xref: /aosp_15_r20/external/libaom/aom_dsp/arm/fwd_txfm_neon.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
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 #include "config/aom_dsp_rtcd.h"
16 
17 #include "aom_dsp/txfm_common.h"
18 #include "aom_dsp/arm/mem_neon.h"
19 #include "aom_dsp/arm/transpose_neon.h"
20 
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)21 static void aom_fdct4x4_helper(const int16_t *input, int stride,
22                                int16x4_t *input_0, int16x4_t *input_1,
23                                int16x4_t *input_2, int16x4_t *input_3) {
24   *input_0 = vshl_n_s16(vld1_s16(input + 0 * stride), 4);
25   *input_1 = vshl_n_s16(vld1_s16(input + 1 * stride), 4);
26   *input_2 = vshl_n_s16(vld1_s16(input + 2 * stride), 4);
27   *input_3 = vshl_n_s16(vld1_s16(input + 3 * stride), 4);
28   // If the very first value != 0, then add 1.
29   if (input[0] != 0) {
30     const int16x4_t one = vreinterpret_s16_s64(vdup_n_s64(1));
31     *input_0 = vadd_s16(*input_0, one);
32   }
33 
34   for (int i = 0; i < 2; ++i) {
35     const int16x8_t input_01 = vcombine_s16(*input_0, *input_1);
36     const int16x8_t input_32 = vcombine_s16(*input_3, *input_2);
37 
38     // in_0 +/- in_3, in_1 +/- in_2
39     const int16x8_t s_01 = vaddq_s16(input_01, input_32);
40     const int16x8_t s_32 = vsubq_s16(input_01, input_32);
41 
42     // step_0 +/- step_1, step_2 +/- step_3
43     const int16x4_t s_0 = vget_low_s16(s_01);
44     const int16x4_t s_1 = vget_high_s16(s_01);
45     const int16x4_t s_2 = vget_high_s16(s_32);
46     const int16x4_t s_3 = vget_low_s16(s_32);
47 
48     // (s_0 +/- s_1) * cospi_16_64
49     // Must expand all elements to s32. See 'needs32' comment in fwd_txfm.c.
50     const int32x4_t s_0_p_s_1 = vaddl_s16(s_0, s_1);
51     const int32x4_t s_0_m_s_1 = vsubl_s16(s_0, s_1);
52     const int32x4_t temp1 = vmulq_n_s32(s_0_p_s_1, (int32_t)cospi_16_64);
53     const int32x4_t temp2 = vmulq_n_s32(s_0_m_s_1, (int32_t)cospi_16_64);
54 
55     // fdct_round_shift
56     int16x4_t out_0 = vrshrn_n_s32(temp1, DCT_CONST_BITS);
57     int16x4_t out_2 = vrshrn_n_s32(temp2, DCT_CONST_BITS);
58 
59     // s_3 * cospi_8_64 + s_2 * cospi_24_64
60     // s_3 * cospi_24_64 - s_2 * cospi_8_64
61     const int32x4_t s_3_cospi_8_64 = vmull_n_s16(s_3, (int32_t)cospi_8_64);
62     const int32x4_t s_3_cospi_24_64 = vmull_n_s16(s_3, (int32_t)cospi_24_64);
63 
64     const int32x4_t temp3 =
65         vmlal_n_s16(s_3_cospi_8_64, s_2, (int32_t)cospi_24_64);
66     const int32x4_t temp4 =
67         vmlsl_n_s16(s_3_cospi_24_64, s_2, (int32_t)cospi_8_64);
68 
69     // fdct_round_shift
70     int16x4_t out_1 = vrshrn_n_s32(temp3, DCT_CONST_BITS);
71     int16x4_t out_3 = vrshrn_n_s32(temp4, DCT_CONST_BITS);
72 
73     // Only transpose the first pass
74     if (i == 0) {
75       transpose_elems_inplace_s16_4x4(&out_0, &out_1, &out_2, &out_3);
76     }
77 
78     *input_0 = out_0;
79     *input_1 = out_1;
80     *input_2 = out_2;
81     *input_3 = out_3;
82   }
83 }
84 
aom_fdct4x4_neon(const int16_t * input,tran_low_t * final_output,int stride)85 void aom_fdct4x4_neon(const int16_t *input, tran_low_t *final_output,
86                       int stride) {
87   // input[M * stride] * 16
88   int16x4_t input_0, input_1, input_2, input_3;
89 
90   aom_fdct4x4_helper(input, stride, &input_0, &input_1, &input_2, &input_3);
91 
92   // Not quite a rounding shift. Only add 1 despite shifting by 2.
93   const int16x8_t one = vdupq_n_s16(1);
94   int16x8_t out_01 = vcombine_s16(input_0, input_1);
95   int16x8_t out_23 = vcombine_s16(input_2, input_3);
96   out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2);
97   out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2);
98   store_s16q_to_tran_low(final_output + 0 * 8, out_01);
99   store_s16q_to_tran_low(final_output + 1 * 8, out_23);
100 }
101 
aom_fdct4x4_lp_neon(const int16_t * input,int16_t * final_output,int stride)102 void aom_fdct4x4_lp_neon(const int16_t *input, int16_t *final_output,
103                          int stride) {
104   // input[M * stride] * 16
105   int16x4_t input_0, input_1, input_2, input_3;
106 
107   aom_fdct4x4_helper(input, stride, &input_0, &input_1, &input_2, &input_3);
108 
109   // Not quite a rounding shift. Only add 1 despite shifting by 2.
110   const int16x8_t one = vdupq_n_s16(1);
111   int16x8_t out_01 = vcombine_s16(input_0, input_1);
112   int16x8_t out_23 = vcombine_s16(input_2, input_3);
113   out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2);
114   out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2);
115   vst1q_s16(final_output + 0 * 8, out_01);
116   vst1q_s16(final_output + 1 * 8, out_23);
117 }
118 
119 #if CONFIG_INTERNAL_STATS
aom_fdct8x8_neon(const int16_t * input,int16_t * final_output,int stride)120 void aom_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) {
121   // stage 1
122   int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2);
123   int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2);
124   int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2);
125   int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2);
126   int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2);
127   int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2);
128   int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2);
129   int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2);
130   for (int i = 0; i < 2; ++i) {
131     int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7;
132     const int16x8_t v_s0 = vaddq_s16(input_0, input_7);
133     const int16x8_t v_s1 = vaddq_s16(input_1, input_6);
134     const int16x8_t v_s2 = vaddq_s16(input_2, input_5);
135     const int16x8_t v_s3 = vaddq_s16(input_3, input_4);
136     const int16x8_t v_s4 = vsubq_s16(input_3, input_4);
137     const int16x8_t v_s5 = vsubq_s16(input_2, input_5);
138     const int16x8_t v_s6 = vsubq_s16(input_1, input_6);
139     const int16x8_t v_s7 = vsubq_s16(input_0, input_7);
140     // fdct4(step, step);
141     int16x8_t v_x0 = vaddq_s16(v_s0, v_s3);
142     int16x8_t v_x1 = vaddq_s16(v_s1, v_s2);
143     int16x8_t v_x2 = vsubq_s16(v_s1, v_s2);
144     int16x8_t v_x3 = vsubq_s16(v_s0, v_s3);
145     // fdct4(step, step);
146     int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
147     int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
148     int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
149     int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
150     int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64);
151     int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64);
152     int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64);
153     int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64);
154     v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64);
155     v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64);
156     v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64);
157     v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64);
158     v_t0_lo = vmulq_n_s32(v_t0_lo, (int32_t)cospi_16_64);
159     v_t0_hi = vmulq_n_s32(v_t0_hi, (int32_t)cospi_16_64);
160     v_t1_lo = vmulq_n_s32(v_t1_lo, (int32_t)cospi_16_64);
161     v_t1_hi = vmulq_n_s32(v_t1_hi, (int32_t)cospi_16_64);
162     {
163       const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
164       const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
165       const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
166       const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
167       const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
168       const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
169       const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
170       const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
171       out_0 = vcombine_s16(a, c);  // 00 01 02 03 40 41 42 43
172       out_2 = vcombine_s16(e, g);  // 20 21 22 23 60 61 62 63
173       out_4 = vcombine_s16(b, d);  // 04 05 06 07 44 45 46 47
174       out_6 = vcombine_s16(f, h);  // 24 25 26 27 64 65 66 67
175     }
176     // Stage 2
177     v_x0 = vsubq_s16(v_s6, v_s5);
178     v_x1 = vaddq_s16(v_s6, v_s5);
179     v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64);
180     v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64);
181     v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64);
182     v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64);
183     {
184       const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
185       const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
186       const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
187       const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
188       const int16x8_t ab = vcombine_s16(a, b);
189       const int16x8_t cd = vcombine_s16(c, d);
190       // Stage 3
191       v_x0 = vaddq_s16(v_s4, ab);
192       v_x1 = vsubq_s16(v_s4, ab);
193       v_x2 = vsubq_s16(v_s7, cd);
194       v_x3 = vaddq_s16(v_s7, cd);
195     }
196     // Stage 4
197     v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64);
198     v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64);
199     v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64);
200     v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64);
201     v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64);
202     v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64);
203     v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64);
204     v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64);
205     v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64);
206     v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64);
207     v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64);
208     v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64);
209     v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64);
210     v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64);
211     v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64);
212     v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64);
213     {
214       const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
215       const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
216       const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
217       const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
218       const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
219       const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
220       const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
221       const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
222       out_1 = vcombine_s16(a, c);  // 10 11 12 13 50 51 52 53
223       out_3 = vcombine_s16(e, g);  // 30 31 32 33 70 71 72 73
224       out_5 = vcombine_s16(b, d);  // 14 15 16 17 54 55 56 57
225       out_7 = vcombine_s16(f, h);  // 34 35 36 37 74 75 76 77
226     }
227     // transpose 8x8
228     {
229       // 00 01 02 03 40 41 42 43
230       // 10 11 12 13 50 51 52 53
231       // 20 21 22 23 60 61 62 63
232       // 30 31 32 33 70 71 72 73
233       // 04 05 06 07 44 45 46 47
234       // 14 15 16 17 54 55 56 57
235       // 24 25 26 27 64 65 66 67
236       // 34 35 36 37 74 75 76 77
237       const int32x4x2_t r02_s32 =
238           vtrnq_s32(vreinterpretq_s32_s16(out_0), vreinterpretq_s32_s16(out_2));
239       const int32x4x2_t r13_s32 =
240           vtrnq_s32(vreinterpretq_s32_s16(out_1), vreinterpretq_s32_s16(out_3));
241       const int32x4x2_t r46_s32 =
242           vtrnq_s32(vreinterpretq_s32_s16(out_4), vreinterpretq_s32_s16(out_6));
243       const int32x4x2_t r57_s32 =
244           vtrnq_s32(vreinterpretq_s32_s16(out_5), vreinterpretq_s32_s16(out_7));
245       const int16x8x2_t r01_s16 =
246           vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]),
247                     vreinterpretq_s16_s32(r13_s32.val[0]));
248       const int16x8x2_t r23_s16 =
249           vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]),
250                     vreinterpretq_s16_s32(r13_s32.val[1]));
251       const int16x8x2_t r45_s16 =
252           vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]),
253                     vreinterpretq_s16_s32(r57_s32.val[0]));
254       const int16x8x2_t r67_s16 =
255           vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]),
256                     vreinterpretq_s16_s32(r57_s32.val[1]));
257       input_0 = r01_s16.val[0];
258       input_1 = r01_s16.val[1];
259       input_2 = r23_s16.val[0];
260       input_3 = r23_s16.val[1];
261       input_4 = r45_s16.val[0];
262       input_5 = r45_s16.val[1];
263       input_6 = r67_s16.val[0];
264       input_7 = r67_s16.val[1];
265       // 00 10 20 30 40 50 60 70
266       // 01 11 21 31 41 51 61 71
267       // 02 12 22 32 42 52 62 72
268       // 03 13 23 33 43 53 63 73
269       // 04 14 24 34 44 54 64 74
270       // 05 15 25 35 45 55 65 75
271       // 06 16 26 36 46 56 66 76
272       // 07 17 27 37 47 57 67 77
273     }
274   }  // for
275   {
276     // from aom_dct_sse2.c
277     // Post-condition (division by two)
278     //    division of two 16 bits signed numbers using shifts
279     //    n / 2 = (n - (n >> 15)) >> 1
280     const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15);
281     const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15);
282     const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15);
283     const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15);
284     const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15);
285     const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15);
286     const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15);
287     const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15);
288     input_0 = vhsubq_s16(input_0, sign_in0);
289     input_1 = vhsubq_s16(input_1, sign_in1);
290     input_2 = vhsubq_s16(input_2, sign_in2);
291     input_3 = vhsubq_s16(input_3, sign_in3);
292     input_4 = vhsubq_s16(input_4, sign_in4);
293     input_5 = vhsubq_s16(input_5, sign_in5);
294     input_6 = vhsubq_s16(input_6, sign_in6);
295     input_7 = vhsubq_s16(input_7, sign_in7);
296     // store results
297     vst1q_s16(&final_output[0 * 8], input_0);
298     vst1q_s16(&final_output[1 * 8], input_1);
299     vst1q_s16(&final_output[2 * 8], input_2);
300     vst1q_s16(&final_output[3 * 8], input_3);
301     vst1q_s16(&final_output[4 * 8], input_4);
302     vst1q_s16(&final_output[5 * 8], input_5);
303     vst1q_s16(&final_output[6 * 8], input_6);
304     vst1q_s16(&final_output[7 * 8], input_7);
305   }
306 }
307 #endif  // CONFIG_INTERNAL_STATS
308