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

Add gemv

Browse source
This commit is contained in:
dominik martinez 2025-05-17 19:15:06 -07:00
parent db60c8841d
commit 29d35b333b
1 changed files with 19 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

19
src/cl-blas.lisp
View file

@ -222,6 +222,25 @@
do (setf m (max m (abs (aref x i))))) do (setf m (max m (abs (aref x i)))))
m)) m))
;;;; Level 2 BLAS: matrix-vector, O(n^2) operations
;;; gemv
(defblas gemv ((A 2) (x 1) (y 1) (alpha 0) (beta 0))
(let* ((n (array-dimension A 1))
(n-block (* (floor n stride) stride))
(alpha-vec (simd alpha)))
(declare (fixnum n n-block)
(simd alpha-vec))
(loop for i fixnum from 0 below (array-dimension A 0)
do (setf (aref y i)
(+ (* beta (aref y i))
(loop for acc of-type simd = (simd 0) then (simd+ acc (simd* alpha-vec (simd-aref A i j) (aref x j)))
for j fixnum from 0 below n-block by stride
finally (return (the float (simd-horizontal+ acc))))
(loop for j fixnum from n-block below n
sum (* alpha (aref A i j) (aref x j)) of-type float))))))
;;;; Level 3 BLAS: matrix-matrix, O(n^3) operations ;;;; Level 3 BLAS: matrix-matrix, O(n^3) operations
(defblas transpose ((A 2)) (defblas transpose ((A 2))