Getting HelloSpatial to work on AWS F1

[Naums Mogers]

Hi,

I've been following the manual on setting up Spatial on an AWS F1 (f1.2xlarge) instance: https://github.com/stanford-ppl/spatial-multiverse/blob/master/docs/aws.md

I've managed to complete all steps successfully with the HelloSpatial example, but encountered a problem at the very last one:

`sudo ./Top 1 1` allocates memories and starts the run, but hangs right after `[run] Begin`. Full output below.

Any pointers on how to proceed? Am I passing wrong arguments? Any advice would be greatly appreciated.

Full `Top` output:

$ sudo ./Top 1 1

[WARNING]: DELITE_NUM_THREADS undefined, defaulting to 1

Executing with 1 thread(s)

Adjusted current_heap_ptr to 256

Allocate mem of size 64 at (nil)

Adjusted current_heap_ptr to 4352

Allocate mem of size 1024 at 0x100

Adjusted current_heap_ptr to 20736

Allocate mem of size 64*64 at 0x1100

Adjusted current_heap_ptr to 21760

Allocate mem of size 2*2*2*2*16 at 0x5100

[memcpy HOST->DEV] hostmem = 0x8cf1a0, devmem = 0, size = 100

[memcpy HOST->DEV] hostmem = 0x8cf2b0, devmem = 100, size = 1000

[memcpy HOST->DEV] hostmem = 0x8d02c0, devmem = 1100, size = 4000

[memcpy HOST->DEV] hostmem = 0x8d42f0, devmem = 5100, size = 400

Adjusted current_heap_ptr to 25856

Allocate mem of size 32*32 at 0x5500

Adjusted current_heap_ptr to 29952

Allocate mem of size 1024 at 0x6500

Naums: start

[run] Begin

// Hangs here

[Naums Mogers]

UPD:

I've modified include/FringeContextAWS.h:run() to not just wait for the status change by polling aws_peek, but also breaking the polling loop after 30 seconds. Now, HelloSpatial returns, but all buffers are filled with "-1", and all tests fail. It's like the Accel block wasn't even executed on the FPGA.

[Matt F]

Hi Naums,

I'm not sure what is happening yet but I'm curious about whether there were any secret errors during synthesis that the flow ignored.  Would you be able to zip up and send the generated directories with all the logs and I can take a look?  

Thanks

[Naums Mogers]

Hi Matt,

Thanks for getting back to me! Here are the archived logs:

• Logs from <..>/aws-fpga/hdk/cl/examples/HelloSpatial/build (5.6 MB):  https://drive.google.com/file/d/1kpzETnjwQLLrfLQ1DpTjKhNwG1Stai3E/view?usp=sharing
• Logs from the S3 bucket (2.6 MB): https://drive.google.com/file/d/12ltHEivJytW-WI_Pf7billezUzVxTdUN/view?usp=sharing

Let me know if I can provide more information. Thanks!

[Matt F]

Hi Naums,

Nothing looks suspicious in the logs, but I have two ideas. 

1) I'm not sure if we tested the scatter/gather engines on the aws backend yet.  I would expect them to work, but they are significantly more complicated than dense dram transactions.  You can try commenting out these lines: https://github.com/stanford-ppl/spatial/blob/c1972756e7a1300b5c2870876059c3e0aaa391c7/apps/src/HelloSpatial.scala#L48-L51

2) Maybe the vivado/AMI/aws-fpga versions are doing something unexpected.   The versions we tested with here are:

• AMI 1.5.0 (includes Vivado 2018.2)
• aws-fpga 1.4.7

or

• AMI 1.6.0 (includes Vivado 2018.3)
• aws-fpga 1.4.8