De1-SoC Performance Different from Example Given !

[mingyi136]

Hai @doonny , I have run inference on De1-SoC board with VEC_SIZE=8 and LANE_NUM=8 (other parameters remain unchanged).

However, the Total kernel runtime is 236.344 ms instead of 149.988 ms given in the example. Is there any additional changes in parameters / coding compared to old version?

Here is my inference result:

***************************************************
PipeCNN: An OpenCL-Based FPGA Accelerator for CNNs
***************************************************

Platform: Intel(R) FPGA SDK for OpenCL(TM)
Totally 1 device(s) are found
  Using Device 0: de1soc_sharedonly : Cyclone V SoC Development Kit
Device OpenCL Version: OpenCL 1.0 Intel(R) FPGA SDK for OpenCL(TM), Version 16.1
Device Max Compute Units: 1
Device Max WorkGroup Size: 2147483647
Device Max WorkItem Size: 2147483647
Device Global Memory Size: 512 MBytes
Device Local Memory Size: 16 KBytes
Device Max Clock Freq: 1000 Mhz

Loading kernel/binary from file conv.aocx
Reprogramming device [0] with handle 1

61063552 total weights read
1024 total output reference read


154587 bytes image data read from binary files

Executing Layer 1:

Launching single work-item kernel winbuffer

Launching single work-item kernel Conv

Launching kernel MemWr with local size: 1, 1, 8  (global size: 55, 55, 96)

Launching single work-item kernel Pool

Launching kernel lrn with local size: 1, 1, 12  (global size: 27, 27, 12)

Executing Layer 2:

Launching single work-item kernel winbuffer

Launching single work-item kernel Conv

Launching kernel MemWr with local size: 1, 1, 8  (global size: 27, 27, 256)

Launching single work-item kernel Pool

Launching kernel lrn with local size: 1, 1, 32  (global size: 13, 13, 32)

Executing Layer 3:

Launching single work-item kernel winbuffer

Launching single work-item kernel Conv

Launching kernel MemWr with local size: 1, 1, 8  (global size: 13, 13, 384)

Executing Layer 4:

Launching single work-item kernel winbuffer

Launching single work-item kernel Conv

Launching kernel MemWr with local size: 1, 1, 8  (global size: 13, 13, 384)

Executing Layer 5:

Launching single work-item kernel winbuffer

Launching single work-item kernel Conv

Launching kernel MemWr with local size: 1, 1, 8  (global size: 13, 13, 256)

Launching single work-item kernel Pool

Executing Layer 6:

Launching single work-item kernel winbuffer

Launching single work-item kernel Conv

Launching kernel MemWr with local size: 1, 1, 8  (global size: 1, 1, 4096)

Executing Layer 7:

Launching single work-item kernel winbuffer

Launching single work-item kernel Conv

Launching kernel MemWr with local size: 1, 1, 8  (global size: 1, 1, 4096)

Executing Layer 8:

Launching single work-item kernel winbuffer

Launching single work-item kernel Conv

Launching kernel MemWr with local size: 1, 1, 8  (global size: 1, 1, 1024)

Copyed all batched results from fc_1 buffers.
Selected item = 0 from the combined batch results in fc buffers

Start verifying results ...

Check Pass !!!

The inference result is n02123045 tabby, tabby cat   (the prob is 56.00)


PipeCNN exited !!!


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

Performance Summary

Kernel runtime summary:
  Layer-1:
    MemRd: 70.738 ms
    Conv : 70.578 ms
    Pool : 70.346 ms
    MemWr: 70.486 ms
    Lrn  : 1.383 ms
  Layer-2:
    MemRd: 56.435 ms
    Conv : 56.304 ms
    Pool : 56.106 ms
    MemWr: 56.241 ms
    Lrn  : 0.456 ms
  Layer-3:
    MemRd: 39.022 ms
    Conv : 38.899 ms
    Pool : 0.000 ms
    MemWr: 38.827 ms
    Lrn  : 0.000 ms
  Layer-4:
    MemRd: 28.978 ms
    Conv : 28.854 ms
    Pool : 0.000 ms
    MemWr: 28.788 ms
    Lrn  : 0.000 ms
  Layer-5:
    MemRd: 19.408 ms
    Conv : 19.272 ms
    Pool : 19.081 ms
    MemWr: 19.209 ms
    Lrn  : 0.000 ms
  Layer-6:
    MemRd: 14.490 ms
    Conv : 14.371 ms
    Pool : 0.000 ms
    MemWr: 14.262 ms
    Lrn  : 0.000 ms
  Layer-7:
    MemRd: 6.538 ms
    Conv : 6.423 ms
    Pool : 0.000 ms
    MemWr: 6.344 ms
    Lrn  : 0.000 ms
  Layer-8:
    MemRd: 1.758 ms
    Conv : 1.642 ms
    Pool : 0.000 ms
    MemWr: 1.562 ms
    Lrn  : 0.000 ms

Total kernel runtime 236.344 ms
Batch size = 1, average process time per batch: 236.344 ms

Total runtime: 0.241783s

[mingyi136]

As referring to this issue: #46 (comment)

Total kernel runtime of 149.988 ms on DE1-Soc Board seem like have been achieved with parameter VEC_SIZE=8 and LANE_NUM=8, too

[mingyi136]

@doonny
I have regenerated both run.exe & conv.aocx using PipeCNN github of 2018 (VEC_SIZE=8 and LANE_NUM=8).

This time I managed to get Total kernal runtime of 157.928 ms. Just wondering, why the latest version of PipeCNN github give a slower inference performance for DE1-SoC board?

[doonny]

May I ask which version of SDK are you using for compilation ?

[sergio14890]

@mingyi136 where you download the bsp for de1soc board?

[mingyi136]

May I ask which version of SDK are you using for compilation ?

@doonny I have compiled conv.aocx using OpenCL SDK 17.1 on Windows, whereas run.exe has been compiled using OpenCL SDK 16.1 on DE1-SoC board (Linux).

[sergio14890]

May I ask which version of SDK are you using for compilation ?

@doonny I have compiled conv.aocx using OpenCL SDK 17.1 on Windows, whereas run.exe has been compiled using OpenCL SDK 16.1 on DE1-SoC board (Linux).

Ahhh okk, i try compile using openCL SDK 17,1, but i have this error:
https://github.com/doonny/PipeCNN/issues/135

You have a license for 16.1 SDK?

[mingyi136]

@sergio14890 , I downloaded Linux SD Card Image (inside has OpenCL 16.1) from here:
https://software.intel.com/content/www/us/en/develop/topics/fpga-academic/learn/tutorials.html
and use it to compile run.exe.

Whereas conv.aocx has been compiled using OpenCL 17.1 & pointing towards DE1-SoC BSP (OpenCL 16.0) which available here:
https://www.terasic.com.tw/cgi-bin/page/archive.pl?Language=English&No=836&PartNo=4

[doonny]

The latest code is optimized for SDK v19.1, in which some features are not supported by older version, like v16.1 and v17.1. We suggest upgrade the SDK to v19.1.