micro: riscv32_mcu build failed with undefined references

[ehirdoy]

System information

• OS Platform and Distribution: Ubuntu 18.04
• TensorFlow installed from: source
• TensorFlow version: 298534b
• Python version: 2.7.15+/3.6.8
• Installed using virtualenv? pip? conda?: pip3

Describe the problem
TF Lite for micro, riscv32 build fails with undefined references.

make -f tensorflow/lite/experimental/micro/tools/make/Makefile TARGET=riscv32_mcu TARGET_ARCH=riscv32 hello_world_bin

/home/ehirdoy/src/tensorflow/tensorflow/lite/experimental/micro/riscv32_mcu/debug_log.cc:18: undefined reference to __wrap_puts' exit.c:(.text.exit+0x2e): undefined reference to __wrap__exit'
sbrkr.c:(.text._sbrk_r+0x12): undefined reference to __wrap__sbrk' writer.c:(.text._write_r+0x16): undefined reference to __wrap__write'
closer.c:(.text._close_r+0x12): undefined reference to __wrap__close' lseekr.c:(.text._lseek_r+0x16): undefined reference to __wrap__lseek'
readr.c:(.text._read_r+0x16): undefined reference to __wrap__read' fstatr.c:(.text._fstat_r+0x14): undefined reference to __wrap__fstat'
isattyr.c:(.text._isatty_r+0x12): undefined reference to `__wrap__isatty'

Provide the exact sequence of commands / steps that you executed before running into the problem

export PATH="$PATH:./tensorflow/lite/experimental/micro/tools/make/downloads/riscv_toolchain/bin"
make -f tensorflow/lite/experimental/micro/tools/make/Makefile clean
make -f tensorflow/lite/experimental/micro/tools/make/Makefile TARGET=riscv32_mcu TARGET_ARCH=riscv32 hello_world_bin

Any other info / logs
make.log

[aselle]

@dansitu, please take a look.

[Zielasko]

I am running into a similar error when trying to compile any of the examples for riscv using the same commands as listed above.
So far i have tested 3 different riscv-toolchains:

• /tensorflow/lite/experimental/micro/tools/make/downloads/riscv_toolchain/bin:
  undefined reference to __wrap__exit, __wrap_puts

• selfbuild toolchain (June 2019):
  undefined reference to __ashldi3, __clzsi2

• sifive toolchain (https://static.dev.sifive.com/dev-tools/riscv64-unknown-elf-gcc-20181030-x86_64-linux-ubuntu14.tar.gz):
  undefined reference to __wrap__exit, __wrap_puts

[zhoupeng]

I write patches which fix this bug, and tested on my machine.
This pull request to fix this issue is #33972
Hi @ehirdoy @Zielasko
Could you have a try of these patches pls? If you find it helpful, pls have a support click.

Best,

[Zielasko]

Thank you for writing a fix @zhoupeng .
I tried to write a workaround in an older version of this repository that let me compile the hello world example and a few examples i have written myself, but they didn't behave correctly.
Using your version lets me compile and successfully run the hello world example on a risc-v simulator with the expected output.
(Using make -f tensorflow/lite/experimental/micro/tools/make/Makefile TARGET=riscv32_mcu hello_world to build the executable)

Unfortunately I still can't build the micro_speech, magic_wand or person_detection examples for
riscv32_mcu.
For the magic_wand example it cant find __wrap_malloc and __wrap_free .

person_detection fails with the error

person_detection section `.bss' will not fit in region `ram' 
[...]
section .stack VMA [0000000080003800,0000000080003fff] overlaps section .bss VMA [00000000800002b8,0000000080012d87]
[...]
region `ram' overflowed by 0 bytes

I have yet to test my own models.

Also, trying to build a linux version of the hello world example with the __dso fix gives me this error:
multiple definition of __dso_handle'`

[zhoupeng]

Hi @Zielasko , Thanks for your response. Sorry for late reply, I am not convenient to handle mail in time, because on a travel of RISC-V tech forum.
1)For the .bss' will not fit in region ram error:
Set the MEMORY.ram option with a larger value in the link script can fix.
The ram option is in the link script tensorflow.git/tensorflow/lite/experimental/micro/tools/make/downloads/sifive_fe310_lib/bsp/env/freedom-e300-hifive1/flash.lds.
In my case, I only try to set ram (wxa!ri) : ORIGIN = 0x80000000, LENGTH = 128K (Maybe 64K or 32K is enough for some examples, but I didn't try)

2)__dso_handle missing problem. Thanks for response, very helpful.
This problem may only exists on non-Linux target (riscv32_mcu,Windows ...), I will submit a new version patch adding target OS check.
Before the new patch, If you try to build a linux version, you can try to comment __dso_handle by replacing void* __dso_handle;
with // void* __dso_handle;, in tensorflow.git/tensorflow/lite/experimental/micro/arduino/abi.cc,

Hi @nkreeger @gbaned @rockyrhodes, sorry for late, thanks @nkreeger for your detail review I will submit a new version patch, catching up on this Fridays after my RISC-V tech forum journey.

[ymodak]

Closing since the associated PR has been merged. Feel free to reopen if necessary. Thanks!

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No