1 /*
2 * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #ifndef VPX_DSP_X86_INV_TXFM_SSE2_H_
12 #define VPX_DSP_X86_INV_TXFM_SSE2_H_
13
14 #include <emmintrin.h> // SSE2
15 #include "./vpx_config.h"
16 #include "vpx/vpx_integer.h"
17 #include "vpx_dsp/inv_txfm.h"
18 #include "vpx_dsp/x86/txfm_common_sse2.h"
19
20 // perform 8x8 transpose
array_transpose_8x8(__m128i * in,__m128i * res)21 static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
22 const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
23 const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
24 const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
25 const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
26 const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
27 const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
28 const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
29 const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
30
31 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
32 const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
33 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
34 const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
35 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
36 const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
37 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
38 const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
39
40 res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
41 res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
42 res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
43 res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
44 res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
45 res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
46 res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
47 res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
48 }
49
50 #define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1) \
51 { \
52 const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
53 const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
54 \
55 in0 = _mm_unpacklo_epi32(tr0_0, tr0_1); /* i1 i0 */ \
56 in1 = _mm_unpackhi_epi32(tr0_0, tr0_1); /* i3 i2 */ \
57 }
58
array_transpose_4X8(__m128i * in,__m128i * out)59 static INLINE void array_transpose_4X8(__m128i *in, __m128i * out) {
60 const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
61 const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
62 const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
63 const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
64
65 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
66 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
67 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
68 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
69
70 out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4);
71 out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4);
72 out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6);
73 out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6);
74 }
75
array_transpose_16x16(__m128i * res0,__m128i * res1)76 static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
77 __m128i tbuf[8];
78 array_transpose_8x8(res0, res0);
79 array_transpose_8x8(res1, tbuf);
80 array_transpose_8x8(res0 + 8, res1);
81 array_transpose_8x8(res1 + 8, res1 + 8);
82
83 res0[8] = tbuf[0];
84 res0[9] = tbuf[1];
85 res0[10] = tbuf[2];
86 res0[11] = tbuf[3];
87 res0[12] = tbuf[4];
88 res0[13] = tbuf[5];
89 res0[14] = tbuf[6];
90 res0[15] = tbuf[7];
91 }
92
93 // Function to allow 8 bit optimisations to be used when profile 0 is used with
94 // highbitdepth enabled
load_input_data(const tran_low_t * data)95 static INLINE __m128i load_input_data(const tran_low_t *data) {
96 #if CONFIG_VP9_HIGHBITDEPTH
97 return octa_set_epi16(data[0], data[1], data[2], data[3], data[4], data[5],
98 data[6], data[7]);
99 #else
100 return _mm_load_si128((const __m128i *)data);
101 #endif
102 }
103
load_buffer_8x16(const tran_low_t * input,__m128i * in)104 static INLINE void load_buffer_8x16(const tran_low_t *input, __m128i *in) {
105 in[0] = load_input_data(input + 0 * 16);
106 in[1] = load_input_data(input + 1 * 16);
107 in[2] = load_input_data(input + 2 * 16);
108 in[3] = load_input_data(input + 3 * 16);
109 in[4] = load_input_data(input + 4 * 16);
110 in[5] = load_input_data(input + 5 * 16);
111 in[6] = load_input_data(input + 6 * 16);
112 in[7] = load_input_data(input + 7 * 16);
113
114 in[8] = load_input_data(input + 8 * 16);
115 in[9] = load_input_data(input + 9 * 16);
116 in[10] = load_input_data(input + 10 * 16);
117 in[11] = load_input_data(input + 11 * 16);
118 in[12] = load_input_data(input + 12 * 16);
119 in[13] = load_input_data(input + 13 * 16);
120 in[14] = load_input_data(input + 14 * 16);
121 in[15] = load_input_data(input + 15 * 16);
122 }
123
124 #define RECON_AND_STORE(dest, in_x) \
125 { \
126 __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
127 d0 = _mm_unpacklo_epi8(d0, zero); \
128 d0 = _mm_add_epi16(in_x, d0); \
129 d0 = _mm_packus_epi16(d0, d0); \
130 _mm_storel_epi64((__m128i *)(dest), d0); \
131 }
132
write_buffer_8x16(uint8_t * dest,__m128i * in,int stride)133 static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) {
134 const __m128i final_rounding = _mm_set1_epi16(1<<5);
135 const __m128i zero = _mm_setzero_si128();
136 // Final rounding and shift
137 in[0] = _mm_adds_epi16(in[0], final_rounding);
138 in[1] = _mm_adds_epi16(in[1], final_rounding);
139 in[2] = _mm_adds_epi16(in[2], final_rounding);
140 in[3] = _mm_adds_epi16(in[3], final_rounding);
141 in[4] = _mm_adds_epi16(in[4], final_rounding);
142 in[5] = _mm_adds_epi16(in[5], final_rounding);
143 in[6] = _mm_adds_epi16(in[6], final_rounding);
144 in[7] = _mm_adds_epi16(in[7], final_rounding);
145 in[8] = _mm_adds_epi16(in[8], final_rounding);
146 in[9] = _mm_adds_epi16(in[9], final_rounding);
147 in[10] = _mm_adds_epi16(in[10], final_rounding);
148 in[11] = _mm_adds_epi16(in[11], final_rounding);
149 in[12] = _mm_adds_epi16(in[12], final_rounding);
150 in[13] = _mm_adds_epi16(in[13], final_rounding);
151 in[14] = _mm_adds_epi16(in[14], final_rounding);
152 in[15] = _mm_adds_epi16(in[15], final_rounding);
153
154 in[0] = _mm_srai_epi16(in[0], 6);
155 in[1] = _mm_srai_epi16(in[1], 6);
156 in[2] = _mm_srai_epi16(in[2], 6);
157 in[3] = _mm_srai_epi16(in[3], 6);
158 in[4] = _mm_srai_epi16(in[4], 6);
159 in[5] = _mm_srai_epi16(in[5], 6);
160 in[6] = _mm_srai_epi16(in[6], 6);
161 in[7] = _mm_srai_epi16(in[7], 6);
162 in[8] = _mm_srai_epi16(in[8], 6);
163 in[9] = _mm_srai_epi16(in[9], 6);
164 in[10] = _mm_srai_epi16(in[10], 6);
165 in[11] = _mm_srai_epi16(in[11], 6);
166 in[12] = _mm_srai_epi16(in[12], 6);
167 in[13] = _mm_srai_epi16(in[13], 6);
168 in[14] = _mm_srai_epi16(in[14], 6);
169 in[15] = _mm_srai_epi16(in[15], 6);
170
171 RECON_AND_STORE(dest + 0 * stride, in[0]);
172 RECON_AND_STORE(dest + 1 * stride, in[1]);
173 RECON_AND_STORE(dest + 2 * stride, in[2]);
174 RECON_AND_STORE(dest + 3 * stride, in[3]);
175 RECON_AND_STORE(dest + 4 * stride, in[4]);
176 RECON_AND_STORE(dest + 5 * stride, in[5]);
177 RECON_AND_STORE(dest + 6 * stride, in[6]);
178 RECON_AND_STORE(dest + 7 * stride, in[7]);
179 RECON_AND_STORE(dest + 8 * stride, in[8]);
180 RECON_AND_STORE(dest + 9 * stride, in[9]);
181 RECON_AND_STORE(dest + 10 * stride, in[10]);
182 RECON_AND_STORE(dest + 11 * stride, in[11]);
183 RECON_AND_STORE(dest + 12 * stride, in[12]);
184 RECON_AND_STORE(dest + 13 * stride, in[13]);
185 RECON_AND_STORE(dest + 14 * stride, in[14]);
186 RECON_AND_STORE(dest + 15 * stride, in[15]);
187 }
188
189 void idct4_sse2(__m128i *in);
190 void idct8_sse2(__m128i *in);
191 void idct16_sse2(__m128i *in0, __m128i *in1);
192 void iadst4_sse2(__m128i *in);
193 void iadst8_sse2(__m128i *in);
194 void iadst16_sse2(__m128i *in0, __m128i *in1);
195
196 #endif // VPX_DSP_X86_INV_TXFM_SSE2_H_
197