[Solved] Bullseye Toolchain Error: selected processor does not support `vsdot.s8 q10,q9,d4[0]' in ARM mode

[distlibs]

I was trying to compile tensorflow-lite 2.7 with Raspberry Pi GCC Cross-Compiler Toolchains (Bullseye). I got error:

...
[ 73%] Building C object _deps/xnnpack-build/CMakeFiles/XNNPACK.dir/src/qc8-igemm/gen/1x16c4-minmax-fp32-neondot.c.o
[ 73%] Building C object _deps/xnnpack-build/CMakeFiles/XNNPACK.dir/src/qc8-igemm/gen/4x8c4-minmax-fp32-neondot.c.o
[ 73%] Building C object _deps/xnnpack-build/CMakeFiles/XNNPACK.dir/src/qc8-igemm/gen/4x16c4-minmax-fp32-neondot.c.o
/tmp/ccLir1WO.s: Assembler messages:
/tmp/ccLir1WO.s:60: Error: selected processor does not support `vsdot.s8 q8,q11,d7[0]' in ARM mode
/tmp/ccLir1WO.s:62: Error: selected processor does not support `vsdot.s8 q10,q9,d7[0]' in ARM mode
/tmp/ccLir1WO.s:65: Error: selected processor does not support `vsdot.s8 q8,q9,d7[1]' in ARM mode
/tmp/ccLir1WO.s:69: Error: selected processor does not support `vsdot.s8 q10,q9,d7[1]' in ARM mode
/tmp/ccLir1WO.s:141: Error: selected processor does not support `vsdot.s8 q10,q9,d7[0]' in ARM mode
/tmp/ccLir1WO.s:144: Error: selected processor does not support `vsdot.s8 q8,q11,d7[0]' in ARM mode
make[2]: *** [_deps/xnnpack-build/CMakeFiles/XNNPACK.dir/build.make:5458: _deps/xnnpack-build/CMakeFiles/XNNPACK.dir/src/qc8-gemm/gen/1x8c4-minmax-fp32-neondot.c.o] Error 1
make[2]: *** Waiting for unfinished jobs....
/tmp/ccpalmhZ.s: Assembler messages:

Using Buster toolchain tensorflow-lite compiles without any errors.

It can be related with:

• https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101723
• Cross-compilation for RPi3 (armv7) fails on assembly google/XNNPACK#1465
• CMake build error on armv7 with latest gcc google/XNNPACK#2003

Maybe you can apply patch like you did here #89

[abhiTronix]

Using Buster toolchain tensorflow-lite compiles without any errors.

It can be related with:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101723
Cross-compilation for RPi3 (armv7) fails on assembly google/XNNPACK#1465
CMake build error on armv7 with latest gcc google/XNNPACK#2003

@distlibs Thanks for bug report. What I'm understanding is that GCC 9-to-11 series is affected by this bug when Binutils >2.33 is used. More specifically Binutils 2.35.1 used in Bullseye is causing this bug in our case. And, this bug is already fixed in master branch of GCC >9 but haven't been included in releases branches like releases/gcc-10.3.0 which was updated 8 months ago.

Solution:

This patch will be released with upcoming version of GCC-10 and GCC-11. Untill then, we can clone patched releases/gcc-10 branch directly for compiling GCC >=10.3.0 binaries for Bullseye to resolve this bug. But we cannot do the same for GCC-10.2.0 since, master branch will also override its version to >10.2.0 rendering it useless for cross-compiling. So just use >=10.3.0 for compiling & cross-compiling tensorflow.

[distlibs]

Binaries cross-compiled with GCC 10.3.0 will be compatible with Raspberry Pi 4 which runs Bullseye with GCC 10.2.1?

[abhiTronix]

Yes, aren't you using Raspberry Pi GCC-10.3.0 Native-Compiler Toolchains (Bullseye) right now? The GCC version doesn't matter, but only GLIBC, KERNEL and BINUTILS version matters, which remains the same.

[distlibs]

I am using right now GCC-10.2.0 Raspberry Pi Cross-Compiler Toolchains (Bullseye). I made mistake, I updated original post. Download link.

[abhiTronix]

Yes, you can use GCC-10.3.0 instead.

[distlibs]

So I need to wait until you clone patched releases/gcc-10 branch directly and pre-compile GCC 10.3.0 binaries for Bullseye? New Raspberry Pi GCC 10.3.0 Cross-Compiler Toolchains (Bullseye) binaries will be uploaded SourceForge where I can download?

[abhiTronix]

You need to wait untill v3.1.0 Toolchains are released, and thereby uploaded to Sourceforge after successful compiling. You should watch this repository to get notify when that happens.

[abhiTronix]

@distlibs Patch has been added. Please wait for CI workflows to finish and then I'll make the release.

[abhiTronix]

@distlibs v3.1.0-alpha toolchains are in build and soon be available on Sourceforge.

[abhiTronix]

This issue has been resolved in the latest Toolchains version v3.1.0-alpha and GCC-10.3.0. Download it from here: https://sourceforge.net/projects/raspberry-pi-cross-compilers/files/Raspberry%20Pi%20GCC%20Cross-Compiler%20Toolchains/Bullseye/GCC%2010.3.0/