FD.io VPP  v21.10.1-2-g0a485f517
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 /* Dummy. Aid making uniform macros */
21 #define vreinterpretq_u8_u8(a) a
22 /* Implement the missing intrinsics to make uniform macros */
23 #define vminvq_u64(x) \
24 ({ \
25  u64 x0 = vgetq_lane_u64(x, 0); \
26  u64 x1 = vgetq_lane_u64(x, 1); \
27  x0 < x1 ? x0 : x1; \
28 })
29 
30 /* Converts all ones/zeros compare mask to bitmap. */
33 {
34  uint8x16_t mask = { 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
35  0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80
36  };
37  /* v --> [0xFF, 0x00, 0xFF, 0xFF, 0xFF, 0x00, 0xFF, 0x00, ... ] */
38  uint8x16_t x = vandq_u8 (v, mask);
39  /* after v & mask,
40  * x --> [0x01, 0x00, 0x04, 0x08, 0x10, 0x00, 0x40, 0x00, ... ] */
41  uint64x2_t x64 = vpaddlq_u32 (vpaddlq_u16 (vpaddlq_u8 (x)));
42  /* after merge, x64 --> [0x5D, 0x.. ] */
43  return (u32) (vgetq_lane_u64 (x64, 0) + (vgetq_lane_u64 (x64, 1) << 8));
44 }
45 
46 /* *INDENT-OFF* */
47 #define foreach_neon_vec128i \
48  _(i,8,16,s8) _(i,16,8,s16) _(i,32,4,s32) _(i,64,2,s64)
49 #define foreach_neon_vec128u \
50  _(u,8,16,u8) _(u,16,8,u16) _(u,32,4,u32) _(u,64,2,u64)
51 #define foreach_neon_vec128f \
52  _(f,32,4,f32) _(f,64,2,f64)
53 
54 #define _(t, s, c, i) \
55  static_always_inline t##s##x##c t##s##x##c##_splat (t##s x) \
56  { \
57  return (t##s##x##c) vdupq_n_##i (x); \
58  } \
59  \
60  static_always_inline t##s##x##c t##s##x##c##_load_unaligned (void *p) \
61  { \
62  return (t##s##x##c) vld1q_##i (p); \
63  } \
64  \
65  static_always_inline void t##s##x##c##_store_unaligned (t##s##x##c v, \
66  void *p) \
67  { \
68  vst1q_##i (p, v); \
69  } \
70  \
71  static_always_inline int t##s##x##c##_is_all_zero (t##s##x##c x) \
72  { \
73  return !!(vminvq_u##s (vceqq_##i (vdupq_n_##i (0), x))); \
74  } \
75  \
76  static_always_inline int t##s##x##c##_is_equal (t##s##x##c a, t##s##x##c b) \
77  { \
78  return !!(vminvq_u##s (vceqq_##i (a, b))); \
79  } \
80  static_always_inline int t##s##x##c##_is_all_equal (t##s##x##c v, t##s x) \
81  { \
82  return t##s##x##c##_is_equal (v, t##s##x##c##_splat (x)); \
83  }; \
84  \
85  static_always_inline u32 t##s##x##c##_zero_byte_mask (t##s##x##c x) \
86  { \
87  uint8x16_t v = vreinterpretq_u8_u##s (vceqq_##i (vdupq_n_##i (0), x)); \
88  return u8x16_compare_byte_mask (v); \
89  } \
90  \
91  static_always_inline u##s##x##c t##s##x##c##_is_greater (t##s##x##c a, \
92  t##s##x##c b) \
93  { \
94  return (u##s##x##c) vcgtq_##i (a, b); \
95  } \
96  \
97  static_always_inline t##s##x##c t##s##x##c##_add_saturate (t##s##x##c a, \
98  t##s##x##c b) \
99  { \
100  return (t##s##x##c) vqaddq_##i (a, b); \
101  } \
102  \
103  static_always_inline t##s##x##c t##s##x##c##_sub_saturate (t##s##x##c a, \
104  t##s##x##c b) \
105  { \
106  return (t##s##x##c) vqsubq_##i (a, b); \
107  } \
108  \
109  static_always_inline t##s##x##c t##s##x##c##_blend ( \
110  t##s##x##c dst, t##s##x##c src, u##s##x##c mask) \
111  { \
112  return (t##s##x##c) vbslq_##i (mask, src, dst); \
113  }
114 
116 
117 #undef _
118 /* *INDENT-ON* */
119 
122 {
123  return (u16x8) vrev16q_u8 ((u8x16) v);
124 }
125 
128 {
129  return (u32x4) vrev32q_u8 ((u8x16) v);
130 }
131 
134 {
135  return (u8x16) vqtbl1q_u8 (v, m);
136 }
137 
140 {
141  return (u32x4) vpaddq_u32 (v1, v2);
142 }
143 
146 {
147  return vmovl_u32 (vget_low_u32 (v));
148 }
149 
152 {
153  return vmovl_high_u32 (v);
154 }
155 
156 /* Creates a mask made up of the MSB of each byte of the source vector */
159 {
160  int8x16_t shift =
161  { -7, -6, -5, -4, -3, -2, -1, 0, -7, -6, -5, -4, -3, -2, -1, 0 };
162  /* v --> [0x80, 0x7F, 0xF0, 0xAF, 0xF0, 0x00, 0xF2, 0x00, ... ] */
163  uint8x16_t x = vshlq_u8 (vandq_u8 (v, vdupq_n_u8 (0x80)), shift);
164  /* after (v & 0x80) >> shift,
165  * x --> [0x01, 0x00, 0x04, 0x08, 0x10, 0x00, 0x40, 0x00, ... ] */
166  uint64x2_t x64 = vpaddlq_u32 (vpaddlq_u16 (vpaddlq_u8 (x)));
167  /* after merge, x64 --> [0x5D, 0x.. ] */
168  return (u16) (vgetq_lane_u64 (x64, 0) + (vgetq_lane_u64 (x64, 1) << 8));
169 }
170 
172 u64x2_gather (void *p0, void *p1)
173 {
174  u64x2 r = vdupq_n_u64 (*(u64 *) p0);
175  r = vsetq_lane_u64 (*(u64 *) p1, r, 1);
176  return r;
177 }
178 
180 u32x4_gather (void *p0, void *p1, void *p2, void *p3)
181 {
182  u32x4 r = vdupq_n_u32 (*(u32 *) p0);
183  r = vsetq_lane_u32 (*(u32 *) p1, r, 1);
184  r = vsetq_lane_u32 (*(u32 *) p2, r, 2);
185  r = vsetq_lane_u32 (*(u32 *) p3, r, 3);
186  return r;
187 }
188 
190 u64x2_scatter (u64x2 r, void *p0, void *p1)
191 {
192  *(u64 *) p0 = vgetq_lane_u64 (r, 0);
193  *(u64 *) p1 = vgetq_lane_u64 (r, 1);
194 }
195 
197 u32x4_scatter (u32x4 r, void *p0, void *p1, void *p2, void *p3)
198 {
199  *(u32 *) p0 = vgetq_lane_u32 (r, 0);
200  *(u32 *) p1 = vgetq_lane_u32 (r, 1);
201  *(u32 *) p2 = vgetq_lane_u32 (r, 2);
202  *(u32 *) p3 = vgetq_lane_u32 (r, 3);
203 }
204 
207 {
208  return vminvq_u32 (v);
209 }
210 
211 #define u8x16_word_shift_left(x,n) vextq_u8(u8x16_splat (0), x, 16 - n)
212 #define u8x16_word_shift_right(x,n) vextq_u8(x, u8x16_splat (0), n)
213 
216 {
217  u8x16 mask = {
218  15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
219  };
220  return (u8x16) vqtbl1q_u8 (v, mask);
221 }
222 
225 {
226 #if __GNUC__ == 8 && __ARM_FEATURE_SHA3 == 1
227  u8x16 r;
228 __asm__ ("eor3 %0.16b,%1.16b,%2.16b,%3.16b": "=w" (r): "0" (a), "w" (b), "w" (c):);
229  return r;
230 #endif
231  return a ^ b ^ c;
232 }
233 
234 #define CLIB_HAVE_VEC128_MSB_MASK
235 
236 #define CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE
237 #define CLIB_VEC128_SPLAT_DEFINED
238 #endif /* included_vector_neon_h */
239 
240 /*
241  * fd.io coding-style-patch-verification: ON
242  *
243  * Local Variables:
244  * eval: (c-set-style "gnu")
245  * End:
246  */
u32x4_scatter
static_always_inline void u32x4_scatter(u32x4 r, void *p0, void *p1, void *p2, void *p3)
Definition: vector_neon.h:197
u64x2_gather
static_always_inline u64x2 u64x2_gather(void *p0, void *p1)
Definition: vector_neon.h:172
u8x16_msb_mask
static_always_inline u16 u8x16_msb_mask(u8x16 v)
Definition: vector_neon.h:158
u16
unsigned short u16
Definition: types.h:57
u64x2
epu8_epi32 epu16_epi32 u64x2
Definition: vector_sse42.h:444
u8x16_xor3
static_always_inline u8x16 u8x16_xor3(u8x16 a, u8x16 b, u8x16 c)
Definition: vector_neon.h:224
r
vnet_hw_if_output_node_runtime_t * r
Definition: interface_output.c:1089
foreach_neon_vec128u
#define foreach_neon_vec128u
Definition: vector_neon.h:49
u32x4_byte_swap
static_always_inline u32x4 u32x4_byte_swap(u32x4 v)
Definition: vector_neon.h:127
u8x16
u8x16
Definition: vector_sse42.h:157
u64x2_from_u32x4
static_always_inline u64x2 u64x2_from_u32x4(u32x4 v)
Definition: vector_neon.h:145
foreach_neon_vec128i
#define foreach_neon_vec128i
Definition: vector_neon.h:47
c
svmdb_client_t * c
Definition: vpp_get_metrics.c:48
static_always_inline
#define static_always_inline
Definition: clib.h:112
u32x4_hadd
static_always_inline u32x4 u32x4_hadd(u32x4 v1, u32x4 v2)
Definition: vector_neon.h:139
u8x16_reflect
static_always_inline u8x16 u8x16_reflect(u8x16 v)
Definition: vector_neon.h:215
mask
vl_api_pnat_mask_t mask
Definition: pnat.api:45
u8x16_shuffle
static_always_inline u8x16 u8x16_shuffle(u8x16 v, u8x16 m)
Definition: vector_neon.h:133
u16x8_byte_swap
foreach_neon_vec128i foreach_neon_vec128u static_always_inline u16x8 u16x8_byte_swap(u16x8 v)
Definition: vector_neon.h:121
always_inline
#define always_inline
Definition: rdma_mlx5dv.h:23
u16x8
_mm_packus_epi16 u16x8
Definition: vector_sse42.h:159
u64
unsigned long u64
Definition: types.h:89
u32
unsigned int u32
Definition: types.h:88
u32x4
unsigned long long u32x4
Definition: ixge.c:28
u8x16_compare_byte_mask
static u32 u8x16_compare_byte_mask(u8x16 v)
Definition: vector_neon.h:32
b
vlib_buffer_t ** b
Definition: nat44_ei_out2in.c:717
a
a
Definition: bitmap.h:525
u64x2_scatter
static_always_inline void u64x2_scatter(u64x2 r, void *p0, void *p1)
Definition: vector_neon.h:190
u32x4_gather
static_always_inline u32x4 u32x4_gather(void *p0, void *p1, void *p2, void *p3)
Definition: vector_neon.h:180
u64x2_from_u32x4_high
static_always_inline u64x2 u64x2_from_u32x4_high(u32x4 v)
Definition: vector_neon.h:151
u32x4_min_scalar
static_always_inline u32 u32x4_min_scalar(u32x4 v)
Definition: vector_neon.h:206