diff options
Diffstat (limited to 'src/abstract_hardware_model.cc')
| -rw-r--r-- | src/abstract_hardware_model.cc | 58 |
1 files changed, 40 insertions, 18 deletions
diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index fe6f8ab..acb376a 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++ ) { @@ -314,12 +316,12 @@ void warp_inst_t::generate_mem_accesses() break; case global_space: case local_space: case param_space_local: - if( m_config->gpgpu_coalesce_arch == 13 ) { - if(isatomic()) - memory_coalescing_arch_13_atomic(is_write, access_type); - else - memory_coalescing_arch_13(is_write, access_type); - } else abort(); + if( m_config->gpgpu_coalesce_arch >= 13 && m_config->gpgpu_coalesce_arch <= 62) { + if(isatomic()) + memory_coalescing_arch_atomic(is_write, access_type); + else + memory_coalescing_arch(is_write, access_type); + } else abort(); break; @@ -343,7 +345,7 @@ void warp_inst_t::generate_mem_accesses() byte_mask.set(idx+i); } for( a=accesses.begin(); a != accesses.end(); ++a ) - m_accessq.push_back( mem_access_t(access_type,a->first,cache_block_size,is_write,a->second,byte_mask) ); + m_accessq.push_back( mem_access_t(access_type,a->first,cache_block_size,is_write,a->second, byte_mask, mem_access_sector_mask_t())); } if ( space.get_type() == global_space ) { @@ -352,15 +354,32 @@ void warp_inst_t::generate_mem_accesses() m_mem_accesses_created=true; } -void warp_inst_t::memory_coalescing_arch_13( bool is_write, mem_access_type access_type ) +void warp_inst_t::memory_coalescing_arch( bool is_write, mem_access_type access_type ) { // see the CUDA manual where it discusses coalescing rules before reading this unsigned segment_size = 0; unsigned warp_parts = m_config->mem_warp_parts; + bool sector_segment_size = false; + + if(m_config->gpgpu_coalesce_arch >= 20 && m_config->gpgpu_coalesce_arch < 39) + { + //Fermi and Kepler, L1 is normal and L2 is sector + if(m_config->gmem_skip_L1D || cache_op == CACHE_GLOBAL) + sector_segment_size = true; + else + sector_segment_size = false; + } + else if(m_config->gpgpu_coalesce_arch >= 40) + { + //Maxwell and Pascal, L1 and L2 are sectors + //all requests should be 32 bytes + sector_segment_size = true; + } + switch( data_size ) { case 1: segment_size = 32; break; - case 2: segment_size = 64; break; - case 4: case 8: case 16: segment_size = 128; break; + case 2: segment_size = sector_segment_size? 32 : 64; break; + case 4: case 8: case 16: segment_size = sector_segment_size? 32 : 128; break; } unsigned subwarp_size = m_config->warp_size / warp_parts; @@ -387,7 +406,8 @@ 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? @@ -410,13 +430,13 @@ void warp_inst_t::memory_coalescing_arch_13( bool is_write, mem_access_type acce new_addr_type addr = t->first; const transaction_info &info = t->second; - memory_coalescing_arch_13_reduce_and_send(is_write, access_type, info, addr, segment_size); + memory_coalescing_arch_reduce_and_send(is_write, access_type, info, addr, segment_size); } } } -void warp_inst_t::memory_coalescing_arch_13_atomic( bool is_write, mem_access_type access_type ) +void warp_inst_t::memory_coalescing_arch_atomic( bool is_write, mem_access_type access_type ) { assert(space.get_type() == global_space); // Atomics allowed only for global memory @@ -485,13 +505,13 @@ void warp_inst_t::memory_coalescing_arch_13_atomic( bool is_write, mem_access_ty for(t=transaction_list.begin(); t!=transaction_list.end(); t++) { // For each transaction const transaction_info &info = *t; - memory_coalescing_arch_13_reduce_and_send(is_write, access_type, info, addr, segment_size); + memory_coalescing_arch_reduce_and_send(is_write, access_type, info, addr, segment_size); } } } } -void warp_inst_t::memory_coalescing_arch_13_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ) +void warp_inst_t::memory_coalescing_arch_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ) { assert( (addr & (segment_size-1)) == 0 ); @@ -540,7 +560,7 @@ void warp_inst_t::memory_coalescing_arch_13_reduce_and_send( bool is_write, mem_ assert(lower_half_used && upper_half_used); } } - m_accessq.push_back( mem_access_t(access_type,addr,size,is_write,info.active,info.bytes) ); + m_accessq.push_back( mem_access_t(access_type,addr,size,is_write,info.active,info.bytes, info.chunks) ); } void warp_inst_t::completed( unsigned long long cycle ) const @@ -574,6 +594,8 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info * //Jin: launch latency management m_launch_latency = g_kernel_launch_latency; + + volta_cache_config_set=false; } kernel_info_t::~kernel_info_t() |