|
Packit |
47f805 |
;
|
|
Packit |
47f805 |
; (C) Frank Klemm 1995,99,2000
|
|
Packit |
47f805 |
; Dedicated to the LAME project
|
|
Packit |
47f805 |
;
|
|
Packit |
47f805 |
;
|
|
Packit |
47f805 |
%include "nasm.h"
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
segment_code
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
; float_t scalar04_float32_i387 (
|
|
Packit |
47f805 |
; const float32_t* const p,
|
|
Packit |
47f805 |
; const float32_t* const q );
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar04_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar08_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 16]
|
|
Packit |
47f805 |
fmul dword [edx + 16]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 20]
|
|
Packit |
47f805 |
fmul dword [edx + 20]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 24]
|
|
Packit |
47f805 |
fmul dword [edx + 24]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 28]
|
|
Packit |
47f805 |
fmul dword [edx + 28]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar12_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 16]
|
|
Packit |
47f805 |
fmul dword [edx + 16]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 20]
|
|
Packit |
47f805 |
fmul dword [edx + 20]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 24]
|
|
Packit |
47f805 |
fmul dword [edx + 24]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 28]
|
|
Packit |
47f805 |
fmul dword [edx + 28]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 32]
|
|
Packit |
47f805 |
fmul dword [edx + 32]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 36]
|
|
Packit |
47f805 |
fmul dword [edx + 36]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 40]
|
|
Packit |
47f805 |
fmul dword [edx + 40]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 44]
|
|
Packit |
47f805 |
fmul dword [edx + 44]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar16_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 16]
|
|
Packit |
47f805 |
fmul dword [edx + 16]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 20]
|
|
Packit |
47f805 |
fmul dword [edx + 20]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 24]
|
|
Packit |
47f805 |
fmul dword [edx + 24]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 28]
|
|
Packit |
47f805 |
fmul dword [edx + 28]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 32]
|
|
Packit |
47f805 |
fmul dword [edx + 32]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 36]
|
|
Packit |
47f805 |
fmul dword [edx + 36]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 40]
|
|
Packit |
47f805 |
fmul dword [edx + 40]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 44]
|
|
Packit |
47f805 |
fmul dword [edx + 44]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 48]
|
|
Packit |
47f805 |
fmul dword [edx + 48]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 52]
|
|
Packit |
47f805 |
fmul dword [edx + 52]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 56]
|
|
Packit |
47f805 |
fmul dword [edx + 56]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 60]
|
|
Packit |
47f805 |
fmul dword [edx + 60]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar20_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 16]
|
|
Packit |
47f805 |
fmul dword [edx + 16]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 20]
|
|
Packit |
47f805 |
fmul dword [edx + 20]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 24]
|
|
Packit |
47f805 |
fmul dword [edx + 24]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 28]
|
|
Packit |
47f805 |
fmul dword [edx + 28]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 32]
|
|
Packit |
47f805 |
fmul dword [edx + 32]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 36]
|
|
Packit |
47f805 |
fmul dword [edx + 36]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 40]
|
|
Packit |
47f805 |
fmul dword [edx + 40]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 44]
|
|
Packit |
47f805 |
fmul dword [edx + 44]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 48]
|
|
Packit |
47f805 |
fmul dword [edx + 48]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 52]
|
|
Packit |
47f805 |
fmul dword [edx + 52]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 56]
|
|
Packit |
47f805 |
fmul dword [edx + 56]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 60]
|
|
Packit |
47f805 |
fmul dword [edx + 60]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 64]
|
|
Packit |
47f805 |
fmul dword [edx + 64]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 68]
|
|
Packit |
47f805 |
fmul dword [edx + 68]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 72]
|
|
Packit |
47f805 |
fmul dword [edx + 72]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 76]
|
|
Packit |
47f805 |
fmul dword [edx + 76]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar24_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 16]
|
|
Packit |
47f805 |
fmul dword [edx + 16]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 20]
|
|
Packit |
47f805 |
fmul dword [edx + 20]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 24]
|
|
Packit |
47f805 |
fmul dword [edx + 24]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 28]
|
|
Packit |
47f805 |
fmul dword [edx + 28]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 32]
|
|
Packit |
47f805 |
fmul dword [edx + 32]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 36]
|
|
Packit |
47f805 |
fmul dword [edx + 36]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 40]
|
|
Packit |
47f805 |
fmul dword [edx + 40]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 44]
|
|
Packit |
47f805 |
fmul dword [edx + 44]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 48]
|
|
Packit |
47f805 |
fmul dword [edx + 48]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 52]
|
|
Packit |
47f805 |
fmul dword [edx + 52]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 56]
|
|
Packit |
47f805 |
fmul dword [edx + 56]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 60]
|
|
Packit |
47f805 |
fmul dword [edx + 60]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 64]
|
|
Packit |
47f805 |
fmul dword [edx + 64]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 68]
|
|
Packit |
47f805 |
fmul dword [edx + 68]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 72]
|
|
Packit |
47f805 |
fmul dword [edx + 72]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 76]
|
|
Packit |
47f805 |
fmul dword [edx + 76]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 80]
|
|
Packit |
47f805 |
fmul dword [edx + 80]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 84]
|
|
Packit |
47f805 |
fmul dword [edx + 84]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 88]
|
|
Packit |
47f805 |
fmul dword [edx + 88]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 92]
|
|
Packit |
47f805 |
fmul dword [edx + 92]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar32_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 16]
|
|
Packit |
47f805 |
fmul dword [edx + 16]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 20]
|
|
Packit |
47f805 |
fmul dword [edx + 20]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 24]
|
|
Packit |
47f805 |
fmul dword [edx + 24]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 28]
|
|
Packit |
47f805 |
fmul dword [edx + 28]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 32]
|
|
Packit |
47f805 |
fmul dword [edx + 32]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 36]
|
|
Packit |
47f805 |
fmul dword [edx + 36]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 40]
|
|
Packit |
47f805 |
fmul dword [edx + 40]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 44]
|
|
Packit |
47f805 |
fmul dword [edx + 44]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 48]
|
|
Packit |
47f805 |
fmul dword [edx + 48]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 52]
|
|
Packit |
47f805 |
fmul dword [edx + 52]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 56]
|
|
Packit |
47f805 |
fmul dword [edx + 56]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 60]
|
|
Packit |
47f805 |
fmul dword [edx + 60]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 64]
|
|
Packit |
47f805 |
fmul dword [edx + 64]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 68]
|
|
Packit |
47f805 |
fmul dword [edx + 68]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 72]
|
|
Packit |
47f805 |
fmul dword [edx + 72]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 76]
|
|
Packit |
47f805 |
fmul dword [edx + 76]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 80]
|
|
Packit |
47f805 |
fmul dword [edx + 80]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 84]
|
|
Packit |
47f805 |
fmul dword [edx + 84]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 88]
|
|
Packit |
47f805 |
fmul dword [edx + 88]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 92]
|
|
Packit |
47f805 |
fmul dword [edx + 92]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 96]
|
|
Packit |
47f805 |
fmul dword [edx + 96]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax +100]
|
|
Packit |
47f805 |
fmul dword [edx +100]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax +104]
|
|
Packit |
47f805 |
fmul dword [edx +104]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax +108]
|
|
Packit |
47f805 |
fmul dword [edx +108]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax +112]
|
|
Packit |
47f805 |
fmul dword [edx +112]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax +116]
|
|
Packit |
47f805 |
fmul dword [edx +116]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax +120]
|
|
Packit |
47f805 |
fmul dword [edx +120]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax +124]
|
|
Packit |
47f805 |
fmul dword [edx +124]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
; float_t scalar4n_float32_i387 (
|
|
Packit |
47f805 |
; const float32_t* const p,
|
|
Packit |
47f805 |
; const float32_t* const q,
|
|
Packit |
47f805 |
; const size_t len );
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar4n_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
%$len arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
mov ecx,[sp(%$len)]
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
dec ecx
|
|
Packit |
47f805 |
jz .ret1
|
|
Packit |
47f805 |
add eax,byte 16
|
|
Packit |
47f805 |
add edx,byte 16
|
|
Packit |
47f805 |
.lbl1
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
add eax,byte 16
|
|
Packit |
47f805 |
add edx,byte 16
|
|
Packit |
47f805 |
dec ecx
|
|
Packit |
47f805 |
jnz .lbl1
|
|
Packit |
47f805 |
.ret1
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
; float_t scalar1n_float32_i387 (
|
|
Packit |
47f805 |
; const float32_t* const p,
|
|
Packit |
47f805 |
; const float32_t* const q,
|
|
Packit |
47f805 |
; const size_t len );
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar1n_float32_i387
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
%$len arg 4
|
|
Packit |
47f805 |
;;; alloc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
mov ecx,[sp(%$len)]
|
|
Packit |
47f805 |
fld0
|
|
Packit |
47f805 |
shr ecx,1
|
|
Packit |
47f805 |
jnc .lbl2
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
add eax,byte 4
|
|
Packit |
47f805 |
add edx,byte 4
|
|
Packit |
47f805 |
.lbl2
|
|
Packit |
47f805 |
shr ecx,1
|
|
Packit |
47f805 |
jnc .lbl3
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
add eax,byte 8
|
|
Packit |
47f805 |
add edx,byte 8
|
|
Packit |
47f805 |
and ecx,ecx
|
|
Packit |
47f805 |
.lbl3
|
|
Packit |
47f805 |
jz .ret2
|
|
Packit |
47f805 |
.lbl4
|
|
Packit |
47f805 |
fld dword [eax]
|
|
Packit |
47f805 |
fmul dword [edx]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 4]
|
|
Packit |
47f805 |
fmul dword [edx + 4]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 8]
|
|
Packit |
47f805 |
fmul dword [edx + 8]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
fld dword [eax + 12]
|
|
Packit |
47f805 |
fmul dword [edx + 12]
|
|
Packit |
47f805 |
faddp st1,st0
|
|
Packit |
47f805 |
add eax,byte 16
|
|
Packit |
47f805 |
add edx,byte 16
|
|
Packit |
47f805 |
dec ecx
|
|
Packit |
47f805 |
jnz .lbl4
|
|
Packit |
47f805 |
.ret2
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar04_float32_3DNow
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm0,qword [eax]
|
|
Packit |
47f805 |
pmov mm1,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm0,qword [edx]
|
|
Packit |
47f805 |
pfmul mm1,qword [edx+8]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pfadd mm0,mm1
|
|
Packit |
47f805 |
pmov qword [sp(%$p)],mm0
|
|
Packit |
47f805 |
femms
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar08_float32_3DNow
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm0,qword [eax]
|
|
Packit |
47f805 |
pmov mm1,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm0,qword [edx]
|
|
Packit |
47f805 |
pfmul mm1,qword [edx+8]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+16]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+24]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+16]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+24]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pfadd mm0,mm1
|
|
Packit |
47f805 |
pmov qword [sp(%$p)],mm0
|
|
Packit |
47f805 |
femms
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar12_float32_3DNow
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm0,qword [eax]
|
|
Packit |
47f805 |
pmov mm1,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm0,qword [edx]
|
|
Packit |
47f805 |
pfmul mm1,qword [edx+8]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+16]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+24]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+16]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+24]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+32]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+40]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+32]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+40]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pfadd mm0,mm1
|
|
Packit |
47f805 |
pmov qword [sp(%$p)],mm0
|
|
Packit |
47f805 |
femms
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar16_float32_3DNow
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm0,qword [eax]
|
|
Packit |
47f805 |
pmov mm1,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm0,qword [edx]
|
|
Packit |
47f805 |
pfmul mm1,qword [edx+8]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+16]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+24]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+16]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+24]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+32]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+40]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+32]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+40]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+48]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+56]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+48]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+56]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pfadd mm0,mm1
|
|
Packit |
47f805 |
pmov qword [sp(%$p)],mm0
|
|
Packit |
47f805 |
femms
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar20_float32_3DNow
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm0,qword [eax]
|
|
Packit |
47f805 |
pmov mm1,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm0,qword [edx]
|
|
Packit |
47f805 |
pfmul mm1,qword [edx+8]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+16]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+24]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+16]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+24]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+32]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+40]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+32]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+40]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+48]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+56]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+48]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+56]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+64]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+72]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+64]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+72]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pfadd mm0,mm1
|
|
Packit |
47f805 |
pmov qword [sp(%$p)],mm0
|
|
Packit |
47f805 |
femms
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar24_float32_3DNow
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm0,qword [eax]
|
|
Packit |
47f805 |
pmov mm1,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm0,qword [edx]
|
|
Packit |
47f805 |
pfmul mm1,qword [edx+8]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+16]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+24]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+16]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+24]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+32]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+40]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+32]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+40]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+48]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+56]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+48]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+56]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+64]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+72]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+64]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+72]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+80]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+88]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+80]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+88]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pfadd mm0,mm1
|
|
Packit |
47f805 |
pmov qword [sp(%$p)],mm0
|
|
Packit |
47f805 |
femms
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar32_float32_3DNow
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm0,qword [eax]
|
|
Packit |
47f805 |
pmov mm1,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm0,qword [edx]
|
|
Packit |
47f805 |
pfmul mm1,qword [edx+8]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+16]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+24]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+16]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+24]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+32]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+40]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+32]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+40]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+48]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+56]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+48]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+56]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+64]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+72]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+64]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+72]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+80]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+88]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+80]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+88]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+96]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+104]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+96]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+104]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm2,qword [eax+112]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+120]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx+112]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+120]
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pfadd mm0,mm1
|
|
Packit |
47f805 |
pmov qword [sp(%$p)],mm0
|
|
Packit |
47f805 |
femms
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar4n_float32_3DNow
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
%$len arg 4
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
mov ecx,[sp(%$len)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
pmov mm0,qword [eax]
|
|
Packit |
47f805 |
pmov mm1,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm0,qword [edx]
|
|
Packit |
47f805 |
pfmul mm1,qword [edx+8]
|
|
Packit |
47f805 |
dec ecx
|
|
Packit |
47f805 |
jz .ret4
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
add eax,byte 16
|
|
Packit |
47f805 |
add edx,byte 16
|
|
Packit |
47f805 |
.lbl4:
|
|
Packit |
47f805 |
pmov mm2,qword [eax]
|
|
Packit |
47f805 |
pmov mm3,qword [eax+8]
|
|
Packit |
47f805 |
pfmul mm2,qword [edx]
|
|
Packit |
47f805 |
pfmul mm3,qword [edx+8]
|
|
Packit |
47f805 |
add eax,byte 16
|
|
Packit |
47f805 |
add edx,byte 16
|
|
Packit |
47f805 |
pfadd mm0,mm2
|
|
Packit |
47f805 |
pfadd mm1,mm3
|
|
Packit |
47f805 |
dec ecx
|
|
Packit |
47f805 |
jnz .lbl4
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
.ret4: pfadd mm0,mm1
|
|
Packit |
47f805 |
pmov qword [sp(%$p)],mm0
|
|
Packit |
47f805 |
femms
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar1n_float32_3DNow
|
|
Packit |
47f805 |
jmp scalar24_float32_i387
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar04_float32_SIMD
|
|
Packit |
47f805 |
jmp scalar04_float32_i387
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar08_float32_SIMD
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm0, [eax]
|
|
Packit |
47f805 |
movups xmm1, [eax+16]
|
|
Packit |
47f805 |
mulps xmm0, [edx]
|
|
Packit |
47f805 |
mulps xmm1, [edx+16]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
addps xmm0,xmm1
|
|
Packit |
47f805 |
sub esp,16
|
|
Packit |
47f805 |
movups [esp],xmm0
|
|
Packit |
47f805 |
fld dword [esp+ 0]
|
|
Packit |
47f805 |
fadd dword [esp+ 4]
|
|
Packit |
47f805 |
fadd dword [esp+ 8]
|
|
Packit |
47f805 |
fadd dword [esp+12]
|
|
Packit |
47f805 |
add esp,16
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar12_float32_SIMD
|
|
Packit |
47f805 |
jmp scalar12_float32_i387
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar16_float32_SIMD
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm0, [eax]
|
|
Packit |
47f805 |
movups xmm1, [eax+16]
|
|
Packit |
47f805 |
mulps xmm0, [edx]
|
|
Packit |
47f805 |
mulps xmm1, [edx+16]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm2, [eax+32]
|
|
Packit |
47f805 |
movups xmm3, [eax+48]
|
|
Packit |
47f805 |
mulps xmm2, [edx+32]
|
|
Packit |
47f805 |
mulps xmm3, [edx+48]
|
|
Packit |
47f805 |
addps xmm0,xmm2
|
|
Packit |
47f805 |
addps xmm1,xmm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
addps xmm0,xmm1
|
|
Packit |
47f805 |
sub esp,16
|
|
Packit |
47f805 |
movups [esp],xmm0
|
|
Packit |
47f805 |
fld dword [esp+ 0]
|
|
Packit |
47f805 |
fadd dword [esp+ 4]
|
|
Packit |
47f805 |
fadd dword [esp+ 8]
|
|
Packit |
47f805 |
fadd dword [esp+12]
|
|
Packit |
47f805 |
add esp,16
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar20_float32_SIMD
|
|
Packit |
47f805 |
jmp scalar20_float32_i387
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar24_float32_SIMD
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm0, [eax]
|
|
Packit |
47f805 |
movups xmm1, [eax+16]
|
|
Packit |
47f805 |
mulps xmm0, [edx]
|
|
Packit |
47f805 |
mulps xmm1, [edx+16]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm2, [eax+32]
|
|
Packit |
47f805 |
movups xmm3, [eax+48]
|
|
Packit |
47f805 |
mulps xmm2, [edx+32]
|
|
Packit |
47f805 |
mulps xmm3, [edx+48]
|
|
Packit |
47f805 |
addps xmm0,xmm2
|
|
Packit |
47f805 |
addps xmm1,xmm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm2, [eax+64]
|
|
Packit |
47f805 |
movups xmm3, [eax+80]
|
|
Packit |
47f805 |
mulps xmm2, [edx+64]
|
|
Packit |
47f805 |
mulps xmm3, [edx+80]
|
|
Packit |
47f805 |
addps xmm0,xmm2
|
|
Packit |
47f805 |
addps xmm1,xmm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
addps xmm0,xmm1
|
|
Packit |
47f805 |
sub esp,16
|
|
Packit |
47f805 |
movups [esp],xmm0
|
|
Packit |
47f805 |
fld dword [esp+ 0]
|
|
Packit |
47f805 |
fadd dword [esp+ 4]
|
|
Packit |
47f805 |
fadd dword [esp+ 8]
|
|
Packit |
47f805 |
fadd dword [esp+12]
|
|
Packit |
47f805 |
add esp,16
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar32_float32_SIMD
|
|
Packit |
47f805 |
%$p arg 4
|
|
Packit |
47f805 |
%$q arg 4
|
|
Packit |
47f805 |
mov eax,[sp(%$p)]
|
|
Packit |
47f805 |
mov edx,[sp(%$q)]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm0, [eax]
|
|
Packit |
47f805 |
movups xmm1, [eax+16]
|
|
Packit |
47f805 |
mulps xmm0, [edx]
|
|
Packit |
47f805 |
mulps xmm1, [edx+16]
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm2, [eax+32]
|
|
Packit |
47f805 |
movups xmm3, [eax+48]
|
|
Packit |
47f805 |
mulps xmm2, [edx+32]
|
|
Packit |
47f805 |
mulps xmm3, [edx+48]
|
|
Packit |
47f805 |
addps xmm0,xmm2
|
|
Packit |
47f805 |
addps xmm1,xmm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm2, [eax+64]
|
|
Packit |
47f805 |
movups xmm3, [eax+80]
|
|
Packit |
47f805 |
mulps xmm2, [edx+64]
|
|
Packit |
47f805 |
mulps xmm3, [edx+80]
|
|
Packit |
47f805 |
addps xmm0,xmm2
|
|
Packit |
47f805 |
addps xmm1,xmm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movups xmm2, [eax+96]
|
|
Packit |
47f805 |
movups xmm3, [eax+112]
|
|
Packit |
47f805 |
mulps xmm2, [edx+96]
|
|
Packit |
47f805 |
mulps xmm3, [edx+112]
|
|
Packit |
47f805 |
addps xmm0,xmm2
|
|
Packit |
47f805 |
addps xmm1,xmm3
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
addps xmm0,xmm1
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
;sub esp,16
|
|
Packit |
47f805 |
;movups [esp],xmm0
|
|
Packit |
47f805 |
;fld dword [esp+ 0]
|
|
Packit |
47f805 |
;fadd dword [esp+ 4]
|
|
Packit |
47f805 |
;fadd dword [esp+ 8]
|
|
Packit |
47f805 |
;fadd dword [esp+12]
|
|
Packit |
47f805 |
;add esp,16
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
movhlps xmm1,xmm0
|
|
Packit |
47f805 |
addps xmm0,xmm1
|
|
Packit |
47f805 |
movlps [sp(%$p)],xmm0
|
|
Packit |
47f805 |
fld dword [sp(%$p)]
|
|
Packit |
47f805 |
fadd dword [sp(%$p)+4]
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar4n_float32_SIMD
|
|
Packit |
47f805 |
jmp scalar4n_float32_i387
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
proc scalar1n_float32_SIMD
|
|
Packit |
47f805 |
jmp scalar1n_float32_i387
|
|
Packit |
47f805 |
endproc
|
|
Packit |
47f805 |
|
|
Packit |
47f805 |
; end of scalar.nas
|