From 355de7f8d033a4b5082cb0816650d880f05d81c8 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 13 Sep 2017 16:03:39 -0400 Subject: Adding sperate dp_unit --- src/cuda-sim/cuda-sim.cc | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'src/cuda-sim') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index d4ace76..c7c32fa 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -678,6 +678,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[0]; initiation_interval = dp_init[0]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -699,6 +700,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[1]; initiation_interval = dp_init[1]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -721,7 +723,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[2]; initiation_interval = dp_init[2]; - op = ALU_SFU_OP; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -744,6 +746,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[3]; initiation_interval = dp_init[3]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: -- cgit v1.3 From 1623ca1036356a8a50fa7e0b504f0f67df054e5c Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 14 Sep 2017 10:28:40 -0400 Subject: adding seperate sfu latency and init variables --- src/cuda-sim/cuda-sim.cc | 27 ++++++++++++++++++++------- 1 file changed, 20 insertions(+), 7 deletions(-) (limited to 'src/cuda-sim') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index c7c32fa..48d1219 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -62,8 +62,8 @@ addr_t g_debug_pc = 0xBEEF1518; unsigned g_ptx_sim_num_insn = 0; unsigned gpgpu_param_num_shaders = 0; -char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp; -char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp; +char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu; +char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp,*opcode_initiation_sfu; char *cdp_latency_str; unsigned cdp_latency[5]; @@ -80,6 +80,10 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode latencies for double precision floating points " "Default 8,8,8,8,335", "8,8,8,8,335"); + option_parser_register(opp, "-opcode_latency_sfu", OPT_CSTR, &opcode_latency_sfu, + "Opcode latencies for SFU instructions" + "Default 8", + "8"); option_parser_register(opp, "-ptx_opcode_initiation_int", OPT_CSTR, &opcode_initiation_int, "Opcode initiation intervals for integers " "Default 1,1,4,4,32", @@ -92,6 +96,10 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode initiation intervals for double precision floating points " "Default 8,8,8,8,130", "8,8,8,8,130"); + option_parser_register(opp, "-ptx_opcode_initiation_sfu", OPT_CSTR, &opcode_initiation_sfu, + "Opcode initiation intervals for sfu instructions" + "Default 8", + "8"); option_parser_register(opp, "-cdp_latency", OPT_CSTR, &cdp_latency_str, "CDP API latency Date: Thu, 26 Oct 2017 11:27:19 -0400 Subject: fixing a typo in config file --- src/cuda-sim/cuda-sim.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/cuda-sim') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 48d1219..a668db1 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -80,7 +80,7 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode latencies for double precision floating points " "Default 8,8,8,8,335", "8,8,8,8,335"); - option_parser_register(opp, "-opcode_latency_sfu", OPT_CSTR, &opcode_latency_sfu, + option_parser_register(opp, "-ptx_opcode_latency_sfu", OPT_CSTR, &opcode_latency_sfu, "Opcode latencies for SFU instructions" "Default 8", "8"); -- cgit v1.3 From 9233f6f9eeea537187deb64add77a320442aa621 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 15:48:26 -0500 Subject: vectoradd is successfully filling the l2 --- src/cuda-sim/cuda-sim.cc | 4 ++++ src/gpgpu-sim/gpu-cache.cc | 23 +++++++++++++++++------ src/gpgpu-sim/gpu-cache.h | 11 +++++++++++ src/gpgpu-sim/gpu-sim.cc | 17 +++++++++++++++++ src/gpgpu-sim/gpu-sim.h | 2 ++ src/gpgpu-sim/l2cache.cc | 9 +++++++++ src/gpgpu-sim/l2cache.h | 6 ++++++ 7 files changed, 66 insertions(+), 6 deletions(-) (limited to 'src/cuda-sim') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index a668db1..685ae53 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -401,6 +401,10 @@ void gpgpu_t::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t coun char *src_data = (char*)src; for (unsigned n=0; n < count; n ++ ) m_global_mem->write(dst_start_addr+n,1, src_data+n,NULL,NULL); + + // Copy into the performance model. + extern gpgpu_sim* g_the_gpu; + g_the_gpu->memcpy_to_gpu(dst_start_addr, src, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 32c2bb1..37fc5ea 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -190,11 +190,17 @@ void tag_array::init( int core_id, int type_id ) m_type_id = type_id; } + enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx, mem_fetch* mf) const { + mem_access_sector_mask_t mask = mf->get_access_sector_mask(); + return probe(addr, idx, mask); +} + + +enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx, mem_access_sector_mask_t mask) const { //assert( m_config.m_write_policy == READ_ONLY ); unsigned set_index = m_config.set_index(addr); new_addr_type tag = m_config.tag(addr); - mem_access_sector_mask_t mask = mf->get_access_sector_mask(); unsigned invalid_line = (unsigned)-1; unsigned valid_line = (unsigned)-1; @@ -310,18 +316,23 @@ enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, void tag_array::fill( new_addr_type addr, unsigned time, mem_fetch* mf) { - assert( m_config.m_alloc_policy == ON_FILL ); + fill(addr, time, mf->get_access_sector_mask()); +} + +void tag_array::fill( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ) +{ + //assert( m_config.m_alloc_policy == ON_FILL ); unsigned idx; - enum cache_request_status status = probe(addr,idx,mf); + enum cache_request_status status = probe(addr,idx,mask); assert(status==MISS||status==SECTOR_MISS); // MSHR should have prevented redundant memory request if(status==MISS) - m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mf->get_access_sector_mask() ); + m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mask ); else if (status==SECTOR_MISS) { assert(m_config.m_cache_type == SECTOR); - ((sector_cache_block*)m_lines[idx])->allocate_sector( time, mf->get_access_sector_mask() ); + ((sector_cache_block*)m_lines[idx])->allocate_sector( time, mask ); } - m_lines[idx]->fill(time, mf->get_access_sector_mask()); + m_lines[idx]->fill(time, mask); } void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf) diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 3e1691a..3713126 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -690,11 +690,13 @@ public: ~tag_array(); enum cache_request_status probe( new_addr_type addr, unsigned &idx, mem_fetch* mf ) const; + enum cache_request_status probe( new_addr_type addr, unsigned &idx, mem_access_sector_mask_t mask ) const; enum cache_request_status access( new_addr_type addr, unsigned time, unsigned &idx, mem_fetch* mf ); enum cache_request_status access( new_addr_type addr, unsigned time, unsigned &idx, bool &wb, evicted_block_info &evicted, mem_fetch* mf ); void fill( new_addr_type addr, unsigned time, mem_fetch* mf ); void fill( unsigned idx, unsigned time, mem_fetch* mf ); + void fill( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ); unsigned size() const { return m_config.get_num_lines();} cache_block_t* get_block(unsigned idx) { return m_lines[idx];} @@ -969,6 +971,15 @@ public: bool data_port_free() const { return m_bandwidth_management.data_port_free(); } bool fill_port_free() const { return m_bandwidth_management.fill_port_free(); } + // This is a gapping hole we are poking in the system to quickly handle + // filling the cache on cudamemcopies. We don't care about anything other than + // L2 state after the memcopy - so just force the tag array to act as though + // something is read or written without doing anything else. + void force_tag_access( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ) + { + m_tag_array->fill( addr, time, mask ); + } + protected: // Constructor that can be used by derived classes with custom tag arrays baseline_cache( const char *name, diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 11ac5df..263cbad 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1595,6 +1595,23 @@ void shader_core_ctx::dump_warp_state( FILE *fout ) const m_warp[w].print(fout); } + +void gpgpu_sim::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ) +{ + assert (dst_start_addr % 32 == 0); + // Right now - I am just going to assume you write the whole last cache line... +// assert (count % 128 == 0); + for ( unsigned counter = 0; counter < count; counter += 32 ) { + const size_t wr_addr = dst_start_addr + counter; + addrdec_t raw_addr; + mem_access_sector_mask_t mask; + mask.set(wr_addr % 128 / 32); + m_memory_config->m_address_mapping.addrdec_tlx( wr_addr, &raw_addr ); + const unsigned partition_id = raw_addr.sub_partition / m_memory_config->m_n_sub_partition_per_memory_channel; + m_memory_partition_unit[ partition_id ]->handle_memcpy_to_gpu( wr_addr, raw_addr.sub_partition, mask ); + } +} + void gpgpu_sim::dump_pipeline( int mask, int s, int m ) const { /* diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index c04648c..6382adf 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -425,6 +425,8 @@ public: void gpu_print_stat(); void dump_pipeline( int mask, int s, int m ) const; + void memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ); + //The next three functions added to be used by the functional simulation function //! Get shader core configuration diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index 8fbf448..1df7858 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -74,6 +74,15 @@ memory_partition_unit::memory_partition_unit( unsigned partition_id, } } +void memory_partition_unit::handle_memcpy_to_gpu( size_t addr, unsigned global_subpart_id, mem_access_sector_mask_t mask ) +{ + unsigned p = global_sub_partition_id_to_local_id(global_subpart_id); + std::string mystring = + mask.to_string(); + MEMPART_DPRINTF("Copy Engine Request Received For Address=%zu, local_subpart=%u, sector_mask=%s \n", addr, p, mystring.c_str()); + m_sub_partition[p]->force_l2_tag_update(addr,gpu_sim_cycle+gpu_tot_sim_cycle, mask); +} + memory_partition_unit::~memory_partition_unit() { delete m_dram; diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 2cc0e76..b613a94 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -72,6 +72,7 @@ public: void print_stat( FILE *fp ) { m_dram->print_stat(fp); } void visualize() const { m_dram->visualize(); } void print( FILE *fp ) const; + void handle_memcpy_to_gpu( size_t dst_start_addr, unsigned subpart_id, mem_access_sector_mask_t mask ); class memory_sub_partition * get_sub_partition(int sub_partition_id) { @@ -178,6 +179,11 @@ public: void accumulate_L2cache_stats(class cache_stats &l2_stats) const; void get_L2cache_sub_stats(struct cache_sub_stats &css) const; + void force_l2_tag_update(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask) + { + m_L2cache->force_tag_access( addr, time, mask ); + } + private: // data unsigned m_id; //< the global sub partition ID -- cgit v1.3 From b9420953855042566ea3371882aef89f66a3f055 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 17:38:24 -0500 Subject: fixing a stupid inheritance bug --- src/cuda-sim/cuda-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.h | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) (limited to 'src/cuda-sim') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 685ae53..5516b1e 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -404,7 +404,7 @@ void gpgpu_t::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t coun // Copy into the performance model. extern gpgpu_sim* g_the_gpu; - g_the_gpu->memcpy_to_gpu(dst_start_addr, src, count); + g_the_gpu->perf_memcpy_to_gpu(dst_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 438769f..0c2ea1c 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1598,7 +1598,7 @@ void shader_core_ctx::dump_warp_state( FILE *fout ) const } -void gpgpu_sim::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ) +void gpgpu_sim::perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ) { if (m_memory_config->m_perf_sim_memcpy) { assert (dst_start_addr % 32 == 0); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 81f13cb..1778008 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -426,7 +426,7 @@ public: void gpu_print_stat(); void dump_pipeline( int mask, int s, int m ) const; - void memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ); + void perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ); //The next three functions added to be used by the functional simulation function -- cgit v1.3 From f125fecc85fdbb29ed846a2acfdb5e8c6cf20e41 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 22:40:27 -0500 Subject: Also initiate our L2 accesses on copies back to the CPU --- src/cuda-sim/cuda-sim.cc | 4 ++++ 1 file changed, 4 insertions(+) (limited to 'src/cuda-sim') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 5516b1e..2f166aa 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -420,6 +420,10 @@ void gpgpu_t::memcpy_from_gpu( void *dst, size_t src_start_addr, size_t count ) unsigned char *dst_data = (unsigned char*)dst; for (unsigned n=0; n < count; n ++ ) m_global_mem->read(src_start_addr+n,1,dst_data+n); + + // Copy into the performance model. + extern gpgpu_sim* g_the_gpu; + g_the_gpu->perf_memcpy_to_gpu(src_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); -- cgit v1.3