2008-12-15 22:12:51 +00:00
|
|
|
/*
|
|
|
|
* Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
|
|
|
|
*
|
2011-03-18 17:35:10 +00:00
|
|
|
* This file is part of Libav.
|
2008-12-15 22:12:51 +00:00
|
|
|
*
|
2011-03-18 17:35:10 +00:00
|
|
|
* Libav is free software; you can redistribute it and/or
|
2008-12-15 22:12:51 +00:00
|
|
|
* modify it under the terms of the GNU Lesser General Public
|
|
|
|
* License as published by the Free Software Foundation; either
|
|
|
|
* version 2.1 of the License, or (at your option) any later version.
|
|
|
|
*
|
2011-03-18 17:35:10 +00:00
|
|
|
* Libav is distributed in the hope that it will be useful,
|
2008-12-15 22:12:51 +00:00
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
|
* Lesser General Public License for more details.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU Lesser General Public
|
2011-03-18 17:35:10 +00:00
|
|
|
* License along with Libav; if not, write to the Free Software
|
2008-12-15 22:12:51 +00:00
|
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "asm.S"
|
|
|
|
|
2008-12-25 23:13:43 +00:00
|
|
|
preserve8
|
2008-12-15 22:12:51 +00:00
|
|
|
|
|
|
|
function ff_h264_idct_add_neon, export=1
|
|
|
|
vld1.64 {d0-d3}, [r1,:128]
|
|
|
|
|
|
|
|
vswp d1, d2
|
|
|
|
vadd.i16 d4, d0, d1
|
|
|
|
vshr.s16 q8, q1, #1
|
|
|
|
vsub.i16 d5, d0, d1
|
|
|
|
vadd.i16 d6, d2, d17
|
|
|
|
vsub.i16 d7, d16, d3
|
|
|
|
vadd.i16 q0, q2, q3
|
|
|
|
vsub.i16 q1, q2, q3
|
|
|
|
|
|
|
|
vtrn.16 d0, d1
|
|
|
|
vtrn.16 d3, d2
|
|
|
|
vtrn.32 d0, d3
|
|
|
|
vtrn.32 d1, d2
|
|
|
|
|
|
|
|
vadd.i16 d4, d0, d3
|
|
|
|
vld1.32 {d18[0]}, [r0,:32], r2
|
|
|
|
vswp d1, d3
|
|
|
|
vshr.s16 q8, q1, #1
|
|
|
|
vld1.32 {d19[1]}, [r0,:32], r2
|
|
|
|
vsub.i16 d5, d0, d1
|
|
|
|
vld1.32 {d18[1]}, [r0,:32], r2
|
|
|
|
vadd.i16 d6, d16, d3
|
|
|
|
vld1.32 {d19[0]}, [r0,:32], r2
|
|
|
|
vsub.i16 d7, d2, d17
|
|
|
|
sub r0, r0, r2, lsl #2
|
|
|
|
vadd.i16 q0, q2, q3
|
|
|
|
vsub.i16 q1, q2, q3
|
|
|
|
|
2009-12-02 00:37:39 +00:00
|
|
|
vrshr.s16 q0, q0, #6
|
|
|
|
vrshr.s16 q1, q1, #6
|
2008-12-15 22:12:51 +00:00
|
|
|
|
|
|
|
vaddw.u8 q0, q0, d18
|
|
|
|
vaddw.u8 q1, q1, d19
|
|
|
|
|
|
|
|
vqmovun.s16 d0, q0
|
|
|
|
vqmovun.s16 d1, q1
|
|
|
|
|
|
|
|
vst1.32 {d0[0]}, [r0,:32], r2
|
|
|
|
vst1.32 {d1[1]}, [r0,:32], r2
|
|
|
|
vst1.32 {d0[1]}, [r0,:32], r2
|
|
|
|
vst1.32 {d1[0]}, [r0,:32], r2
|
|
|
|
|
|
|
|
bx lr
|
2010-03-09 16:17:56 +00:00
|
|
|
endfunc
|
2008-12-15 22:12:54 +00:00
|
|
|
|
|
|
|
function ff_h264_idct_dc_add_neon, export=1
|
|
|
|
vld1.16 {d2[],d3[]}, [r1,:16]
|
|
|
|
vrshr.s16 q1, q1, #6
|
|
|
|
vld1.32 {d0[0]}, [r0,:32], r2
|
|
|
|
vld1.32 {d0[1]}, [r0,:32], r2
|
|
|
|
vaddw.u8 q2, q1, d0
|
|
|
|
vld1.32 {d1[0]}, [r0,:32], r2
|
|
|
|
vld1.32 {d1[1]}, [r0,:32], r2
|
|
|
|
vaddw.u8 q1, q1, d1
|
|
|
|
vqmovun.s16 d0, q2
|
|
|
|
vqmovun.s16 d1, q1
|
|
|
|
sub r0, r0, r2, lsl #2
|
|
|
|
vst1.32 {d0[0]}, [r0,:32], r2
|
|
|
|
vst1.32 {d0[1]}, [r0,:32], r2
|
|
|
|
vst1.32 {d1[0]}, [r0,:32], r2
|
|
|
|
vst1.32 {d1[1]}, [r0,:32], r2
|
|
|
|
bx lr
|
2010-03-09 16:17:56 +00:00
|
|
|
endfunc
|
2008-12-25 23:13:43 +00:00
|
|
|
|
|
|
|
function ff_h264_idct_add16_neon, export=1
|
|
|
|
push {r4-r8,lr}
|
|
|
|
mov r4, r0
|
|
|
|
mov r5, r1
|
|
|
|
mov r1, r2
|
|
|
|
mov r2, r3
|
|
|
|
ldr r6, [sp, #24]
|
2008-12-30 03:13:52 +00:00
|
|
|
movrel r7, scan8
|
2008-12-25 23:13:43 +00:00
|
|
|
mov ip, #16
|
|
|
|
1: ldrb r8, [r7], #1
|
|
|
|
ldr r0, [r5], #4
|
|
|
|
ldrb r8, [r6, r8]
|
|
|
|
subs r8, r8, #1
|
|
|
|
blt 2f
|
|
|
|
ldrsh lr, [r1]
|
|
|
|
add r0, r0, r4
|
2011-06-14 10:29:48 +00:00
|
|
|
it ne
|
2008-12-25 23:13:43 +00:00
|
|
|
movne lr, #0
|
|
|
|
cmp lr, #0
|
2011-06-14 10:29:48 +00:00
|
|
|
ite ne
|
|
|
|
adrne lr, ff_h264_idct_dc_add_neon + CONFIG_THUMB
|
|
|
|
adreq lr, ff_h264_idct_add_neon + CONFIG_THUMB
|
2008-12-25 23:13:43 +00:00
|
|
|
blx lr
|
|
|
|
2: subs ip, ip, #1
|
|
|
|
add r1, r1, #32
|
|
|
|
bne 1b
|
|
|
|
pop {r4-r8,pc}
|
2010-03-09 16:17:56 +00:00
|
|
|
endfunc
|
2008-12-25 23:13:43 +00:00
|
|
|
|
|
|
|
function ff_h264_idct_add16intra_neon, export=1
|
|
|
|
push {r4-r8,lr}
|
|
|
|
mov r4, r0
|
|
|
|
mov r5, r1
|
|
|
|
mov r1, r2
|
|
|
|
mov r2, r3
|
|
|
|
ldr r6, [sp, #24]
|
2008-12-30 03:13:52 +00:00
|
|
|
movrel r7, scan8
|
2008-12-25 23:13:43 +00:00
|
|
|
mov ip, #16
|
|
|
|
1: ldrb r8, [r7], #1
|
|
|
|
ldr r0, [r5], #4
|
|
|
|
ldrb r8, [r6, r8]
|
|
|
|
add r0, r0, r4
|
|
|
|
cmp r8, #0
|
|
|
|
ldrsh r8, [r1]
|
2011-06-14 10:29:48 +00:00
|
|
|
iteet ne
|
|
|
|
adrne lr, ff_h264_idct_add_neon + CONFIG_THUMB
|
|
|
|
adreq lr, ff_h264_idct_dc_add_neon + CONFIG_THUMB
|
2008-12-25 23:13:43 +00:00
|
|
|
cmpeq r8, #0
|
|
|
|
blxne lr
|
|
|
|
subs ip, ip, #1
|
|
|
|
add r1, r1, #32
|
|
|
|
bne 1b
|
|
|
|
pop {r4-r8,pc}
|
2010-03-09 16:17:56 +00:00
|
|
|
endfunc
|
2008-12-25 23:13:43 +00:00
|
|
|
|
|
|
|
function ff_h264_idct_add8_neon, export=1
|
|
|
|
push {r4-r10,lr}
|
|
|
|
ldm r0, {r4,r9}
|
|
|
|
add r5, r1, #16*4
|
|
|
|
add r1, r2, #16*32
|
|
|
|
mov r2, r3
|
2011-06-15 11:58:00 +00:00
|
|
|
mov r3, r1
|
2008-12-25 23:13:43 +00:00
|
|
|
ldr r6, [sp, #32]
|
2008-12-30 03:13:52 +00:00
|
|
|
movrel r7, scan8+16
|
2011-06-15 11:58:00 +00:00
|
|
|
mov r12, #0
|
|
|
|
1: ldrb r8, [r7, r12]
|
|
|
|
ldr r0, [r5, r12, lsl #2]
|
2008-12-25 23:13:43 +00:00
|
|
|
ldrb r8, [r6, r8]
|
2011-06-15 11:58:00 +00:00
|
|
|
add r0, r0, r4
|
|
|
|
add r1, r3, r12, lsl #5
|
2008-12-25 23:13:43 +00:00
|
|
|
cmp r8, #0
|
|
|
|
ldrsh r8, [r1]
|
2011-06-14 10:29:48 +00:00
|
|
|
iteet ne
|
|
|
|
adrne lr, ff_h264_idct_add_neon + CONFIG_THUMB
|
|
|
|
adreq lr, ff_h264_idct_dc_add_neon + CONFIG_THUMB
|
2008-12-25 23:13:43 +00:00
|
|
|
cmpeq r8, #0
|
|
|
|
blxne lr
|
2011-06-15 11:58:00 +00:00
|
|
|
add r12, r12, #1
|
|
|
|
cmp r12, #4
|
2011-06-14 10:29:48 +00:00
|
|
|
itt eq
|
2011-06-15 11:58:00 +00:00
|
|
|
moveq r12, #16
|
|
|
|
moveq r4, r9
|
|
|
|
cmp r12, #20
|
|
|
|
blt 1b
|
2008-12-25 23:13:43 +00:00
|
|
|
pop {r4-r10,pc}
|
2010-03-09 16:17:56 +00:00
|
|
|
endfunc
|
2008-12-25 23:13:43 +00:00
|
|
|
|
2010-08-05 19:45:57 +00:00
|
|
|
.macro idct8x8_cols pass
|
|
|
|
.if \pass == 0
|
|
|
|
qa .req q2
|
|
|
|
qb .req q14
|
|
|
|
vshr.s16 q2, q10, #1
|
|
|
|
vadd.i16 q0, q8, q12
|
|
|
|
vld1.16 {q14-q15},[r1,:128]!
|
|
|
|
vsub.i16 q1, q8, q12
|
|
|
|
vshr.s16 q3, q14, #1
|
|
|
|
vsub.i16 q2, q2, q14
|
|
|
|
vadd.i16 q3, q3, q10
|
|
|
|
.else
|
|
|
|
qa .req q14
|
|
|
|
qb .req q2
|
|
|
|
vtrn.32 q8, q10
|
|
|
|
vtrn.16 q12, q13
|
|
|
|
vtrn.32 q9, q11
|
|
|
|
vtrn.32 q12, q2
|
|
|
|
vtrn.32 q13, q15
|
|
|
|
vswp d21, d4
|
|
|
|
vshr.s16 q14, q10, #1
|
|
|
|
vswp d17, d24
|
|
|
|
vshr.s16 q3, q2, #1
|
|
|
|
vswp d19, d26
|
|
|
|
vadd.i16 q0, q8, q12
|
|
|
|
vswp d23, d30
|
|
|
|
vsub.i16 q1, q8, q12
|
|
|
|
vsub.i16 q14, q14, q2
|
|
|
|
vadd.i16 q3, q3, q10
|
|
|
|
.endif
|
|
|
|
vadd.i16 q10, q1, qa
|
|
|
|
vsub.i16 q12, q1, qa
|
|
|
|
vadd.i16 q8, q0, q3
|
|
|
|
vsub.i16 qb, q0, q3
|
|
|
|
vsub.i16 q0, q13, q11
|
|
|
|
vadd.i16 q1, q15, q9
|
|
|
|
vsub.i16 qa, q15, q9
|
|
|
|
vadd.i16 q3, q13, q11
|
|
|
|
vsub.i16 q0, q0, q15
|
|
|
|
vsub.i16 q1, q1, q11
|
|
|
|
vadd.i16 qa, qa, q13
|
|
|
|
vadd.i16 q3, q3, q9
|
|
|
|
vshr.s16 q9, q9, #1
|
|
|
|
vshr.s16 q11, q11, #1
|
|
|
|
vshr.s16 q13, q13, #1
|
|
|
|
vshr.s16 q15, q15, #1
|
|
|
|
vsub.i16 q0, q0, q15
|
|
|
|
vsub.i16 q1, q1, q11
|
|
|
|
vadd.i16 qa, qa, q13
|
|
|
|
vadd.i16 q3, q3, q9
|
|
|
|
vshr.s16 q9, q0, #2
|
|
|
|
vshr.s16 q11, q1, #2
|
|
|
|
vshr.s16 q13, qa, #2
|
|
|
|
vshr.s16 q15, q3, #2
|
|
|
|
vsub.i16 q3, q3, q9
|
|
|
|
vsub.i16 qa, q11, qa
|
|
|
|
vadd.i16 q1, q1, q13
|
|
|
|
vadd.i16 q0, q0, q15
|
|
|
|
.if \pass == 0
|
|
|
|
vsub.i16 q15, q8, q3
|
|
|
|
vadd.i16 q8, q8, q3
|
|
|
|
vadd.i16 q9, q10, q2
|
|
|
|
vsub.i16 q2, q10, q2
|
|
|
|
vtrn.16 q8, q9
|
|
|
|
vadd.i16 q10, q12, q1
|
|
|
|
vtrn.16 q2, q15
|
|
|
|
vadd.i16 q11, q14, q0
|
|
|
|
vsub.i16 q13, q12, q1
|
|
|
|
vtrn.16 q10, q11
|
|
|
|
vsub.i16 q12, q14, q0
|
|
|
|
.else
|
|
|
|
vsub.i16 q15, q8, q3
|
|
|
|
vadd.i16 q8, q8, q3
|
|
|
|
vadd.i16 q9, q10, q14
|
|
|
|
vsub.i16 q14, q10, q14
|
|
|
|
vadd.i16 q10, q12, q1
|
|
|
|
vsub.i16 q13, q12, q1
|
|
|
|
vadd.i16 q11, q2, q0
|
|
|
|
vsub.i16 q12, q2, q0
|
|
|
|
.endif
|
|
|
|
.unreq qa
|
|
|
|
.unreq qb
|
|
|
|
.endm
|
|
|
|
|
|
|
|
function ff_h264_idct8_add_neon, export=1
|
|
|
|
vld1.16 {q8-q9}, [r1,:128]!
|
|
|
|
vld1.16 {q10-q11},[r1,:128]!
|
|
|
|
vld1.16 {q12-q13},[r1,:128]!
|
|
|
|
|
|
|
|
idct8x8_cols 0
|
|
|
|
idct8x8_cols 1
|
|
|
|
|
|
|
|
mov r3, r0
|
|
|
|
vrshr.s16 q8, q8, #6
|
|
|
|
vld1.8 {d0}, [r0,:64], r2
|
|
|
|
vrshr.s16 q9, q9, #6
|
|
|
|
vld1.8 {d1}, [r0,:64], r2
|
|
|
|
vrshr.s16 q10, q10, #6
|
|
|
|
vld1.8 {d2}, [r0,:64], r2
|
|
|
|
vrshr.s16 q11, q11, #6
|
|
|
|
vld1.8 {d3}, [r0,:64], r2
|
|
|
|
vrshr.s16 q12, q12, #6
|
|
|
|
vld1.8 {d4}, [r0,:64], r2
|
|
|
|
vrshr.s16 q13, q13, #6
|
|
|
|
vld1.8 {d5}, [r0,:64], r2
|
|
|
|
vrshr.s16 q14, q14, #6
|
|
|
|
vld1.8 {d6}, [r0,:64], r2
|
|
|
|
vrshr.s16 q15, q15, #6
|
|
|
|
vld1.8 {d7}, [r0,:64], r2
|
|
|
|
vaddw.u8 q8, q8, d0
|
|
|
|
vaddw.u8 q9, q9, d1
|
|
|
|
vaddw.u8 q10, q10, d2
|
|
|
|
vqmovun.s16 d0, q8
|
|
|
|
vaddw.u8 q11, q11, d3
|
|
|
|
vqmovun.s16 d1, q9
|
|
|
|
vaddw.u8 q12, q12, d4
|
|
|
|
vqmovun.s16 d2, q10
|
|
|
|
vst1.8 {d0}, [r3,:64], r2
|
|
|
|
vaddw.u8 q13, q13, d5
|
|
|
|
vqmovun.s16 d3, q11
|
|
|
|
vst1.8 {d1}, [r3,:64], r2
|
|
|
|
vaddw.u8 q14, q14, d6
|
|
|
|
vqmovun.s16 d4, q12
|
|
|
|
vst1.8 {d2}, [r3,:64], r2
|
|
|
|
vaddw.u8 q15, q15, d7
|
|
|
|
vqmovun.s16 d5, q13
|
|
|
|
vst1.8 {d3}, [r3,:64], r2
|
|
|
|
vqmovun.s16 d6, q14
|
|
|
|
vqmovun.s16 d7, q15
|
|
|
|
vst1.8 {d4}, [r3,:64], r2
|
|
|
|
vst1.8 {d5}, [r3,:64], r2
|
|
|
|
vst1.8 {d6}, [r3,:64], r2
|
|
|
|
vst1.8 {d7}, [r3,:64], r2
|
|
|
|
|
|
|
|
sub r1, r1, #128
|
|
|
|
bx lr
|
|
|
|
endfunc
|
|
|
|
|
|
|
|
function ff_h264_idct8_dc_add_neon, export=1
|
|
|
|
vld1.16 {d30[],d31[]},[r1,:16]
|
|
|
|
vld1.32 {d0}, [r0,:64], r2
|
|
|
|
vrshr.s16 q15, q15, #6
|
|
|
|
vld1.32 {d1}, [r0,:64], r2
|
|
|
|
vld1.32 {d2}, [r0,:64], r2
|
|
|
|
vaddw.u8 q8, q15, d0
|
|
|
|
vld1.32 {d3}, [r0,:64], r2
|
|
|
|
vaddw.u8 q9, q15, d1
|
|
|
|
vld1.32 {d4}, [r0,:64], r2
|
|
|
|
vaddw.u8 q10, q15, d2
|
|
|
|
vld1.32 {d5}, [r0,:64], r2
|
|
|
|
vaddw.u8 q11, q15, d3
|
|
|
|
vld1.32 {d6}, [r0,:64], r2
|
|
|
|
vaddw.u8 q12, q15, d4
|
|
|
|
vld1.32 {d7}, [r0,:64], r2
|
|
|
|
vaddw.u8 q13, q15, d5
|
|
|
|
vaddw.u8 q14, q15, d6
|
|
|
|
vaddw.u8 q15, q15, d7
|
|
|
|
vqmovun.s16 d0, q8
|
|
|
|
vqmovun.s16 d1, q9
|
|
|
|
vqmovun.s16 d2, q10
|
|
|
|
vqmovun.s16 d3, q11
|
|
|
|
sub r0, r0, r2, lsl #3
|
|
|
|
vst1.32 {d0}, [r0,:64], r2
|
|
|
|
vqmovun.s16 d4, q12
|
|
|
|
vst1.32 {d1}, [r0,:64], r2
|
|
|
|
vqmovun.s16 d5, q13
|
|
|
|
vst1.32 {d2}, [r0,:64], r2
|
|
|
|
vqmovun.s16 d6, q14
|
|
|
|
vst1.32 {d3}, [r0,:64], r2
|
|
|
|
vqmovun.s16 d7, q15
|
|
|
|
vst1.32 {d4}, [r0,:64], r2
|
|
|
|
vst1.32 {d5}, [r0,:64], r2
|
|
|
|
vst1.32 {d6}, [r0,:64], r2
|
|
|
|
vst1.32 {d7}, [r0,:64], r2
|
|
|
|
bx lr
|
|
|
|
endfunc
|
|
|
|
|
|
|
|
function ff_h264_idct8_add4_neon, export=1
|
|
|
|
push {r4-r8,lr}
|
|
|
|
mov r4, r0
|
|
|
|
mov r5, r1
|
|
|
|
mov r1, r2
|
|
|
|
mov r2, r3
|
|
|
|
ldr r6, [sp, #24]
|
|
|
|
movrel r7, scan8
|
|
|
|
mov r12, #16
|
|
|
|
1: ldrb r8, [r7], #4
|
|
|
|
ldr r0, [r5], #16
|
|
|
|
ldrb r8, [r6, r8]
|
|
|
|
subs r8, r8, #1
|
|
|
|
blt 2f
|
|
|
|
ldrsh lr, [r1]
|
|
|
|
add r0, r0, r4
|
2011-06-14 10:29:48 +00:00
|
|
|
it ne
|
2010-08-05 19:45:57 +00:00
|
|
|
movne lr, #0
|
|
|
|
cmp lr, #0
|
2011-06-14 10:29:48 +00:00
|
|
|
ite ne
|
|
|
|
adrne lr, ff_h264_idct8_dc_add_neon + CONFIG_THUMB
|
|
|
|
adreq lr, ff_h264_idct8_add_neon + CONFIG_THUMB
|
2010-08-05 19:45:57 +00:00
|
|
|
blx lr
|
|
|
|
2: subs r12, r12, #4
|
|
|
|
add r1, r1, #128
|
|
|
|
bne 1b
|
|
|
|
pop {r4-r8,pc}
|
|
|
|
endfunc
|
|
|
|
|
2011-07-09 11:37:08 +00:00
|
|
|
const scan8
|
|
|
|
.byte 4+ 1*8, 5+ 1*8, 4+ 2*8, 5+ 2*8
|
2011-06-15 11:58:00 +00:00
|
|
|
.byte 6+ 1*8, 7+ 1*8, 6+ 2*8, 7+ 2*8
|
|
|
|
.byte 4+ 3*8, 5+ 3*8, 4+ 4*8, 5+ 4*8
|
|
|
|
.byte 6+ 3*8, 7+ 3*8, 6+ 4*8, 7+ 4*8
|
|
|
|
.byte 4+ 6*8, 5+ 6*8, 4+ 7*8, 5+ 7*8
|
|
|
|
.byte 6+ 6*8, 7+ 6*8, 6+ 7*8, 7+ 7*8
|
|
|
|
.byte 4+ 8*8, 5+ 8*8, 4+ 9*8, 5+ 9*8
|
|
|
|
.byte 6+ 8*8, 7+ 8*8, 6+ 9*8, 7+ 9*8
|
|
|
|
.byte 4+11*8, 5+11*8, 4+12*8, 5+12*8
|
|
|
|
.byte 6+11*8, 7+11*8, 6+12*8, 7+12*8
|
|
|
|
.byte 4+13*8, 5+13*8, 4+14*8, 5+14*8
|
|
|
|
.byte 6+13*8, 7+13*8, 6+14*8, 7+14*8
|
2011-07-09 11:37:08 +00:00
|
|
|
endconst
|