cblas_isamin on a dataset containing 0.0 and a non-1 incX returns 7

[RossBrunton]

The following example shows the problem:

int a = 30;
std::vector<float> x(a * 5);
for (int i = 0; i < a * 5; i ++) {
  x[i] = i + 1;
}
x[10 * 5] = 0;
std::cout << cblas_isamin(a, x.data(), 5) << "\n";

This prints out 7, rather than 10 as I would expect. Changing the 10 to any value larger than 7 also prints 7. Changing incX to 1, or removing the x[10 * 5] = 0 line causes this to output 0 as expected.

I'm using commit 6a8b42 .

[martin-frbg]

Oh joy, this is probably a historical bug from GotoBLAS... I assume you are running this on x86_64 hardware ?

[RossBrunton]

This is x86_64, yes.

[martin-frbg]

Quick solution seems to be to change the ISAMINKERNEL entry in kernel/x86_64/KERNEL from iamax_sse.S to iamax.S - have not checked yet if related functions are similarly affected.

[martin-frbg]

Seems both isamin and ismin give wrong results for non-1 incX, although they use the same code as their (working) -max counterparts with only the comparison inverted.