diff options
| author | Xiaolin Zhao <zhaoxiaolin@loongson.cn> | 2025-08-06 11:34:12 +0800 |
|---|---|---|
| committer | abner chenc <chenguoqi@loongson.cn> | 2026-01-28 17:01:37 -0800 |
| commit | 7f0f67195194cb07122315d5ab563eb617dbe21a (patch) | |
| tree | f8726a984eff7a4de1b88d6d309923f2c08972fc | |
| parent | 985b0b3fe26661c10a3201470e80685765656363 (diff) | |
| download | go-7f0f67195194cb07122315d5ab563eb617dbe21a.tar.xz | |
math: optimize the floating-point pipeline on loong64
Using the FSEL instruction on loong64 to eliminate branches and reduce
pipeline interruptions.
On the Loongson CPU 3A6000, there is a 0.09% performance improvement, as follows:
goos: linux
goarch: loong64
pkg: math/big
cpu: Loongson-3A6000-HV @ 2500.00MHz
│ old.bench │ new.bench │
│ sec/op │ sec/op vs base │
Exp 7.748m ± 0% 7.740m ± 0% -0.10% (p=0.001 n=10)
Exp2 7.747m ± 0% 7.741m ± 0% -0.09% (p=0.002 n=10)
geomean 7.747m 7.740m -0.09%
Change-Id: If62f2e81bf345c83a1fa9350ace131240cfa3b9b
Reviewed-on: https://go-review.googlesource.com/c/go/+/693458
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: abner chenc <chenguoqi@loongson.cn>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Meidan Li <limeidan@loongson.cn>
| -rw-r--r-- | src/math/exp_loong64.s | 16 |
1 files changed, 4 insertions, 12 deletions
diff --git a/src/math/exp_loong64.s b/src/math/exp_loong64.s index bf2823f888..8420a2254e 100644 --- a/src/math/exp_loong64.s +++ b/src/math/exp_loong64.s @@ -62,13 +62,9 @@ TEXT ·archExp(SB),$0-16 MOVD 8(R10), F3 MOVD 48(R10), F2 CMPGTD F0, F5, FCC0 - BFPT add // x > 0 -sub: - FMSUBD F3, F2, F0, F3 // Log2e*x - 0.5 - JMP 2(PC) -add: + FMSUBD F3, F2, F0, F4 // Log2e*x - 0.5 FMADDD F3, F2, F0, F3 // Log2e*x + 0.5 - + FSEL FCC0, F3, F4, F3 FTINTRZVD F3, F4 // float64 -> int64 MOVV F4, R5 // R5 = int(k) FFINTDV F4, F3 // int64 -> float64 @@ -162,13 +158,9 @@ TEXT ·archExp2(SB),$0-16 MOVD 0(R10), F10 MOVD 8(R10), F2 CMPGTD F0, F10, FCC0 - BFPT add -sub: - SUBD F2, F0, F3 // x - 0.5 - JMP 2(PC) -add: + SUBD F2, F0, F4 // x - 0.5 ADDD F2, F0, F3 // x + 0.5 - + FSEL FCC0, F3, F4, F3 FTINTRZVD F3, F4 MOVV F4, R5 FFINTDV F4, F3 |