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_AOM_DSP_X86_FWD_TXFM_SSE2_H_
13 #define AOM_AOM_DSP_X86_FWD_TXFM_SSE2_H_
14
15 #ifdef __cplusplus
16 extern "C" {
17 #endif
18
k_madd_epi32(__m128i a,__m128i b)19 static inline __m128i k_madd_epi32(__m128i a, __m128i b) {
20 __m128i buf0, buf1;
21 buf0 = _mm_mul_epu32(a, b);
22 a = _mm_srli_epi64(a, 32);
23 b = _mm_srli_epi64(b, 32);
24 buf1 = _mm_mul_epu32(a, b);
25 return _mm_add_epi64(buf0, buf1);
26 }
27
k_packs_epi64(__m128i a,__m128i b)28 static inline __m128i k_packs_epi64(__m128i a, __m128i b) {
29 __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
30 __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
31 return _mm_unpacklo_epi64(buf0, buf1);
32 }
33
check_epi16_overflow_x2(const __m128i * preg0,const __m128i * preg1)34 static inline int check_epi16_overflow_x2(const __m128i *preg0,
35 const __m128i *preg1) {
36 const __m128i max_overflow = _mm_set1_epi16(0x7fff);
37 const __m128i min_overflow = _mm_set1_epi16((short)0x8000);
38 __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
39 _mm_cmpeq_epi16(*preg0, min_overflow));
40 __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
41 _mm_cmpeq_epi16(*preg1, min_overflow));
42 cmp0 = _mm_or_si128(cmp0, cmp1);
43 return _mm_movemask_epi8(cmp0);
44 }
45
check_epi16_overflow_x4(const __m128i * preg0,const __m128i * preg1,const __m128i * preg2,const __m128i * preg3)46 static inline int check_epi16_overflow_x4(const __m128i *preg0,
47 const __m128i *preg1,
48 const __m128i *preg2,
49 const __m128i *preg3) {
50 const __m128i max_overflow = _mm_set1_epi16(0x7fff);
51 const __m128i min_overflow = _mm_set1_epi16((short)0x8000);
52 __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
53 _mm_cmpeq_epi16(*preg0, min_overflow));
54 __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
55 _mm_cmpeq_epi16(*preg1, min_overflow));
56 __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(*preg2, max_overflow),
57 _mm_cmpeq_epi16(*preg2, min_overflow));
58 __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(*preg3, max_overflow),
59 _mm_cmpeq_epi16(*preg3, min_overflow));
60 cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3));
61 return _mm_movemask_epi8(cmp0);
62 }
63
check_epi16_overflow_x8(const __m128i * preg0,const __m128i * preg1,const __m128i * preg2,const __m128i * preg3,const __m128i * preg4,const __m128i * preg5,const __m128i * preg6,const __m128i * preg7)64 static inline int check_epi16_overflow_x8(
65 const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
66 const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
67 const __m128i *preg6, const __m128i *preg7) {
68 int res0, res1;
69 res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
70 res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
71 return res0 + res1;
72 }
73
check_epi16_overflow_x12(const __m128i * preg0,const __m128i * preg1,const __m128i * preg2,const __m128i * preg3,const __m128i * preg4,const __m128i * preg5,const __m128i * preg6,const __m128i * preg7,const __m128i * preg8,const __m128i * preg9,const __m128i * preg10,const __m128i * preg11)74 static inline int check_epi16_overflow_x12(
75 const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
76 const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
77 const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
78 const __m128i *preg9, const __m128i *preg10, const __m128i *preg11) {
79 int res0, res1;
80 res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
81 res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
82 if (!res0) res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
83 return res0 + res1;
84 }
85
check_epi16_overflow_x16(const __m128i * preg0,const __m128i * preg1,const __m128i * preg2,const __m128i * preg3,const __m128i * preg4,const __m128i * preg5,const __m128i * preg6,const __m128i * preg7,const __m128i * preg8,const __m128i * preg9,const __m128i * preg10,const __m128i * preg11,const __m128i * preg12,const __m128i * preg13,const __m128i * preg14,const __m128i * preg15)86 static inline int check_epi16_overflow_x16(
87 const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
88 const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
89 const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
90 const __m128i *preg9, const __m128i *preg10, const __m128i *preg11,
91 const __m128i *preg12, const __m128i *preg13, const __m128i *preg14,
92 const __m128i *preg15) {
93 int res0, res1;
94 res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
95 res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
96 if (!res0) {
97 res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
98 if (!res1) res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
99 }
100 return res0 + res1;
101 }
102
check_epi16_overflow_x32(const __m128i * preg0,const __m128i * preg1,const __m128i * preg2,const __m128i * preg3,const __m128i * preg4,const __m128i * preg5,const __m128i * preg6,const __m128i * preg7,const __m128i * preg8,const __m128i * preg9,const __m128i * preg10,const __m128i * preg11,const __m128i * preg12,const __m128i * preg13,const __m128i * preg14,const __m128i * preg15,const __m128i * preg16,const __m128i * preg17,const __m128i * preg18,const __m128i * preg19,const __m128i * preg20,const __m128i * preg21,const __m128i * preg22,const __m128i * preg23,const __m128i * preg24,const __m128i * preg25,const __m128i * preg26,const __m128i * preg27,const __m128i * preg28,const __m128i * preg29,const __m128i * preg30,const __m128i * preg31)103 static inline int check_epi16_overflow_x32(
104 const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
105 const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
106 const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
107 const __m128i *preg9, const __m128i *preg10, const __m128i *preg11,
108 const __m128i *preg12, const __m128i *preg13, const __m128i *preg14,
109 const __m128i *preg15, const __m128i *preg16, const __m128i *preg17,
110 const __m128i *preg18, const __m128i *preg19, const __m128i *preg20,
111 const __m128i *preg21, const __m128i *preg22, const __m128i *preg23,
112 const __m128i *preg24, const __m128i *preg25, const __m128i *preg26,
113 const __m128i *preg27, const __m128i *preg28, const __m128i *preg29,
114 const __m128i *preg30, const __m128i *preg31) {
115 int res0, res1;
116 res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
117 res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
118 if (!res0) {
119 res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
120 if (!res1) {
121 res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
122 if (!res0) {
123 res0 = check_epi16_overflow_x4(preg16, preg17, preg18, preg19);
124 if (!res1) {
125 res1 = check_epi16_overflow_x4(preg20, preg21, preg22, preg23);
126 if (!res0) {
127 res0 = check_epi16_overflow_x4(preg24, preg25, preg26, preg27);
128 if (!res1)
129 res1 = check_epi16_overflow_x4(preg28, preg29, preg30, preg31);
130 }
131 }
132 }
133 }
134 }
135 return res0 + res1;
136 }
137
store_output(const __m128i * poutput,tran_low_t * dst_ptr)138 static inline void store_output(const __m128i *poutput, tran_low_t *dst_ptr) {
139 const __m128i zero = _mm_setzero_si128();
140 const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
141 __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
142 __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
143 _mm_store_si128((__m128i *)(dst_ptr), out0);
144 _mm_store_si128((__m128i *)(dst_ptr + 4), out1);
145 }
146
storeu_output(const __m128i * poutput,tran_low_t * dst_ptr)147 static inline void storeu_output(const __m128i *poutput, tran_low_t *dst_ptr) {
148 const __m128i zero = _mm_setzero_si128();
149 const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
150 __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
151 __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
152 _mm_storeu_si128((__m128i *)(dst_ptr), out0);
153 _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1);
154 }
155
156 #ifdef __cplusplus
157 } // extern "C"
158 #endif
159
160 #endif // AOM_AOM_DSP_X86_FWD_TXFM_SSE2_H_
161