Yuri Victorovich
19 lines
698 B
C++
19 lines
698 B
C++
--- test/taylor_atan.cpp.orig
|
|
+++ test/taylor_atan.cpp
|
|
@@ -48,7 +48,16 @@
|
|
>{};
|
|
|
|
-constexpr bool skip_batch_ld = false;
|
|
+// On x86/x86-64, long double is 80-bit extended precision. Batch Taylor
|
|
+// integration uses double-length (DL) arithmetic via EFT internally, which
|
|
+// relies on fmal() being an exact fused multiply-add. Since there is no
|
|
+// hardware FMA instruction for 80-bit extended precision on x86, the system
|
|
+// fmal() is not exact, making batch results unreliable for long double.
|
|
+#if defined(HEYOKA_ARCH_X86)
|
|
+constexpr bool skip_batch_ld = true;
|
|
+#else
|
|
+constexpr bool skip_batch_ld = false;
|
|
+#endif
|
|
|
|
TEST_CASE("taylor atan decompose bug 00")
|