1 /*
2 * AltiVec optimizations for libjpeg-turbo
3 *
4 * Copyright (C) 2014, D. R. Commander.
5 * All rights reserved.
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 /* SLOW INTEGER FORWARD DCT */
24
25 #include "jsimd_altivec.h"
26
27
28 #define F_0_298 2446 /* FIX(0.298631336) */
29 #define F_0_390 3196 /* FIX(0.390180644) */
30 #define F_0_541 4433 /* FIX(0.541196100) */
31 #define F_0_765 6270 /* FIX(0.765366865) */
32 #define F_0_899 7373 /* FIX(0.899976223) */
33 #define F_1_175 9633 /* FIX(1.175875602) */
34 #define F_1_501 12299 /* FIX(1.501321110) */
35 #define F_1_847 15137 /* FIX(1.847759065) */
36 #define F_1_961 16069 /* FIX(1.961570560) */
37 #define F_2_053 16819 /* FIX(2.053119869) */
38 #define F_2_562 20995 /* FIX(2.562915447) */
39 #define F_3_072 25172 /* FIX(3.072711026) */
40
41 #define CONST_BITS 13
42 #define PASS1_BITS 2
43 #define DESCALE_P1 (CONST_BITS - PASS1_BITS)
44 #define DESCALE_P2 (CONST_BITS + PASS1_BITS)
45
46
47 #define DO_FDCT_COMMON(PASS) \
48 { \
49 /* (Original) \
50 * z1 = (tmp12 + tmp13) * 0.541196100; \
51 * data2 = z1 + tmp13 * 0.765366865; \
52 * data6 = z1 + tmp12 * -1.847759065; \
53 * \
54 * (This implementation) \
55 * data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100; \
56 * data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065); \
57 */ \
58 \
59 tmp1312l = vec_mergeh(tmp13, tmp12); \
60 tmp1312h = vec_mergel(tmp13, tmp12); \
61 \
62 out2l = vec_msums(tmp1312l, pw_f130_f054, pd_descale_p##PASS); \
63 out2h = vec_msums(tmp1312h, pw_f130_f054, pd_descale_p##PASS); \
64 out6l = vec_msums(tmp1312l, pw_f054_mf130, pd_descale_p##PASS); \
65 out6h = vec_msums(tmp1312h, pw_f054_mf130, pd_descale_p##PASS); \
66 \
67 out2l = vec_sra(out2l, descale_p##PASS); \
68 out2h = vec_sra(out2h, descale_p##PASS); \
69 out6l = vec_sra(out6l, descale_p##PASS); \
70 out6h = vec_sra(out6h, descale_p##PASS); \
71 \
72 out2 = vec_pack(out2l, out2h); \
73 out6 = vec_pack(out6l, out6h); \
74 \
75 /* Odd part */ \
76 \
77 z3 = vec_add(tmp4, tmp6); \
78 z4 = vec_add(tmp5, tmp7); \
79 \
80 /* (Original) \
81 * z5 = (z3 + z4) * 1.175875602; \
82 * z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; \
83 * z3 += z5; z4 += z5; \
84 * \
85 * (This implementation) \
86 * z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; \
87 * z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); \
88 */ \
89 \
90 z34l = vec_mergeh(z3, z4); \
91 z34h = vec_mergel(z3, z4); \
92 \
93 z3l = vec_msums(z34l, pw_mf078_f117, pd_descale_p##PASS); \
94 z3h = vec_msums(z34h, pw_mf078_f117, pd_descale_p##PASS); \
95 z4l = vec_msums(z34l, pw_f117_f078, pd_descale_p##PASS); \
96 z4h = vec_msums(z34h, pw_f117_f078, pd_descale_p##PASS); \
97 \
98 /* (Original) \
99 * z1 = tmp4 + tmp7; z2 = tmp5 + tmp6; \
100 * tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869; \
101 * tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110; \
102 * z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; \
103 * data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4; \
104 * data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4; \
105 * \
106 * (This implementation) \
107 * tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223; \
108 * tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447; \
109 * tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447); \
110 * tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223); \
111 * data7 = tmp4 + z3; data5 = tmp5 + z4; \
112 * data3 = tmp6 + z3; data1 = tmp7 + z4; \
113 */ \
114 \
115 tmp47l = vec_mergeh(tmp4, tmp7); \
116 tmp47h = vec_mergel(tmp4, tmp7); \
117 \
118 out7l = vec_msums(tmp47l, pw_mf060_mf089, z3l); \
119 out7h = vec_msums(tmp47h, pw_mf060_mf089, z3h); \
120 out1l = vec_msums(tmp47l, pw_mf089_f060, z4l); \
121 out1h = vec_msums(tmp47h, pw_mf089_f060, z4h); \
122 \
123 out7l = vec_sra(out7l, descale_p##PASS); \
124 out7h = vec_sra(out7h, descale_p##PASS); \
125 out1l = vec_sra(out1l, descale_p##PASS); \
126 out1h = vec_sra(out1h, descale_p##PASS); \
127 \
128 out7 = vec_pack(out7l, out7h); \
129 out1 = vec_pack(out1l, out1h); \
130 \
131 tmp56l = vec_mergeh(tmp5, tmp6); \
132 tmp56h = vec_mergel(tmp5, tmp6); \
133 \
134 out5l = vec_msums(tmp56l, pw_mf050_mf256, z4l); \
135 out5h = vec_msums(tmp56h, pw_mf050_mf256, z4h); \
136 out3l = vec_msums(tmp56l, pw_mf256_f050, z3l); \
137 out3h = vec_msums(tmp56h, pw_mf256_f050, z3h); \
138 \
139 out5l = vec_sra(out5l, descale_p##PASS); \
140 out5h = vec_sra(out5h, descale_p##PASS); \
141 out3l = vec_sra(out3l, descale_p##PASS); \
142 out3h = vec_sra(out3h, descale_p##PASS); \
143 \
144 out5 = vec_pack(out5l, out5h); \
145 out3 = vec_pack(out3l, out3h); \
146 }
147
148 #define DO_FDCT_PASS1() \
149 { \
150 /* Even part */ \
151 \
152 tmp10 = vec_add(tmp0, tmp3); \
153 tmp13 = vec_sub(tmp0, tmp3); \
154 tmp11 = vec_add(tmp1, tmp2); \
155 tmp12 = vec_sub(tmp1, tmp2); \
156 \
157 out0 = vec_add(tmp10, tmp11); \
158 out0 = vec_sl(out0, pass1_bits); \
159 out4 = vec_sub(tmp10, tmp11); \
160 out4 = vec_sl(out4, pass1_bits); \
161 \
162 DO_FDCT_COMMON(1); \
163 }
164
165 #define DO_FDCT_PASS2() \
166 { \
167 /* Even part */ \
168 \
169 tmp10 = vec_add(tmp0, tmp3); \
170 tmp13 = vec_sub(tmp0, tmp3); \
171 tmp11 = vec_add(tmp1, tmp2); \
172 tmp12 = vec_sub(tmp1, tmp2); \
173 \
174 out0 = vec_add(tmp10, tmp11); \
175 out0 = vec_add(out0, pw_descale_p2x); \
176 out0 = vec_sra(out0, pass1_bits); \
177 out4 = vec_sub(tmp10, tmp11); \
178 out4 = vec_add(out4, pw_descale_p2x); \
179 out4 = vec_sra(out4, pass1_bits); \
180 \
181 DO_FDCT_COMMON(2); \
182 }
183
184
185 void
jsimd_fdct_islow_altivec(DCTELEM * data)186 jsimd_fdct_islow_altivec (DCTELEM *data)
187 {
188 __vector short row0, row1, row2, row3, row4, row5, row6, row7,
189 col0, col1, col2, col3, col4, col5, col6, col7,
190 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp10, tmp11, tmp12, tmp13,
191 tmp47l, tmp47h, tmp56l, tmp56h, tmp1312l, tmp1312h,
192 z3, z4, z34l, z34h,
193 out0, out1, out2, out3, out4, out5, out6, out7;
194 __vector int z3l, z3h, z4l, z4h,
195 out1l, out1h, out2l, out2h, out3l, out3h, out5l, out5h, out6l, out6h,
196 out7l, out7h;
197
198 /* Constants */
199 __vector short
200 pw_f130_f054 = { __4X2(F_0_541 + F_0_765, F_0_541) },
201 pw_f054_mf130 = { __4X2(F_0_541, F_0_541 - F_1_847) },
202 pw_mf078_f117 = { __4X2(F_1_175 - F_1_961, F_1_175) },
203 pw_f117_f078 = { __4X2(F_1_175, F_1_175 - F_0_390) },
204 pw_mf060_mf089 = { __4X2(F_0_298 - F_0_899, -F_0_899) },
205 pw_mf089_f060 = { __4X2(-F_0_899, F_1_501 - F_0_899) },
206 pw_mf050_mf256 = { __4X2(F_2_053 - F_2_562, -F_2_562) },
207 pw_mf256_f050 = { __4X2(-F_2_562, F_3_072 - F_2_562) },
208 pw_descale_p2x = { __8X(1 << (PASS1_BITS - 1)) };
209 __vector unsigned short pass1_bits = { __8X(PASS1_BITS) };
210 __vector int pd_descale_p1 = { __4X(1 << (DESCALE_P1 - 1)) },
211 pd_descale_p2 = { __4X(1 << (DESCALE_P2 - 1)) };
212 __vector unsigned int descale_p1 = { __4X(DESCALE_P1) },
213 descale_p2 = { __4X(DESCALE_P2) };
214
215 /* Pass 1: process rows */
216
217 row0 = vec_ld(0, data);
218 row1 = vec_ld(16, data);
219 row2 = vec_ld(32, data);
220 row3 = vec_ld(48, data);
221 row4 = vec_ld(64, data);
222 row5 = vec_ld(80, data);
223 row6 = vec_ld(96, data);
224 row7 = vec_ld(112, data);
225
226 TRANSPOSE(row, col);
227
228 tmp0 = vec_add(col0, col7);
229 tmp7 = vec_sub(col0, col7);
230 tmp1 = vec_add(col1, col6);
231 tmp6 = vec_sub(col1, col6);
232 tmp2 = vec_add(col2, col5);
233 tmp5 = vec_sub(col2, col5);
234 tmp3 = vec_add(col3, col4);
235 tmp4 = vec_sub(col3, col4);
236
237 DO_FDCT_PASS1();
238
239 /* Pass 2: process columns */
240
241 TRANSPOSE(out, row);
242
243 tmp0 = vec_add(row0, row7);
244 tmp7 = vec_sub(row0, row7);
245 tmp1 = vec_add(row1, row6);
246 tmp6 = vec_sub(row1, row6);
247 tmp2 = vec_add(row2, row5);
248 tmp5 = vec_sub(row2, row5);
249 tmp3 = vec_add(row3, row4);
250 tmp4 = vec_sub(row3, row4);
251
252 DO_FDCT_PASS2();
253
254 vec_st(out0, 0, data);
255 vec_st(out1, 16, data);
256 vec_st(out2, 32, data);
257 vec_st(out3, 48, data);
258 vec_st(out4, 64, data);
259 vec_st(out5, 80, data);
260 vec_st(out6, 96, data);
261 vec_st(out7, 112, data);
262 }
263