Index: /branches/MESI/modules/vci_cc_vcache_wrapper/caba/source/include/vci_cc_vcache_wrapper.h =================================================================== --- /branches/MESI/modules/vci_cc_vcache_wrapper/caba/source/include/vci_cc_vcache_wrapper.h (revision 699) +++ /branches/MESI/modules/vci_cc_vcache_wrapper/caba/source/include/vci_cc_vcache_wrapper.h (revision 700) @@ -560,4 +560,5 @@ sc_signal r_dcache_read_state; sc_signal r_dcache_read_for_modify; // a command intent to write + sc_signal r_dcache_getm_hit; // getm for a shared line sc_signal r_dcache_rsp_state; Index: /branches/MESI/modules/vci_cc_vcache_wrapper/caba/source/src/vci_cc_vcache_wrapper.cpp =================================================================== --- /branches/MESI/modules/vci_cc_vcache_wrapper/caba/source/src/vci_cc_vcache_wrapper.cpp (revision 699) +++ /branches/MESI/modules/vci_cc_vcache_wrapper/caba/source/src/vci_cc_vcache_wrapper.cpp (revision 700) @@ -2330,5 +2330,6 @@ &cache_way, &cache_set, - &cache_word ); + &cache_word + ); r_dcache.write( r_dcache_save_cache_way.read(), @@ -2921,4 +2922,5 @@ r_dcache_vci_miss_req = true; r_dcache_read_for_modify = true; + r_dcache_getm_hit = true; r_dcache_miss_type = PROC_MISS; r_dcache_fsm = DCACHE_MISS_WAIT; @@ -2947,4 +2949,5 @@ r_dcache_fsm = DCACHE_MISS_SELECT; r_dcache_read_for_modify = true; + r_dcache_getm_hit = false; } else @@ -6543,7 +6546,15 @@ p_vci.trdid = 0; if(r_dcache_read_for_modify.read()) - p_vci.pktid = TYPE_WRITE; + { + if(r_dcache_getm_hit.read()) + p_vci.pktid = TYPE_WRITE + 0x8; + else + p_vci.pktid = TYPE_WRITE; + + } else + { p_vci.pktid = TYPE_READ_DATA_MISS; + } p_vci.plen = m_dcache_words << 2; p_vci.cmd = vci_param::CMD_READ; Index: /branches/MESI/modules/vci_mem_cache/caba/source/include/vci_mem_cache.h =================================================================== --- /branches/MESI/modules/vci_mem_cache/caba/source/include/vci_mem_cache.h (revision 699) +++ /branches/MESI/modules/vci_mem_cache/caba/source/include/vci_mem_cache.h (revision 700) @@ -290,5 +290,5 @@ CAS_DIR_HIT_COMPARE, CAS_DIR_HIT_WRITE, - CAS_MULTI_INVAL_IVT_LOCK, + CAS_INVAL_IVT_LOCK, CAS_WAIT, CAS_MULTI_INVAL_HEAP_LOCK, @@ -296,9 +296,4 @@ CAS_HEAP_NEXT, CAS_HEAP_LAST, - CAS_BC_TRT_LOCK, - CAS_BC_IVT_LOCK, - CAS_BC_DIR_INVAL, - CAS_BC_CC_SEND, - CAS_BC_XRAM_REQ, CAS_RSP_FAIL, CAS_RSP_SUCCESS, @@ -361,12 +356,4 @@ }; -// /* States of the ALLOC_UPT fsm */ -// enum alloc_upt_fsm_state_e -// { -// ALLOC_UPT_WRITE, -// ALLOC_UPT_CAS, -// ALLOC_UPT_MULTI_ACK -// }; -// /* States of the ALLOC_IVT fsm */ enum alloc_ivt_fsm_state_e @@ -435,7 +422,14 @@ // Counters accessible in software (not yet but eventually) uint32_t m_cpt_reset_count; // Last cycle at which counters have been reset - uint32_t m_cpt_read_local; // Number of local READ transactions - uint32_t m_cpt_read_remote; // number of remote READ transactions - uint32_t m_cpt_read_cost; // Number of (flits * distance) for READs + uint32_t m_cpt_read_miss_local; // Number of local READ transactions + uint32_t m_cpt_read_miss_remote; // number of remote READ transactions + uint32_t m_cpt_read_miss_cost; // Number of (flits * distance) for READs + + uint32_t m_cpt_getm_miss_local; // Number of local getm miss transactions + uint32_t m_cpt_getm_miss_remote; // number of remote getm miss transactions + + uint32_t m_cpt_getm_hit_local; // Number of local getm hit transactions + uint32_t m_cpt_getm_hit_remote; // number of remote getm hit transactions + uint32_t m_cpt_getm_cost; // Number of (flits * distance) for getm uint32_t m_cpt_write_local; // Number of local WRITE transactions @@ -471,11 +465,27 @@ uint32_t m_cpt_cleanup_local; // Number of local CLEANUP transactions uint32_t m_cpt_cleanup_remote; // Number of remote CLEANUP transactions + uint32_t m_cpt_cleanup_with_data; uint32_t m_cpt_cleanup_cost; // Number of (flits * distance) for CLEANUPs + uint32_t m_cpt_multi_ack_miss; // Number of local CLEANUP transactions + uint32_t m_cpt_multi_ack_hit; // Number of local CLEANUP transactions + uint32_t m_cpt_multi_ack_hit_with_data; // Number of remote CLEANUP transactions + // Counters not accessible by software uint32_t m_cpt_read_miss; // Number of MISS READ - uint32_t m_cpt_write_miss; // Number of MISS WRITE - uint32_t m_cpt_write_dirty; // Cumulated length for WRITE transactions + uint32_t m_cpt_getm_miss; // Number of getm miss + uint32_t m_cpt_getm_broadcast; // Number of BROADCAST INVAL because getm + uint32_t m_cpt_getm_minval; // Number of MULTI INVAL because getm + uint32_t m_cpt_write_miss; // Number of MISS WRITE + uint32_t m_cpt_write_dirty; // Cumulated length for WRITE transactions uint32_t m_cpt_write_broadcast;// Number of BROADCAST INVAL because write + uint32_t m_cpt_write_minval; // Number of MULTI INVAL because write + uint32_t m_cpt_cas_broadcast; // Number of BROADCAST INVAL because cas + uint32_t m_cpt_cas_minval; // Number of MULTI INVAL because cas + uint32_t m_cpt_cas_miss; + + uint32_t m_cpt_read_locked_rb; // Read blocked by a locked ligne + uint32_t m_cpt_cas_locked_rb; // cas blocked by a locked ligne + uint32_t m_cpt_write_locked_rb; // wt blocked by a locked ligne uint32_t m_cpt_trt_rb; // Read blocked by a hit in trt @@ -492,5 +502,7 @@ uint32_t m_cpt_cleanup_fsm_dir_lock; // wait DIR LOCK uint32_t m_cpt_cleanup_fsm_n_dir_lock; // NB DIR LOCK - + uint32_t m_cpt_multi_ack_fsm_dir_lock; // wait DIR LOCK + uint32_t m_cpt_multi_ack_fsm_n_dir_lock; // NB DIR LOCK + uint32_t m_cpt_dir_unused; // NB cycles DIR LOCK unused uint32_t m_cpt_read_fsm_dir_used; // NB cycles DIR LOCK used @@ -499,4 +511,5 @@ uint32_t m_cpt_xram_rsp_fsm_dir_used; // NB cycles DIR LOCK used uint32_t m_cpt_cleanup_fsm_dir_used; // NB cycles DIR LOCK used + uint32_t m_cpt_multi_ack_fsm_dir_used; // NB cycles DIR LOCK used uint32_t m_cpt_read_fsm_trt_lock; // wait TRT LOCK @@ -510,5 +523,6 @@ uint32_t m_cpt_cas_fsm_n_trt_lock; // NB TRT LOCK uint32_t m_cpt_xram_rsp_fsm_n_trt_lock; // NB TRT LOCK - uint32_t m_cpt_ixr_fsm_n_trt_lock; // NB TRT LOCK + uint32_t m_cpt_ixr_cmd_fsm_n_trt_lock; // NB TRT LOCK + uint32_t m_cpt_ixr_rsp_fsm_n_trt_lock; // NB TRT LOCK uint32_t m_cpt_read_fsm_trt_used; // NB cycles TRT LOCK used @@ -516,28 +530,17 @@ uint32_t m_cpt_cas_fsm_trt_used; // NB cycles TRT LOCK used uint32_t m_cpt_xram_rsp_fsm_trt_used; // NB cycles TRT LOCK used - uint32_t m_cpt_ixr_fsm_trt_used; // NB cycles TRT LOCK used + uint32_t m_cpt_ixr_cmd_fsm_trt_used; // NB cycles TRT LOCK used + uint32_t m_cpt_ixr_rsp_fsm_trt_used; // NB cycles TRT LOCK used + uint32_t m_cpt_cleanup_fsm_trt_used; // NB cycles TRT LOCK used uint32_t m_cpt_trt_unused; // NB cycles TRT LOCK unused - uint32_t m_cpt_write_fsm_upt_lock; // wait UPT LOCK - uint32_t m_cpt_xram_rsp_fsm_upt_lock; // wait UPT LOCK - uint32_t m_cpt_multi_ack_fsm_upt_lock; // wait UPT LOCK - uint32_t m_cpt_cleanup_fsm_ivt_lock; // wait UPT LOCK - uint32_t m_cpt_cas_fsm_upt_lock; // wait UPT LOCK + uint32_t m_cpt_cleanup_fsm_ivt_lock; // wait ivt LOCK + uint32_t m_cpt_cleanup_fsm_n_ivt_lock; // NB cycles UPT LOCK used - uint32_t m_cpt_write_fsm_n_upt_lock; // NB UPT LOCK - uint32_t m_cpt_xram_rsp_fsm_n_upt_lock; // NB UPT LOCK - uint32_t m_cpt_multi_ack_fsm_n_upt_lock; // NB UPT LOCK - uint32_t m_cpt_cleanup_fsm_n_upt_lock; // NB UPT LOCK - uint32_t m_cpt_cas_fsm_n_upt_lock; // NB UPT LOCK - - uint32_t m_cpt_write_fsm_upt_used; // NB cycles UPT LOCK used - uint32_t m_cpt_xram_rsp_fsm_upt_used; // NB cycles UPT LOCK used - uint32_t m_cpt_multi_ack_fsm_upt_used; // NB cycles UPT LOCK used - uint32_t m_cpt_cleanup_fsm_ivt_used; // NB cycles UPT LOCK used - uint32_t m_cpt_cas_fsm_upt_used; // NB cycles UPT LOCK used - + uint32_t m_cpt_multi_ack_fsm_ivt_lock; // wait ivt LOCK + uint32_t m_cpt_multi_ack_fsm_n_ivt_lock; // NB cycles UPT LOCK used + uint32_t m_cpt_ivt_unused; // NB cycles UPT LOCK unused - uint32_t m_cpt_upt_unused; // NB cycles UPT LOCK unused uint32_t m_cpt_read_fsm_heap_lock; // wait HEAP LOCK @@ -546,4 +549,5 @@ uint32_t m_cpt_cleanup_fsm_heap_lock; // wait HEAP LOCK uint32_t m_cpt_xram_rsp_fsm_heap_lock; // wait HEAP LOCK + uint32_t m_cpt_multi_ack_fsm_heap_lock; // wait HEAP LOCK uint32_t m_cpt_read_fsm_n_heap_lock; // NB HEAP LOCK @@ -552,4 +556,5 @@ uint32_t m_cpt_cleanup_fsm_n_heap_lock; // NB HEAP LOCK uint32_t m_cpt_xram_rsp_fsm_n_heap_lock; // NB HEAP LOCK + uint32_t m_cpt_multi_ack_fsm_n_heap_lock; // NB HEAP LOCK uint32_t m_cpt_read_fsm_heap_used; // NB cycles HEAP LOCK used @@ -558,29 +563,8 @@ uint32_t m_cpt_cleanup_fsm_heap_used; // NB cycles HEAP LOCK used uint32_t m_cpt_xram_rsp_fsm_heap_used; // NB cycles HEAP LOCK used + uint32_t m_cpt_multi_ack_fsm_heap_used; // NB cycles HEAP LOCK used uint32_t m_cpt_heap_unused; // NB cycles HEAP LOCK unused - - //RWT - uint32_t m_cpt_cleanup_data; - uint32_t m_cpt_ncc_to_cc_read; // NB change from NCC to CC caused by a READ - uint32_t m_cpt_ncc_to_cc_write; // NB change from NCC to CC caused by a WRITE - uint32_t m_cpt_ncc_to_cc; // NB change from NCC to CC - - uint32_t m_cpt_read_data_unc; - uint32_t m_cpt_read_data_miss_CC; - uint32_t m_cpt_read_ins_unc; - uint32_t m_cpt_read_ins_miss; - uint32_t m_cpt_read_ll_CC; - uint32_t m_cpt_read_data_miss_NCC; - uint32_t m_cpt_read_ll_NCC; - uint32_t m_cpt_read_WTF; - - uint32_t m_cpt_update_flits; // Number of flits for UPDATEs - uint32_t m_cpt_inval_cost; // Number of (flits * distance) for INVALs - - uint32_t m_cpt_get; - uint32_t m_cpt_put; - - size_t m_prev_count; + uint32_t m_cpt_slot_inval; protected: Index: /branches/MESI/modules/vci_mem_cache/caba/source/src/vci_mem_cache.cpp =================================================================== --- /branches/MESI/modules/vci_mem_cache/caba/source/src/vci_mem_cache.cpp (revision 699) +++ /branches/MESI/modules/vci_mem_cache/caba/source/src/vci_mem_cache.cpp (revision 700) @@ -250,5 +250,5 @@ "CAS_DIR_HIT_COMPARE", "CAS_DIR_HIT_WRITE", - "CAS_MULTI_INVAL_IVT_LOCK", + "CAS_INVAL_IVT_LOCK", "CAS_WAIT", "CAS_MULTI_INVAL_HEAP_LOCK", @@ -256,9 +256,4 @@ "CAS_HEAP_NEXT", "CAS_HEAP_LAST", - "CAS_BC_TRT_LOCK", - "CAS_BC_IVT_LOCK", - "CAS_BC_DIR_INVAL", - "CAS_BC_CC_SEND", - "CAS_BC_XRAM_REQ", "CAS_RSP_FAIL", "CAS_RSP_SUCCESS", @@ -314,10 +309,5 @@ "ALLOC_TRT_CONFIG" }; -// const char *alloc_upt_fsm_str[] = -// { -// "ALLOC_UPT_WRITE", -// "ALLOC_UPT_CAS", -// "ALLOC_UPT_MULTI_ACK" -// }; + const char *alloc_ivt_fsm_str[] = { @@ -716,5 +706,5 @@ // LOCAL - case MEMC_LOCAL_READ_LO : rdata = m_cpt_read_local ; break; + case MEMC_LOCAL_READ_LO : rdata = m_cpt_read_miss_local ; break; case MEMC_LOCAL_WRITE_LO : rdata = m_cpt_write_flits_local ; break; case MEMC_LOCAL_LL_LO : rdata = m_cpt_ll_local ; break; @@ -729,5 +719,5 @@ // REMOTE - case MEMC_REMOTE_READ_LO : rdata = m_cpt_read_remote ; break; + case MEMC_REMOTE_READ_LO : rdata = m_cpt_read_miss_remote ; break; case MEMC_REMOTE_WRITE_LO : rdata = m_cpt_write_flits_remote ; break; case MEMC_REMOTE_LL_LO : rdata = m_cpt_ll_remote ; break; @@ -742,5 +732,5 @@ // COST - case MEMC_COST_READ_LO : rdata = m_cpt_read_cost ; break; + case MEMC_COST_READ_LO : rdata = m_cpt_read_miss_cost ; break; case MEMC_COST_WRITE_LO : rdata = m_cpt_write_cost; break; case MEMC_COST_LL_LO : rdata = m_cpt_ll_cost ; break; @@ -838,43 +828,151 @@ { m_cpt_reset_count = m_cpt_cycles; - m_cpt_read_local = 0; - m_cpt_read_remote = 0; - m_cpt_read_cost = 0; - m_cpt_write_local = 0; - m_cpt_write_remote = 0; - m_cpt_write_flits_local = 0; - m_cpt_write_flits_remote = 0; - m_cpt_write_cost = 0; - m_cpt_ll_local = 0; - m_cpt_ll_remote = 0; - m_cpt_ll_cost = 0; - m_cpt_sc_local = 0; - m_cpt_sc_remote = 0; - m_cpt_sc_cost = 0; - m_cpt_cas_local = 0; - m_cpt_cas_remote = 0; - m_cpt_cas_cost = 0; - m_cpt_update = 0; - m_cpt_update_local = 0; - m_cpt_update_remote = 0; - m_cpt_update_cost = 0; - m_cpt_minval = 0; - m_cpt_minval_local = 0; - m_cpt_minval_remote = 0; - m_cpt_minval_cost = 0; - m_cpt_binval = 0; - m_cpt_cleanup_local = 0; - m_cpt_cleanup_remote = 0; - m_cpt_cleanup_cost = 0; - m_cpt_read_miss = 0; - m_cpt_write_miss = 0; - m_cpt_write_dirty = 0; - m_cpt_trt_rb = 0; - m_cpt_trt_full = 0; - m_cpt_get = 0; - m_cpt_put = 0; - m_cpt_ncc_to_cc_read = 0; - m_cpt_ncc_to_cc_write = 0; - m_cpt_ncc_to_cc = 0; + + m_cpt_read_miss_local = 0 ; + m_cpt_read_miss_remote = 0 ; + m_cpt_read_miss_cost = 0 ; + + m_cpt_getm_miss_local = 0 ; + m_cpt_getm_miss_remote = 0 ; + + m_cpt_getm_hit_local = 0 ; + m_cpt_getm_hit_remote = 0 ; + m_cpt_getm_cost = 0 ; + + m_cpt_write_local = 0 ; + m_cpt_write_remote = 0 ; + m_cpt_write_flits_local = 0 ; + m_cpt_write_flits_remote = 0 ; + m_cpt_write_cost = 0 ; + + m_cpt_ll_local = 0 ; + m_cpt_ll_remote = 0 ; + m_cpt_ll_cost = 0 ; + + m_cpt_sc_local = 0 ; + m_cpt_sc_remote = 0 ; + m_cpt_sc_cost = 0 ; + + m_cpt_cas_local = 0 ; + m_cpt_cas_remote = 0 ; + m_cpt_cas_cost = 0 ; + + m_cpt_update = 0 ; + m_cpt_update_local = 0 ; + m_cpt_update_remote = 0 ; + m_cpt_update_cost = 0 ; + + m_cpt_minval = 0 ; + m_cpt_minval_local = 0 ; + m_cpt_minval_remote = 0 ; + m_cpt_minval_cost = 0 ; + + m_cpt_binval = 0 ; + + m_cpt_cleanup_local = 0 ; + m_cpt_cleanup_remote = 0 ; + m_cpt_cleanup_with_data = 0 ; + m_cpt_cleanup_cost = 0 ; + + m_cpt_multi_ack_miss = 0 ; + m_cpt_multi_ack_hit = 0 ; + m_cpt_multi_ack_hit_with_data = 0 ; + + + m_cpt_read_miss = 0 ; + m_cpt_getm_miss = 0 ; + m_cpt_getm_broadcast = 0 ; + m_cpt_getm_minval = 0 ; + m_cpt_write_miss = 0 ; + m_cpt_write_dirty = 0 ; + m_cpt_write_broadcast = 0 ; + m_cpt_write_minval = 0 ; + m_cpt_cas_broadcast = 0 ; + m_cpt_cas_minval = 0 ; + m_cpt_cas_miss = 0 ; + + m_cpt_read_locked_rb = 0 ; + m_cpt_cas_locked_rb = 0 ; + m_cpt_write_locked_rb = 0 ; + + m_cpt_trt_rb = 0 ; + m_cpt_trt_full = 0 ; + + m_cpt_read_fsm_dir_lock = 0 ; + m_cpt_read_fsm_n_dir_lock = 0 ; + m_cpt_write_fsm_dir_lock = 0 ; + m_cpt_write_fsm_n_dir_lock = 0 ; + m_cpt_xram_rsp_fsm_dir_lock = 0 ; + m_cpt_xram_rsp_fsm_n_dir_lock = 0 ; + m_cpt_cas_fsm_dir_lock = 0 ; + m_cpt_cas_fsm_n_dir_lock = 0 ; + m_cpt_cleanup_fsm_dir_lock = 0 ; + m_cpt_cleanup_fsm_n_dir_lock = 0 ; + m_cpt_multi_ack_fsm_dir_lock = 0 ; + m_cpt_multi_ack_fsm_n_dir_lock = 0 ; + + m_cpt_dir_unused = 0 ; + m_cpt_read_fsm_dir_used = 0 ; + m_cpt_write_fsm_dir_used = 0 ; + m_cpt_cas_fsm_dir_used = 0 ; + m_cpt_xram_rsp_fsm_dir_used = 0 ; + m_cpt_cleanup_fsm_dir_used = 0 ; + m_cpt_multi_ack_fsm_dir_used = 0 ; + + m_cpt_read_fsm_trt_lock = 0 ; + m_cpt_write_fsm_trt_lock = 0 ; + m_cpt_cas_fsm_trt_lock = 0 ; + m_cpt_xram_rsp_fsm_trt_lock = 0 ; + m_cpt_ixr_fsm_trt_lock = 0 ; + + m_cpt_read_fsm_n_trt_lock = 0 ; + m_cpt_write_fsm_n_trt_lock = 0 ; + m_cpt_cas_fsm_n_trt_lock = 0 ; + m_cpt_xram_rsp_fsm_n_trt_lock = 0 ; + m_cpt_ixr_cmd_fsm_n_trt_lock = 0 ; + m_cpt_ixr_rsp_fsm_n_trt_lock = 0 ; + + m_cpt_read_fsm_trt_used = 0 ; + m_cpt_write_fsm_trt_used = 0 ; + m_cpt_cas_fsm_trt_used = 0 ; + m_cpt_xram_rsp_fsm_trt_used = 0 ; + m_cpt_ixr_cmd_fsm_trt_used = 0 ; + m_cpt_ixr_rsp_fsm_trt_used = 0 ; + m_cpt_cleanup_fsm_trt_used = 0 ; + + m_cpt_trt_unused = 0 ; + + m_cpt_cleanup_fsm_ivt_lock = 0 ; + m_cpt_cleanup_fsm_n_ivt_lock = 0 ; + + m_cpt_multi_ack_fsm_ivt_lock = 0 ; + m_cpt_multi_ack_fsm_n_ivt_lock = 0 ; + + m_cpt_ivt_unused = 0 ; + + m_cpt_read_fsm_heap_lock = 0 ; + m_cpt_write_fsm_heap_lock = 0 ; + m_cpt_cas_fsm_heap_lock = 0 ; + m_cpt_cleanup_fsm_heap_lock = 0 ; + m_cpt_xram_rsp_fsm_heap_lock = 0 ; + m_cpt_multi_ack_fsm_heap_lock = 0 ; + + m_cpt_read_fsm_n_heap_lock = 0 ; + m_cpt_write_fsm_n_heap_lock = 0 ; + m_cpt_cas_fsm_n_heap_lock = 0 ; + m_cpt_cleanup_fsm_n_heap_lock = 0 ; + m_cpt_xram_rsp_fsm_n_heap_lock = 0 ; + m_cpt_multi_ack_fsm_n_heap_lock = 0 ; + + m_cpt_read_fsm_heap_used = 0 ; + m_cpt_write_fsm_heap_used = 0 ; + m_cpt_cas_fsm_heap_used = 0 ; + m_cpt_cleanup_fsm_heap_used = 0 ; + m_cpt_xram_rsp_fsm_heap_used = 0 ; + m_cpt_multi_ack_fsm_heap_used = 0 ; + + m_cpt_heap_unused = 0 ; + m_cpt_slot_inval = 0 ; + } @@ -894,56 +992,72 @@ << "[001] NUMBER OF CYCLES = " << m_cpt_cycles << std::endl << std::endl - << "[002] LOCAL READ = " << m_cpt_read_local << std::endl - << "[003] REMOTE READ = " << m_cpt_read_remote << std::endl - << "[004] READ COST (FLITS * DIST) = " << m_cpt_read_cost << std::endl + << "[002] LOCAL READ = " << m_cpt_read_miss_local << std::endl + << "[003] REMOTE READ = " << m_cpt_read_miss_remote << std::endl + << "[004] READ COST (FLITS * DIST) = " << m_cpt_read_miss_cost << std::endl << std::endl - << "[005] LOCAL WRITE = " << m_cpt_write_local << std::endl - << "[006] REMOTE WRITE = " << m_cpt_write_remote << std::endl - << "[007] WRITE FLITS LOCAL = " << m_cpt_write_flits_local << std::endl - << "[008] WRITE FLITS REMOTE = " << m_cpt_write_flits_remote << std::endl - << "[009] WRITE COST (FLITS * DIST) = " << m_cpt_write_cost << std::endl + << "[005] LOCAL GETM MISS = " << m_cpt_getm_miss_local << std::endl + << "[006] REMOTE GETM MISS = " << m_cpt_getm_miss_remote << std::endl + << "[007] LOCAL GETM HIT = " << m_cpt_getm_hit_local << std::endl + << "[008] REMOTE GETM HIT = " << m_cpt_getm_hit_remote << std::endl + << "[009] GETM COST (FLITS * DIST) = " << m_cpt_getm_cost << std::endl << std::endl - << "[010] LOCAL LL = " << m_cpt_ll_local << std::endl - << "[011] REMOTE LL = " << m_cpt_ll_remote << std::endl - << "[012] LL COST (FLITS * DIST) = " << m_cpt_ll_cost << std::endl + << "[010] LOCAL WRITE = " << m_cpt_write_local << std::endl + << "[011] REMOTE WRITE = " << m_cpt_write_remote << std::endl + << "[012] WRITE FLITS LOCAL = " << m_cpt_write_flits_local << std::endl + << "[013] WRITE FLITS REMOTE = " << m_cpt_write_flits_remote << std::endl + << "[014] WRITE COST (FLITS * DIST) = " << m_cpt_write_cost << std::endl << std::endl - << "[013] LOCAL SC = " << m_cpt_sc_local << std::endl - << "[014] REMOTE SC = " << m_cpt_sc_remote << std::endl - << "[015] SC COST (FLITS * DIST) = " << m_cpt_sc_cost << std::endl + << "[015] LOCAL LL = " << m_cpt_ll_local << std::endl + << "[016] REMOTE LL = " << m_cpt_ll_remote << std::endl + << "[017] LL COST (FLITS * DIST) = " << m_cpt_ll_cost << std::endl << std::endl - << "[016] LOCAL CAS = " << m_cpt_cas_local << std::endl - << "[017] REMOTE CAS = " << m_cpt_cas_remote << std::endl - << "[018] CAS COST (FLITS * DIST) = " << m_cpt_cas_cost << std::endl + << "[018] LOCAL SC = " << m_cpt_sc_local << std::endl + << "[019] REMOTE SC = " << m_cpt_sc_remote << std::endl + << "[020] SC COST (FLITS * DIST) = " << m_cpt_sc_cost << std::endl << std::endl - << "[019] REQUESTS TRIG. UPDATE = " << m_cpt_update << std::endl - << "[020] LOCAL UPDATE = " << m_cpt_update_local << std::endl - << "[021] REMOTE UPDATE = " << m_cpt_update_remote << std::endl - << "[022] UPDT COST (FLITS * DIST) = " << m_cpt_update_cost << std::endl + << "[021] LOCAL CAS = " << m_cpt_cas_local << std::endl + << "[022] REMOTE CAS = " << m_cpt_cas_remote << std::endl + << "[023] CAS COST (FLITS * DIST) = " << m_cpt_cas_cost << std::endl << std::endl - << "[023] REQUESTS TRIG. M_INV = " << m_cpt_minval << std::endl - << "[024] LOCAL M_INV = " << m_cpt_minval_local << std::endl - << "[025] REMOTE M_INV = " << m_cpt_minval_remote << std::endl - << "[026] M_INV COST (FLITS * DIST) = " << m_cpt_minval_cost << std::endl + << "[024] REQUESTS TRIG. UPDATE = " << m_cpt_update << std::endl + << "[025] LOCAL UPDATE = " << m_cpt_update_local << std::endl + << "[026] REMOTE UPDATE = " << m_cpt_update_remote << std::endl + << "[027] UPDT COST (FLITS * DIST) = " << m_cpt_update_cost << std::endl << std::endl - << "[027] BROADCAT INVAL = " << m_cpt_binval << std::endl + << "[028] REQUESTS TRIG. M_INV = " << m_cpt_minval << std::endl + << "[029] LOCAL M_INV = " << m_cpt_minval_local << std::endl + << "[030] REMOTE M_INV = " << m_cpt_minval_remote << std::endl + << "[031] M_INV COST (FLITS * DIST) = " << m_cpt_minval_cost << std::endl << std::endl - << "[028] LOCAL CLEANUP = " << m_cpt_cleanup_local << std::endl - << "[029] REMOTE CLEANUP = " << m_cpt_cleanup_remote << std::endl - << "[030] CLNUP COST (FLITS * DIST) = " << m_cpt_cleanup_cost << std::endl + << "[032] BROADCAT INVAL = " << m_cpt_binval << std::endl << std::endl + << "[033] LOCAL CLEANUP = " << m_cpt_cleanup_local << std::endl + << "[034] REMOTE CLEANUP = " << m_cpt_cleanup_remote << std::endl + << "[035] CLEANUP DATA = " << m_cpt_cleanup_with_data << std::endl + << "[036] CLNUP COST (FLITS * DIST) = " << m_cpt_cleanup_cost << std::endl << std::endl - << "[031] READ MISS = " << m_cpt_read_miss << std::endl - << "[032] WRITE MISS = " << m_cpt_write_miss << std::endl - << "[033] WRITE DIRTY = " << m_cpt_write_dirty << std::endl - << "[034] RD BLOCKED BY HIT IN TRT = " << m_cpt_trt_rb << std::endl - << "[035] TRANS BLOCKED BY FULL TRT = " << m_cpt_trt_full << std::endl - << "[036] PUT (UNIMPLEMENTED) = " << m_cpt_put << std::endl - << "[037] GET (UNIMPLEMENTED) = " << m_cpt_get << std::endl - << "[038] WRITE BROADCAST = " << m_cpt_write_broadcast << std::endl + << "[037] MULTI_ACK_MISS = " << m_cpt_multi_ack_miss << std::endl + << "[038] MULTI_ACK_HIT = " << m_cpt_multi_ack_hit << std::endl + << "[039] MULTI_ACK_DATA = " << m_cpt_multi_ack_hit_with_data << std::endl << std::endl - << "[039] CLEANUP DATA (FLITS) = " << m_cpt_cleanup_data * 16 << std::endl - << "[040] NCC TO CC (READ) = " << m_cpt_ncc_to_cc_read << std::endl - << "[041] NCC TO CC (WRITE) = " << m_cpt_ncc_to_cc_write << std::endl - << "[042] NCC TO CC (TOTAL) = " << m_cpt_ncc_to_cc << std::endl + << "[040] READ MISS = " << m_cpt_read_miss << std::endl + << "[041] GETM MISS = " << m_cpt_getm_miss << std::endl + << "[042] WRITE MISS = " << m_cpt_write_miss << std::endl + << "[043] CAS MISS = " << m_cpt_cas_miss << std::endl + << "[044] WRITE DIRTY = " << m_cpt_write_dirty << std::endl + << "[045] RD BLOCKED BY HIT IN TRT = " << m_cpt_trt_rb << std::endl + << "[046] TRANS BLOCKED BY FULL TRT = " << m_cpt_trt_full << std::endl + << "[047] WRITE BROADCAST = " << m_cpt_write_broadcast << std::endl + << "[048] WRITE MINVAL = " << m_cpt_write_minval << std::endl + << "[049] CAS BROADCAST = " << m_cpt_cas_broadcast << std::endl + << "[050] CAS MINVAL = " << m_cpt_cas_minval << std::endl + << "[051] GETM BROADCAST = " << m_cpt_getm_broadcast << std::endl + << "[052] GETM MINVAL = " << m_cpt_getm_minval << std::endl + << std::endl + << "[053] READ LOCKED WAIT = " << m_cpt_read_locked_rb << std::endl + << "[054] CAS LOCKED WAIT = " << m_cpt_cas_locked_rb << std::endl + << "[055] WRITE LOCKED WAIT = " << m_cpt_write_locked_rb << std::endl + << "[056] EVICTED SLOT = " << m_cpt_slot_inval << std::endl + << std::endl; } @@ -954,9 +1068,22 @@ std::cout << "----------------------------------" << std::dec << std::endl; std::cout - << "[100] READ TOTAL = " << m_cpt_read_local + m_cpt_read_remote << std::endl - << "[101] READ RATE = " << (double) (m_cpt_read_local + m_cpt_read_remote) / m_cpt_cycles << std::endl - << "[102] LOCAL READ RATE = " << (double) m_cpt_read_local / m_cpt_cycles << std::endl - << "[103] REMOTE READ RATE = " << (double) m_cpt_read_remote / m_cpt_cycles << std::endl - << "[104] READ MISS RATE = " << (double) m_cpt_read_miss / (m_cpt_read_local + m_cpt_read_remote) << std::endl + << "[100] READ TOTAL = " << m_cpt_read_miss_local + m_cpt_read_miss_remote << std::endl + << "[101] READ RATE = " << (double) (m_cpt_read_miss_local + m_cpt_read_miss_remote) / m_cpt_cycles << std::endl + << "[102] LOCAL READ RATE = " << (double) m_cpt_read_miss_local / m_cpt_cycles << std::endl + << "[103] REMOTE READ RATE = " << (double) m_cpt_read_miss_remote / m_cpt_cycles << std::endl + << "[104] READ MISS RATE = " << (double) m_cpt_read_miss / (m_cpt_read_miss_local + m_cpt_read_miss_remote) << std::endl + << "[100] GETM MISS TOTAL = " << m_cpt_getm_miss_local + m_cpt_getm_miss_remote << std::endl + << "[101] GETM MISS RATE = " << (double) (m_cpt_getm_miss_local + m_cpt_getm_miss_remote) / m_cpt_cycles << std::endl + << "[102] LOCAL GETM MISS RATE = " << (double) m_cpt_getm_miss_local / m_cpt_cycles << std::endl + << "[103] REMOTE GETM MISS RATE = " << (double) m_cpt_getm_miss_remote / m_cpt_cycles << std::endl + << "[100] GETM HIT TOTAL = " << m_cpt_getm_hit_local + m_cpt_getm_hit_remote << std::endl + << "[101] GETM HIT RATE = " << (double) (m_cpt_getm_hit_local + m_cpt_getm_hit_remote) / m_cpt_cycles << std::endl + << "[102] LOCAL GETM HIT RATE = " << (double) m_cpt_getm_hit_local / m_cpt_cycles << std::endl + << "[103] REMOTE GETM HIT RATE = " << (double) m_cpt_getm_hit_remote / m_cpt_cycles << std::endl + + << "[100] GETM TOTAL = " << m_cpt_getm_hit_local + m_cpt_getm_hit_remote + m_cpt_getm_miss_remote + m_cpt_getm_miss_local << std::endl + + << "[104] GETM MISS RATE = " << (double) m_cpt_read_miss / (m_cpt_getm_hit_local + m_cpt_getm_hit_remote + m_cpt_getm_miss_remote + m_cpt_getm_miss_local) << std::endl + << std::endl << "[105] WRITE TOTAL = " << m_cpt_write_local + m_cpt_write_remote << std::endl @@ -995,4 +1122,5 @@ << "[134] LOCAL CAS RATE = " << (double) m_cpt_cas_local / m_cpt_cycles << std::endl << "[135] REMOTE CAS RATE = " << (double) m_cpt_cas_remote / m_cpt_cycles << std::endl + << "[124] MULTI ACK RATE = " << (double) (m_cpt_multi_ack_hit + m_cpt_multi_ack_miss) / m_cpt_cycles << std::endl << std::endl << std::endl; @@ -1187,91 +1315,150 @@ // Activity counters - m_cpt_reset_count = 0; - m_cpt_cycles = 0; - m_cpt_read_local = 0; - m_cpt_read_remote = 0; - m_cpt_read_cost = 0; - m_cpt_write_local = 0; - m_cpt_write_remote = 0; - m_cpt_write_flits_local = 0; - m_cpt_write_flits_remote = 0; - m_cpt_write_cost = 0; - m_cpt_ll_local = 0; - m_cpt_ll_remote = 0; - m_cpt_ll_cost = 0; - m_cpt_sc_local = 0; - m_cpt_sc_remote = 0; - m_cpt_sc_cost = 0; - m_cpt_cas_local = 0; - m_cpt_cas_remote = 0; - m_cpt_cas_cost = 0; - m_cpt_update = 0; - m_cpt_update_local = 0; - m_cpt_update_remote = 0; - m_cpt_update_cost = 0; - m_cpt_minval = 0; - m_cpt_minval_local = 0; - m_cpt_minval_remote = 0; - m_cpt_minval_cost = 0; - m_cpt_binval = 0; - m_cpt_cleanup_local = 0; - m_cpt_cleanup_remote = 0; - m_cpt_cleanup_cost = 0; - - m_cpt_read_miss = 0; - m_cpt_write_miss = 0; - m_cpt_write_dirty = 0; - m_cpt_write_broadcast = 0; - m_cpt_trt_rb = 0; - m_cpt_trt_full = 0; - m_cpt_get = 0; - m_cpt_put = 0; - m_cpt_dir_unused = 0; - m_cpt_upt_unused = 0; - m_cpt_ivt_unused = 0; - m_cpt_heap_unused = 0; - m_cpt_trt_unused = 0; - m_cpt_read_fsm_n_dir_lock = 0; - m_cpt_read_fsm_dir_lock = 0; - m_cpt_read_fsm_dir_used = 0; - m_cpt_read_fsm_trt_lock = 0; - m_cpt_read_fsm_heap_lock = 0; - m_cpt_write_fsm_dir_lock = 0; - m_cpt_write_fsm_n_dir_lock = 0; - m_cpt_write_fsm_upt_lock = 0; - m_cpt_write_fsm_heap_lock = 0; - m_cpt_write_fsm_dir_used = 0; - m_cpt_write_fsm_trt_lock = 0; - m_cpt_cas_fsm_n_dir_lock = 0; - m_cpt_cas_fsm_dir_lock = 0; - m_cpt_cas_fsm_upt_lock = 0; - m_cpt_cas_fsm_heap_lock = 0; - m_cpt_cas_fsm_trt_lock = 0; - m_cpt_cas_fsm_dir_used = 0; - m_cpt_xram_rsp_fsm_n_dir_lock = 0; - m_cpt_xram_rsp_fsm_dir_lock = 0; - m_cpt_xram_rsp_fsm_trt_lock = 0; - m_cpt_xram_rsp_fsm_upt_lock = 0; - m_cpt_xram_rsp_fsm_heap_lock = 0; - m_cpt_xram_rsp_fsm_dir_used = 0; - m_cpt_cleanup_fsm_dir_lock = 0; - m_cpt_cleanup_fsm_n_dir_lock = 0; - m_cpt_cleanup_fsm_heap_lock = 0; - m_cpt_cleanup_fsm_ivt_lock = 0; - m_cpt_cleanup_fsm_dir_used = 0; - m_cpt_ixr_fsm_trt_lock = 0; - m_cpt_multi_ack_fsm_upt_lock = 0; - m_cpt_read_data_unc = 0; - m_cpt_read_data_miss_CC = 0; - m_cpt_read_ins_unc = 0; - m_cpt_read_ins_miss = 0; - m_cpt_read_ll_CC = 0; - m_cpt_read_data_miss_NCC = 0; - m_cpt_read_ll_NCC = 0; - m_cpt_read_WTF = 0; - m_cpt_cleanup_data = 0; - m_cpt_ncc_to_cc_read = 0; - m_cpt_ncc_to_cc_write = 0; - m_cpt_ncc_to_cc = 0; + m_cpt_reset_count = 0 ; + m_cpt_read_miss_local = 0 ; + m_cpt_read_miss_remote = 0 ; + m_cpt_read_miss_cost = 0 ; + + m_cpt_getm_miss_local = 0 ; + m_cpt_getm_miss_remote = 0 ; + + m_cpt_getm_hit_local = 0 ; + m_cpt_getm_hit_remote = 0 ; + m_cpt_getm_cost = 0 ; + + m_cpt_write_local = 0 ; + m_cpt_write_remote = 0 ; + m_cpt_write_flits_local = 0 ; + m_cpt_write_flits_remote = 0 ; + m_cpt_write_cost = 0 ; + + m_cpt_ll_local = 0 ; + m_cpt_ll_remote = 0 ; + m_cpt_ll_cost = 0 ; + + m_cpt_sc_local = 0 ; + m_cpt_sc_remote = 0 ; + m_cpt_sc_cost = 0 ; + + m_cpt_cas_local = 0 ; + m_cpt_cas_remote = 0 ; + m_cpt_cas_cost = 0 ; + + m_cpt_update = 0 ; + m_cpt_update_local = 0 ; + m_cpt_update_remote = 0 ; + m_cpt_update_cost = 0 ; + + m_cpt_minval = 0 ; + m_cpt_minval_local = 0 ; + m_cpt_minval_remote = 0 ; + m_cpt_minval_cost = 0 ; + + m_cpt_binval = 0 ; + + m_cpt_cleanup_local = 0 ; + m_cpt_cleanup_remote = 0 ; + m_cpt_cleanup_with_data = 0 ; + m_cpt_cleanup_cost = 0 ; + + m_cpt_multi_ack_miss = 0 ; + m_cpt_multi_ack_hit = 0 ; + m_cpt_multi_ack_hit_with_data = 0 ; + + + m_cpt_read_miss = 0 ; + m_cpt_getm_miss = 0 ; + m_cpt_getm_broadcast = 0 ; + m_cpt_getm_minval = 0 ; + m_cpt_write_miss = 0 ; + m_cpt_write_dirty = 0 ; + m_cpt_write_broadcast = 0 ; + m_cpt_write_minval = 0 ; + m_cpt_cas_broadcast = 0 ; + m_cpt_cas_minval = 0 ; + m_cpt_cas_miss = 0 ; + + m_cpt_read_locked_rb = 0 ; + m_cpt_cas_locked_rb = 0 ; + m_cpt_write_locked_rb = 0 ; + + m_cpt_trt_rb = 0 ; + m_cpt_trt_full = 0 ; + + m_cpt_read_fsm_dir_lock = 0 ; + m_cpt_read_fsm_n_dir_lock = 0 ; + m_cpt_write_fsm_dir_lock = 0 ; + m_cpt_write_fsm_n_dir_lock = 0 ; + m_cpt_xram_rsp_fsm_dir_lock = 0 ; + m_cpt_xram_rsp_fsm_n_dir_lock = 0 ; + m_cpt_cas_fsm_dir_lock = 0 ; + m_cpt_cas_fsm_n_dir_lock = 0 ; + m_cpt_cleanup_fsm_dir_lock = 0 ; + m_cpt_cleanup_fsm_n_dir_lock = 0 ; + m_cpt_multi_ack_fsm_dir_lock = 0 ; + m_cpt_multi_ack_fsm_n_dir_lock = 0 ; + + m_cpt_dir_unused = 0 ; + m_cpt_read_fsm_dir_used = 0 ; + m_cpt_write_fsm_dir_used = 0 ; + m_cpt_cas_fsm_dir_used = 0 ; + m_cpt_xram_rsp_fsm_dir_used = 0 ; + m_cpt_cleanup_fsm_dir_used = 0 ; + m_cpt_multi_ack_fsm_dir_used = 0 ; + + m_cpt_read_fsm_trt_lock = 0 ; + m_cpt_write_fsm_trt_lock = 0 ; + m_cpt_cas_fsm_trt_lock = 0 ; + m_cpt_xram_rsp_fsm_trt_lock = 0 ; + m_cpt_ixr_fsm_trt_lock = 0 ; + + m_cpt_read_fsm_n_trt_lock = 0 ; + m_cpt_write_fsm_n_trt_lock = 0 ; + m_cpt_cas_fsm_n_trt_lock = 0 ; + m_cpt_xram_rsp_fsm_n_trt_lock = 0 ; + m_cpt_ixr_cmd_fsm_n_trt_lock = 0 ; + m_cpt_ixr_rsp_fsm_n_trt_lock = 0 ; + + m_cpt_read_fsm_trt_used = 0 ; + m_cpt_write_fsm_trt_used = 0 ; + m_cpt_cas_fsm_trt_used = 0 ; + m_cpt_xram_rsp_fsm_trt_used = 0 ; + m_cpt_ixr_cmd_fsm_trt_used = 0 ; + m_cpt_ixr_rsp_fsm_trt_used = 0 ; + m_cpt_cleanup_fsm_trt_used = 0 ; + + m_cpt_trt_unused = 0 ; + + m_cpt_cleanup_fsm_ivt_lock = 0 ; + m_cpt_cleanup_fsm_n_ivt_lock = 0 ; + + m_cpt_multi_ack_fsm_ivt_lock = 0 ; + m_cpt_multi_ack_fsm_n_ivt_lock = 0 ; + + m_cpt_ivt_unused = 0 ; + + m_cpt_read_fsm_heap_lock = 0 ; + m_cpt_write_fsm_heap_lock = 0 ; + m_cpt_cas_fsm_heap_lock = 0 ; + m_cpt_cleanup_fsm_heap_lock = 0 ; + m_cpt_xram_rsp_fsm_heap_lock = 0 ; + m_cpt_multi_ack_fsm_heap_lock = 0 ; + + m_cpt_read_fsm_n_heap_lock = 0 ; + m_cpt_write_fsm_n_heap_lock = 0 ; + m_cpt_cas_fsm_n_heap_lock = 0 ; + m_cpt_cleanup_fsm_n_heap_lock = 0 ; + m_cpt_xram_rsp_fsm_n_heap_lock = 0 ; + m_cpt_multi_ack_fsm_n_heap_lock = 0 ; + + m_cpt_read_fsm_heap_used = 0 ; + m_cpt_write_fsm_heap_used = 0 ; + m_cpt_cas_fsm_heap_used = 0 ; + m_cpt_cleanup_fsm_heap_used = 0 ; + m_cpt_xram_rsp_fsm_heap_used = 0 ; + m_cpt_multi_ack_fsm_heap_used = 0 ; + + m_cpt_heap_unused = 0 ; + m_cpt_slot_inval = 0 ; return; } @@ -1830,7 +2017,24 @@ } else { - if (is_local_req(p_vci_tgt.srcid.read())) m_cpt_read_local++; - else m_cpt_read_remote++; - m_cpt_read_cost += m_words * req_distance(p_vci_tgt.srcid.read()); + if ((p_vci_tgt.pktid.read() & 0x7) == TYPE_WRITE) + { + if ((p_vci_tgt.pktid.read() & 0x8) == 0x8) + { + if (is_local_req(p_vci_tgt.srcid.read())) m_cpt_getm_hit_local++; + else m_cpt_getm_hit_remote++; + } + else + { + if (is_local_req(p_vci_tgt.srcid.read())) m_cpt_getm_miss_local++; + else m_cpt_getm_miss_remote++; + } + m_cpt_getm_cost += req_distance(p_vci_tgt.srcid.read()); + } + else + { + if (is_local_req(p_vci_tgt.srcid.read())) m_cpt_read_miss_local++; + else m_cpt_read_miss_remote++; + m_cpt_read_miss_cost += req_distance(p_vci_tgt.srcid.read()); + } } // @@ -1963,4 +2167,9 @@ r_multi_ack_miss_updt = miss_updt; + if(miss_updt) + m_cpt_multi_ack_miss ++; + else + m_cpt_multi_ack_hit ++; + if( not eop) { @@ -1968,4 +2177,5 @@ r_multi_ack_need_data = true; r_multi_ack_data_index = 0; + m_cpt_multi_ack_hit_with_data ++; } else @@ -2028,5 +2238,9 @@ case MULTI_ACK_IVT_LOCK: { + m_cpt_multi_ack_fsm_ivt_lock ++; if(r_alloc_ivt_fsm.read() != ALLOC_IVT_MULTI_ACK) break; + + m_cpt_multi_ack_fsm_n_ivt_lock ++; + assert( r_alloc_dir_fsm.read() == ALLOC_DIR_MULTI_ACK && " BAD ALLOC_DIR_MULTI_ACK IN MULTI_ACK_IVT_LOCK STATE"); size_t index = r_multi_ack_ivt_index.read(); @@ -2090,8 +2304,10 @@ case MULTI_ACK_DIR_REQ: // Get the lock to the directory { + m_cpt_multi_ack_fsm_dir_lock ++; if(r_alloc_dir_fsm.read() != ALLOC_DIR_MULTI_ACK) break; r_multi_ack_fsm = MULTI_ACK_IVT_LOCK; //std::cout << " MEM_CACHE : CLEANUP_DIR_REQ" << std::endl; + m_cpt_multi_ack_fsm_n_dir_lock ++; #if DEBUG_MEMC_CLEANUP @@ -2333,8 +2549,11 @@ case MULTI_ACK_HEAP_REQ: // Get the lock to the HEAP directory { + m_cpt_multi_ack_fsm_n_heap_lock ++; if(r_alloc_heap_fsm.read() == ALLOC_HEAP_MULTI_ACK) { r_multi_ack_fsm = MULTI_ACK_HEAP_LOCK; } + + m_cpt_multi_ack_fsm_heap_lock ++; #if DEBUG_MEMC_MULTI_ACK if(m_debug) @@ -3173,4 +3392,5 @@ else if(entry.state == ENTRY_LOCKED) // the line is locked, can't use it at this time { + m_cpt_read_locked_rb ++; r_read_fsm = READ_WAIT; } @@ -3287,4 +3507,9 @@ cmd_read_fifo_get = true; } + + if(multi_req) + m_cpt_getm_minval ++; + else + m_cpt_getm_broadcast ++; #if DEBUG_MEMC_READ if(m_debug) @@ -3850,5 +4075,10 @@ else // missing line is requested to the XRAM { - m_cpt_read_miss++; + + if ( (m_cmd_read_pktid_fifo.read() & 0x7) == TYPE_WRITE ) + m_cpt_getm_miss ++; + else + m_cpt_read_miss++; + r_read_trt_index = index; r_read_fsm = READ_TRT_SET; @@ -4059,5 +4289,9 @@ // Get the lock to the directory // and access the llsc_global_table - if (r_alloc_dir_fsm.read() != ALLOC_DIR_WRITE ) break; + if (r_alloc_dir_fsm.read() != ALLOC_DIR_WRITE ) + { + m_cpt_write_fsm_n_dir_lock ++; + break; + } if ((r_write_pktid.read() & 0x7) == TYPE_SC) @@ -4087,4 +4321,6 @@ r_write_fsm = WRITE_DIR_LOCK; } + + m_cpt_write_fsm_dir_lock ++; #if DEBUG_MEMC_WRITE @@ -4129,4 +4365,5 @@ else if (entry.state == ENTRY_LOCKED)//the line is locked { + m_cpt_write_locked_rb ++; r_write_fsm = WRITE_WAIT; } @@ -4439,4 +4676,5 @@ } + m_cpt_write_fsm_n_trt_lock ++; assert(false && "VCI_MEM_CACHE ERROR: this part must not be reached"); } @@ -4582,6 +4820,9 @@ false, // is a command write index); - /*// std::cout << " trt wait nline =" << std::hex << nline - << " write data " << std::hex << std::endl;*/ + + if(is_brdcast) + m_cpt_write_broadcast ++; + else + m_cpt_write_minval ++; #if DEBUG_MEMC_WRITE if( m_debug and wok ) @@ -4627,5 +4868,4 @@ } - m_cpt_write_fsm_upt_lock++; break; @@ -4844,4 +5084,5 @@ for( size_t i=0 ; i Line is locked" + << " CAS_INVAL_IVT_LOCK> Line is locked" << " / wok = " << wok << " / set = " << set @@ -7472,196 +7734,5 @@ break; } - - ///////////////////// - case CAS_BC_TRT_LOCK: // check the TRT to register a PUT transaction - { - assert( (r_alloc_dir_fsm.read() == ALLOC_DIR_CAS) and - "MEMC ERROR in CAS_BC_TRT_LOCK state: Bas DIR allocation"); - - if(r_alloc_trt_fsm.read() == ALLOC_TRT_CAS) - { - size_t wok_index = 0; - bool wok = !m_trt.full(wok_index); - if( wok ) - { - r_cas_trt_index = wok_index; - r_cas_fsm = CAS_BC_IVT_LOCK; - } - else - { - r_cas_fsm = CAS_WAIT; - m_cpt_cas_fsm_n_trt_lock++; - } - -#if DEBUG_MEMC_CAS - if(m_debug) - std::cout << " Check TRT" - << " : wok = " << wok << " / index = " << wok_index << std::endl; -#endif - } - m_cpt_cas_fsm_trt_lock++; - - break; - } - ///////////////////// - case CAS_BC_IVT_LOCK: // register a broadcast inval transaction in IVT - // write data in cache in case of successful registration - { - if(r_alloc_ivt_fsm.read() == ALLOC_IVT_CAS) - { - bool wok = false; - size_t index = 0; - size_t srcid = m_cmd_cas_srcid_fifo.read(); - size_t trdid = m_cmd_cas_trdid_fifo.read(); - size_t pktid = m_cmd_cas_pktid_fifo.read(); - addr_t nline = m_nline[(addr_t)(m_cmd_cas_addr_fifo.read())]; - size_t nb_copies = r_cas_count.read(); - - // register a broadcast inval transaction in IVT - wok = m_ivt.set(false, // it's an inval transaction - true, // it's a broadcast - true, // response required - false, // no acknowledge required - srcid, - trdid, - pktid, - nline, - nb_copies, - false, // is a command write - index); - - if(wok) // IVT not full - { - // cache update - size_t way = r_cas_way.read(); - size_t set = m_y[(addr_t)(m_cmd_cas_addr_fifo.read())]; - size_t word = m_x[(addr_t)(m_cmd_cas_addr_fifo.read())]; - - m_cache_data.write(way, set, word, r_cas_wdata.read()); - if(r_cas_cpt.read() ==4) - m_cache_data.write(way, set, word+1, m_cmd_cas_wdata_fifo.read()); - - r_cas_ivt_index = index; - r_cas_fsm = CAS_BC_DIR_INVAL; - -#if DEBUG_MEMC_CAS - if(m_debug) - std::cout << " Register a broadcast inval transaction in IVT" - << " / nline = " << std::hex << nline - << " / count = " << std::dec << nb_copies - << " / ivt_index = " << index << std::endl; -#endif - } - else // releases the lock protecting IVT - { - r_cas_fsm = CAS_WAIT; - } - m_cpt_cas_fsm_n_upt_lock++; - } - - m_cpt_cas_fsm_upt_lock++; - - break; - } - ////////////////////// - case CAS_BC_DIR_INVAL: // Register the PUT transaction in TRT, and inval the DIR entry - { - assert( (r_alloc_dir_fsm.read() == ALLOC_DIR_CAS) and - "MEMC ERROR in CAS_BC_DIR_INVAL state: Bad DIR allocation"); - - assert( (r_alloc_trt_fsm.read() == ALLOC_TRT_CAS) and - "MEMC ERROR in CAS_BC_DIR_INVAL state: Bad TRT allocation"); - - assert( (r_alloc_ivt_fsm.read() == ALLOC_IVT_CAS) and - "MEMC ERROR in CAS_BC_DIR_INVAL state: Bad IVT allocation"); - - std::vector data_vector; - data_vector.clear(); - size_t word = m_x[(addr_t)(m_cmd_cas_addr_fifo.read())]; - - for(size_t i=0; i (m_words,0), - data_vector ); - - // invalidate directory entry - DirectoryEntry entry; - entry.state = ENTRY_INVALID; - entry.dirty = false; - entry.tag = 0; - entry.is_cnt = false; - entry.lock = false; - entry.count = 0; - entry.owner.srcid = 0; - entry.owner.inst = false; - entry.ptr = 0; - size_t set = m_y[(addr_t)(m_cmd_cas_addr_fifo.read())]; - size_t way = r_cas_way.read(); - m_cache_directory.write(set, way, entry); - - r_cas_fsm = CAS_BC_CC_SEND; - -#if DEBUG_MEMC_CAS - if(m_debug) - std::cout << " Inval DIR & register in TRT:" - << " address = " << m_cmd_cas_addr_fifo.read() << std::endl; -#endif - break; - } - /////////////////// - case CAS_BC_CC_SEND: // Request the broadcast inval to CC_SEND FSM - { - if( not r_cas_to_cc_send_multi_req.read() and - not r_cas_to_cc_send_brdcast_req.read()) - { - r_cas_to_cc_send_multi_req = false; - r_cas_to_cc_send_brdcast_req = true; - r_cas_to_cc_send_nline = m_nline[(addr_t)(m_cmd_cas_addr_fifo.read())]; - - r_cas_fsm = CAS_BC_XRAM_REQ; - } - break; - } - //////////////////// - case CAS_BC_XRAM_REQ: // request the IXR FSM to start a put transaction - { - if( not r_cas_to_ixr_cmd_req.read() ) - { - r_cas_to_ixr_cmd_req = true; - r_cas_to_ixr_cmd_index = r_cas_trt_index.read(); - r_cas_fsm = CAS_IDLE; - cmd_cas_fifo_get = true; - r_cas_cpt = 0; - -#if DEBUG_MEMC_CAS - if(m_debug) - std::cout << " Request a PUT transaction to IXR_CMD FSM" << std::hex - << " / address = " << (addr_t) m_cmd_cas_addr_fifo.read() - << " / trt_index = " << r_cas_trt_index.read() << std::endl; -#endif - } - - break; - } - ///////////////// case CAS_RSP_FAIL: // request TGT_RSP FSM to send a failure response { @@ -7738,4 +7809,5 @@ r_cas_trt_index = index; r_cas_fsm = CAS_MISS_TRT_SET; + m_cpt_cas_miss ++; } m_cpt_cas_fsm_n_trt_lock++; @@ -8312,4 +8384,7 @@ { if(not p_dspin_m2p.read) break; + // + m_cpt_binval++; + // r_cc_send_fsm = CC_SEND_READ_BRDCAST_NLINE; break; @@ -8327,4 +8402,26 @@ { if(not p_dspin_m2p.read) break; + // + if (is_local_req(m_read_to_cc_send_srcid_fifo.read())) + { + if(r_read_to_cc_send_type.read()) + m_cpt_minval_local ++; + else + m_cpt_update_local ++; + } + else + { + if(r_read_to_cc_send_type.read()) + { + m_cpt_minval_remote++; + m_cpt_minval_cost += req_distance(m_read_to_cc_send_srcid_fifo.read()); + } + else + { + m_cpt_update_remote++; + m_cpt_update_cost += req_distance(m_read_to_cc_send_srcid_fifo.read()); + } + } + // r_cc_send_fsm = CC_SEND_READ_MULTI_NLINE; break; @@ -8335,5 +8432,8 @@ r_read_to_cc_send_multi_req = false; } - + // + if(r_read_to_cc_send_type.read()) + m_cpt_minval++; + // r_cc_send_fsm = CC_SEND_READ_IDLE; break; @@ -8392,4 +8492,15 @@ { if(not p_dspin_m2p.read) break; + // + if (is_local_req(m_write_to_cc_send_srcid_fifo.read())) + { + m_cpt_minval_local++; + } + else + { + m_cpt_minval_remote++; + m_cpt_minval_cost += req_distance(m_write_to_cc_send_srcid_fifo.read()); + } + // r_cc_send_fsm = CC_SEND_WRITE_INVAL_NLINE; break; @@ -8400,7 +8511,6 @@ r_write_to_cc_send_multi_req = false; } - // - m_cpt_update++; + m_cpt_minval++; // r_cc_send_fsm = CC_SEND_WRITE_IDLE; @@ -8437,5 +8547,7 @@ // - r_cas_to_cc_send_brdcast_req = false; + // + m_cpt_minval++; + // r_cas_to_cc_send_brdcast_req = false; r_cc_send_fsm = CC_SEND_CAS_IDLE; @@ -8454,4 +8566,15 @@ { if(not p_dspin_m2p.read) break; + // + if (is_local_req(m_cas_to_cc_send_srcid_fifo.read())) + { + m_cpt_minval_local++; + } + else + { + m_cpt_minval_remote++; + m_cpt_minval_cost += req_distance(m_cas_to_cc_send_srcid_fifo.read()); + } + // r_cc_send_fsm = CC_SEND_CAS_INVAL_NLINE; break; @@ -8463,5 +8586,7 @@ r_cas_to_cc_send_multi_req = false; } - + // + m_cpt_minval++; + // r_cc_send_fsm = CC_SEND_CAS_IDLE; break; @@ -8902,36 +9027,4 @@ r_read_to_tgt_rsp_req = false; r_tgt_rsp_fsm = TGT_RSP_READ_IDLE; - if (r_read_to_tgt_rsp_pktid.read() == 0x0) - { - m_cpt_read_data_unc ++; - } - else if (r_read_to_tgt_rsp_pktid.read() == 0x1) - { - m_cpt_read_data_miss_CC ++; - } - else if (r_read_to_tgt_rsp_pktid.read() == 0x2) - { - m_cpt_read_ins_unc ++; - } - else if (r_read_to_tgt_rsp_pktid.read() == 0x3) - { - m_cpt_read_ins_miss ++; - } - else if (r_read_to_tgt_rsp_pktid.read() == 0x6) - { - m_cpt_read_ll_CC ++; - } - else if (r_read_to_tgt_rsp_pktid.read() == 0x9) - { - m_cpt_read_data_miss_NCC ++; - } - else if (r_read_to_tgt_rsp_pktid.read() == 0x14) - { - m_cpt_read_ll_NCC ++; - } - else - { - m_cpt_read_WTF ++; - } } else @@ -8994,38 +9087,4 @@ r_cleanup_to_tgt_rsp_req = false; r_tgt_rsp_key_sent = false; - - if (r_cleanup_to_tgt_rsp_pktid.read() == 0x0) - { - m_cpt_read_data_unc ++; - } - else if (r_cleanup_to_tgt_rsp_pktid.read() == 0x1) - { - m_cpt_read_data_miss_CC ++; - } - else if (r_cleanup_to_tgt_rsp_pktid.read() == 0x2) - { - m_cpt_read_ins_unc ++; - } - else if (r_cleanup_to_tgt_rsp_pktid.read() == 0x3) - { - m_cpt_read_ins_miss ++; - } - else if (r_cleanup_to_tgt_rsp_pktid.read() == 0x6) - { - m_cpt_read_ll_CC ++; - } - else if (r_cleanup_to_tgt_rsp_pktid.read() == 0x9) - { - m_cpt_read_data_miss_NCC ++; - } - else if (r_cleanup_to_tgt_rsp_pktid.read() == 0x14) - { - m_cpt_read_ll_NCC ++; - } - else if (!r_cleanup_to_tgt_rsp_type.read()) - { - m_cpt_read_WTF ++; - } - } else @@ -9090,39 +9149,4 @@ r_xram_rsp_to_tgt_rsp_req = false; r_tgt_rsp_fsm = TGT_RSP_XRAM_IDLE; - - - if (r_xram_rsp_to_tgt_rsp_pktid.read() == 0x0) - { - m_cpt_read_data_unc ++; - } - else if (r_xram_rsp_to_tgt_rsp_pktid.read() == 0x1) - { - m_cpt_read_data_miss_CC ++; - } - else if (r_xram_rsp_to_tgt_rsp_pktid.read() == 0x2) - { - m_cpt_read_ins_unc ++; - } - else if (r_xram_rsp_to_tgt_rsp_pktid.read() == 0x3) - { - m_cpt_read_ins_miss ++; - } - else if (r_xram_rsp_to_tgt_rsp_pktid.read() == 0x6) - { - m_cpt_read_ll_CC ++; - } - else if (r_xram_rsp_to_tgt_rsp_pktid.read() == 0x9) - { - m_cpt_read_data_miss_NCC ++; - } - else if (r_xram_rsp_to_tgt_rsp_pktid.read() == 0x14) - { - m_cpt_read_ll_NCC ++; - } - else - { - m_cpt_read_WTF ++; - } - } else @@ -9217,6 +9241,5 @@ r_alloc_ivt_fsm = ALLOC_IVT_CLEANUP; - else if ((r_cas_fsm.read() == CAS_MULTI_INVAL_IVT_LOCK) or - (r_cas_fsm.read() == CAS_BC_IVT_LOCK)) + else if (r_cas_fsm.read() == CAS_INVAL_IVT_LOCK) r_alloc_ivt_fsm = ALLOC_IVT_CAS; @@ -9244,6 +9267,5 @@ r_alloc_ivt_fsm = ALLOC_IVT_CLEANUP; - else if ((r_cas_fsm.read() == CAS_MULTI_INVAL_IVT_LOCK) or - (r_cas_fsm.read() == CAS_BC_IVT_LOCK)) + else if (r_cas_fsm.read() == CAS_INVAL_IVT_LOCK) r_alloc_ivt_fsm = ALLOC_IVT_CAS; @@ -9270,6 +9292,5 @@ r_alloc_ivt_fsm = ALLOC_IVT_CLEANUP; - else if ((r_cas_fsm.read() == CAS_MULTI_INVAL_IVT_LOCK) or - (r_cas_fsm.read() == CAS_BC_IVT_LOCK)) + else if (r_cas_fsm.read() == CAS_INVAL_IVT_LOCK) r_alloc_ivt_fsm = ALLOC_IVT_CAS; @@ -9299,6 +9320,5 @@ (r_cleanup_fsm.read() != CLEANUP_LOCKED_IVT_DECREMENT)) { - if ((r_cas_fsm.read() == CAS_MULTI_INVAL_IVT_LOCK) or - (r_cas_fsm.read() == CAS_BC_IVT_LOCK)) + if (r_cas_fsm.read() == CAS_INVAL_IVT_LOCK) r_alloc_ivt_fsm = ALLOC_IVT_CAS; @@ -9326,6 +9346,5 @@ ////////////////////////// case ALLOC_IVT_CAS: // allocated to CAS FSM - if ((r_cas_fsm.read() != CAS_MULTI_INVAL_IVT_LOCK) and - (r_cas_fsm.read() != CAS_BC_IVT_LOCK)) + if (r_cas_fsm.read() != CAS_INVAL_IVT_LOCK) { if (r_config_fsm.read() == CONFIG_IVT_LOCK) @@ -9375,6 +9394,5 @@ r_alloc_ivt_fsm = ALLOC_IVT_CLEANUP; - else if ((r_cas_fsm.read() == CAS_MULTI_INVAL_IVT_LOCK) or - (r_cas_fsm.read() == CAS_BC_IVT_LOCK)) + else if (r_cas_fsm.read() == CAS_INVAL_IVT_LOCK) r_alloc_ivt_fsm = ALLOC_IVT_CAS; @@ -9405,6 +9423,5 @@ r_alloc_ivt_fsm = ALLOC_IVT_CLEANUP; - else if ((r_cas_fsm.read() == CAS_MULTI_INVAL_IVT_LOCK) or - (r_cas_fsm.read() == CAS_BC_IVT_LOCK)) + else if (r_cas_fsm.read() == CAS_INVAL_IVT_LOCK) r_alloc_ivt_fsm = ALLOC_IVT_CAS; @@ -9562,8 +9579,6 @@ (r_cas_fsm.read() != CAS_DIR_HIT_COMPARE) and (r_cas_fsm.read() != CAS_DIR_HIT_WRITE) and - (r_cas_fsm.read() != CAS_BC_TRT_LOCK) and - (r_cas_fsm.read() != CAS_BC_IVT_LOCK) and (r_cas_fsm.read() != CAS_MISS_TRT_LOCK) and - (r_cas_fsm.read() != CAS_MULTI_INVAL_IVT_LOCK) and + (r_cas_fsm.read() != CAS_INVAL_IVT_LOCK) and (r_cas_fsm.read() != CAS_MULTI_INVAL_HEAP_LOCK)) or @@ -9633,5 +9648,5 @@ } else - m_cpt_xram_rsp_fsm_dir_used++; + m_cpt_multi_ack_fsm_dir_used++; break; /////////////////////// @@ -9720,6 +9735,5 @@ r_alloc_trt_fsm = ALLOC_TRT_WRITE; - else if((r_cas_fsm.read() == CAS_MISS_TRT_LOCK) or - (r_cas_fsm.read() == CAS_BC_TRT_LOCK)) + else if(r_cas_fsm.read() == CAS_MISS_TRT_LOCK) r_alloc_trt_fsm = ALLOC_TRT_CAS; @@ -9758,6 +9772,5 @@ (r_write_fsm.read() != WRITE_INVAL_IVT_LOCK)) { - if((r_cas_fsm.read() == CAS_MISS_TRT_LOCK) or - (r_cas_fsm.read() == CAS_BC_TRT_LOCK)) + if(r_cas_fsm.read() == CAS_MISS_TRT_LOCK) r_alloc_trt_fsm = ALLOC_TRT_CAS; @@ -9796,7 +9809,5 @@ //////////////////// case ALLOC_TRT_CAS: - if((r_cas_fsm.read() != CAS_MISS_TRT_LOCK) and - (r_cas_fsm.read() != CAS_BC_TRT_LOCK) and - (r_cas_fsm.read() != CAS_BC_IVT_LOCK)) + if(r_cas_fsm.read() != CAS_MISS_TRT_LOCK) { if((r_ixr_cmd_fsm.read() == IXR_CMD_READ_TRT) or @@ -9863,8 +9874,14 @@ r_alloc_trt_fsm = ALLOC_TRT_WRITE; - else if((r_cas_fsm.read() == CAS_MISS_TRT_LOCK) or - (r_cas_fsm.read() == CAS_BC_TRT_LOCK)) + else if(r_cas_fsm.read() == CAS_MISS_TRT_LOCK) r_alloc_trt_fsm = ALLOC_TRT_CAS; - } + + else + m_cpt_trt_unused++; + + } + else + m_cpt_ixr_cmd_fsm_trt_used++; + break; @@ -9893,6 +9910,5 @@ r_alloc_trt_fsm = ALLOC_TRT_WRITE; - else if((r_cas_fsm.read() == CAS_MISS_TRT_LOCK) or - (r_cas_fsm.read() == CAS_BC_TRT_LOCK)) + else if(r_cas_fsm.read() == CAS_MISS_TRT_LOCK) r_alloc_trt_fsm = ALLOC_TRT_CAS; @@ -9929,6 +9945,5 @@ r_alloc_trt_fsm = ALLOC_TRT_WRITE; - else if((r_cas_fsm.read() == CAS_MISS_TRT_LOCK) || - (r_cas_fsm.read() == CAS_BC_TRT_LOCK)) + else if(r_cas_fsm.read() == CAS_MISS_TRT_LOCK) r_alloc_trt_fsm = ALLOC_TRT_CAS; @@ -9949,5 +9964,5 @@ } else - m_cpt_ixr_fsm_trt_used++; + m_cpt_ixr_rsp_fsm_trt_used++; break; @@ -9966,6 +9981,5 @@ r_alloc_trt_fsm = ALLOC_TRT_WRITE; - else if((r_cas_fsm.read() == CAS_MISS_TRT_LOCK) or - (r_cas_fsm.read() == CAS_BC_TRT_LOCK)) + else if(r_cas_fsm.read() == CAS_MISS_TRT_LOCK) r_alloc_trt_fsm = ALLOC_TRT_CAS; @@ -9998,6 +10012,5 @@ r_alloc_trt_fsm = ALLOC_TRT_WRITE; - else if((r_cas_fsm.read() == CAS_MISS_TRT_LOCK) or - (r_cas_fsm.read() == CAS_BC_TRT_LOCK)) + else if(r_cas_fsm.read() == CAS_MISS_TRT_LOCK) r_alloc_trt_fsm = ALLOC_TRT_CAS; @@ -10020,5 +10033,10 @@ else if( r_config_fsm.read() == CONFIG_TRT_LOCK ) r_alloc_trt_fsm = ALLOC_TRT_CONFIG; - } + + else + m_cpt_trt_unused ++; + } + else + m_cpt_cleanup_fsm_trt_used++; break; @@ -10208,5 +10226,5 @@ } else - m_cpt_cleanup_fsm_heap_used++; + m_cpt_multi_ack_fsm_heap_used++; break;