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 <arm_neon.h>
13
14 #include "config/aom_config.h"
15
16 #include "aom_dsp/txfm_common.h"
17 #include "aom_dsp/arm/mem_neon.h"
18 #include "aom_dsp/arm/transpose_neon.h"
19
aom_fdct4x4_helper(const int16_t * input,int stride,int16x4_t * input_0,int16x4_t * input_1,int16x4_t * input_2,int16x4_t * input_3)20 static void aom_fdct4x4_helper(const int16_t *input, int stride,
21 int16x4_t *input_0, int16x4_t *input_1,
22 int16x4_t *input_2, int16x4_t *input_3) {
23 *input_0 = vshl_n_s16(vld1_s16(input + 0 * stride), 4);
24 *input_1 = vshl_n_s16(vld1_s16(input + 1 * stride), 4);
25 *input_2 = vshl_n_s16(vld1_s16(input + 2 * stride), 4);
26 *input_3 = vshl_n_s16(vld1_s16(input + 3 * stride), 4);
27 // If the very first value != 0, then add 1.
28 if (input[0] != 0) {
29 const int16x4_t one = vreinterpret_s16_s64(vdup_n_s64(1));
30 *input_0 = vadd_s16(*input_0, one);
31 }
32
33 for (int i = 0; i < 2; ++i) {
34 const int16x8_t input_01 = vcombine_s16(*input_0, *input_1);
35 const int16x8_t input_32 = vcombine_s16(*input_3, *input_2);
36
37 // in_0 +/- in_3, in_1 +/- in_2
38 const int16x8_t s_01 = vaddq_s16(input_01, input_32);
39 const int16x8_t s_32 = vsubq_s16(input_01, input_32);
40
41 // step_0 +/- step_1, step_2 +/- step_3
42 const int16x4_t s_0 = vget_low_s16(s_01);
43 const int16x4_t s_1 = vget_high_s16(s_01);
44 const int16x4_t s_2 = vget_high_s16(s_32);
45 const int16x4_t s_3 = vget_low_s16(s_32);
46
47 // (s_0 +/- s_1) * cospi_16_64
48 // Must expand all elements to s32. See 'needs32' comment in fwd_txfm.c.
49 const int32x4_t s_0_p_s_1 = vaddl_s16(s_0, s_1);
50 const int32x4_t s_0_m_s_1 = vsubl_s16(s_0, s_1);
51 const int32x4_t temp1 = vmulq_n_s32(s_0_p_s_1, (int32_t)cospi_16_64);
52 const int32x4_t temp2 = vmulq_n_s32(s_0_m_s_1, (int32_t)cospi_16_64);
53
54 // fdct_round_shift
55 int16x4_t out_0 = vrshrn_n_s32(temp1, DCT_CONST_BITS);
56 int16x4_t out_2 = vrshrn_n_s32(temp2, DCT_CONST_BITS);
57
58 // s_3 * cospi_8_64 + s_2 * cospi_24_64
59 // s_3 * cospi_24_64 - s_2 * cospi_8_64
60 const int32x4_t s_3_cospi_8_64 = vmull_n_s16(s_3, (int32_t)cospi_8_64);
61 const int32x4_t s_3_cospi_24_64 = vmull_n_s16(s_3, (int32_t)cospi_24_64);
62
63 const int32x4_t temp3 =
64 vmlal_n_s16(s_3_cospi_8_64, s_2, (int32_t)cospi_24_64);
65 const int32x4_t temp4 =
66 vmlsl_n_s16(s_3_cospi_24_64, s_2, (int32_t)cospi_8_64);
67
68 // fdct_round_shift
69 int16x4_t out_1 = vrshrn_n_s32(temp3, DCT_CONST_BITS);
70 int16x4_t out_3 = vrshrn_n_s32(temp4, DCT_CONST_BITS);
71
72 // Only transpose the first pass
73 if (i == 0) {
74 transpose_elems_inplace_s16_4x4(&out_0, &out_1, &out_2, &out_3);
75 }
76
77 *input_0 = out_0;
78 *input_1 = out_1;
79 *input_2 = out_2;
80 *input_3 = out_3;
81 }
82 }
83
aom_fdct4x4_neon(const int16_t * input,tran_low_t * final_output,int stride)84 void aom_fdct4x4_neon(const int16_t *input, tran_low_t *final_output,
85 int stride) {
86 // input[M * stride] * 16
87 int16x4_t input_0, input_1, input_2, input_3;
88
89 aom_fdct4x4_helper(input, stride, &input_0, &input_1, &input_2, &input_3);
90
91 // Not quite a rounding shift. Only add 1 despite shifting by 2.
92 const int16x8_t one = vdupq_n_s16(1);
93 int16x8_t out_01 = vcombine_s16(input_0, input_1);
94 int16x8_t out_23 = vcombine_s16(input_2, input_3);
95 out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2);
96 out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2);
97 store_s16q_to_tran_low(final_output + 0 * 8, out_01);
98 store_s16q_to_tran_low(final_output + 1 * 8, out_23);
99 }
100
aom_fdct4x4_lp_neon(const int16_t * input,int16_t * final_output,int stride)101 void aom_fdct4x4_lp_neon(const int16_t *input, int16_t *final_output,
102 int stride) {
103 // input[M * stride] * 16
104 int16x4_t input_0, input_1, input_2, input_3;
105
106 aom_fdct4x4_helper(input, stride, &input_0, &input_1, &input_2, &input_3);
107
108 // Not quite a rounding shift. Only add 1 despite shifting by 2.
109 const int16x8_t one = vdupq_n_s16(1);
110 int16x8_t out_01 = vcombine_s16(input_0, input_1);
111 int16x8_t out_23 = vcombine_s16(input_2, input_3);
112 out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2);
113 out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2);
114 vst1q_s16(final_output + 0 * 8, out_01);
115 vst1q_s16(final_output + 1 * 8, out_23);
116 }
117
aom_fdct8x8_neon(const int16_t * input,int16_t * final_output,int stride)118 void aom_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) {
119 // stage 1
120 int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2);
121 int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2);
122 int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2);
123 int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2);
124 int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2);
125 int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2);
126 int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2);
127 int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2);
128 for (int i = 0; i < 2; ++i) {
129 int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7;
130 const int16x8_t v_s0 = vaddq_s16(input_0, input_7);
131 const int16x8_t v_s1 = vaddq_s16(input_1, input_6);
132 const int16x8_t v_s2 = vaddq_s16(input_2, input_5);
133 const int16x8_t v_s3 = vaddq_s16(input_3, input_4);
134 const int16x8_t v_s4 = vsubq_s16(input_3, input_4);
135 const int16x8_t v_s5 = vsubq_s16(input_2, input_5);
136 const int16x8_t v_s6 = vsubq_s16(input_1, input_6);
137 const int16x8_t v_s7 = vsubq_s16(input_0, input_7);
138 // fdct4(step, step);
139 int16x8_t v_x0 = vaddq_s16(v_s0, v_s3);
140 int16x8_t v_x1 = vaddq_s16(v_s1, v_s2);
141 int16x8_t v_x2 = vsubq_s16(v_s1, v_s2);
142 int16x8_t v_x3 = vsubq_s16(v_s0, v_s3);
143 // fdct4(step, step);
144 int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
145 int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
146 int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
147 int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
148 int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64);
149 int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64);
150 int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64);
151 int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64);
152 v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64);
153 v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64);
154 v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64);
155 v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64);
156 v_t0_lo = vmulq_n_s32(v_t0_lo, (int32_t)cospi_16_64);
157 v_t0_hi = vmulq_n_s32(v_t0_hi, (int32_t)cospi_16_64);
158 v_t1_lo = vmulq_n_s32(v_t1_lo, (int32_t)cospi_16_64);
159 v_t1_hi = vmulq_n_s32(v_t1_hi, (int32_t)cospi_16_64);
160 {
161 const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
162 const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
163 const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
164 const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
165 const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
166 const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
167 const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
168 const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
169 out_0 = vcombine_s16(a, c); // 00 01 02 03 40 41 42 43
170 out_2 = vcombine_s16(e, g); // 20 21 22 23 60 61 62 63
171 out_4 = vcombine_s16(b, d); // 04 05 06 07 44 45 46 47
172 out_6 = vcombine_s16(f, h); // 24 25 26 27 64 65 66 67
173 }
174 // Stage 2
175 v_x0 = vsubq_s16(v_s6, v_s5);
176 v_x1 = vaddq_s16(v_s6, v_s5);
177 v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64);
178 v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64);
179 v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64);
180 v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64);
181 {
182 const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
183 const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
184 const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
185 const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
186 const int16x8_t ab = vcombine_s16(a, b);
187 const int16x8_t cd = vcombine_s16(c, d);
188 // Stage 3
189 v_x0 = vaddq_s16(v_s4, ab);
190 v_x1 = vsubq_s16(v_s4, ab);
191 v_x2 = vsubq_s16(v_s7, cd);
192 v_x3 = vaddq_s16(v_s7, cd);
193 }
194 // Stage 4
195 v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64);
196 v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64);
197 v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64);
198 v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64);
199 v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64);
200 v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64);
201 v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64);
202 v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64);
203 v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64);
204 v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64);
205 v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64);
206 v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64);
207 v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64);
208 v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64);
209 v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64);
210 v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64);
211 {
212 const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
213 const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
214 const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
215 const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
216 const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
217 const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
218 const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
219 const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
220 out_1 = vcombine_s16(a, c); // 10 11 12 13 50 51 52 53
221 out_3 = vcombine_s16(e, g); // 30 31 32 33 70 71 72 73
222 out_5 = vcombine_s16(b, d); // 14 15 16 17 54 55 56 57
223 out_7 = vcombine_s16(f, h); // 34 35 36 37 74 75 76 77
224 }
225 // transpose 8x8
226 {
227 // 00 01 02 03 40 41 42 43
228 // 10 11 12 13 50 51 52 53
229 // 20 21 22 23 60 61 62 63
230 // 30 31 32 33 70 71 72 73
231 // 04 05 06 07 44 45 46 47
232 // 14 15 16 17 54 55 56 57
233 // 24 25 26 27 64 65 66 67
234 // 34 35 36 37 74 75 76 77
235 const int32x4x2_t r02_s32 =
236 vtrnq_s32(vreinterpretq_s32_s16(out_0), vreinterpretq_s32_s16(out_2));
237 const int32x4x2_t r13_s32 =
238 vtrnq_s32(vreinterpretq_s32_s16(out_1), vreinterpretq_s32_s16(out_3));
239 const int32x4x2_t r46_s32 =
240 vtrnq_s32(vreinterpretq_s32_s16(out_4), vreinterpretq_s32_s16(out_6));
241 const int32x4x2_t r57_s32 =
242 vtrnq_s32(vreinterpretq_s32_s16(out_5), vreinterpretq_s32_s16(out_7));
243 const int16x8x2_t r01_s16 =
244 vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]),
245 vreinterpretq_s16_s32(r13_s32.val[0]));
246 const int16x8x2_t r23_s16 =
247 vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]),
248 vreinterpretq_s16_s32(r13_s32.val[1]));
249 const int16x8x2_t r45_s16 =
250 vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]),
251 vreinterpretq_s16_s32(r57_s32.val[0]));
252 const int16x8x2_t r67_s16 =
253 vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]),
254 vreinterpretq_s16_s32(r57_s32.val[1]));
255 input_0 = r01_s16.val[0];
256 input_1 = r01_s16.val[1];
257 input_2 = r23_s16.val[0];
258 input_3 = r23_s16.val[1];
259 input_4 = r45_s16.val[0];
260 input_5 = r45_s16.val[1];
261 input_6 = r67_s16.val[0];
262 input_7 = r67_s16.val[1];
263 // 00 10 20 30 40 50 60 70
264 // 01 11 21 31 41 51 61 71
265 // 02 12 22 32 42 52 62 72
266 // 03 13 23 33 43 53 63 73
267 // 04 14 24 34 44 54 64 74
268 // 05 15 25 35 45 55 65 75
269 // 06 16 26 36 46 56 66 76
270 // 07 17 27 37 47 57 67 77
271 }
272 } // for
273 {
274 // from aom_dct_sse2.c
275 // Post-condition (division by two)
276 // division of two 16 bits signed numbers using shifts
277 // n / 2 = (n - (n >> 15)) >> 1
278 const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15);
279 const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15);
280 const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15);
281 const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15);
282 const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15);
283 const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15);
284 const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15);
285 const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15);
286 input_0 = vhsubq_s16(input_0, sign_in0);
287 input_1 = vhsubq_s16(input_1, sign_in1);
288 input_2 = vhsubq_s16(input_2, sign_in2);
289 input_3 = vhsubq_s16(input_3, sign_in3);
290 input_4 = vhsubq_s16(input_4, sign_in4);
291 input_5 = vhsubq_s16(input_5, sign_in5);
292 input_6 = vhsubq_s16(input_6, sign_in6);
293 input_7 = vhsubq_s16(input_7, sign_in7);
294 // store results
295 vst1q_s16(&final_output[0 * 8], input_0);
296 vst1q_s16(&final_output[1 * 8], input_1);
297 vst1q_s16(&final_output[2 * 8], input_2);
298 vst1q_s16(&final_output[3 * 8], input_3);
299 vst1q_s16(&final_output[4 * 8], input_4);
300 vst1q_s16(&final_output[5 * 8], input_5);
301 vst1q_s16(&final_output[6 * 8], input_6);
302 vst1q_s16(&final_output[7 * 8], input_7);
303 }
304 }
305