[CI][Docker] Add riscv-gnu-toolchain to ci_riscv

[katebern-grovety]

Currently the ci_riscv docker image uses GNU toolchain for Xuantie RISC-V CPU, which does not contain V-extension instructions. This PR adds GNU toolchain for RISC-V from RISC-V Collaboration. This update is required for the PR

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @Mousius, @areusch, @leandron _{See #10317 for details}

_{Generated by tvm-bot}

[katebern-grovety]

cc @Mousius, @areusch, @leandron, @PhilippvK, @alter-xp

[PhilippvK]

In contrast to the Xuantie GCC, this is a non-multilib build of the RISC-V GNU Toolchain. Support is probably limited to rv64gc/lp64d.

In the long term, we should find a way to support riscv32 and riscv64 (eventually, with different levels of floating-point support).

[katebern-grovety]

Thanks @PhilippvK! I'll replace it with a multilib build.

[PhilippvK]

Do you have an guess on how long the from source build will take? I am not sure If this would be acceptable… Maybe there is a multilib build available to use here: https://github.com/sifive/freedom-tools/releases

[katebern-grovety]

I guess it will take a few hours, by the link you specified there are builds from two years ago. As a solution, I can publish a release in forked riscv-gnu-toolchain repository.

[PhilippvK]

Same issue here. RISC-V proxy kernel needs to be built separately for rv32gc and rv64gc.

[katebern-grovety]

I'll add build for rv32gc and rv64gc.

[PhilippvK]

The RISC-V GCC is only supposed to be used for linking and providing the libc headers to Clang right?

[katebern-grovety]

The RISC-V GCC is only supposed to be used for linking and providing the libc headers to Clang right?

Yes, RISC-V GCC is used for linking and providing libs.

[alter-xp]

The current version of Xuantie GCC supports V-extension. However, as I mentioned in the PR, is it syntactically correct to directly use vwredsum ?

[katebern-grovety]

The current version of Xuantie GCC supports V-extension. However, as I mentioned in the PR, is it syntactically correct to directly use vwredsum ?

But on the current version of the toolchain, the error
/opt/riscv/riscv64-unknown-linux-gnu/bin/riscv64-unknown-linux-gnu-ld: -march=rv64i2p0_m2p0_a2p0_f2p0_d2p0_c2p0_v1p0_zve32f1p0_zve32x1p0_zve64d1p0_zve64f1p0_zve64x1p0_zvl128b1p0_zvl32b1p0_zvl64b1p0: Invalid or unknown z ISA extension: 'zve' occurs, and there are no errors with the updated toolchain.

Instead of vwredsum I'll use vwredsum_vs_i16m1_i32m1.