On Icebreaker, TFLM unit test fails, and microbenchmarks hang about cfu-playground HOT 3 CLOSED

I tried inserting printfs in the "cached loads" benchmark, after reducing BUF_SIZE to 2048 (edit: this is 8kB, since the buffer has BUF_SIZE int32 elements), and captured the following output before it hang: buf: 0x00dee0-0x00fedc acc: 0x00dedc Edit: these are locally allocated on the stack.

Using nm to look at symbol addresses in .data, the highest locations I could find were:

00003468 b rx_buf
000034e8 b tx_buf
00003568 B __malloc_current_mallinfo
00003590 B errno
00003598 B _ebss
00003598 B _end

So it looks like there's plenty of space...maybe it's not actually hanging, or my theory is wrong.

OH -- it didn't hang, it was just very slow --- 742 cycles per load when executing out of spiflash!

For this board, the VexRiscv build has NO Icache or Dcache, and the code is located in spiflash, not main_ram or sram.

from cfu-playground.

Good catch on the TFLM unit tests, I was confused because I thought I tested them before submitting the PR. It turns out the unit tests fail only when no models are included and only for the Icebreaker (at least on the PR branch). I'll need to look into why the unit tests are behaving that way, I'm currently not sure why they'd pass when a model is included and only fail for the icebreaker when one isn't.

As for the benchmarks, a buffer size of 128kB is a bit big for the Icebreaker, I agree we should think about including a configuration that decreases the size a bit.

from cfu-playground.

Fixed with #136.

from cfu-playground.