/src/aom/aom_dsp/x86/quantize_x86.h
Line | Count | Source |
1 | | /* |
2 | | * Copyright (c) 2018, Alliance for Open Media. All rights reserved. |
3 | | * |
4 | | * This source code is subject to the terms of the BSD 2 Clause License and |
5 | | * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
6 | | * was not distributed with this source code in the LICENSE file, you can |
7 | | * obtain it at www.aomedia.org/license/software. If the Alliance for Open |
8 | | * Media Patent License 1.0 was not distributed with this source code in the |
9 | | * PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
10 | | */ |
11 | | |
12 | | #include <assert.h> |
13 | | #include <emmintrin.h> |
14 | | |
15 | | #include "aom/aom_integer.h" |
16 | | |
17 | | static inline void load_b_values(const int16_t *zbin_ptr, __m128i *zbin, |
18 | | const int16_t *round_ptr, __m128i *round, |
19 | | const int16_t *quant_ptr, __m128i *quant, |
20 | | const int16_t *dequant_ptr, __m128i *dequant, |
21 | 0 | const int16_t *shift_ptr, __m128i *shift) { |
22 | 0 | *zbin = _mm_load_si128((const __m128i *)zbin_ptr); |
23 | 0 | *round = _mm_load_si128((const __m128i *)round_ptr); |
24 | 0 | *quant = _mm_load_si128((const __m128i *)quant_ptr); |
25 | 0 | *zbin = _mm_sub_epi16(*zbin, _mm_set1_epi16(1)); |
26 | 0 | *dequant = _mm_load_si128((const __m128i *)dequant_ptr); |
27 | 0 | *shift = _mm_load_si128((const __m128i *)shift_ptr); |
28 | 0 | } Unexecuted instantiation: quantize_sse2.c:load_b_values Unexecuted instantiation: adaptive_quantize_sse2.c:load_b_values Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:load_b_values Unexecuted instantiation: quantize_ssse3.c:load_b_values Unexecuted instantiation: aom_quantize_avx.c:load_b_values Unexecuted instantiation: adaptive_quantize_avx2.c:load_b_values Unexecuted instantiation: quantize_avx2.c:load_b_values Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:load_b_values Unexecuted instantiation: av1_quantize_sse2.c:load_b_values |
29 | | |
30 | | // With ssse3 and later abs() and sign() are preferred. |
31 | 0 | static inline __m128i invert_sign_sse2(__m128i a, __m128i sign) { |
32 | 0 | a = _mm_xor_si128(a, sign); |
33 | 0 | return _mm_sub_epi16(a, sign); |
34 | 0 | } Unexecuted instantiation: quantize_sse2.c:invert_sign_sse2 Unexecuted instantiation: adaptive_quantize_sse2.c:invert_sign_sse2 Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:invert_sign_sse2 Unexecuted instantiation: quantize_ssse3.c:invert_sign_sse2 Unexecuted instantiation: aom_quantize_avx.c:invert_sign_sse2 Unexecuted instantiation: adaptive_quantize_avx2.c:invert_sign_sse2 Unexecuted instantiation: quantize_avx2.c:invert_sign_sse2 Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:invert_sign_sse2 Unexecuted instantiation: av1_quantize_sse2.c:invert_sign_sse2 |
35 | | |
36 | 0 | static inline __m128i invert_sign_32_sse2(__m128i a, __m128i sign) { |
37 | 0 | a = _mm_xor_si128(a, sign); |
38 | 0 | return _mm_sub_epi32(a, sign); |
39 | 0 | } Unexecuted instantiation: quantize_sse2.c:invert_sign_32_sse2 Unexecuted instantiation: adaptive_quantize_sse2.c:invert_sign_32_sse2 Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:invert_sign_32_sse2 Unexecuted instantiation: quantize_ssse3.c:invert_sign_32_sse2 Unexecuted instantiation: aom_quantize_avx.c:invert_sign_32_sse2 Unexecuted instantiation: adaptive_quantize_avx2.c:invert_sign_32_sse2 Unexecuted instantiation: quantize_avx2.c:invert_sign_32_sse2 Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:invert_sign_32_sse2 Unexecuted instantiation: av1_quantize_sse2.c:invert_sign_32_sse2 |
40 | | |
41 | | static inline void calculate_qcoeff(__m128i *coeff, const __m128i round, |
42 | 0 | const __m128i quant, const __m128i shift) { |
43 | 0 | __m128i tmp, qcoeff; |
44 | 0 | qcoeff = _mm_adds_epi16(*coeff, round); |
45 | 0 | tmp = _mm_mulhi_epi16(qcoeff, quant); |
46 | 0 | qcoeff = _mm_add_epi16(tmp, qcoeff); |
47 | 0 | *coeff = _mm_mulhi_epi16(qcoeff, shift); |
48 | 0 | } Unexecuted instantiation: quantize_sse2.c:calculate_qcoeff Unexecuted instantiation: adaptive_quantize_sse2.c:calculate_qcoeff Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:calculate_qcoeff Unexecuted instantiation: quantize_ssse3.c:calculate_qcoeff Unexecuted instantiation: aom_quantize_avx.c:calculate_qcoeff Unexecuted instantiation: adaptive_quantize_avx2.c:calculate_qcoeff Unexecuted instantiation: quantize_avx2.c:calculate_qcoeff Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:calculate_qcoeff Unexecuted instantiation: av1_quantize_sse2.c:calculate_qcoeff |
49 | | |
50 | | static inline void calculate_qcoeff_log_scale(__m128i *coeff, |
51 | | const __m128i round, |
52 | | const __m128i quant, |
53 | | const __m128i *shift, |
54 | 0 | const int *log_scale) { |
55 | 0 | __m128i tmp, tmp1, qcoeff; |
56 | 0 | qcoeff = _mm_adds_epi16(*coeff, round); |
57 | 0 | tmp = _mm_mulhi_epi16(qcoeff, quant); |
58 | 0 | qcoeff = _mm_add_epi16(tmp, qcoeff); |
59 | 0 | tmp = _mm_mullo_epi16(qcoeff, *shift); |
60 | 0 | tmp = _mm_srli_epi16(tmp, (16 - *log_scale)); |
61 | 0 | tmp1 = _mm_mulhi_epi16(qcoeff, *shift); |
62 | 0 | tmp1 = _mm_slli_epi16(tmp1, *log_scale); |
63 | 0 | *coeff = _mm_or_si128(tmp, tmp1); |
64 | 0 | } Unexecuted instantiation: quantize_sse2.c:calculate_qcoeff_log_scale Unexecuted instantiation: adaptive_quantize_sse2.c:calculate_qcoeff_log_scale Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:calculate_qcoeff_log_scale Unexecuted instantiation: quantize_ssse3.c:calculate_qcoeff_log_scale Unexecuted instantiation: aom_quantize_avx.c:calculate_qcoeff_log_scale Unexecuted instantiation: adaptive_quantize_avx2.c:calculate_qcoeff_log_scale Unexecuted instantiation: quantize_avx2.c:calculate_qcoeff_log_scale Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:calculate_qcoeff_log_scale Unexecuted instantiation: av1_quantize_sse2.c:calculate_qcoeff_log_scale |
65 | | |
66 | 0 | static inline __m128i calculate_dqcoeff(__m128i qcoeff, __m128i dequant) { |
67 | 0 | return _mm_mullo_epi16(qcoeff, dequant); |
68 | 0 | } Unexecuted instantiation: quantize_sse2.c:calculate_dqcoeff Unexecuted instantiation: adaptive_quantize_sse2.c:calculate_dqcoeff Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:calculate_dqcoeff Unexecuted instantiation: quantize_ssse3.c:calculate_dqcoeff Unexecuted instantiation: aom_quantize_avx.c:calculate_dqcoeff Unexecuted instantiation: adaptive_quantize_avx2.c:calculate_dqcoeff Unexecuted instantiation: quantize_avx2.c:calculate_dqcoeff Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:calculate_dqcoeff Unexecuted instantiation: av1_quantize_sse2.c:calculate_dqcoeff |
69 | | |
70 | | static inline void calculate_dqcoeff_and_store_log_scale(__m128i qcoeff, |
71 | | __m128i dequant, |
72 | | const __m128i zero, |
73 | | tran_low_t *dqcoeff, |
74 | 0 | const int *log_scale) { |
75 | | // calculate abs |
76 | 0 | __m128i coeff_sign = _mm_srai_epi16(qcoeff, 15); |
77 | 0 | __m128i coeff = invert_sign_sse2(qcoeff, coeff_sign); |
78 | |
|
79 | 0 | const __m128i sign_0 = _mm_unpacklo_epi16(coeff_sign, zero); |
80 | 0 | const __m128i sign_1 = _mm_unpackhi_epi16(coeff_sign, zero); |
81 | |
|
82 | 0 | const __m128i low = _mm_mullo_epi16(coeff, dequant); |
83 | 0 | const __m128i high = _mm_mulhi_epi16(coeff, dequant); |
84 | 0 | __m128i dqcoeff32_0 = _mm_unpacklo_epi16(low, high); |
85 | 0 | __m128i dqcoeff32_1 = _mm_unpackhi_epi16(low, high); |
86 | |
|
87 | 0 | dqcoeff32_0 = _mm_srli_epi32(dqcoeff32_0, *log_scale); |
88 | 0 | dqcoeff32_1 = _mm_srli_epi32(dqcoeff32_1, *log_scale); |
89 | |
|
90 | 0 | dqcoeff32_0 = invert_sign_32_sse2(dqcoeff32_0, sign_0); |
91 | 0 | dqcoeff32_1 = invert_sign_32_sse2(dqcoeff32_1, sign_1); |
92 | |
|
93 | 0 | _mm_store_si128((__m128i *)(dqcoeff), dqcoeff32_0); |
94 | 0 | _mm_store_si128((__m128i *)(dqcoeff + 4), dqcoeff32_1); |
95 | 0 | } Unexecuted instantiation: quantize_sse2.c:calculate_dqcoeff_and_store_log_scale Unexecuted instantiation: adaptive_quantize_sse2.c:calculate_dqcoeff_and_store_log_scale Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:calculate_dqcoeff_and_store_log_scale Unexecuted instantiation: quantize_ssse3.c:calculate_dqcoeff_and_store_log_scale Unexecuted instantiation: aom_quantize_avx.c:calculate_dqcoeff_and_store_log_scale Unexecuted instantiation: adaptive_quantize_avx2.c:calculate_dqcoeff_and_store_log_scale Unexecuted instantiation: quantize_avx2.c:calculate_dqcoeff_and_store_log_scale Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:calculate_dqcoeff_and_store_log_scale Unexecuted instantiation: av1_quantize_sse2.c:calculate_dqcoeff_and_store_log_scale |
96 | | |
97 | | // Scan 16 values for eob reference in scan_ptr. Use masks (-1) from comparing |
98 | | // to zbin to add 1 to the index in 'scan'. |
99 | | static inline __m128i scan_for_eob(__m128i *coeff0, __m128i *coeff1, |
100 | | const __m128i zbin_mask0, |
101 | | const __m128i zbin_mask1, |
102 | | const int16_t *scan_ptr, const int index, |
103 | 0 | const __m128i zero) { |
104 | 0 | const __m128i zero_coeff0 = _mm_cmpeq_epi16(*coeff0, zero); |
105 | 0 | const __m128i zero_coeff1 = _mm_cmpeq_epi16(*coeff1, zero); |
106 | 0 | __m128i scan0 = _mm_load_si128((const __m128i *)(scan_ptr + index)); |
107 | 0 | __m128i scan1 = _mm_load_si128((const __m128i *)(scan_ptr + index + 8)); |
108 | 0 | __m128i eob0, eob1; |
109 | | // Add one to convert from indices to counts |
110 | 0 | scan0 = _mm_sub_epi16(scan0, zbin_mask0); |
111 | 0 | scan1 = _mm_sub_epi16(scan1, zbin_mask1); |
112 | 0 | eob0 = _mm_andnot_si128(zero_coeff0, scan0); |
113 | 0 | eob1 = _mm_andnot_si128(zero_coeff1, scan1); |
114 | 0 | return _mm_max_epi16(eob0, eob1); |
115 | 0 | } Unexecuted instantiation: quantize_sse2.c:scan_for_eob Unexecuted instantiation: adaptive_quantize_sse2.c:scan_for_eob Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:scan_for_eob Unexecuted instantiation: quantize_ssse3.c:scan_for_eob Unexecuted instantiation: aom_quantize_avx.c:scan_for_eob Unexecuted instantiation: adaptive_quantize_avx2.c:scan_for_eob Unexecuted instantiation: quantize_avx2.c:scan_for_eob Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:scan_for_eob Unexecuted instantiation: av1_quantize_sse2.c:scan_for_eob |
116 | | |
117 | 0 | static inline int16_t accumulate_eob(__m128i eob) { |
118 | 0 | __m128i eob_shuffled; |
119 | 0 | eob_shuffled = _mm_shuffle_epi32(eob, 0xe); |
120 | 0 | eob = _mm_max_epi16(eob, eob_shuffled); |
121 | 0 | eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); |
122 | 0 | eob = _mm_max_epi16(eob, eob_shuffled); |
123 | 0 | eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); |
124 | 0 | eob = _mm_max_epi16(eob, eob_shuffled); |
125 | 0 | return _mm_extract_epi16(eob, 1); |
126 | 0 | } Unexecuted instantiation: quantize_sse2.c:accumulate_eob Unexecuted instantiation: adaptive_quantize_sse2.c:accumulate_eob Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:accumulate_eob Unexecuted instantiation: quantize_ssse3.c:accumulate_eob Unexecuted instantiation: aom_quantize_avx.c:accumulate_eob Unexecuted instantiation: adaptive_quantize_avx2.c:accumulate_eob Unexecuted instantiation: quantize_avx2.c:accumulate_eob Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:accumulate_eob Unexecuted instantiation: av1_quantize_sse2.c:accumulate_eob |
127 | | |
128 | 0 | static inline __m128i load_coefficients(const tran_low_t *coeff_ptr) { |
129 | 0 | static_assert(sizeof(tran_low_t) == 4, ""); |
130 | 0 | const __m128i coeff1 = _mm_load_si128((__m128i *)(coeff_ptr)); |
131 | 0 | const __m128i coeff2 = _mm_load_si128((__m128i *)(coeff_ptr + 4)); |
132 | 0 | return _mm_packs_epi32(coeff1, coeff2); |
133 | 0 | } Unexecuted instantiation: quantize_sse2.c:load_coefficients Unexecuted instantiation: adaptive_quantize_sse2.c:load_coefficients Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:load_coefficients Unexecuted instantiation: quantize_ssse3.c:load_coefficients Unexecuted instantiation: aom_quantize_avx.c:load_coefficients Unexecuted instantiation: adaptive_quantize_avx2.c:load_coefficients Unexecuted instantiation: quantize_avx2.c:load_coefficients Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:load_coefficients Unexecuted instantiation: av1_quantize_sse2.c:load_coefficients |
134 | | |
135 | | static inline void store_coefficients(__m128i coeff_vals, |
136 | 0 | tran_low_t *coeff_ptr) { |
137 | 0 | static_assert(sizeof(tran_low_t) == 4, ""); |
138 | |
|
139 | 0 | __m128i one = _mm_set1_epi16(1); |
140 | 0 | __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one); |
141 | 0 | __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one); |
142 | 0 | __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi); |
143 | 0 | __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi); |
144 | 0 | _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1); |
145 | 0 | _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2); |
146 | 0 | } Unexecuted instantiation: quantize_sse2.c:store_coefficients Unexecuted instantiation: adaptive_quantize_sse2.c:store_coefficients Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:store_coefficients Unexecuted instantiation: quantize_ssse3.c:store_coefficients Unexecuted instantiation: aom_quantize_avx.c:store_coefficients Unexecuted instantiation: adaptive_quantize_avx2.c:store_coefficients Unexecuted instantiation: quantize_avx2.c:store_coefficients Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:store_coefficients Unexecuted instantiation: av1_quantize_sse2.c:store_coefficients |
147 | | |
148 | | static inline void update_mask1(__m128i *cmp_mask0, __m128i *cmp_mask1, |
149 | | const int16_t *iscan_ptr, int *is_found, |
150 | 0 | __m128i *mask) { |
151 | 0 | __m128i all_zero; |
152 | 0 | __m128i temp_mask = _mm_setzero_si128(); |
153 | 0 | all_zero = _mm_or_si128(*cmp_mask0, *cmp_mask1); |
154 | 0 | if (_mm_movemask_epi8(all_zero)) { |
155 | 0 | __m128i iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr)); |
156 | 0 | __m128i mask0 = _mm_and_si128(*cmp_mask0, iscan0); |
157 | 0 | __m128i iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8)); |
158 | 0 | __m128i mask1 = _mm_and_si128(*cmp_mask1, iscan1); |
159 | 0 | temp_mask = _mm_max_epi16(mask0, mask1); |
160 | 0 | *is_found = 1; |
161 | 0 | } |
162 | 0 | *mask = _mm_max_epi16(temp_mask, *mask); |
163 | 0 | } Unexecuted instantiation: quantize_sse2.c:update_mask1 Unexecuted instantiation: adaptive_quantize_sse2.c:update_mask1 Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:update_mask1 Unexecuted instantiation: quantize_ssse3.c:update_mask1 Unexecuted instantiation: aom_quantize_avx.c:update_mask1 Unexecuted instantiation: adaptive_quantize_avx2.c:update_mask1 Unexecuted instantiation: quantize_avx2.c:update_mask1 Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:update_mask1 Unexecuted instantiation: av1_quantize_sse2.c:update_mask1 |
164 | | |
165 | | static inline void update_mask0(__m128i *qcoeff0, __m128i *qcoeff1, |
166 | | __m128i *threshold, const int16_t *iscan_ptr, |
167 | 0 | int *is_found, __m128i *mask) { |
168 | 0 | __m128i zero = _mm_setzero_si128(); |
169 | 0 | __m128i coeff[4], cmp_mask0, cmp_mask1, cmp_mask2, cmp_mask3; |
170 | |
|
171 | 0 | coeff[0] = _mm_unpacklo_epi16(*qcoeff0, zero); |
172 | 0 | coeff[1] = _mm_unpackhi_epi16(*qcoeff0, zero); |
173 | 0 | coeff[2] = _mm_unpacklo_epi16(*qcoeff1, zero); |
174 | 0 | coeff[3] = _mm_unpackhi_epi16(*qcoeff1, zero); |
175 | |
|
176 | 0 | coeff[0] = _mm_slli_epi32(coeff[0], AOM_QM_BITS); |
177 | 0 | cmp_mask0 = _mm_cmpgt_epi32(coeff[0], threshold[0]); |
178 | 0 | coeff[1] = _mm_slli_epi32(coeff[1], AOM_QM_BITS); |
179 | 0 | cmp_mask1 = _mm_cmpgt_epi32(coeff[1], threshold[1]); |
180 | 0 | coeff[2] = _mm_slli_epi32(coeff[2], AOM_QM_BITS); |
181 | 0 | cmp_mask2 = _mm_cmpgt_epi32(coeff[2], threshold[1]); |
182 | 0 | coeff[3] = _mm_slli_epi32(coeff[3], AOM_QM_BITS); |
183 | 0 | cmp_mask3 = _mm_cmpgt_epi32(coeff[3], threshold[1]); |
184 | |
|
185 | 0 | cmp_mask0 = _mm_packs_epi32(cmp_mask0, cmp_mask1); |
186 | 0 | cmp_mask1 = _mm_packs_epi32(cmp_mask2, cmp_mask3); |
187 | |
|
188 | 0 | update_mask1(&cmp_mask0, &cmp_mask1, iscan_ptr, is_found, mask); |
189 | 0 | } Unexecuted instantiation: quantize_sse2.c:update_mask0 Unexecuted instantiation: adaptive_quantize_sse2.c:update_mask0 Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:update_mask0 Unexecuted instantiation: quantize_ssse3.c:update_mask0 Unexecuted instantiation: aom_quantize_avx.c:update_mask0 Unexecuted instantiation: adaptive_quantize_avx2.c:update_mask0 Unexecuted instantiation: quantize_avx2.c:update_mask0 Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:update_mask0 Unexecuted instantiation: av1_quantize_sse2.c:update_mask0 |
190 | | |
191 | 0 | static inline int calculate_non_zero_count(__m128i mask) { |
192 | 0 | __m128i mask0, mask1; |
193 | 0 | int non_zero_count = 0; |
194 | 0 | mask0 = _mm_unpackhi_epi64(mask, mask); |
195 | 0 | mask1 = _mm_max_epi16(mask0, mask); |
196 | 0 | mask0 = _mm_shuffle_epi32(mask1, 1); |
197 | 0 | mask0 = _mm_max_epi16(mask0, mask1); |
198 | 0 | mask1 = _mm_srli_epi32(mask0, 16); |
199 | 0 | mask0 = _mm_max_epi16(mask0, mask1); |
200 | 0 | non_zero_count = _mm_extract_epi16(mask0, 0) + 1; |
201 | |
|
202 | 0 | return non_zero_count; |
203 | 0 | } Unexecuted instantiation: quantize_sse2.c:calculate_non_zero_count Unexecuted instantiation: adaptive_quantize_sse2.c:calculate_non_zero_count Unexecuted instantiation: highbd_adaptive_quantize_sse2.c:calculate_non_zero_count Unexecuted instantiation: quantize_ssse3.c:calculate_non_zero_count Unexecuted instantiation: aom_quantize_avx.c:calculate_non_zero_count Unexecuted instantiation: adaptive_quantize_avx2.c:calculate_non_zero_count Unexecuted instantiation: quantize_avx2.c:calculate_non_zero_count Unexecuted instantiation: highbd_adaptive_quantize_avx2.c:calculate_non_zero_count Unexecuted instantiation: av1_quantize_sse2.c:calculate_non_zero_count |