xref: /freebsd/contrib/arm-optimized-routines/networking/aarch64/chksum_simd.c (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
1 /*
2  * AArch64-specific checksum implementation using NEON
3  *
4  * Copyright (c) 2020, Arm Limited.
5  * SPDX-License-Identifier: MIT
6  */
7 
8 #include "networking.h"
9 #include "../chksum_common.h"
10 
11 #ifndef __ARM_NEON
12 #pragma GCC target("+simd")
13 #endif
14 
15 #include <arm_neon.h>
16 
17 always_inline
18 static inline uint64_t
19 slurp_head64(const void **pptr, uint32_t *nbytes)
20 {
21     Assert(*nbytes >= 8);
22     uint64_t sum = 0;
23     uint32_t off = (uintptr_t) *pptr % 8;
24     if (likely(off != 0))
25     {
26 	/* Get rid of bytes 0..off-1 */
27 	const unsigned char *ptr64 = align_ptr(*pptr, 8);
28 	uint64_t mask = ALL_ONES << (CHAR_BIT * off);
29 	uint64_t val = load64(ptr64) & mask;
30 	/* Fold 64-bit sum to 33 bits */
31 	sum = val >> 32;
32 	sum += (uint32_t) val;
33 	*pptr = ptr64 + 8;
34 	*nbytes -= 8 - off;
35     }
36     return sum;
37 }
38 
39 always_inline
40 static inline uint64_t
41 slurp_tail64(uint64_t sum, const void *ptr, uint32_t nbytes)
42 {
43     Assert(nbytes < 8);
44     if (likely(nbytes != 0))
45     {
46 	/* Get rid of bytes 7..nbytes */
47 	uint64_t mask = ALL_ONES >> (CHAR_BIT * (8 - nbytes));
48 	Assert(__builtin_popcountl(mask) / CHAR_BIT == nbytes);
49 	uint64_t val = load64(ptr) & mask;
50 	sum += val >> 32;
51 	sum += (uint32_t) val;
52 	nbytes = 0;
53     }
54     Assert(nbytes == 0);
55     return sum;
56 }
57 
58 unsigned short
59 __chksum_aarch64_simd(const void *ptr, unsigned int nbytes)
60 {
61     bool swap = (uintptr_t) ptr & 1;
62     uint64_t sum;
63 
64     if (unlikely(nbytes < 50))
65     {
66 	sum = slurp_small(ptr, nbytes);
67 	swap = false;
68 	goto fold;
69     }
70 
71     /* 8-byte align pointer */
72     Assert(nbytes >= 8);
73     sum = slurp_head64(&ptr, &nbytes);
74     Assert(((uintptr_t) ptr & 7) == 0);
75 
76     const uint32_t *may_alias ptr32 = ptr;
77 
78     uint64x2_t vsum0 = { 0, 0 };
79     uint64x2_t vsum1 = { 0, 0 };
80     uint64x2_t vsum2 = { 0, 0 };
81     uint64x2_t vsum3 = { 0, 0 };
82 
83     /* Sum groups of 64 bytes */
84     for (uint32_t i = 0; i < nbytes / 64; i++)
85     {
86 	uint32x4_t vtmp0 = vld1q_u32(ptr32);
87 	uint32x4_t vtmp1 = vld1q_u32(ptr32 + 4);
88 	uint32x4_t vtmp2 = vld1q_u32(ptr32 + 8);
89 	uint32x4_t vtmp3 = vld1q_u32(ptr32 + 12);
90 	vsum0 = vpadalq_u32(vsum0, vtmp0);
91 	vsum1 = vpadalq_u32(vsum1, vtmp1);
92 	vsum2 = vpadalq_u32(vsum2, vtmp2);
93 	vsum3 = vpadalq_u32(vsum3, vtmp3);
94 	ptr32 += 16;
95     }
96     nbytes %= 64;
97 
98     /* Fold vsum2 and vsum3 into vsum0 and vsum1 */
99     vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum2));
100     vsum1 = vpadalq_u32(vsum1, vreinterpretq_u32_u64(vsum3));
101 
102     /* Add any trailing group of 32 bytes */
103     if (nbytes & 32)
104     {
105 	uint32x4_t vtmp0 = vld1q_u32(ptr32);
106 	uint32x4_t vtmp1 = vld1q_u32(ptr32 + 4);
107 	vsum0 = vpadalq_u32(vsum0, vtmp0);
108 	vsum1 = vpadalq_u32(vsum1, vtmp1);
109 	ptr32 += 8;
110 	nbytes -= 32;
111     }
112     Assert(nbytes < 32);
113 
114     /* Fold vsum1 into vsum0 */
115     vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum1));
116 
117     /* Add any trailing group of 16 bytes */
118     if (nbytes & 16)
119     {
120 	uint32x4_t vtmp = vld1q_u32(ptr32);
121 	vsum0 = vpadalq_u32(vsum0, vtmp);
122 	ptr32 += 4;
123 	nbytes -= 16;
124     }
125     Assert(nbytes < 16);
126 
127     /* Add any trailing group of 8 bytes */
128     if (nbytes & 8)
129     {
130 	uint32x2_t vtmp = vld1_u32(ptr32);
131 	vsum0 = vaddw_u32(vsum0, vtmp);
132 	ptr32 += 2;
133 	nbytes -= 8;
134     }
135     Assert(nbytes < 8);
136 
137     uint64_t val = vaddlvq_u32(vreinterpretq_u32_u64(vsum0));
138     sum += val >> 32;
139     sum += (uint32_t) val;
140 
141     /* Handle any trailing 0..7 bytes */
142     sum = slurp_tail64(sum, ptr32, nbytes);
143 
144 fold:
145     return fold_and_swap(sum, swap);
146 }
147