gemmini icon indicating copy to clipboard operation
gemmini copied to clipboard

system freeze while running tests on FPGA.

Open 12ff7a6 opened this issue 4 years ago • 10 comments

hello, I have deployed 2 rocketcores and Gemmini accelerator on Xilinx VC707 FPGA by using the repo https://github.com/eugene-tarassov/vivado-risc-v. But I got problems when I was running Gemmini tests. For example, if I run conv_with_pool-linux, then I get

debian@debian:/boot/bareMetalC$ ./conv_with_pool-linux
Output dimension: 112
Pooling output dimension: 56

Randomize inputs...
Randomize weights...
Randomize bias...
CPU conv...
CPU conv took 4275198029 cycles
CPU pool...
CPU pool took 122327864 cycles
CPU conv+pool took 4397525893 cycles
Flatten w[  911.674284] systemd[1]: Starting Cleanup of Temporary Directories...
[ 1027.062508] systemd[1]: systemd-udevd.service: State 'stop-watchdog' timed out. Killing.
[ 1027.078076] systemd[1]: systemd-udevd.service: Killing process 158 (systemd-udevd) with signal SIGKILL.

or if I run mvin_mvout-linux then I get no response and the system freeze forever. I must restart the FPGA. Does anyone know how to get rid of this problem?

12ff7a6 avatar Oct 13 '21 18:10 12ff7a6

Hi there, I faced similar issues and noticed that the pointer mechanism is buggy. When I use a pointer to call functions like tiled_conv_A_stride, the system freezes. but when i call the functions directly, it seems to pass the test. Would you mind sharing your updates on this issue? Thanks!

johnny-wang16 avatar Nov 08 '21 01:11 johnny-wang16

@12ff7a6 Hmm, we haven't tried to run on FPGAs before. Are you able to replicate this issue on Verilator/VCS or Firesim?

@johnny-wang16 I'm not familiar with that error. Do you have an example code where you run into that issue? Calling tiled_conv_A_stride with pointers should be fine...

hngenc avatar Nov 11 '21 00:11 hngenc

hi @hngenc thanks for the reply! It turns out that I had an illegal stack pointer initialization. after i modified _stack_start, the issue is resolved. Thanks.

johnny-wang16 avatar Nov 11 '21 09:11 johnny-wang16

@12ff7a6 could you share how you were able to get Gemmini running on VC707?

gnodipac886 avatar Nov 14 '22 05:11 gnodipac886

@gnodipac886 try this repo https://github.com/eugene-tarassov/vivado-risc-v

12ff7a6 avatar Nov 14 '22 07:11 12ff7a6

thanks!

gnodipac886 avatar Nov 14 '22 18:11 gnodipac886

hi @hngenc thanks for the reply! It turns out that I had an illegal stack pointer initialization. after i modified _stack_start, the issue is resolved. Thanks.

@johnny-wang16 Do you happen to remember how you were able to solve this issue? Thanks in advance!

gnodipac886 avatar Nov 22 '22 02:11 gnodipac886

Hi there, it's been awhile but i think the problem i faced atm was different from yours. if my memory serves me right, there's a config file or somewhere that can specify the memory map. i had to change that config file to solve the issue.

johnny-wang16 avatar Nov 22 '22 08:11 johnny-wang16

ah, ok, thank you, I'll try to see if I can find it :)

gnodipac886 avatar Nov 23 '22 00:11 gnodipac886