From f852a7108e691045dd3910065836a817babcde8c Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 7 Sep 2018 22:01:31 -0400 Subject: adding streamin cache + fixing TEX cache + adding l1 latency and smem latency --- src/abstract_hardware_model.h | 1 + src/gpgpu-sim/gpu-cache.cc | 20 ++++++- src/gpgpu-sim/gpu-cache.h | 36 ++++++++++-- src/gpgpu-sim/gpu-sim.cc | 6 ++ src/gpgpu-sim/l2cache.cc | 8 ++- src/gpgpu-sim/shader.cc | 126 ++++++++++++++++++++++++++++++++++++++++-- src/gpgpu-sim/shader.h | 6 ++ 7 files changed, 191 insertions(+), 12 deletions(-) (limited to 'src') diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 1b764e2..a70b077 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -392,6 +392,7 @@ protected: #define LOCAL_MEM_SIZE_MAX (8*1024) #define MAX_STREAMING_MULTIPROCESSORS 64 #define MAX_THREAD_PER_SM 2048 +#define MAX_WARP_PER_SM 64 #define TOTAL_LOCAL_MEM_PER_SM (MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) #define TOTAL_SHARED_MEM (MAX_STREAMING_MULTIPROCESSORS*SHARED_MEM_SIZE_MAX) #define TOTAL_LOCAL_MEM (MAX_STREAMING_MULTIPROCESSORS*MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index f181f20..a11853a 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -71,6 +71,7 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ switch(m_set_index_function){ case FERMI_HASH_SET_FUNCTION: + case BITWISE_XORING_FUNCTION: /* * Set Indexing function from "A Detailed GPU Cache Model Based on Reuse Distance Theory" * Cedric Nugteren et al. @@ -1581,7 +1582,7 @@ enum cache_request_status tex_cache::access( new_addr_type addr, mem_fetch *mf, if ( status == MISS ) { // we need to send a memory request... unsigned rob_index = m_rob.push( rob_entry(cache_index, mf, block_addr) ); - m_extra_mf_fields[mf] = extra_mf_fields(rob_index); + m_extra_mf_fields[mf] = extra_mf_fields(rob_index, m_config); mf->set_data_size(m_config.get_line_sz()); m_tags.fill(cache_index,time,mf); // mark block as valid m_request_fifo.push(mf); @@ -1636,6 +1637,23 @@ void tex_cache::cycle(){ /// Place returning cache block into reorder buffer void tex_cache::fill( mem_fetch *mf, unsigned time ) { + if(m_config.m_mshr_type == SECTOR_TEX_FIFO) { + assert(mf->get_original_mf()); + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->get_original_mf()); + assert( e != m_extra_mf_fields.end() ); + e->second.pending_read--; + + if(e->second.pending_read > 0) { + //wait for the other requests to come back + delete mf; + return; + } else { + mem_fetch *temp = mf; + mf = mf->get_original_mf(); + delete temp; + } + } + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf); assert( e != m_extra_mf_fields.end() ); assert( e->second.m_valid ); diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index dee43f2..4ed382c 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -455,7 +455,8 @@ enum write_policy_t { enum allocation_policy_t { ON_MISS, - ON_FILL + ON_FILL, + STREAMING }; @@ -467,8 +468,9 @@ enum write_allocate_policy_t { }; enum mshr_config_t { - TEX_FIFO, + TEX_FIFO, // Tex cache ASSOC, // normal cache + SECTOR_TEX_FIFO, //Tex cache sends requests to high-level sector cache SECTOR_ASSOC // normal cache sends requests to high-level sector cache }; @@ -485,6 +487,12 @@ enum cache_type{ SECTOR }; +#define MAX_WARP_PER_SHADER 64 +#define INCT_TOTAL_BUFFER 64 +#define L2_TOTAL 64 +#define MAX_WARP_PER_SHADER 64 +#define MAX_WARP_PER_SHADER 64 + class cache_config { public: cache_config() @@ -544,10 +552,27 @@ public: switch (ap) { case 'm': m_alloc_policy = ON_MISS; break; case 'f': m_alloc_policy = ON_FILL; break; + case 's': m_alloc_policy = STREAMING; break; default: exit_parse_error(); } + if(m_alloc_policy == STREAMING) { + //For streaming cache, we set the alloc policy to be on-fill to remove all line_alloc_fail stalls + //we set the MSHRs to be equal to the cache line. This is possible by moving TAG to be shared between cache line and MSHR enrty (i.e. for each cache line, there is an MSHR rntey associated with it) + // This is the easiest think we can think about to model (mimics) L1 streaming cache in Pascal and Volta + //Based on our microbenchmakrs, MSHRs entries have been increasing substantially in Pascal and Volta + //For more information about streaming cache, see: + // http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + // https://ieeexplore.ieee.org/document/8344474/ + + m_alloc_policy = ON_FILL; + m_mshr_entries = m_nset*m_assoc; + if(m_cache_type == SECTOR) + m_mshr_entries *= SECTOR_CHUNCK_SIZE; + m_mshr_max_merge = MAX_WARP_PER_SM; + } switch (mshr_type) { case 'F': m_mshr_type = TEX_FIFO; assert(ntok==14); break; + case 'T': m_mshr_type = SECTOR_TEX_FIFO; assert(ntok==14); break; case 'A': m_mshr_type = ASSOC; break; case 'S' : m_mshr_type = SECTOR_ASSOC; break; default: exit_parse_error(); @@ -722,6 +747,7 @@ class l1d_cache_config : public cache_config{ public: l1d_cache_config() : cache_config(){} virtual unsigned set_index(new_addr_type addr) const; + unsigned l1_latency; }; class l2_cache_config : public cache_config { @@ -1442,7 +1468,7 @@ public: m_result_fifo(config.m_result_fifo_entries) { m_name = name; - assert(config.m_mshr_type == TEX_FIFO); + assert(config.m_mshr_type == TEX_FIFO || config.m_mshr_type == SECTOR_TEX_FIFO ); assert(config.m_write_policy == READ_ONLY); assert(config.m_alloc_policy == ON_MISS); m_memport=memport; @@ -1595,13 +1621,15 @@ private: struct extra_mf_fields { extra_mf_fields() { m_valid = false;} - extra_mf_fields( unsigned i ) + extra_mf_fields( unsigned i, const cache_config &m_config ) { m_valid = true; m_rob_index = i; + pending_read = m_config.m_mshr_type == SECTOR_TEX_FIFO? m_config.m_line_sz/SECTOR_SIZE : 0; } bool m_valid; unsigned m_rob_index; + unsigned pending_read; }; cache_stats m_stats; diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index ea2dfba..08d4525 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -250,6 +250,12 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); + option_parser_register(opp, "-l1_latency", OPT_UINT32, &m_L1D_config.l1_latency, + "L1 Hit Latency", + "0"); + option_parser_register(opp, "-smem_latency", OPT_UINT32, &smem_latency, + "smem Latency", + "3"); option_parser_register(opp, "-gpgpu_cache:dl1PrefL1", OPT_CSTR, &m_L1D_config.m_config_stringPrefL1, "per-shader L1 data cache config " " {::,:::,::, | none}", diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index f42610f..25da107 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -609,7 +609,7 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques } else { printf("Invalid sector received, address = 0x%06x, sector mask = %s, data size = %d", - mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size(), mf->get_data_size()); + mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size()); assert(0 && "Undefined sector mask is received"); } @@ -640,7 +640,11 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques result.push_back(n_mf); byte_sector_mask <<= SECTOR_SIZE; } - } else assert(0 && "Undefined data size is received"); + } else { + printf("Invalid sector received, address = 0x%06x, sector mask = %d, byte mask = , data size = %d", + mf->get_addr(), mf->get_access_sector_mask().count(), mf->get_data_size()); + assert(0 && "Undefined data size is received"); + } return result; } diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index b59e5d2..51689e3 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1488,6 +1488,111 @@ mem_stage_stall_type ldst_unit::process_memory_access_queue( cache_t *cache, war return process_cache_access( cache, mf->get_addr(), inst, events, mf, status ); } +mem_stage_stall_type ldst_unit::process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst ) +{ + mem_stage_stall_type result = NO_RC_FAIL; + if( inst.accessq_empty() ) + return result; + + mem_fetch *mf = m_mf_allocator->alloc(inst,inst.accessq_back()); + + if(m_config->m_L1D_config.l1_latency > 0) + { + if((l1_latency_queue[m_config->m_L1D_config.l1_latency-1]) == NULL) + { + l1_latency_queue[m_config->m_L1D_config.l1_latency-1] = mf; + + if( mf->get_inst().is_store() ) { + unsigned inc_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf->get_data_size()/SECTOR_SIZE) : 1; + + for(unsigned i=0; i< inc_ack; ++i) + m_core->inc_store_req( inst.warp_id() ); + } + + inst.accessq_pop_back(); + } + else + { + result = BK_CONF; + delete mf; + } + if( !inst.accessq_empty() && result !=BK_CONF) + result = COAL_STALL; + return result; + } + else + { + std::list events; + enum cache_request_status status = cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle,events); + return process_cache_access( cache, mf->get_addr(), inst, events, mf, status ); + } +} + +void ldst_unit::L1_latency_queue_cycle() +{ + //std::deque< std::pair >::iterator it = m_latency_queue.begin(); + if((l1_latency_queue[0]) != NULL) + { + mem_fetch* mf_next = l1_latency_queue[0]; + std::list events; + enum cache_request_status status = m_L1D->access(mf_next->get_addr(),mf_next,gpu_sim_cycle+gpu_tot_sim_cycle,events); + + bool write_sent = was_write_sent(events); + bool read_sent = was_read_sent(events); + + if ( status == HIT ) { + assert( !read_sent ); + l1_latency_queue[0] = NULL; + if ( mf_next->get_inst().is_load() ) { + for ( unsigned r=0; r < 4; r++) + if (mf_next->get_inst().out[r] > 0) + { + assert(m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]>0); + unsigned still_pending = --m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]; + if(!still_pending) + { + m_pending_writes[mf_next->get_inst().warp_id()].erase(mf_next->get_inst().out[r]); + m_scoreboard->releaseRegister(mf_next->get_inst().warp_id(),mf_next->get_inst().out[r]); + m_core->warp_inst_complete(mf_next->get_inst()); + } + } + } + + //For write hit in WB policy + if(mf_next->get_inst().is_store() && !write_sent) + { + unsigned dec_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf_next->get_data_size()/SECTOR_SIZE) : 1; + + mf_next->set_reply(); + + for(unsigned i=0; i< dec_ack; ++i) + m_core->store_ack(mf_next); + } + + if( !write_sent ) + delete mf_next; + + } else if ( status == RESERVATION_FAIL ) { + assert( !read_sent ); + assert( !write_sent ); + } else { + assert( status == MISS || status == HIT_RESERVED ); + l1_latency_queue[0] = NULL; + } + } + + for( unsigned stage = 0; stagem_L1D_config.l1_latency-1; ++stage) + if( l1_latency_queue[stage] == NULL) { + l1_latency_queue[stage] = l1_latency_queue[stage+1] ; + l1_latency_queue[stage+1] = NULL; + } + +} + + + bool ldst_unit::constant_cycle( warp_inst_t &inst, mem_stage_stall_type &rc_fail, mem_stage_access_type &fail_type) { if( inst.empty() || ((inst.space.get_type() != const_space) && (inst.space.get_type() != param_space_kernel)) ) @@ -1561,7 +1666,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea } } else { assert( CACHE_UNDEFINED != inst.cache_op ); - stall_cond = process_memory_access_queue(m_L1D,inst); + stall_cond = process_memory_access_queue_l1cache(m_L1D,inst); } if( !inst.accessq_empty() && stall_cond == NO_RC_FAIL) stall_cond = COAL_STALL; @@ -1771,8 +1876,9 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, const memory_config *mem_config, shader_core_stats *stats, unsigned sid, - unsigned tpc ) : pipelined_simd_unit(NULL,config,3,core), m_next_wb(config) + unsigned tpc ) : pipelined_simd_unit(NULL,config,config->smem_latency,core), m_next_wb(config) { + assert(config->smem_latency > 1); init( icnt, mf_allocator, core, @@ -1793,6 +1899,12 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, m_icnt, m_mf_allocator, IN_L1D_MISS_QUEUE ); + + if(m_config->m_L1D_config.l1_latency > 0) + { + for(int i=0; im_L1D_config.l1_latency; i++ ) + l1_latency_queue.push_back((mem_fetch*)NULL); + } } } @@ -2019,7 +2131,11 @@ void ldst_unit::cycle() m_L1T->cycle(); m_L1C->cycle(); - if( m_L1D ) m_L1D->cycle(); + if( m_L1D ) { + m_L1D->cycle(); + if(m_config->m_L1D_config.l1_latency > 0) + L1_latency_queue_cycle(); + } warp_inst_t &pipe_reg = *m_dispatch_reg; enum mem_stage_stall_type rc_fail = NO_RC_FAIL; @@ -2042,9 +2158,9 @@ void ldst_unit::cycle() unsigned warp_id = pipe_reg.warp_id(); if( pipe_reg.is_load() ) { if( pipe_reg.space.get_type() == shared_space ) { - if( m_pipeline_reg[2]->empty() ) { + if( m_pipeline_reg[m_config->smem_latency-1]->empty() ) { // new shared memory request - move_warp(m_pipeline_reg[2],m_dispatch_reg); + move_warp(m_pipeline_reg[m_config->smem_latency-1],m_dispatch_reg); m_dispatch_reg->clear(); } } else { diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index cc441b3..e07096e 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1220,6 +1220,7 @@ protected: mem_fetch *mf, enum cache_request_status status ); mem_stage_stall_type process_memory_access_queue( cache_t *cache, warp_inst_t &inst ); + mem_stage_stall_type process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst ); const memory_config *m_memory_config; class mem_fetch_interface *m_icnt; @@ -1248,6 +1249,9 @@ protected: // for debugging unsigned long long m_last_inst_gpu_sim_cycle; unsigned long long m_last_inst_gpu_tot_sim_cycle; + + std::deque l1_latency_queue; + void L1_latency_queue_cycle(); }; enum pipeline_stage_name_t { @@ -1399,6 +1403,8 @@ struct shader_core_config : public core_config int simt_core_sim_order; + unsigned smem_latency; + unsigned mem2device(unsigned memid) const { return memid + n_simt_clusters; } //Jin: concurrent kernel on sm -- cgit v1.3