diff options
| author | aamir <[email protected]> | 2018-10-24 21:18:27 -0700 |
|---|---|---|
| committer | aamir <[email protected]> | 2018-10-24 21:18:27 -0700 |
| commit | 7c441c450e40bf07bdf1acfe1eb2258952e1f7b7 (patch) | |
| tree | ff4f3ab1e5d42c8284d419209afb2ccc5f25603b /src/abstract_hardware_model.cc | |
| parent | 68134d5eb326552fc1ef4b02b2eb21103266283b (diff) | |
| parent | 09e6092ace5213a5d5a49bf80b052802c06a4268 (diff) | |
merged tensor-cores code
Diffstat (limited to 'src/abstract_hardware_model.cc')
| -rw-r--r-- | src/abstract_hardware_model.cc | 9 |
1 files changed, 6 insertions, 3 deletions
diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index fe6f8ab..f7f1016 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -184,7 +184,7 @@ void warp_inst_t::generate_mem_accesses() { if( empty() || op == MEMORY_BARRIER_OP || m_mem_accesses_created ) return; - if ( !((op == LOAD_OP) || (op == STORE_OP)) ) + if (!((op == LOAD_OP) || (op==TENSOR_CORE_LOAD_OP) || (op == STORE_OP)||(op==TENSOR_CORE_STORE_OP))) return; if( m_warp_active_mask.count() == 0 ) return; // predicated off @@ -213,6 +213,7 @@ void warp_inst_t::generate_mem_accesses() access_type = is_write? LOCAL_ACC_W: LOCAL_ACC_R; break; case shared_space: break; + case sstarr_space: break; default: assert(0); break; } @@ -220,7 +221,8 @@ void warp_inst_t::generate_mem_accesses() new_addr_type cache_block_size = 0; // in bytes switch( space.get_type() ) { - case shared_space: { + case shared_space: + case sstarr_space: { unsigned subwarp_size = m_config->warp_size / m_config->mem_warp_parts; unsigned total_accesses=0; for( unsigned subwarp=0; subwarp < m_config->mem_warp_parts; subwarp++ ) { @@ -387,7 +389,8 @@ void warp_inst_t::memory_coalescing_arch_13( bool is_write, mem_access_type acce assert(num_accesses <= MAX_ACCESSES_PER_INSN_PER_THREAD); - for(unsigned access=0; access<num_accesses; access++) { +// for(unsigned access=0; access<num_accesses; access++) { + for(unsigned access=0; (access<MAX_ACCESSES_PER_INSN_PER_THREAD)&&(m_per_scalar_thread[thread].memreqaddr[access]!=0); access++) { new_addr_type addr = m_per_scalar_thread[thread].memreqaddr[access]; unsigned block_address = line_size_based_tag_func(addr,segment_size); unsigned chunk = (addr&127)/32; // which 32-byte chunk within in a 128-byte chunk does this thread access? |