Forum Discussion

DongWang-BJTU's avatar
DongWang-BJTU
Icon for Occasional Contributor rankOccasional Contributor
6 years ago

Emulator v19.3 is much slower than v19.1/18.1

I have currently upgraded to OpenCLSDK v19.3 and found that the legacy emulator is much slower than previous versions. I am using the same OpenCL code with the same compilation configurations. Is t...
MEIYAN_L_Intel's avatar
MEIYAN_L_Intel
Icon for Frequent Contributor rankFrequent Contributor
6 years ago

Hi,

May I know the information as below for confirmation:

  1. Does the compile boardtest.cl for emulation pass?
  2. Does the emulator work for the example provided by developer eg: vector_add / hello_world?
  3. There is a change for OpenCL SDK v19.3 which is "Changed CL_CONTEXT_EMULATOR_DEVICE_INTELFPGA to CL_CONFIG_CPU_EMULATE_DEVICES". Do you used "CL_CONFIG_CPU_EMULATE_DEVICES" as a command to run emulation?
  4. what is the command you have used for emulation?

eg: aoc march=emulator-legacy-emulator

Thanks

DongWang-BJTU's avatar
DongWang-BJTU
Icon for Occasional Contributor rankOccasional Contributor
6 years ago

Hi, I can confirm the boardtest.cl and the example works fine for the emulator as the following log shows:

---------------------------

Initializing OpenCL

Platform: Intel(R) FPGA Emulation Platform for OpenCL(TM)

Using 1 device(s)

Intel(R) FPGA Emulation Device

Using AOCX: vector_add.aocx

Launching for device 0 (1000000 elements)

Time: 4.284 ms

Kernel time (device 0): 1.197 ms

Verification: PASS

--------------------------

I am fully aware of the new ENV for v19.3 and the variable was correctly set.

The official examples provided are too small that only execute for a few seconds that makes the performance difference between the two emulators unnoticable.

You could use our opensource project PipeCNN (https://github.com/doonny/PipeCNN) to test the emulator, please try to run VGG16/ResNet-50 benchmarks and you will see the difference in emulator execution time between v19.1 and v19.3.

Our inhourse project has a much bigger longer runtime when using the new emulator. The project uses several autorun kernels, I susbect the autorun kernels caused this problem. I have captured the CPU utilization as follow. You can clear see there is a big difference: the legacy emualtor uses more threads compared to the new one.

  1. emulaor v19.3 (new emulator)

2. emulator v19.1 (lagecy one)

I am using Ubuntu 16.04.6 (Linux fpga-z390 4.15.0-72-generic #81~16.04.1-Ubuntu SMP Tue Nov 26 16:34:21 UTC 2019 x86_64 x86_64 x86_64 GNU/Linux

)

on a i9-9900K CPU with 64GB memory.