google / google/CFU-Playground
Tensorflow model evaluation hangs when CFU support is turned off
- Dominant language
- Verilog
- Stars
- 565
- Forks
- 160
- PR merge metrics
- No merged PRs in 30d
Description
I was building for HPS with CFU support turned off in the design, using a local hack like this:
```patch
--- a/soc/hps.mk
+++ b/soc/hps.mk
@@ -34,7 +34,7 @@ endif
PROJ_DIR:= $(CFU_ROOT)/proj/$(PROJ)
CFU_V:= $(PROJ_DIR)/cfu.v
-CFU_ARGS:= --cpu-cfu $(CFU_V)
+#CFU_ARGS:= --cpu-cfu $(CFU_V)
SOC_NAME:= hps.$(PROJ)
OUT_DIR:= build/$(SOC_NAME)
```
and then running:
make -C proj/proj_template PLATFORM=hps bitstream software
If I use the menu to try and execute CFU op0, the application freezes. This is expected, since the CPU tried to execute an illegal instruction.
If I use the menu to try and evaluate pdti8, the application also freezes here:
```
Tests for pdti8 model
=====================
1: Run with zeros input
2: Run with no-person input
3: Run with person input
g: Run golden tests (check for expected outputs)
x: eXit to previous menu
pdti8> 3
Running Run with person input
Classify Person
Copied 9216 bytes at 0x0x40005680
Running pdti8
```
Since I am building `proj/proj_template` there are no customisations being applied to Tensorflow so I don't think it could be executing a CFU op. I was expecting it to evaluate the model as normal.
Contributor guide
Assessment
This issue has not been assessed yet.