media/libvpx/vp8/common/arm/neon/dequant_idct_neon.asm@ac0c01689b40 (annotated)
media/libvpx/vp8/common/arm/neon/dequant_idct_neon.asm
Thu, 15 Jan 2015 15:59:08 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Thu, 15 Jan 2015 15:59:08 +0100
- branch
- TOR_BUG_9701
- changeset 10
- ac0c01689b40
- permissions
- -rw-r--r--
Implement a real Private Browsing Mode condition by changing the API/ABI;
This solves Tor bug #9701, complying with disk avoidance documented in
https://www.torproject.org/projects/torbrowser/design/#disk-avoidance.
| michael@0 | 1 | ; |
| michael@0 | 2 | ; Copyright (c) 2010 The WebM project authors. All Rights Reserved. |
| michael@0 | 3 | ; |
| michael@0 | 4 | ; Use of this source code is governed by a BSD-style license |
| michael@0 | 5 | ; that can be found in the LICENSE file in the root of the source |
| michael@0 | 6 | ; tree. An additional intellectual property rights grant can be found |
| michael@0 | 7 | ; in the file PATENTS. All contributing project authors may |
| michael@0 | 8 | ; be found in the AUTHORS file in the root of the source tree. |
| michael@0 | 9 | ; |
| michael@0 | 10 | |
| michael@0 | 11 | |
| michael@0 | 12 | EXPORT |vp8_dequant_idct_add_neon| |
| michael@0 | 13 | ARM |
| michael@0 | 14 | REQUIRE8 |
| michael@0 | 15 | PRESERVE8 |
| michael@0 | 16 | |
| michael@0 | 17 | AREA ||.text||, CODE, READONLY, ALIGN=2 |
| michael@0 | 18 | ;void vp8_dequant_idct_add_neon(short *input, short *dq, |
| michael@0 | 19 | ; unsigned char *dest, int stride) |
| michael@0 | 20 | ; r0 short *input, |
| michael@0 | 21 | ; r1 short *dq, |
| michael@0 | 22 | ; r2 unsigned char *dest |
| michael@0 | 23 | ; r3 int stride |
| michael@0 | 24 | |
| michael@0 | 25 | |vp8_dequant_idct_add_neon| PROC |
| michael@0 | 26 | vld1.16 {q3, q4}, [r0] |
| michael@0 | 27 | vld1.16 {q5, q6}, [r1] |
| michael@0 | 28 | |
| michael@0 | 29 | add r1, r2, r3 ; r1 = dest + stride |
| michael@0 | 30 | lsl r3, #1 ; 2x stride |
| michael@0 | 31 | |
| michael@0 | 32 | vld1.32 {d14[0]}, [r2], r3 |
| michael@0 | 33 | vld1.32 {d14[1]}, [r1], r3 |
| michael@0 | 34 | vld1.32 {d15[0]}, [r2] |
| michael@0 | 35 | vld1.32 {d15[1]}, [r1] |
| michael@0 | 36 | |
| michael@0 | 37 | adr r12, cospi8sqrt2minus1 ; pointer to the first constant |
| michael@0 | 38 | |
| michael@0 | 39 | vmul.i16 q1, q3, q5 ;input for short_idct4x4llm_neon |
| michael@0 | 40 | vmul.i16 q2, q4, q6 |
| michael@0 | 41 | |
| michael@0 | 42 | ;|short_idct4x4llm_neon| PROC |
| michael@0 | 43 | vld1.16 {d0}, [r12] |
| michael@0 | 44 | vswp d3, d4 ;q2(vp[4] vp[12]) |
| michael@0 | 45 | |
| michael@0 | 46 | vqdmulh.s16 q3, q2, d0[2] |
| michael@0 | 47 | vqdmulh.s16 q4, q2, d0[0] |
| michael@0 | 48 | |
| michael@0 | 49 | vqadd.s16 d12, d2, d3 ;a1 |
| michael@0 | 50 | vqsub.s16 d13, d2, d3 ;b1 |
| michael@0 | 51 | |
| michael@0 | 52 | vshr.s16 q3, q3, #1 |
| michael@0 | 53 | vshr.s16 q4, q4, #1 |
| michael@0 | 54 | |
| michael@0 | 55 | vqadd.s16 q3, q3, q2 |
| michael@0 | 56 | vqadd.s16 q4, q4, q2 |
| michael@0 | 57 | |
| michael@0 | 58 | vqsub.s16 d10, d6, d9 ;c1 |
| michael@0 | 59 | vqadd.s16 d11, d7, d8 ;d1 |
| michael@0 | 60 | |
| michael@0 | 61 | vqadd.s16 d2, d12, d11 |
| michael@0 | 62 | vqadd.s16 d3, d13, d10 |
| michael@0 | 63 | vqsub.s16 d4, d13, d10 |
| michael@0 | 64 | vqsub.s16 d5, d12, d11 |
| michael@0 | 65 | |
| michael@0 | 66 | vtrn.32 d2, d4 |
| michael@0 | 67 | vtrn.32 d3, d5 |
| michael@0 | 68 | vtrn.16 d2, d3 |
| michael@0 | 69 | vtrn.16 d4, d5 |
| michael@0 | 70 | |
| michael@0 | 71 | ; memset(input, 0, 32) -- 32bytes |
| michael@0 | 72 | vmov.i16 q14, #0 |
| michael@0 | 73 | |
| michael@0 | 74 | vswp d3, d4 |
| michael@0 | 75 | vqdmulh.s16 q3, q2, d0[2] |
| michael@0 | 76 | vqdmulh.s16 q4, q2, d0[0] |
| michael@0 | 77 | |
| michael@0 | 78 | vqadd.s16 d12, d2, d3 ;a1 |
| michael@0 | 79 | vqsub.s16 d13, d2, d3 ;b1 |
| michael@0 | 80 | |
| michael@0 | 81 | vmov q15, q14 |
| michael@0 | 82 | |
| michael@0 | 83 | vshr.s16 q3, q3, #1 |
| michael@0 | 84 | vshr.s16 q4, q4, #1 |
| michael@0 | 85 | |
| michael@0 | 86 | vqadd.s16 q3, q3, q2 |
| michael@0 | 87 | vqadd.s16 q4, q4, q2 |
| michael@0 | 88 | |
| michael@0 | 89 | vqsub.s16 d10, d6, d9 ;c1 |
| michael@0 | 90 | vqadd.s16 d11, d7, d8 ;d1 |
| michael@0 | 91 | |
| michael@0 | 92 | vqadd.s16 d2, d12, d11 |
| michael@0 | 93 | vqadd.s16 d3, d13, d10 |
| michael@0 | 94 | vqsub.s16 d4, d13, d10 |
| michael@0 | 95 | vqsub.s16 d5, d12, d11 |
| michael@0 | 96 | |
| michael@0 | 97 | vst1.16 {q14, q15}, [r0] |
| michael@0 | 98 | |
| michael@0 | 99 | vrshr.s16 d2, d2, #3 |
| michael@0 | 100 | vrshr.s16 d3, d3, #3 |
| michael@0 | 101 | vrshr.s16 d4, d4, #3 |
| michael@0 | 102 | vrshr.s16 d5, d5, #3 |
| michael@0 | 103 | |
| michael@0 | 104 | vtrn.32 d2, d4 |
| michael@0 | 105 | vtrn.32 d3, d5 |
| michael@0 | 106 | vtrn.16 d2, d3 |
| michael@0 | 107 | vtrn.16 d4, d5 |
| michael@0 | 108 | |
| michael@0 | 109 | vaddw.u8 q1, q1, d14 |
| michael@0 | 110 | vaddw.u8 q2, q2, d15 |
| michael@0 | 111 | |
| michael@0 | 112 | sub r2, r2, r3 |
| michael@0 | 113 | sub r1, r1, r3 |
| michael@0 | 114 | |
| michael@0 | 115 | vqmovun.s16 d0, q1 |
| michael@0 | 116 | vqmovun.s16 d1, q2 |
| michael@0 | 117 | |
| michael@0 | 118 | vst1.32 {d0[0]}, [r2], r3 |
| michael@0 | 119 | vst1.32 {d0[1]}, [r1], r3 |
| michael@0 | 120 | vst1.32 {d1[0]}, [r2] |
| michael@0 | 121 | vst1.32 {d1[1]}, [r1] |
| michael@0 | 122 | |
| michael@0 | 123 | bx lr |
| michael@0 | 124 | |
| michael@0 | 125 | ENDP ; |vp8_dequant_idct_add_neon| |
| michael@0 | 126 | |
| michael@0 | 127 | ; Constant Pool |
| michael@0 | 128 | cospi8sqrt2minus1 DCD 0x4e7b4e7b |
| michael@0 | 129 | sinpi8sqrt2 DCD 0x8a8c8a8c |
| michael@0 | 130 | |
| michael@0 | 131 | END |
