Also initiate our L2 accesses on copies back to the CPU

1 files changed, 4 insertions, 0 deletions

diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index 5516b1e..2f166aa 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -420,6 +420,10 @@ void gpgpu_t::memcpy_from_gpu( void *dst, size_t src_start_addr, size_t count )
    unsigned char *dst_data = (unsigned char*)dst;
    for (unsigned n=0; n < count; n ++ ) 
       m_global_mem->read(src_start_addr+n,1,dst_data+n);
+
+   // Copy into the performance model.
+   extern gpgpu_sim* g_the_gpu; 
+   g_the_gpu->perf_memcpy_to_gpu(src_start_addr, count);
    if(g_debug_execution >= 3) {
       printf( " done.\n");
       fflush(stdout);