diff options
Diffstat (limited to 'src')
| -rw-r--r-- | src/gpgpu-sim/gpu-cache.cc | 12 | ||||
| -rw-r--r-- | src/gpgpu-sim/gpu-cache.h | 37 |
2 files changed, 37 insertions, 12 deletions
diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 2cc75bb..46813e7 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -392,11 +392,11 @@ enum cache_request_status tag_array::access(new_addr_type addr, unsigned time, } void tag_array::fill(new_addr_type addr, unsigned time, mem_fetch *mf) { - fill(addr, time, mf->get_access_sector_mask()); + fill(addr, time, mf->get_access_sector_mask(), mf->get_access_byte_mask()); } void tag_array::fill(new_addr_type addr, unsigned time, - mem_access_sector_mask_t mask) { + mem_access_sector_mask_t mask, mem_access_byte_mask_t byte_mask) { // assert( m_config.m_alloc_policy == ON_FILL ); unsigned idx; enum cache_request_status status = probe(addr, idx, mask); @@ -410,12 +410,12 @@ void tag_array::fill(new_addr_type addr, unsigned time, ((sector_cache_block *)m_lines[idx])->allocate_sector(time, mask); } - m_lines[idx]->fill(time, mask); + m_lines[idx]->fill(time, mask, byte_mask); } void tag_array::fill(unsigned index, unsigned time, mem_fetch *mf) { assert(m_config.m_alloc_policy == ON_MISS); - m_lines[index]->fill(time, mf->get_access_sector_mask()); + m_lines[index]->fill(time, mf->get_access_sector_mask(), mf->get_access_byte_mask()); } // TODO: we need write back the flushed data to the upper level @@ -1432,6 +1432,7 @@ enum cache_request_status data_cache::wr_miss_wa_fetch_on_write( cache_block_t *block = m_tag_array->get_block(cache_index); block->set_modified_on_fill(true, mf->get_access_sector_mask()); + block->set_byte_mask_on_fill(true); events.push_back(cache_event(WRITE_ALLOCATE_SENT)); @@ -1483,8 +1484,7 @@ enum cache_request_status data_cache::wr_miss_wa_lazy_fetch_on_read( if (m_status == HIT_RESERVED) { block->set_ignore_on_fill(true, mf->get_access_sector_mask()); block->set_modified_on_fill(true, mf->get_access_sector_mask()); - } else { - block->set_status(MODIFIED, mf->get_access_sector_mask()); + block->set_byte_mask_on_fill(true); } if (mf->get_access_byte_mask().count() == m_config.get_atom_sz()) { diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index eb811d7..a84ddd1 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -121,7 +121,8 @@ struct cache_block_t { virtual void allocate(new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask) = 0; - virtual void fill(unsigned time, mem_access_sector_mask_t sector_mask) = 0; + virtual void fill(unsigned time, mem_access_sector_mask_t sector_mask, + mem_access_byte_mask_t byte_mask) = 0; virtual bool is_invalid_line() = 0; virtual bool is_valid_line() = 0; @@ -133,6 +134,7 @@ struct cache_block_t { virtual void set_status(enum cache_block_state m_status, mem_access_sector_mask_t sector_mask) = 0; virtual void set_byte_mask(mem_fetch *mf) = 0; + virtual void set_byte_mask(mem_access_byte_mask_t byte_mask) = 0; virtual mem_access_byte_mask_t get_byte_mask() = 0; virtual mem_access_sector_mask_t get_sector_mask() = 0; virtual unsigned long long get_last_access_time() = 0; @@ -145,6 +147,7 @@ struct cache_block_t { mem_access_sector_mask_t sector_mask) = 0; virtual void set_readable_on_fill(bool readable, mem_access_sector_mask_t sector_mask) = 0; + virtual void set_byte_mask_on_fill(bool m_modified) = 0; virtual unsigned get_modified_size() = 0; virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask) = 0; @@ -178,8 +181,10 @@ struct line_cache_block : public cache_block_t { m_ignore_on_fill_status = false; m_set_modified_on_fill = false; m_set_readable_on_fill = false; + m_set_byte_mask_on_fill = false; } - void fill(unsigned time, mem_access_sector_mask_t sector_mask) { + virtual void fill(unsigned time, mem_access_sector_mask_t sector_mask, + mem_access_byte_mask_t byte_mask) { // if(!m_ignore_on_fill_status) // assert( m_status == RESERVED ); @@ -187,6 +192,7 @@ struct line_cache_block : public cache_block_t { if (m_set_readable_on_fill) m_readable = true; + if (m_set_byte_mask_on_fill) set_byte_mask(byte_mask); m_fill_time = time; } @@ -206,6 +212,9 @@ struct line_cache_block : public cache_block_t { virtual void set_byte_mask(mem_fetch *mf) { m_byte_mask = m_byte_mask | mf->get_access_byte_mask(); } + virtual void set_byte_mask(mem_access_byte_mask_t byte_mask) { + m_byte_mask = m_byte_mask | byte_mask; + } virtual mem_access_byte_mask_t get_byte_mask() { return m_byte_mask; } @@ -234,6 +243,9 @@ struct line_cache_block : public cache_block_t { mem_access_sector_mask_t sector_mask) { m_set_readable_on_fill = readable; } + virtual void set_byte_mask_on_fill(bool m_modified) { + m_set_byte_mask_on_fill = m_modified; + } virtual unsigned get_modified_size() { return SECTOR_CHUNCK_SIZE * SECTOR_SIZE; // i.e. cache line size } @@ -256,6 +268,7 @@ struct line_cache_block : public cache_block_t { bool m_ignore_on_fill_status; bool m_set_modified_on_fill; bool m_set_readable_on_fill; + bool m_set_byte_mask_on_fill; bool m_readable; mem_access_byte_mask_t m_byte_mask; }; @@ -303,6 +316,7 @@ struct sector_cache_block : public cache_block_t { m_ignore_on_fill_status[sidx] = false; m_set_modified_on_fill[sidx] = false; m_set_readable_on_fill[sidx] = false; + m_set_byte_mask_on_fill = false; // set line stats m_line_alloc_time = time; // only set this for the first allocated sector @@ -337,7 +351,8 @@ struct sector_cache_block : public cache_block_t { m_line_fill_time = 0; } - virtual void fill(unsigned time, mem_access_sector_mask_t sector_mask) { + virtual void fill(unsigned time, mem_access_sector_mask_t sector_mask, + mem_access_byte_mask_t byte_mask) { unsigned sidx = get_sector_index(sector_mask); // if(!m_ignore_on_fill_status[sidx]) @@ -348,6 +363,7 @@ struct sector_cache_block : public cache_block_t { m_readable[sidx] = true; m_set_readable_on_fill[sidx] = false; } + if (m_set_byte_mask_on_fill) set_byte_mask(byte_mask); m_sector_fill_time[sidx] = time; m_line_fill_time = time; @@ -389,7 +405,10 @@ struct sector_cache_block : public cache_block_t { } virtual void set_byte_mask(mem_fetch *mf) { - m_byte_mask = m_byte_mask | mf->get_access_byte_mask();; + m_byte_mask = m_byte_mask | mf->get_access_byte_mask(); + } + virtual void set_byte_mask(mem_access_byte_mask_t byte_mask) { + m_byte_mask = m_byte_mask | byte_mask; } virtual mem_access_byte_mask_t get_byte_mask() { return m_byte_mask; @@ -427,6 +446,9 @@ struct sector_cache_block : public cache_block_t { unsigned sidx = get_sector_index(sector_mask); m_set_modified_on_fill[sidx] = m_modified; } + virtual void set_byte_mask_on_fill(bool m_modified) { + m_set_byte_mask_on_fill = m_modified; + } virtual void set_readable_on_fill(bool readable, mem_access_sector_mask_t sector_mask) { @@ -468,6 +490,7 @@ struct sector_cache_block : public cache_block_t { bool m_ignore_on_fill_status[SECTOR_CHUNCK_SIZE]; bool m_set_modified_on_fill[SECTOR_CHUNCK_SIZE]; bool m_set_readable_on_fill[SECTOR_CHUNCK_SIZE]; + bool m_set_byte_mask_on_fill; bool m_readable[SECTOR_CHUNCK_SIZE]; mem_access_byte_mask_t m_byte_mask; @@ -904,7 +927,8 @@ class tag_array { void fill(new_addr_type addr, unsigned time, mem_fetch *mf); void fill(unsigned idx, unsigned time, mem_fetch *mf); - void fill(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask); + void fill(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask, + mem_access_byte_mask_t byte_mask); unsigned size() const { return m_config.get_num_lines(); } cache_block_t *get_block(unsigned idx) { return m_lines[idx]; } @@ -1291,7 +1315,8 @@ class baseline_cache : public cache_t { // something is read or written without doing anything else. void force_tag_access(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask) { - m_tag_array->fill(addr, time, mask); + mem_access_byte_mask_t byte_mask; + m_tag_array->fill(addr, time, mask, byte_mask); } protected: |