diff options
Diffstat (limited to 'src/gpgpu-sim/shader.cc')
| -rw-r--r-- | src/gpgpu-sim/shader.cc | 62 |
1 files changed, 43 insertions, 19 deletions
diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 14d9044..6229d16 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1225,22 +1225,6 @@ void scheduler_unit::cycle() { previous_issued_inst_exec_type = exec_unit_type_t::MEM; } } else { - bool sp_pipe_avail = - (m_shader->m_config->gpgpu_num_sp_units > 0) && - m_sp_out->has_free(m_shader->m_config->sub_core_model, m_id); - bool sfu_pipe_avail = - (m_shader->m_config->gpgpu_num_sfu_units > 0) && - m_sfu_out->has_free(m_shader->m_config->sub_core_model, m_id); - bool tensor_core_pipe_avail = - (m_shader->m_config->gpgpu_num_tensor_core_units > 0) && - m_tensor_core_out->has_free( - m_shader->m_config->sub_core_model, m_id); - bool dp_pipe_avail = - (m_shader->m_config->gpgpu_num_dp_units > 0) && - m_dp_out->has_free(m_shader->m_config->sub_core_model, m_id); - bool int_pipe_avail = - (m_shader->m_config->gpgpu_num_int_units > 0) && - m_int_out->has_free(m_shader->m_config->sub_core_model, m_id); // This code need to be refactored if (pI->op != TENSOR_CORE_OP && pI->op != SFU_OP && @@ -1248,6 +1232,13 @@ void scheduler_unit::cycle() { bool execute_on_SP = false; bool execute_on_INT = false; + bool sp_pipe_avail = + (m_shader->m_config->gpgpu_num_sp_units > 0) && + m_sp_out->has_free(m_shader->m_config->sub_core_model, m_id); + bool int_pipe_avail = + (m_shader->m_config->gpgpu_num_int_units > 0) && + m_int_out->has_free(m_shader->m_config->sub_core_model, m_id); + // if INT unit pipline exist, then execute ALU and INT // operations on INT unit and SP-FPU on SP unit (like in Volta) // if INT unit pipline does not exist, then execute all ALU, INT @@ -1308,6 +1299,11 @@ void scheduler_unit::cycle() { (pI->op == DP_OP) && !(diff_exec_units && previous_issued_inst_exec_type == exec_unit_type_t::DP)) { + + bool dp_pipe_avail = + (m_shader->m_config->gpgpu_num_dp_units > 0) && + m_dp_out->has_free(m_shader->m_config->sub_core_model, m_id); + if (dp_pipe_avail) { m_shader->issue_warp(*m_dp_out, pI, active_mask, warp_id, m_id); @@ -1323,6 +1319,11 @@ void scheduler_unit::cycle() { (pI->op == SFU_OP) || (pI->op == ALU_SFU_OP)) && !(diff_exec_units && previous_issued_inst_exec_type == exec_unit_type_t::SFU)) { + + bool sfu_pipe_avail = + (m_shader->m_config->gpgpu_num_sfu_units > 0) && + m_sfu_out->has_free(m_shader->m_config->sub_core_model, m_id); + if (sfu_pipe_avail) { m_shader->issue_warp(*m_sfu_out, pI, active_mask, warp_id, m_id); @@ -1334,6 +1335,12 @@ void scheduler_unit::cycle() { } else if ((pI->op == TENSOR_CORE_OP) && !(diff_exec_units && previous_issued_inst_exec_type == exec_unit_type_t::TENSOR)) { + + bool tensor_core_pipe_avail = + (m_shader->m_config->gpgpu_num_tensor_core_units > 0) && + m_tensor_core_out->has_free( + m_shader->m_config->sub_core_model, m_id); + if (tensor_core_pipe_avail) { m_shader->issue_warp(*m_tensor_core_out, pI, active_mask, warp_id, m_id); @@ -1566,7 +1573,10 @@ void swl_scheduler::order_warps() { } } -void shader_core_ctx::read_operands() {} +void shader_core_ctx::read_operands() { + for (int i = 0; i < m_config->reg_file_port_throughput; ++i) + m_operand_collector.step(); +} address_type coalesced_segment(address_type addr, unsigned segment_size_lg2bytes) { @@ -2578,8 +2588,7 @@ inst->space.get_type() != shared_space) { unsigned warp_id = inst->warp_id(); */ void ldst_unit::cycle() { writeback(); - for (int i = 0; i < m_config->reg_file_port_throughput; ++i) - m_operand_collector->step(); + for (unsigned stage = 0; (stage + 1) < m_pipeline_depth; stage++) if (m_pipeline_reg[stage]->empty() && !m_pipeline_reg[stage + 1]->empty()) move_warp(m_pipeline_reg[stage], m_pipeline_reg[stage + 1]); @@ -3336,6 +3345,21 @@ unsigned int shader_core_config::max_cta(const kernel_info_t &k) const { m_L1D_config.get_total_size_inKB()); } + if(m_L1D_config.is_streaming()) { + //for streaming cache, if the whole memory is allocated + //to the L1 cache, then make the allocation to be on_MISS + //otherwise, make it ON_FILL to eliminate line allocation fails + //i.e. MSHR throughput is the same, independent on the L1 cache size/associativity + if(total_shmed == 0) { + m_L1D_config.set_allocation_policy(ON_MISS); + printf("GPGPU-Sim: Reconfigure L1 allocation to ON_MISS\n"); + } + else { + m_L1D_config.set_allocation_policy(ON_FILL); + printf("GPGPU-Sim: Reconfigure L1 allocation to ON_FILL\n"); + } + } + k.cache_config_set = true; } |