Hello,
I found that the test case failure is not due to incorrect
behavior but rather seems like it is because of platform-specific
floating-point precision differences on ppc64le. The numerical
error is marginal and well within acceptable limits for
floating-point operations involving matrix decompositions and
pseudo-inverses.
This appears to be a case where
the test's absolute epsilon (1e-5)
is a bit strict for some platforms, especially when fallback to
the normal pseudo-inverse occurs (as the fast pseudo-inverse
fails). The same issue also appears to be for arm64.
To resolve this, I suggest for relaxing the epsilon slightly.
Following is the diff for that.
diff --git a/unittest/ztensor_cp_test.cc
b/unittest/ztensor_cp_test.cc
index e9f9a0f..4cd8446 100644
--- a/unittest/ztensor_cp_test.cc
+++ b/unittest/ztensor_cp_test.cc
@@ -26,6 +26,9 @@ TEST_CASE("ZCP") {
// double epsilon = fmax(1e-10,
std::numeric_limits<double>::epsilon());
double epsilon = 1e-5;
+ #if defined(__PPC64__) || defined(__powerpc64__)
+ epsilon = 2e-5;
+ #endif
ztensor Z3(3, 2, 4);
std::ifstream inp3(__dirname + "/z-mat3D.txt");
if (inp3.is_open()) {
debian@ltc-zz14-lp9:~/test-btas/btas-1.0.0/debian/patches$
Thanks and regards.