Files
vectorscan/src/nfa/arm/truffle.hpp
Yoan Picchi 7054378c93 Speed up truffle with 256b TBL instructions
256b wide SVE vectors allow some simplification of truffle.
Up to 40% speedup on graviton3. Going from 12500 MB/s to 17000 MB/s
onhe microbenchmark.
SVE2 also offer this capability for 128b vector with a speedup around
25% compared to normal SVE

Add unit tests and benchmark for this wide variant

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>
2024-05-22 16:13:53 +00:00

262 lines
13 KiB
C++

/*
* Copyright (c) 2015-2017, Intel Corporation
* Copyright (c) 2020-2021, VectorCamp PC
* Copyright (c) 2023, Arm Limited
*
* 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 Truffle: character class acceleration.
*
*/
#ifdef HAVE_SVE
#ifdef HAVE_SVE2
/*
* blockSingleMask takes in a character set (as masks) and a string and return for each character
* of the string wether or not it is part of the set.
*
* 'shuf_mask_lo_highclear' and 'shuf_mask_lo_highset' are 128-bit masks where each bit
* represents whether or not a character is in the character set. The 'highclear' and
* 'highset' in the name refers to the MSb of the byte of the character (allowing two
* 128-bit masks to cover all 256 values).
*
* The mask is an array of 32 bytes and is encoded this way:
* Let C be a character in the set. The bit describing that character is at byte[C%32] and
* within that byte, it's at bit[C/32]
* As example, 'a' = 0x61, so the resulting mask will be: 0x00 0x08 0x00 0x00 0x00 ...
*
* Assume the mask is in one of those configurations:
* - both masks are exactly 128b wide
* - the first mask is exactly 256b wide and the second is zeroed.
* - the first mask is more than 256b wide, with bits past the 256th being zero, and the second mask is zeroed.
*/
static really_inline
svuint8_t blockSingleMaskWideSVE2(svuint8_t shuf_mask_lo_highclear, svuint8_t shuf_mask_lo_highset, svuint8_t chars) {
const svuint8_t pshub_mask = svdup_u8(0x1f);
const svuint8_t unique_bit_per_lane_mask = svreinterpret_u8(svdup_u64(0x8040201008040201));
svuint8x2_t shuf_mask_32 = svcreate2(shuf_mask_lo_highclear, shuf_mask_lo_highset);
/*
* svtbl2 does a table lookup. Each byte in the second argument indexes into the array of bytes
* in shuf_mask_32 and saves the result in the corresponding byte of byte_select.
* We mask the chars so that we are using the low nibble of char as the index.
*/
svuint8_t byte_select = svtbl2(shuf_mask_32, svand_x(svptrue_b8(), chars, pshub_mask));
/*
* We now have selected the byte that contain the bit corresponding to the char. We need to
* further filter it, otherwise we'd get a match for any character % 32 to a searched character
*
* The low nibble was used previously to select the byte out of the mask. The high nibble is
* used to select the bit out of the byte. So we shift everything right by 5.
*
* Using svtbl, we can make an array where each element is a different bit. Using the high
* nibble we can get a mask selecting only the bit out of a byte that may have the relevant
* charset char.
*/
svuint8_t char_high_nibble = svlsr_x(svptrue_b8(), chars, 5);
svuint8_t bit_select = svtbl(unique_bit_per_lane_mask, char_high_nibble);
/*
* We apply the bit_select mask onto the selected byte. What is left is the bit in the charset
* encoding the character in char. A non zero value means the char was in the charset
*
* The _x suffix only works if we process a full char vector. If we were to use a partial
* vector, then _z and a mask would be required on this svand only. Otherwise, the disabled
* lanes may have arbitrary values
*/
return svand_x(svptrue_b8(), byte_select, bit_select);
}
#endif //HAVE_SVE2
/*
* blockSingleMask takes in a character set (as masks) and a string and return for each character
* of the string wether or not it is part of the set.
*
* 'shuf_mask_lo_highclear' and 'shuf_mask_lo_highset' are 128-bit masks where each bit
* represents whether or not a character is in the character set. The 'highclear' and
* 'highset' in the name refers to the MSb of the byte of the character (allowing two
* 128-bit masks to cover all 256 values).
*
* The masks are arrays of 16 bytes each and are encoded this way:
* Let C be a character in the set. The bit describing that character is at byte[C%16] and
* within that byte, it's at bit[C/16]
* As example, 'a' = 0x61, so the resulting mask will be: 0x00 0x40 0x00 0x00 0x00 ...
*
* Assume both mask are 128b wide. If they are larger, the additional bits must be zero
*/
static really_inline
svuint8_t blockSingleMaskSVE(svuint8_t shuf_mask_lo_highclear, svuint8_t shuf_mask_lo_highset, svuint8_t chars) {
const svuint8_t highconst = svdup_u8(0x80);
const svuint8_t pshub_mask = svdup_u8(0x8f);
const svuint8_t unique_bit_per_lane_mask = svreinterpret_u8(svdup_u64(0x8040201008040201));
/*
* svtbl does a table lookup. Each byte in the second argument indexes into the array of bytes
* in shuf_mask_lo_highclear and saves the result in the corresponding byte of byte_select_low.
* We mask the chars so that we are using the low nibble of char as the index but we keep the
* MSb so that high characters (not represented by the highclear mask) become an index out of
* bounds and result in a 0.
*/
svuint8_t byte_select_low = svtbl(shuf_mask_lo_highclear, svand_x(svptrue_b8(), chars, pshub_mask));
/*
* We flip the MSb of the chars and do the same table lookup with the highset mask.
* This way it's the characters with MSb cleared that will result in out of bands indexes.
* This allows us to cover the full range (0-127 and 128-255)
*/
svuint8_t char_high_flipped = sveor_x(svptrue_b8(), chars, highconst);
svuint8_t byte_select_high = svtbl(shuf_mask_lo_highset, svand_x(svptrue_b8(), char_high_flipped, pshub_mask));
/*
* We now have selected the byte that contain the bit corresponding to the char. We need to
* further filter it, otherwise we'd get a match for any character % 16 to a searched character
*
* The low nibble was used previously to select the byte out of the mask. The high nibble is
* used to select the bit out of the byte. So we shift everything right by 4.
*
* Using svtbl, we can make an array where each element is a different bit. Using the high
* nibble we can get a mask selecting only the bit out of a byte that may have the relevant
* charset char.
*/
svuint8_t char_high_nibble = svlsr_x(svptrue_b8(), chars, 4);
svuint8_t bit_select = svtbl(unique_bit_per_lane_mask, char_high_nibble);
/*
* For every lane, only one of the byte selected may have a value, so we can OR them. We
* then apply the bit_select mask. What is left is the bit in the charset encoding the
* character in char. A non zero value means the char was in the charset
*
* The _x suffix only works if we process a full char vector. If we were to use a partial
* vector, then _z and a mask would be required on this svand only. Otherwise, the disabled
* lanes may have arbitrary values
*/
return svand_x(svptrue_b8(), svorr_x(svptrue_b8(), byte_select_low, byte_select_high), bit_select);
}
/*
* blockSingleMask takes in a character set (as masks) and a string and return for each character
* of the string wether or not it is part of the set.
*
* 'shuf_mask_32' is a 256-bit masks where each bit represents whether or not a character is in
* the character set.
*
* The mask is an array of 32 bytes and is encoded this way:
* Let C be a character in the set. The bit describing that character is at byte[C%32] and
* within that byte, it's at bit[C/32]
* As example, 'a' = 0x61, so the resulting mask will be: 0x00 0x08 0x00 0x00 0x00 ...
*
* Assume both mask are 128b wide. If they are larger, the additional bits must be zero
*/
static really_inline
svuint8_t blockSingleMaskWideSVE(svuint8_t shuf_mask_32, svuint8_t chars) {//TODO I might have issues with the type
const svuint8_t pshub_mask = svdup_u8(0x1f);
const svuint8_t unique_bit_per_lane_mask = svreinterpret_u8(svdup_u64(0x8040201008040201));
/*
* svtbl does a table lookup. Each byte in the second argument indexes into the array of bytes
* in shuf_mask_32 and saves the result in the corresponding byte of byte_select.
* We mask the chars so that we are using the low nibble of char as the index.
*/
svuint8_t byte_select = svtbl(shuf_mask_32, svand_x(svptrue_b8(), chars, pshub_mask));
/*
* We now have selected the byte that contain the bit corresponding to the char. We need to
* further filter it, otherwise we'd get a match for any character % 32 to a searched character
*
* The low nibble was used previously to select the byte out of the mask. The high nibble is
* used to select the bit out of the byte. So we shift everything right by 5.
*
* Using svtbl, we can make an array where each element is a different bit. Using the high
* nibble we can get a mask selecting only the bit out of a byte that may have the relevant
* charset char.
*/
svuint8_t char_high_nibble = svlsr_x(svptrue_b8(), chars, 5);
svuint8_t bit_select = svtbl(unique_bit_per_lane_mask, char_high_nibble);
/*
* We apply the bit_select mask onto the selected byte. What is left is the bit in the charset
* encoding the character in char. A non zero value means the char was in the charset
*
* The _x suffix only works if we process a full char vector. If we were to use a partial
* vector, then _z and a mask would be required on this svand only. Otherwise, the disabled
* lanes may have arbitrary values
*/
return svand_x(svptrue_b8(), byte_select, bit_select);
}
/* require normal truffle compilation. The 256b mask is split between the two parameters */
static really_inline
svuint8_t blockSingleMask(svuint8_t shuf_mask_lo_highclear, svuint8_t shuf_mask_lo_highset, svuint8_t chars) {
return blockSingleMaskSVE(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
}
/* require wide truffle compilation. The 256b mask is fully contained in the first parameter */
static really_inline
svuint8_t blockSingleMaskWide32(svuint8_t shuf_mask_32, svuint8_t chars) {
return blockSingleMaskWideSVE(shuf_mask_32, chars);
}
#ifdef HAVE_SVE2
/* require wide truffle compilation. The 256b mask is split between the two parameters if the vector is 128b,
* or fully contained in the first parameter is it's 256b and more*/
static really_inline
svuint8_t blockSingleMaskWide(svuint8_t shuf_mask_lo_highclear, svuint8_t shuf_mask_lo_highset, svuint8_t chars) {
return blockSingleMaskWideSVE2(shuf_mask_lo_highclear, shuf_mask_lo_highset, chars);
}
#endif //HAVE_SVE2
#endif //HAVE_SVE
/* require normal truffle compilation. The 256b mask is split between the two parameters */
template <uint16_t S>
static really_inline
const SuperVector<S> blockSingleMask(SuperVector<S> shuf_mask_lo_highclear, SuperVector<S> shuf_mask_lo_highset, SuperVector<S> chars) {
chars.print8("chars");
shuf_mask_lo_highclear.print8("shuf_mask_lo_highclear");
shuf_mask_lo_highset.print8("shuf_mask_lo_highset");
SuperVector<S> highconst = SuperVector<S>::dup_u8(0x80);
highconst.print8("highconst");
SuperVector<S> shuf_mask_hi = SuperVector<S>::dup_u64(0x8040201008040201);
shuf_mask_hi.print8("shuf_mask_hi");
SuperVector<S> shuf1 = shuf_mask_lo_highclear.pshufb(chars);
shuf1.print8("shuf1");
SuperVector<S> t1 = chars ^ highconst;
t1.print8("t1");
SuperVector<S> shuf2 = shuf_mask_lo_highset.pshufb(t1);
shuf2.print8("shuf2");
SuperVector<S> t2 = highconst.opandnot(chars.template vshr_64_imm<4>());
t2.print8("t2");
SuperVector<S> shuf3 = shuf_mask_hi.pshufb(t2);
shuf3.print8("shuf3");
SuperVector<S> res = (shuf1 | shuf2) & shuf3;
res.print8("(shuf1 | shuf2) & shuf3");
return !res.eq(SuperVector<S>::Zeroes());
}