Toolchain issues #82
Comments
|
I am sorry for the inconvenience… I currently provide various GCC (13/14) & LLVM (15-19) builds for Ubuntu 20.04 & 22.04. Unfortunately I ran out of the 50GB auf space I can use on the could service provider. To also upload 24.04 variants of the tools I would need ~5GB more which is currently difficult to free up. There is a likelihood that the 22.04 binaries also work on 24.04, so you could give it a try by hardcoding the Ubuntu version detected by the script. Regarding multilib support: Actually I have uploaded multilib toolschains to the Sync&Share yesterday. Feel free to give them a try, but it is highly likely that they will not work together with LLVM because RISC-V multilib support is basically broken since forever. This is the reason why we have to maintain many non-multilib builds instead of only one multilib-GCC. I would love to eliminate this issue, but guess it’s up to the RISC-V & LLVM community to deal with that. Some further notes:
If you need guidance in setting up muRISCV-NN and Vicuna, please feel free to reach out to us! |
|
Here you can browse for the available Toolchains: https://syncandshare.lrz.de/getlink/fiWBtDLWz17RBc1Yd4VDW7/ |
|
Hi, I actually tried rv32im_zve32x_zicsr_zifencei and an error message appeared as below. Could you comment on how this could be solved? Thanks in advance. ===========================================vincent@vincent-debug:~/Proj/muriscv-nn/build$ cmake -DSIMULATOR=Vicuna -DRISCV_GCC_PREFIX=$(pwd)/../Toolchain/rv32im_zicsr_zifencei_zve32x_ilp32 -DENABLE_INTG_TESTS=ON -DTOOLCHAIN=GCC -DUSE_VEXT=ON .. is not able to compile a simple test program. It fails with the following output: CMake will not be able to correctly generate this project. -- Configuring incomplete, errors occurred! |
@chchenl Of course. Since the TC was built with This should fix your build issue: diff --git a/CMakeLists.txt b/CMakeLists.txt
index 17300ed..93c72f7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -44,7 +44,7 @@ option(USE_PEXT "Use the RISC-V P Packed Extension" OFF)
# - rv32imzve32x / ilp32 -> integer only and embedded vector support (GCC only supports rv32imv!)
if("${SIMULATOR}" STREQUAL "Vicuna") # Vicuna only supports a reduced set of extensions
if(USE_VEXT)
- set(RISCV_ARCH "rv32imzve32x" CACHE STRING "march argument to the compiler, using the vector extension")
+ set(RISCV_ARCH "rv32im_zicsr_zifencei_zve32x" CACHE STRING "march argument to the compiler, using the vector extension")
elseif(USE_PEXT)
message(FATAL_ERROR "Vicuna does not support packed instructions.")
else() |
|
Hi, I modified the CMakeLists.txt accordingly, but it seems the error message remains the same.There seems to be 3 suspicious messagesIs this the reason why the cmake command failed? Thanks for info.(BTW, don't know why the font size here gets so large--sorry about that. --Fix now, the font size is much smaller.)================ cmake -DSIMULATOR=Vicuna -DRISCV_GCC_PREFIX=$(pwd)/../Toolchain/rv32im_zicsr_zifencei_zve32x_ilp32 -DENABLE_INTG_TESTS=ON -DTOOLCHAIN=GCC -DUSE_VEXT=ON .. is not able to compile a simple test program. It fails with the following output: CMake will not be able to correctly generate this project. |
|
Another message near the end...: /home/vincent/Proj/muriscv-nn/Toolchain/rv32im_zicsr_zifencei_zve32x_ilp32/bin/riscv32-unknown-elf-gcc -march=rv32imzve32x -mabi=ilp32 -mcmodel=medany -march=rv32imzve32x -mabi=ilp32 -mcmodel=medany CMakeFiles/cmTC_b7209.dir/testCCompiler.c.obj -o cmTC_b7209 |
|
@chchenl That’s odd. I tried the fix and it worked for me as expected. Here is some advice for further investigation:
|
|
Hi, Here is a new message, which seems to indicate I skipped some steps(downloading something?)--but I cannot tell if I skipped something from the readme file. I suspect it resuled from no-running Integration/tflm/download_tflm.sh along the way. But isn't this supposed to run automatically or invoked by some other (intermediate) script? I tried to manually issue 'download_tflm.sh', but the directory structure generated does not coincide with the directory structure the 'error message' expected, therefore you may have to manually adjust its structures. I guess this is not a 'standard' approach, though. cmake -DSIMULATOR=Vicuna -DRISCV_GCC_PREFIX=$(pwd)/../Toolchain/rv32im_zicsr_zifencei_zve32x_ilp32 -DENABLE_INTG_TESTS=ON -DTOOLCHAIN=GCC -DUSE_VEXT=ON .. Tried extensions .c .C .c++ .cc .cpp .cxx .cu .mpp .m .M .mm .ixx .cppm CMake Error at Integration/tflm/CMakeLists.txt:19 (add_library): |
|
Hi PhilippvK, Another idea occurred to me. Or if I am on the wrong path to that goal, could you show me a proper path/sequence to that goal? Thanks. |
For reference an setting up and running the integration tests please refer to the CI scripts: https://github.com/tum-ei-eda/muriscv-nn/blob/master/.github/workflows/integration_tests.yml Instead of invoking cmake manually, you can use the Due to the various limitations/bugs in the upstream version of Vicuna ist it unlikely that the integration tests will work out-of-the box… |
The mentioned in the last post, the best „documentation/example“ on how to setup the flow is to look at the ci scripts which should be working 100% this week. It currently just uses Spike as Somulator, but we can look into integrating Vicuna there as well. |
PhilippvK
added
bug
The dowload_xxx.sh files seem not-working(using ubuntu 23.x or 24.x)
In vicuna RTL sim, if using built-in downloaded toolchain(version 22.x),
when the below command is issued(as prescribed in Sim/Vicuna/Readme file):
"cmake -DSIMULATOR=Vicuna -DRISCV_GCC_PREFIX=$(pwd)/../Toolchain/rv32imv/ -DENABLE_INTG_TESTS=ON -DTOOLCHAIN=GCC -DUSE_VEXT=ON .."
an error message saying that it cannot find 'multilib' required for the architecture rv32im_zvxxxx_zlxxx in the toolchain rv32imv (I cannot capture the screen right now. Probably I'll upload it sometime later.)
Or rather, a more detailed error message is: "cannot find suitable multilib set for -march=rv32im_zve32x_zvl32b'/' -mabi=ilp32.....", then this cmake command failed and terminated.
This raises another issue: is the built-in(downloaded) toolchain buit with '--enable-multilib' option enabled?
From the readme file, I guess the recommended approach for building one's own toolchain does not enable
this option.
Or if this issue is related to versions of gcc or any other utilities? I use gcc-13.2.
Thanks for any info.
--vincent
The text was updated successfully, but these errors were encountered: