...
Run Format

Source file src/hash/crc32/crc32_amd64.go

Documentation: hash/crc32

  // Copyright 2011 The Go Authors. All rights reserved.
  // Use of this source code is governed by a BSD-style
  // license that can be found in the LICENSE file.
  
  // AMD64-specific hardware-assisted CRC32 algorithms. See crc32.go for a
  // description of the interface that each architecture-specific file
  // implements.
  
  package crc32
  
  import (
  	"internal/cpu"
  	"unsafe"
  )
  
  // This file contains the code to call the SSE 4.2 version of the Castagnoli
  // and IEEE CRC.
  
  // castagnoliSSE42 is defined in crc32_amd64.s and uses the SSE 4.2 CRC32
  // instruction.
  //go:noescape
  func castagnoliSSE42(crc uint32, p []byte) uint32
  
  // castagnoliSSE42Triple is defined in crc32_amd64.s and uses the SSE 4.2 CRC32
  // instruction.
  //go:noescape
  func castagnoliSSE42Triple(
  	crcA, crcB, crcC uint32,
  	a, b, c []byte,
  	rounds uint32,
  ) (retA uint32, retB uint32, retC uint32)
  
  // ieeeCLMUL is defined in crc_amd64.s and uses the PCLMULQDQ
  // instruction as well as SSE 4.1.
  //go:noescape
  func ieeeCLMUL(crc uint32, p []byte) uint32
  
  const castagnoliK1 = 168
  const castagnoliK2 = 1344
  
  type sse42Table [4]Table
  
  var castagnoliSSE42TableK1 *sse42Table
  var castagnoliSSE42TableK2 *sse42Table
  
  func archAvailableCastagnoli() bool {
  	return cpu.X86.HasSSE42
  }
  
  func archInitCastagnoli() {
  	if !cpu.X86.HasSSE42 {
  		panic("arch-specific Castagnoli not available")
  	}
  	castagnoliSSE42TableK1 = new(sse42Table)
  	castagnoliSSE42TableK2 = new(sse42Table)
  	// See description in updateCastagnoli.
  	//    t[0][i] = CRC(i000, O)
  	//    t[1][i] = CRC(0i00, O)
  	//    t[2][i] = CRC(00i0, O)
  	//    t[3][i] = CRC(000i, O)
  	// where O is a sequence of K zeros.
  	var tmp [castagnoliK2]byte
  	for b := 0; b < 4; b++ {
  		for i := 0; i < 256; i++ {
  			val := uint32(i) << uint32(b*8)
  			castagnoliSSE42TableK1[b][i] = castagnoliSSE42(val, tmp[:castagnoliK1])
  			castagnoliSSE42TableK2[b][i] = castagnoliSSE42(val, tmp[:])
  		}
  	}
  }
  
  // castagnoliShift computes the CRC32-C of K1 or K2 zeroes (depending on the
  // table given) with the given initial crc value. This corresponds to
  // CRC(crc, O) in the description in updateCastagnoli.
  func castagnoliShift(table *sse42Table, crc uint32) uint32 {
  	return table[3][crc>>24] ^
  		table[2][(crc>>16)&0xFF] ^
  		table[1][(crc>>8)&0xFF] ^
  		table[0][crc&0xFF]
  }
  
  func archUpdateCastagnoli(crc uint32, p []byte) uint32 {
  	if !cpu.X86.HasSSE42 {
  		panic("not available")
  	}
  
  	// This method is inspired from the algorithm in Intel's white paper:
  	//    "Fast CRC Computation for iSCSI Polynomial Using CRC32 Instruction"
  	// The same strategy of splitting the buffer in three is used but the
  	// combining calculation is different; the complete derivation is explained
  	// below.
  	//
  	// -- The basic idea --
  	//
  	// The CRC32 instruction (available in SSE4.2) can process 8 bytes at a
  	// time. In recent Intel architectures the instruction takes 3 cycles;
  	// however the processor can pipeline up to three instructions if they
  	// don't depend on each other.
  	//
  	// Roughly this means that we can process three buffers in about the same
  	// time we can process one buffer.
  	//
  	// The idea is then to split the buffer in three, CRC the three pieces
  	// separately and then combine the results.
  	//
  	// Combining the results requires precomputed tables, so we must choose a
  	// fixed buffer length to optimize. The longer the length, the faster; but
  	// only buffers longer than this length will use the optimization. We choose
  	// two cutoffs and compute tables for both:
  	//  - one around 512: 168*3=504
  	//  - one around 4KB: 1344*3=4032
  	//
  	// -- The nitty gritty --
  	//
  	// Let CRC(I, X) be the non-inverted CRC32-C of the sequence X (with
  	// initial non-inverted CRC I). This function has the following properties:
  	//   (a) CRC(I, AB) = CRC(CRC(I, A), B)
  	//   (b) CRC(I, A xor B) = CRC(I, A) xor CRC(0, B)
  	//
  	// Say we want to compute CRC(I, ABC) where A, B, C are three sequences of
  	// K bytes each, where K is a fixed constant. Let O be the sequence of K zero
  	// bytes.
  	//
  	// CRC(I, ABC) = CRC(I, ABO xor C)
  	//             = CRC(I, ABO) xor CRC(0, C)
  	//             = CRC(CRC(I, AB), O) xor CRC(0, C)
  	//             = CRC(CRC(I, AO xor B), O) xor CRC(0, C)
  	//             = CRC(CRC(I, AO) xor CRC(0, B), O) xor CRC(0, C)
  	//             = CRC(CRC(CRC(I, A), O) xor CRC(0, B), O) xor CRC(0, C)
  	//
  	// The castagnoliSSE42Triple function can compute CRC(I, A), CRC(0, B),
  	// and CRC(0, C) efficiently.  We just need to find a way to quickly compute
  	// CRC(uvwx, O) given a 4-byte initial value uvwx. We can precompute these
  	// values; since we can't have a 32-bit table, we break it up into four
  	// 8-bit tables:
  	//
  	//    CRC(uvwx, O) = CRC(u000, O) xor
  	//                   CRC(0v00, O) xor
  	//                   CRC(00w0, O) xor
  	//                   CRC(000x, O)
  	//
  	// We can compute tables corresponding to the four terms for all 8-bit
  	// values.
  
  	crc = ^crc
  
  	// If a buffer is long enough to use the optimization, process the first few
  	// bytes to align the buffer to an 8 byte boundary (if necessary).
  	if len(p) >= castagnoliK1*3 {
  		delta := int(uintptr(unsafe.Pointer(&p[0])) & 7)
  		if delta != 0 {
  			delta = 8 - delta
  			crc = castagnoliSSE42(crc, p[:delta])
  			p = p[delta:]
  		}
  	}
  
  	// Process 3*K2 at a time.
  	for len(p) >= castagnoliK2*3 {
  		// Compute CRC(I, A), CRC(0, B), and CRC(0, C).
  		crcA, crcB, crcC := castagnoliSSE42Triple(
  			crc, 0, 0,
  			p, p[castagnoliK2:], p[castagnoliK2*2:],
  			castagnoliK2/24)
  
  		// CRC(I, AB) = CRC(CRC(I, A), O) xor CRC(0, B)
  		crcAB := castagnoliShift(castagnoliSSE42TableK2, crcA) ^ crcB
  		// CRC(I, ABC) = CRC(CRC(I, AB), O) xor CRC(0, C)
  		crc = castagnoliShift(castagnoliSSE42TableK2, crcAB) ^ crcC
  		p = p[castagnoliK2*3:]
  	}
  
  	// Process 3*K1 at a time.
  	for len(p) >= castagnoliK1*3 {
  		// Compute CRC(I, A), CRC(0, B), and CRC(0, C).
  		crcA, crcB, crcC := castagnoliSSE42Triple(
  			crc, 0, 0,
  			p, p[castagnoliK1:], p[castagnoliK1*2:],
  			castagnoliK1/24)
  
  		// CRC(I, AB) = CRC(CRC(I, A), O) xor CRC(0, B)
  		crcAB := castagnoliShift(castagnoliSSE42TableK1, crcA) ^ crcB
  		// CRC(I, ABC) = CRC(CRC(I, AB), O) xor CRC(0, C)
  		crc = castagnoliShift(castagnoliSSE42TableK1, crcAB) ^ crcC
  		p = p[castagnoliK1*3:]
  	}
  
  	// Use the simple implementation for what's left.
  	crc = castagnoliSSE42(crc, p)
  	return ^crc
  }
  
  func archAvailableIEEE() bool {
  	return cpu.X86.HasPCLMULQDQ && cpu.X86.HasSSE41
  }
  
  var archIeeeTable8 *slicing8Table
  
  func archInitIEEE() {
  	if !cpu.X86.HasPCLMULQDQ || !cpu.X86.HasSSE41 {
  		panic("not available")
  	}
  	// We still use slicing-by-8 for small buffers.
  	archIeeeTable8 = slicingMakeTable(IEEE)
  }
  
  func archUpdateIEEE(crc uint32, p []byte) uint32 {
  	if !cpu.X86.HasPCLMULQDQ || !cpu.X86.HasSSE41 {
  		panic("not available")
  	}
  
  	if len(p) >= 64 {
  		left := len(p) & 15
  		do := len(p) - left
  		crc = ^ieeeCLMUL(^crc, p[:do])
  		p = p[do:]
  	}
  	if len(p) == 0 {
  		return crc
  	}
  	return slicingUpdate(crc, archIeeeTable8, p)
  }
  

View as plain text