Move KernelTimer.H to ablastr/parallelization (#3863)

* move KernelTimer to ablastr

* add missing folder

* fixed bug

* fixed bug

* really fixed the bug

* kernelTimer -> KernelTimer

1 files changed, 89 insertions, 0 deletions

diff --git a/Source/ablastr/parallelization/KernelTimer.H b/Source/ablastr/parallelization/KernelTimer.H
new file mode 100644
index 000000000..0d11220da
--- /dev/null
+++ b/Source/ablastr/parallelization/KernelTimer.H
@@ -0,0 +1,89 @@
+/* Copyright 2019-2020 Michael Rowan, Axel Huebl, Kevin Gott
+ *
+ * This file is part of WarpX.
+ *
+ * License: BSD-3-Clause-LBNL
+ */
+#ifndef ABLASTR_KERNELTIMER_H_
+#define ABLASTR_KERNELTIMER_H_
+
+#include "ablastr/utils/TextMsg.H"
+
+#include <AMReX.H>
+#include <AMReX_BLassert.H>
+#include <AMReX_GpuAtomic.H>
+#include <AMReX_GpuQualifiers.H>
+#include <AMReX_REAL.H>
+
+#include <climits>
+
+namespace ablastr::parallelization
+{
+
+/**
+ * \brief Defines a timer object to be used on GPU; measures summed thread cycles.
+ */
+class KernelTimer
+{
+public:
+    /** Constructor.
+     * \param[in] do_timing Controls whether timer is active.
+     * \param[in,out] cost Pointer to cost which holds summed thread cycles
+     * (for performance, it is recommended to allocate pinned host memory).
+     */
+    AMREX_GPU_DEVICE
+    KernelTimer
+    (const bool do_timing, amrex::Real* cost)
+#if (defined AMREX_USE_GPU)
+        : m_do_timing(do_timing)
+#endif
+    {
+#if (defined AMREX_USE_GPU)
+    if (m_do_timing && cost) {
+        m_cost = cost;
+#   if defined(AMREX_USE_CUDA) || defined(AMREX_USE_HIP)
+        // Start the timer
+        m_wt = clock64();
+
+#   elif defined(AMREX_USE_DPCPP)
+        // To be updated
+        ABLASTR_ALWAYS_ASSERT_WITH_MESSAGE(m_do_timing == false,
+            "KernelTimer not yet supported for this hardware.");
+#   endif
+    }
+#else  // AMREX_USE_GPU
+        amrex::ignore_unused(do_timing, cost);
+#endif // AMREX_USE_GPU
+    }
+
+    //! Destructor.
+    AMREX_GPU_DEVICE
+    ~KernelTimer () {
+#if (defined AMREX_USE_GPU)
+        if (m_do_timing && m_cost) {
+#   if defined(AMREX_USE_CUDA) || defined(AMREX_USE_HIP)
+            m_wt = clock64() - m_wt;
+            amrex::Gpu::Atomic::Add( m_cost, amrex::Real(m_wt));
+#   elif defined(AMREX_USE_DPCPP)
+            // To be updated
+#   endif
+        }
+#endif //AMREX_USE_GPU
+    }
+
+#if (defined AMREX_USE_GPU)
+private:
+    //! Stores whether kernel timer is active.
+    bool m_do_timing;
+
+    //! Location in which to accumulate costs from all threads.
+    amrex::Real* m_cost = nullptr;
+
+    //! Store the time difference (cost) from a single thread.
+    long long int m_wt;
+#endif //AMREX_USE_GPU
+};
+
+} // namespace ablastr::parallelization
+
+#endif // ABLASTR_KERNELTIMER_H_