biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | inline | side by side
Make DeviceStream into a class
[alexxy/gromacs.git] / src / gromacs / nbnxm / opencl / nbnxm_ocl.cpp
diff --git a/src/gromacs/nbnxm/opencl/nbnxm_ocl.cpp b/src/gromacs/nbnxm/opencl/nbnxm_ocl.cpp
index 013bd093a01e3aff9f67f7706e199bedcf4aab2c..ba0c2ee93974aecee7408f2e3675400d4e019e6b 100644 (file)
--- a/src/gromacs/nbnxm/opencl/nbnxm_ocl.cpp
+++ b/src/gromacs/nbnxm/opencl/nbnxm_ocl.cpp
@@ -484,10 +484,10 @@ void gpu_copy_xq_to_gpu(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom, const Atom
     /* local/nonlocal offset and length used for xq and f */
     int adat_begin, adat_len;
 
-    cl_atomdata_t*   adat   = nb->atdat;
-    cl_plist_t*      plist  = nb->plist[iloc];
-    cl_timers_t*     t      = nb->timers;
-    cl_command_queue stream = nb->stream[iloc];
+    cl_atomdata_t*      adat         = nb->atdat;
+    cl_plist_t*         plist        = nb->plist[iloc];
+    cl_timers_t*        t            = nb->timers;
+    const DeviceStream& deviceStream = nb->deviceStreams[iloc];
 
     bool bDoTime = nb->bDoTime;
 
@@ -522,17 +522,17 @@ void gpu_copy_xq_to_gpu(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom, const Atom
     /* beginning of timed HtoD section */
     if (bDoTime)
     {
-        t->xf[atomLocality].nb_h2d.openTimingRegion(stream);
+        t->xf[atomLocality].nb_h2d.openTimingRegion(deviceStream);
     }
 
     /* HtoD x, q */
-    ocl_copy_H2D_async(adat->xq, nbatom->x().data() + adat_begin * 4,
-                       adat_begin * sizeof(float) * 4, adat_len * sizeof(float) * 4, stream,
+    ocl_copy_H2D_async(adat->xq, nbatom->x().data() + adat_begin * 4, adat_begin * sizeof(float) * 4,
+                       adat_len * sizeof(float) * 4, deviceStream.stream(),
                        bDoTime ? t->xf[atomLocality].nb_h2d.fetchNextEvent() : nullptr);
 
     if (bDoTime)
     {
-        t->xf[atomLocality].nb_h2d.closeTimingRegion(stream);
+        t->xf[atomLocality].nb_h2d.closeTimingRegion(deviceStream);
     }
 
     /* When we get here all misc operations issues in the local stream as well as
@@ -543,7 +543,7 @@ void gpu_copy_xq_to_gpu(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom, const Atom
         if (iloc == InteractionLocality::Local)
         {
             cl_int gmx_used_in_debug cl_error = clEnqueueMarkerWithWaitList(
-                    stream, 0, nullptr, &(nb->misc_ops_and_local_H2D_done));
+                    deviceStream.stream(), 0, nullptr, &(nb->misc_ops_and_local_H2D_done));
             GMX_ASSERT(cl_error == CL_SUCCESS,
                        ("clEnqueueMarkerWithWaitList failed: " + ocl_get_error_string(cl_error)).c_str());
 
@@ -551,13 +551,13 @@ void gpu_copy_xq_to_gpu(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom, const Atom
              * in the local stream in order to be able to sync with the above event
              * from the non-local stream.
              */
-            cl_error = clFlush(stream);
+            cl_error = clFlush(deviceStream.stream());
             GMX_ASSERT(cl_error == CL_SUCCESS,
                        ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
         }
         else
         {
-            sync_ocl_event(stream, &(nb->misc_ops_and_local_H2D_done));
+            sync_ocl_event(deviceStream.stream(), &(nb->misc_ops_and_local_H2D_done));
         }
     }
 }
@@ -583,11 +583,11 @@ void gpu_copy_xq_to_gpu(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom, const Atom
  */
 void gpu_launch_kernel(NbnxmGpu* nb, const gmx::StepWorkload& stepWork, const Nbnxm::InteractionLocality iloc)
 {
-    cl_atomdata_t*   adat   = nb->atdat;
-    cl_nbparam_t*    nbp    = nb->nbparam;
-    cl_plist_t*      plist  = nb->plist[iloc];
-    cl_timers_t*     t      = nb->timers;
-    cl_command_queue stream = nb->stream[iloc];
+    cl_atomdata_t*      adat         = nb->atdat;
+    cl_nbparam_t*       nbp          = nb->nbparam;
+    cl_plist_t*         plist        = nb->plist[iloc];
+    cl_timers_t*        t            = nb->timers;
+    const DeviceStream& deviceStream = nb->deviceStreams[iloc];
 
     bool bDoTime = nb->bDoTime;
 
@@ -628,14 +628,14 @@ void gpu_launch_kernel(NbnxmGpu* nb, const gmx::StepWorkload& stepWork, const Nb
     /* beginning of timed nonbonded calculation section */
     if (bDoTime)
     {
-        t->interaction[iloc].nb_k.openTimingRegion(stream);
+        t->interaction[iloc].nb_k.openTimingRegion(deviceStream);
     }
 
     /* kernel launch config */
 
     KernelLaunchConfig config;
     config.sharedMemorySize = calc_shmem_required_nonbonded(nbp->vdwtype, nb->bPrefetchLjParam);
-    config.stream           = stream;
+    config.stream           = deviceStream.stream();
     config.blockSize[0]     = c_clSize;
     config.blockSize[1]     = c_clSize;
     config.gridSize[0]      = plist->nsci;
@@ -686,7 +686,7 @@ void gpu_launch_kernel(NbnxmGpu* nb, const gmx::StepWorkload& stepWork, const Nb
 
     if (bDoTime)
     {
-        t->interaction[iloc].nb_k.closeTimingRegion(stream);
+        t->interaction[iloc].nb_k.closeTimingRegion(deviceStream);
     }
 }
 
@@ -722,12 +722,12 @@ static inline int calc_shmem_required_prune(const int num_threads_z)
  */
 void gpu_launch_kernel_pruneonly(NbnxmGpu* nb, const InteractionLocality iloc, const int numParts)
 {
-    cl_atomdata_t*   adat    = nb->atdat;
-    cl_nbparam_t*    nbp     = nb->nbparam;
-    cl_plist_t*      plist   = nb->plist[iloc];
-    cl_timers_t*     t       = nb->timers;
-    cl_command_queue stream  = nb->stream[iloc];
-    bool             bDoTime = nb->bDoTime;
+    cl_atomdata_t*      adat         = nb->atdat;
+    cl_nbparam_t*       nbp          = nb->nbparam;
+    cl_plist_t*         plist        = nb->plist[iloc];
+    cl_timers_t*        t            = nb->timers;
+    const DeviceStream& deviceStream = nb->deviceStreams[iloc];
+    bool                bDoTime      = nb->bDoTime;
 
     if (plist->haveFreshList)
     {
@@ -781,7 +781,7 @@ void gpu_launch_kernel_pruneonly(NbnxmGpu* nb, const InteractionLocality iloc, c
     /* beginning of timed prune calculation section */
     if (bDoTime)
     {
-        timer->openTimingRegion(stream);
+        timer->openTimingRegion(deviceStream);
     }
 
     /* Kernel launch config:
@@ -795,7 +795,7 @@ void gpu_launch_kernel_pruneonly(NbnxmGpu* nb, const InteractionLocality iloc, c
     /* kernel launch config */
     KernelLaunchConfig config;
     config.sharedMemorySize = calc_shmem_required_prune(num_threads_z);
-    config.stream           = stream;
+    config.stream           = deviceStream.stream();
     config.blockSize[0]     = c_clSize;
     config.blockSize[1]     = c_clSize;
     config.blockSize[2]     = num_threads_z;
@@ -840,7 +840,7 @@ void gpu_launch_kernel_pruneonly(NbnxmGpu* nb, const InteractionLocality iloc, c
 
     if (bDoTime)
     {
-        timer->closeTimingRegion(stream);
+        timer->closeTimingRegion(deviceStream);
     }
 }
 
@@ -861,10 +861,10 @@ void gpu_launch_cpyback(NbnxmGpu*                nb,
     /* determine interaction locality from atom locality */
     const InteractionLocality iloc = gpuAtomToInteractionLocality(aloc);
 
-    cl_atomdata_t*   adat    = nb->atdat;
-    cl_timers_t*     t       = nb->timers;
-    bool             bDoTime = nb->bDoTime;
-    cl_command_queue stream  = nb->stream[iloc];
+    cl_atomdata_t*      adat         = nb->atdat;
+    cl_timers_t*        t            = nb->timers;
+    bool                bDoTime      = nb->bDoTime;
+    const DeviceStream& deviceStream = nb->deviceStreams[iloc];
 
     /* don't launch non-local copy-back if there was no non-local work to do */
     if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
@@ -886,24 +886,24 @@ void gpu_launch_cpyback(NbnxmGpu*                nb,
     /* beginning of timed D2H section */
     if (bDoTime)
     {
-        t->xf[aloc].nb_d2h.openTimingRegion(stream);
+        t->xf[aloc].nb_d2h.openTimingRegion(deviceStream);
     }
 
     /* With DD the local D2H transfer can only start after the non-local
        has been launched. */
     if (iloc == InteractionLocality::Local && nb->bNonLocalStreamActive)
     {
-        sync_ocl_event(stream, &(nb->nonlocal_done));
+        sync_ocl_event(deviceStream.stream(), &(nb->nonlocal_done));
     }
 
     /* DtoH f */
     ocl_copy_D2H_async(nbatom->out[0].f.data() + adat_begin * DIM, adat->f,
                        adat_begin * DIM * sizeof(nbatom->out[0].f[0]),
-                       adat_len * DIM * sizeof(nbatom->out[0].f[0]), stream,
+                       adat_len * DIM * sizeof(nbatom->out[0].f[0]), deviceStream.stream(),
                        bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
 
     /* kick off work */
-    cl_error = clFlush(stream);
+    cl_error = clFlush(deviceStream.stream());
     GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
 
     /* After the non-local D2H is launched the nonlocal_done event can be
@@ -912,7 +912,7 @@ void gpu_launch_cpyback(NbnxmGpu*                nb,
        data back first. */
     if (iloc == InteractionLocality::NonLocal)
     {
-        cl_error = clEnqueueMarkerWithWaitList(stream, 0, nullptr, &(nb->nonlocal_done));
+        cl_error = clEnqueueMarkerWithWaitList(deviceStream.stream(), 0, nullptr, &(nb->nonlocal_done));
         GMX_ASSERT(cl_error == CL_SUCCESS,
                    ("clEnqueueMarkerWithWaitList failed: " + ocl_get_error_string(cl_error)).c_str());
         nb->bNonLocalStreamActive = CL_TRUE;
@@ -924,24 +924,25 @@ void gpu_launch_cpyback(NbnxmGpu*                nb,
         /* DtoH fshift when virial is needed */
         if (stepWork.computeVirial)
         {
-            ocl_copy_D2H_async(nb->nbst.fshift, adat->fshift, 0, SHIFTS * sizeof(nb->nbst.fshift[0]),
-                               stream, bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
+            ocl_copy_D2H_async(nb->nbst.fshift, adat->fshift, 0,
+                               SHIFTS * sizeof(nb->nbst.fshift[0]), deviceStream.stream(),
+                               bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
         }
 
         /* DtoH energies */
         if (stepWork.computeEnergy)
         {
-            ocl_copy_D2H_async(nb->nbst.e_lj, adat->e_lj, 0, sizeof(float), stream,
+            ocl_copy_D2H_async(nb->nbst.e_lj, adat->e_lj, 0, sizeof(float), deviceStream.stream(),
                                bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
 
-            ocl_copy_D2H_async(nb->nbst.e_el, adat->e_el, 0, sizeof(float), stream,
+            ocl_copy_D2H_async(nb->nbst.e_el, adat->e_el, 0, sizeof(float), deviceStream.stream(),
                                bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
         }
     }
 
     if (bDoTime)
     {
-        t->xf[aloc].nb_d2h.closeTimingRegion(stream);
+        t->xf[aloc].nb_d2h.closeTimingRegion(deviceStream);
     }
 }
 
Local GROMACS mirror with custom stuff