;
cl_program m_searchProgram = nullptr;
cl_kernel m_searchKernel = nullptr; cl_kernel m_searchKernel = nullptr;
size_t m_workGroupSize = 256; size_t m_workGroupSize = 256;

45
src/backend/opencl/runners/tools/OclKawPow.cpp
View file

@ -54,8 +54,9 @@ namespace xmrig {
class KawPowCacheEntry class KawPowCacheEntry
{ {
public: public:
inline KawPowCacheEntry(const Algorithm &algo, uint64_t period, uint32_t worksize, uint32_t index, cl_program program) : inline KawPowCacheEntry(const Algorithm &algo, uint64_t period, uint32_t worksize, uint32_t index, cl_program program, cl_kernel kernel) :
program(program), program(program),
kernel(kernel),
m_algo(algo), m_algo(algo),
m_index(index), m_index(index),
m_period(period), m_period(period),
@ -65,9 +66,10 @@ public:
inline bool isExpired(uint64_t period) const { return m_period + 1 < period; } inline bool isExpired(uint64_t period) const { return m_period + 1 < period; }
inline bool match(const Algorithm &algo, uint64_t period, uint32_t worksize, uint32_t index) const { return m_algo == algo && m_period == period && m_worksize == worksize && m_index == index; } inline bool match(const Algorithm &algo, uint64_t period, uint32_t worksize, uint32_t index) const { return m_algo == algo && m_period == period && m_worksize == worksize && m_index == index; }
inline bool match(const IOclRunner &runner, uint64_t period, uint32_t worksize) const { return match(runner.algorithm(), period, worksize, runner.deviceIndex()); } inline bool match(const IOclRunner &runner, uint64_t period, uint32_t worksize) const { return match(runner.algorithm(), period, worksize, runner.deviceIndex()); }
inline void release() { OclLib::release(program); } inline void release() { OclLib::release(kernel); OclLib::release(program); }
cl_program program; cl_program program;
cl_kernel kernel;
private: private:
Algorithm m_algo; Algorithm m_algo;
@ -82,16 +84,16 @@ class KawPowCache
public: public:
KawPowCache() = default; KawPowCache() = default;
inline cl_program search(const IOclRunner &runner, uint64_t period, uint32_t worksize) { return search(runner.algorithm(), period, worksize, runner.deviceIndex()); } inline cl_kernel search(const IOclRunner &runner, uint64_t period, uint32_t worksize) { return search(runner.algorithm(), period, worksize, runner.deviceIndex()); }
inline cl_program search(const Algorithm &algo, uint64_t period, uint32_t worksize, uint32_t index) inline cl_kernel search(const Algorithm &algo, uint64_t period, uint32_t worksize, uint32_t index)
{ {
std::lock_guard<std::mutex> lock(m_mutex); std::lock_guard<std::mutex> lock(m_mutex);
for (const auto &entry : m_data) { for (const auto &entry : m_data) {
if (entry.match(algo, period, worksize, index)) { if (entry.match(algo, period, worksize, index)) {
return entry.program; return entry.kernel;
} }
} }
@ -99,9 +101,10 @@ public:
} }
void add(const Algorithm &algo, uint64_t period, uint32_t worksize, uint32_t index, cl_program program) void add(const Algorithm &algo, uint64_t period, uint32_t worksize, uint32_t index, cl_program program, cl_kernel kernel)
{ {
if (search(algo, period, worksize, index)) { if (search(algo, period, worksize, index)) {
OclLib::release(kernel);
OclLib::release(program); OclLib::release(program);
return; return;
} }
@ -109,7 +112,7 @@ public:
std::lock_guard<std::mutex> lock(m_mutex); std::lock_guard<std::mutex> lock(m_mutex);
gc(period); gc(period);
m_data.emplace_back(algo, period, worksize, index, program); m_data.emplace_back(algo, period, worksize, index, program, kernel);
} }
@ -159,15 +162,15 @@ static KawPowCache cache;
class KawPowBuilder class KawPowBuilder
{ {
public: public:
cl_program build(const IOclRunner &runner, uint64_t period, uint32_t worksize) cl_kernel build(const IOclRunner &runner, uint64_t period, uint32_t worksize)
{ {
std::lock_guard<std::mutex> lock(m_mutex); std::lock_guard<std::mutex> lock(m_mutex);
const uint64_t ts = Chrono::steadyMSecs(); const uint64_t ts = Chrono::steadyMSecs();
cl_program program = cache.search(runner, period, worksize); cl_kernel kernel = cache.search(runner, period, worksize);
if (program) { if (kernel) {
return program; return kernel;
} }
cl_int ret; cl_int ret;
@ -175,7 +178,7 @@ public:
cl_device_id device = runner.data().device.id(); cl_device_id device = runner.data().device.id();
const char *s = source.c_str(); const char *s = source.c_str();
program = OclLib::createProgramWithSource(runner.ctx(), 1, &s, nullptr, &ret); cl_program program = OclLib::createProgramWithSource(runner.ctx(), 1, &s, nullptr, &ret);
if (ret != CL_SUCCESS) { if (ret != CL_SUCCESS) {
return nullptr; return nullptr;
} }
@ -199,11 +202,17 @@ public:
return nullptr; return nullptr;
} }
kernel = OclLib::createKernel(program, "progpow_search", &ret);
if (ret != CL_SUCCESS) {
OclLib::release(program);
return nullptr;
}
LOG_INFO("%s " YELLOW("KawPow") " program for period " WHITE_BOLD("%" PRIu64) " compiled " BLACK_BOLD("(%" PRIu64 "ms)"), Tags::opencl(), period, Chrono::steadyMSecs() - ts); LOG_INFO("%s " YELLOW("KawPow") " program for period " WHITE_BOLD("%" PRIu64) " compiled " BLACK_BOLD("(%" PRIu64 "ms)"), Tags::opencl(), period, Chrono::steadyMSecs() - ts);
cache.add(runner.algorithm(), period, worksize, runner.deviceIndex(), program); cache.add(runner.algorithm(), period, worksize, runner.deviceIndex(), program, kernel);
return program; return kernel;
} }
@ -382,7 +391,7 @@ public:
static KawPowBuilder builder; static KawPowBuilder builder;
cl_program OclKawPow::get(const IOclRunner &runner, uint64_t height, uint32_t worksize) cl_kernel OclKawPow::get(const IOclRunner &runner, uint64_t height, uint32_t worksize)
{ {
const uint64_t period = height / KPHash::PERIOD_LENGTH; const uint64_t period = height / KPHash::PERIOD_LENGTH;
@ -396,9 +405,9 @@ cl_program OclKawPow::get(const IOclRunner &runner, uint64_t height, uint32_t wo
[](uv_work_t *req, int) { delete static_cast<KawPowBaton*>(req->data); } [](uv_work_t *req, int) { delete static_cast<KawPowBaton*>(req->data); }
); );
cl_program program = cache.search(runner, period, worksize); cl_kernel kernel = cache.search(runner, period, worksize);
if (program) { if (kernel) {
return program; return kernel;
} }
return builder.build(runner, period, worksize); return builder.build(runner, period, worksize);

4
src/backend/opencl/runners/tools/OclKawPow.h
View file

@ -30,7 +30,7 @@
#include <cstdint> #include <cstdint>
using cl_program = struct _cl_program *; using cl_kernel = struct _cl_kernel *;
namespace xmrig { namespace xmrig {
@ -42,7 +42,7 @@ class IOclRunner;
class OclKawPow class OclKawPow
{ {
public: public:
static cl_program get(const IOclRunner &runner, uint64_t height, uint32_t worksize); static cl_kernel get(const IOclRunner &runner, uint64_t height, uint32_t worksize);
static void clear(); static void clear();
}; };