frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Caffe2] Improve SVE batch box cox by 2% (#163360)

Browse Source 
Summary:
Improve bound checking on exp computation, decreasing the longest dependency chain by 1.

Box-cox benchmarks show about 2% of improved throughput.
Precision remains unaltered.

before:

NonZeroLambdaBatch                                        155.30us     6.44K

after:

NonZeroLambdaBatch                                        151.78us     6.59K

Test Plan:
Correctness:

buck2 test @//mode/opt //koski/functions_contrib/df4ai/tests:batch_box_cox_test

Performance:

buck2 run @//mode/opt //koski/functions_contrib/df4ai/benchmark:boxcox_benchmark

Differential Revision:
D82847111

Privacy Context Container: L1208939

Pull Request resolved: https://github.com/pytorch/pytorch/pull/163360
Approved by: https://github.com/Skylion007
This commit is contained in:
Nicolas De Carli
2025-09-20 06:42:23 +00:00
committed by PyTorch MergeBot
parent 5050cfa363
commit eb11d172e3
1 changed files with 12 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
caffe2/perfkernels/batch_box_cox_sve128.cc
View File

@ -114,14 +114,20 @@ inline float32x4_t vexpq_f32(float32x4_t x) {
auto poly = svset_neonq(svundef_f32(), vfmaq_f32(scale, p12345, scale));
auto pHigh = svcmpgt_f32(svptrue_b8(), svset_neonq(svundef_f32(), x), max_input);
auto pLow = svcmplt_f32(svptrue_b8(), svset_neonq(svundef_f32(), x), min_input);
auto bound = svsel_f32(
pHigh,
inf,
zero);
auto pCombined = svorr_b_z(svptrue_b8(), pLow, pHigh);
// Handle underflow and overflow.
poly = svsel_f32(
svcmplt_f32(svptrue_b8(), svset_neonq(svundef_f32(), x), min_input),
zero,
poly);
poly = svsel_f32(
svcmpgt_f32(svptrue_b8(), svset_neonq(svundef_f32(), x), max_input),
inf,
pCombined,
bound,
poly);
return svget_neonq(poly);