1 /*
2 * jfdctfst-neon.c - fast integer FDCT (Arm Neon)
3 *
4 * Copyright (C) 2020, Arm Limited. All Rights Reserved.
5 *
6 * This software is provided 'as-is', without any express or implied
7 * warranty. In no event will the authors be held liable for any damages
8 * arising from the use of this software.
9 *
10 * Permission is granted to anyone to use this software for any purpose,
11 * including commercial applications, and to alter it and redistribute it
12 * freely, subject to the following restrictions:
13 *
14 * 1. The origin of this software must not be misrepresented; you must not
15 * claim that you wrote the original software. If you use this software
16 * in a product, an acknowledgment in the product documentation would be
17 * appreciated but is not required.
18 * 2. Altered source versions must be plainly marked as such, and must not be
19 * misrepresented as being the original software.
20 * 3. This notice may not be removed or altered from any source distribution.
21 */
22
23 #define JPEG_INTERNALS
24 #include "../../jinclude.h"
25 #include "../../jpeglib.h"
26 #include "../../jsimd.h"
27 #include "../../jdct.h"
28 #include "../../jsimddct.h"
29 #include "../jsimd.h"
30 #include "align.h"
31
32 #include <arm_neon.h>
33
34
35 /* jsimd_fdct_ifast_neon() performs a fast, not so accurate forward DCT
36 * (Discrete Cosine Transform) on one block of samples. It uses the same
37 * calculations and produces exactly the same output as IJG's original
38 * jpeg_fdct_ifast() function, which can be found in jfdctfst.c.
39 *
40 * Scaled integer constants are used to avoid floating-point arithmetic:
41 * 0.382683433 = 12544 * 2^-15
42 * 0.541196100 = 17795 * 2^-15
43 * 0.707106781 = 23168 * 2^-15
44 * 0.306562965 = 9984 * 2^-15
45 *
46 * See jfdctfst.c for further details of the DCT algorithm. Where possible,
47 * the variable names and comments here in jsimd_fdct_ifast_neon() match up
48 * with those in jpeg_fdct_ifast().
49 */
50
51 #define F_0_382 12544
52 #define F_0_541 17792
53 #define F_0_707 23168
54 #define F_0_306 9984
55
56
57 ALIGN(16) static const int16_t jsimd_fdct_ifast_neon_consts[] = {
58 F_0_382, F_0_541, F_0_707, F_0_306
59 };
60
jsimd_fdct_ifast_neon(DCTELEM * data)61 void jsimd_fdct_ifast_neon(DCTELEM *data)
62 {
63 /* Load an 8x8 block of samples into Neon registers. De-interleaving loads
64 * are used, followed by vuzp to transpose the block such that we have a
65 * column of samples per vector - allowing all rows to be processed at once.
66 */
67 int16x8x4_t data1 = vld4q_s16(data);
68 int16x8x4_t data2 = vld4q_s16(data + 4 * DCTSIZE);
69
70 int16x8x2_t cols_04 = vuzpq_s16(data1.val[0], data2.val[0]);
71 int16x8x2_t cols_15 = vuzpq_s16(data1.val[1], data2.val[1]);
72 int16x8x2_t cols_26 = vuzpq_s16(data1.val[2], data2.val[2]);
73 int16x8x2_t cols_37 = vuzpq_s16(data1.val[3], data2.val[3]);
74
75 int16x8_t col0 = cols_04.val[0];
76 int16x8_t col1 = cols_15.val[0];
77 int16x8_t col2 = cols_26.val[0];
78 int16x8_t col3 = cols_37.val[0];
79 int16x8_t col4 = cols_04.val[1];
80 int16x8_t col5 = cols_15.val[1];
81 int16x8_t col6 = cols_26.val[1];
82 int16x8_t col7 = cols_37.val[1];
83
84 /* Pass 1: process rows. */
85
86 /* Load DCT conversion constants. */
87 const int16x4_t consts = vld1_s16(jsimd_fdct_ifast_neon_consts);
88
89 int16x8_t tmp0 = vaddq_s16(col0, col7);
90 int16x8_t tmp7 = vsubq_s16(col0, col7);
91 int16x8_t tmp1 = vaddq_s16(col1, col6);
92 int16x8_t tmp6 = vsubq_s16(col1, col6);
93 int16x8_t tmp2 = vaddq_s16(col2, col5);
94 int16x8_t tmp5 = vsubq_s16(col2, col5);
95 int16x8_t tmp3 = vaddq_s16(col3, col4);
96 int16x8_t tmp4 = vsubq_s16(col3, col4);
97
98 /* Even part */
99 int16x8_t tmp10 = vaddq_s16(tmp0, tmp3); /* phase 2 */
100 int16x8_t tmp13 = vsubq_s16(tmp0, tmp3);
101 int16x8_t tmp11 = vaddq_s16(tmp1, tmp2);
102 int16x8_t tmp12 = vsubq_s16(tmp1, tmp2);
103
104 col0 = vaddq_s16(tmp10, tmp11); /* phase 3 */
105 col4 = vsubq_s16(tmp10, tmp11);
106
107 int16x8_t z1 = vqdmulhq_lane_s16(vaddq_s16(tmp12, tmp13), consts, 2);
108 col2 = vaddq_s16(tmp13, z1); /* phase 5 */
109 col6 = vsubq_s16(tmp13, z1);
110
111 /* Odd part */
112 tmp10 = vaddq_s16(tmp4, tmp5); /* phase 2 */
113 tmp11 = vaddq_s16(tmp5, tmp6);
114 tmp12 = vaddq_s16(tmp6, tmp7);
115
116 int16x8_t z5 = vqdmulhq_lane_s16(vsubq_s16(tmp10, tmp12), consts, 0);
117 int16x8_t z2 = vqdmulhq_lane_s16(tmp10, consts, 1);
118 z2 = vaddq_s16(z2, z5);
119 int16x8_t z4 = vqdmulhq_lane_s16(tmp12, consts, 3);
120 z5 = vaddq_s16(tmp12, z5);
121 z4 = vaddq_s16(z4, z5);
122 int16x8_t z3 = vqdmulhq_lane_s16(tmp11, consts, 2);
123
124 int16x8_t z11 = vaddq_s16(tmp7, z3); /* phase 5 */
125 int16x8_t z13 = vsubq_s16(tmp7, z3);
126
127 col5 = vaddq_s16(z13, z2); /* phase 6 */
128 col3 = vsubq_s16(z13, z2);
129 col1 = vaddq_s16(z11, z4);
130 col7 = vsubq_s16(z11, z4);
131
132 /* Transpose to work on columns in pass 2. */
133 int16x8x2_t cols_01 = vtrnq_s16(col0, col1);
134 int16x8x2_t cols_23 = vtrnq_s16(col2, col3);
135 int16x8x2_t cols_45 = vtrnq_s16(col4, col5);
136 int16x8x2_t cols_67 = vtrnq_s16(col6, col7);
137
138 int32x4x2_t cols_0145_l = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[0]),
139 vreinterpretq_s32_s16(cols_45.val[0]));
140 int32x4x2_t cols_0145_h = vtrnq_s32(vreinterpretq_s32_s16(cols_01.val[1]),
141 vreinterpretq_s32_s16(cols_45.val[1]));
142 int32x4x2_t cols_2367_l = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[0]),
143 vreinterpretq_s32_s16(cols_67.val[0]));
144 int32x4x2_t cols_2367_h = vtrnq_s32(vreinterpretq_s32_s16(cols_23.val[1]),
145 vreinterpretq_s32_s16(cols_67.val[1]));
146
147 int32x4x2_t rows_04 = vzipq_s32(cols_0145_l.val[0], cols_2367_l.val[0]);
148 int32x4x2_t rows_15 = vzipq_s32(cols_0145_h.val[0], cols_2367_h.val[0]);
149 int32x4x2_t rows_26 = vzipq_s32(cols_0145_l.val[1], cols_2367_l.val[1]);
150 int32x4x2_t rows_37 = vzipq_s32(cols_0145_h.val[1], cols_2367_h.val[1]);
151
152 int16x8_t row0 = vreinterpretq_s16_s32(rows_04.val[0]);
153 int16x8_t row1 = vreinterpretq_s16_s32(rows_15.val[0]);
154 int16x8_t row2 = vreinterpretq_s16_s32(rows_26.val[0]);
155 int16x8_t row3 = vreinterpretq_s16_s32(rows_37.val[0]);
156 int16x8_t row4 = vreinterpretq_s16_s32(rows_04.val[1]);
157 int16x8_t row5 = vreinterpretq_s16_s32(rows_15.val[1]);
158 int16x8_t row6 = vreinterpretq_s16_s32(rows_26.val[1]);
159 int16x8_t row7 = vreinterpretq_s16_s32(rows_37.val[1]);
160
161 /* Pass 2: process columns. */
162
163 tmp0 = vaddq_s16(row0, row7);
164 tmp7 = vsubq_s16(row0, row7);
165 tmp1 = vaddq_s16(row1, row6);
166 tmp6 = vsubq_s16(row1, row6);
167 tmp2 = vaddq_s16(row2, row5);
168 tmp5 = vsubq_s16(row2, row5);
169 tmp3 = vaddq_s16(row3, row4);
170 tmp4 = vsubq_s16(row3, row4);
171
172 /* Even part */
173 tmp10 = vaddq_s16(tmp0, tmp3); /* phase 2 */
174 tmp13 = vsubq_s16(tmp0, tmp3);
175 tmp11 = vaddq_s16(tmp1, tmp2);
176 tmp12 = vsubq_s16(tmp1, tmp2);
177
178 row0 = vaddq_s16(tmp10, tmp11); /* phase 3 */
179 row4 = vsubq_s16(tmp10, tmp11);
180
181 z1 = vqdmulhq_lane_s16(vaddq_s16(tmp12, tmp13), consts, 2);
182 row2 = vaddq_s16(tmp13, z1); /* phase 5 */
183 row6 = vsubq_s16(tmp13, z1);
184
185 /* Odd part */
186 tmp10 = vaddq_s16(tmp4, tmp5); /* phase 2 */
187 tmp11 = vaddq_s16(tmp5, tmp6);
188 tmp12 = vaddq_s16(tmp6, tmp7);
189
190 z5 = vqdmulhq_lane_s16(vsubq_s16(tmp10, tmp12), consts, 0);
191 z2 = vqdmulhq_lane_s16(tmp10, consts, 1);
192 z2 = vaddq_s16(z2, z5);
193 z4 = vqdmulhq_lane_s16(tmp12, consts, 3);
194 z5 = vaddq_s16(tmp12, z5);
195 z4 = vaddq_s16(z4, z5);
196 z3 = vqdmulhq_lane_s16(tmp11, consts, 2);
197
198 z11 = vaddq_s16(tmp7, z3); /* phase 5 */
199 z13 = vsubq_s16(tmp7, z3);
200
201 row5 = vaddq_s16(z13, z2); /* phase 6 */
202 row3 = vsubq_s16(z13, z2);
203 row1 = vaddq_s16(z11, z4);
204 row7 = vsubq_s16(z11, z4);
205
206 vst1q_s16(data + 0 * DCTSIZE, row0);
207 vst1q_s16(data + 1 * DCTSIZE, row1);
208 vst1q_s16(data + 2 * DCTSIZE, row2);
209 vst1q_s16(data + 3 * DCTSIZE, row3);
210 vst1q_s16(data + 4 * DCTSIZE, row4);
211 vst1q_s16(data + 5 * DCTSIZE, row5);
212 vst1q_s16(data + 6 * DCTSIZE, row6);
213 vst1q_s16(data + 7 * DCTSIZE, row7);
214 }
215