diff options
Diffstat (limited to 'src/gpgpu-sim/l2cache.h')
| -rw-r--r-- | src/gpgpu-sim/l2cache.h | 14 |
1 files changed, 14 insertions, 0 deletions
diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 2cc0e76..2d13918 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -72,6 +72,7 @@ public: void print_stat( FILE *fp ) { m_dram->print_stat(fp); } void visualize() const { m_dram->visualize(); } void print( FILE *fp ) const; + void handle_memcpy_to_gpu( size_t dst_start_addr, unsigned subpart_id, mem_access_sector_mask_t mask ); class memory_sub_partition * get_sub_partition(int sub_partition_id) { @@ -178,6 +179,12 @@ public: void accumulate_L2cache_stats(class cache_stats &l2_stats) const; void get_L2cache_sub_stats(struct cache_sub_stats &css) const; + void force_l2_tag_update(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask) + { + m_L2cache->force_tag_access( addr, m_memcpy_cycle_offset + time, mask ); + m_memcpy_cycle_offset += 1; + } + private: // data unsigned m_id; //< the global sub partition ID @@ -210,6 +217,13 @@ private: friend class L2interface; std::vector<mem_fetch*> breakdown_request_to_sector_requests(mem_fetch* mf); + + // This is a cycle offset that has to be applied to the l2 accesses to account for + // the cudamemcpy read/writes. We want GPGPU-Sim to only count cycles for kernel execution + // but we want cudamemcpy to go through the L2. Everytime an access is made from cudamemcpy + // this counter is incremented, and when the l2 is accessed (in both cudamemcpyies and otherwise) + // this value is added to the gpgpu-sim cycle counters. + unsigned m_memcpy_cycle_offset; }; class L2interface : public mem_fetch_interface { |