From: Kyrylo Tkachov <[email protected]>
The test builds a reference sum by adding the elements in order and
compares it against the vaddv reduction with an absolute tolerance of
1e-6. vaddvq_f32 expands to two faddp, i.e. a pairwise sum, so the two
sides associate differently and agree only to within a few ULP. At the
magnitudes in input_float32[] one float32 ULP is 6.1e-5 to 2.4e-4, tens
to hundreds of times larger than the tolerance, so the comparison only
ever passed when both sides happened to round the same way.
The test compiles with -ffast-math, and since r17-2478-g3adb33259541
("Improve BB vectorization of reductions", PR tree-optimization/126028)
the reference chain is reassociated too, which changes the rounding and
the test aborts in test_vaddvf32_float32x4_t. That commit removed the
forced-even-lane trim in vect_slp_check_for_roots and added the
non-matching-lane split retry in vect_build_slp_instance, so the four
input_float32[] lanes are now discovered as a group once the out_l[0]
lane is split off. Both association orders are legal here and both land
within 1 ULP of the exact sum.
This is the third time the test has broken this way. PR testsuite/101506
added the asm barrier on the reference accumulator for the same reason.
Compare relatively instead. Every reference value in the test is
non-zero, so a plain relative test is safe, and at 1e-6 relative the
float32 checks still have about a factor of ten of margin over the
worst-case three-ULP reassociation error for this data.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/vect-vaddv.c (EQUALF, EQUALD): Compare
relative to the reference value.
Signed-off-by: Kyrylo Tkachov <[email protected]>
---
gcc/testsuite/gcc.target/aarch64/vect-vaddv.c | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/gcc/testsuite/gcc.target/aarch64/vect-vaddv.c
b/gcc/testsuite/gcc.target/aarch64/vect-vaddv.c
index 3a12ae9706a..47c1f77a5bc 100644
--- a/gcc/testsuite/gcc.target/aarch64/vect-vaddv.c
+++ b/gcc/testsuite/gcc.target/aarch64/vect-vaddv.c
@@ -41,8 +41,11 @@ double input_float64[] = {0.1, -0.1, 0.4, 10.3,
7.9, -870.0, 10.4, 310.11,
0.0, -865.0, -2213.0, -1.5};
-#define EQUALF(a, b) (fabsf (a - b) < DELTA)
-#define EQUALD(a, b) (fabs (a - b) < DELTA)
+/* The reference sum and the vaddv reduction associate the addends
+ differently, so they agree only to within a few ULP. Compare with a
+ relative tolerance: every reference value here is non-zero. */
+#define EQUALF(a, b) (fabsf (a - b) <= DELTA * fabsf (b))
+#define EQUALD(a, b) (fabs (a - b) <= DELTA * fabs (b))
#define EQUALL(a, b) (a == b)
#define TEST(SUFFIX, Q, TYPE, LANES, FLOAT) \
--
2.50.1 (Apple Git-155)
