Source file src/hash/crc32/crc32_amd64.go

     1  // Copyright 2011 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // AMD64-specific hardware-assisted CRC32 algorithms. See crc32.go for a
     6  // description of the interface that each architecture-specific file
     7  // implements.
     8  
     9  package crc32
    10  
    11  import (
    12  	"internal/cpu"
    13  	"unsafe"
    14  )
    15  
    16  // This file contains the code to call the SSE 4.2 version of the Castagnoli
    17  // and IEEE CRC.
    18  
    19  // castagnoliSSE42 is defined in crc32_amd64.s and uses the SSE 4.2 CRC32
    20  // instruction.
    21  //go:noescape
    22  func castagnoliSSE42(crc uint32, p []byte) uint32
    23  
    24  // castagnoliSSE42Triple is defined in crc32_amd64.s and uses the SSE 4.2 CRC32
    25  // instruction.
    26  //go:noescape
    27  func castagnoliSSE42Triple(
    28  	crcA, crcB, crcC uint32,
    29  	a, b, c []byte,
    30  	rounds uint32,
    31  ) (retA uint32, retB uint32, retC uint32)
    32  
    33  // ieeeCLMUL is defined in crc_amd64.s and uses the PCLMULQDQ
    34  // instruction as well as SSE 4.1.
    35  //go:noescape
    36  func ieeeCLMUL(crc uint32, p []byte) uint32
    37  
    38  const castagnoliK1 = 168
    39  const castagnoliK2 = 1344
    40  
    41  type sse42Table [4]Table
    42  
    43  var castagnoliSSE42TableK1 *sse42Table
    44  var castagnoliSSE42TableK2 *sse42Table
    45  
    46  func archAvailableCastagnoli() bool {
    47  	return cpu.X86.HasSSE42
    48  }
    49  
    50  func archInitCastagnoli() {
    51  	if !cpu.X86.HasSSE42 {
    52  		panic("arch-specific Castagnoli not available")
    53  	}
    54  	castagnoliSSE42TableK1 = new(sse42Table)
    55  	castagnoliSSE42TableK2 = new(sse42Table)
    56  	// See description in updateCastagnoli.
    57  	//    t[0][i] = CRC(i000, O)
    58  	//    t[1][i] = CRC(0i00, O)
    59  	//    t[2][i] = CRC(00i0, O)
    60  	//    t[3][i] = CRC(000i, O)
    61  	// where O is a sequence of K zeros.
    62  	var tmp [castagnoliK2]byte
    63  	for b := 0; b < 4; b++ {
    64  		for i := 0; i < 256; i++ {
    65  			val := uint32(i) << uint32(b*8)
    66  			castagnoliSSE42TableK1[b][i] = castagnoliSSE42(val, tmp[:castagnoliK1])
    67  			castagnoliSSE42TableK2[b][i] = castagnoliSSE42(val, tmp[:])
    68  		}
    69  	}
    70  }
    71  
    72  // castagnoliShift computes the CRC32-C of K1 or K2 zeroes (depending on the
    73  // table given) with the given initial crc value. This corresponds to
    74  // CRC(crc, O) in the description in updateCastagnoli.
    75  func castagnoliShift(table *sse42Table, crc uint32) uint32 {
    76  	return table[3][crc>>24] ^
    77  		table[2][(crc>>16)&0xFF] ^
    78  		table[1][(crc>>8)&0xFF] ^
    79  		table[0][crc&0xFF]
    80  }
    81  
    82  func archUpdateCastagnoli(crc uint32, p []byte) uint32 {
    83  	if !cpu.X86.HasSSE42 {
    84  		panic("not available")
    85  	}
    86  
    87  	// This method is inspired from the algorithm in Intel's white paper:
    88  	//    "Fast CRC Computation for iSCSI Polynomial Using CRC32 Instruction"
    89  	// The same strategy of splitting the buffer in three is used but the
    90  	// combining calculation is different; the complete derivation is explained
    91  	// below.
    92  	//
    93  	// -- The basic idea --
    94  	//
    95  	// The CRC32 instruction (available in SSE4.2) can process 8 bytes at a
    96  	// time. In recent Intel architectures the instruction takes 3 cycles;
    97  	// however the processor can pipeline up to three instructions if they
    98  	// don't depend on each other.
    99  	//
   100  	// Roughly this means that we can process three buffers in about the same
   101  	// time we can process one buffer.
   102  	//
   103  	// The idea is then to split the buffer in three, CRC the three pieces
   104  	// separately and then combine the results.
   105  	//
   106  	// Combining the results requires precomputed tables, so we must choose a
   107  	// fixed buffer length to optimize. The longer the length, the faster; but
   108  	// only buffers longer than this length will use the optimization. We choose
   109  	// two cutoffs and compute tables for both:
   110  	//  - one around 512: 168*3=504
   111  	//  - one around 4KB: 1344*3=4032
   112  	//
   113  	// -- The nitty gritty --
   114  	//
   115  	// Let CRC(I, X) be the non-inverted CRC32-C of the sequence X (with
   116  	// initial non-inverted CRC I). This function has the following properties:
   117  	//   (a) CRC(I, AB) = CRC(CRC(I, A), B)
   118  	//   (b) CRC(I, A xor B) = CRC(I, A) xor CRC(0, B)
   119  	//
   120  	// Say we want to compute CRC(I, ABC) where A, B, C are three sequences of
   121  	// K bytes each, where K is a fixed constant. Let O be the sequence of K zero
   122  	// bytes.
   123  	//
   124  	// CRC(I, ABC) = CRC(I, ABO xor C)
   125  	//             = CRC(I, ABO) xor CRC(0, C)
   126  	//             = CRC(CRC(I, AB), O) xor CRC(0, C)
   127  	//             = CRC(CRC(I, AO xor B), O) xor CRC(0, C)
   128  	//             = CRC(CRC(I, AO) xor CRC(0, B), O) xor CRC(0, C)
   129  	//             = CRC(CRC(CRC(I, A), O) xor CRC(0, B), O) xor CRC(0, C)
   130  	//
   131  	// The castagnoliSSE42Triple function can compute CRC(I, A), CRC(0, B),
   132  	// and CRC(0, C) efficiently.  We just need to find a way to quickly compute
   133  	// CRC(uvwx, O) given a 4-byte initial value uvwx. We can precompute these
   134  	// values; since we can't have a 32-bit table, we break it up into four
   135  	// 8-bit tables:
   136  	//
   137  	//    CRC(uvwx, O) = CRC(u000, O) xor
   138  	//                   CRC(0v00, O) xor
   139  	//                   CRC(00w0, O) xor
   140  	//                   CRC(000x, O)
   141  	//
   142  	// We can compute tables corresponding to the four terms for all 8-bit
   143  	// values.
   144  
   145  	crc = ^crc
   146  
   147  	// If a buffer is long enough to use the optimization, process the first few
   148  	// bytes to align the buffer to an 8 byte boundary (if necessary).
   149  	if len(p) >= castagnoliK1*3 {
   150  		delta := int(uintptr(unsafe.Pointer(&p[0])) & 7)
   151  		if delta != 0 {
   152  			delta = 8 - delta
   153  			crc = castagnoliSSE42(crc, p[:delta])
   154  			p = p[delta:]
   155  		}
   156  	}
   157  
   158  	// Process 3*K2 at a time.
   159  	for len(p) >= castagnoliK2*3 {
   160  		// Compute CRC(I, A), CRC(0, B), and CRC(0, C).
   161  		crcA, crcB, crcC := castagnoliSSE42Triple(
   162  			crc, 0, 0,
   163  			p, p[castagnoliK2:], p[castagnoliK2*2:],
   164  			castagnoliK2/24)
   165  
   166  		// CRC(I, AB) = CRC(CRC(I, A), O) xor CRC(0, B)
   167  		crcAB := castagnoliShift(castagnoliSSE42TableK2, crcA) ^ crcB
   168  		// CRC(I, ABC) = CRC(CRC(I, AB), O) xor CRC(0, C)
   169  		crc = castagnoliShift(castagnoliSSE42TableK2, crcAB) ^ crcC
   170  		p = p[castagnoliK2*3:]
   171  	}
   172  
   173  	// Process 3*K1 at a time.
   174  	for len(p) >= castagnoliK1*3 {
   175  		// Compute CRC(I, A), CRC(0, B), and CRC(0, C).
   176  		crcA, crcB, crcC := castagnoliSSE42Triple(
   177  			crc, 0, 0,
   178  			p, p[castagnoliK1:], p[castagnoliK1*2:],
   179  			castagnoliK1/24)
   180  
   181  		// CRC(I, AB) = CRC(CRC(I, A), O) xor CRC(0, B)
   182  		crcAB := castagnoliShift(castagnoliSSE42TableK1, crcA) ^ crcB
   183  		// CRC(I, ABC) = CRC(CRC(I, AB), O) xor CRC(0, C)
   184  		crc = castagnoliShift(castagnoliSSE42TableK1, crcAB) ^ crcC
   185  		p = p[castagnoliK1*3:]
   186  	}
   187  
   188  	// Use the simple implementation for what's left.
   189  	crc = castagnoliSSE42(crc, p)
   190  	return ^crc
   191  }
   192  
   193  func archAvailableIEEE() bool {
   194  	return cpu.X86.HasPCLMULQDQ && cpu.X86.HasSSE41
   195  }
   196  
   197  var archIeeeTable8 *slicing8Table
   198  
   199  func archInitIEEE() {
   200  	if !cpu.X86.HasPCLMULQDQ || !cpu.X86.HasSSE41 {
   201  		panic("not available")
   202  	}
   203  	// We still use slicing-by-8 for small buffers.
   204  	archIeeeTable8 = slicingMakeTable(IEEE)
   205  }
   206  
   207  func archUpdateIEEE(crc uint32, p []byte) uint32 {
   208  	if !cpu.X86.HasPCLMULQDQ || !cpu.X86.HasSSE41 {
   209  		panic("not available")
   210  	}
   211  
   212  	if len(p) >= 64 {
   213  		left := len(p) & 15
   214  		do := len(p) - left
   215  		crc = ^ieeeCLMUL(^crc, p[:do])
   216  		p = p[do:]
   217  	}
   218  	if len(p) == 0 {
   219  		return crc
   220  	}
   221  	return slicingUpdate(crc, archIeeeTable8, p)
   222  }
   223  

View as plain text