Hardware Reference
In-Depth Information
Instructions
Time
Skyeye
M3SCoPE
Error %
Skyeye
M3SCoPE
Speed-up
Bubble
5200180007
5200180007
0
4m45,53s
3,76s
x71
Factorial
2747041
2996535
9,1
1,28s
0,02s
x64
Hanoi
18481575
17695142
4,3
11,15s
0,11s
x100
Coder GSM
13466069
14066581
4,4
10,6s
0,09s
x117
H264
5601674012
5800347641
3,5
5m25,6s
3,92s
x80
Fig. 2.14 Analysis of accuracy and speed of the estimation technique. The results obtained with
M3-SCoPE have been compared with SKyeye [ 56 ], a no cycle-accurate ISS; the faster type of
processor simulator
Data Cache Misses
Instruction Cache Misses
M3-SCoPE
Example
ISS (SkyEye)
Error (%)
M3-SCoPE
ISS (SkyEye)
Error (%)
Bubble
25
27
8
5204878
5199772
0.09
Hanoi
20
18
10
45
38
15.55
Factorial
8
7
12
500
375
25
GSM - 1 frame
2
660
670
1.49
3738
3663
GSM - 4 frames
15324
14814
3.4
2370
2452
3.46
GSM - 7 frames
26333
25730
2.3
4104
4235
3.19
GSM - 10 frames
37425
36624
2.2
5915
6026
1.84
Fig. 2.15 Accuracy of the cache models. The error is more representative in large examples, since
in small examples, the number of errors is too small for comparisons
As can be shown (Fig. 2.14 ), the time estimation error of the native co-simulation
technique is lower than 10%. Regarding the errors in the estimation of the number
of misses, which has a high importance in bus and network modeling, the errors are
in a similar range (Fig. 2.15 ).
In benchmarks with a small number of function calls, (i.e the bubble) the miss
rate estimation errors is less than 1%, but when the number of calls to function is
increased the miss rate grows too. Hanoi benchmark has a large percent of function
calls in his instructions, his miss rate estimation error is 15% and in Factorial, which
is an extreme code that only has function calls, the miss rate estimation error is 25%.
This error is due to the big number of function calls, the difference in the compilers
for different architectures produces that the number of registers saved in a function
 
Search WWH ::




Custom Search