dom/cl-blas
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Compare commits

base: dom:7c46a4e35a9af06c5ab69d39286fd5ff3fb7dc56
Branches Tags
dom:main
..
compare: dom:e4e449f56254bc271d4c321a7735653d5e8d05c4
Branches Tags
dom:main

No commits in common. "7c46a4e35a9af06c5ab69d39286fd5ff3fb7dc56" and "e4e449f56254bc271d4c321a7735653d5e8d05c4" have entirely different histories.
7c46a4e35a ... e4e449f562

1 changed files with 5 additions and 7 deletions
Show stats Expand all files Collapse all files

12
src/cl-blas.lisp
View file

@ -90,14 +90,12 @@
(declaim (inline simd-abs-f32.8 simd-abs-f64.4))
(defun simd-abs-f32.8 (pack)
(declare (f32.8 pack)
(optimize . #.*optimize-qualities*))
(f32.8-if (f32.8< pack (f32.8 0.0)) (f32.8* pack (f32.8 -1.0)) pack))
(declare (f32.8 pack))
(f32.8-sqrt (f32.8* pack pack)))
(defun simd-abs-f64.4 (pack)
(declare (f64.4 pack)
(optimize . #.*optimize-qualities*))
(f64.4-if (f64.4< pack 0) (f64.4* pack -1) pack))
(declare (f64.4 pack))
(f64.4-sqrt (f64.4* pack pack)))
(push '(simd-abs simd-abs-f32.8) *single-float-fn-generic*)
(push '(simd-abs simd-abs-f64.4) *double-float-fn-generic*)
@ -307,7 +305,7 @@
(defblas gemm-serial ((A 2) (B 2) (C 2) (alpha 0) (beta 0))
(loop with m fixnum = (array-dimension C 0)
with n fixnum = (array-dimension C 1)
with r-block fixnum = (* (floor (array-dimension A 1) stride) stride)
with r-block fixnum = (* (floor (array-dimension A 1) 8) 8)
with r fixnum = (array-dimension A 1)
with Bt of-type (simple-array float 2) = (transpose B)
for i fixnum from 0 below m