HADOOP-7446. Implement CRC32C native code using SSE4.2 instructions. Contributed by Kihwal Lee and Todd Lipcon.

git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/branches/branch-0.23@1190649 13f79535-47bb-0310-9956-ffa450edef68

hadoop-common-project/hadoop-common/CHANGES.txt

@@ -444,6 +444,9 @@ Release 0.23.0 - Unreleased
     HADOOP-7753. Support fadvise and sync_file_range in NativeIO. Add
     ReadaheadPool infrastructure for use in HDFS and MR. (todd)
 
+    HADOOP-7446. Implement CRC32C native code using SSE4.2 instructions.
+    (Kihwal Lee and todd via todd)
+
   BUG FIXES
 
     HADOOP-7740. Fixed security audit logger configuration. (Arpit Gupta via Eric Yang)

hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/util/NativeCrc32.c

@@ -124,6 +124,11 @@ JNIEXPORT void JNICALL Java_org_apache_hadoop_util_NativeCrc32_nativeVerifyChunk
       "bad offsets or lengths");
     return;
   }
+  if (unlikely(bytes_per_checksum) <= 0) {
+    THROW(env, "java/lang/IllegalArgumentException",
+      "invalid bytes_per_checksum");
+    return;
+  }
 
   uint32_t *sums = (uint32_t *)(sums_addr + sums_offset);
   uint8_t *data = data_addr + data_offset;

hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/util/bulk_crc32.c

@@ -21,6 +21,7 @@
  *   All rights reserved. Use of this source code is governed by a
  *   BSD-style license that can be found in the LICENSE file.
  */
+#include <assert.h>
 #include <arpa/inet.h>
 #include <stdint.h>
 #include <unistd.h>
@@ -30,47 +31,124 @@
 #include "bulk_crc32.h"
 #include "gcc_optimizations.h"
 
+#define USE_PIPELINED
+
 typedef uint32_t (*crc_update_func_t)(uint32_t, const uint8_t *, size_t);
 static uint32_t crc_init();
 static uint32_t crc_val(uint32_t crc);
 static uint32_t crc32_zlib_sb8(uint32_t crc, const uint8_t *buf, size_t length);
 static uint32_t crc32c_sb8(uint32_t crc, const uint8_t *buf, size_t length);
 
+#ifdef USE_PIPELINED
+static void pipelined_crc32c(uint32_t *crc1, uint32_t *crc2, uint32_t *crc3, const uint8_t *p_buf, size_t block_size, int num_blocks);
+#endif USE_PIPELINED
+static int cached_cpu_supports_crc32; // initialized by constructor below
+static uint32_t crc32c_hardware(uint32_t crc, const uint8_t* data, size_t length);
+
 int bulk_verify_crc(const uint8_t *data, size_t data_len,
                     const uint32_t *sums, int checksum_type,
                     int bytes_per_checksum,
                     crc32_error_t *error_info) {
 
+#ifdef USE_PIPELINED
+  uint32_t crc1, crc2, crc3;
+  int n_blocks = data_len / bytes_per_checksum;
+  int remainder = data_len % bytes_per_checksum;
+  int do_pipelined = 0;
+#endif
+  uint32_t crc;
   crc_update_func_t crc_update_func;
   switch (checksum_type) {
     case CRC32_ZLIB_POLYNOMIAL:
       crc_update_func = crc32_zlib_sb8;
       break;
     case CRC32C_POLYNOMIAL:
-      crc_update_func = crc32c_sb8;
+      if (likely(cached_cpu_supports_crc32)) {
+        crc_update_func = crc32c_hardware;
+#ifdef USE_PIPELINED
+        do_pipelined = 1;
+#endif
+      } else {
+        crc_update_func = crc32c_sb8;
+      }
       break;
     default:
       return INVALID_CHECKSUM_TYPE;
   }
 
+#ifdef USE_PIPELINED
+  if (do_pipelined) {
+    /* Process three blocks at a time */
+    while (likely(n_blocks >= 3)) {
+      crc1 = crc2 = crc3 = crc_init();  
+      pipelined_crc32c(&crc1, &crc2, &crc3, data, bytes_per_checksum, 3);
+
+      crc = ntohl(crc_val(crc1));
+      if ((crc = ntohl(crc_val(crc1))) != *sums)
+        goto return_crc_error;
+      sums++;
+      data += bytes_per_checksum;
+      if ((crc = ntohl(crc_val(crc2))) != *sums)
+        goto return_crc_error;
+      sums++;
+      data += bytes_per_checksum;
+      if ((crc = ntohl(crc_val(crc3))) != *sums)
+        goto return_crc_error;
+      sums++;
+      data += bytes_per_checksum;
+      n_blocks -= 3;
+    }
+
+    /* One or two blocks */
+    if (n_blocks) {
+      crc1 = crc2 = crc_init();
+      pipelined_crc32c(&crc1, &crc2, &crc3, data, bytes_per_checksum, n_blocks);
+
+      if ((crc = ntohl(crc_val(crc1))) != *sums)
+        goto return_crc_error;
+      data += bytes_per_checksum;
+      sums++;
+      if (n_blocks == 2) {
+        if ((crc = ntohl(crc_val(crc2))) != *sums)
+          goto return_crc_error;
+        sums++;
+        data += bytes_per_checksum;
+      }
+    }
+ 
+    /* For something smaller than a block */
+    if (remainder) {
+      crc1 = crc_init();
+      pipelined_crc32c(&crc1, &crc2, &crc3, data, remainder, 1);
+
+      if ((crc = ntohl(crc_val(crc1))) != *sums)
+        goto return_crc_error;
+    }
+    return CHECKSUMS_VALID;
+  }
+#endif
+
   while (likely(data_len > 0)) {
     int len = likely(data_len >= bytes_per_checksum) ? bytes_per_checksum : data_len;
-    uint32_t crc = crc_init();
+    crc = crc_init();
     crc = crc_update_func(crc, data, len);
     crc = ntohl(crc_val(crc));
     if (unlikely(crc != *sums)) {
-      if (error_info != NULL) {
-        error_info->got_crc = crc;
-        error_info->expected_crc = *sums;
-        error_info->bad_data = data;
-      }
-      return INVALID_CHECKSUM_DETECTED;
+      goto return_crc_error;
     }
     data += len;
     data_len -= len;
     sums++;
   }
   return CHECKSUMS_VALID;
+
+return_crc_error:
+  if (error_info != NULL) {
+    error_info->got_crc = crc;
+    error_info->expected_crc = *sums;
+    error_info->bad_data = data;
+  }
+  return INVALID_CHECKSUM_DETECTED;
 }
 
 
@@ -154,3 +232,417 @@ static uint32_t crc32_zlib_sb8(
   }
   return crc;    
 }
+
+///////////////////////////////////////////////////////////////////////////
+// Begin code for SSE4.2 specific hardware support of CRC32C
+///////////////////////////////////////////////////////////////////////////
+
+#if (defined(__amd64__) || defined(__i386)) && defined(__GNUC__)
+#  define SSE42_FEATURE_BIT (1 << 20)
+#  define CPUID_FEATURES 1
+/**
+ * Call the cpuid instruction to determine CPU feature flags.
+ */
+static uint32_t cpuid(uint32_t eax_in) {
+  uint32_t eax, ebx, ecx, edx;
+#  if defined(__PIC__) && !defined(__LP64__)
+// 32-bit PIC code uses the ebx register for the base offset --
+// have to save and restore it on the stack
+  asm("pushl %%ebx\n\t"
+      "cpuid\n\t"
+      "movl %%ebx, %[ebx]\n\t"
+      "popl %%ebx" : "=a" (eax), [ebx] "=r"(ebx),  "=c"(ecx), "=d"(edx) : "a" (eax_in)
+      : "cc");
+#  else
+  asm("cpuid" : "=a" (eax), "=b"(ebx), "=c"(ecx), "=d"(edx) : "a"(eax_in)
+      : "cc");
+#  endif
+
+  return ecx;
+}
+
+/**
+ * On library load, initiailize the cached value above for
+ * whether the cpu supports SSE4.2's crc32 instruction.
+ */
+void __attribute__ ((constructor)) init_cpu_support_flag(void) {
+  uint32_t ecx = cpuid(CPUID_FEATURES);
+  cached_cpu_supports_crc32 = ecx & SSE42_FEATURE_BIT;
+}
+
+
+//
+// Definitions of the SSE4.2 crc32 operations. Using these instead of
+// the GCC __builtin_* intrinsics allows this code to compile without
+// -msse4.2, since we do dynamic CPU detection at runtime.
+//
+
+#  ifdef __LP64__
+inline uint64_t _mm_crc32_u64(uint64_t crc, uint64_t value) {
+  asm("crc32q %[value], %[crc]\n" : [crc] "+r" (crc) : [value] "rm" (value));
+  return crc;
+}
+#  endif
+
+inline uint32_t _mm_crc32_u32(uint32_t crc, uint32_t value) {
+  asm("crc32l %[value], %[crc]\n" : [crc] "+r" (crc) : [value] "rm" (value));
+  return crc;
+}
+
+inline uint32_t _mm_crc32_u16(uint32_t crc, uint16_t value) {
+  asm("crc32w %[value], %[crc]\n" : [crc] "+r" (crc) : [value] "rm" (value));
+  return crc;
+}
+
+inline uint32_t _mm_crc32_u8(uint32_t crc, uint8_t value) {
+  asm("crc32b %[value], %[crc]\n" : [crc] "+r" (crc) : [value] "rm" (value));
+  return crc;
+}
+ 
+
+#  ifdef __LP64__
+/**
+ * Hardware-accelerated CRC32C calculation using the 64-bit instructions.
+ */
+static uint32_t crc32c_hardware(uint32_t crc, const uint8_t* p_buf, size_t length) {
+  // start directly at p_buf, even if it's an unaligned address. According
+  // to the original author of this code, doing a small run of single bytes
+  // to word-align the 64-bit instructions doesn't seem to help, but
+  // we haven't reconfirmed those benchmarks ourselves.
+  uint64_t crc64bit = crc;
+  size_t i;
+  for (i = 0; i < length / sizeof(uint64_t); i++) {
+    crc64bit = _mm_crc32_u64(crc64bit, *(uint64_t*) p_buf);
+    p_buf += sizeof(uint64_t);
+  }
+
+  // This ugly switch is slightly faster for short strings than the straightforward loop
+  uint32_t crc32bit = (uint32_t) crc64bit;
+  length &= sizeof(uint64_t) - 1;
+  switch (length) {
+    case 7:
+      crc32bit = _mm_crc32_u8(crc32bit, *p_buf++);
+    case 6:
+      crc32bit = _mm_crc32_u16(crc32bit, *(uint16_t*) p_buf);
+      p_buf += 2;
+    // case 5 is below: 4 + 1
+    case 4:
+      crc32bit = _mm_crc32_u32(crc32bit, *(uint32_t*) p_buf);
+      break;
+    case 3:
+      crc32bit = _mm_crc32_u8(crc32bit, *p_buf++);
+    case 2:
+      crc32bit = _mm_crc32_u16(crc32bit, *(uint16_t*) p_buf);
+      break;
+    case 5:
+      crc32bit = _mm_crc32_u32(crc32bit, *(uint32_t*) p_buf);
+      p_buf += 4;
+    case 1:
+      crc32bit = _mm_crc32_u8(crc32bit, *p_buf);
+      break;
+    case 0:
+      break;
+    default:
+      // This should never happen; enable in debug code
+      assert(0 && "ended up with 8 or more bytes at tail of calculation");
+  }
+
+  return crc32bit;
+}
+
+#ifdef USE_PIPELINED
+/**
+ * Pipelined version of hardware-accelerated CRC32C calculation using
+ * the 64 bit crc32q instruction. 
+ * One crc32c instruction takes three cycles, but two more with no data
+ * dependency can be in the pipeline to achieve something close to single 
+ * instruction/cycle. Here we feed three blocks in RR.
+ *
+ *   crc1, crc2, crc3 : Store initial checksum for each block before
+ *           calling. When it returns, updated checksums are stored.
+ *   p_buf : The base address of the data buffer. The buffer should be
+ *           at least as big as block_size * num_blocks.
+ *   block_size : The size of each block in bytes.
+ *   num_blocks : The number of blocks to work on. Min = 1, Max = 3
+ */
+static void pipelined_crc32c(uint32_t *crc1, uint32_t *crc2, uint32_t *crc3, const uint8_t *p_buf, size_t block_size, int num_blocks) {
+  uint64_t c1 = *crc1;
+  uint64_t c2 = *crc2;
+  uint64_t c3 = *crc3;
+  uint64_t *data = (uint64_t*)p_buf;
+  int counter = block_size / sizeof(uint64_t);
+  int remainder = block_size % sizeof(uint64_t);
+  uint8_t *bdata;
+
+  /* We do switch here because the loop has to be tight in order
+   * to fill the pipeline. Any other statement inside the loop
+   * or inbetween crc32 instruction can slow things down. Calling
+   * individual crc32 instructions three times from C also causes
+   * gcc to insert other instructions inbetween.
+   *
+   * Do not rearrange the following code unless you have verified
+   * the generated machine code is as efficient as before.
+   */
+  switch (num_blocks) {
+    case 3:
+      /* Do three blocks */
+      while (likely(counter)) {
+        __asm__ __volatile__(
+        "crc32q (%7), %0;\n\t"
+        "crc32q (%7,%6,1), %1;\n\t"
+        "crc32q (%7,%6,2), %2;\n\t"
+         : "=r"(c1), "=r"(c2), "=r"(c3)
+         : "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(data)
+        );
+        data++;
+        counter--;
+      }
+
+      /* Take care of the remainder. They are only up to three bytes,
+       * so performing byte-level crc32 won't take much time.
+       */
+      bdata = (uint8_t*)data;
+      while (likely(remainder)) {
+        __asm__ __volatile__(
+        "crc32b (%7), %0;\n\t"
+        "crc32b (%7,%6,1), %1;\n\t"
+        "crc32b (%7,%6,2), %2;\n\t"
+         : "=r"(c1), "=r"(c2), "=r"(c3)
+         : "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(bdata)
+        );
+        bdata++;
+        remainder--;
+      }
+      break;
+    case 2:
+      /* Do two blocks */
+      while (likely(counter)) {
+        __asm__ __volatile__(
+        "crc32q (%5), %0;\n\t"
+        "crc32q (%5,%4,1), %1;\n\t"
+         : "=r"(c1), "=r"(c2) 
+         : "r"(c1), "r"(c2), "r"(block_size), "r"(data)
+        );
+        data++;
+        counter--;
+      }
+
+      bdata = (uint8_t*)data;
+      while (likely(remainder)) {
+        __asm__ __volatile__(
+        "crc32b (%5), %0;\n\t"
+        "crc32b (%5,%4,1), %1;\n\t"
+         : "=r"(c1), "=r"(c2) 
+         : "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(bdata)
+        );
+        bdata++;
+        remainder--;
+      }
+      break;
+    case 1:
+      /* single block */
+      while (likely(counter)) {
+        __asm__ __volatile__(
+        "crc32q (%2), %0;\n\t"
+         : "=r"(c1) 
+         : "r"(c1), "r"(data)
+        );
+        data++;
+        counter--;
+      }
+      bdata = (uint8_t*)data;
+      while (likely(remainder)) {
+        __asm__ __volatile__(
+        "crc32b (%2), %0;\n\t"
+         : "=r"(c1) 
+         : "r"(c1), "r"(bdata)
+        );
+        bdata++;
+        remainder--;
+      }
+      break;
+    case 0:
+      return;
+    default:
+      assert(0 && "BUG: Invalid number of checksum blocks");
+  }
+
+  *crc1 = c1;
+  *crc2 = c2;
+  *crc3 = c3;
+  return;
+}
+#endif /* USE_PIPELINED */
+
+# else  // 32-bit
+
+/**
+ * Hardware-accelerated CRC32C calculation using the 32-bit instructions.
+ */
+static uint32_t crc32c_hardware(uint32_t crc, const uint8_t* p_buf, size_t length) {
+  // start directly at p_buf, even if it's an unaligned address. According
+  // to the original author of this code, doing a small run of single bytes
+  // to word-align the 64-bit instructions doesn't seem to help, but
+  // we haven't reconfirmed those benchmarks ourselves.
+  size_t i;
+  for (i = 0; i < length / sizeof(uint32_t); i++) {
+    crc = _mm_crc32_u32(crc, *(uint32_t*) p_buf);
+    p_buf += sizeof(uint32_t);
+  }
+
+  // This ugly switch is slightly faster for short strings than the straightforward loop
+  length &= sizeof(uint32_t) - 1;
+  switch (length) {
+    case 3:
+      crc = _mm_crc32_u8(crc, *p_buf++);
+    case 2:
+      crc = _mm_crc32_u16(crc, *(uint16_t*) p_buf);
+      break;
+    case 1:
+      crc = _mm_crc32_u8(crc, *p_buf);
+      break;
+    case 0:
+      break;
+    default:
+      // This should never happen; enable in debug code
+      assert(0 && "ended up with 4 or more bytes at tail of calculation");
+  }
+
+  return crc;
+}
+
+#ifdef USE_PIPELINED
+/**
+ * Pipelined version of hardware-accelerated CRC32C calculation using
+ * the 32 bit crc32l instruction. 
+ * One crc32c instruction takes three cycles, but two more with no data
+ * dependency can be in the pipeline to achieve something close to single 
+ * instruction/cycle. Here we feed three blocks in RR.
+ *
+ *   crc1, crc2, crc3 : Store initial checksum for each block before
+ *                calling. When it returns, updated checksums are stored.
+ *   data       : The base address of the data buffer. The buffer should be
+ *                at least as big as block_size * num_blocks.
+ *   block_size : The size of each block in bytes. 
+ *   num_blocks : The number of blocks to work on. Min = 1, Max = 3
+ */
+static void pipelined_crc32c(uint32_t *crc1, uint32_t *crc2, uint32_t *crc3, const uint8_t *p_buf, size_t block_size, int num_blocks) {
+  uint32_t c1 = *crc1;
+  uint32_t c2 = *crc2;
+  uint32_t c3 = *crc3;
+  int counter = block_size / sizeof(uint32_t);
+  int remainder = block_size % sizeof(uint32_t);
+  uint32_t *data = (uint32_t*)p_buf;
+  uint8_t *bdata;
+
+  /* We do switch here because the loop has to be tight in order
+   * to fill the pipeline. Any other statement inside the loop
+   * or inbetween crc32 instruction can slow things down. Calling
+   * individual crc32 instructions three times from C also causes
+   * gcc to insert other instructions inbetween.
+   *
+   * Do not rearrange the following code unless you have verified
+   * the generated machine code is as efficient as before.
+   */
+  switch (num_blocks) {
+    case 3:
+      /* Do three blocks */
+      while (likely(counter)) {
+        __asm__ __volatile__(
+        "crc32l (%7), %0;\n\t"
+        "crc32l (%7,%6,1), %1;\n\t"
+        "crc32l (%7,%6,2), %2;\n\t"
+         : "=r"(c1), "=r"(c2), "=r"(c3)
+         : "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(data)
+        );
+        data++;
+        counter--;
+      }
+      /* Take care of the remainder. They are only up to three bytes,
+       * so performing byte-level crc32 won't take much time.
+       */
+      bdata = (uint8_t*)data;
+      while (likely(remainder)) {
+        __asm__ __volatile__(
+        "crc32b (%7), %0;\n\t"
+        "crc32b (%7,%6,1), %1;\n\t"
+        "crc32b (%7,%6,2), %2;\n\t"
+         : "=r"(c1), "=r"(c2), "=r"(c3)
+         : "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(bdata)
+        );
+        bdata++;
+        remainder--;
+      }
+      break;
+    case 2:
+      /* Do two blocks */
+      while (likely(counter)) {
+        __asm__ __volatile__(
+        "crc32l (%5), %0;\n\t"
+        "crc32l (%5,%4,1), %1;\n\t"
+         : "=r"(c1), "=r"(c2) 
+         : "r"(c1), "r"(c2), "r"(block_size), "r"(data)
+        );
+        data++;
+        counter--;
+      }
+
+      bdata = (uint8_t*)data;
+      while (likely(remainder)) {
+        __asm__ __volatile__(
+        "crc32b (%5), %0;\n\t"
+        "crc32b (%5,%4,1), %1;\n\t"
+         : "=r"(c1), "=r"(c2) 
+         : "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(bdata)
+        );
+        bdata++;
+        remainder--;
+      }
+      break;
+    case 1:
+      /* single block */
+      while (likely(counter)) {
+        __asm__ __volatile__(
+        "crc32l (%2), %0;\n\t"
+         : "=r"(c1) 
+         : "r"(c1), "r"(data)
+        );
+        data++;
+        counter--;
+      }
+      bdata = (uint8_t*)data;
+      while (likely(remainder)) {
+        __asm__ __volatile__(
+        "crc32b (%2), %0;\n\t"
+         : "=r"(c1) 
+         : "r"(c1), "r"(bdata)
+        );
+        bdata++;
+        remainder--;
+      }
+      break;
+    case 0:
+       return;
+    default:
+      assert(0 && "BUG: Invalid number of checksum blocks");
+  }
+
+  *crc1 = c1;
+  *crc2 = c2;
+  *crc3 = c3;
+  return;
+}
+
+#endif /* USE_PIPELINED */
+
+# endif // 64-bit vs 32-bit
+
+#else // end x86 architecture
+
+static uint32_t crc32c_hardware(uint32_t crc, const uint8_t* data, size_t length) {
+  // never called!
+  assert(0 && "hardware crc called on an unsupported platform");
+  return 0;
+}
+
+#endif