Bug in index output of max and min on gpuArray

There is a bug in the index output of max and min for gpuArrays. I am using 2018b and the bug also appears in 2019b. The index of the max (the 2nd output) is wrong for various lengths of arrays. For example:

x = randn(8680, 2, 'gpuArray');
[m, k] = max(x, [], 1);
disp(k)
[mCpu, kCpu] = max(gather(x), [], 1);
disp(kCpu)
isequal(k, kCpu) % different values!!

x = randn(8681, 2, 'gpuArray');
[m, k] = max(x, [], 1);
disp(k)
[mCpu, kCpu] = max(gather(x), [], 1);
disp(kCpu)
isequal(k, kCpu) % nagically now the same value!

OK let's do a test to see which lengths of x faile

n = 1 : 10e3;
bad = [];
for i = 1 : length(n)
    x = randn(n(i), 2, 'single', 'gpuArray');
    [m1, k1] = max(x, [], 1);
    [m2, k2] = max(gather(x), [], 1);
    if ~isequal(k1, k2)
        bad(end+1) = n(i);
    end
end

figure
plot(bad) % there are 410 values between 1 and 10e3 that fail!!

Now let's try again with different 2nd dim lengths of x

n = 1 : 10e3;
bad = [];
for i = 1 : length(n)
    x = randn(n(i), 4, 'single', 'gpuArray');
    [m1, k1] = max(x, [], 1);
    [m2, k2] = max(gather(x), [], 1);
    if ~isequal(k1, k2)
        bad(end+1) = n(i);
    end
end

figure
plot(bad) % now 147 that fail.

Answer 1

[Srivardhan Gadila]

The issue has been reported to the concerned team.

The workaround would be to transpose the matrix and reducing along the columns:

[m1, k1] = max(x', [], 2);
m1 = m1';
k1 = k1';