Clear device force buffer if it was reallocated

The force buffer on the device is assumed to be clear at the start of
the step. This was not ensured if the buffer was realocated after e.g.
domain decomposition.

Change-Id: I4677f71edc9479d972a6d9471b8cbe2c377f8827

diff --git a/src/gromacs/mdtypes/state_propagator_data_gpu.h b/src/gromacs/mdtypes/state_propagator_data_gpu.h
index a0d9309f6714ec14ad949b613c54441e3a8e98e2..c6ae19c589c0101759ddd7b135f2161f282872e8 100644 (file)
--- a/src/gromacs/mdtypes/state_propagator_data_gpu.h
+++ b/src/gromacs/mdtypes/state_propagator_data_gpu.h
@@ -144,6 +144,8 @@ public:
     ~StatePropagatorDataGpu();
 
     /*! \brief Set the ranges for local and non-local atoms and reallocates buffers.
+     *
+     * Reallocates coordinate, velocities and force buffers on the device.
      *
      * \note
      * The coordinates buffer is (re)allocated, when required by PME, with a padding,
@@ -151,6 +153,10 @@ public:
      * is scheduled in the \p pmeStream_ (unlike the coordinates H2D) as only the PME
      * task uses this padding area.
      *
+     * \note
+     * The force buffer is cleared if its size increases, so that previously unused
+     * memory is cleared before forces are accumulated.
+     *
      *  \param[in] numAtomsLocal  Number of atoms in local domain.
      *  \param[in] numAtomsAll    Total number of atoms to handle.
      */

diff --git a/src/gromacs/mdtypes/state_propagator_data_gpu_impl.h b/src/gromacs/mdtypes/state_propagator_data_gpu_impl.h
index 70365ac62eba349060f56b5f89c8970cec324f5b..af073b6284dde35d568465192004b375be8b2dd0 100644 (file)
--- a/src/gromacs/mdtypes/state_propagator_data_gpu_impl.h
+++ b/src/gromacs/mdtypes/state_propagator_data_gpu_impl.h
@@ -137,6 +137,8 @@ public:
 
 
     /*! \brief Set the ranges for local and non-local atoms and reallocates buffers.
+     *
+     * Reallocates coordinate, velocities and force buffers on the device.
      *
      * \note
      * The coordinates buffer is (re)allocated, when required by PME, with a padding,
@@ -144,6 +146,10 @@ public:
      * is scheduled in the \p pmeStream_ (unlike the coordinates H2D) as only the PME
      * task uses this padding area.
      *
+     * \note
+     * The force buffer is cleared if its size increases, so that previously unused
+     * memory is cleared before forces are accumulated.
+     *
      *  \param[in] numAtomsLocal  Number of atoms in local domain.
      *  \param[in] numAtomsAll    Total number of atoms to handle.
      */

diff --git a/src/gromacs/mdtypes/state_propagator_data_gpu_impl_gpu.cpp b/src/gromacs/mdtypes/state_propagator_data_gpu_impl_gpu.cpp
index 2a8a394bdaf57232db08f773f39fb5e996ea4ff3..f42ad7230e9b220eaca5be3cae2ebb401fa8390f 100644 (file)
--- a/src/gromacs/mdtypes/state_propagator_data_gpu_impl_gpu.cpp
+++ b/src/gromacs/mdtypes/state_propagator_data_gpu_impl_gpu.cpp
@@ -194,7 +194,15 @@ void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
     }
 
     reallocateDeviceBuffer(&d_v_, DIM * numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
+    const int d_fOldCapacity = d_fCapacity_;
     reallocateDeviceBuffer(&d_f_, DIM * numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
+    // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
+    // the force accumulation stage before syncing with the local stream. Only done in CUDA,
+    // since the force buffer ops are not implemented in OpenCL.
+    if (GMX_GPU == GMX_GPU_CUDA && d_fCapacity_ != d_fOldCapacity)
+    {
+        clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, localStream_);
+    }
 }
 
 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)

diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda.cu b/src/gromacs/nbnxm/cuda/nbnxm_cuda.cu
index 45963daba6f4fca5121086d82501cdc7a93aaa42..04d0dfd3838fa29fd09171261ea4f42c5cd0b08d 100644 (file)
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda.cu
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda.cu
@@ -875,7 +875,12 @@ void nbnxn_gpu_x_to_nbat_x(const Nbnxm::Grid&        grid,
     nbnxnInsertNonlocalGpuDependency(nb, interactionLoc);
 }
 
-/* F buffer operations on GPU: performs force summations and conversion from nb to rvec format. */
+/* F buffer operations on GPU: performs force summations and conversion from nb to rvec format.
+ *
+ * NOTE: When the total force device buffer is reallocated and its size increases, it is cleared in
+ *       Local stream. Hence, if accumulateForce is true, NonLocal stream should start accumulating
+ *       forces only after Local stream already done so.
+ */
 void nbnxn_gpu_add_nbat_f_to_f(const AtomLocality                         atomLocality,
                                DeviceBuffer<float>                        totalForcesDevice,
                                gmx_nbnxn_gpu_t*                           nb,