kolibrios-gitea/kernel/branches/Kolibri-acpi/drivers/mix_sse2.inc

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;                                                              ;;
;; Copyright (C) KolibriOS team 2004-2011. All rights reserved. ;;
;; Distributed under terms of the GNU General Public License    ;;
;;                                                              ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

if used mmx128_mix_2

align 4
mmx128_mix_2:
        prefetcht1 [eax+128]
        prefetcht1 [ebx+128]

        movaps  xmm0, [eax]
        movaps  xmm1, [eax+16]
        movaps  xmm2, [eax+32]
        movaps  xmm3, [eax+48]
        movaps  xmm4, [eax+64]
        movaps  xmm5, [eax+80]
        movaps  xmm6, [eax+96]
        movaps  xmm7, [eax+112]

        paddsw  xmm0, [ebx]
        movaps  [edi], xmm0
        paddsw  xmm1, [ebx+16]
        movaps  [edi+16], xmm1
        paddsw  xmm2, [ebx+32]
        movaps  [edi+32], xmm2
        paddsw  xmm3, [ebx+48]
        movaps  [edi+48], xmm3
        paddsw  xmm4, [ebx+64]
        movaps  [edi+64], xmm4
        paddsw  xmm5, [ebx+80]
        movaps  [edi+80], xmm5
        paddsw  xmm6, [ebx+96]
        movaps  [edi+96], xmm6
        paddsw  xmm7, [ebx+112]
        movaps  [edi+112], xmm7
        ret

align 4
mmx128_mix_3:
        prefetcht1 [eax+128]
        prefetcht1 [ebx+128]
        prefetcht1 [ecx+128]

        movaps  xmm0, [eax]
        movaps  xmm1, [eax+16]
        movaps  xmm2, [eax+32]
        movaps  xmm3, [eax+48]
        movaps  xmm4, [eax+64]
        movaps  xmm5, [eax+80]
        movaps  xmm6, [eax+96]
        movaps  xmm7, [eax+112]

        paddsw  xmm0, [ebx]
        paddsw  xmm1, [ebx+16]
        paddsw  xmm2, [ebx+32]
        paddsw  xmm3, [ebx+48]
        paddsw  xmm4, [ebx+64]
        paddsw  xmm5, [ebx+80]
        paddsw  xmm6, [ebx+96]
        paddsw  xmm7, [ebx+112]

        paddsw  xmm0, [ecx]
        movaps  [edi], xmm0
        paddsw  xmm1, [ecx+16]
        movaps  [edi+16], xmm1
        paddsw  xmm2, [ecx+32]
        movaps  [edi+32], xmm2
        paddsw  xmm3, [ecx+48]
        movaps  [edi+48], xmm3
        paddsw  xmm4, [ecx+64]
        movaps  [edi+64], xmm4
        paddsw  xmm5, [ecx+80]
        movaps  [edi+80], xmm5
        paddsw  xmm6, [ecx+96]
        movaps  [edi+96], xmm6
        paddsw  xmm7, [ecx+112]
        movaps  [edi+112], xmm7
        ret

align 4
mmx128_mix_4:
        prefetcht1 [eax+128]
        prefetcht1 [ebx+128]
        prefetcht1 [ecx+128]
        prefetcht1 [edx+128]

        movaps  xmm0, [eax]
        movaps  xmm2, [eax+16]
        movaps  xmm4, [eax+32]
        movaps  xmm6, [eax+48]
        movaps  xmm1, [ebx]
        movaps  xmm3, [ebx+16]
        movaps  xmm5, [ebx+32]
        movaps  xmm7, [ebx+48]

        paddsw  xmm0, [ecx]
        paddsw  xmm2, [ecx+16]
        paddsw  xmm4, [ecx+32]
        paddsw  xmm6, [ecx+48]
        paddsw  xmm1, [edx]
        paddsw  xmm3, [edx+16]
        paddsw  xmm5, [edx+32]
        paddsw  xmm7, [edx+48]

        paddsw  xmm0, xmm1
        movaps  [edi], xmm0
        paddsw  xmm2, xmm3
        movaps  [edi+16], xmm2
        paddsw  xmm4, xmm5
        movaps  [edi+32], xmm4
        paddsw  xmm6, xmm7
        movaps  [edi+48], xmm6

        movaps  xmm0, [eax+64]
        movaps  xmm2, [eax+80]
        movaps  xmm4, [eax+96]
        movaps  xmm6, [eax+112]

        movaps  xmm1, [ebx+64]
        movaps  xmm3, [ebx+80]
        movaps  xmm5, [ebx+96]
        movaps  xmm7, [ebx+112]
        paddsw  xmm0, [ecx+64]
        paddsw  xmm2, [ecx+80]
        paddsw  xmm4, [ecx+96]
        paddsw  xmm6, [ecx+112]

        paddsw  xmm1, [edx+64]
        paddsw  xmm3, [edx+80]
        paddsw  xmm5, [edx+96]
        paddsw  xmm7, [edx+112]
        paddsw  xmm0, xmm1
        movaps  [edi+64], xmm0
        paddsw  xmm2, xmm3
        movaps  [edi+80], xmm2
        paddsw  xmm4, xmm5
        movaps  [edi+96], xmm4
        paddsw  xmm6, xmm7
        movaps  [edi+112], xmm6
        ret
end if