github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-06-28 16:41:01 +03:00
Code Issues Packages Projects Releases Wiki Activity

Truffle simd vectorized

Browse Source
This commit is contained in:
apostolos 2021-07-02 17:12:47 +03:00 committed by Konstantinos Margaritis
parent d1009e8830
commit 1ce5e17ce9
7 changed files with 413 additions and 609 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
CMakeLists.txt
View File

@ -696,7 +696,7 @@ set (hs_exec_SRCS
src/nfa/tamarama.c src/nfa/tamarama.c
src/nfa/tamarama.h src/nfa/tamarama.h
src/nfa/tamarama_internal.h src/nfa/tamarama_internal.h
src/nfa/truffle.c src/nfa/truffle.cpp
src/nfa/truffle.h src/nfa/truffle.h
src/nfa/vermicelli.h src/nfa/vermicelli.h
src/nfa/vermicelli_run.h src/nfa/vermicelli_run.h

608
src/nfa/truffle.c
View File

@ -1,608 +0,0 @@
/*
* Copyright (c) 2015-2017, Intel Corporation
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* * Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of Intel Corporation nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
/*
* Matches a byte in a charclass using three shuffles
*/
#include "ue2common.h"
#include "truffle.h"
#include "util/arch.h"
#include "util/bitutils.h"
#include "util/simd_utils.h"
#if !defined(HAVE_AVX2)
static really_inline
const u8 *lastMatch(const u8 *buf, u32 z) {
if (unlikely(z != 0xffff)) {
u32 pos = clz32(~z & 0xffff);
assert(pos >= 16 && pos < 32);
return buf + (31 - pos);
}
return NULL; // no match
}
static really_inline
const u8 *firstMatch(const u8 *buf, u32 z) {
if (unlikely(z != 0xffff)) {
u32 pos = ctz32(~z & 0xffff);
assert(pos < 16);
return buf + pos;
}
return NULL; // no match
}
static really_inline
u32 block(m128 shuf_mask_lo_highclear, m128 shuf_mask_lo_highset, m128 v) {
m128 highconst = set1_16x8(0x80);
m128 shuf_mask_hi = set1_2x64(0x8040201008040201);
// and now do the real work
m128 shuf1 = pshufb_m128(shuf_mask_lo_highclear, v);
m128 t1 = xor128(v, highconst);
m128 shuf2 = pshufb_m128(shuf_mask_lo_highset, t1);
m128 t2 = andnot128(highconst, rshift64_m128(v, 4));
m128 shuf3 = pshufb_m128(shuf_mask_hi, t2);
m128 tmp = and128(or128(shuf1, shuf2), shuf3);
m128 tmp2 = eq128(tmp, zeroes128());
u32 z = movemask128(tmp2);
return z;
}
static
const u8 *truffleMini(m128 shuf_mask_lo_highclear, m128 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
uintptr_t len = buf_end - buf;
assert(len < 16);
m128 chars = zeroes128();
memcpy(&chars, buf, len);
u32 z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
// can't be these bytes in z
u32 mask = (0xffff >> (16 - len)) ^ 0xffff;
const u8 *rv = firstMatch(buf, z | mask);
if (rv) {
return rv;
} else {
return buf_end;
}
}
static really_inline
const u8 *fwdBlock(m128 shuf_mask_lo_highclear, m128 shuf_mask_lo_highset,
m128 v, const u8 *buf) {
u32 z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, v);
return firstMatch(buf, z);
}
static really_inline
const u8 *revBlock(m128 shuf_mask_lo_highclear, m128 shuf_mask_lo_highset,
m128 v, const u8 *buf) {
u32 z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, v);
return lastMatch(buf, z);
}
const u8 *truffleExec(m128 shuf_mask_lo_highclear,
m128 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
DEBUG_PRINTF("len %zu\n", buf_end - buf);
assert(buf && buf_end);
assert(buf < buf_end);
const u8 *rv;
if (buf_end - buf < 16) {
return truffleMini(shuf_mask_lo_highclear, shuf_mask_lo_highset, buf,
buf_end);
}
size_t min = (size_t)buf % 16;
assert(buf_end - buf >= 16);
// Preconditioning: most of the time our buffer won't be aligned.
m128 chars = loadu128(buf);
rv = fwdBlock(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars, buf);
if (rv) {
return rv;
}
buf += (16 - min);
const u8 *last_block = buf_end - 16;
while (buf < last_block) {
m128 lchars = load128(buf);
rv = fwdBlock(shuf_mask_lo_highclear, shuf_mask_lo_highset, lchars,
buf);
if (rv) {
return rv;
}
buf += 16;
}
// Use an unaligned load to mop up the last 16 bytes and get an accurate
// picture to buf_end.
assert(buf <= buf_end && buf >= buf_end - 16);
chars = loadu128(buf_end - 16);
rv = fwdBlock(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars,
buf_end - 16);
if (rv) {
return rv;
}
return buf_end;
}
static
const u8 *truffleRevMini(m128 shuf_mask_lo_highclear,
m128 shuf_mask_lo_highset, const u8 *buf,
const u8 *buf_end) {
uintptr_t len = buf_end - buf;
assert(len < 16);
m128 chars = zeroes128();
memcpy(&chars, buf, len);
u32 mask = (0xffff >> (16 - len)) ^ 0xffff;
u32 z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
const u8 *rv = lastMatch(buf, z | mask);
if (rv) {
return rv;
}
return buf - 1;
}
const u8 *rtruffleExec(m128 shuf_mask_lo_highclear,
m128 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
assert(buf && buf_end);
assert(buf < buf_end);
const u8 *rv;
DEBUG_PRINTF("len %zu\n", buf_end - buf);
if (buf_end - buf < 16) {
return truffleRevMini(shuf_mask_lo_highclear, shuf_mask_lo_highset, buf,
buf_end);
}
assert(buf_end - buf >= 16);
// Preconditioning: most of the time our buffer won't be aligned.
m128 chars = loadu128(buf_end - 16);
rv = revBlock(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars,
buf_end - 16);
if (rv) {
return rv;
}
buf_end = (const u8 *)((size_t)buf_end & ~((size_t)0xf));
const u8 *last_block = buf + 16;
while (buf_end > last_block) {
buf_end -= 16;
m128 lchars = load128(buf_end);
rv = revBlock(shuf_mask_lo_highclear, shuf_mask_lo_highset, lchars,
buf_end);
if (rv) {
return rv;
}
}
// Use an unaligned load to mop up the last 16 bytes and get an accurate
// picture to buf_end.
chars = loadu128(buf);
rv = revBlock(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars, buf);
if (rv) {
return rv;
}
return buf - 1;
}
#elif !defined(HAVE_AVX512)
// AVX2
static really_inline
const u8 *lastMatch(const u8 *buf, u32 z) {
if (unlikely(z != 0xffffffff)) {
u32 pos = clz32(~z);
assert(pos < 32);
return buf + (31 - pos);
}
return NULL; // no match
}
static really_inline
const u8 *firstMatch(const u8 *buf, u32 z) {
if (unlikely(z != 0xffffffff)) {
u32 pos = ctz32(~z);
assert(pos < 32);
return buf + pos;
}
return NULL; // no match
}
static really_inline
u32 block(m256 shuf_mask_lo_highclear, m256 shuf_mask_lo_highset, m256 v) {
m256 highconst = set1_32x8(0x80);
m256 shuf_mask_hi = set1_4x64(0x8040201008040201);
// and now do the real work
m256 shuf1 = pshufb_m256(shuf_mask_lo_highclear, v);
m256 t1 = xor256(v, highconst);
m256 shuf2 = pshufb_m256(shuf_mask_lo_highset, t1);
m256 t2 = andnot256(highconst, rshift64_m256(v, 4));
m256 shuf3 = pshufb_m256(shuf_mask_hi, t2);
m256 tmp = and256(or256(shuf1, shuf2), shuf3);
m256 tmp2 = eq256(tmp, zeroes256());
u32 z = movemask256(tmp2);
return z;
}
static
const u8 *truffleMini(m256 shuf_mask_lo_highclear, m256 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
uintptr_t len = buf_end - buf;
assert(len < 32);
m256 chars = zeroes256();
memcpy(&chars, buf, len);
u32 z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
// can't be these bytes in z
u32 mask = (0xffffffff >> (32 - len)) ^ 0xffffffff;
const u8 *rv = firstMatch(buf, z | mask);
if (rv) {
return rv;
} else {
return buf_end;
}
}
static really_inline
const u8 *fwdBlock(m256 shuf_mask_lo_highclear, m256 shuf_mask_lo_highset,
m256 v, const u8 *buf) {
u32 z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, v);
return firstMatch(buf, z);
}
static really_inline
const u8 *revBlock(m256 shuf_mask_lo_highclear, m256 shuf_mask_lo_highset,
m256 v, const u8 *buf) {
u32 z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, v);
return lastMatch(buf, z);
}
const u8 *truffleExec(m128 shuf_mask_lo_highclear,
m128 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
DEBUG_PRINTF("len %zu\n", buf_end - buf);
const m256 wide_clear = set1_2x128(shuf_mask_lo_highclear);
const m256 wide_set = set1_2x128(shuf_mask_lo_highset);
assert(buf && buf_end);
assert(buf < buf_end);
const u8 *rv;
if (buf_end - buf < 32) {
return truffleMini(wide_clear, wide_set, buf, buf_end);
}
size_t min = (size_t)buf % 32;
assert(buf_end - buf >= 32);
// Preconditioning: most of the time our buffer won't be aligned.
m256 chars = loadu256(buf);
rv = fwdBlock(wide_clear, wide_set, chars, buf);
if (rv) {
return rv;
}
buf += (32 - min);
const u8 *last_block = buf_end - 32;
while (buf < last_block) {
m256 lchars = load256(buf);
rv = fwdBlock(wide_clear, wide_set, lchars, buf);
if (rv) {
return rv;
}
buf += 32;
}
// Use an unaligned load to mop up the last 32 bytes and get an accurate
// picture to buf_end.
assert(buf <= buf_end && buf >= buf_end - 32);
chars = loadu256(buf_end - 32);
rv = fwdBlock(wide_clear, wide_set, chars, buf_end - 32);
if (rv) {
return rv;
}
return buf_end;
}
static
const u8 *truffleRevMini(m256 shuf_mask_lo_highclear,
m256 shuf_mask_lo_highset, const u8 *buf,
const u8 *buf_end) {
uintptr_t len = buf_end - buf;
assert(len < 32);
m256 chars = zeroes256();
memcpy(&chars, buf, len);
u32 mask = (0xffffffff >> (32 - len)) ^ 0xffffffff;
u32 z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
const u8 *rv = lastMatch(buf, z | mask);
if (rv) {
return rv;
}
return buf - 1;
}
const u8 *rtruffleExec(m128 shuf_mask_lo_highclear,
m128 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
const m256 wide_clear = set1_2x128(shuf_mask_lo_highclear);
const m256 wide_set = set1_2x128(shuf_mask_lo_highset);
assert(buf && buf_end);
assert(buf < buf_end);
const u8 *rv;
DEBUG_PRINTF("len %zu\n", buf_end - buf);
if (buf_end - buf < 32) {
return truffleRevMini(wide_clear, wide_set, buf, buf_end);
}
assert(buf_end - buf >= 32);
// Preconditioning: most of the time our buffer won't be aligned.
m256 chars = loadu256(buf_end - 32);
rv = revBlock(wide_clear, wide_set, chars,
buf_end - 32);
if (rv) {
return rv;
}
buf_end = (const u8 *)((size_t)buf_end & ~((size_t)0x1f));
const u8 *last_block = buf + 32;
while (buf_end > last_block) {
buf_end -= 32;
m256 lchars = load256(buf_end);
rv = revBlock(wide_clear, wide_set, lchars, buf_end);
if (rv) {
return rv;
}
}
// Use an unaligned load to mop up the last 32 bytes and get an accurate
// picture to buf_end.
chars = loadu256(buf);
rv = revBlock(wide_clear, wide_set, chars, buf);
if (rv) {
return rv;
}
return buf - 1;
}
#else // AVX512
static really_inline
const u8 *lastMatch(const u8 *buf, u64a z) {
if (unlikely(z != ~0ULL)) {
u64a pos = clz64(~z);
assert(pos < 64);
return buf + (63 - pos);
}
return NULL; // no match
}
static really_inline
const u8 *firstMatch(const u8 *buf, u64a z) {
if (unlikely(z != ~0ULL)) {
u64a pos = ctz64(~z);
assert(pos < 64);
DEBUG_PRINTF("pos %llu\n", pos);
return buf + pos;
}
return NULL; // no match
}
static really_inline
u64a block(m512 shuf_mask_lo_highclear, m512 shuf_mask_lo_highset, m512 v) {
m512 highconst = set1_64x8(0x80);
m512 shuf_mask_hi = set1_8x64(0x8040201008040201);
// and now do the real work
m512 shuf1 = pshufb_m512(shuf_mask_lo_highclear, v);
m512 t1 = xor512(v, highconst);
m512 shuf2 = pshufb_m512(shuf_mask_lo_highset, t1);
m512 t2 = andnot512(highconst, rshift64_m512(v, 4));
m512 shuf3 = pshufb_m512(shuf_mask_hi, t2);
m512 tmp = and512(or512(shuf1, shuf2), shuf3);
u64a z = eq512mask(tmp, zeroes512());
return z;
}
static really_inline
const u8 *truffleMini(m512 shuf_mask_lo_highclear, m512 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
uintptr_t len = buf_end - buf;
assert(len <= 64);
__mmask64 mask = (~0ULL) >> (64 - len);
m512 chars = loadu_maskz_m512(mask, buf);
u64a z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
const u8 *rv = firstMatch(buf, z | ~mask);
return rv;
}
static really_inline
const u8 *fwdBlock(m512 shuf_mask_lo_highclear, m512 shuf_mask_lo_highset,
m512 v, const u8 *buf) {
u64a z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, v);
return firstMatch(buf, z);
}
static really_inline
const u8 *revBlock(m512 shuf_mask_lo_highclear, m512 shuf_mask_lo_highset,
m512 v, const u8 *buf) {
u64a z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, v);
return lastMatch(buf, z);
}
const u8 *truffleExec(m128 shuf_mask_lo_highclear, m128 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
DEBUG_PRINTF("len %zu\n", buf_end - buf);
const m512 wide_clear = set1_4x128(shuf_mask_lo_highclear);
const m512 wide_set = set1_4x128(shuf_mask_lo_highset);
assert(buf && buf_end);
assert(buf < buf_end);
const u8 *rv;
if (buf_end - buf <= 64) {
rv = truffleMini(wide_clear, wide_set, buf, buf_end);
return rv ? rv : buf_end;
}
assert(buf_end - buf >= 64);
if ((uintptr_t)buf % 64) {
// Preconditioning: most of the time our buffer won't be aligned.
rv = truffleMini(wide_clear, wide_set, buf, ROUNDUP_PTR(buf, 64));
if (rv) {
return rv;
}
buf = ROUNDUP_PTR(buf, 64);
}
const u8 *last_block = buf_end - 64;
while (buf < last_block) {
m512 lchars = load512(buf);
rv = fwdBlock(wide_clear, wide_set, lchars, buf);
if (rv) {
return rv;
}
buf += 64;
}
// Use an unaligned load to mop up the last 64 bytes and get an accurate
// picture to buf_end.
assert(buf <= buf_end && buf >= buf_end - 64);
m512 chars = loadu512(buf_end - 64);
rv = fwdBlock(wide_clear, wide_set, chars, buf_end - 64);
if (rv) {
return rv;
}
return buf_end;
}
static really_inline
const u8 *truffleRevMini(m512 shuf_mask_lo_highclear, m512 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
uintptr_t len = buf_end - buf;
assert(len < 64);
__mmask64 mask = (~0ULL) >> (64 - len);
m512 chars = loadu_maskz_m512(mask, buf);
u64a z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
DEBUG_PRINTF("mask 0x%016llx z 0x%016llx\n", mask, z);
const u8 *rv = lastMatch(buf, z | ~mask);
if (rv) {
return rv;
}
return buf - 1;
}
const u8 *rtruffleExec(m128 shuf_mask_lo_highclear, m128 shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end) {
const m512 wide_clear = set1_4x128(shuf_mask_lo_highclear);
const m512 wide_set = set1_4x128(shuf_mask_lo_highset);
assert(buf && buf_end);
assert(buf < buf_end);
const u8 *rv;
DEBUG_PRINTF("len %zu\n", buf_end - buf);
if (buf_end - buf < 64) {
return truffleRevMini(wide_clear, wide_set, buf, buf_end);
}
assert(buf_end - buf >= 64);
// Preconditioning: most of the time our buffer won't be aligned.
m512 chars = loadu512(buf_end - 64);
rv = revBlock(wide_clear, wide_set, chars, buf_end - 64);
if (rv) {
return rv;
}
buf_end = (const u8 *)ROUNDDOWN_N((uintptr_t)buf_end, 64);
const u8 *last_block = buf + 64;
while (buf_end > last_block) {
buf_end -= 64;
m512 lchars = load512(buf_end);
rv = revBlock(wide_clear, wide_set, lchars, buf_end);
if (rv) {
return rv;
}
}
// Use an unaligned load to mop up the last 64 bytes and get an accurate
// picture to buf_end.
chars = loadu512(buf);
rv = revBlock(wide_clear, wide_set, chars, buf);
if (rv) {
return rv;
}
return buf - 1;
}
#endif

54
src/nfa/truffle.cpp Normal file
View File

@ -0,0 +1,54 @@
/*
* Copyright (c) 2015-2017, Intel Corporation
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* * Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of Intel Corporation nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
/** \file
* \brief Shufti: character class acceleration.
*
* Utilises the SSSE3 pshufb shuffle instruction
*/
#include "truffle.h"
#include "ue2common.h"
#include "util/arch.h"
#include "util/bitutils.h"
#if !defined(HAVE_SVE)
#include "truffle_simd.hpp"
const u8 *truffleExec(m128 mask_lo, m128 mask_hi, const u8 *buf,
const u8 *buf_end) {
return truffleExecReal<VECTORSIZE>(mask_lo, mask_hi, buf, buf_end);
}
const u8 *rtruffleExec(m128 mask_lo, m128 mask_hi, const u8 *buf,
const u8 *buf_end) {
return rtruffleExecReal<VECTORSIZE>(mask_lo, mask_hi, buf, buf_end);
}
#endif

304
src/nfa/truffle_simd.hpp Normal file
View File

@ -0,0 +1,304 @@
/*
* Copyright (c) 2015-2017, Intel Corporation
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* * Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of Intel Corporation nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
/** \file
* \brief Shufti: character class acceleration.
*
* Utilises the SSSE3 pshufb shuffle instruction
*/
#include "truffle.h"
#include "ue2common.h"
#include "util/arch.h"
#include "util/bitutils.h"
#include "util/unaligned.h"
#include "util/simd/types.hpp"
template <uint16_t S>
const u8 *firstMatch(const u8 *buf, typename SuperVector<S>::movemask_type z);
template <uint16_t S>
const u8 *lastMatch(const u8 *buf, typename SuperVector<S>::movemask_type z);
template <>
really_inline
const u8 *firstMatch<16>(const u8 *buf, typename SuperVector<16>::movemask_type z) {
DEBUG_PRINTF("buf %p z %08x \n", buf, z);
DEBUG_PRINTF("z %08x\n", z);
if (unlikely(z != 0xffff)) {
u32 pos = ctz32(~z & 0xffff);
DEBUG_PRINTF("~z %08x\n", ~z);
DEBUG_PRINTF("match @ pos %u\n", pos);
assert(pos < 16);
return buf + pos;
} else {
return NULL; // no match
}
}
template <>
really_inline
const u8 *firstMatch<64>(const u8 *buf, typename SuperVector<64>::movemask_type z) {
DEBUG_PRINTF("z 0x%016llx\n", z);
if (unlikely(z != ~0ULL)) {
u32 pos = ctz64(~z);
DEBUG_PRINTF("match @ pos %u\n", pos);
assert(pos < 64);
return buf + pos;
} else {
return NULL; // no match
}
}
template <>
really_inline
const u8 *lastMatch<16>(const u8 *buf, typename SuperVector<16>::movemask_type z) {
DEBUG_PRINTF("buf %p z %08x \n", buf, z);
DEBUG_PRINTF("z %08x\n", z);
if (unlikely(z != 0xffff)) {
u32 pos = clz32(~z & 0xffff);
DEBUG_PRINTF("~z %08x\n", ~z);
DEBUG_PRINTF("match @ pos %u\n", pos);
assert(pos >= 16 && pos < 32);
return buf + (31 - pos);
} else {
return NULL; // no match
}
}
template <>
really_inline
const u8 *lastMatch<64>(const u8 *buf, typename SuperVector<64>::movemask_type z) {
DEBUG_PRINTF("z 0x%016llx\n", z);
if (unlikely(z != ~0ULL)) {
u32 pos = clz64(~z);
DEBUG_PRINTF("match @ pos %u\n", pos);
assert(pos < 64);
return buf + pos;
} else {
return NULL; // no match
}
}
template <uint16_t S>
static really_inline
typename SuperVector<S>::movemask_type block(SuperVector<S> shuf_mask_lo_highclear, SuperVector<S> shuf_mask_lo_highset,
SuperVector<S> v){
SuperVector<S> highconst = SuperVector<S>::set1_16x8(0x80);
print_m128_16x8("highconst", highconst.u.v128[0]);
SuperVector<S> shuf_mask_hi = SuperVector<S>::set1_2x64(0x8040201008040201);
print_m128_2x64("shuf_mask_hi", shuf_mask_hi.u.v128[0]);
SuperVector<S> shuf1 = shuf_mask_lo_highclear.pshufb(v);
SuperVector<S> t1 = v ^ highconst;
SuperVector<S> shuf2 = shuf_mask_lo_highset.pshufb(t1);
SuperVector<S> t2 = highconst.opandnot(v.rshift64(4));
SuperVector<S> shuf3 = shuf_mask_hi.pshufb(t2);
SuperVector<S> tmp = shuf3 & (shuf1 | shuf2);
return tmp.eqmask(SuperVector<S>::Zeroes());
}
template <uint16_t S>
static really_inline const u8 *truffleMini(SuperVector<S> shuf_mask_lo_highclear, SuperVector<S> shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end){
uintptr_t len = buf_end - buf;
assert(len < 16);
SuperVector<S> chars = SuperVector<S>::loadu(buf);
u32 mask = (0xffff >> (16 - len)) ^ 0xffff;
typename SuperVector<S>::movemask_type z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
const u8 *rv = firstMatch<S>(buf, z | mask);
if (rv) {
return rv;
} else {
return buf_end;
}
}
template <uint16_t S>
static really_inline
const u8 *fwdBlock(SuperVector<S> shuf_mask_lo_highclear, SuperVector<S> shuf_mask_lo_highset, SuperVector<S> v,
const u8 *buf) {
typename SuperVector<S>::movemask_type z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, v);
DEBUG_PRINTF("z %08x\n", z);
return firstMatch<S>(buf, z);
}
template <uint16_t S>
const u8 *truffleExecReal(SuperVector<S> shuf_mask_lo_highclear, SuperVector<S> shuf_mask_lo_highset, const u8 *buf, const u8 *buf_end) {
assert(buf && buf_end);
assert(buf < buf_end);
DEBUG_PRINTF("truffle %p len %zu\n", buf, buf_end - buf);
DEBUG_PRINTF("b %s\n", buf);
const SuperVector<S> wide_shuf_mask_lo_highclear(shuf_mask_lo_highclear);
const SuperVector<S> wide_shuf_mask_lo_highset(shuf_mask_lo_highset);
const u8 *d = buf;
const u8 *rv;
DEBUG_PRINTF("start %p end %p \n", d, buf_end);
assert(d < buf_end);
if (d + S <= buf_end) {
// peel off first part to cacheline boundary
const u8 *d1 = ROUNDUP_PTR(d, S);
DEBUG_PRINTF("until aligned %p \n", d1);
if (d1 != d) {
rv = truffleMini(shuf_mask_lo_highclear, shuf_mask_lo_highset, d, d1);
if (rv != d1) {
return rv;
}
d = d1;
}
size_t loops = (buf_end - d) / S;
DEBUG_PRINTF("loops %ld \n", loops);
for (size_t i = 0; i < loops; i++, d+= S) {
DEBUG_PRINTF("d %p \n", d);
const u8 *base = ROUNDUP_PTR(d, S);
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(base + 256);
SuperVector<S> chars = SuperVector<S>::load(d);
rv = fwdBlock(wide_shuf_mask_lo_highclear, wide_shuf_mask_lo_highset, chars, d);
if (rv) return rv;
}
}
DEBUG_PRINTF("d %p e %p \n", d, buf_end);
// finish off tail
rv = buf_end;
if (d != buf_end) {
rv = truffleMini(shuf_mask_lo_highclear, shuf_mask_lo_highset, d, buf_end);
DEBUG_PRINTF("rv %p \n", rv);
}
return rv;
}
template <uint16_t S>
static really_inline const u8 *truffleRevMini(SuperVector<S> shuf_mask_lo_highclear, SuperVector<S> shuf_mask_lo_highset,
const u8 *buf, const u8 *buf_end){
uintptr_t len = buf_end - buf;
assert(len < 16);
SuperVector<S> chars = SuperVector<S>::loadu(buf);
u32 mask = (0xffff >> (16 - len)) ^ 0xffff;
typename SuperVector<S>::movemask_type z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
const u8 *rv = lastMatch<S>(buf,z | mask);
if (rv) {
return rv;
}
return buf - 1;
}
template <uint16_t S>
static really_inline
const u8 *revBlock(SuperVector<S> shuf_mask_lo_highclear, SuperVector<S> shuf_mask_lo_highset, SuperVector<S> v,
const u8 *buf) {
typename SuperVector<S>::movemask_type z = block(shuf_mask_lo_highclear, shuf_mask_lo_highset, v);
DEBUG_PRINTF("z %08x\n", z);
return lastMatch<S>(buf, z);
}
template <uint16_t S>
const u8 *rtruffleExecReal(SuperVector<S> shuf_mask_lo_highclear, SuperVector<S> shuf_mask_lo_highset, const u8 *buf, const u8 *buf_end){
assert(buf && buf_end);
assert(buf < buf_end);
DEBUG_PRINTF("trufle %p len %zu\n", buf, buf_end - buf);
DEBUG_PRINTF("b %s\n", buf);
const SuperVector<S> wide_shuf_mask_lo_highclear(shuf_mask_lo_highclear);
const SuperVector<S> wide_shuf_mask_lo_highset(shuf_mask_lo_highset);
const u8 *d = buf_end;
const u8 *rv;
DEBUG_PRINTF("start %p end %p \n", buf, d);
assert(d > buf);
if (d - S >= buf) {
// peel off first part to cacheline boundary
const u8 *d1 = ROUNDDOWN_PTR(d, S);
DEBUG_PRINTF("until aligned %p \n", d1);
if (d1 != d) {
rv = truffleRevMini(shuf_mask_lo_highclear, shuf_mask_lo_highset, d1, d);
if (rv != d1 - 1) return rv;
d = d1;
}
while (d - S >= buf) {
d -= S;
DEBUG_PRINTF("d %p \n", d);
const u8 *base = ROUNDDOWN_PTR(buf, S);
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(base + 256);
SuperVector<S> chars = SuperVector<S>::load(d);
rv = revBlock(wide_shuf_mask_lo_highclear, wide_shuf_mask_lo_highset, chars, d);
if (rv) return rv;
}
}
DEBUG_PRINTF("d %p e %p \n", buf, d);
// finish off tail
if (d != buf) {
rv = truffleRevMini(shuf_mask_lo_highclear, shuf_mask_lo_highset, buf, d);
DEBUG_PRINTF("rv %p \n", rv);
if (rv != d - 1) return rv;
}
return buf;
}

13
src/util/simd/arch/x86/impl.cpp
View File

@ -164,18 +164,31 @@ really_inline SuperVector<16> SuperVector<16>::operator|(SuperVector<16> const b
return {_mm_or_si128(u.v128[0], b.u.v128[0])}; return {_mm_or_si128(u.v128[0], b.u.v128[0])};
} }
template <>
really_inline SuperVector<16> SuperVector<16>::operator^(SuperVector<16> const b) const
{
return {_mm_xor_si128(u.v128[0], b.u.v128[0])};
}
template <> template <>
really_inline SuperVector<16> SuperVector<16>::opand(SuperVector<16> const b) const really_inline SuperVector<16> SuperVector<16>::opand(SuperVector<16> const b) const
{ {
return *this & b; return *this & b;
} }
template <>
really_inline SuperVector<16> SuperVector<16>::opxor(SuperVector<16> const b) const
{
return *this ^ b;
}
template <> template <>
really_inline SuperVector<16> SuperVector<16>::opandnot(SuperVector<16> const b) const really_inline SuperVector<16> SuperVector<16>::opandnot(SuperVector<16> const b) const
{ {
return {_mm_andnot_si128(u.v128[0], b.u.v128[0])}; return {_mm_andnot_si128(u.v128[0], b.u.v128[0])};
} }
template <> template <>
really_inline SuperVector<16> SuperVector<16>::eq(SuperVector<16> const b) const really_inline SuperVector<16> SuperVector<16>::eq(SuperVector<16> const b) const
{ {

4
src/util/simd/types.hpp
View File

@ -175,10 +175,14 @@ public:
SuperVector operator&(SuperVector const b) const; SuperVector operator&(SuperVector const b) const;
SuperVector operator|(SuperVector const b) const; SuperVector operator|(SuperVector const b) const;
SuperVector operator^(SuperVector const b) const;
SuperVector opand(SuperVector const b) const; SuperVector opand(SuperVector const b) const;
SuperVector opor(SuperVector const b) const; SuperVector opor(SuperVector const b) const;
SuperVector opandnot(SuperVector const b) const; SuperVector opandnot(SuperVector const b) const;
SuperVector opxor(SuperVector const b) const;
SuperVector eq(SuperVector const b) const; SuperVector eq(SuperVector const b) const;
SuperVector operator<<(uint8_t const N) const; SuperVector operator<<(uint8_t const N) const;

37
unit/internal/supervector.cpp
View File

@ -117,6 +117,43 @@ TEST(SuperVectorUtilsTest,OR128c){
} }
} }
TEST(SuperVectorUtilsTest,XOR128c){
srand (time(NULL));
u8 vec[16];
for (int i=0; i<16; i++) {
vec[i] = rand() % 100 + 1;
}
u8 vec2[16];
for (int i=0; i<16; i++) {
vec2[i] = rand() % 100 + 1;
}
auto SP1 = SuperVector<16>::loadu(vec);
auto SP2 = SuperVector<16>::loadu(vec2);
auto SPResult = SP1 ^ SP2;
for (int i=0; i<16; i++) {
ASSERT_EQ(SPResult.u.u8[i],vec[i] ^ vec2[i]);
}
}
TEST(SuperVectorUtilsTest,OPXOR128c){
srand (time(NULL));
u8 vec[16];
for (int i=0; i<16; i++) {
vec[i] = rand() % 100 + 1;
}
u8 vec2[16];
for (int i=0; i<16; i++) {
vec2[i] = rand() % 100 + 1;
}
auto SP1 = SuperVector<16>::loadu(vec);
auto SP2 = SuperVector<16>::loadu(vec2);
auto SPResult = SP1.opxor(SP2);
for (int i=0; i<16; i++) {
ASSERT_EQ(SPResult.u.u8[i],vec[i] ^ vec2[i]);
}
}
TEST(SuperVectorUtilsTest,OPANDNOT128c){ TEST(SuperVectorUtilsTest,OPANDNOT128c){
auto SP1 = SuperVector<16>::Zeroes(); auto SP1 = SuperVector<16>::Zeroes();
auto SP2 = SuperVector<16>::Ones(); auto SP2 = SuperVector<16>::Ones();