diff options
Diffstat (limited to 'src')
-rw-r--r-- | src/runtime/CPP/CPPScheduler.cpp | 29 | ||||
-rw-r--r-- | src/runtime/CPP/SingleThreadScheduler.cpp | 4 | ||||
-rw-r--r-- | src/runtime/OMP/OMPScheduler.cpp | 6 |
3 files changed, 28 insertions, 11 deletions
diff --git a/src/runtime/CPP/CPPScheduler.cpp b/src/runtime/CPP/CPPScheduler.cpp index ab91b1071c..0da9892cb2 100644 --- a/src/runtime/CPP/CPPScheduler.cpp +++ b/src/runtime/CPP/CPPScheduler.cpp @@ -270,12 +270,12 @@ void CPPScheduler::run_workloads(std::vector<IScheduler::Workload> &workloads) } } -void CPPScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension) +void CPPScheduler::schedule(ICPPKernel *kernel, const Hints &hints) { ARM_COMPUTE_ERROR_ON_MSG(!kernel, "The child class didn't set the kernel"); const Window &max_window = kernel->window(); - const unsigned int num_iterations = max_window.num_iterations(split_dimension); + const unsigned int num_iterations = max_window.num_iterations(hints.split_dimension()); const unsigned int num_threads = std::min(num_iterations, _num_threads); if(num_iterations == 0) @@ -291,12 +291,29 @@ void CPPScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension) } else { - std::vector<IScheduler::Workload> workloads(num_threads); - for(unsigned int t = 0; t < num_threads; t++) + unsigned int num_windows = 0; + switch(hints.strategy()) { - workloads[t] = [&](const ThreadInfo & info) + case StrategyHint::STATIC: + num_windows = num_threads; + break; + case StrategyHint::DYNAMIC: { - Window win = max_window.split_window(split_dimension, info.thread_id, info.num_threads); + // Make sure we don't use some windows which are too small as this might create some contention on the ThreadFeeder + const unsigned int max_iterations = static_cast<unsigned int>(_num_threads) * 3; + num_windows = num_iterations > max_iterations ? max_iterations : num_iterations; + break; + } + default: + ARM_COMPUTE_ERROR("Unknown strategy"); + } + std::vector<IScheduler::Workload> workloads(num_windows); + for(unsigned int t = 0; t < num_windows; t++) + { + //Capture 't' by copy, all the other variables by reference: + workloads[t] = [t, &hints, &max_window, &num_windows, &kernel](const ThreadInfo & info) + { + Window win = max_window.split_window(hints.split_dimension(), t, num_windows); win.validate(); kernel->run(win, info); }; diff --git a/src/runtime/CPP/SingleThreadScheduler.cpp b/src/runtime/CPP/SingleThreadScheduler.cpp index 6099e2cab5..37011595fd 100644 --- a/src/runtime/CPP/SingleThreadScheduler.cpp +++ b/src/runtime/CPP/SingleThreadScheduler.cpp @@ -41,9 +41,9 @@ void SingleThreadScheduler::set_num_threads(unsigned int num_threads) ARM_COMPUTE_ERROR_ON(num_threads != 1); } -void SingleThreadScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension) +void SingleThreadScheduler::schedule(ICPPKernel *kernel, const Hints &hints) { - ARM_COMPUTE_UNUSED(split_dimension); + ARM_COMPUTE_UNUSED(hints); ThreadInfo info; info.cpu_info = &_cpu_info; kernel->run(kernel->window(), info); diff --git a/src/runtime/OMP/OMPScheduler.cpp b/src/runtime/OMP/OMPScheduler.cpp index 795c96caf0..43a8a6de54 100644 --- a/src/runtime/OMP/OMPScheduler.cpp +++ b/src/runtime/OMP/OMPScheduler.cpp @@ -56,7 +56,7 @@ void OMPScheduler::set_num_threads(unsigned int num_threads) _num_threads = (num_threads == 0) ? num_cores : num_threads; } -void OMPScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension) +void OMPScheduler::schedule(ICPPKernel *kernel, const Hints &hints) { ARM_COMPUTE_ERROR_ON_MSG(!kernel, "The child class didn't set the kernel"); @@ -64,7 +64,7 @@ void OMPScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension) info.cpu_info = &_cpu_info; const Window &max_window = kernel->window(); - const unsigned int num_iterations = max_window.num_iterations(split_dimension); + const unsigned int num_iterations = max_window.num_iterations(hints.split_dimension()); info.num_threads = std::min(num_iterations, _num_threads); if(!kernel->is_parallelisable() || info.num_threads == 1) @@ -76,7 +76,7 @@ void OMPScheduler::schedule(ICPPKernel *kernel, unsigned int split_dimension) #pragma omp parallel firstprivate(info) num_threads(info.num_threads) { const int tid = omp_get_thread_num(); - Window win = max_window.split_window(split_dimension, tid, info.num_threads); + Window win = max_window.split_window(hints.split_dimension(), tid, info.num_threads); info.thread_id = tid; kernel->run(win, info); } |