Files
wuffs/std/crc32/common_up_x86_avx2.wuffs
T
Patrick Georgi f563f32f82 Add SPDX identifier to license headers
SPDX is an ISO standard for describing licenses in a computer readable
format, which is useful when applying license compliance tools and the
like to source code.

In addition to ensuring that the Wuffs C distribution is properly
annotated to the benefit of projects using Wuffs and SPDX-supporting
tools, add it everywhere else where there's a license in this repo for
consistency, too.
2023-10-06 20:52:36 +02:00

135 lines
5.2 KiB
Plaintext
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
// Copyright 2021 The Wuffs Authors.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.
//
// SPDX-License-Identifier: Apache-2.0 OR MIT
// --------
// See "SIMD Implementations" in README.md for a link to Gopal et al. "Fast CRC
// Computation for Generic Polynomials Using PCLMULQDQ Instruction".
// up_x86_avx2 is exactly the same as up_x86_sse42 except for the "choose
// cpu_arch >= x86_avx2". With AVX, PCLMULQDQ has a three-operand form, not
// just a two-operand form: https://www.felixcloutier.com/x86/pclmulqdq
pri func ieee_hasher.up_x86_avx2!(x: roslice base.u8),
choose cpu_arch >= x86_avx2,
{
var s : base.u32
var p : roslice base.u8
var util : base.x86_sse42_utility
var k : base.x86_m128i
var x0 : base.x86_m128i
var x1 : base.x86_m128i
var x2 : base.x86_m128i
var x3 : base.x86_m128i
var y0 : base.x86_m128i
var y1 : base.x86_m128i
var y2 : base.x86_m128i
var y3 : base.x86_m128i
var tail_index : base.u64
s = 0xFFFF_FFFF ^ this.state
// Align to a 16-byte boundary.
while (args.x.length() > 0) and ((15 & args.x.uintptr_low_12_bits()) <> 0) {
s = IEEE_TABLE[0][((s & 0xFF) as base.u8) ^ args.x[0]] ^ (s >> 8)
args.x = args.x[1 ..]
} endwhile
// For short inputs, just do a simple loop.
if args.x.length() < 64 {
iterate (p = args.x)(length: 1, advance: 1, unroll: 1) {
s = IEEE_TABLE[0][((s & 0xFF) as base.u8) ^ p[0]] ^ (s >> 8)
}
this.state = 0xFFFF_FFFF ^ s
return nothing
}
// Load 128×4 = 512 bits from the first 64-byte chunk.
x0 = util.make_m128i_slice128(a: args.x[0x00 .. 0x10])
x1 = util.make_m128i_slice128(a: args.x[0x10 .. 0x20])
x2 = util.make_m128i_slice128(a: args.x[0x20 .. 0x30])
x3 = util.make_m128i_slice128(a: args.x[0x30 .. 0x40])
// Combine with the initial state.
x0 = x0._mm_xor_si128(b: util.make_m128i_single_u32(a: s))
// Process the remaining 64-byte chunks.
k = util.make_m128i_slice128(a: IEEE_X86_SSE42_K1K2[.. 16])
iterate (p = args.x[64 ..])(length: 64, advance: 64, unroll: 1) {
y0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x00)
y1 = x1._mm_clmulepi64_si128(b: k, imm8: 0x00)
y2 = x2._mm_clmulepi64_si128(b: k, imm8: 0x00)
y3 = x3._mm_clmulepi64_si128(b: k, imm8: 0x00)
x0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x11)
x1 = x1._mm_clmulepi64_si128(b: k, imm8: 0x11)
x2 = x2._mm_clmulepi64_si128(b: k, imm8: 0x11)
x3 = x3._mm_clmulepi64_si128(b: k, imm8: 0x11)
x0 = x0._mm_xor_si128(b: y0)._mm_xor_si128(b: util.make_m128i_slice128(a: p[0x00 .. 0x10]))
x1 = x1._mm_xor_si128(b: y1)._mm_xor_si128(b: util.make_m128i_slice128(a: p[0x10 .. 0x20]))
x2 = x2._mm_xor_si128(b: y2)._mm_xor_si128(b: util.make_m128i_slice128(a: p[0x20 .. 0x30]))
x3 = x3._mm_xor_si128(b: y3)._mm_xor_si128(b: util.make_m128i_slice128(a: p[0x30 .. 0x40]))
}
// Reduce 128×4 = 512 bits to 128 bits.
k = util.make_m128i_slice128(a: IEEE_X86_SSE42_K3K4[.. 16])
y0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x00)
x0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x11)
x0 = x0._mm_xor_si128(b: x1)
x0 = x0._mm_xor_si128(b: y0)
y0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x00)
x0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x11)
x0 = x0._mm_xor_si128(b: x2)
x0 = x0._mm_xor_si128(b: y0)
y0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x00)
x0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x11)
x0 = x0._mm_xor_si128(b: x3)
x0 = x0._mm_xor_si128(b: y0)
// Reduce 128 bits to 64 bits.
x1 = x0._mm_clmulepi64_si128(b: k, imm8: 0x10)
x2 = util.make_m128i_multiple_u32(
a00: 0xFFFF_FFFF,
a01: 0x0000_0000,
a02: 0xFFFF_FFFF,
a03: 0x0000_0000)
x0 = x0._mm_srli_si128(imm8: 8)
x0 = x0._mm_xor_si128(b: x1)
k = util.make_m128i_slice128(a: IEEE_X86_SSE42_K5ZZ[.. 16])
x1 = x0._mm_srli_si128(imm8: 4)
x0 = x0._mm_and_si128(b: x2)
x0 = x0._mm_clmulepi64_si128(b: k, imm8: 0x00)
x0 = x0._mm_xor_si128(b: x1)
// Reduce 64 bits to 32 bits (Barrett Reduction) and extract.
//
// Barrett Reduction is Algorithm 1 (page 14) of Gopal et al., after
// adjusting for bit-reflection as per Figure 12 (page 21).
k = util.make_m128i_slice128(a: IEEE_X86_SSE42_PXMU[.. 16])
x1 = x0._mm_and_si128(b: x2)
x1 = x1._mm_clmulepi64_si128(b: k, imm8: 0x10)
x1 = x1._mm_and_si128(b: x2)
x1 = x1._mm_clmulepi64_si128(b: k, imm8: 0x00)
x0 = x0._mm_xor_si128(b: x1)
s = x0._mm_extract_epi32(imm8: 1)
// Handle the tail of args.x that wasn't a complete 64-byte chunk.
tail_index = args.x.length() & 0xFFFF_FFFF_FFFF_FFC0 // And-not 64.
if tail_index < args.x.length() {
iterate (p = args.x[tail_index ..])(length: 1, advance: 1, unroll: 1) {
s = IEEE_TABLE[0][((s & 0xFF) as base.u8) ^ p[0]] ^ (s >> 8)
}
}
this.state = 0xFFFF_FFFF ^ s
}