The Tor Browser / file comparison
comparison: media/libvpx/vp8/encoder/x86/denoising_sse2.c
media/libvpx/vp8/encoder/x86/denoising_sse2.c
- branch
- TOR_BUG_9701
- changeset 10
- ac0c01689b40
equal
deleted
inserted
replaced
| -1:000000000000 | 0:81aea887062d |
|---|---|
| 1 /* | |
| 2 * Copyright (c) 2012 The WebM project authors. All Rights Reserved. | |
| 3 * | |
| 4 * Use of this source code is governed by a BSD-style license | |
| 5 * that can be found in the LICENSE file in the root of the source | |
| 6 * tree. An additional intellectual property rights grant can be found | |
| 7 * in the file PATENTS. All contributing project authors may | |
| 8 * be found in the AUTHORS file in the root of the source tree. | |
| 9 */ | |
| 10 | |
| 11 #include "vp8/encoder/denoising.h" | |
| 12 #include "vp8/common/reconinter.h" | |
| 13 #include "vpx/vpx_integer.h" | |
| 14 #include "vpx_mem/vpx_mem.h" | |
| 15 #include "vp8_rtcd.h" | |
| 16 | |
| 17 #include <emmintrin.h> | |
| 18 #include "vpx_ports/emmintrin_compat.h" | |
| 19 | |
| 20 union sum_union { | |
| 21 __m128i v; | |
| 22 signed char e[16]; | |
| 23 }; | |
| 24 | |
| 25 int vp8_denoiser_filter_sse2(YV12_BUFFER_CONFIG *mc_running_avg, | |
| 26 YV12_BUFFER_CONFIG *running_avg, | |
| 27 MACROBLOCK *signal, unsigned int motion_magnitude, | |
| 28 int y_offset, int uv_offset) | |
| 29 { | |
| 30 unsigned char *sig = signal->thismb; | |
| 31 int sig_stride = 16; | |
| 32 unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset; | |
| 33 int mc_avg_y_stride = mc_running_avg->y_stride; | |
| 34 unsigned char *running_avg_y = running_avg->y_buffer + y_offset; | |
| 35 int avg_y_stride = running_avg->y_stride; | |
| 36 int r; | |
| 37 __m128i acc_diff = _mm_setzero_si128(); | |
| 38 const __m128i k_0 = _mm_setzero_si128(); | |
| 39 const __m128i k_4 = _mm_set1_epi8(4); | |
| 40 const __m128i k_8 = _mm_set1_epi8(8); | |
| 41 const __m128i k_16 = _mm_set1_epi8(16); | |
| 42 /* Modify each level's adjustment according to motion_magnitude. */ | |
| 43 const __m128i l3 = _mm_set1_epi8( | |
| 44 (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 : 6); | |
| 45 /* Difference between level 3 and level 2 is 2. */ | |
| 46 const __m128i l32 = _mm_set1_epi8(2); | |
| 47 /* Difference between level 2 and level 1 is 1. */ | |
| 48 const __m128i l21 = _mm_set1_epi8(1); | |
| 49 | |
| 50 for (r = 0; r < 16; ++r) | |
| 51 { | |
| 52 /* Calculate differences */ | |
| 53 const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0])); | |
| 54 const __m128i v_mc_running_avg_y = _mm_loadu_si128( | |
| 55 (__m128i *)(&mc_running_avg_y[0])); | |
| 56 __m128i v_running_avg_y; | |
| 57 const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig); | |
| 58 const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y); | |
| 59 /* Obtain the sign. FF if diff is negative. */ | |
| 60 const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0); | |
| 61 /* Clamp absolute difference to 16 to be used to get mask. Doing this | |
| 62 * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */ | |
| 63 const __m128i clamped_absdiff = _mm_min_epu8( | |
| 64 _mm_or_si128(pdiff, ndiff), k_16); | |
| 65 /* Get masks for l2 l1 and l0 adjustments */ | |
| 66 const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff); | |
| 67 const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff); | |
| 68 const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff); | |
| 69 /* Get adjustments for l2, l1, and l0 */ | |
| 70 __m128i adj2 = _mm_and_si128(mask2, l32); | |
| 71 const __m128i adj1 = _mm_and_si128(mask1, l21); | |
| 72 const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff); | |
| 73 __m128i adj, padj, nadj; | |
| 74 | |
| 75 /* Combine the adjustments and get absolute adjustments. */ | |
| 76 adj2 = _mm_add_epi8(adj2, adj1); | |
| 77 adj = _mm_sub_epi8(l3, adj2); | |
| 78 adj = _mm_andnot_si128(mask0, adj); | |
| 79 adj = _mm_or_si128(adj, adj0); | |
| 80 | |
| 81 /* Restore the sign and get positive and negative adjustments. */ | |
| 82 padj = _mm_andnot_si128(diff_sign, adj); | |
| 83 nadj = _mm_and_si128(diff_sign, adj); | |
| 84 | |
| 85 /* Calculate filtered value. */ | |
| 86 v_running_avg_y = _mm_adds_epu8(v_sig, padj); | |
| 87 v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj); | |
| 88 _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y); | |
| 89 | |
| 90 /* Adjustments <=7, and each element in acc_diff can fit in signed | |
| 91 * char. | |
| 92 */ | |
| 93 acc_diff = _mm_adds_epi8(acc_diff, padj); | |
| 94 acc_diff = _mm_subs_epi8(acc_diff, nadj); | |
| 95 | |
| 96 /* Update pointers for next iteration. */ | |
| 97 sig += sig_stride; | |
| 98 mc_running_avg_y += mc_avg_y_stride; | |
| 99 running_avg_y += avg_y_stride; | |
| 100 } | |
| 101 | |
| 102 { | |
| 103 /* Compute the sum of all pixel differences of this MB. */ | |
| 104 union sum_union s; | |
| 105 int sum_diff = 0; | |
| 106 s.v = acc_diff; | |
| 107 sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5] | |
| 108 + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11] | |
| 109 + s.e[12] + s.e[13] + s.e[14] + s.e[15]; | |
| 110 | |
| 111 if (abs(sum_diff) > SUM_DIFF_THRESHOLD) | |
| 112 { | |
| 113 return COPY_BLOCK; | |
| 114 } | |
| 115 } | |
| 116 | |
| 117 vp8_copy_mem16x16(running_avg->y_buffer + y_offset, avg_y_stride, | |
| 118 signal->thismb, sig_stride); | |
| 119 return FILTER_BLOCK; | |
| 120 } |
