2 * Copyright 2017 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #include <folly/Checksum.h>
18 #include <boost/crc.hpp>
19 #include <folly/CpuId.h>
20 #include <folly/detail/ChecksumDetail.h>
24 #if FOLLY_X64 && (__SSE4_2__ || defined(__clang__) || __GNUC_PREREQ(4, 9))
25 #include <nmmintrin.h>
33 crc32c_sw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum);
34 #if FOLLY_X64 && (__SSE4_2__ || defined(__clang__) || __GNUC_PREREQ(4, 9))
36 // Fast SIMD implementation of CRC-32C for x86 with SSE 4.2
37 FOLLY_TARGET_ATTRIBUTE("sse4.2")
38 uint32_t crc32c_hw(const uint8_t *data, size_t nbytes,
39 uint32_t startingChecksum) {
40 uint32_t sum = startingChecksum;
43 // Process bytes one at a time until we reach an 8-byte boundary and can
44 // start doing aligned 64-bit reads.
45 static uintptr_t ALIGN_MASK = sizeof(uint64_t) - 1;
46 size_t mask = (size_t)((uintptr_t)data & ALIGN_MASK);
48 size_t limit = std::min(nbytes, sizeof(uint64_t) - mask);
49 while (offset < limit) {
50 sum = (uint32_t)_mm_crc32_u8(sum, data[offset]);
55 // Process 8 bytes at a time until we have fewer than 8 bytes left.
56 while (offset + sizeof(uint64_t) <= nbytes) {
57 const uint64_t* src = (const uint64_t*)(data + offset);
58 sum = uint32_t(_mm_crc32_u64(sum, *src));
59 offset += sizeof(uint64_t);
62 // Process any bytes remaining after the last aligned 8-byte block.
63 while (offset < nbytes) {
64 sum = (uint32_t)_mm_crc32_u8(sum, data[offset]);
71 crc32_sw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum);
73 // Fast SIMD implementation of CRC-32 for x86 with pclmul
75 crc32_hw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum) {
76 uint32_t sum = startingChecksum;
79 // Process unaligned bytes
80 if ((uintptr_t)data & 15) {
81 size_t limit = std::min(nbytes, -(uintptr_t)data & 15);
82 sum = crc32_sw(data, limit, sum);
88 sum = crc32_hw_aligned(sum, (const __m128i*)(data + offset), nbytes / 16);
89 offset += nbytes & ~15;
93 // Remaining unaligned bytes
94 return crc32_sw(data + offset, nbytes, sum);
97 bool crc32c_hw_supported() {
98 static folly::CpuId id;
102 bool crc32_hw_supported() {
103 static folly::CpuId id;
109 uint32_t crc32c_hw(const uint8_t *data, size_t nbytes,
110 uint32_t startingChecksum) {
111 throw std::runtime_error("crc32_hw is not implemented on this platform");
114 bool crc32c_hw_supported() {
118 bool crc32_hw_supported() {
123 template <uint32_t CRC_POLYNOMIAL>
124 uint32_t crc_sw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum) {
125 // Reverse the bits in the starting checksum so they'll be in the
126 // right internal format for Boost's CRC engine.
127 // O(1)-time, branchless bit reversal algorithm from
128 // http://graphics.stanford.edu/~seander/bithacks.html
129 startingChecksum = ((startingChecksum >> 1) & 0x55555555) |
130 ((startingChecksum & 0x55555555) << 1);
131 startingChecksum = ((startingChecksum >> 2) & 0x33333333) |
132 ((startingChecksum & 0x33333333) << 2);
133 startingChecksum = ((startingChecksum >> 4) & 0x0f0f0f0f) |
134 ((startingChecksum & 0x0f0f0f0f) << 4);
135 startingChecksum = ((startingChecksum >> 8) & 0x00ff00ff) |
136 ((startingChecksum & 0x00ff00ff) << 8);
137 startingChecksum = (startingChecksum >> 16) |
138 (startingChecksum << 16);
140 boost::crc_optimal<32, CRC_POLYNOMIAL, ~0U, 0, true, true> sum(
142 sum.process_bytes(data, nbytes);
143 return sum.checksum();
147 crc32c_sw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum) {
148 constexpr uint32_t CRC32C_POLYNOMIAL = 0x1EDC6F41;
149 return crc_sw<CRC32C_POLYNOMIAL>(data, nbytes, startingChecksum);
153 crc32_sw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum) {
154 constexpr uint32_t CRC32_POLYNOMIAL = 0x04C11DB7;
155 return crc_sw<CRC32_POLYNOMIAL>(data, nbytes, startingChecksum);
160 uint32_t crc32c(const uint8_t *data, size_t nbytes,
161 uint32_t startingChecksum) {
162 if (detail::crc32c_hw_supported()) {
163 return detail::crc32c_hw(data, nbytes, startingChecksum);
165 return detail::crc32c_sw(data, nbytes, startingChecksum);
169 uint32_t crc32(const uint8_t* data, size_t nbytes, uint32_t startingChecksum) {
170 if (detail::crc32_hw_supported()) {
171 return detail::crc32_hw(data, nbytes, startingChecksum);
173 return detail::crc32_sw(data, nbytes, startingChecksum);