• State Resolved
  • Locked Locked
  • Replies 1 reply
  • Subscribers 53 subscribers
  • Views 543 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Difference in number of cycles between COFF and ELF binaries?

B.J. Buchalter
Intellectual 875 points
Other Parts Discussed in Thread: SYSBIOS

Hi Folks,

I am starting to get my feet wet with developing code for the 674x DSP core. I am using the TI816x device, and, as such I am working with the xdc tools and cgt packages that are included with the ezsdk. I have been working through the RTSC tutorial and I have been trying to convert the samples to create elf binaries (as required by the syslink/sysbios system on the TI816x device. In the process I have found that the same "Hello World" sample app in lesson 5 takes a different number cycles to execute in the simulator depending whether it is compiled to COFF or to ELF (with the ELF taking more cycles). This is not whole-program optimization. Here are the results: 

$ /home/bj/ti-ezsdk_dm816x-evm_5_01_00_77/xdctools_3_20_08_88/packages/ti/platforms/sim64Pxx/Linux/kelvin prog.xe674 
simulating Joule FP ISA
Hello World
Simulation done:

  Total cycles:                             4391
  Core cycles (excl. stalls):               4390  ( 99.98%)
  Nop cycles:                               1977  ( 45.02%)
  Stall cycles and overlapped stall cycles
    Total stall cycles:                        1  (  0.02%)
      XP :                                     1  (  0.02%)

[snip]

$ /home/bj/ti-ezsdk_dm816x-evm_5_01_00_77/xdctools_3_20_08_88/packages/ti/platforms/sim64Pxx/Linux/kelvin prog.x674 
simulating Joule FP ISA
Hello World
Simulation done:

  Total cycles:                             3480
  Core cycles (excl. stalls):               3455  ( 99.28%)
  Nop cycles:                               1471  ( 42.27%)
  Stall cycles and overlapped stall cycles
    Total stall cycles:                       25  (  0.72%)
      XP :                                    25  (  0.72%)

[snip]

That is a pretty significant overhead (roughly 25%) just for a different executable format. Does anyone have any idea why this would be? Would whole-program optimization make this go away (I am still trying to figure out how to turn that on).

Are there any suggestions as to how to make this difference go away?

TIA, B.J.

  • 0
    TI__Guru* 84285 points

    For the most part, there's negligible performance difference bewteen COFF and ELF, but there are a few places where the ELF ABI (EABI) is just different, and the the compiler has to do something different with different performance.  In particular, EABI ensures that all uninitialized global variables are initialized to 0 before main begins, whereas COFF does not.  This probably accounts for the difference for this test case.