media/libvpx/vp9/encoder/x86/vp9_error_sse2.asm@ac0c01689b40 (annotated)
media/libvpx/vp9/encoder/x86/vp9_error_sse2.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 | %include "third_party/x86inc/x86inc.asm" |
| michael@0 | 12 | |
| michael@0 | 13 | SECTION .text |
| michael@0 | 14 | |
| michael@0 | 15 | ; int64_t vp9_block_error(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size, |
| michael@0 | 16 | ; int64_t *ssz) |
| michael@0 | 17 | |
| michael@0 | 18 | INIT_XMM sse2 |
| michael@0 | 19 | cglobal block_error, 3, 3, 8, uqc, dqc, size, ssz |
| michael@0 | 20 | pxor m4, m4 ; sse accumulator |
| michael@0 | 21 | pxor m6, m6 ; ssz accumulator |
| michael@0 | 22 | pxor m5, m5 ; dedicated zero register |
| michael@0 | 23 | lea uqcq, [uqcq+sizeq*2] |
| michael@0 | 24 | lea dqcq, [dqcq+sizeq*2] |
| michael@0 | 25 | neg sizeq |
| michael@0 | 26 | .loop: |
| michael@0 | 27 | mova m2, [uqcq+sizeq*2] |
| michael@0 | 28 | mova m0, [dqcq+sizeq*2] |
| michael@0 | 29 | mova m3, [uqcq+sizeq*2+mmsize] |
| michael@0 | 30 | mova m1, [dqcq+sizeq*2+mmsize] |
| michael@0 | 31 | psubw m0, m2 |
| michael@0 | 32 | psubw m1, m3 |
| michael@0 | 33 | ; individual errors are max. 15bit+sign, so squares are 30bit, and |
| michael@0 | 34 | ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit) |
| michael@0 | 35 | pmaddwd m0, m0 |
| michael@0 | 36 | pmaddwd m1, m1 |
| michael@0 | 37 | pmaddwd m2, m2 |
| michael@0 | 38 | pmaddwd m3, m3 |
| michael@0 | 39 | ; accumulate in 64bit |
| michael@0 | 40 | punpckldq m7, m0, m5 |
| michael@0 | 41 | punpckhdq m0, m5 |
| michael@0 | 42 | paddq m4, m7 |
| michael@0 | 43 | punpckldq m7, m1, m5 |
| michael@0 | 44 | paddq m4, m0 |
| michael@0 | 45 | punpckhdq m1, m5 |
| michael@0 | 46 | paddq m4, m7 |
| michael@0 | 47 | punpckldq m7, m2, m5 |
| michael@0 | 48 | paddq m4, m1 |
| michael@0 | 49 | punpckhdq m2, m5 |
| michael@0 | 50 | paddq m6, m7 |
| michael@0 | 51 | punpckldq m7, m3, m5 |
| michael@0 | 52 | paddq m6, m2 |
| michael@0 | 53 | punpckhdq m3, m5 |
| michael@0 | 54 | paddq m6, m7 |
| michael@0 | 55 | paddq m6, m3 |
| michael@0 | 56 | add sizeq, mmsize |
| michael@0 | 57 | jl .loop |
| michael@0 | 58 | |
| michael@0 | 59 | ; accumulate horizontally and store in return value |
| michael@0 | 60 | movhlps m5, m4 |
| michael@0 | 61 | movhlps m7, m6 |
| michael@0 | 62 | paddq m4, m5 |
| michael@0 | 63 | paddq m6, m7 |
| michael@0 | 64 | %if ARCH_X86_64 |
| michael@0 | 65 | movq rax, m4 |
| michael@0 | 66 | movq [sszq], m6 |
| michael@0 | 67 | %else |
| michael@0 | 68 | mov eax, sszm |
| michael@0 | 69 | pshufd m5, m4, 0x1 |
| michael@0 | 70 | movq [eax], m6 |
| michael@0 | 71 | movd eax, m4 |
| michael@0 | 72 | movd edx, m5 |
| michael@0 | 73 | %endif |
| michael@0 | 74 | RET |
