A demo of my WIP SIMD library: https://github.com/marcoheisig/sb-simd

simd-sum.lisp

(defun simd-sum (array &aux (n (array-total-size array)))
  "Compute the sum of the elements of the supplied simple double-float ARRAY."
  (declare (type (simple-array double-float 1) array)
           (optimize speed (safety 0)))
  (do ((index 0 (the (integer 0 #.(- array-total-size-limit 16)) (+ index 16)))
       (acc1 (make-f64.4 0 0 0 0) (f64.4+ acc1 (f64.4-row-major-aref array (+ index 0))))
       (acc2 (make-f64.4 0 0 0 0) (f64.4+ acc2 (f64.4-row-major-aref array (+ index 4))))
       (acc3 (make-f64.4 0 0 0 0) (f64.4+ acc3 (f64.4-row-major-aref array (+ index 8))))
       (acc4 (make-f64.4 0 0 0 0) (f64.4+ acc4 (f64.4-row-major-aref array (+ index 12)))))
      ((>= index (- n 16))
       (do ((result (multiple-value-call #'+ (f64.4-values (f64.4+ acc1 acc2 acc3 acc4)))
                    (+ result (row-major-aref array index)))
            (index index (1+ index)))
           ((>= index n) result)))))

; disassembly for simd-sum
; Size: 272 bytes. Origin: #x52B83286                         ; simd-sum
; 286:       488B4AF9         mov RCX, [RDX-7]
; 28A:       31C0             xor EAX, EAX
; 28C:       C5D457ED         vxorps YMM5, YMM5, YMM5
; 290:       C5F457C9         vxorps YMM1, YMM1, YMM1
; 294:       C5EC57D2         vxorps YMM2, YMM2, YMM2
; 298:       C5E457DB         vxorps YMM3, YMM3, YMM3
; 29C:       EB3D             jmp L1
; 29E:       6690             nop
; 2A0: L0:   488D7020         lea RSI, [RAX+32]
; 2A4:       C5FD10648201     vmovupd YMM4, [RDX+RAX*4+1]
; 2AA:       C5D558EC         vaddpd YMM5, YMM5, YMM4
; 2AE:       488D5808         lea RBX, [RAX+8]
; 2B2:       C5FD10449A01     vmovupd YMM0, [RDX+RBX*4+1]
; 2B8:       C5F558C8         vaddpd YMM1, YMM1, YMM0
; 2BC:       488D5810         lea RBX, [RAX+16]
; 2C0:       C5FD10449A01     vmovupd YMM0, [RDX+RBX*4+1]
; 2C6:       C5ED58D0         vaddpd YMM2, YMM2, YMM0
; 2CA:       4883C018         add RAX, 24
; 2CE:       C5FD10448201     vmovupd YMM0, [RDX+RAX*4+1]
; 2D4:       C5E558D8         vaddpd YMM3, YMM3, YMM0
; 2D8:       488BC6           mov RAX, RSI
; 2DB: L1:   488D59E0         lea RBX, [RCX-32]
; 2DF:       4839D8           cmp RAX, RBX
; 2E2:       7CBC             jl L0
; 2E4:       C5D558E9         vaddpd YMM5, YMM5, YMM1
; 2E8:       C5D558EA         vaddpd YMM5, YMM5, YMM2
; 2EC:       C5D558EB         vaddpd YMM5, YMM5, YMM3
; 2F0:       C5FC28C5         vmovaps YMM0, YMM5
; 2F4:       C5F157C9         vxorpd XMM1, XMM1, XMM1
; 2F8:       F20F10C8         movsd XMM1, XMM0
; 2FC:       C5FC28C5         vmovaps YMM0, YMM5
; 300:       C5FD73D808       vpsrldq YMM0, YMM0, 8
; 305:       C5E157DB         vxorpd XMM3, XMM3, XMM3
; 309:       F20F10D8         movsd XMM3, XMM0
; 30D:       C4E37D19E801     vextractf128 YMM0, YMM5, 1
; 313:       C5D957E4         vxorpd XMM4, XMM4, XMM4
; 317:       F20F10E0         movsd XMM4, XMM0
; 31B:       C4E37D19E801     vextractf128 YMM0, YMM5, 1
; 321:       C5FD73D808       vpsrldq YMM0, YMM0, 8
; 326:       C5D157ED         vxorpd XMM5, XMM5, XMM5
; 32A:       F20F10E8         movsd XMM5, XMM0
; 32E:       F20F58D9         addsd XMM3, XMM1
; 332:       F20F58E3         addsd XMM4, XMM3
; 336:       F20F58EC         addsd XMM5, XMM4
; 33A:       EB1A             jmp L3
; 33C:       0F1F4000         nop
; 340: L2:   F20F10548201     movsd XMM2, [RDX+RAX*4+1]
; 346:       660F28CD         movapd XMM1, XMM5
; 34A:       F20F58CA         addsd XMM1, XMM2
; 34E:       4883C002         add RAX, 2
; 352:       660F28E9         movapd XMM5, XMM1
; 356: L3:   4839C8           cmp RAX, RCX
; 359:       7CE5             jl L2
; 35B:       49896D28         mov [R13+40], RBP               ; thread.pseudo-atomic-bits
; 35F:       498B5568         mov RDX, [R13+104]              ; thread.alloc-region
; 363:       4C8D5A10         lea R11, [RDX+16]
; 367:       4D3B5D70         cmp R11, [R13+112]
; 36B:       771F             jnbe L6
; 36D:       4D895D68         mov [R13+104], R11              ; thread.alloc-region
; 371: L4:   66C7021D01       mov word ptr [RDX], 285
; 376:       80CA0F           or DL, 15
; 379:       49316D28         xor [R13+40], RBP               ; thread.pseudo-atomic-bits
; 37D:       7402             jeq L5
; 37F:       CC09             int3 9                          ; pending interrupt trap
; 381: L5:   F20F116AF9       movsd [RDX-7], XMM5
; 386:       488BE5           mov RSP, RBP
; 389:       F8               clc
; 38A:       5D               pop RBP
; 38B:       C3               ret
; 38C: L6:   6A10             push 16
; 38E:       E86DCE57FF       call #x52100200                 ; ALLOC-TRAMP
; 393:       5A               pop RDX
; 394:       EBDB             jmp L4