1/*****************************************************************************
2 * McPAT
3 * SOFTWARE LICENSE AGREEMENT
4 * Copyright 2012 Hewlett-Packard Development Company, L.P.
5 * All Rights Reserved
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions are
9 * met: redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer;
11 * redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
--- 7 unchanged lines hidden (view full) ---
20 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.”
29 *
30 ***************************************************************************/
31
32#define GLOBALVAR
33#include <cassert>
34#include <cmath>
35#include <iostream>
36
37#include "area.h"
38#include "array.h"
39#include "decoder.h"
40#include "globalvar.h"
41#include "parameter.h"
42
43using namespace std;
44
45ArrayST::ArrayST(const InputParameter *configure_interface,
46 string _name,
47 enum Device_ty device_ty_,
48 bool opt_local_,
49 enum Core_type core_ty_,
50 bool _is_default)
51:l_ip(*configure_interface),
52 name(_name),
53 device_ty(device_ty_),
54 opt_local(opt_local_),
55 core_ty(core_ty_),
56 is_default(_is_default)
57 {
58
59 if (l_ip.cache_sz<64) l_ip.cache_sz=64;
60 l_ip.error_checking();//not only do the error checking but also fill some missing parameters
61 optimize_array();
62
63}
64
65
66void ArrayST::compute_base_power()
67 {
68 //l_ip.out_w =l_ip.line_sz*8;
69 local_result=cacti_interface(&l_ip);
70
71 }
72
73void ArrayST::optimize_array()
74{
75 list<uca_org_t > candidate_solutions(0);
76 list<uca_org_t >::iterator candidate_iter, min_dynamic_energy_iter;
77
78 uca_org_t * temp_res = 0;
79 local_result.valid=false;
80
81 double throughput=l_ip.throughput, latency=l_ip.latency;
82 double area_efficiency_threshold = 20.0;
83 bool throughput_overflow=true, latency_overflow=true;
84 compute_base_power();
85
86 if ((local_result.cycle_time - throughput) <= 1e-10 )
87 throughput_overflow=false;
88 if ((local_result.access_time - latency)<= 1e-10)
89 latency_overflow=false;
90
91 if (opt_for_clk && opt_local)
92 {
93 if (throughput_overflow || latency_overflow)
94 {
95 l_ip.ed=0;
96
97 l_ip.delay_wt = 100;//Fixed number, make sure timing can be satisfied.
98 l_ip.cycle_time_wt = 1000;
99
100 l_ip.area_wt = 10;//Fixed number, This is used to exhaustive search for individual components.
101 l_ip.dynamic_power_wt = 10;//Fixed number, This is used to exhaustive search for individual components.
102 l_ip.leakage_power_wt = 10;
103
104 l_ip.delay_dev = 1000000;//Fixed number, make sure timing can be satisfied.
105 l_ip.cycle_time_dev = 100;
106
107 l_ip.area_dev = 1000000;//Fixed number, This is used to exhaustive search for individual components.
108 l_ip.dynamic_power_dev = 1000000;//Fixed number, This is used to exhaustive search for individual components.
109 l_ip.leakage_power_dev = 1000000;
110
111 throughput_overflow=true; //Reset overflow flag before start optimization iterations
112 latency_overflow=true;
113
114 temp_res = &local_result; //Clean up the result for optimized for ED^2P
115 temp_res->cleanup();
116 }
117
118
119 while ((throughput_overflow || latency_overflow)&&l_ip.cycle_time_dev > 10)// && l_ip.delay_dev > 10
120 {
121 compute_base_power();
122
123 l_ip.cycle_time_dev-=10;//This is the time_dev to be used for next iteration
124
125 // from best area to worst area -->worst timing to best timing
126 if ((((local_result.cycle_time - throughput) <= 1e-10 ) && (local_result.access_time - latency)<= 1e-10)||
127 (local_result.data_array2->area_efficiency < area_efficiency_threshold && l_ip.assoc == 0))
128 { //if no satisfiable solution is found,the most aggressive one is left
129 candidate_solutions.push_back(local_result);
130 //output_data_csv(candidate_solutions.back());
131 if (((local_result.cycle_time - throughput) <= 1e-10) && ((local_result.access_time - latency)<= 1e-10))
132 //ensure stop opt not because of cam
133 {
134 throughput_overflow=false;
135 latency_overflow=false;
136 }
137
138 }
139 else
140 {
141 //TODO: whether checking the partial satisfied results too, or just change the mark???
142 if ((local_result.cycle_time - throughput) <= 1e-10)
143 throughput_overflow=false;
144 if ((local_result.access_time - latency)<= 1e-10)
145 latency_overflow=false;
146
147 if (l_ip.cycle_time_dev > 10)
148 { //if not >10 local_result is the last result, it cannot be cleaned up
149 temp_res = &local_result; //Only solutions not saved in the list need to be cleaned up
150 temp_res->cleanup();
151 }
152 }
153// l_ip.cycle_time_dev-=10;
154// l_ip.delay_dev-=10;
155
156 }
157
158
159 if (l_ip.assoc > 0)
160 {
161 //For array structures except CAM and FA, Give warning but still provide a result with best timing found
162 if (throughput_overflow==true)
163 cout<< "Warning: " << name<<" array structure cannot satisfy throughput constraint." << endl;
164 if (latency_overflow==true)
165 cout<< "Warning: " << name<<" array structure cannot satisfy latency constraint." << endl;
166 }
167
168// else
169// {
170// /*According to "Content-Addressable Memory (CAM) Circuits and
171// Architectures": A Tutorial and Survey
172// by Kostas Pagiamtzis et al.
173// CAM structures can be heavily pipelined and use look-ahead techniques,
174// therefore timing can be relaxed. But McPAT does not model the advanced
175// techniques. If continue optimizing, the area efficiency will be too low
176// */
177// //For CAM and FA, stop opt if area efficiency is too low
178// if (throughput_overflow==true)
179// cout<< "Warning: " <<" McPAT stopped optimization on throughput for "<< name
180// <<" array structure because its area efficiency is below "<<area_efficiency_threshold<<"% " << endl;
181// if (latency_overflow==true)
182// cout<< "Warning: " <<" McPAT stopped optimization on latency for "<< name
183// <<" array structure because its area efficiency is below "<<area_efficiency_threshold<<"% " << endl;
184// }
185
186 //double min_dynamic_energy, min_dynamic_power, min_leakage_power, min_cycle_time;
187 double min_dynamic_energy=BIGNUM;
188 if (candidate_solutions.empty()==false)
189 {
190 local_result.valid=true;
191 for (candidate_iter = candidate_solutions.begin(); candidate_iter != candidate_solutions.end(); ++candidate_iter)
192
193 {
194 if (min_dynamic_energy > (candidate_iter)->power.readOp.dynamic)
195 {
196 min_dynamic_energy = (candidate_iter)->power.readOp.dynamic;
197 min_dynamic_energy_iter = candidate_iter;
198 local_result = *(min_dynamic_energy_iter);
199 //TODO: since results are reordered results and l_ip may miss match. Therefore, the final output spread sheets may show the miss match.
200
201 }
202 else
203 {
204 candidate_iter->cleanup() ;
205 }
206
207 }
208
209
210 }
211 candidate_solutions.clear();
212 }
213
214 double long_channel_device_reduction = longer_channel_device_reduction(device_ty,core_ty);
215
216 double macro_layout_overhead = g_tp.macro_layout_overhead;
217 double chip_PR_overhead = g_tp.chip_layout_overhead;
218 double total_overhead = macro_layout_overhead*chip_PR_overhead;
219 local_result.area *= total_overhead;
220
221 //maintain constant power density
222 double pppm_t[4] = {total_overhead,1,1,total_overhead};
223
224 double sckRation = g_tp.sckt_co_eff;
225 local_result.power.readOp.dynamic *= sckRation;
226 local_result.power.writeOp.dynamic *= sckRation;
227 local_result.power.searchOp.dynamic *= sckRation;
228 local_result.power.readOp.leakage *= l_ip.nbanks;
229 local_result.power.readOp.longer_channel_leakage =
230 local_result.power.readOp.leakage*long_channel_device_reduction;
231 local_result.power = local_result.power* pppm_t;
232
233 local_result.data_array2->power.readOp.dynamic *= sckRation;
234 local_result.data_array2->power.writeOp.dynamic *= sckRation;
235 local_result.data_array2->power.searchOp.dynamic *= sckRation;
236 local_result.data_array2->power.readOp.leakage *= l_ip.nbanks;
237 local_result.data_array2->power.readOp.longer_channel_leakage =
238 local_result.data_array2->power.readOp.leakage*long_channel_device_reduction;
239 local_result.data_array2->power = local_result.data_array2->power* pppm_t;
240
241
242 if (!(l_ip.pure_cam || l_ip.pure_ram || l_ip.fully_assoc) && l_ip.is_cache)
243 {
244 local_result.tag_array2->power.readOp.dynamic *= sckRation;
245 local_result.tag_array2->power.writeOp.dynamic *= sckRation;
246 local_result.tag_array2->power.searchOp.dynamic *= sckRation;
247 local_result.tag_array2->power.readOp.leakage *= l_ip.nbanks;
248 local_result.tag_array2->power.readOp.longer_channel_leakage =
249 local_result.tag_array2->power.readOp.leakage*long_channel_device_reduction;
250 local_result.tag_array2->power = local_result.tag_array2->power* pppm_t;
251 }
252
253
254}
255
256void ArrayST::leakage_feedback(double temperature)
257{
258 // Update the temperature. l_ip is already set and error-checked in the creator function.
259 l_ip.temp = (unsigned int)round(temperature/10.0)*10;
260
261 // This corresponds to cacti_interface() in the initialization process. Leakage power is updated here.
--- 29 unchanged lines hidden (view full) ---
291 local_result.tag_array2->power.writeOp.dynamic *= sckRation;
292 local_result.tag_array2->power.searchOp.dynamic *= sckRation;
293 local_result.tag_array2->power.readOp.leakage *= l_ip.nbanks;
294 local_result.tag_array2->power.readOp.longer_channel_leakage = local_result.tag_array2->power.readOp.leakage*long_channel_device_reduction;
295 local_result.tag_array2->power = local_result.tag_array2->power* pppm_t;
296 }
297}
298
299ArrayST:: ~ArrayST()
300{
301 local_result.cleanup();
302}
2 * McPAT
3 * SOFTWARE LICENSE AGREEMENT
4 * Copyright 2012 Hewlett-Packard Development Company, L.P.
5 * All Rights Reserved
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions are
9 * met: redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer;
11 * redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
--- 7 unchanged lines hidden (view full) ---
20 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.”
29 *
30 ***************************************************************************/
31
32#define GLOBALVAR
33#include <cassert>
34#include <cmath>
35#include <iostream>
36
37#include "area.h"
38#include "array.h"
39#include "decoder.h"
40#include "globalvar.h"
41#include "parameter.h"
42
43using namespace std;
44
45ArrayST::ArrayST(const InputParameter *configure_interface,
46 string _name,
47 enum Device_ty device_ty_,
48 bool opt_local_,
49 enum Core_type core_ty_,
50 bool _is_default)
51:l_ip(*configure_interface),
52 name(_name),
53 device_ty(device_ty_),
54 opt_local(opt_local_),
55 core_ty(core_ty_),
56 is_default(_is_default)
57 {
58
59 if (l_ip.cache_sz<64) l_ip.cache_sz=64;
60 l_ip.error_checking();//not only do the error checking but also fill some missing parameters
61 optimize_array();
62
63}
64
65
66void ArrayST::compute_base_power()
67 {
68 //l_ip.out_w =l_ip.line_sz*8;
69 local_result=cacti_interface(&l_ip);
70
71 }
72
73void ArrayST::optimize_array()
74{
75 list<uca_org_t > candidate_solutions(0);
76 list<uca_org_t >::iterator candidate_iter, min_dynamic_energy_iter;
77
78 uca_org_t * temp_res = 0;
79 local_result.valid=false;
80
81 double throughput=l_ip.throughput, latency=l_ip.latency;
82 double area_efficiency_threshold = 20.0;
83 bool throughput_overflow=true, latency_overflow=true;
84 compute_base_power();
85
86 if ((local_result.cycle_time - throughput) <= 1e-10 )
87 throughput_overflow=false;
88 if ((local_result.access_time - latency)<= 1e-10)
89 latency_overflow=false;
90
91 if (opt_for_clk && opt_local)
92 {
93 if (throughput_overflow || latency_overflow)
94 {
95 l_ip.ed=0;
96
97 l_ip.delay_wt = 100;//Fixed number, make sure timing can be satisfied.
98 l_ip.cycle_time_wt = 1000;
99
100 l_ip.area_wt = 10;//Fixed number, This is used to exhaustive search for individual components.
101 l_ip.dynamic_power_wt = 10;//Fixed number, This is used to exhaustive search for individual components.
102 l_ip.leakage_power_wt = 10;
103
104 l_ip.delay_dev = 1000000;//Fixed number, make sure timing can be satisfied.
105 l_ip.cycle_time_dev = 100;
106
107 l_ip.area_dev = 1000000;//Fixed number, This is used to exhaustive search for individual components.
108 l_ip.dynamic_power_dev = 1000000;//Fixed number, This is used to exhaustive search for individual components.
109 l_ip.leakage_power_dev = 1000000;
110
111 throughput_overflow=true; //Reset overflow flag before start optimization iterations
112 latency_overflow=true;
113
114 temp_res = &local_result; //Clean up the result for optimized for ED^2P
115 temp_res->cleanup();
116 }
117
118
119 while ((throughput_overflow || latency_overflow)&&l_ip.cycle_time_dev > 10)// && l_ip.delay_dev > 10
120 {
121 compute_base_power();
122
123 l_ip.cycle_time_dev-=10;//This is the time_dev to be used for next iteration
124
125 // from best area to worst area -->worst timing to best timing
126 if ((((local_result.cycle_time - throughput) <= 1e-10 ) && (local_result.access_time - latency)<= 1e-10)||
127 (local_result.data_array2->area_efficiency < area_efficiency_threshold && l_ip.assoc == 0))
128 { //if no satisfiable solution is found,the most aggressive one is left
129 candidate_solutions.push_back(local_result);
130 //output_data_csv(candidate_solutions.back());
131 if (((local_result.cycle_time - throughput) <= 1e-10) && ((local_result.access_time - latency)<= 1e-10))
132 //ensure stop opt not because of cam
133 {
134 throughput_overflow=false;
135 latency_overflow=false;
136 }
137
138 }
139 else
140 {
141 //TODO: whether checking the partial satisfied results too, or just change the mark???
142 if ((local_result.cycle_time - throughput) <= 1e-10)
143 throughput_overflow=false;
144 if ((local_result.access_time - latency)<= 1e-10)
145 latency_overflow=false;
146
147 if (l_ip.cycle_time_dev > 10)
148 { //if not >10 local_result is the last result, it cannot be cleaned up
149 temp_res = &local_result; //Only solutions not saved in the list need to be cleaned up
150 temp_res->cleanup();
151 }
152 }
153// l_ip.cycle_time_dev-=10;
154// l_ip.delay_dev-=10;
155
156 }
157
158
159 if (l_ip.assoc > 0)
160 {
161 //For array structures except CAM and FA, Give warning but still provide a result with best timing found
162 if (throughput_overflow==true)
163 cout<< "Warning: " << name<<" array structure cannot satisfy throughput constraint." << endl;
164 if (latency_overflow==true)
165 cout<< "Warning: " << name<<" array structure cannot satisfy latency constraint." << endl;
166 }
167
168// else
169// {
170// /*According to "Content-Addressable Memory (CAM) Circuits and
171// Architectures": A Tutorial and Survey
172// by Kostas Pagiamtzis et al.
173// CAM structures can be heavily pipelined and use look-ahead techniques,
174// therefore timing can be relaxed. But McPAT does not model the advanced
175// techniques. If continue optimizing, the area efficiency will be too low
176// */
177// //For CAM and FA, stop opt if area efficiency is too low
178// if (throughput_overflow==true)
179// cout<< "Warning: " <<" McPAT stopped optimization on throughput for "<< name
180// <<" array structure because its area efficiency is below "<<area_efficiency_threshold<<"% " << endl;
181// if (latency_overflow==true)
182// cout<< "Warning: " <<" McPAT stopped optimization on latency for "<< name
183// <<" array structure because its area efficiency is below "<<area_efficiency_threshold<<"% " << endl;
184// }
185
186 //double min_dynamic_energy, min_dynamic_power, min_leakage_power, min_cycle_time;
187 double min_dynamic_energy=BIGNUM;
188 if (candidate_solutions.empty()==false)
189 {
190 local_result.valid=true;
191 for (candidate_iter = candidate_solutions.begin(); candidate_iter != candidate_solutions.end(); ++candidate_iter)
192
193 {
194 if (min_dynamic_energy > (candidate_iter)->power.readOp.dynamic)
195 {
196 min_dynamic_energy = (candidate_iter)->power.readOp.dynamic;
197 min_dynamic_energy_iter = candidate_iter;
198 local_result = *(min_dynamic_energy_iter);
199 //TODO: since results are reordered results and l_ip may miss match. Therefore, the final output spread sheets may show the miss match.
200
201 }
202 else
203 {
204 candidate_iter->cleanup() ;
205 }
206
207 }
208
209
210 }
211 candidate_solutions.clear();
212 }
213
214 double long_channel_device_reduction = longer_channel_device_reduction(device_ty,core_ty);
215
216 double macro_layout_overhead = g_tp.macro_layout_overhead;
217 double chip_PR_overhead = g_tp.chip_layout_overhead;
218 double total_overhead = macro_layout_overhead*chip_PR_overhead;
219 local_result.area *= total_overhead;
220
221 //maintain constant power density
222 double pppm_t[4] = {total_overhead,1,1,total_overhead};
223
224 double sckRation = g_tp.sckt_co_eff;
225 local_result.power.readOp.dynamic *= sckRation;
226 local_result.power.writeOp.dynamic *= sckRation;
227 local_result.power.searchOp.dynamic *= sckRation;
228 local_result.power.readOp.leakage *= l_ip.nbanks;
229 local_result.power.readOp.longer_channel_leakage =
230 local_result.power.readOp.leakage*long_channel_device_reduction;
231 local_result.power = local_result.power* pppm_t;
232
233 local_result.data_array2->power.readOp.dynamic *= sckRation;
234 local_result.data_array2->power.writeOp.dynamic *= sckRation;
235 local_result.data_array2->power.searchOp.dynamic *= sckRation;
236 local_result.data_array2->power.readOp.leakage *= l_ip.nbanks;
237 local_result.data_array2->power.readOp.longer_channel_leakage =
238 local_result.data_array2->power.readOp.leakage*long_channel_device_reduction;
239 local_result.data_array2->power = local_result.data_array2->power* pppm_t;
240
241
242 if (!(l_ip.pure_cam || l_ip.pure_ram || l_ip.fully_assoc) && l_ip.is_cache)
243 {
244 local_result.tag_array2->power.readOp.dynamic *= sckRation;
245 local_result.tag_array2->power.writeOp.dynamic *= sckRation;
246 local_result.tag_array2->power.searchOp.dynamic *= sckRation;
247 local_result.tag_array2->power.readOp.leakage *= l_ip.nbanks;
248 local_result.tag_array2->power.readOp.longer_channel_leakage =
249 local_result.tag_array2->power.readOp.leakage*long_channel_device_reduction;
250 local_result.tag_array2->power = local_result.tag_array2->power* pppm_t;
251 }
252
253
254}
255
256void ArrayST::leakage_feedback(double temperature)
257{
258 // Update the temperature. l_ip is already set and error-checked in the creator function.
259 l_ip.temp = (unsigned int)round(temperature/10.0)*10;
260
261 // This corresponds to cacti_interface() in the initialization process. Leakage power is updated here.
--- 29 unchanged lines hidden (view full) ---
291 local_result.tag_array2->power.writeOp.dynamic *= sckRation;
292 local_result.tag_array2->power.searchOp.dynamic *= sckRation;
293 local_result.tag_array2->power.readOp.leakage *= l_ip.nbanks;
294 local_result.tag_array2->power.readOp.longer_channel_leakage = local_result.tag_array2->power.readOp.leakage*long_channel_device_reduction;
295 local_result.tag_array2->power = local_result.tag_array2->power* pppm_t;
296 }
297}
298
299ArrayST:: ~ArrayST()
300{
301 local_result.cleanup();
302}