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 #ifndef AOM_DSP_X86_INV_TXFM_SSE2_H_
13 #define AOM_DSP_X86_INV_TXFM_SSE2_H_
14
15 #include <emmintrin.h> // SSE2
16 #include "./aom_config.h"
17 #include "aom/aom_integer.h"
18 #include "aom_dsp/inv_txfm.h"
19 #include "aom_dsp/x86/txfm_common_sse2.h"
20
21 // perform 8x8 transpose
array_transpose_4x4(__m128i * res)22 static INLINE void array_transpose_4x4(__m128i *res) {
23 const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
24 const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
25
26 res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1);
27 res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1);
28 }
29
array_transpose_8x8(__m128i * in,__m128i * res)30 static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
31 const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
32 const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
33 const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
34 const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
35 const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
36 const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
37 const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
38 const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
39
40 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
41 const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
42 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
43 const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
44 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
45 const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
46 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
47 const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
48
49 res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
50 res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
51 res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
52 res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
53 res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
54 res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
55 res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
56 res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
57 }
58
59 #define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
60 out2, out3, out4, out5, out6, out7) \
61 { \
62 const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
63 const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
64 const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1); \
65 const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3); \
66 const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5); \
67 const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \
68 const __m128i tr0_6 = _mm_unpackhi_epi16(in4, in5); \
69 const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7); \
70 \
71 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
72 const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); \
73 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
74 const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); \
75 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
76 const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); \
77 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
78 const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); \
79 \
80 out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
81 out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
82 out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
83 out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
84 out4 = _mm_unpacklo_epi64(tr1_1, tr1_5); \
85 out5 = _mm_unpackhi_epi64(tr1_1, tr1_5); \
86 out6 = _mm_unpacklo_epi64(tr1_3, tr1_7); \
87 out7 = _mm_unpackhi_epi64(tr1_3, tr1_7); \
88 }
89
90 #define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1) \
91 { \
92 const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
93 const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
94 \
95 in0 = _mm_unpacklo_epi32(tr0_0, tr0_1); /* i1 i0 */ \
96 in1 = _mm_unpackhi_epi32(tr0_0, tr0_1); /* i3 i2 */ \
97 }
98
array_transpose_4X8(__m128i * in,__m128i * out)99 static INLINE void array_transpose_4X8(__m128i *in, __m128i *out) {
100 const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
101 const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
102 const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
103 const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
104
105 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
106 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
107 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
108 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
109
110 out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4);
111 out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4);
112 out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6);
113 out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6);
114 }
115
array_transpose_16x16(__m128i * res0,__m128i * res1)116 static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
117 __m128i tbuf[8];
118 array_transpose_8x8(res0, res0);
119 array_transpose_8x8(res1, tbuf);
120 array_transpose_8x8(res0 + 8, res1);
121 array_transpose_8x8(res1 + 8, res1 + 8);
122
123 res0[8] = tbuf[0];
124 res0[9] = tbuf[1];
125 res0[10] = tbuf[2];
126 res0[11] = tbuf[3];
127 res0[12] = tbuf[4];
128 res0[13] = tbuf[5];
129 res0[14] = tbuf[6];
130 res0[15] = tbuf[7];
131 }
132
133 // Function to allow 8 bit optimisations to be used when profile 0 is used with
134 // highbitdepth enabled
load_input_data(const tran_low_t * data)135 static INLINE __m128i load_input_data(const tran_low_t *data) {
136 if (sizeof(tran_low_t) == 4) {
137 return octa_set_epi16(data[0], data[1], data[2], data[3], data[4], data[5],
138 data[6], data[7]);
139 } else {
140 return _mm_load_si128((const __m128i *)data);
141 }
142 }
143
load_buffer_8x16(const tran_low_t * input,__m128i * in)144 static INLINE void load_buffer_8x16(const tran_low_t *input, __m128i *in) {
145 in[0] = load_input_data(input + 0 * 16);
146 in[1] = load_input_data(input + 1 * 16);
147 in[2] = load_input_data(input + 2 * 16);
148 in[3] = load_input_data(input + 3 * 16);
149 in[4] = load_input_data(input + 4 * 16);
150 in[5] = load_input_data(input + 5 * 16);
151 in[6] = load_input_data(input + 6 * 16);
152 in[7] = load_input_data(input + 7 * 16);
153
154 in[8] = load_input_data(input + 8 * 16);
155 in[9] = load_input_data(input + 9 * 16);
156 in[10] = load_input_data(input + 10 * 16);
157 in[11] = load_input_data(input + 11 * 16);
158 in[12] = load_input_data(input + 12 * 16);
159 in[13] = load_input_data(input + 13 * 16);
160 in[14] = load_input_data(input + 14 * 16);
161 in[15] = load_input_data(input + 15 * 16);
162 }
163
164 #define RECON_AND_STORE(dest, in_x) \
165 { \
166 __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
167 d0 = _mm_unpacklo_epi8(d0, zero); \
168 d0 = _mm_add_epi16(in_x, d0); \
169 d0 = _mm_packus_epi16(d0, d0); \
170 _mm_storel_epi64((__m128i *)(dest), d0); \
171 }
172
write_buffer_8x16(uint8_t * dest,__m128i * in,int stride)173 static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) {
174 const __m128i final_rounding = _mm_set1_epi16(1 << 5);
175 const __m128i zero = _mm_setzero_si128();
176 // Final rounding and shift
177 in[0] = _mm_adds_epi16(in[0], final_rounding);
178 in[1] = _mm_adds_epi16(in[1], final_rounding);
179 in[2] = _mm_adds_epi16(in[2], final_rounding);
180 in[3] = _mm_adds_epi16(in[3], final_rounding);
181 in[4] = _mm_adds_epi16(in[4], final_rounding);
182 in[5] = _mm_adds_epi16(in[5], final_rounding);
183 in[6] = _mm_adds_epi16(in[6], final_rounding);
184 in[7] = _mm_adds_epi16(in[7], final_rounding);
185 in[8] = _mm_adds_epi16(in[8], final_rounding);
186 in[9] = _mm_adds_epi16(in[9], final_rounding);
187 in[10] = _mm_adds_epi16(in[10], final_rounding);
188 in[11] = _mm_adds_epi16(in[11], final_rounding);
189 in[12] = _mm_adds_epi16(in[12], final_rounding);
190 in[13] = _mm_adds_epi16(in[13], final_rounding);
191 in[14] = _mm_adds_epi16(in[14], final_rounding);
192 in[15] = _mm_adds_epi16(in[15], final_rounding);
193
194 in[0] = _mm_srai_epi16(in[0], 6);
195 in[1] = _mm_srai_epi16(in[1], 6);
196 in[2] = _mm_srai_epi16(in[2], 6);
197 in[3] = _mm_srai_epi16(in[3], 6);
198 in[4] = _mm_srai_epi16(in[4], 6);
199 in[5] = _mm_srai_epi16(in[5], 6);
200 in[6] = _mm_srai_epi16(in[6], 6);
201 in[7] = _mm_srai_epi16(in[7], 6);
202 in[8] = _mm_srai_epi16(in[8], 6);
203 in[9] = _mm_srai_epi16(in[9], 6);
204 in[10] = _mm_srai_epi16(in[10], 6);
205 in[11] = _mm_srai_epi16(in[11], 6);
206 in[12] = _mm_srai_epi16(in[12], 6);
207 in[13] = _mm_srai_epi16(in[13], 6);
208 in[14] = _mm_srai_epi16(in[14], 6);
209 in[15] = _mm_srai_epi16(in[15], 6);
210
211 RECON_AND_STORE(dest + 0 * stride, in[0]);
212 RECON_AND_STORE(dest + 1 * stride, in[1]);
213 RECON_AND_STORE(dest + 2 * stride, in[2]);
214 RECON_AND_STORE(dest + 3 * stride, in[3]);
215 RECON_AND_STORE(dest + 4 * stride, in[4]);
216 RECON_AND_STORE(dest + 5 * stride, in[5]);
217 RECON_AND_STORE(dest + 6 * stride, in[6]);
218 RECON_AND_STORE(dest + 7 * stride, in[7]);
219 RECON_AND_STORE(dest + 8 * stride, in[8]);
220 RECON_AND_STORE(dest + 9 * stride, in[9]);
221 RECON_AND_STORE(dest + 10 * stride, in[10]);
222 RECON_AND_STORE(dest + 11 * stride, in[11]);
223 RECON_AND_STORE(dest + 12 * stride, in[12]);
224 RECON_AND_STORE(dest + 13 * stride, in[13]);
225 RECON_AND_STORE(dest + 14 * stride, in[14]);
226 RECON_AND_STORE(dest + 15 * stride, in[15]);
227 }
228
229 #define TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, out0, out1, out2, out3) \
230 { \
231 const __m128i tr0_0 = _mm_unpackhi_epi16(tmp0, tmp1); \
232 const __m128i tr0_1 = _mm_unpacklo_epi16(tmp1, tmp0); \
233 const __m128i tr0_4 = _mm_unpacklo_epi16(tmp2, tmp3); \
234 const __m128i tr0_5 = _mm_unpackhi_epi16(tmp3, tmp2); \
235 \
236 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
237 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
238 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
239 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
240 \
241 out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
242 out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
243 out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
244 out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
245 }
246
247 #define TRANSPOSE_8X8_10(in0, in1, in2, in3, out0, out1) \
248 { \
249 const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
250 const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
251 out0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
252 out1 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
253 }
254
255 void iadst16_8col(__m128i *in);
256 void idct16_8col(__m128i *in);
257 void aom_idct4_sse2(__m128i *in);
258 void aom_idct8_sse2(__m128i *in);
259 void aom_idct16_sse2(__m128i *in0, __m128i *in1);
260 void aom_iadst4_sse2(__m128i *in);
261 void aom_iadst8_sse2(__m128i *in);
262 void aom_iadst16_sse2(__m128i *in0, __m128i *in1);
263 void idct32_8col(__m128i *in0, __m128i *in1);
264
265 #endif // AOM_DSP_X86_INV_TXFM_SSE2_H_
266