Test results of QMKL (https://github.com/Terminus-IMRC/qmkl)

rpi2.md

sgemm

P = 96
Q = 363
R = 3072
ALPHA = 1.000000
BETA = 1.000000
==== sgemm example (ALPHA * 96x363 * 363x3072 + BETA * 96x3072) ====
GPU: 0.02727 [s], 7.88379e+09 [flop/s]
CPU (4 threads): 4.8888 [s], 4.39762e+07 [flop/s]
Minimum absolute error: 0
Maximum absolute error: 0.000892639
Minimum relative error: 0
Maximum relative error: 1.44661
CPU with NEON (4 threads): 0.458766 [s], 4.68629e+08 [flop/s]
Minimum absolute error: 0
Maximum absolute error: 2.28882e-05
Minimum relative error: 0
Maximum relative error: 0.118797

rpi3.md

sgemm

P = 96
Q = 363
R = 3072
ALPHA = 1.000000
BETA = 1.000000
==== sgemm example (ALPHA * 96x363 * 363x3072 + BETA * 96x3072) ====
GPU: 0.022868 [s], 9.40138e+09 [flop/s]
CPU (4 threads): 5.20537 [s], 4.13017e+07 [flop/s]
Minimum absolute error: 0
Maximum absolute error: 0.000938416
Minimum relative error: 0
Maximum relative error: 0.311594
CPU with NEON (4 threads): 0.547971 [s], 3.9234e+08 [flop/s]
Minimum absolute error: 0
Maximum absolute error: 2.67029e-05
Minimum relative error: 0
Maximum relative error: 0.0215263

run-tests.sh

#!/usr/bin/env bash

BUILDDIR="$1"

if ! [ -d "$BUILDDIR" ]; then
	echo "No such directory: $BUILDDIR"
	exit 1
fi

set -e

run_test() {
	echo
	echo "## $1"
	echo
	echo '```'
	sudo "./$1"
	echo '```'
}

cd "$BUILDDIR/test"
run_test sgemm
run_test scopy
run_test vsAbs

I knew the service because piwheels uses that, but I didn't know it provides "dedicated" Raspberry Pi host, and it supports power cycling over its control panel. Thanks again!