RVV Bench on ARA #17
Comments
|
Yes, I tried about half a year ago, but ara still had some bugs, which broke most of the benchmarks. |
|
@camel-cdr I do have a docker container for the ARA. Here is the repo for it. https://github.com/rseac/pulp-ara-docker |
|
@rseac Thanks a lot! I'll try to use it. |
|
@rseac You docker worked great, ara unfortunately didn't. I tried running my instruction throughout benchmark, but about 5% of instructions hang the simulation with certain valid vtype configurations. The most problematic examples were the integer comparisons (vmseq,...), except for vmsgtu/vmsgt, which hung with SEW=16 and logical mask instructions. Otheres are vrgather.vx, vcompress.vm, vredsum.vs, vmadc, zext vrgather.vv works, but is extremely slow at 6-8 cycles per element. It's still measuring the last handful of instructions, but if you want, I can share the results of the instructions that worked once it completed. |
|
@camel-cdr Yes, I'd be happy to take a look at the results. I could even try to post some of these problems you notice on the ARA issues once I understand them myself. |
|
@camel-cdr Did you use the same scripts as linked earlier? Or specific ones to target the ARA core? |
|
@rseac Sorry for the late response. Here are the measurements I managed to run: log.txt I used your buildscript with small modifications: FROM ubuntu:22.04
ENV DEBIAN_FRONTEND=noninteractive
RUN apt -y update && apt -y upgrade && apt-get --no-install-recommends -y install \
build-essential git zlib1g zlib1g-dev pkg-config cmake vim \
ninja-build python3 texinfo device-tree-compiler \
autoconf automake bc bison clang flex \
ca-certificates ccache libfl2 libfl-dev help2man \
curl libelf-dev python3-numpy \
&& apt-get clean \
&& rm -rf /var/lib/apt/lists/*
RUN git clone https://github.com/pulp-platform/ara.git
WORKDIR /ara
RUN git config --global url."https://github.com/".insteadOf "[email protected]:";
RUN git submodule update --init --recursive
RUN make toolchain-llvm
RUN make riscv-isa-sim
RUN make verilator
RUN /usr/bin/install -c /ara/install/verilator/bin/verilator_bin /ara/install/verilator/share/verilator/
RUN cd hardware && make checkout && make apply-patches && make verilate
WORKDIR /ara/apps
RUN git clone https://github.com/camel-cdr/rvv-bench \
&& cp rvv-bench/nolibc.h . \
&& mkdir rvv \
&& cp rvv-bench/instructions/rvv/gen.S . \
&& cp rvv-bench/instructions/rvv/config.h rvv-bench/instructions/rvv/main.c rvv \
&& sed -e '2a#define CUSTOM_HOST 1' -e '2a#include "printf.h"' -e '2a#include <string.h>' -i nolibc.h \
&& sed 's/main/nolibc_main/g;s/_start/main/g;s/nolibc_main();/\0\n#define main nolibc_main/g' -i nolibc.h \
&& sed 's/\(memwrite(.*\)}/\1printf("%.*s",len,ptr);}/g' -i nolibc.h \
&& sed 's/WARMUP.*$/WARMUP 1/g;s/UNROLL.*$/UNROLL 4/g;s/LOOP.*$/LOOP 8/1;s/RUNS.*$/RUNS 1/g' -i rvv/config.h \
&& sed 's/\.\.\/nolibc/nolibc/g' -i rvv/main.c
RUN echo 'echo "vim gen.S && m4 gen.S > rvv/gen.S && ( make clean; make bin/rvv ) 2>&1 >/dev/null && app=rvv make -C /ara/hardware simv"' > ~/.bashrcYou can just execute the command printed once you run the container. The following were the ones that failed to execute correctly: m_mask($1, bench_vrgathervx, T_A, m_mod_t0_vl, vrgather.vx, v8, v16, t0)
m_mask($1, bench_vrgathervi, T_A, m_nop, vrgather.vi, v8, v16, 3)
m_mask($1, bench_vredsumvs, T_A, m_nop, vredsum.vs, v8, v16, v24)
m_mask($1, bench_vredandvs, T_A, m_nop, vredand.vs, v8, v16, v24)
m_mask($1, bench_vredorvs, T_A, m_nop, vredor.vs, v8, v16, v24)
m_mask($1, bench_vredxorvs, T_A, m_nop, vredxor.vs, v8, v16, v24)
m_mask($1, bench_vredminuvs, T_A, m_nop, vredminu.vs, v8, v16, v24)
m_mask($1, bench_vredminvs, T_A, m_nop, vredmin.vs, v8, v16, v24)
m_mask($1, bench_vredmaxuvs, T_A, m_nop, vredmaxu.vs, v8, v16, v24)
m_mask($1, bench_vredmaxvs, T_A, m_nop, vredmax.vs, v8, v16, v24)
m_bench_vxim($1, T_A, vmadc)
m_bench_vxm($1, T_A, vmsbc)
m_bench_vxi($1, T_A, vmseq)
m_bench_vxi($1, T_A, vmsne)
m_bench_vx($1, T_A, vmsltu)
m_bench_vx($1, T_A, vmslt)
m_bench_vxi($1, T_A, vmsleu)
m_bench_vxi($1, T_A, vmsle)
m_$1(bench_vcompressvm, T_A, m_nop, vcompress.vm, v8, v16, v24)
m_$1(bench_vmandnmm, T_m1, m_nop, vmandn.mm, v8, v16, v24)
m_$1(bench_vmandmm, T_m1, m_nop, vmand.mm, v8, v16, v24)
m_$1(bench_vmormm, T_m1, m_nop, vmor.mm, v8, v16, v24)
m_$1(bench_vmxormm, T_m1, m_nop, vmxor.mm, v8, v16, v24)
m_$1(bench_vmornmm, T_m1, m_nop, vmorn.mm, v8, v16, v24)
m_$1(bench_vmnandmm, T_m1, m_nop, vmnand.mm, v8, v16, v24)
m_$1(bench_vmnormm, T_m1, m_nop, vmnor.mm, v8, v16, v24)
m_$1(bench_vmxnormm, T_m1, m_nop, vmxnor.mm, v8, v16, v24)
m_mask($1, bench_vfredosumvs, T_F, m_nop, vfredosum.vs, v8, v16, v24)
m_mask($1, bench_vwredsumuvs, T_WR, m_nop, vwredsumu.vs, v8, v16, v24)
m_mask($1, bench_vwredsumvs, T_WR, m_nop, vwredsum.vs, v8, v16, v24)
m_mask($1, bench_vfwredosumvs, T_FWR, m_nop, vfwredosum.vs, v8, v16, v24)
m_mask($1, bench_vfwredusumvs, T_FWR, m_nop, vfwredusum.vs, v8, v16, v24)
m_mask($1, bench_vfirstm, T_m1, m_1bit, vfirst.m, t0, v8)
m_mask($1, bench_vzextvf2, T_E2, m_1bit, vzext.vf2, v8, v16)
m_mask($1, bench_vsextvf2, T_E2, m_1bit, vsext.vf2, v8, v16)
m_mask($1, bench_vzextvf4, T_E4, m_1bit, vzext.vf4, v8, v16)
m_mask($1, bench_vsextvf4, T_E4, m_1bit, vsext.vf4, v8, v16)
m_mask($1, bench_vzextvf8, T_E8, m_1bit, vzext.vf8, v8, v16)
m_mask($1, bench_vsextvf8, T_E8, m_1bit, vsext.vf8, v8, v16) |
|
Hey @camel-cdr, @rseac, I should have fixed many of the old bugs you initially reported and open-sourced the missing RVV instruction support. I am almost done with basic Linux support and my next priority is verification. I will start from these instructions, thanks for reporting! |
|
@mp-17 Thanks, thats great to hear, I closed the old issue. Is the vrgather performance I measured expected and/or how does the current implementation work/is supposed to work? I saw the new AraXL paper, is that a fork of Ara or further development? |
|
@camel-cdr Thanks for providing this. These are single instruction tests, is that correct? I'm assuming that you'd want these single instruction tests to pass before trying the rev-bench benchmarks themselves? Do any of the benchmarks in bench/ go through? Or has that not been tried yet. |
|
@rseac Yes, I prefer to get the instructions themselves working. I tried a few of the benchmarks, here are results from a 4-lane configuration, for the things that worked: It's interesting to see that alignment matters a lot. |
Is there interest in trying to get these benchmarks work on the ARA vector engine? Has this already been started?
The text was updated successfully, but these errors were encountered: