run-all.sh

#!/bin/bash
set -e
# BW1000
export HIP_VISIBLE_DEVICES=1
BIND_CMD="numactl -N 1 -m 1"

# BW150
# export HIP_VISIBLE_DEVICES=1
# BIND_CMD="numactl -N 0 -m 0"

make clean
CXX=hipcc make GPU_ARCH=gfx936
# CXX=nvcc make GPU_ARCH=sm_80

W1="--verify 1 -M 11264 -K 4096  --alpha 1 --beta 0 -B 128"
W2="--verify 1 -M 4096  -K 11264 --alpha 1 --beta 0 -B 128"
W3="--verify 1 -M 12288 -K 4096  --alpha 1 --beta 0 -B 128"
W4="--verify 1 -M 4096  -K 4096  --alpha 1 --beta 1 -B 128"

if [[ "$*" == *"--pmc"* ]]; then
    PROF_CMD="hipprof --trace-off --pmc"
    ${PROF_CMD} -o log/pmc-w1 ${BIND_CMD} ./gemv_bench --warmups 10 --loops 20 ${W1}
    ${PROF_CMD} -o log/pmc-w2 ${BIND_CMD} ./gemv_bench --warmups 10 --loops 20 ${W2}
    ${PROF_CMD} -o log/pmc-w3 ${BIND_CMD} ./gemv_bench --warmups 10 --loops 20 ${W3}
    ${PROF_CMD} -o log/pmc-w4 ${BIND_CMD} ./gemv_bench --warmups 10 --loops 20 ${W4}
elif [[ "$*" == *"--trace"* ]]; then
    PROF_CMD="hipprof --hip-trace"
    ${PROF_CMD} -o log/trace-w1 ${BIND_CMD} ./gemv_bench --warmups 100 --loops 1000 ${W1}
    ${PROF_CMD} -o log/trace-w2 ${BIND_CMD} ./gemv_bench --warmups 100 --loops 1000 ${W2}
    ${PROF_CMD} -o log/trace-w3 ${BIND_CMD} ./gemv_bench --warmups 100 --loops 1000 ${W3}
    ${PROF_CMD} -o log/trace-w4 ${BIND_CMD} ./gemv_bench --warmups 100 --loops 1000 ${W4}
else
    ${BIND_CMD} ./gemv_bench --warmups 100 --loops 2000 ${W1}
    ${BIND_CMD} ./gemv_bench --warmups 100 --loops 2000 ${W2}
    ${BIND_CMD} ./gemv_bench --warmups 100 --loops 2000 ${W3}
    ${BIND_CMD} ./gemv_bench --warmups 100 --loops 2000 ${W4}
fi