1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
|
Porting Cpuburn on ARM
* Cortex A8
To begin, I studied the Cortex-A8 and I have found that the Neon is the entity which consume the more.
So all my code is build with Neon's instructions.
There is only two pipes in the Cortex-A8 so I try to always have this pipes full.
In the first part I initialize my loop by entering variable in Neons registers.
Then I begin my loop with Neon's Instructions.
At the begining, I had this loop :
loop:
subs r1, #1
bne loop
I know that my processor is at 600mHz and I do 4 000 000 000 iterations so I can find the number of cycle.And I saw that this loop takes 3 cycles.
After that, I added instructions in the loop to see which instructions cost 0 cycle or which instructions can be parallelized.
This leads me to the actual loop which takes 4 cycles.
At Texas Instruments, I have boards which allow me to measure temperature and power.
Here is the results for temperature :
idle : 35 C
empty loop : 38 C
Burn : 43 C
And the power measure gives:
idle : 262 mW
empty loop : 365 mW
burn : 693 mW
Measurement of power are done on the Omap3630 in the cpu's rail and measure of temperature with the band gap temperature sensor.
When i quit the loop, I do a test to see if the computation is good, if not the programm exit.
To compile this program, I use codesourcery toolchain : arm2009q1.
* Cortex A9
The Neon is the entity which consumes the more like on Cortex A8.
So I began my research on the Neon with the same method as Cortex A8.
Here the empty loop take only one cycle due to "Fast loop mode".
I managed to use two Neon's entity at the same time : load and multiply.
We can't use more Neon's instructions in parallel so we need to use other instructions to complete the pipeline.
So I try with Arm instructions and I managed to add two ARM instructions in parallel of Neon's instructions.
At the end, I have a loop of 3 cycles with 4 instructions in parallel.
I get this results on an omap4 board with two Cortex-A9 :
--------------- -----------------------------------------------
| | Delta of temperature | Delta of consumption |
|---------------|---------------------------------------------- |
|1*burnCortexA9 | + 12 C | + 0.6 W |
| | | |
|1*burnCortexA8 | + 11 C | + 0.56 W |
| | | |
|2*burnCortexA9 | + 26 C | + 1.26 W |
| | | |
|2*burnCortexA8 | + 28 C | + 1.16 W |
| | | |
----------------------------------------------------------------
I tested burnCortexA9 on omap3 too and I get a delta of 560 mW between burnCortexA8 and burnCortexA9. (BurnCortexA8 is still the best on omap3).
Gregory Herrero
9 Jul 2010
|
