From 5111264954e2d1a4d3e91d23a0869a0d7105be4c Mon Sep 17 00:00:00 2001
From: morgolock <pablo.tello@arm.com>
Date: Thu, 20 Aug 2020 14:51:39 +0100
Subject: COMPMID-3661: Added multidimension support to OMP scheduler.

Change-Id: Iedacf7094896f08d7c2847c8fb99bd7153deba2c
Signed-off-by: morgolock <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3809
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
---
 src/runtime/IScheduler.cpp | 117 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 117 insertions(+)

(limited to 'src/runtime/IScheduler.cpp')

diff --git a/src/runtime/IScheduler.cpp b/src/runtime/IScheduler.cpp
index 6b961d7dfc..53df3699b0 100644
--- a/src/runtime/IScheduler.cpp
+++ b/src/runtime/IScheduler.cpp
@@ -23,8 +23,11 @@
  */
 #include "arm_compute/runtime/IScheduler.h"
 
+#include "arm_compute/core/CPP/ICPPKernel.h"
 #include "arm_compute/core/Error.h"
+#include "arm_compute/core/Window.h"
 #include "arm_compute/runtime/CPUUtils.h"
+#include "arm_compute/runtime/SchedulerUtils.h"
 
 namespace arm_compute
 {
@@ -51,6 +54,120 @@ unsigned int IScheduler::num_threads_hint() const
 {
     return _num_threads_hint;
 }
+
+void IScheduler::schedule_common(ICPPKernel *kernel, const Hints &hints, ITensorPack &tensors)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(!kernel, "The child class didn't set the kernel");
+    ARM_COMPUTE_UNUSED(kernel);
+    ARM_COMPUTE_UNUSED(hints);
+    ARM_COMPUTE_UNUSED(tensors);
+#ifndef BARE_METAL
+    const Window &max_window = kernel->window();
+    if(hints.split_dimension() == IScheduler::split_dimensions_all)
+    {
+        /*
+         * if the split dim is size_t max then this signals we should parallelise over
+         * all dimensions
+         */
+        const std::size_t m = max_window.num_iterations(Window::DimX);
+        const std::size_t n = max_window.num_iterations(Window::DimY);
+
+        //in c++17 this can be swapped for   auto [ m_threads, n_threads ] = split_2d(...
+        unsigned m_threads, n_threads;
+        std::tie(m_threads, n_threads) = split_2d(this->num_threads(), m, n);
+
+        std::vector<IScheduler::Workload> workloads;
+        for(unsigned int ni = 0; ni != n_threads; ++ni)
+        {
+            for(unsigned int mi = 0; mi != m_threads; ++mi)
+            {
+                workloads.push_back(
+                    [ni, mi, m_threads, n_threads, &max_window, &kernel](const ThreadInfo & info)
+                {
+                    //narrow the window to our mi-ni workload
+                    Window win = max_window.split_window(Window::DimX, mi, m_threads)
+                                 .split_window(Window::DimY, ni, n_threads);
+
+                    win.validate();
+
+                    Window thread_locator;
+                    thread_locator.set(Window::DimX, Window::Dimension(mi, m_threads));
+                    thread_locator.set(Window::DimY, Window::Dimension(ni, n_threads));
+
+                    thread_locator.validate();
+
+                    kernel->run_nd(win, info, thread_locator);
+                });
+            }
+        }
+        run_workloads(workloads);
+    }
+    else
+    {
+        const unsigned int num_iterations = max_window.num_iterations(hints.split_dimension());
+        const unsigned int num_threads    = std::min(num_iterations, this->num_threads());
+
+        if(num_iterations == 0)
+        {
+            return;
+        }
+
+        if(!kernel->is_parallelisable() || num_threads == 1)
+        {
+            ThreadInfo info;
+            info.cpu_info = &_cpu_info;
+            if(tensors.empty())
+            {
+                kernel->run(max_window, info);
+            }
+            else
+            {
+                kernel->run_op(tensors, max_window, info);
+            }
+        }
+        else
+        {
+            unsigned int num_windows = 0;
+            switch(hints.strategy())
+            {
+                case StrategyHint::STATIC:
+                    num_windows = num_threads;
+                    break;
+                case StrategyHint::DYNAMIC:
+                {
+                    const unsigned int granule_threshold = (hints.threshold() <= 0) ? num_threads : static_cast<unsigned int>(hints.threshold());
+                    // Make sure we don't use some windows which are too small as this might create some contention on the ThreadFeeder
+                    num_windows = num_iterations > granule_threshold ? granule_threshold : 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, &tensors](const ThreadInfo & info)
+                {
+                    Window win = max_window.split_window(hints.split_dimension(), t, num_windows);
+                    win.validate();
+
+                    if(tensors.empty())
+                    {
+                        kernel->run(win, info);
+                    }
+                    else
+                    {
+                        kernel->run_op(tensors, win, info);
+                    }
+                };
+            }
+            run_workloads(workloads);
+        }
+    }
+#endif /* !BARE_METAL */
+}
+
 void IScheduler::run_tagged_workloads(std::vector<Workload> &workloads, const char *tag)
 {
     ARM_COMPUTE_UNUSED(tag);
-- 
cgit v1.2.1