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 (void)uv_offset; 38 __m128i acc_diff = _mm_setzero_si128(); 39 const __m128i k_0 = _mm_setzero_si128(); 40 const __m128i k_4 = _mm_set1_epi8(4); 41 const __m128i k_8 = _mm_set1_epi8(8); 42 const __m128i k_16 = _mm_set1_epi8(16); 43 /* Modify each level's adjustment according to motion_magnitude. */ 44 const __m128i l3 = _mm_set1_epi8( 45 (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 : 6); 46 /* Difference between level 3 and level 2 is 2. */ 47 const __m128i l32 = _mm_set1_epi8(2); 48 /* Difference between level 2 and level 1 is 1. */ 49 const __m128i l21 = _mm_set1_epi8(1); 50 51 for (r = 0; r < 16; ++r) 52 { 53 /* Calculate differences */ 54 const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0])); 55 const __m128i v_mc_running_avg_y = _mm_loadu_si128( 56 (__m128i *)(&mc_running_avg_y[0])); 57 __m128i v_running_avg_y; 58 const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig); 59 const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y); 60 /* Obtain the sign. FF if diff is negative. */ 61 const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0); 62 /* Clamp absolute difference to 16 to be used to get mask. Doing this 63 * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */ 64 const __m128i clamped_absdiff = _mm_min_epu8( 65 _mm_or_si128(pdiff, ndiff), k_16); 66 /* Get masks for l2 l1 and l0 adjustments */ 67 const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff); 68 const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff); 69 const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff); 70 /* Get adjustments for l2, l1, and l0 */ 71 __m128i adj2 = _mm_and_si128(mask2, l32); 72 const __m128i adj1 = _mm_and_si128(mask1, l21); 73 const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff); 74 __m128i adj, padj, nadj; 75 76 /* Combine the adjustments and get absolute adjustments. */ 77 adj2 = _mm_add_epi8(adj2, adj1); 78 adj = _mm_sub_epi8(l3, adj2); 79 adj = _mm_andnot_si128(mask0, adj); 80 adj = _mm_or_si128(adj, adj0); 81 82 /* Restore the sign and get positive and negative adjustments. */ 83 padj = _mm_andnot_si128(diff_sign, adj); 84 nadj = _mm_and_si128(diff_sign, adj); 85 86 /* Calculate filtered value. */ 87 v_running_avg_y = _mm_adds_epu8(v_sig, padj); 88 v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj); 89 _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y); 90 91 /* Adjustments <=7, and each element in acc_diff can fit in signed 92 * char. 93 */ 94 acc_diff = _mm_adds_epi8(acc_diff, padj); 95 acc_diff = _mm_subs_epi8(acc_diff, nadj); 96 97 /* Update pointers for next iteration. */ 98 sig += sig_stride; 99 mc_running_avg_y += mc_avg_y_stride; 100 running_avg_y += avg_y_stride; 101 } 102 103 { 104 /* Compute the sum of all pixel differences of this MB. */ 105 union sum_union s; 106 int sum_diff = 0; 107 s.v = acc_diff; 108 sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5] 109 + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11] 110 + s.e[12] + s.e[13] + s.e[14] + s.e[15]; 111 112 if (abs(sum_diff) > SUM_DIFF_THRESHOLD) 113 { 114 return COPY_BLOCK; 115 } 116 } 117 118 vp8_copy_mem16x16(running_avg->y_buffer + y_offset, avg_y_stride, 119 signal->thismb, sig_stride); 120 return FILTER_BLOCK; 121 } 122
