io.c.svn-base

模拟多核状态下龙芯处理器的功能

   printf("	   dyn. read energy (nJ): %g\n",result->data_output_power.readOp.dynamic*1e9);
   printf("	   leak. read power (mW): %g\n",result->data_output_power.readOp.leakage*1e3);
   printf(" total_out_driver (ns): %g\n", result->total_out_driver_delay_data/1e-9);
   printf("	 dyn. read energy (nJ): %g\n", result->total_out_driver_power_data.readOp.dynamic*1e9);
   printf("	 leak. read power (mW): %g\n", result->total_out_driver_power_data.readOp.leakage*1e3);

   printf(" total data path (without output driver) (ns): %g\n",result->subbank_address_routing_delay/1e-9+result->decoder_delay_data/1e-9+result->wordline_delay_data/1e-9+result->bitline_delay_data/1e-9+result->sense_amp_delay_data/1e-9);
   if (!parameters->fully_assoc)
     {
       if (A==1)
         printf(" total tag path is dm (ns): %g\n", result->subbank_address_routing_delay/1e-9+result->decoder_delay_tag/1e-9+result->wordline_delay_tag/1e-9+result->bitline_delay_tag/1e-9+result->sense_amp_delay_tag/1e-9+result->compare_part_delay/1e-9);
       else
         printf(" total tag path is set assoc (ns): %g\n", result->subbank_address_routing_delay/1e-9+result->decoder_delay_tag/1e-9+result->wordline_delay_tag/1e-9+result->bitline_delay_tag/1e-9+result->sense_amp_delay_tag/1e-9+result->compare_part_delay/1e-9+result->drive_mux_delay/1e-9+result->selb_delay/1e-9);
     }
}

void output_area_components(arearesult_type *arearesult, parameter_type *parameters)
{
    printf("\nArea Components:\n\n");

	//v4.1: No longer using calculate_area function as area has already been
	//computed for the given tech node. 
/*
    printf("Aspect Ratio Data height/width: %f\n", aspect_ratio_data);
    printf("Aspect Ratio Tag height/width: %f\n", aspect_ratio_tag);
    printf("Aspect Ratio Subbank height/width: %f\n", aspect_ratio_subbank);
    printf("Aspect Ratio Total height/width: %f\n\n", aspect_ratio_total);
*/
    printf("Aspect Ratio Total height/width: %f\n\n", arearesult->aspect_ratio_total);

	printf("Data array (mm^2): %f\n",arearesult->dataarray_area.scaled_area);
	printf("Data predecode (mm^2): %f\n",arearesult->datapredecode_area.scaled_area);
	printf("Data colmux predecode (mm^2): %f\n",arearesult->datacolmuxpredecode_area.scaled_area);
	printf("Data colmux post decode (mm^2): %f\n",arearesult->datacolmuxpostdecode_area.scaled_area);
	printf("Data write signal (mm^2): %f\n",arearesult->datawritesig_area.scaled_area);

	printf("\nTag array (mm^2): %f\n",arearesult->tagarray_area.scaled_area);
	printf("Tag predecode (mm^2): %f\n",arearesult->tagpredecode_area.scaled_area);
	printf("Tag colmux predecode (mm^2): %f\n",arearesult->tagcolmuxpredecode_area.scaled_area);
	printf("Tag colmux post decode (mm^2): %f\n",arearesult->tagcolmuxpostdecode_area.scaled_area);
	printf("Tag output driver decode (mm^2): %f\n",arearesult->tagoutdrvdecode_area.scaled_area);
	printf("Tag output driver enable signals (mm^2): %f\n",arearesult->tagoutdrvsig_area.scaled_area);

	printf("\nPercentage of data ramcells alone of total area: %f %%\n", arearesult->perc_data);
    printf("Percentage of tag ramcells alone of total area: %f %%\n",arearesult->perc_tag);
    printf("Percentage of total control/routing alone of total area: %f %%\n",arearesult->perc_cont);
	printf("\nSubbank Efficiency : %f\n", arearesult->sub_eff);
	printf("Total Efficiency : %f\n",arearesult->total_eff);
	printf("\nTotal area One bank (mm^2): %f\n",arearesult->totalarea);
	printf("Total area subbanked (mm^2): %f\n",arearesult->subbankarea);

}


void output_data(result_type *result,arearesult_type *arearesult, parameter_type *parameters)
{
   double datapath,tagpath;

   FILE *stream;

   stream=fopen("cache_params.aux", "w");

   datapath = result->subbank_address_routing_delay+result->decoder_delay_data+result->wordline_delay_data+result->bitline_delay_data+result->sense_amp_delay_data+result->total_out_driver_delay_data+result->data_output_delay;
   if (parameters->tag_associativity == 1) {
         tagpath = result->subbank_address_routing_delay+result->decoder_delay_tag+result->wordline_delay_tag+result->bitline_delay_tag+result->sense_amp_delay_tag+result->compare_part_delay+result->drive_valid_delay;
   } else {
         tagpath = result->subbank_address_routing_delay+result->decoder_delay_tag+result->wordline_delay_tag+result->bitline_delay_tag+result->sense_amp_delay_tag+result->compare_part_delay+
			 //result->drive_mux_delay+
			 result->selb_delay+result->data_output_delay+result->total_out_driver_delay_data;
   }

   if (stream){
	   fprintf(stream, "#define PARAMNDWL %d\n#define PARAMNDBL %d\n#define PARAMNSPD %f\n#define PARAMNTWL %d\n#define PARAMNTBL %d\n#define PARAMNTSPD %d\n#define PARAMSENSESCALE %f\n#define PARAMGS %d\n#define PARAMDNOR %d\n#define PARAMTNOR %d\n#define PARAMRPORTS %d\n#define PARAMWPORTS %d\n#define PARAMRWPORTS %d\n#define PARAMMUXOVER %d\n", result->best_Ndwl, result->best_Ndbl, result->best_Nspd, result->best_Ntwl, result->best_Ntbl, result->best_Ntspd, result->senseext_scale, (result->senseext_scale==1.0), result->data_nor_inputs, result->tag_nor_inputs, parameters->num_read_ports, parameters->num_write_ports, parameters->num_readwrite_ports, result->best_muxover);
   }
 
    fclose(stream);
#  if OUTPUTTYPE == LONG
      printf("\n---------- CACTI version 4.1 ----------\n");
      printf("\nCache Parameters:\n");
	  printf("  Number of banks: %d\n",(int)result->subbanks);
      printf("  Total Cache Size: %d\n",(int) (parameters->cache_size));
      printf("  Size in bytes of a bank: %d\n",parameters->cache_size / (int)result->subbanks);
      printf("  Number of sets per bank: %d\n",parameters->number_of_sets);
      if (parameters->fully_assoc)
        printf("  Associativity: fully associative\n");
      else
        {
          if (parameters->tag_associativity==1)
            printf("  Associativity: direct mapped\n");
          else
            printf("  Associativity: %d\n",parameters->tag_associativity);
        }
      printf("  Block Size (bytes): %d\n",parameters->block_size);
      printf("  Read/Write Ports: %d\n",parameters->num_readwrite_ports);
      printf("  Read Ports: %d\n",parameters->num_read_ports);
      printf("  Write Ports: %d\n",parameters->num_write_ports);
      printf("  Technology Size: %2.2fum\n", parameters->tech_size);
      printf("  Vdd: %2.1fV\n", parameters->VddPow);

      printf("\nAccess Time (ns): %g\n",result->access_time*1e9);
      printf("Cycle Time (ns):  %g\n",result->cycle_time*1e9);
      //if (parameters->fully_assoc)
        //{
			printf("Total dynamic Read Power at max. freq. (W): %g\n",result->total_power_allbanks.readOp.dynamic/result->cycle_time);
			printf("Total leakage Read/Write Power all Banks (mW): %g\n",result->total_power_allbanks.readOp.leakage*1e3);
			printf("Total dynamic Read Energy all Banks (nJ): %g\n",result->total_power_allbanks.readOp.dynamic*1e9);
			printf("Total dynamic Write Energy all Banks (nJ): %g\n",result->total_power_allbanks.writeOp.dynamic*1e9);
			printf("Total dynamic Read Energy Without Routing (nJ): %g\n",result->total_power_without_routing.readOp.dynamic*1e9);
			printf("Total dynamic Write Energy Without Routing (nJ): %g\n",result->total_power_without_routing.writeOp.dynamic*1e9);
			printf("Total dynamic Routing Energy (nJ): %g\n",result->total_routing_power.readOp.dynamic*1e9);
			printf("Total leakage Read/Write Power Without Routing (mW): %g\n",result->total_power_without_routing.readOp.leakage*1e3);
			//printf("Total leakage Write Power all Banks (mW): %g\n",result->total_power_allbanks.writeOp.leakage*1e3);
			//printf("Total leakage Write Power Without Routing (mW): %g\n",result->total_power_without_routing.writeOp.leakage*1e3);
			printf("Total leakage Read/Write Routing Power (mW): %g\n",result->total_routing_power.readOp.leakage*1e3);

          //printf("Maximum Bank Power (nJ):  %g\n",(result->subbank_address_routing_power+result->decoder_power_data+result->wordline_power_data+result->bitline_power_data+result->sense_amp_power_data+result->data_output_power+result->total_out_driver_power_data)*1e9);
          //      printf("Power (W) - 500MHz:  %g\n",(result->decoder_power_data+result->wordline_power_data+result->bitline_power_data+result->sense_amp_power_data+result->data_output_power)*500*1e6);
        //}
      /*else
        {
	  printf("Total dynamic Read Power at max. freq. (W): %g\n",result->total_power_allbanks.readOp.dynamic/result->cycle_time);
	  printf("Total dynamic Read Energy all Banks (nJ): %g\n",result->total_power_allbanks.readOp.dynamic*1e9);
	  printf("Total leakage Read Power all Banks (mW): %g\n",result->total_power_allbanks.readOp.leakage*1e3);
	  printf("Total dynamic Write Energy all Banks (nJ): %g\n",result->total_power_allbanks.writeOp.dynamic*1e9);
	  printf("Total leakage Write Power all Banks (mW): %g\n",result->total_power_allbanks.writeOp.leakage*1e3);
	  printf("Total dynamic Read Energy Without Routing (nJ): %g\n",result->total_power_without_routing.readOp.dynamic*1e9);
	  printf("Total leakage Read Power Without Routing (mW): %g\n",result->total_power_without_routing.readOp.leakage*1e3);
	  printf("Total dynamic Write Energy Without Routing (nJ): %g\n",result->total_power_without_routing.writeOp.dynamic*1e9);
	  printf("Total leakage Write Power Without Routing (mW): %g\n",result->total_power_without_routing.writeOp.leakage*1e3);
	  printf("Total dynamic Routing Energy (nJ): %g\n",result->total_routing_power.readOp.dynamic*1e9);
	  printf("Total leakage Routing Power (mW): %g\n",result->total_routing_power.readOp.leakage*1e3);*/
          //printf("Maximum Bank Power (nJ):  %g\n",(result->subbank_address_routing_power+result->decoder_power_data+result->wordline_power_data+result->bitline_power_data+result->sense_amp_power_data+result->total_out_driver_power_data+result->decoder_power_tag+result->wordline_power_tag+result->bitline_power_tag+result->sense_amp_power_tag+result->compare_part_power+result->drive_valid_power+result->drive_mux_power+result->selb_power+result->data_output_power)*1e9);
          //      printf("Power (W) - 500MHz:  %g\n",(result->decoder_power_data+result->wordline_power_data+result->bitline_power_data+result->sense_amp_power_data+result->total_out_driver_power_data+result->decoder_power_tag+result->wordline_power_tag+result->bitline_power_tag+result->sense_amp_power_tag+result->compare_part_power+result->drive_valid_power+result->drive_mux_power+result->selb_power+result->data_output_power)*500*1e6);
        //}
      printf("\nBest Ndwl (L1): %d\n",result->best_Ndwl);
      printf("Best Ndbl (L1): %d\n",result->best_Ndbl);
      printf("Best Nspd (L1): %f\n",result->best_Nspd);
      printf("Best Ntwl (L1): %d\n",result->best_Ntwl);
      printf("Best Ntbl (L1): %d\n",result->best_Ntbl);
      printf("Best Ntspd (L1): %d\n",result->best_Ntspd);
      //printf("Nor inputs (data): %d\n",result->data_nor_inputs);
      //printf("Nor inputs (tag): %d\n",result->tag_nor_inputs);

	  output_area_components(arearesult,parameters);
      printf("\nTime Components:\n");
     
      printf(" data side (with Output driver) (ns): %g\n",datapath/1e-9);
      if (!parameters->fully_assoc)
        printf(" tag side (with Output driver) (ns): %g\n",(tagpath)/1e-9);
      output_time_components(result,parameters);

#  else
      printf("%d %d %d  %d %d %d %d %d %d  %e %e %e %e  %e %e %e %e  %e %e %e %e  %e %e %e %e  %e %e\n",
               parameters->cache_size,
               parameters->block_size,
               parameters->associativity,
               result->best_Ndwl,
               result->best_Ndbl,
               result->best_Nspd,
               result->best_Ntwl,
               result->best_Ntbl,
               result->best_Ntspd,
               result->access_time,
               result->cycle_time,
               datapath,
               tagpath,
               result->decoder_delay_data,
               result->wordline_delay_data,
               result->bitline_delay_data,
               result->sense_amp_delay_data,
               result->decoder_delay_tag,
               result->wordline_delay_tag,
               result->bitline_delay_tag,
               result->sense_amp_delay_tag,
               result->compare_part_delay,
               result->drive_mux_delay,
               result->selb_delay,
               result->drive_valid_delay,
               result->data_output_delay,
               result->precharge_delay);



# endif

}


total_result_type cacti_interface(
		int cache_size,
		int line_size,
		int associativity,
		int rw_ports,
		int excl_read_ports,
		int excl_write_ports,
		int single_ended_read_ports,
		int banks,
		double tech_node,
		int output_width,
		int specific_tag,
		int tag_width,
		int access_mode,
		int pure_sram,
        double Nspd_predef,
        int Ndwl_predef,
        int Ndbl_predef,
        int optimize)
{
   int C,B,A,ERP,EWP,RWP,NSER;
   double tech;
   double logbanks, assoc;
   double logbanksfloor, assocfloor;
   int seq_access = 0;
   int fast_access = 0;
   int bits_output = output_width;
   int nr_args = 9;
   double NSubbanks = (double)banks;

   double ratioofbankstoports;

   extern int force_tag, force_tag_size;

   

   total_result_type endresult;

   result_type result;
   arearesult_type arearesult;
   area_type arearesult_subbanked;
   parameter_type parameters; 

   /* input parameters:
         C B A ERP EWP */

   /*dt: make sure we're using some simple leakage reduction */
   dualVt = FALSE;

    force_tag = 0;

   if(specific_tag) {
	   force_tag = 1;
	   force_tag_size = tag_width;
   }
   switch (access_mode){
	   case 0:
		   seq_access = fast_access = FALSE;
		   break;
	   case 1:
		   seq_access = TRUE;
		   fast_access = FALSE;
		   break;
	   case 2:
		   seq_access = FALSE;
		   fast_access = TRUE;
		   break;
   }

   B = line_size;
   if ((B < 1)) {
       printf("Block size must >=1\n");
       return endresult;
	   //exit(1);
   }

   if ((B*8 < bits_output)) {
       printf("Block size must be at least %d\n", bits_output/8);
       return endresult;
	   //exit(1);
   }
   

   tech = tech_node;
   if ((tech <= 0)) {
       printf("Feature size must be > 0\n");
       return endresult;
	   //exit(1);
   }
   if ((tech > 0.8)) {
       printf("Feature size must be <= 0.80 (um)\n");
       return endresult;
	   //exit(1);
   }

   if (nr_args ==9)
     {
       RWP = rw_ports;
       ERP = excl_read_ports;
       EWP = excl_write_ports;
       NSER = single_ended_read_ports;


       if ((RWP < 0) || (EWP < 0) || (ERP < 0)) {
			printf("Ports must >=0\n");
			return endresult;
			//exit(1);
       }
       if (RWP > 2) {
			printf("Maximum of 2 read/write ports\n");
			return endresult;
			//exit(1);
       }