media/libvpx/vp8/encoder/x86/denoising_sse2.c@ac0c01689b40 (annotated)
media/libvpx/vp8/encoder/x86/denoising_sse2.c
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) 2012 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 "vp8/encoder/denoising.h" |
| michael@0 | 12 | #include "vp8/common/reconinter.h" |
| michael@0 | 13 | #include "vpx/vpx_integer.h" |
| michael@0 | 14 | #include "vpx_mem/vpx_mem.h" |
| michael@0 | 15 | #include "vp8_rtcd.h" |
| michael@0 | 16 | |
| michael@0 | 17 | #include <emmintrin.h> |
| michael@0 | 18 | #include "vpx_ports/emmintrin_compat.h" |
| michael@0 | 19 | |
| michael@0 | 20 | union sum_union { |
| michael@0 | 21 | __m128i v; |
| michael@0 | 22 | signed char e[16]; |
| michael@0 | 23 | }; |
| michael@0 | 24 | |
| michael@0 | 25 | int vp8_denoiser_filter_sse2(YV12_BUFFER_CONFIG *mc_running_avg, |
| michael@0 | 26 | YV12_BUFFER_CONFIG *running_avg, |
| michael@0 | 27 | MACROBLOCK *signal, unsigned int motion_magnitude, |
| michael@0 | 28 | int y_offset, int uv_offset) |
| michael@0 | 29 | { |
| michael@0 | 30 | unsigned char *sig = signal->thismb; |
| michael@0 | 31 | int sig_stride = 16; |
| michael@0 | 32 | unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset; |
| michael@0 | 33 | int mc_avg_y_stride = mc_running_avg->y_stride; |
| michael@0 | 34 | unsigned char *running_avg_y = running_avg->y_buffer + y_offset; |
| michael@0 | 35 | int avg_y_stride = running_avg->y_stride; |
| michael@0 | 36 | int r; |
| michael@0 | 37 | __m128i acc_diff = _mm_setzero_si128(); |
| michael@0 | 38 | const __m128i k_0 = _mm_setzero_si128(); |
| michael@0 | 39 | const __m128i k_4 = _mm_set1_epi8(4); |
| michael@0 | 40 | const __m128i k_8 = _mm_set1_epi8(8); |
| michael@0 | 41 | const __m128i k_16 = _mm_set1_epi8(16); |
| michael@0 | 42 | /* Modify each level's adjustment according to motion_magnitude. */ |
| michael@0 | 43 | const __m128i l3 = _mm_set1_epi8( |
| michael@0 | 44 | (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 : 6); |
| michael@0 | 45 | /* Difference between level 3 and level 2 is 2. */ |
| michael@0 | 46 | const __m128i l32 = _mm_set1_epi8(2); |
| michael@0 | 47 | /* Difference between level 2 and level 1 is 1. */ |
| michael@0 | 48 | const __m128i l21 = _mm_set1_epi8(1); |
| michael@0 | 49 | |
| michael@0 | 50 | for (r = 0; r < 16; ++r) |
| michael@0 | 51 | { |
| michael@0 | 52 | /* Calculate differences */ |
| michael@0 | 53 | const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0])); |
| michael@0 | 54 | const __m128i v_mc_running_avg_y = _mm_loadu_si128( |
| michael@0 | 55 | (__m128i *)(&mc_running_avg_y[0])); |
| michael@0 | 56 | __m128i v_running_avg_y; |
| michael@0 | 57 | const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig); |
| michael@0 | 58 | const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y); |
| michael@0 | 59 | /* Obtain the sign. FF if diff is negative. */ |
| michael@0 | 60 | const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0); |
| michael@0 | 61 | /* Clamp absolute difference to 16 to be used to get mask. Doing this |
| michael@0 | 62 | * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */ |
| michael@0 | 63 | const __m128i clamped_absdiff = _mm_min_epu8( |
| michael@0 | 64 | _mm_or_si128(pdiff, ndiff), k_16); |
| michael@0 | 65 | /* Get masks for l2 l1 and l0 adjustments */ |
| michael@0 | 66 | const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff); |
| michael@0 | 67 | const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff); |
| michael@0 | 68 | const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff); |
| michael@0 | 69 | /* Get adjustments for l2, l1, and l0 */ |
| michael@0 | 70 | __m128i adj2 = _mm_and_si128(mask2, l32); |
| michael@0 | 71 | const __m128i adj1 = _mm_and_si128(mask1, l21); |
| michael@0 | 72 | const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff); |
| michael@0 | 73 | __m128i adj, padj, nadj; |
| michael@0 | 74 | |
| michael@0 | 75 | /* Combine the adjustments and get absolute adjustments. */ |
| michael@0 | 76 | adj2 = _mm_add_epi8(adj2, adj1); |
| michael@0 | 77 | adj = _mm_sub_epi8(l3, adj2); |
| michael@0 | 78 | adj = _mm_andnot_si128(mask0, adj); |
| michael@0 | 79 | adj = _mm_or_si128(adj, adj0); |
| michael@0 | 80 | |
| michael@0 | 81 | /* Restore the sign and get positive and negative adjustments. */ |
| michael@0 | 82 | padj = _mm_andnot_si128(diff_sign, adj); |
| michael@0 | 83 | nadj = _mm_and_si128(diff_sign, adj); |
| michael@0 | 84 | |
| michael@0 | 85 | /* Calculate filtered value. */ |
| michael@0 | 86 | v_running_avg_y = _mm_adds_epu8(v_sig, padj); |
| michael@0 | 87 | v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj); |
| michael@0 | 88 | _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y); |
| michael@0 | 89 | |
| michael@0 | 90 | /* Adjustments <=7, and each element in acc_diff can fit in signed |
| michael@0 | 91 | * char. |
| michael@0 | 92 | */ |
| michael@0 | 93 | acc_diff = _mm_adds_epi8(acc_diff, padj); |
| michael@0 | 94 | acc_diff = _mm_subs_epi8(acc_diff, nadj); |
| michael@0 | 95 | |
| michael@0 | 96 | /* Update pointers for next iteration. */ |
| michael@0 | 97 | sig += sig_stride; |
| michael@0 | 98 | mc_running_avg_y += mc_avg_y_stride; |
| michael@0 | 99 | running_avg_y += avg_y_stride; |
| michael@0 | 100 | } |
| michael@0 | 101 | |
| michael@0 | 102 | { |
| michael@0 | 103 | /* Compute the sum of all pixel differences of this MB. */ |
| michael@0 | 104 | union sum_union s; |
| michael@0 | 105 | int sum_diff = 0; |
| michael@0 | 106 | s.v = acc_diff; |
| michael@0 | 107 | sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5] |
| michael@0 | 108 | + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11] |
| michael@0 | 109 | + s.e[12] + s.e[13] + s.e[14] + s.e[15]; |
| michael@0 | 110 | |
| michael@0 | 111 | if (abs(sum_diff) > SUM_DIFF_THRESHOLD) |
| michael@0 | 112 | { |
| michael@0 | 113 | return COPY_BLOCK; |
| michael@0 | 114 | } |
| michael@0 | 115 | } |
| michael@0 | 116 | |
| michael@0 | 117 | vp8_copy_mem16x16(running_avg->y_buffer + y_offset, avg_y_stride, |
| michael@0 | 118 | signal->thismb, sig_stride); |
| michael@0 | 119 | return FILTER_BLOCK; |
| michael@0 | 120 | } |
