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

KawPow: tuned work group size for OpenCL

Browse source
This commit is contained in:
SChernykh 2020-05-28 10:58:06 +02:00
parent 22b937cc1c
commit e3d727cdb6
11 changed files with 162 additions and 117 deletions
Download patch file Download diff file Expand all files Collapse all files

22
src/backend/opencl/runners/OclKawPowRunner.cpp
View file

@ -47,9 +47,19 @@ namespace xmrig {
OclKawPowRunner::OclKawPowRunner(size_t index, const OclLaunchData &data) : OclBaseRunner(index, data)
{
switch (data.thread.worksize())
{
case 64:
case 128:
case 256:
case 512:
m_workGroupSize = data.thread.worksize();
break;
}
if (data.device.vendorId() == OclVendor::OCL_VENDOR_NVIDIA) {
m_options += " -DPLATFORM=OPENCL_PLATFORM_NVIDIA";
m_workGroupSize = 32;
m_dagWorkGroupSize = 32;
}
}
@ -69,9 +79,9 @@ OclKawPowRunner::~OclKawPowRunner()
void OclKawPowRunner::run(uint32_t nonce, uint32_t *hashOutput)
{
const size_t local_work_size = 128;
const size_t local_work_size = m_workGroupSize;
const size_t global_work_offset = nonce;
const size_t global_work_size = m_intensity - (m_intensity % local_work_size);
const size_t global_work_size = m_intensity - (m_intensity % m_workGroupSize);
enqueueWriteBuffer(m_input, CL_FALSE, 0, 40, m_blob);
@ -101,7 +111,7 @@ void OclKawPowRunner::run(uint32_t nonce, uint32_t *hashOutput)
void OclKawPowRunner::set(const Job &job, uint8_t *blob)
{
m_blockHeight = static_cast<uint32_t>(job.height());
m_searchProgram = OclKawPow::get(*this, m_blockHeight);
m_searchProgram = OclKawPow::get(*this, m_blockHeight, m_workGroupSize);
m_searchKernel = OclLib::createKernel(m_searchProgram, "progpow_search");
const uint32_t epoch = m_blockHeight / KPHash::EPOCH_LENGTH;
@ -138,11 +148,11 @@ void OclKawPowRunner::set(const Job &job, uint8_t *blob)
const uint32_t dag_words = dag_size / sizeof(node);
m_calculateDagKernel->setArgs(0, m_lightCache, m_dag, dag_words, m_lightCacheSize / sizeof(node));
constexpr uint32_t N = 1 << 18;
constexpr uint32_t N = 1 << 20;
for (uint32_t start = 0; start < dag_words; start += N) {
m_calculateDagKernel->setArg(0, sizeof(start), &start);
m_calculateDagKernel->enqueue(m_queue, N, m_workGroupSize);
m_calculateDagKernel->enqueue(m_queue, N, m_dagWorkGroupSize);
}
OclLib::finish(m_queue);