Index: trunk/modules/vci_mem_cache_v4/caba/source/include/vci_mem_cache_v4.h =================================================================== --- trunk/modules/vci_mem_cache_v4/caba/source/include/vci_mem_cache_v4.h (revision 340) +++ trunk/modules/vci_mem_cache_v4/caba/source/include/vci_mem_cache_v4.h (revision 361) @@ -29,12 +29,4 @@ * alexandre.joannou@lip6.fr * - * Modifications done by Christophe Choichillon on the 7/04/2009: - * - Adding new states in the CLEANUP FSM : CLEANUP_UPT_LOCK and CLEANUP_UPT_WRITE - * - Adding a new VCI target port for the CLEANUP network - * - Adding new state in the ALLOC_UPT_FSM : ALLOC_UPT_CLEANUP - * - * Modifications to do : - * - Adding new variables used by the CLEANUP FSM - * */ @@ -106,11 +98,11 @@ INIT_CMD_INVAL_NLINE, INIT_CMD_XRAM_BRDCAST, + INIT_CMD_WRITE_BRDCAST, + INIT_CMD_CAS_BRDCAST, INIT_CMD_UPDT_IDLE, - INIT_CMD_WRITE_BRDCAST, INIT_CMD_UPDT_NLINE, INIT_CMD_UPDT_INDEX, INIT_CMD_UPDT_DATA, INIT_CMD_CAS_UPDT_IDLE, - INIT_CMD_CAS_BRDCAST, INIT_CMD_CAS_UPDT_NLINE, INIT_CMD_CAS_UPDT_INDEX, @@ -215,4 +207,5 @@ CAS_DIR_LOCK, CAS_DIR_HIT_READ, + CAS_DIR_HIT_COMPARE, CAS_DIR_HIT_WRITE, CAS_UPT_LOCK, @@ -372,4 +365,22 @@ soclib::caba::VciInitiator p_vci_ixr; +#if MONITOR_MEMCACHE_FSM + sc_out p_read_fsm; + sc_out p_write_fsm; + sc_out p_xram_rsp_fsm; + sc_out p_cas_fsm; + sc_out p_cleanup_fsm; + sc_out p_alloc_heap_fsm; + sc_out p_alloc_dir_fsm; + sc_out p_alloc_trt_fsm; + sc_out p_alloc_upt_fsm; + sc_out p_tgt_cmd_fsm; + sc_out p_tgt_rsp_fsm; + sc_out p_ixr_cmd_fsm; + sc_out p_ixr_rsp_fsm; + sc_out p_init_cmd_fsm; + sc_out p_init_rsp_fsm; +#endif + VciMemCacheV4( sc_module_name name, // Instance Name @@ -660,4 +671,5 @@ sc_signal r_cas_trt_index; // Transaction Table index sc_signal r_cas_upt_index; // Update Table index + sc_signal * r_cas_data; // cache line data // Buffer between CAS fsm and INIT_CMD fsm (XRAM read) @@ -770,4 +782,5 @@ sc_signal r_tgt_rsp_fsm; sc_signal r_tgt_rsp_cpt; + sc_signal r_tgt_rsp_key_sent; //////////////////////////////////////////////////// Index: trunk/modules/vci_mem_cache_v4/caba/source/include/xram_transaction_v4.h =================================================================== --- trunk/modules/vci_mem_cache_v4/caba/source/include/xram_transaction_v4.h (revision 340) +++ trunk/modules/vci_mem_cache_v4/caba/source/include/xram_transaction_v4.h (revision 361) @@ -78,4 +78,5 @@ wdata.assign(source.wdata.begin(),source.wdata.end()); rerror = source.rerror; + ll_key = source.ll_key; } Index: trunk/modules/vci_mem_cache_v4/caba/source/src/vci_mem_cache_v4.cpp =================================================================== --- trunk/modules/vci_mem_cache_v4/caba/source/src/vci_mem_cache_v4.cpp (revision 340) +++ trunk/modules/vci_mem_cache_v4/caba/source/src/vci_mem_cache_v4.cpp (revision 361) @@ -80,11 +80,11 @@ "INIT_CMD_INVAL_NLINE", "INIT_CMD_XRAM_BRDCAST", + "INIT_CMD_WRITE_BRDCAST", + "INIT_CMD_CAS_BRDCAST", "INIT_CMD_UPDT_IDLE", - "INIT_CMD_WRITE_BRDCAST", "INIT_CMD_UPDT_NLINE", "INIT_CMD_UPDT_INDEX", "INIT_CMD_UPDT_DATA", "INIT_CMD_CAS_UPDT_IDLE", - "INIT_CMD_CAS_BRDCAST", "INIT_CMD_CAS_UPDT_NLINE", "INIT_CMD_CAS_UPDT_INDEX", @@ -175,4 +175,5 @@ "CAS_DIR_LOCK", "CAS_DIR_HIT_READ", + "CAS_DIR_HIT_COMPARE", "CAS_DIR_HIT_WRITE", "CAS_UPT_LOCK", @@ -372,4 +373,23 @@ r_alloc_heap_fsm("r_alloc_heap_fsm"), r_alloc_heap_reset_cpt("r_alloc_heap_reset_cpt") + +#if MONITOR_MEMCACHE_FSM + , + p_read_fsm ("p_read_fsm"), + p_write_fsm ("p_write_fsm"), + p_xram_rsp_fsm ("p_xram_rsp_fsm"), + p_cas_fsm ("p_cas_fsm"), + p_cleanup_fsm ("p_cleanup_fsm"), + p_alloc_heap_fsm ("p_alloc_heap_fsm"), + p_alloc_dir_fsm ("p_alloc_dir_fsm"), + p_alloc_trt_fsm ("p_alloc_trt_fsm"), + p_alloc_upt_fsm ("p_alloc_upt_fsm"), + p_tgt_cmd_fsm ("p_tgt_cmd_fsm"), + p_tgt_rsp_fsm ("p_tgt_rsp_fsm"), + p_ixr_cmd_fsm ("p_ixr_cmd_fsm"), + p_ixr_rsp_fsm ("p_ixr_rsp_fsm"), + p_init_cmd_fsm ("p_init_cmd_fsm"), + p_init_rsp_fsm ("p_init_rsp_fsm") +#endif { assert(IS_POW_OF_2(nsets)); @@ -434,6 +454,7 @@ // Allocation for CAS FSM - r_cas_to_ixr_cmd_data = new sc_signal[nwords]; - r_cas_rdata = new sc_signal[2]; + r_cas_to_ixr_cmd_data = new sc_signal[nwords]; + r_cas_data = new sc_signal[nwords]; + r_cas_rdata = new sc_signal[2]; @@ -684,4 +705,6 @@ r_copies_limit = 3; + + r_tgt_rsp_key_sent = false; // Activity counters @@ -4359,6 +4382,6 @@ r_cas_count = entry.count; - if ( entry.valid ) r_cas_fsm = CAS_DIR_HIT_READ; - else r_cas_fsm = CAS_MISS_TRT_LOCK; + if ( entry.valid ) r_cas_fsm = CAS_DIR_HIT_READ; + else r_cas_fsm = CAS_MISS_TRT_LOCK; #if DEBUG_MEMC_CAS @@ -4387,9 +4410,8 @@ ///////////////////// case CAS_DIR_HIT_READ: // update directory for lock and dirty bit - // and check data change in cache + // and check data change in cache { size_t way = r_cas_way.read(); size_t set = m_y[(vci_addr_t)(m_cmd_cas_addr_fifo.read())]; - size_t word = m_x[(vci_addr_t)(m_cmd_cas_addr_fifo.read())]; // update directory (lock & dirty bits) @@ -4399,5 +4421,5 @@ entry.dirty = true; entry.lock = true; - entry.tag = r_cas_tag.read(); + entry.tag = r_cas_tag.read(); entry.owner.srcid = r_cas_copy.read(); #if L1_MULTI_CACHE @@ -4410,18 +4432,42 @@ m_cache_directory.write(set, way, entry); - // read data in cache & check data change - bool ok = ( r_cas_rdata[0].read() == m_cache_data.read(way, set, word) ); - if ( r_cas_cpt.read()==4 ) // 64 bits CAS - ok &= ( r_cas_rdata[1] == m_cache_data.read(way, set, word+1)); + // Stored data from cache in buffer to do the comparison in next state + m_cache_data.read_line(way, set, r_cas_data); + + r_cas_fsm = CAS_DIR_HIT_COMPARE; + +#if DEBUG_MEMC_CAS + if(m_debug_cas_fsm) + { + std::cout + << " Read data from " + << " cache and store it in buffer" + << std::endl; + } +#endif + break; + } + + case CAS_DIR_HIT_COMPARE: + { + size_t word = m_x[(vci_addr_t)(m_cmd_cas_addr_fifo.read())]; + + // Read data in buffer & check data change + bool ok = (r_cas_rdata[0].read() == r_cas_data[word].read()); + + if(r_cas_cpt.read() == 4) // 64 bits CAS + ok &= (r_cas_rdata[1] == r_cas_data[word+1]); // to avoid livelock, force the atomic access to fail pseudo-randomly - bool forced_fail = ( (r_cas_lfsr % (64) == 0) && RANDOMIZE_CAS ); - r_cas_lfsr = (r_cas_lfsr >> 1) ^ ((-(r_cas_lfsr & 1)) & 0xd0000001); - - if( ok and not forced_fail ) // no data change + bool forced_fail = ((r_cas_lfsr % (64) == 0) && RANDOMIZE_CAS); + r_cas_lfsr = (r_cas_lfsr >> 1) ^ ((- (r_cas_lfsr & 1)) & 0xd0000001); + + // cas success + if(ok and not forced_fail) { r_cas_fsm = CAS_DIR_HIT_WRITE; } - else // return failure + // cas failure + else { r_cas_fsm = CAS_RSP_FAIL; @@ -4429,11 +4475,13 @@ #if DEBUG_MEMC_CAS -if( m_debug_cas_fsm ) -{ - std::cout << " Test if CAS success:" - << " / expected value = " << r_cas_rdata[0].read() - << " / actual value = " << m_cache_data.read(way, set, word) - << " / forced_fail = " << forced_fail << std::endl; -} + if(m_debug_cas_fsm) + { + std::cout + << " Compare the old" + << " and the new data" + << " / expected value = " << r_cas_rdata[0].read() + << " / actual value = " << r_cas_data[word].read() + << " / forced_fail = " << forced_fail << std::endl; + } #endif break; @@ -5493,12 +5541,24 @@ } #endif - if ( r_tgt_rsp_cpt.read() == (r_read_to_tgt_rsp_word.read()+r_read_to_tgt_rsp_length-1) ) + + uint32_t last_word_idx = r_read_to_tgt_rsp_word.read() + r_read_to_tgt_rsp_length.read() - 1; + bool is_last_word = (r_tgt_rsp_cpt.read() == last_word_idx); + bool is_ll = ((r_read_to_tgt_rsp_pktid.read() & 0x7) == TYPE_LL); + + if ( (is_last_word and not is_ll) or (r_tgt_rsp_key_sent.read() and is_ll)) { - r_tgt_rsp_fsm = TGT_RSP_READ_IDLE; + // Last word in case of READ or second flit in case if LL + r_tgt_rsp_key_sent = false; r_read_to_tgt_rsp_req = false; + r_tgt_rsp_fsm = TGT_RSP_READ_IDLE; } else { - r_tgt_rsp_cpt = r_tgt_rsp_cpt.read() + 1; + if (is_ll) { + r_tgt_rsp_key_sent = true; // Send second flit of ll + } + else { + r_tgt_rsp_cpt = r_tgt_rsp_cpt.read() + 1; // Send next word of read + } } } @@ -5583,14 +5643,27 @@ } #endif - if ( (r_tgt_rsp_cpt.read() == - (r_xram_rsp_to_tgt_rsp_word.read()+r_xram_rsp_to_tgt_rsp_length.read()-1)) - || r_xram_rsp_to_tgt_rsp_rerror.read() ) + uint32_t last_word_idx = r_xram_rsp_to_tgt_rsp_word.read() + r_xram_rsp_to_tgt_rsp_length.read() - 1; + bool is_last_word = (r_tgt_rsp_cpt.read() == last_word_idx); + bool is_ll = ((r_xram_rsp_to_tgt_rsp_pktid.read() & 0x7) == TYPE_LL); + bool is_error = r_xram_rsp_to_tgt_rsp_rerror.read(); + + if (((is_last_word or is_error) and not is_ll) or + (r_tgt_rsp_key_sent.read() and is_ll)) { - r_tgt_rsp_fsm = TGT_RSP_XRAM_IDLE; + // Last word sent in case of READ or second flit sent in case if LL + r_tgt_rsp_key_sent = false; r_xram_rsp_to_tgt_rsp_req = false; + r_tgt_rsp_fsm = TGT_RSP_XRAM_IDLE; } else { - r_tgt_rsp_cpt = r_tgt_rsp_cpt.read() + 1; + if (is_ll) + { + r_tgt_rsp_key_sent = true; // Send second flit of ll + } + else + { + r_tgt_rsp_cpt = r_tgt_rsp_cpt.read() + 1; // Send next word of read + } } } @@ -5819,19 +5892,20 @@ //////////////////// case ALLOC_DIR_CAS: - if ((( r_cas_fsm.read() != CAS_DIR_REQ ) && - ( r_cas_fsm.read() != CAS_DIR_LOCK ) && - ( r_cas_fsm.read() != CAS_DIR_HIT_READ ) && - ( r_cas_fsm.read() != CAS_DIR_HIT_WRITE ) && - ( r_cas_fsm.read() != CAS_BC_TRT_LOCK ) && - ( r_cas_fsm.read() != CAS_BC_UPT_LOCK ) && - ( r_cas_fsm.read() != CAS_MISS_TRT_LOCK ) && - ( r_cas_fsm.read() != CAS_UPT_LOCK ) && - ( r_cas_fsm.read() != CAS_UPT_HEAP_LOCK )) + if ((( r_cas_fsm.read() != CAS_DIR_REQ ) && + ( r_cas_fsm.read() != CAS_DIR_LOCK ) && + ( r_cas_fsm.read() != CAS_DIR_HIT_READ ) && + ( r_cas_fsm.read() != CAS_DIR_HIT_COMPARE) && + ( r_cas_fsm.read() != CAS_DIR_HIT_WRITE ) && + ( r_cas_fsm.read() != CAS_BC_TRT_LOCK ) && + ( r_cas_fsm.read() != CAS_BC_UPT_LOCK ) && + ( r_cas_fsm.read() != CAS_MISS_TRT_LOCK ) && + ( r_cas_fsm.read() != CAS_UPT_LOCK ) && + ( r_cas_fsm.read() != CAS_UPT_HEAP_LOCK )) || - (( r_cas_fsm.read() == CAS_UPT_HEAP_LOCK ) && - ( r_alloc_heap_fsm.read() == ALLOC_HEAP_CAS )) + (( r_cas_fsm.read() == CAS_UPT_HEAP_LOCK ) && + ( r_alloc_heap_fsm.read() == ALLOC_HEAP_CAS )) || - (( r_cas_fsm.read() == CAS_MISS_TRT_LOCK ) && - ( r_alloc_trt_fsm.read() == ALLOC_TRT_CAS ))) + (( r_cas_fsm.read() == CAS_MISS_TRT_LOCK ) && + ( r_alloc_trt_fsm.read() == ALLOC_TRT_CAS ))) { if ( r_cleanup_fsm.read() == CLEANUP_DIR_REQ ) @@ -6328,4 +6402,21 @@ ///////////////////////////// { +#if MONITOR_MEMCACHE_FSM + p_read_fsm.write (r_read_fsm.read()); + p_write_fsm.write (r_write_fsm.read()); + p_xram_rsp_fsm.write (r_xram_rsp_fsm.read()); + p_cas_fsm.write (r_cas_fsm.read()); + p_cleanup_fsm.write (r_cleanup_fsm.read()); + p_alloc_heap_fsm.write(r_alloc_heap_fsm.read()); + p_alloc_dir_fsm.write (r_alloc_dir_fsm.read()); + p_alloc_trt_fsm.write (r_alloc_trt_fsm.read()); + p_alloc_upt_fsm.write (r_alloc_upt_fsm.read()); + p_tgt_cmd_fsm.write (r_tgt_cmd_fsm.read()); + p_tgt_rsp_fsm.write (r_tgt_rsp_fsm.read()); + p_ixr_cmd_fsm.write (r_ixr_cmd_fsm.read()); + p_ixr_rsp_fsm.write (r_ixr_rsp_fsm.read()); + p_init_cmd_fsm.write (r_init_cmd_fsm.read()); + p_init_rsp_fsm.write (r_init_rsp_fsm.read()); +#endif //////////////////////////////////////////////////////////// // Command signals on the p_vci_ixr port @@ -6455,19 +6546,32 @@ p_vci_tgt.reop = false; break; + case TGT_RSP_READ: - p_vci_tgt.rspval = true; - if( ((r_read_to_tgt_rsp_pktid.read() & 0x7) == TYPE_LL) - && (r_tgt_rsp_cpt.read() == (r_read_to_tgt_rsp_word.read()+r_read_to_tgt_rsp_length-1)) ) - p_vci_tgt.rdata = r_read_to_tgt_rsp_data[r_tgt_rsp_cpt.read()-1].read(); - else if ((r_read_to_tgt_rsp_pktid.read() & 0x7) == TYPE_LL) - p_vci_tgt.rdata = r_read_to_tgt_rsp_ll_key.read(); + { + uint32_t last_word_idx = r_read_to_tgt_rsp_word.read() + r_read_to_tgt_rsp_length - 1; + bool is_last_word = (r_tgt_rsp_cpt.read() == last_word_idx); + bool is_ll = ((r_read_to_tgt_rsp_pktid.read() & 0x7) == TYPE_LL); + + p_vci_tgt.rspval = true; + + if ( is_ll and not r_tgt_rsp_key_sent.read() ) + { + // LL response first flit + p_vci_tgt.rdata = r_read_to_tgt_rsp_ll_key.read(); + } else - p_vci_tgt.rdata = r_read_to_tgt_rsp_data[r_tgt_rsp_cpt.read()].read(); - p_vci_tgt.rsrcid = r_read_to_tgt_rsp_srcid.read(); - p_vci_tgt.rtrdid = r_read_to_tgt_rsp_trdid.read(); - p_vci_tgt.rpktid = r_read_to_tgt_rsp_pktid.read(); - p_vci_tgt.rerror = 0; - p_vci_tgt.reop = ( r_tgt_rsp_cpt.read() == (r_read_to_tgt_rsp_word.read()+r_read_to_tgt_rsp_length-1) ); + { + // LL response second flit or READ response + p_vci_tgt.rdata = r_read_to_tgt_rsp_data[r_tgt_rsp_cpt.read()].read(); + } + + p_vci_tgt.rsrcid = r_read_to_tgt_rsp_srcid.read(); + p_vci_tgt.rtrdid = r_read_to_tgt_rsp_trdid.read(); + p_vci_tgt.rpktid = r_read_to_tgt_rsp_pktid.read(); + p_vci_tgt.rerror = 0; + p_vci_tgt.reop = (is_last_word and not is_ll) or (r_tgt_rsp_key_sent.read() and is_ll); break; + } + case TGT_RSP_WRITE: /*if( ((r_write_to_tgt_rsp_pktid.read() & 0x7) == TYPE_SC) ) @@ -6487,4 +6591,5 @@ p_vci_tgt.reop = true; break; + case TGT_RSP_CLEANUP: p_vci_tgt.rspval = true; @@ -6496,4 +6601,5 @@ p_vci_tgt.reop = true; break; + case TGT_RSP_CAS: p_vci_tgt.rspval = true; @@ -6505,19 +6611,32 @@ p_vci_tgt.reop = true; break; + case TGT_RSP_XRAM: - p_vci_tgt.rspval = true; - if( ((r_xram_rsp_to_tgt_rsp_pktid.read() & 0x7) == TYPE_LL) - && (r_tgt_rsp_cpt.read() == (r_xram_rsp_to_tgt_rsp_word.read()+r_xram_rsp_to_tgt_rsp_length-1)) ) - p_vci_tgt.rdata = r_xram_rsp_to_tgt_rsp_ll_key.read(); - else - p_vci_tgt.rdata = r_xram_rsp_to_tgt_rsp_data[r_tgt_rsp_cpt.read()].read(); - p_vci_tgt.rsrcid = r_xram_rsp_to_tgt_rsp_srcid.read(); - p_vci_tgt.rtrdid = r_xram_rsp_to_tgt_rsp_trdid.read(); - p_vci_tgt.rpktid = r_xram_rsp_to_tgt_rsp_pktid.read(); - p_vci_tgt.rerror = r_xram_rsp_to_tgt_rsp_rerror.read(); - p_vci_tgt.reop = (( r_tgt_rsp_cpt.read() - == (r_xram_rsp_to_tgt_rsp_word.read()+r_xram_rsp_to_tgt_rsp_length.read()-1)) - || r_xram_rsp_to_tgt_rsp_rerror.read()); + { + uint32_t last_word_idx = r_xram_rsp_to_tgt_rsp_word.read() + r_xram_rsp_to_tgt_rsp_length.read() - 1; + bool is_last_word = (r_tgt_rsp_cpt.read() == last_word_idx); + bool is_ll = ((r_xram_rsp_to_tgt_rsp_pktid.read() & 0x7) == TYPE_LL); + bool is_error = r_xram_rsp_to_tgt_rsp_rerror.read(); + + p_vci_tgt.rspval = true; + + if( is_ll and not r_tgt_rsp_key_sent.read() ) { + // LL response first flit + p_vci_tgt.rdata = r_xram_rsp_to_tgt_rsp_ll_key.read(); + } + else { + // LL response second flit or READ response + p_vci_tgt.rdata = r_xram_rsp_to_tgt_rsp_data[r_tgt_rsp_cpt.read()].read(); + } + + p_vci_tgt.rsrcid = r_xram_rsp_to_tgt_rsp_srcid.read(); + p_vci_tgt.rtrdid = r_xram_rsp_to_tgt_rsp_trdid.read(); + p_vci_tgt.rpktid = r_xram_rsp_to_tgt_rsp_pktid.read(); + p_vci_tgt.rerror = is_error; + p_vci_tgt.reop = (((is_last_word or is_error) and not is_ll) or + (r_tgt_rsp_key_sent.read() and is_ll)); break; + } + case TGT_RSP_INIT: p_vci_tgt.rspval = true;