1 // Copyright 2014 Google Inc. All Rights Reserved.
2 //
3 // Use of this source code is governed by a BSD-style license
4 // that can be found in the COPYING file in the root of the source
5 // tree. An additional intellectual property rights grant can be found
6 // in the file PATENTS. All contributing project authors may
7 // be found in the AUTHORS file in the root of the source tree.
8 // -----------------------------------------------------------------------------
9 //
10 // NEON common code.
11
12 #ifndef WEBP_DSP_NEON_H_
13 #define WEBP_DSP_NEON_H_
14
15 #include "src/dsp/dsp.h"
16
17 #if defined(WEBP_USE_NEON)
18
19 #include <arm_neon.h>
20
21 // Right now, some intrinsics functions seem slower, so we disable them
22 // everywhere except newer clang/gcc or aarch64 where the inline assembly is
23 // incompatible.
24 #if LOCAL_CLANG_PREREQ(3,8) || LOCAL_GCC_PREREQ(4,9) || defined(__aarch64__)
25 #define WEBP_USE_INTRINSICS // use intrinsics when possible
26 #endif
27
28 #define INIT_VECTOR2(v, a, b) do { \
29 v.val[0] = a; \
30 v.val[1] = b; \
31 } while (0)
32
33 #define INIT_VECTOR3(v, a, b, c) do { \
34 v.val[0] = a; \
35 v.val[1] = b; \
36 v.val[2] = c; \
37 } while (0)
38
39 #define INIT_VECTOR4(v, a, b, c, d) do { \
40 v.val[0] = a; \
41 v.val[1] = b; \
42 v.val[2] = c; \
43 v.val[3] = d; \
44 } while (0)
45
46 // if using intrinsics, this flag avoids some functions that make gcc-4.6.3
47 // crash ("internal compiler error: in immed_double_const, at emit-rtl.").
48 // (probably similar to gcc.gnu.org/bugzilla/show_bug.cgi?id=48183)
49 #if !(LOCAL_CLANG_PREREQ(3,8) || LOCAL_GCC_PREREQ(4,8) || defined(__aarch64__))
50 #define WORK_AROUND_GCC
51 #endif
52
Transpose4x4_NEON(const int32x4x4_t rows)53 static WEBP_INLINE int32x4x4_t Transpose4x4_NEON(const int32x4x4_t rows) {
54 uint64x2x2_t row01, row23;
55
56 row01.val[0] = vreinterpretq_u64_s32(rows.val[0]);
57 row01.val[1] = vreinterpretq_u64_s32(rows.val[1]);
58 row23.val[0] = vreinterpretq_u64_s32(rows.val[2]);
59 row23.val[1] = vreinterpretq_u64_s32(rows.val[3]);
60 // Transpose 64-bit values (there's no vswp equivalent)
61 {
62 const uint64x1_t row0h = vget_high_u64(row01.val[0]);
63 const uint64x1_t row2l = vget_low_u64(row23.val[0]);
64 const uint64x1_t row1h = vget_high_u64(row01.val[1]);
65 const uint64x1_t row3l = vget_low_u64(row23.val[1]);
66 row01.val[0] = vcombine_u64(vget_low_u64(row01.val[0]), row2l);
67 row23.val[0] = vcombine_u64(row0h, vget_high_u64(row23.val[0]));
68 row01.val[1] = vcombine_u64(vget_low_u64(row01.val[1]), row3l);
69 row23.val[1] = vcombine_u64(row1h, vget_high_u64(row23.val[1]));
70 }
71 {
72 const int32x4x2_t out01 = vtrnq_s32(vreinterpretq_s32_u64(row01.val[0]),
73 vreinterpretq_s32_u64(row01.val[1]));
74 const int32x4x2_t out23 = vtrnq_s32(vreinterpretq_s32_u64(row23.val[0]),
75 vreinterpretq_s32_u64(row23.val[1]));
76 int32x4x4_t out;
77 out.val[0] = out01.val[0];
78 out.val[1] = out01.val[1];
79 out.val[2] = out23.val[0];
80 out.val[3] = out23.val[1];
81 return out;
82 }
83 }
84
85 #if 0 // Useful debug macro.
86 #include <stdio.h>
87 #define PRINT_REG(REG, SIZE) do { \
88 int i; \
89 printf("%s \t[%d]: 0x", #REG, SIZE); \
90 if (SIZE == 8) { \
91 uint8_t _tmp[8]; \
92 vst1_u8(_tmp, (REG)); \
93 for (i = 0; i < 8; ++i) printf("%.2x ", _tmp[i]); \
94 } else if (SIZE == 16) { \
95 uint16_t _tmp[4]; \
96 vst1_u16(_tmp, (REG)); \
97 for (i = 0; i < 4; ++i) printf("%.4x ", _tmp[i]); \
98 } \
99 printf("\n"); \
100 } while (0)
101 #endif
102
103 #endif // WEBP_USE_NEON
104 #endif // WEBP_DSP_NEON_H_
105