204 lines
3.9 KiB
NASM
204 lines
3.9 KiB
NASM
;========================================================================
|
|
;= =
|
|
;= Fast Hartley Transform demo for KolibriOS =
|
|
;= =
|
|
;= Copyright (C) 2010, Artem Jerdev <kolibri@jerdev.co.uk> =
|
|
;= =
|
|
;= refer to wiki.kolibtios.org for all details =
|
|
;= =
|
|
;========================================================================
|
|
|
|
|
|
|
|
|
|
use32
|
|
|
|
org 0x0
|
|
|
|
db 'MENUET01' ; 8 byte id
|
|
dd 0x01 ; header version
|
|
dd START ; start of code
|
|
dd I_END ; size of image
|
|
dd 0x100000 ; memory for app
|
|
dd 0xbfffc ; esp
|
|
dd 0x0 , 0x0 ; I_Param , I_Icon
|
|
|
|
include '../../macros.inc'
|
|
include '../debug.inc'
|
|
include 'fht4code.asm'
|
|
|
|
|
|
START: ; start of execution
|
|
|
|
call main
|
|
|
|
|
|
mov eax,-1 ; close this program
|
|
int 0x40
|
|
|
|
|
|
;=============================================================
|
|
;Func: calculates a simple function
|
|
; ff = (512*2^(-t) * cos (2.5*t))
|
|
; uses: eax, ebx
|
|
;------------
|
|
Func:
|
|
|
|
; 9 : {
|
|
|
|
; 10 : double x,t;
|
|
; 11 : int f;
|
|
; 12 :
|
|
; 13 : x = (i < N2) ? i : i - NUM_POINTS;
|
|
mov eax, [ii]
|
|
cmp eax, 512
|
|
jge .index_negative
|
|
jmp .index_correct
|
|
.index_negative:
|
|
sub eax, 1024
|
|
.index_correct:
|
|
mov [temp], eax
|
|
; fild [temp]
|
|
|
|
; 14 : t = x / 16.0;
|
|
; f2xm1 argument (abs) must be less than 1, so
|
|
mov [t_mod], eax
|
|
and [t_mod], 0x0F ; x % 16
|
|
shr eax, 4 ; x / 16
|
|
mov [t_div], eax
|
|
fild [temp]
|
|
|
|
; 15 : if (t<0) t = -t;
|
|
fabs
|
|
exp_ok:
|
|
; 16 : f = (512*2^(-t) * cos (2.5*t));
|
|
fchs
|
|
f2xm1
|
|
fmul [f500]
|
|
fstp [tv93]
|
|
fld [f2_5]
|
|
fmul [tt]
|
|
fcos
|
|
fmul [tv93]
|
|
fstp [tt]
|
|
mov bx, word[tt+6]
|
|
shr bx,4
|
|
and bx,0x07FF
|
|
add ax,bx
|
|
shl ax,4
|
|
and word[tt+6], 0x800F
|
|
or word[tt+6], ax
|
|
fld [tt]
|
|
fstp [ff]
|
|
|
|
; 17 : return f;
|
|
; 18 : }
|
|
ret
|
|
;---------------------------------------------------------
|
|
; test data filler
|
|
;
|
|
; uses eax, ebx, ecx
|
|
FillData:
|
|
; 29 : for (i=0; i<NUM_POINTS; i++)
|
|
; here : ecx = i
|
|
xor ecx, ecx
|
|
.funcloop:
|
|
; 30 : {
|
|
; 31 : ia[i] = Func(i);
|
|
mov [ii], ecx
|
|
call Func
|
|
fld [ff]
|
|
fstp qword [edx+ecx*8]
|
|
; 32 : }
|
|
inc ecx
|
|
cmp ecx, [_in] ; ecx == N ?
|
|
jne .funcloop
|
|
ret
|
|
|
|
;====================================================================
|
|
; main
|
|
;====================================================================
|
|
|
|
_ia dd 0
|
|
_ii dd 0
|
|
_ip dd 0
|
|
_in dd 0
|
|
_it dd 0
|
|
;-----------------
|
|
|
|
main:
|
|
mov eax, 68
|
|
mov ebx, 11
|
|
int 0x40
|
|
fninit
|
|
mov cl, 2 ; power of 4
|
|
mov byte[_ip], cl
|
|
mov eax, 1
|
|
shl eax, cl
|
|
shl eax, cl
|
|
mov [_in], eax
|
|
mov dl, cl
|
|
call CreateSinCosTable
|
|
mov [_it], edx
|
|
mov ecx, [_in]
|
|
shl ecx, 3
|
|
mov ebx, 12
|
|
mov eax, 68
|
|
int 0x40
|
|
mov [_ia], eax
|
|
mov edx, eax
|
|
|
|
call FillData
|
|
|
|
cpuid
|
|
rdtsc
|
|
mov [t_0], eax
|
|
|
|
push [_it]
|
|
push [_ia]
|
|
push [_ip]
|
|
push [_in]
|
|
call FHT_4
|
|
add esp, 16
|
|
cpuid
|
|
rdtsc
|
|
mov [t_1], eax
|
|
|
|
sub eax, [t_0]
|
|
debug_print_hex eax
|
|
print '<- fht time'
|
|
|
|
mov edx, [_it]
|
|
call DestroySinCosTable
|
|
mov ecx, [_ia]
|
|
mov ebx, 13
|
|
mov eax, 68
|
|
int 0x40
|
|
ret
|
|
|
|
|
|
; ========================================================
|
|
; static data
|
|
;----------------
|
|
|
|
align 8
|
|
|
|
;f18 dq 0x4032000000000000
|
|
f256 dq 256.01f
|
|
f14_2 dq 14.2f
|
|
f500 dq 0x407f400000000000
|
|
f2_5 dq 0x4004000000000000
|
|
tt dq ?
|
|
tv93 dq ?
|
|
t_div dd ?
|
|
t_mod dd ?
|
|
temp dd ?
|
|
ff dq ? ; return value (int)
|
|
ii dd ? ; argument (int) = array index
|
|
t_1 dd ?
|
|
t_0 dd ?
|
|
fcontrol dw 0x0037f
|
|
title db ' Fast Hartley Transform Test - A.Jerdev 2010'
|
|
|
|
I_END:
|