blob: d80c691e3d905e37d5675abe8a76e9f345d47cf1 [file] [log] [blame]
Christophe Fontaine33e81952016-12-19 14:41:52 +01001/*
2 * Copyright (c) 2015 Cisco and/or its affiliates.
3 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at:
6 *
7 * http://www.apache.org/licenses/LICENSE-2.0
8 *
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS,
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License.
14 */
15
16#ifndef included_vector_neon_h
17#define included_vector_neon_h
18#include <arm_neon.h>
19
Christophe Fontaine33e81952016-12-19 14:41:52 +010020/* Arithmetic */
Christophe Fontaine33e81952016-12-19 14:41:52 +010021#define u16x8_sub_saturate(a,b) vsubq_u16(a,b)
22#define i16x8_sub_saturate(a,b) vsubq_s16(a,b)
Lijian Zhangf5942d52018-10-31 13:35:20 +080023/* Dummy. Aid making uniform macros */
24#define vreinterpretq_u8_u8(a) a
Lijian.Zhange6a47cf2019-03-12 18:32:39 +080025/* Implement the missing intrinsics to make uniform macros */
26#define vminvq_u64(x) \
27({ \
28 u64 x0 = vgetq_lane_u64(x, 0); \
29 u64 x1 = vgetq_lane_u64(x, 1); \
30 x0 < x1 ? x0 : x1; \
31})
Christophe Fontaine33e81952016-12-19 14:41:52 +010032
Gabriel Ganneb81831d2017-12-05 17:33:37 +010033/* Converts all ones/zeros compare mask to bitmap. */
34always_inline u32
Lijian Zhanga22ba882018-10-25 18:50:33 +080035u8x16_compare_byte_mask (u8x16 v)
Gabriel Ganneb81831d2017-12-05 17:33:37 +010036{
Lijian Zhanga22ba882018-10-25 18:50:33 +080037 uint8x16_t mask = { 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
38 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80
39 };
40 /* v --> [0xFF, 0x00, 0xFF, 0xFF, 0xFF, 0x00, 0xFF, 0x00, ... ] */
41 uint8x16_t x = vandq_u8 (v, mask);
42 /* after v & mask,
43 * x --> [0x01, 0x00, 0x04, 0x08, 0x10, 0x00, 0x40, 0x00, ... ] */
44 uint64x2_t x64 = vpaddlq_u32 (vpaddlq_u16 (vpaddlq_u8 (x)));
45 /* after merge, x64 --> [0x5D, 0x.. ] */
46 return (u32) (vgetq_lane_u64 (x64, 0) + (vgetq_lane_u64 (x64, 1) << 8));
Gabriel Ganneb81831d2017-12-05 17:33:37 +010047}
Christophe Fontaine33e81952016-12-19 14:41:52 +010048
Sirshak Das536953d2018-06-26 13:08:46 -050049/* *INDENT-OFF* */
50#define foreach_neon_vec128i \
51 _(i,8,16,s8) _(i,16,8,s16) _(i,32,4,s32) _(i,64,2,s64)
52#define foreach_neon_vec128u \
53 _(u,8,16,u8) _(u,16,8,u16) _(u,32,4,u32) _(u,64,2,u64)
54#define foreach_neon_vec128f \
55 _(f,32,4,f32) _(f,64,2,f64)
Adrian Oanca22ac59b2018-02-23 16:27:41 +010056
Sirshak Das536953d2018-06-26 13:08:46 -050057#define _(t, s, c, i) \
58static_always_inline t##s##x##c \
59t##s##x##c##_splat (t##s x) \
60{ return (t##s##x##c) vdupq_n_##i (x); } \
61\
62static_always_inline t##s##x##c \
63t##s##x##c##_load_unaligned (void *p) \
64{ return (t##s##x##c) vld1q_##i (p); } \
65\
66static_always_inline void \
67t##s##x##c##_store_unaligned (t##s##x##c v, void *p) \
68{ vst1q_##i (p, v); } \
69\
70static_always_inline int \
71t##s##x##c##_is_all_zero (t##s##x##c x) \
Lijian.Zhange6a47cf2019-03-12 18:32:39 +080072{ return !!(vminvq_u##s (vceqq_##i (vdupq_n_##i(0), x))); } \
Sirshak Das536953d2018-06-26 13:08:46 -050073\
74static_always_inline int \
75t##s##x##c##_is_equal (t##s##x##c a, t##s##x##c b) \
Lijian.Zhange6a47cf2019-03-12 18:32:39 +080076{ return !!(vminvq_u##s (vceqq_##i (a, b))); } \
Sirshak Das536953d2018-06-26 13:08:46 -050077\
78static_always_inline int \
79t##s##x##c##_is_all_equal (t##s##x##c v, t##s x) \
80{ return t##s##x##c##_is_equal (v, t##s##x##c##_splat (x)); }; \
Lijian Zhangf5942d52018-10-31 13:35:20 +080081\
82static_always_inline u32 \
83t##s##x##c##_zero_byte_mask (t##s##x##c x) \
84{ uint8x16_t v = vreinterpretq_u8_u##s (vceqq_##i (vdupq_n_##i(0), x)); \
85 return u8x16_compare_byte_mask (v); } \
Lijian.Zhang37c83782019-04-04 15:26:26 +080086\
87static_always_inline u##s##x##c \
88t##s##x##c##_is_greater (t##s##x##c a, t##s##x##c b) \
89{ return (u##s##x##c) vcgtq_##i (a, b); } \
90\
91static_always_inline t##s##x##c \
92t##s##x##c##_blend (t##s##x##c dst, t##s##x##c src, u##s##x##c mask) \
93{ return (t##s##x##c) vbslq_##i (mask, src, dst); }
Adrian Oanca22ac59b2018-02-23 16:27:41 +010094
Sirshak Das536953d2018-06-26 13:08:46 -050095foreach_neon_vec128i foreach_neon_vec128u
96
97#undef _
98/* *INDENT-ON* */
99
Sirshak Dasafc725a2018-07-27 01:13:33 -0500100static_always_inline u16x8
101u16x8_byte_swap (u16x8 v)
102{
Sirshak Das6da42de2018-08-22 12:02:04 +0800103 return (u16x8) vrev16q_u8 ((u8x16) v);
Sirshak Dasafc725a2018-07-27 01:13:33 -0500104}
105
Damjan Mariondd648aa2020-03-12 11:56:00 +0100106static_always_inline u32x4
107u32x4_byte_swap (u32x4 v)
108{
109 return vrev64q_u32 (v);
110}
111
Sirshak Dasafc725a2018-07-27 01:13:33 -0500112static_always_inline u8x16
113u8x16_shuffle (u8x16 v, u8x16 m)
114{
115 return (u8x16) vqtbl1q_u8 (v, m);
116}
117
Sirshak Das61f325d2018-08-03 11:24:51 -0500118static_always_inline u32x4
119u32x4_hadd (u32x4 v1, u32x4 v2)
120{
121 return (u32x4) vpaddq_u32 (v1, v2);
122}
123
Sirshak Das8e5d5db2018-08-22 14:04:33 +0800124static_always_inline u64x2
125u32x4_extend_to_u64x2 (u32x4 v)
126{
127 return vmovl_u32 (vget_low_u32 (v));
128}
129
Sirshak Das759226e2018-08-22 08:46:52 +0800130static_always_inline u64x2
131u32x4_extend_to_u64x2_high (u32x4 v)
132{
Sirshak Das4f611172018-10-09 11:28:44 -0500133 return vmovl_high_u32 (v);
Sirshak Das759226e2018-08-22 08:46:52 +0800134}
135
Lijian Zhanga22ba882018-10-25 18:50:33 +0800136/* Creates a mask made up of the MSB of each byte of the source vector */
137static_always_inline u16
138u8x16_msb_mask (u8x16 v)
139{
140 int8x16_t shift =
141 { -7, -6, -5, -4, -3, -2, -1, 0, -7, -6, -5, -4, -3, -2, -1, 0 };
142 /* v --> [0x80, 0x7F, 0xF0, 0xAF, 0xF0, 0x00, 0xF2, 0x00, ... ] */
143 uint8x16_t x = vshlq_u8 (vandq_u8 (v, vdupq_n_u8 (0x80)), shift);
144 /* after (v & 0x80) >> shift,
145 * x --> [0x01, 0x00, 0x04, 0x08, 0x10, 0x00, 0x40, 0x00, ... ] */
146 uint64x2_t x64 = vpaddlq_u32 (vpaddlq_u16 (vpaddlq_u8 (x)));
147 /* after merge, x64 --> [0x5D, 0x.. ] */
148 return (u16) (vgetq_lane_u64 (x64, 0) + (vgetq_lane_u64 (x64, 1) << 8));
149}
150
Lijian.Zhang107aa832019-05-29 17:08:47 +0800151static_always_inline u64x2
152u64x2_gather (void *p0, void *p1)
153{
154 u64x2 r = vdupq_n_u64 (*(u64 *) p0);
155 r = vsetq_lane_u64 (*(u64 *) p1, r, 1);
156 return r;
157}
158
159static_always_inline u32x4
160u32x4_gather (void *p0, void *p1, void *p2, void *p3)
161{
162 u32x4 r = vdupq_n_u32 (*(u32 *) p0);
163 r = vsetq_lane_u32 (*(u32 *) p1, r, 1);
164 r = vsetq_lane_u32 (*(u32 *) p2, r, 2);
165 r = vsetq_lane_u32 (*(u32 *) p3, r, 3);
166 return r;
167}
168
Lijian.Zhang0f984512019-05-29 17:13:42 +0800169static_always_inline void
170u64x2_scatter (u64x2 r, void *p0, void *p1)
171{
172 *(u64 *) p0 = vgetq_lane_u64 (r, 0);
173 *(u64 *) p1 = vgetq_lane_u64 (r, 1);
174}
175
176static_always_inline void
177u32x4_scatter (u32x4 r, void *p0, void *p1, void *p2, void *p3)
178{
179 *(u32 *) p0 = vgetq_lane_u32 (r, 0);
180 *(u32 *) p1 = vgetq_lane_u32 (r, 1);
181 *(u32 *) p2 = vgetq_lane_u32 (r, 2);
182 *(u32 *) p3 = vgetq_lane_u32 (r, 3);
183}
184
Damjan Marion776644e2020-01-31 10:24:07 +0100185static_always_inline u32
186u32x4_min_scalar (u32x4 v)
187{
188 return vminvq_u32 (v);
189}
190
Damjan Marion4339c362020-02-18 15:12:07 +0100191#define u8x16_word_shift_left(x,n) vextq_u8(u8x16_splat (0), x, 16 - n)
192#define u8x16_word_shift_right(x,n) vextq_u8(x, u8x16_splat (0), n)
Damjan Marion93723742020-02-12 20:27:46 +0100193
Damjan Marion622b5ce2020-02-12 10:59:14 +0100194static_always_inline u8x16
195u8x16_reflect (u8x16 v)
196{
197 u8x16 mask = {
198 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
199 };
200 return (u8x16) vqtbl1q_u8 (v, mask);
201}
202
Damjan Marionf75defa2020-02-13 18:14:06 +0100203static_always_inline u8x16
204u8x16_xor3 (u8x16 a, u8x16 b, u8x16 c)
205{
206#if __GNUC__ == 8 && __ARM_FEATURE_SHA3 == 1
207 u8x16 r;
208__asm__ ("eor3 %0.16b,%1.16b,%2.16b,%3.16b": "=w" (r): "0" (a), "w" (b), "w" (c):);
209 return r;
210#endif
211 return a ^ b ^ c;
212}
213
Lijian Zhanga22ba882018-10-25 18:50:33 +0800214#define CLIB_HAVE_VEC128_MSB_MASK
215
Sirshak Das536953d2018-06-26 13:08:46 -0500216#define CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE
217#define CLIB_VEC128_SPLAT_DEFINED
Christophe Fontaine33e81952016-12-19 14:41:52 +0100218#endif /* included_vector_neon_h */
219
220/*
221 * fd.io coding-style-patch-verification: ON
222 *
223 * Local Variables:
224 * eval: (c-set-style "gnu")
225 * End:
226 */