diff options
| author | Mahmoud <[email protected]> | 2019-04-11 22:32:08 -0400 |
|---|---|---|
| committer | Mahmoud <[email protected]> | 2019-04-11 22:32:08 -0400 |
| commit | ec2ada3b5f2ef91341d30e8f56f3c4d9ecafb2a4 (patch) | |
| tree | 8dc5adbbde6c74c5665e348bfea7e2bc7f5988f5 /src | |
| parent | bb7ed349e146ea6b9ef1100d81fa7ac5ef2751cb (diff) | |
adding some stats and random hashing
Diffstat (limited to 'src')
| -rw-r--r-- | src/gpgpu-sim/addrdec.cc | 29 | ||||
| -rw-r--r-- | src/gpgpu-sim/addrdec.h | 5 | ||||
| -rw-r--r-- | src/gpgpu-sim/gpu-sim.cc | 2 | ||||
| -rw-r--r-- | src/gpgpu-sim/gpu-sim.h | 3 | ||||
| -rw-r--r-- | src/gpgpu-sim/local_interconnect.cc | 65 | ||||
| -rw-r--r-- | src/gpgpu-sim/local_interconnect.h | 17 |
6 files changed, 95 insertions, 26 deletions
diff --git a/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc index 8651869..c4c0f53 100644 --- a/src/gpgpu-sim/addrdec.cc +++ b/src/gpgpu-sim/addrdec.cc @@ -165,6 +165,27 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ assert(tlx->chip < m_n_channel); break; } + case RANDOM: + { + new_addr_type chip_address = (addr>>ADDR_CHIP_S); + tr1_hash_map<new_addr_type,unsigned>::const_iterator got = address_random_interleaving.find (chip_address); + if ( got == address_random_interleaving.end() ) { + unsigned new_chip_id = rand() % (m_n_channel*m_n_sub_partition_in_channel); + address_random_interleaving[chip_address] = new_chip_id; + tlx->chip = new_chip_id/m_n_sub_partition_in_channel; + tlx->sub_partition = new_chip_id; + } + else { + unsigned new_chip_id = got->second; + tlx->chip = new_chip_id/m_n_sub_partition_in_channel; + tlx->sub_partition = new_chip_id; + } + + assert(tlx->chip < m_n_channel); + assert(tlx->sub_partition < m_n_channel*m_n_sub_partition_in_channel); + return; + break; + } case CUSTOM: /* No custom set function implemented */ //Do you custom index here @@ -175,9 +196,9 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ } // combine the chip address and the lower bits of DRAM bank address to form the subpartition ID - unsigned sub_partition_addr_mask = m_n_sub_partition_in_channel - 1; + unsigned sub_partition_addr_mask = m_n_sub_partition_in_channel - 1; tlx->sub_partition = tlx->chip * m_n_sub_partition_in_channel - + (tlx->bk & sub_partition_addr_mask); + + (tlx->bk & sub_partition_addr_mask); } void linear_to_raw_address_translation::addrdec_parseoption(const char *option) @@ -396,6 +417,10 @@ void linear_to_raw_address_translation::init(unsigned int n_channel, unsigned in if (run_test) { sweep_test(); } + + if(memory_partition_indexing == RANDOM) + srand (time(NULL)); + } #include "../tr1_hash_map.h" diff --git a/src/gpgpu-sim/addrdec.h b/src/gpgpu-sim/addrdec.h index bdc5fec..a5333fb 100644 --- a/src/gpgpu-sim/addrdec.h +++ b/src/gpgpu-sim/addrdec.h @@ -40,6 +40,7 @@ enum partition_index_function{ BITWISE_PERMUTATION, IPOLY, PAE, + RANDOM, CUSTOM }; @@ -55,6 +56,7 @@ struct addrdec_t { unsigned sub_partition; }; + class linear_to_raw_address_translation { public: linear_to_raw_address_translation(); @@ -62,7 +64,7 @@ public: void init(unsigned int n_channel, unsigned int n_sub_partition_in_channel); // accessors - void addrdec_tlx(new_addr_type addr, addrdec_t *tlx) const; + void addrdec_tlx(new_addr_type addr, addrdec_t *tlx) const; new_addr_type partition_address( new_addr_type addr ) const; private: @@ -92,6 +94,7 @@ private: unsigned int gap; int m_n_channel; int m_n_sub_partition_in_channel; + }; #endif diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 81d9b69..8837aef 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -99,6 +99,8 @@ unsigned long long gpu_tot_sim_cycle_parition_util = 0; unsigned long long partiton_replys_in_parallel = 0; unsigned long long partiton_replys_in_parallel_total = 0; +tr1_hash_map<new_addr_type,unsigned> address_random_interleaving; + /* Clock Domains */ #define CORE 0x01 diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 6ce5524..efda65a 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -62,8 +62,7 @@ #define SAMPLELOG 222 #define DUMPLOG 333 - - +extern tr1_hash_map<new_addr_type,unsigned> address_random_interleaving; enum dram_ctrl_t { diff --git a/src/gpgpu-sim/local_interconnect.cc b/src/gpgpu-sim/local_interconnect.cc index 1d93fe6..21b44ed 100644 --- a/src/gpgpu-sim/local_interconnect.cc +++ b/src/gpgpu-sim/local_interconnect.cc @@ -36,9 +36,10 @@ #include "local_interconnect.h" #include "mem_fetch.h" -xbar_router::xbar_router(unsigned router_id, unsigned n_shader, unsigned n_mem, unsigned m_in_buffer_limit, unsigned m_out_buffer_limit) +xbar_router::xbar_router(unsigned router_id, enum Interconnect_type m_type, unsigned n_shader, unsigned n_mem, unsigned m_in_buffer_limit, unsigned m_out_buffer_limit) { m_id=router_id; + router_type=m_type; _n_mem = n_mem; _n_shader = n_shader; total_nodes = n_shader+n_mem; @@ -47,10 +48,21 @@ xbar_router::xbar_router(unsigned router_id, unsigned n_shader, unsigned n_mem, next_node=0; in_buffer_limit = m_in_buffer_limit; out_buffer_limit = m_out_buffer_limit; + if(m_type == REQ_NET) { + active_in_buffers=n_shader; + active_out_buffers=n_mem; + } + else if(m_type == REPLY_NET) { + active_in_buffers=n_mem; + active_out_buffers=n_shader; + } cycles = 0; conflicts= 0; out_buffer_full=0; + in_buffer_full=0; + out_buffer_util=0; + in_buffer_util=0; packets_num=0; } @@ -81,14 +93,19 @@ void* xbar_router::Pop(unsigned ouput_deviceID) } -bool xbar_router::Has_Buffer_In(unsigned input_deviceID, unsigned size) const{ +bool xbar_router::Has_Buffer_In(unsigned input_deviceID, unsigned size, bool update_counter){ assert(input_deviceID < total_nodes); - return (in_buffers[input_deviceID].size() + size <= in_buffer_limit); + + bool has_buffer = (in_buffers[input_deviceID].size() + size <= in_buffer_limit); + if(update_counter && !has_buffer) + in_buffer_full++; + + return has_buffer; } -bool xbar_router::Has_Buffer_Out(unsigned output_deviceID, unsigned size) const{ +bool xbar_router::Has_Buffer_Out(unsigned output_deviceID, unsigned size){ return (out_buffers[output_deviceID].size() + size <= out_buffer_limit); } @@ -118,6 +135,12 @@ void xbar_router::Advance() { } next_node = (++next_node % total_nodes); + + //collect some stats about buffer util + for(unsigned i=0; i<total_nodes; ++i){ + in_buffer_util+=in_buffers[i].size(); + out_buffer_util+=out_buffers[i].size(); + } } bool xbar_router::Busy() const { @@ -169,7 +192,7 @@ void LocalInterconnect::CreateInterconnect(unsigned m_n_shader, unsigned m_n_mem net.resize(n_subnets); for (unsigned i = 0; i < n_subnets; ++i) { - net[i] = new xbar_router( i, m_n_shader, m_n_mem, m_inct_config.in_buffer_limit, m_inct_config.out_buffer_limit ); + net[i] = new xbar_router( i, static_cast<Interconnect_type>(i), m_n_shader, m_n_mem, m_inct_config.in_buffer_limit, m_inct_config.out_buffer_limit ); } } @@ -236,10 +259,10 @@ bool LocalInterconnect::HasBuffer(unsigned deviceID, unsigned int size) const{ bool has_buffer = false; - has_buffer = net[0]->Has_Buffer_In(deviceID, size); - if ((n_subnets>1) && deviceID >= n_shader) // deviceID is memory node - has_buffer = net[1]->Has_Buffer_In(deviceID, size); + has_buffer = net[REPLY_NET]->Has_Buffer_In(deviceID, 1, true); + else + has_buffer = net[REQ_NET]->Has_Buffer_In(deviceID, 1, true); return has_buffer; @@ -247,18 +270,24 @@ bool LocalInterconnect::HasBuffer(unsigned deviceID, unsigned int size) const{ void LocalInterconnect::DisplayStats() const{ - cout<<"Req_Network_injected_packets_num = "<<net[0]->packets_num<<endl; - cout<<"Req_Network_cycles = "<<net[0]->cycles<<endl; - cout<<"Req_Network_injected_packets_per_cycle = "<<(float)(net[0]->packets_num) / (net[0]->cycles)<<endl; - cout<<"Req_Network_conflicts_per_cycle = "<<(float)(net[0]->conflicts) / (net[0]->cycles)<<endl; - cout<<"Req_Network_out_buffer_full_per_cycle = "<<(float)(net[0]->out_buffer_full) / (net[0]->cycles)<<endl; + cout<<"Req_Network_injected_packets_num = "<<net[REQ_NET]->packets_num<<endl; + cout<<"Req_Network_cycles = "<<net[REQ_NET]->cycles<<endl; + cout<<"Req_Network_injected_packets_per_cycle = "<<(float)(net[REQ_NET]->packets_num) / (net[REQ_NET]->cycles)<<endl; + cout<<"Req_Network_conflicts_per_cycle = "<<(float)(net[REQ_NET]->conflicts) / (net[REQ_NET]->cycles)<<endl; + cout<<"Req_Network_in_buffer_full_per_cycle = "<<(float)(net[REQ_NET]->in_buffer_full) / (net[REQ_NET]->cycles)<<endl; + cout<<"Req_Network_in_buffer_avg_util = "<<((float)(net[REQ_NET]->in_buffer_util) / (net[REQ_NET]->cycles) / net[REQ_NET]->active_in_buffers)<<endl; + cout<<"Req_Network_out_buffer_full_per_cycle = "<<(float)(net[REQ_NET]->out_buffer_full) / (net[REQ_NET]->cycles)<<endl; + cout<<"Req_Network_out_buffer_avg_util = "<<((float)(net[REQ_NET]->out_buffer_util) / (net[REQ_NET]->cycles) / net[REQ_NET]->active_out_buffers)<<endl; cout<<endl; - cout<<"Reply_Network_injected_packets_num = "<<net[1]->packets_num<<endl; - cout<<"ReplyNetwork_cycles = "<<net[1]->cycles<<endl; - cout<<"ReplyNetwork_injected_packets_per_cycle = "<<(float)(net[1]->packets_num) / (net[1]->cycles)<<endl; - cout<<"ReplyNetwork_conflicts_per_cycle = "<<(float)(net[1]->conflicts) / (net[1]->cycles)<<endl; - cout<<"ReplyNetwork_out_buffer_full_per_cycle = "<<(float)(net[1]->out_buffer_full) / (net[1]->cycles)<<endl; + cout<<"Reply_Network_injected_packets_num = "<<net[REPLY_NET]->packets_num<<endl; + cout<<"Reply_Network_cycles = "<<net[REPLY_NET]->cycles<<endl; + cout<<"Reply_Network_injected_packets_per_cycle = "<<(float)(net[REPLY_NET]->packets_num) / (net[REPLY_NET]->cycles)<<endl; + cout<<"Reply_Network_conflicts_per_cycle = "<<(float)(net[REPLY_NET]->conflicts) / (net[REPLY_NET]->cycles)<<endl; + cout<<"Reply_Network_in_buffer_full_per_cycle = "<<(float)(net[REPLY_NET]->in_buffer_full) / (net[REPLY_NET]->cycles)<<endl; + cout<<"Reply_Network_in_buffer_avg_util = "<<((float)(net[REPLY_NET]->in_buffer_util) / (net[REPLY_NET]->cycles) / net[REPLY_NET]->active_in_buffers)<<endl; + cout<<"Reply_Network_out_buffer_full_per_cycle = "<<(float)(net[REPLY_NET]->out_buffer_full) / (net[REPLY_NET]->cycles)<<endl; + cout<<"Reply_Network_out_buffer_avg_util= "<<((float)(net[REPLY_NET]->out_buffer_util) / (net[REPLY_NET]->cycles) / net[REPLY_NET]->active_out_buffers)<<endl; } diff --git a/src/gpgpu-sim/local_interconnect.h b/src/gpgpu-sim/local_interconnect.h index 727a782..0f83d05 100644 --- a/src/gpgpu-sim/local_interconnect.h +++ b/src/gpgpu-sim/local_interconnect.h @@ -44,22 +44,29 @@ struct inct_config unsigned subnets; }; +enum Interconnect_type { + REQ_NET=0, + REPLY_NET=1 +}; class xbar_router { public: - xbar_router(unsigned router_id, unsigned n_shader, unsigned n_mem, unsigned m_in_buffer_limit, unsigned m_out_buffer_limit); + xbar_router(unsigned router_id, enum Interconnect_type m_type, unsigned n_shader, unsigned n_mem, unsigned m_in_buffer_limit, unsigned m_out_buffer_limit); ~xbar_router(); void Push(unsigned input_deviceID, unsigned output_deviceID, void* data, unsigned int size); void* Pop(unsigned ouput_deviceID); void Advance(); bool Busy() const; - bool Has_Buffer_In(unsigned input_deviceID, unsigned size) const; - bool Has_Buffer_Out(unsigned output_deviceID, unsigned size) const; + bool Has_Buffer_In(unsigned input_deviceID, unsigned size, bool update_counter=false); + bool Has_Buffer_Out(unsigned output_deviceID, unsigned size); //some stats unsigned long long cycles; unsigned long long conflicts; unsigned long long out_buffer_full; + unsigned long long out_buffer_util; + unsigned long long in_buffer_full; + unsigned long long in_buffer_util; unsigned long long packets_num; private: @@ -77,6 +84,10 @@ private: unsigned in_buffer_limit, out_buffer_limit; unsigned next_node; unsigned m_id; + enum Interconnect_type router_type; + unsigned active_in_buffers,active_out_buffers; + + friend class LocalInterconnect; }; |
