phynexis-deps/eigen
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fp16/bf16: fix NumTraits<half>::epsilon, half GPU raw-bit interop, RVV10 bf16 masks, and 16-bit normal RNG

Browse Source 
libeigen/eigen!2704

Co-authored-by: Rasmus Munk Larsen <rmlarsen@gmail.com>
This commit is contained in:
Rasmus Munk Larsen
2026-07-10 14:55:30 -07:00
co-authored by Rasmus Munk Larsen
parent c453a87216
commit 40e7e38646
12 changed files with 311 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/half_float.cpp
+8
View File
@@ -127,6 +127,14 @@ void test_numtraits() {
VERIFY(NumTraits<half>::IsSigned);
// The fp16 machine epsilon is 2^-10 = 0x1400, and NumTraits must agree with
// std::numeric_limits (NumTraits used to return 2^-13, making every
// epsilon-based threshold 8x too tight).
VERIFY_HALF_BITS_EQUAL(NumTraits<half>::epsilon(), 0x1400);
VERIFY_HALF_BITS_EQUAL(std::numeric_limits<half>::epsilon(), 0x1400);
VERIFY_IS_EQUAL(numext::bit_cast<numext::uint16_t>(NumTraits<half>::epsilon()),
numext::bit_cast<numext::uint16_t>(std::numeric_limits<half>::epsilon()));
VERIFY_IS_EQUAL(numext::bit_cast<numext::uint16_t>(std::numeric_limits<half>::infinity()),
numext::bit_cast<numext::uint16_t>(half(std::numeric_limits<float>::infinity())));
// There is no guarantee that casting a 32-bit NaN to 16-bit has a precise