3-way crc32c

Summary: Current folly version uses a single crc32c.  crc32c has a latency of 3, so pipelining three in a row makes it nearly 3x faster (for data all in cache).

Reviewed By: yfeldblum

Differential Revision: D5418228

fbshipit-source-id: d3a250e1b4fe1f0bc99b44c660df94cf233aebd6

diff --git a/folly/Checksum.cpp b/folly/Checksum.cpp
index e800b0d53c96da1ef4a9901c6ce52af8e74db39d..1b784f3f97f3af365ca456f696e7fc75e84de0f0 100644 (file)
--- a/folly/Checksum.cpp
+++ b/folly/Checksum.cpp
@@ -33,40 +33,6 @@ uint32_t
 crc32c_sw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum);
 #if FOLLY_SSE_PREREQ(4, 2)
 
-// Fast SIMD implementation of CRC-32C for x86 with SSE 4.2
-FOLLY_TARGET_ATTRIBUTE("sse4.2")
-uint32_t crc32c_hw(const uint8_t *data, size_t nbytes,
-    uint32_t startingChecksum) {
-  uint32_t sum = startingChecksum;
-  size_t offset = 0;
-
-  // Process bytes one at a time until we reach an 8-byte boundary and can
-  // start doing aligned 64-bit reads.
-  static uintptr_t ALIGN_MASK = sizeof(uint64_t) - 1;
-  size_t mask = (size_t)((uintptr_t)data & ALIGN_MASK);
-  if (mask != 0) {
-    size_t limit = std::min(nbytes, sizeof(uint64_t) - mask);
-    while (offset < limit) {
-      sum = (uint32_t)_mm_crc32_u8(sum, data[offset]);
-      offset++;
-    }
-  }
-
-  // Process 8 bytes at a time until we have fewer than 8 bytes left.
-  while (offset + sizeof(uint64_t) <= nbytes) {
-    const uint64_t* src = (const uint64_t*)(data + offset);
-    sum = uint32_t(_mm_crc32_u64(sum, *src));
-    offset += sizeof(uint64_t);
-  }
-
-  // Process any bytes remaining after the last aligned 8-byte block.
-  while (offset < nbytes) {
-    sum = (uint32_t)_mm_crc32_u8(sum, data[offset]);
-    offset++;
-  }
-  return sum;
-}
-
 uint32_t
 crc32_sw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum);
 
@@ -106,11 +72,6 @@ bool crc32_hw_supported() {
 
 #else
 
-uint32_t crc32c_hw(const uint8_t *data, size_t nbytes,
-    uint32_t startingChecksum) {
-  throw std::runtime_error("crc32_hw is not implemented on this platform");
-}
-
 uint32_t crc32_hw(const uint8_t *data, size_t nbytes,
     uint32_t startingChecksum) {
   throw std::runtime_error("crc32_hw is not implemented on this platform");

diff --git a/folly/Makefile.am b/folly/Makefile.am
index 20683ce654479f9a9fcd4643fab0563a6f89ea84..a1607f040733b2c6667982e91734d95ab76ac247 100644 (file)
--- a/folly/Makefile.am
+++ b/folly/Makefile.am
@@ -440,6 +440,7 @@ GroupVarintTables.cpp: build/generate_varint_tables.py
 CLEANFILES += GroupVarintTables.cpp
 
 libfollybasesse42_la_SOURCES = \
+       detail/Crc32cDetail.cpp \
        detail/ChecksumDetail.cpp \
        detail/RangeSse42.cpp
 

diff --git a/folly/detail/Crc32cDetail.cpp b/folly/detail/Crc32cDetail.cpp
new file mode 100644 (file)
index 0000000..af8fb2b
--- /dev/null
+++ b/folly/detail/Crc32cDetail.cpp
@@ -0,0 +1,289 @@
+/*
+ * Copyright 2016 Ferry Toth, Exalon Delft BV, The Netherlands
+ *  This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the author be held liable for any damages
+ * arising from the use of this software.
+ *  Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *  1. The origin of this software must not be misrepresented; you must not
+ *   claim that you wrote the original software. If you use this software
+ *   in a product, an acknowledgment in the product documentation would be
+ *   appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *   misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ *  Ferry Toth
+ * ftoth@exalondelft.nl
+ *
+ * https://github.com/htot/crc32c
+ *
+ * Modified by Facebook
+ *
+ * Original intel whitepaper:
+ * "Fast CRC Computation for iSCSI Polynomial Using CRC32 Instruction"
+ * https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/crc-iscsi-polynomial-crc32-instruction-paper.pdf
+ *
+ * 32-bit support dropped
+ * use intrinsics instead of inline asm
+ * other code cleanup
+ */
+
+#include <folly/detail/ChecksumDetail.h>
+
+#include <folly/CppAttributes.h>
+
+#include <boost/preprocessor/arithmetic/add.hpp>
+#include <boost/preprocessor/arithmetic/sub.hpp>
+#include <boost/preprocessor/repetition/repeat_from_to.hpp>
+
+namespace folly {
+namespace crc32_detail {
+
+#if FOLLY_SSE_PREREQ(4, 2)
+
+#define CRCtriplet(crc, buf, offset)                  \
+  crc##0 = _mm_crc32_u64(crc##0, *(buf##0 + offset)); \
+  crc##1 = _mm_crc32_u64(crc##1, *(buf##1 + offset)); \
+  crc##2 = _mm_crc32_u64(crc##2, *(buf##2 + offset)); \
+  FOLLY_FALLTHROUGH;
+
+#define CRCduplet(crc, buf, offset)                   \
+  crc##0 = _mm_crc32_u64(crc##0, *(buf##0 + offset)); \
+  crc##1 = _mm_crc32_u64(crc##1, *(buf##1 + offset));
+
+#define CRCsinglet(crc, buf, offset)                    \
+  crc = _mm_crc32_u64(crc, *(uint64_t*)(buf + offset)); \
+  FOLLY_FALLTHROUGH;
+
+#define CASEREPEAT_TRIPLET(unused, count, total)    \
+  case BOOST_PP_ADD(1, BOOST_PP_SUB(total, count)): \
+    CRCtriplet(crc, next, -BOOST_PP_ADD(1, BOOST_PP_SUB(total, count)));
+
+#define CASEREPEAT_SINGLET(unused, count, total) \
+  case BOOST_PP_SUB(total, count):               \
+    CRCsinglet(crc0, next, -BOOST_PP_SUB(total, count) * 8);
+
+// Numbers taken directly from intel whitepaper.
+const __m128i clmul_constants[] = {
+    {0x14cd00bd6, 0x105ec76f0}, {0x0ba4fc28e, 0x14cd00bd6},
+    {0x1d82c63da, 0x0f20c0dfe}, {0x09e4addf8, 0x0ba4fc28e},
+    {0x039d3b296, 0x1384aa63a}, {0x102f9b8a2, 0x1d82c63da},
+    {0x14237f5e6, 0x01c291d04}, {0x00d3b6092, 0x09e4addf8},
+    {0x0c96cfdc0, 0x0740eef02}, {0x18266e456, 0x039d3b296},
+    {0x0daece73e, 0x0083a6eec}, {0x0ab7aff2a, 0x102f9b8a2},
+    {0x1248ea574, 0x1c1733996}, {0x083348832, 0x14237f5e6},
+    {0x12c743124, 0x02ad91c30}, {0x0b9e02b86, 0x00d3b6092},
+    {0x018b33a4e, 0x06992cea2}, {0x1b331e26a, 0x0c96cfdc0},
+    {0x17d35ba46, 0x07e908048}, {0x1bf2e8b8a, 0x18266e456},
+    {0x1a3e0968a, 0x11ed1f9d8}, {0x0ce7f39f4, 0x0daece73e},
+    {0x061d82e56, 0x0f1d0f55e}, {0x0d270f1a2, 0x0ab7aff2a},
+    {0x1c3f5f66c, 0x0a87ab8a8}, {0x12ed0daac, 0x1248ea574},
+    {0x065863b64, 0x08462d800}, {0x11eef4f8e, 0x083348832},
+    {0x1ee54f54c, 0x071d111a8}, {0x0b3e32c28, 0x12c743124},
+    {0x0064f7f26, 0x0ffd852c6}, {0x0dd7e3b0c, 0x0b9e02b86},
+    {0x0f285651c, 0x0dcb17aa4}, {0x010746f3c, 0x018b33a4e},
+    {0x1c24afea4, 0x0f37c5aee}, {0x0271d9844, 0x1b331e26a},
+    {0x08e766a0c, 0x06051d5a2}, {0x093a5f730, 0x17d35ba46},
+    {0x06cb08e5c, 0x11d5ca20e}, {0x06b749fb2, 0x1bf2e8b8a},
+    {0x1167f94f2, 0x021f3d99c}, {0x0cec3662e, 0x1a3e0968a},
+    {0x19329634a, 0x08f158014}, {0x0e6fc4e6a, 0x0ce7f39f4},
+    {0x08227bb8a, 0x1a5e82106}, {0x0b0cd4768, 0x061d82e56},
+    {0x13c2b89c4, 0x188815ab2}, {0x0d7a4825c, 0x0d270f1a2},
+    {0x10f5ff2ba, 0x105405f3e}, {0x00167d312, 0x1c3f5f66c},
+    {0x0f6076544, 0x0e9adf796}, {0x026f6a60a, 0x12ed0daac},
+    {0x1a2adb74e, 0x096638b34}, {0x19d34af3a, 0x065863b64},
+    {0x049c3cc9c, 0x1e50585a0}, {0x068bce87a, 0x11eef4f8e},
+    {0x1524fa6c6, 0x19f1c69dc}, {0x16cba8aca, 0x1ee54f54c},
+    {0x042d98888, 0x12913343e}, {0x1329d9f7e, 0x0b3e32c28},
+    {0x1b1c69528, 0x088f25a3a}, {0x02178513a, 0x0064f7f26},
+    {0x0e0ac139e, 0x04e36f0b0}, {0x0170076fa, 0x0dd7e3b0c},
+    {0x141a1a2e2, 0x0bd6f81f8}, {0x16ad828b4, 0x0f285651c},
+    {0x041d17b64, 0x19425cbba}, {0x1fae1cc66, 0x010746f3c},
+    {0x1a75b4b00, 0x18db37e8a}, {0x0f872e54c, 0x1c24afea4},
+    {0x01e41e9fc, 0x04c144932}, {0x086d8e4d2, 0x0271d9844},
+    {0x160f7af7a, 0x052148f02}, {0x05bb8f1bc, 0x08e766a0c},
+    {0x0a90fd27a, 0x0a3c6f37a}, {0x0b3af077a, 0x093a5f730},
+    {0x04984d782, 0x1d22c238e}, {0x0ca6ef3ac, 0x06cb08e5c},
+    {0x0234e0b26, 0x063ded06a}, {0x1d88abd4a, 0x06b749fb2},
+    {0x04597456a, 0x04d56973c}, {0x0e9e28eb4, 0x1167f94f2},
+    {0x07b3ff57a, 0x19385bf2e}, {0x0c9c8b782, 0x0cec3662e},
+    {0x13a9cba9e, 0x0e417f38a}, {0x093e106a4, 0x19329634a},
+    {0x167001a9c, 0x14e727980}, {0x1ddffc5d4, 0x0e6fc4e6a},
+    {0x00df04680, 0x0d104b8fc}, {0x02342001e, 0x08227bb8a},
+    {0x00a2a8d7e, 0x05b397730}, {0x168763fa6, 0x0b0cd4768},
+    {0x1ed5a407a, 0x0e78eb416}, {0x0d2c3ed1a, 0x13c2b89c4},
+    {0x0995a5724, 0x1641378f0}, {0x19b1afbc4, 0x0d7a4825c},
+    {0x109ffedc0, 0x08d96551c}, {0x0f2271e60, 0x10f5ff2ba},
+    {0x00b0bf8ca, 0x00bf80dd2}, {0x123888b7a, 0x00167d312},
+    {0x1e888f7dc, 0x18dcddd1c}, {0x002ee03b2, 0x0f6076544},
+    {0x183e8d8fe, 0x06a45d2b2}, {0x133d7a042, 0x026f6a60a},
+    {0x116b0f50c, 0x1dd3e10e8}, {0x05fabe670, 0x1a2adb74e},
+    {0x130004488, 0x0de87806c}, {0x000bcf5f6, 0x19d34af3a},
+    {0x18f0c7078, 0x014338754}, {0x017f27698, 0x049c3cc9c},
+    {0x058ca5f00, 0x15e3e77ee}, {0x1af900c24, 0x068bce87a},
+    {0x0b5cfca28, 0x0dd07448e}, {0x0ded288f8, 0x1524fa6c6},
+    {0x059f229bc, 0x1d8048348}, {0x06d390dec, 0x16cba8aca},
+    {0x037170390, 0x0a3e3e02c}, {0x06353c1cc, 0x042d98888},
+    {0x0c4584f5c, 0x0d73c7bea}, {0x1f16a3418, 0x1329d9f7e},
+    {0x0531377e2, 0x185137662}, {0x1d8d9ca7c, 0x1b1c69528},
+    {0x0b25b29f2, 0x18a08b5bc}, {0x19fb2a8b0, 0x02178513a},
+    {0x1a08fe6ac, 0x1da758ae0}, {0x045cddf4e, 0x0e0ac139e},
+    {0x1a91647f2, 0x169cf9eb0}, {0x1a0f717c4, 0x0170076fa},
+};
+
+/*
+ * CombineCRC performs pclmulqdq multiplication of 2 partial CRC's and a well
+ * chosen constant and xor's these with the remaining CRC.
+ */
+uint64_t CombineCRC(
+    unsigned long block_size,
+    uint64_t crc0,
+    uint64_t crc1,
+    uint64_t crc2,
+    const uint64_t* next2) {
+  const auto multiplier = *(clmul_constants + block_size - 1);
+  const auto crc0_xmm = _mm_set_epi64x(0, crc0);
+  const auto res0 = _mm_clmulepi64_si128(crc0_xmm, multiplier, 0x00);
+  const auto crc1_xmm = _mm_set_epi64x(0, crc1);
+  const auto res1 = _mm_clmulepi64_si128(crc1_xmm, multiplier, 0x10);
+  const auto res = _mm_xor_si128(res0, res1);
+  crc0 = _mm_cvtsi128_si64(res);
+  crc0 = crc0 ^ *((uint64_t*)next2 - 1);
+  crc2 = _mm_crc32_u64(crc2, crc0);
+  return crc2;
+}
+
+// Generates a block that will crc up to 7 bytes of unaligned data.
+// Always inline to avoid overhead on small crc sizes.
+FOLLY_ALWAYS_INLINE void align_to_8(
+    unsigned long align,
+    uint64_t& crc0, // crc so far, updated on return
+    const unsigned char*& next) { // next data pointer, updated on return
+  uint32_t crc32bit = crc0;
+  if (align & 0x04) {
+    crc32bit = _mm_crc32_u32(crc32bit, *(uint32_t*)next);
+    next += sizeof(uint32_t);
+  }
+  if (align & 0x02) {
+    crc32bit = _mm_crc32_u16(crc32bit, *(uint16_t*)next);
+    next += sizeof(uint16_t);
+  }
+  if (align & 0x01) {
+    crc32bit = _mm_crc32_u8(crc32bit, *(next));
+    next++;
+  }
+  crc0 = crc32bit;
+}
+
+// The main loop for large crc sizes. Generates three crc32c
+// streams, of varying block sizes, using a duff's device.
+void triplet_loop(
+    unsigned long block_size,
+    uint64_t& crc0, // crc so far, updated on return
+    const unsigned char*& next, // next data pointer, updated on return
+    unsigned long n) { // block count
+  uint64_t crc1 = 0, crc2 = 0;
+  // points to the first byte of the next block
+  const uint64_t* next0 = (uint64_t*)next + block_size;
+  const uint64_t* next1 = next0 + block_size;
+  const uint64_t* next2 = next1 + block_size;
+
+  // Use Duff's device, a for() loop inside a switch()
+  // statement. This needs to execute at least once, round len
+  // down to nearest triplet multiple
+  switch (block_size) {
+    case 128:
+      do {
+        // jumps here for a full block of len 128
+        CRCtriplet(crc, next, -128);
+
+        // Generates case statements from 127 to 2 of form:
+        // case 127:
+        //    CRCtriplet(crc, next, -127);
+        BOOST_PP_REPEAT_FROM_TO(0, 126, CASEREPEAT_TRIPLET, 126);
+
+        // For the last byte, the three crc32c streams must be combined
+        // using carry-less multiplication.
+        case 1:
+          CRCduplet(crc, next, -1); // the final triplet is actually only 2
+          crc0 = CombineCRC(block_size, crc0, crc1, crc2, next2);
+          if (--n > 0) {
+            crc1 = crc2 = 0;
+            block_size = 128;
+            // points to the first byte of the next block
+            next0 = next2 + 128;
+            next1 = next0 + 128; // from here on all blocks are 128 long
+            next2 = next1 + 128;
+          }
+          FOLLY_FALLTHROUGH;
+        case 0:;
+      } while (n > 0);
+  }
+
+  next = (const unsigned char*)next2;
+}
+
+} // namespace crc32c_detail
+
+namespace detail {
+
+/* Compute CRC-32C using the Intel hardware instruction. */
+FOLLY_TARGET_ATTRIBUTE("sse4.2")
+uint32_t crc32c_hw(const uint8_t* buf, size_t len, uint32_t crc) {
+  const unsigned char* next = (const unsigned char*)buf;
+  unsigned long count;
+  uint64_t crc0;
+  crc0 = crc;
+
+  if (len >= 8) {
+    // if len > 216 then align and use triplets
+    if (len > 216) {
+      unsigned long align = (8 - (uintptr_t)next) & 7;
+      crc32_detail::align_to_8(align, crc0, next);
+      len -= align;
+
+      count = len / 24; // number of triplets
+      len %= 24; // bytes remaining
+      unsigned long n = count >> 7; // #blocks = first block + full blocks
+      unsigned long block_size = count & 127;
+      if (block_size == 0) {
+        block_size = 128;
+      } else {
+        n++;
+      }
+
+      // This is a separate function call mainly to stop
+      // clang from spilling registers.
+      crc32_detail::triplet_loop(block_size, crc0, next, n);
+    }
+
+    unsigned count2 = len >> 3;
+    len = len & 7;
+    next += (count2 * 8);
+
+    // Generates a duff device for the last 128 bytes of aligned data.
+    switch (count2) {
+      // Generates case statements of the form:
+      // case 27:
+      //   CRCsinglet(crc0, next, -27 * 8);
+      BOOST_PP_REPEAT_FROM_TO(0, 27, CASEREPEAT_SINGLET, 27);
+      case 0:;
+    }
+  }
+
+  // compute the crc for up to seven trailing bytes
+  crc32_detail::align_to_8(len, crc0, next);
+  return (uint32_t)crc0;
+}
+
+#else
+
+uint32_t
+crc32c_hw(const uint8_t* data, size_t nbytes, uint32_t startingChecksum) {
+  throw std::runtime_error("crc32_hw is not implemented on this platform");
+}
+
+#endif
+}
+} // namespace

diff --git a/folly/test/ChecksumTest.cpp b/folly/test/ChecksumTest.cpp
index b54e528e1e898f53020279766d5a1adf699af80b..16489dce970410fa9d54e974945f00671dede894 100644 (file)
--- a/folly/test/ChecksumTest.cpp
+++ b/folly/test/ChecksumTest.cpp
@@ -121,6 +121,19 @@ TEST(Checksum, crc32c_hardware) {
   }
 }
 
+TEST(Checksum, crc32c_hardware_eq) {
+  if (folly::detail::crc32c_hw_supported()) {
+    for (int i = 0; i < 1000; i++) {
+      auto sw = folly::detail::crc32c_sw(buffer, i, 0);
+      auto hw = folly::detail::crc32c_hw(buffer, i, 0);
+      EXPECT_EQ(sw, hw);
+    }
+  } else {
+    LOG(WARNING) << "skipping hardware-accelerated CRC-32C tests"
+                 << " (not supported on this CPU)";
+  }
+}
+
 TEST(Checksum, crc32c_continuation_hardware) {
   if (folly::detail::crc32c_hw_supported()) {
     testCRC32CContinuation(folly::detail::crc32c_hw);