FD.io VPP  v20.09-64-g4f7b92f0a
Vector Packet Processing
vector_neon.h
Go to the documentation of this file.
1 /*
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 
20 /* Arithmetic */
21 #define u16x8_sub_saturate(a,b) vsubq_u16(a,b)
22 #define i16x8_sub_saturate(a,b) vsubq_s16(a,b)
23 /* Dummy. Aid making uniform macros */
24 #define vreinterpretq_u8_u8(a) a
25 /* 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 })
32 
33 /* Converts all ones/zeros compare mask to bitmap. */
36 {
37  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));
47 }
48 
49 /* *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)
56 
57 #define _(t, s, c, i) \
58 static_always_inline t##s##x##c \
59 t##s##x##c##_splat (t##s x) \
60 { return (t##s##x##c) vdupq_n_##i (x); } \
61 \
62 static_always_inline t##s##x##c \
63 t##s##x##c##_load_unaligned (void *p) \
64 { return (t##s##x##c) vld1q_##i (p); } \
65 \
66 static_always_inline void \
67 t##s##x##c##_store_unaligned (t##s##x##c v, void *p) \
68 { vst1q_##i (p, v); } \
69 \
70 static_always_inline int \
71 t##s##x##c##_is_all_zero (t##s##x##c x) \
72 { return !!(vminvq_u##s (vceqq_##i (vdupq_n_##i(0), x))); } \
73 \
74 static_always_inline int \
75 t##s##x##c##_is_equal (t##s##x##c a, t##s##x##c b) \
76 { return !!(vminvq_u##s (vceqq_##i (a, b))); } \
77 \
78 static_always_inline int \
79 t##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)); }; \
81 \
82 static_always_inline u32 \
83 t##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); } \
86 \
87 static_always_inline u##s##x##c \
88 t##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 \
91 static_always_inline t##s##x##c \
92 t##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); }
94 
96 
97 #undef _
98 /* *INDENT-ON* */
99 
102 {
103  return (u16x8) vrev16q_u8 ((u8x16) v);
104 }
105 
108 {
109  return (u32x4) vrev32q_u8 ((u8x16) v);
110 }
111 
113 u8x16_shuffle (u8x16 v, u8x16 m)
114 {
115  return (u8x16) vqtbl1q_u8 (v, m);
116 }
117 
120 {
121  return (u32x4) vpaddq_u32 (v1, v2);
122 }
123 
126 {
127  return vmovl_u32 (vget_low_u32 (v));
128 }
129 
132 {
133  return vmovl_high_u32 (v);
134 }
135 
136 /* Creates a mask made up of the MSB of each byte of the source vector */
138 u8x16_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 
152 u64x2_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 
160 u32x4_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 
170 u64x2_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 
177 u32x4_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 
187 {
188  return vminvq_u32 (v);
189 }
190 
191 #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)
193 
195 u8x16_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 
204 u8x16_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 
214 #define CLIB_HAVE_VEC128_MSB_MASK
215 
216 #define CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE
217 #define CLIB_VEC128_SPLAT_DEFINED
218 #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  */
#define foreach_neon_vec128i
Definition: vector_neon.h:50
static_always_inline u32x4 u32x4_byte_swap(u32x4 v)
Definition: vector_neon.h:107
a
Definition: bitmap.h:538
static_always_inline u64x2 u64x2_from_u32x4_high(u32x4 v)
Definition: vector_neon.h:131
unsigned long u64
Definition: types.h:89
static_always_inline u8x16 u8x16_reflect(u8x16 v)
Definition: vector_neon.h:195
static_always_inline u32x4 u32x4_hadd(u32x4 v1, u32x4 v2)
Definition: vector_neon.h:119
u16 mask
Definition: flow_types.api:52
#define static_always_inline
Definition: clib.h:108
unsigned int u32
Definition: types.h:88
epu8_epi32 epu16_epi32 u64x2
Definition: vector_sse42.h:691
unsigned short u16
Definition: types.h:57
#define always_inline
Definition: ipsec.h:28
static_always_inline u8x16 u8x16_xor3(u8x16 a, u8x16 b, u8x16 c)
Definition: vector_neon.h:204
static_always_inline u32x4 u32x4_gather(void *p0, void *p1, void *p2, void *p3)
Definition: vector_neon.h:160
static_always_inline u16 u8x16_msb_mask(u8x16 v)
Definition: vector_neon.h:138
static u32 u8x16_compare_byte_mask(u8x16 v)
Definition: vector_neon.h:35
svmdb_client_t * c
static_always_inline void u64x2_scatter(u64x2 r, void *p0, void *p1)
Definition: vector_neon.h:170
static_always_inline u32 u32x4_min_scalar(u32x4 v)
Definition: vector_neon.h:186
#define foreach_neon_vec128u
Definition: vector_neon.h:52
static_always_inline u64x2 u64x2_from_u32x4(u32x4 v)
Definition: vector_neon.h:125
static_always_inline u8x16 u8x16_shuffle(u8x16 v, u8x16 m)
Definition: vector_neon.h:113
static_always_inline void u32x4_scatter(u32x4 r, void *p0, void *p1, void *p2, void *p3)
Definition: vector_neon.h:177
static_always_inline u64x2 u64x2_gather(void *p0, void *p1)
Definition: vector_neon.h:152
unsigned long long u32x4
Definition: ixge.c:28
foreach_neon_vec128i foreach_neon_vec128u static_always_inline u16x8 u16x8_byte_swap(u16x8 v)
Definition: vector_neon.h:101