|
Packit |
a4058c |
/*
|
|
Packit |
a4058c |
* Copyright (C) 2000 Red Hat, Inc
|
|
Packit |
a4058c |
*
|
|
Packit |
a4058c |
* This library is free software; you can redistribute it and/or
|
|
Packit |
a4058c |
* modify it under the terms of the GNU Lesser General Public
|
|
Packit |
a4058c |
* License as published by the Free Software Foundation; either
|
|
Packit |
a4058c |
* version 2 of the License, or (at your option) any later version.
|
|
Packit |
a4058c |
*
|
|
Packit |
a4058c |
* This library is distributed in the hope that it will be useful,
|
|
Packit |
a4058c |
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
Packit |
a4058c |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Packit |
a4058c |
* Lesser General Public License for more details.
|
|
Packit |
a4058c |
*
|
|
Packit |
a4058c |
* You should have received a copy of the GNU Lesser General Public
|
|
Packit |
a4058c |
* License along with this library; if not, see <http://www.gnu.org/licenses/>.
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
.file "scale_line_22_33_mmx.S"
|
|
Packit |
a4058c |
.version "01.01"
|
|
Packit |
a4058c |
gcc2_compiled.:
|
|
Packit |
a4058c |
.text
|
|
Packit |
a4058c |
.align 16
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
#if !defined(__MINGW32__) && !defined(__CYGWIN__) && !defined(__INTERIX)
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/* Magic indicating no need for an executable stack */
|
|
Packit |
a4058c |
#if !defined __powerpc64__ && !defined __ia64__
|
|
Packit |
a4058c |
.section .note.GNU-stack; .previous
|
|
Packit |
a4058c |
#endif
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
.globl _pixops_scale_line_22_33_mmx
|
|
Packit |
a4058c |
.type _pixops_scale_line_22_33_mmx,@function
|
|
Packit |
a4058c |
_pixops_scale_line_22_33_mmx:
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
#else
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
.globl __pixops_scale_line_22_33_mmx
|
|
Packit |
a4058c |
__pixops_scale_line_22_33_mmx:
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
#endif
|
|
Packit |
a4058c |
/*
|
|
Packit |
a4058c |
* Arguments
|
|
Packit |
a4058c |
*
|
|
Packit |
a4058c |
* weights: 8(%ebp)
|
|
Packit |
a4058c |
* p: 12(%ebp) %esi
|
|
Packit |
a4058c |
* q1: 16(%ebp)
|
|
Packit |
a4058c |
* q2: 20(%ebp)
|
|
Packit |
a4058c |
* xstep: 24(%ebp)
|
|
Packit |
a4058c |
* p_end: 28(%ebp)
|
|
Packit |
a4058c |
* xinit: 32(%ebp)
|
|
Packit |
a4058c |
*
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/*
|
|
Packit |
a4058c |
* Function call entry
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
pushl %ebp
|
|
Packit |
a4058c |
movl %esp,%ebp
|
|
Packit |
a4058c |
subl $28,%esp
|
|
Packit |
a4058c |
pushl %edi
|
|
Packit |
a4058c |
pushl %esi
|
|
Packit |
a4058c |
pushl %ebx
|
|
Packit |
a4058c |
/* Locals:
|
|
Packit |
a4058c |
* int x %ebx
|
|
Packit |
a4058c |
* int x_scaled -24(%ebp)
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/*
|
|
Packit |
a4058c |
* Setup
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
/* Initialize variables */
|
|
Packit |
a4058c |
movl 32(%ebp),%ebx
|
|
Packit |
a4058c |
movl 32(%ebp),%edx
|
|
Packit |
a4058c |
sarl $16,%edx
|
|
Packit |
a4058c |
movl 12(%ebp),%esi
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
cmpl 28(%ebp),%esi
|
|
Packit |
a4058c |
jnb .out
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/* For the body of this loop, %mm01, %mm1, %mm2, %mm3 hold the 4 adjoining
|
|
Packit |
a4058c |
* points we are interpolating between, as:
|
|
Packit |
a4058c |
*
|
|
Packit |
a4058c |
* 000000BB00GG00RR
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/* Load initial values into %mm1, %mm3 */
|
|
Packit |
a4058c |
leal (%edx,%edx,2),%edx # Multiply by 3
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
movl 16(%ebp),%edi
|
|
Packit |
a4058c |
pxor %mm4, %mm4
|
|
Packit |
a4058c |
movzbl 2(%edi,%edx),%ecx
|
|
Packit |
a4058c |
shll $16,%ecx
|
|
Packit |
a4058c |
movzwl (%edi,%edx),%eax
|
|
Packit |
a4058c |
orl %eax,%ecx
|
|
Packit |
a4058c |
movd %ecx, %mm1
|
|
Packit |
a4058c |
punpcklbw %mm4, %mm1
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
movl 20(%ebp),%edi
|
|
Packit |
a4058c |
movzbl 2(%edi,%edx),%ecx
|
|
Packit |
a4058c |
shll $16,%ecx
|
|
Packit |
a4058c |
movzwl (%edi,%edx),%eax
|
|
Packit |
a4058c |
orl %eax,%ecx
|
|
Packit |
a4058c |
movd %ecx, %mm3
|
|
Packit |
a4058c |
punpcklbw %mm4, %mm3
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
addl $65536,%ebx
|
|
Packit |
a4058c |
movl %ebx,%edx
|
|
Packit |
a4058c |
sarl $16,%edx
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
jmp .newx
|
|
Packit |
a4058c |
.p2align 4,,7
|
|
Packit |
a4058c |
.loop:
|
|
Packit |
a4058c |
/* short *pixel_weights = weights + ((x >> (SCALE_SHIFT - SUBSAMPLE_BITS)) & SUBSAMPLE_MASK) * n_x * n_y
|
|
Packit |
a4058c |
* 16 4 0xf 2 2
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
movl %ebx,%eax
|
|
Packit |
a4058c |
andl $0xf000,%eax
|
|
Packit |
a4058c |
shrl $7,%eax
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/* At this point, %edi holds weights. Load the 4 weights into %mm4,%mm5,%mm6,%mm7, multiply and
|
|
Packit |
a4058c |
* accumulate.
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
movq (%edi,%eax),%mm4
|
|
Packit |
a4058c |
pmullw %mm0,%mm4
|
|
Packit |
a4058c |
movq 8(%edi,%eax),%mm5
|
|
Packit |
a4058c |
pmullw %mm1,%mm5
|
|
Packit |
a4058c |
movq 16(%edi,%eax),%mm6
|
|
Packit |
a4058c |
movq 24(%edi,%eax),%mm7
|
|
Packit |
a4058c |
pmullw %mm2,%mm6
|
|
Packit |
a4058c |
pmullw %mm3,%mm7
|
|
Packit |
a4058c |
paddw %mm4, %mm5
|
|
Packit |
a4058c |
paddw %mm6, %mm7
|
|
Packit |
a4058c |
paddw %mm5, %mm7
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/* %mm7 holds the accumulated sum. Compute (C + 0x80) / 256
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
pxor %mm4, %mm4
|
|
Packit |
a4058c |
movl $8421504, %eax # 0x00808080
|
|
Packit |
a4058c |
movd %eax, %mm6
|
|
Packit |
a4058c |
punpcklbw %mm4, %mm6
|
|
Packit |
a4058c |
paddw %mm6, %mm7
|
|
Packit |
a4058c |
psrlw $8, %mm7
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/* Pack into %eax and store result
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
packuswb %mm7, %mm7
|
|
Packit |
a4058c |
movd %mm7, %eax
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
movb %al, (%esi)
|
|
Packit |
a4058c |
shrl $8, %eax
|
|
Packit |
a4058c |
movw %ax, 1(%esi)
|
|
Packit |
a4058c |
addl $3, %esi
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
cmpl %esi,28(%ebp)
|
|
Packit |
a4058c |
je .out
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
/* x += x_step; */
|
|
Packit |
a4058c |
addl 24(%ebp),%ebx
|
|
Packit |
a4058c |
/* x_scaled = x >> 16; */
|
|
Packit |
a4058c |
movl %ebx,%edx
|
|
Packit |
a4058c |
sarl $16,%edx
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
cmpl %edx,-24(%ebp)
|
|
Packit |
a4058c |
je .loop
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
.newx:
|
|
Packit |
a4058c |
movl %edx,-24(%ebp)
|
|
Packit |
a4058c |
/*
|
|
Packit |
a4058c |
* Load the two new values into %mm1, %mm3, move old values into %mm0, %mm2
|
|
Packit |
a4058c |
*/
|
|
Packit |
a4058c |
movq %mm1, %mm0
|
|
Packit |
a4058c |
movq %mm3, %mm2
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
leal (%edx,%edx,2),%edx # Multiply by 3
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
movl 16(%ebp),%edi
|
|
Packit |
a4058c |
movzbl 2(%edi,%edx),%ecx
|
|
Packit |
a4058c |
shll $16,%ecx
|
|
Packit |
a4058c |
movzwl (%edi,%edx),%eax
|
|
Packit |
a4058c |
orl %eax,%ecx
|
|
Packit |
a4058c |
movd %ecx, %mm1
|
|
Packit |
a4058c |
punpcklbw %mm4, %mm1
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
movl 20(%ebp),%edi
|
|
Packit |
a4058c |
movzbl 2(%edi,%edx),%ecx
|
|
Packit |
a4058c |
shll $16,%ecx
|
|
Packit |
a4058c |
movzwl (%edi,%edx),%eax
|
|
Packit |
a4058c |
orl %eax,%ecx
|
|
Packit |
a4058c |
movd %ecx, %mm3
|
|
Packit |
a4058c |
punpcklbw %mm4, %mm3
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
movl 8(%ebp),%edi
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
jmp .loop
|
|
Packit |
a4058c |
|
|
Packit |
a4058c |
.out:
|
|
Packit |
a4058c |
movl %esi,%eax
|
|
Packit |
a4058c |
emms
|
|
Packit |
a4058c |
leal -40(%ebp),%esp
|
|
Packit |
a4058c |
popl %ebx
|
|
Packit |
a4058c |
popl %esi
|
|
Packit |
a4058c |
popl %edi
|
|
Packit |
a4058c |
movl %ebp,%esp
|
|
Packit |
a4058c |
popl %ebp
|
|
Packit |
a4058c |
ret
|