bench/Makefile.am

@@ -29,7 +29,4 @@ test_cublas_parallel_atrip_SOURCES = test-cublas-parallel-atrip.cxx
 
 bin_PROGRAMS += test-cuda-sanity
 test_cuda_sanity_SOURCES = test-cuda-sanity.cxx
-
-bin_PROGRAMS += test-cublas-parallel
-test_cublas_parallel_SOURCES = test-cublas-parallel.cxx
 endif

bench/test-cublas-parallel-atrip.cxx

@@ -158,8 +158,8 @@ int main(int  argc, char** argv) {
       MPI_Barrier(MPI_COMM_WORLD);
       timings["memcpy"].start();
       for (size_t _s = 0; _s < slices_d.size(); _s++) {
-        for (size_t _b = 0; _b < slices_h[_s].size(); _b++) {
+        // for (size_t _b = 0; _b < slices_h[_s].size(); _b++) {
-        // for (size_t _b = 0; _b < 1 ; _b++) {
+        for (size_t _b = 0; _b < 1 ; _b++) {
           auto device = (*slices_d[_s])[_b];
           auto host   = slices_h[_s][_b].data();
           cuMemcpyHtoD(device, host, sizes[_s]);

include/atrip/SliceUnion.hpp

@@ -208,14 +208,15 @@ template <typename F=double>
             auto dataPointer = freePointers.begin();
             freePointers.erase(dataPointer);
             blank.data = *dataPointer;
-            WITH_CHRONO("cuda:memcpy",
+            //
-            WITH_CHRONO("cuda:memcpy:self-sufficient",
+            //
-                        _CHECK_CUDA_SUCCESS("copying mpi data to device",
+            // TODO [#A]: do cuMemcpy of
-                                            cuMemcpyHtoD(blank.data,
+            //      sources[from.source].data() ⇒ blank.data
-                                                         (void*)sources[from.source].data(),
+            // Do this when everything else is working.
-                                                         sizeof(F) * sources[from.source].size()));
+            // This will probably be a bottleneck of the H-to-D communication,
-                        ))
+            // as most slices are SelfSufficient.
-
+            //
+            //
 #else
             blank.data = sources[from.source].data();
 #endif

include/atrip/Unions.hpp

@@ -52,8 +52,6 @@ namespace atrip {
                  , origin_.up.data()
                  , 1.0);
     memcpy(v.data(), toSlice.data, sizeof(F) * v.size());
-#else
-#  pragma message("WARNING: COMPILING WITHOUT SLICING THE TENSORS")
 #endif
 
   }