/src/aom/av1/encoder/x86/av1_quantize_sse2.c
Line | Count | Source |
1 | | /* |
2 | | * Copyright (c) 2016, 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 <emmintrin.h> |
13 | | #include <xmmintrin.h> |
14 | | |
15 | | #include "config/av1_rtcd.h" |
16 | | |
17 | | #include "aom/aom_integer.h" |
18 | | #include "aom_dsp/x86/quantize_x86.h" |
19 | | |
20 | | static inline void read_coeff(const tran_low_t *coeff, intptr_t offset, |
21 | 0 | __m128i *c0, __m128i *c1) { |
22 | 0 | const tran_low_t *addr = coeff + offset; |
23 | 0 | if (sizeof(tran_low_t) == 4) { |
24 | 0 | const __m128i x0 = _mm_load_si128((const __m128i *)addr); |
25 | 0 | const __m128i x1 = _mm_load_si128((const __m128i *)addr + 1); |
26 | 0 | const __m128i x2 = _mm_load_si128((const __m128i *)addr + 2); |
27 | 0 | const __m128i x3 = _mm_load_si128((const __m128i *)addr + 3); |
28 | 0 | *c0 = _mm_packs_epi32(x0, x1); |
29 | 0 | *c1 = _mm_packs_epi32(x2, x3); |
30 | 0 | } else { |
31 | 0 | *c0 = _mm_load_si128((const __m128i *)addr); |
32 | 0 | *c1 = _mm_load_si128((const __m128i *)addr + 1); |
33 | 0 | } |
34 | 0 | } |
35 | | |
36 | | static inline void write_qcoeff(const __m128i *qc0, const __m128i *qc1, |
37 | 0 | tran_low_t *qcoeff, intptr_t offset) { |
38 | 0 | tran_low_t *addr = qcoeff + offset; |
39 | 0 | if (sizeof(tran_low_t) == 4) { |
40 | 0 | const __m128i zero = _mm_setzero_si128(); |
41 | 0 | __m128i sign_bits = _mm_cmplt_epi16(*qc0, zero); |
42 | 0 | __m128i y0 = _mm_unpacklo_epi16(*qc0, sign_bits); |
43 | 0 | __m128i y1 = _mm_unpackhi_epi16(*qc0, sign_bits); |
44 | 0 | _mm_store_si128((__m128i *)addr, y0); |
45 | 0 | _mm_store_si128((__m128i *)addr + 1, y1); |
46 | |
|
47 | 0 | sign_bits = _mm_cmplt_epi16(*qc1, zero); |
48 | 0 | y0 = _mm_unpacklo_epi16(*qc1, sign_bits); |
49 | 0 | y1 = _mm_unpackhi_epi16(*qc1, sign_bits); |
50 | 0 | _mm_store_si128((__m128i *)addr + 2, y0); |
51 | 0 | _mm_store_si128((__m128i *)addr + 3, y1); |
52 | 0 | } else { |
53 | 0 | _mm_store_si128((__m128i *)addr, *qc0); |
54 | 0 | _mm_store_si128((__m128i *)addr + 1, *qc1); |
55 | 0 | } |
56 | 0 | } |
57 | | |
58 | 0 | static inline void write_zero(tran_low_t *qcoeff, intptr_t offset) { |
59 | 0 | const __m128i zero = _mm_setzero_si128(); |
60 | 0 | tran_low_t *addr = qcoeff + offset; |
61 | 0 | if (sizeof(tran_low_t) == 4) { |
62 | 0 | _mm_store_si128((__m128i *)addr, zero); |
63 | 0 | _mm_store_si128((__m128i *)addr + 1, zero); |
64 | 0 | _mm_store_si128((__m128i *)addr + 2, zero); |
65 | 0 | _mm_store_si128((__m128i *)addr + 3, zero); |
66 | 0 | } else { |
67 | 0 | _mm_store_si128((__m128i *)addr, zero); |
68 | 0 | _mm_store_si128((__m128i *)addr + 1, zero); |
69 | 0 | } |
70 | 0 | } |
71 | | |
72 | | static inline void quantize(const int16_t *iscan_ptr, |
73 | | const tran_low_t *coeff_ptr, intptr_t n_coeffs, |
74 | | tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, |
75 | | const __m128i *round0, const __m128i *round1, |
76 | | const __m128i *quant0, const __m128i *quant1, |
77 | | const __m128i *dequant0, const __m128i *dequant1, |
78 | | const __m128i *thr0, const __m128i *thr1, |
79 | 0 | __m128i *eob) { |
80 | 0 | __m128i coeff0, coeff1; |
81 | | // Do DC and first 15 AC |
82 | 0 | read_coeff(coeff_ptr, n_coeffs, &coeff0, &coeff1); |
83 | | |
84 | | // Poor man's sign extract |
85 | 0 | const __m128i coeff0_sign = _mm_srai_epi16(coeff0, 15); |
86 | 0 | const __m128i coeff1_sign = _mm_srai_epi16(coeff1, 15); |
87 | 0 | __m128i qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); |
88 | 0 | __m128i qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); |
89 | 0 | qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); |
90 | 0 | qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); |
91 | 0 | const __m128i mask0 = _mm_or_si128(_mm_cmpgt_epi16(qcoeff0, *thr0), |
92 | 0 | _mm_cmpeq_epi16(qcoeff0, *thr0)); |
93 | 0 | const __m128i mask1 = _mm_or_si128(_mm_cmpgt_epi16(qcoeff1, *thr1), |
94 | 0 | _mm_cmpeq_epi16(qcoeff1, *thr1)); |
95 | 0 | const int nzflag = _mm_movemask_epi8(mask0) | _mm_movemask_epi8(mask1); |
96 | |
|
97 | 0 | if (nzflag) { |
98 | 0 | qcoeff0 = _mm_adds_epi16(qcoeff0, *round0); |
99 | 0 | qcoeff1 = _mm_adds_epi16(qcoeff1, *round1); |
100 | 0 | const __m128i qtmp0 = _mm_mulhi_epi16(qcoeff0, *quant0); |
101 | 0 | const __m128i qtmp1 = _mm_mulhi_epi16(qcoeff1, *quant1); |
102 | | |
103 | | // Reinsert signs |
104 | 0 | qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign); |
105 | 0 | qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign); |
106 | 0 | qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); |
107 | 0 | qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); |
108 | |
|
109 | 0 | write_qcoeff(&qcoeff0, &qcoeff1, qcoeff_ptr, n_coeffs); |
110 | |
|
111 | 0 | coeff0 = _mm_mullo_epi16(qcoeff0, *dequant0); |
112 | 0 | coeff1 = _mm_mullo_epi16(qcoeff1, *dequant1); |
113 | |
|
114 | 0 | write_qcoeff(&coeff0, &coeff1, dqcoeff_ptr, n_coeffs); |
115 | |
|
116 | 0 | const __m128i zero = _mm_setzero_si128(); |
117 | | // Scan for eob |
118 | 0 | const __m128i zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); |
119 | 0 | const __m128i zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); |
120 | 0 | const __m128i nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); |
121 | 0 | const __m128i nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); |
122 | 0 | const __m128i iscan0 = |
123 | 0 | _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); |
124 | 0 | const __m128i iscan1 = |
125 | 0 | _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); |
126 | | // Add one to convert from indices to counts |
127 | 0 | const __m128i iscan0_nz = _mm_sub_epi16(iscan0, nzero_coeff0); |
128 | 0 | const __m128i iscan1_nz = _mm_sub_epi16(iscan1, nzero_coeff1); |
129 | 0 | const __m128i eob0 = _mm_and_si128(iscan0_nz, nzero_coeff0); |
130 | 0 | const __m128i eob1 = _mm_and_si128(iscan1_nz, nzero_coeff1); |
131 | 0 | const __m128i eob2 = _mm_max_epi16(eob0, eob1); |
132 | 0 | *eob = _mm_max_epi16(*eob, eob2); |
133 | 0 | } else { |
134 | 0 | write_zero(qcoeff_ptr, n_coeffs); |
135 | 0 | write_zero(dqcoeff_ptr, n_coeffs); |
136 | 0 | } |
137 | 0 | } |
138 | | |
139 | | void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, |
140 | | const int16_t *zbin_ptr, const int16_t *round_ptr, |
141 | | const int16_t *quant_ptr, |
142 | | const int16_t *quant_shift_ptr, |
143 | | tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, |
144 | | const int16_t *dequant_ptr, uint16_t *eob_ptr, |
145 | 0 | const int16_t *scan_ptr, const int16_t *iscan_ptr) { |
146 | 0 | (void)scan_ptr; |
147 | 0 | (void)zbin_ptr; |
148 | 0 | (void)quant_shift_ptr; |
149 | |
|
150 | 0 | coeff_ptr += n_coeffs; |
151 | 0 | iscan_ptr += n_coeffs; |
152 | 0 | qcoeff_ptr += n_coeffs; |
153 | 0 | dqcoeff_ptr += n_coeffs; |
154 | 0 | n_coeffs = -n_coeffs; |
155 | |
|
156 | 0 | const __m128i round0 = _mm_load_si128((const __m128i *)round_ptr); |
157 | 0 | const __m128i round1 = _mm_unpackhi_epi64(round0, round0); |
158 | 0 | const __m128i quant0 = _mm_load_si128((const __m128i *)quant_ptr); |
159 | 0 | const __m128i quant1 = _mm_unpackhi_epi64(quant0, quant0); |
160 | 0 | const __m128i dequant0 = _mm_load_si128((const __m128i *)dequant_ptr); |
161 | 0 | const __m128i dequant1 = _mm_unpackhi_epi64(dequant0, dequant0); |
162 | 0 | const __m128i thr0 = _mm_srai_epi16(dequant0, 1); |
163 | 0 | const __m128i thr1 = _mm_srai_epi16(dequant1, 1); |
164 | 0 | __m128i eob = _mm_setzero_si128(); |
165 | |
|
166 | 0 | quantize(iscan_ptr, coeff_ptr, n_coeffs, qcoeff_ptr, dqcoeff_ptr, &round0, |
167 | 0 | &round1, &quant0, &quant1, &dequant0, &dequant1, &thr0, &thr1, &eob); |
168 | |
|
169 | 0 | n_coeffs += 8 * 2; |
170 | | |
171 | | // AC only loop |
172 | 0 | while (n_coeffs < 0) { |
173 | 0 | quantize(iscan_ptr, coeff_ptr, n_coeffs, qcoeff_ptr, dqcoeff_ptr, &round1, |
174 | 0 | &round1, &quant1, &quant1, &dequant1, &dequant1, &thr1, &thr1, |
175 | 0 | &eob); |
176 | 0 | n_coeffs += 8 * 2; |
177 | 0 | } |
178 | | |
179 | | // Accumulate EOB |
180 | 0 | { |
181 | 0 | __m128i eob_shuffled; |
182 | 0 | eob_shuffled = _mm_shuffle_epi32(eob, 0xe); |
183 | 0 | eob = _mm_max_epi16(eob, eob_shuffled); |
184 | 0 | eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); |
185 | 0 | eob = _mm_max_epi16(eob, eob_shuffled); |
186 | 0 | eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); |
187 | 0 | eob = _mm_max_epi16(eob, eob_shuffled); |
188 | 0 | *eob_ptr = _mm_extract_epi16(eob, 1); |
189 | 0 | } |
190 | 0 | } |
191 | | |
192 | | static inline void quantize_lp(const int16_t *iscan_ptr, |
193 | | const int16_t *coeff_ptr, intptr_t n_coeffs, |
194 | | int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, |
195 | | const __m128i *round0, const __m128i *round1, |
196 | | const __m128i *quant0, const __m128i *quant1, |
197 | | const __m128i *dequant0, const __m128i *dequant1, |
198 | 0 | __m128i *eob) { |
199 | 0 | const int16_t *read = coeff_ptr + n_coeffs; |
200 | 0 | __m128i coeff0 = _mm_load_si128((const __m128i *)read); |
201 | 0 | __m128i coeff1 = _mm_load_si128((const __m128i *)read + 1); |
202 | | |
203 | | // Poor man's sign extract |
204 | 0 | const __m128i coeff0_sign = _mm_srai_epi16(coeff0, 15); |
205 | 0 | const __m128i coeff1_sign = _mm_srai_epi16(coeff1, 15); |
206 | 0 | __m128i qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); |
207 | 0 | __m128i qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); |
208 | 0 | qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); |
209 | 0 | qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); |
210 | |
|
211 | 0 | qcoeff0 = _mm_adds_epi16(qcoeff0, *round0); |
212 | 0 | qcoeff1 = _mm_adds_epi16(qcoeff1, *round1); |
213 | 0 | const __m128i qtmp0 = _mm_mulhi_epi16(qcoeff0, *quant0); |
214 | 0 | const __m128i qtmp1 = _mm_mulhi_epi16(qcoeff1, *quant1); |
215 | | |
216 | | // Reinsert signs |
217 | 0 | qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign); |
218 | 0 | qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign); |
219 | 0 | qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); |
220 | 0 | qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); |
221 | |
|
222 | 0 | int16_t *addr = qcoeff_ptr + n_coeffs; |
223 | 0 | _mm_store_si128((__m128i *)addr, qcoeff0); |
224 | 0 | _mm_store_si128((__m128i *)addr + 1, qcoeff1); |
225 | |
|
226 | 0 | coeff0 = _mm_mullo_epi16(qcoeff0, *dequant0); |
227 | 0 | coeff1 = _mm_mullo_epi16(qcoeff1, *dequant1); |
228 | |
|
229 | 0 | addr = dqcoeff_ptr + n_coeffs; |
230 | 0 | _mm_store_si128((__m128i *)addr, coeff0); |
231 | 0 | _mm_store_si128((__m128i *)addr + 1, coeff1); |
232 | |
|
233 | 0 | const __m128i zero = _mm_setzero_si128(); |
234 | | // Scan for eob |
235 | 0 | const __m128i zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); |
236 | 0 | const __m128i zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); |
237 | 0 | const __m128i nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); |
238 | 0 | const __m128i nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); |
239 | |
|
240 | 0 | const __m128i iscan0 = |
241 | 0 | _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); |
242 | 0 | const __m128i iscan1 = |
243 | 0 | _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); |
244 | | |
245 | | // Add one to convert from indices to counts |
246 | 0 | const __m128i iscan0_nz = _mm_sub_epi16(iscan0, nzero_coeff0); |
247 | 0 | const __m128i iscan1_nz = _mm_sub_epi16(iscan1, nzero_coeff1); |
248 | 0 | const __m128i eob0 = _mm_and_si128(iscan0_nz, nzero_coeff0); |
249 | 0 | const __m128i eob1 = _mm_and_si128(iscan1_nz, nzero_coeff1); |
250 | 0 | const __m128i eob2 = _mm_max_epi16(eob0, eob1); |
251 | 0 | *eob = _mm_max_epi16(*eob, eob2); |
252 | 0 | } |
253 | | |
254 | | void av1_quantize_lp_sse2(const int16_t *coeff_ptr, intptr_t n_coeffs, |
255 | | const int16_t *round_ptr, const int16_t *quant_ptr, |
256 | | int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, |
257 | | const int16_t *dequant_ptr, uint16_t *eob_ptr, |
258 | 0 | const int16_t *scan, const int16_t *iscan) { |
259 | 0 | (void)scan; |
260 | 0 | coeff_ptr += n_coeffs; |
261 | 0 | iscan += n_coeffs; |
262 | 0 | qcoeff_ptr += n_coeffs; |
263 | 0 | dqcoeff_ptr += n_coeffs; |
264 | 0 | n_coeffs = -n_coeffs; |
265 | | |
266 | | // Setup global values |
267 | 0 | const __m128i round0 = _mm_load_si128((const __m128i *)round_ptr); |
268 | 0 | const __m128i round1 = _mm_unpackhi_epi64(round0, round0); |
269 | 0 | const __m128i quant0 = _mm_load_si128((const __m128i *)quant_ptr); |
270 | 0 | const __m128i quant1 = _mm_unpackhi_epi64(quant0, quant0); |
271 | 0 | const __m128i dequant0 = _mm_load_si128((const __m128i *)dequant_ptr); |
272 | 0 | const __m128i dequant1 = _mm_unpackhi_epi64(dequant0, dequant0); |
273 | 0 | __m128i eob = _mm_setzero_si128(); |
274 | | |
275 | | // DC and first 15 AC |
276 | 0 | quantize_lp(iscan, coeff_ptr, n_coeffs, qcoeff_ptr, dqcoeff_ptr, &round0, |
277 | 0 | &round1, &quant0, &quant1, &dequant0, &dequant1, &eob); |
278 | 0 | n_coeffs += 8 * 2; |
279 | | |
280 | | // AC only loop |
281 | 0 | while (n_coeffs < 0) { |
282 | 0 | quantize_lp(iscan, coeff_ptr, n_coeffs, qcoeff_ptr, dqcoeff_ptr, &round1, |
283 | 0 | &round1, &quant1, &quant1, &dequant1, &dequant1, &eob); |
284 | 0 | n_coeffs += 8 * 2; |
285 | 0 | } |
286 | | |
287 | | // Accumulate EOB |
288 | 0 | *eob_ptr = accumulate_eob(eob); |
289 | 0 | } |