cblas_zdotc_sub returns incorrect results on ppc64le

[susilehtola]

Bug reported on Fedora: https://bugzilla.redhat.com/show_bug.cgi?id=1878449

cblas_zdotc_sub returns incorrect results on ppc64le on Fedora 32 and above (all with GCC 10.2.1). The correct result, however, is obtained on Fedora 31 (GCC 9.3.1).

The test is the following: since zdotc_sub(x,y) = conjugate(x) . y, one should have zdotc_sub(y,x) = conjugate(zdotc_sub(x,y)) but this property is violated by the numerics.

Test program from original report below

#include <stdio.h>
#include <cblas.h>

int
main(void)
{
	const double x[16] = {
		-0.4759506499688366,	+0.4656291053809647,
		-0.6826320556910688,	-0.7694515466337011,
		-0.4437469610112805,	-0.2274498627312811,
		-0.0813662255708667,	+0.25700235907942326,
		-0.35799891895966574,	-0.7498841473288012,
		+0.036785641195074215,	+0.9670972102872823,
		-0.4761141488697107,	-0.11355026270974378,
		+0.9521705697548672,	+0.5791166840478064,
	};
	const double y[16] = {
		+0.5882371515831732,	-7.0347725597508237e-04,
		-0.27747685516513854,	-6.4286264567294582e-01,
		-0.167792120858004,	-1.7371739715068224e-01,
		+0.1685162562206428,	-6.0160952958980318e-01,
		+0.5203435476117149,	+6.3398807174073424e-02,
		-0.6243831891471687,	+6.6474140613596511e-01,
		-0.42366570437592865,	-6.2949810985591026e-01,
		+0.34043771632046815,	+9.1471844734252139e-01,
	};
	double ret[2];
	cblas_zdotc_sub(8, x, 1, y, 1, ret);
	printf("Result: %.15f%+.15fj\n", ret[0], ret[1]);
	cblas_zdotc_sub(8, y, 1, x, 1, ret);
	printf("Result: %.15f%+.15fj\n", ret[0], ret[1]);
	return 0;
}

[martin-frbg]

Seems to be an optimizer bug that comes and goes... I think I saw this with gcc 10.1 but the patch (pragma GCC optimize "O1") seemed to be unnecessary with gcc 10.2 by the time my openpower account was reactivated. I'll take another look.

[martin-frbg]

Not reproducible with current develop and gcc 8.4.0, gcc 9.3.0 or gcc 10.2.0 on Debian11 ppc64le. (Also not the problem it first reminded me of)

[martin-frbg]

Not reproduced with develop, but reproduced with 0.3.10 on Fedora32 - though the only relevant change would seem to be my replacing the poorly portable __real, __imag initializations with OPENBLAS_MAKE_COMPLEX_FLOAT

[martin-frbg]

Definitely the switch to OPENBLAS_MAKE_COMPLEX_FLOAT in the PPC zdot.c that fixed this

[susilehtola]

You mean 661c6bf and 21072e5?

[martin-frbg]

Yes (where the exclusion of the microkernel has nothing to do with this particular problem)