Implement f{32,64}.nearest instructions #486
Conversation
chfast
commented
Aug 13, 2020
•
chfast
commented
chfast
force-pushed
the
fp_ceil_floor_trunc
branch
6 times, most recently
from
0c96df1
to
91fa494
Compare
chfast
force-pushed
the
fp_ceil_floor_trunc
branch
2 times, most recently
from
e098abc
to
e4c2652
Compare
Codecov Report
@@ Coverage Diff @@
## master #486 +/- ##
==========================================
- Coverage 99.68% 99.67% -0.01%
==========================================
Files 54 54
Lines 17218 17108 -110
==========================================
- Hits 17163 17053 -110
Misses 55 55 |
chfast
force-pushed
the
fp_ceil_floor_trunc
branch
from
e4c2652
to
1fcc1db
Compare
lib/fizzy/execute.cpp
Outdated
| @@ -551,6 +552,18 @@ inline T ftrunc(T value) noexcept | |||
| return std::trunc(value); | |||
| } | |||
|
|
|||
| template <typename T> | |||
| T fnearest(T a) noexcept | |||
There was a problem hiding this comment.
This doesn't seem to enforce the same NaN rules as #481, yet the spec seems to be the same regarding it?
There was a problem hiding this comment.
We have them tested. I was expecting this will fail and I will need if (isnan()) case, but seems not required. Will check against the C spec though.
There was a problem hiding this comment.
This is the same as in other related functions: for NaN input it returns unspecified NaN.
There was a problem hiding this comment.
Are we making it deterministic for our case?
There was a problem hiding this comment.
Yes, we always return canonical NaN now as in other similar functions.
I also noticed that at least glibc is just setting the "quiet" bit in NaN to 1, leaving all others (including the sign) unchanged.
| @@ -829,223 +829,6 @@ TEST(execute, memory_copy_32bytes) | |||
| EXPECT_EQ(output, input); | |||
| } | |||
|
|
|||
| TEST(execute, fp_instructions) | |||
| auto instance = instantiate(parse(this->get_unop_code(Instr::f32_nearest))); | ||
| const auto exec = [&](auto arg) { return execute(*instance, 0, {arg}); }; | ||
|
|
||
| for (const auto& [arg, expected_trunc] : this->positive_trunc_tests) |
There was a problem hiding this comment.
Note to self:
fnearest(+-inf) = +-inf
fnearest(+-0) = +-0
are included in positive_trunc_tests.
There was a problem hiding this comment.
Can you leave this here as a comment?
There was a problem hiding this comment.
Please add this comment for clarity.
chfast
force-pushed
the
fp_ceil_floor_trunc
branch
5 times, most recently
from
08def8c
to
b3282b4
Compare
chfast
force-pushed
the
fp_nearest
branch
4 times, most recently
from
778c3f2
to
4c85f0f
Compare
chfast
force-pushed
the
fp_nearest
branch
2 times, most recently
from
4c862ed
to
bc2bfea
Compare
|
This is a working version with We may spend another week to:
|
The test is to detect unimplemented floating-point instructions, but all are implemented now.
