mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2024-12-27 03:14:05 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity
75339a5528
opencv/3rdparty/zlib-ng/arch/power/crc32_power8.c
Letu Ren 0de26fd78e Add zlib-ng as an alternative zlib implementation 
Zlib-ng is zlib replacement with optimizations for "next generation" systems. Its optimization may benifits image library decode and encode speed such as libpng. In our tests, if using zlib-ng and libpng combination on a x86_64 machine with AVX2, the time of `imdecode` amd `imencode` will drop 20% approximately. This patch enables zlib-ng's optimization if `CV_DISABLE_OPTIMIZATION` is OFF. Since Zlib-ng can dispatch intrinsics on the fly, port work is much easier.

Related discussion: https://github.com/opencv/opencv/issues/22573
2024-01-14 14:58:47 +08:00

590 lines
24 KiB
C
Raw Blame History

/* crc32 for POWER8 using VSX instructions
* Copyright (C) 2021 IBM Corporation
*
* Author: Rogerio Alves <rogealve@br.ibm.com>
*
* For conditions of distribution and use, see copyright notice in zlib.h
*
* Calculate the checksum of data that is 16 byte aligned and a multiple of
* 16 bytes.
*
* The first step is to reduce it to 1024 bits. We do this in 8 parallel
* chunks in order to mask the latency of the vpmsum instructions. If we
* have more than 32 kB of data to checksum we repeat this step multiple
* times, passing in the previous 1024 bits.
*
* The next step is to reduce the 1024 bits to 64 bits. This step adds
* 32 bits of 0s to the end - this matches what a CRC does. We just
* calculate constants that land the data in this 32 bits.
*
* We then use fixed point Barrett reduction to compute a mod n over GF(2)
* for n = CRC using POWER8 instructions. We use x = 32.
*
* http://en.wikipedia.org/wiki/Barrett_reduction
*
* This code uses gcc vector builtins instead using assembly directly.
*/
#include <altivec.h>
#include "zendian.h"
#include "zbuild.h"
#include "crc32_constants.h"
#include "crc32_braid_tbl.h"
#if defined (__clang__)
#include "fallback_builtins.h"
#endif
#define MAX_SIZE 32768
#define VMX_ALIGN 16
#define VMX_ALIGN_MASK (VMX_ALIGN-1)
static unsigned int crc32_align(unsigned int crc, const unsigned char *p, unsigned long len) {
while (len--)
crc = crc_table[(crc ^ *p++) & 0xff] ^ (crc >> 8);
return crc;
}
static unsigned int ALIGNED_(32) __crc32_vpmsum(unsigned int crc, const void* p, unsigned long len);
Z_INTERNAL uint32_t crc32_power8(uint32_t crc, const unsigned char *p, size_t _len) {
unsigned int prealign;
unsigned int tail;
unsigned long len = (unsigned long) _len;
if (p == (const unsigned char *) 0x0)
return 0;
crc ^= 0xffffffff;
if (len < VMX_ALIGN + VMX_ALIGN_MASK) {
crc = crc32_align(crc, p, len);
goto out;
}
if ((unsigned long)p & VMX_ALIGN_MASK) {
prealign = VMX_ALIGN - ((unsigned long)p & VMX_ALIGN_MASK);
crc = crc32_align(crc, p, prealign);
len -= prealign;
p += prealign;
}
crc = __crc32_vpmsum(crc, p, len & ~VMX_ALIGN_MASK);
tail = len & VMX_ALIGN_MASK;
if (tail) {
p += len & ~VMX_ALIGN_MASK;
crc = crc32_align(crc, p, tail);
}
out:
crc ^= 0xffffffff;
return crc;
}
/* When we have a load-store in a single-dispatch group and address overlap
* such that forward is not allowed (load-hit-store) the group must be flushed.
* A group ending NOP prevents the flush.
*/
#define GROUP_ENDING_NOP __asm__("ori 2,2,0" ::: "memory")
#if BYTE_ORDER == BIG_ENDIAN
#define BYTESWAP_DATA
#endif
#ifdef BYTESWAP_DATA
#define VEC_PERM(vr, va, vb, vc) vr = vec_perm(va, vb, (__vector unsigned char) vc)
#if BYTE_ORDER == LITTLE_ENDIAN
/* Byte reverse permute constant LE. */
static const __vector unsigned long long vperm_const ALIGNED_(16) = { 0x08090A0B0C0D0E0FUL, 0x0001020304050607UL };
#else
static const __vector unsigned long long vperm_const ALIGNED_(16) = { 0x0F0E0D0C0B0A0908UL, 0X0706050403020100UL };
#endif
#else
#define VEC_PERM(vr, va, vb, vc)
#endif
static unsigned int ALIGNED_(32) __crc32_vpmsum(unsigned int crc, const void* p, unsigned long len) {
const __vector unsigned long long vzero = {0,0};
const __vector unsigned long long vones = {0xffffffffffffffffUL, 0xffffffffffffffffUL};
const __vector unsigned long long vmask_32bit =
(__vector unsigned long long)vec_sld((__vector unsigned char)vzero, (__vector unsigned char)vones, 4);
const __vector unsigned long long vmask_64bit =
(__vector unsigned long long)vec_sld((__vector unsigned char)vzero, (__vector unsigned char)vones, 8);
__vector unsigned long long vcrc;
__vector unsigned long long vconst1, vconst2;
/* vdata0-vdata7 will contain our data (p). */
__vector unsigned long long vdata0, vdata1, vdata2, vdata3, vdata4, vdata5, vdata6, vdata7;
/* v0-v7 will contain our checksums */
__vector unsigned long long v0 = {0,0};
__vector unsigned long long v1 = {0,0};
__vector unsigned long long v2 = {0,0};
__vector unsigned long long v3 = {0,0};
__vector unsigned long long v4 = {0,0};
__vector unsigned long long v5 = {0,0};
__vector unsigned long long v6 = {0,0};
__vector unsigned long long v7 = {0,0};
/* Vector auxiliary variables. */
__vector unsigned long long va0, va1, va2, va3, va4, va5, va6, va7;
unsigned int offset; /* Constant table offset. */
unsigned long i; /* Counter. */
unsigned long chunks;
unsigned long block_size;
int next_block = 0;
/* Align by 128 bits. The last 128 bit block will be processed at end. */
unsigned long length = len & 0xFFFFFFFFFFFFFF80UL;
vcrc = (__vector unsigned long long)__builtin_pack_vector_int128(0UL, crc);
/* Short version. */
if (len < 256) {
/* Calculate where in the constant table we need to start. */
offset = 256 - len;
vconst1 = vec_ld(offset, vcrc_short_const);
vdata0 = vec_ld(0, (__vector unsigned long long*) p);
VEC_PERM(vdata0, vdata0, vconst1, vperm_const);
/* xor initial value */
vdata0 = vec_xor(vdata0, vcrc);
vdata0 = (__vector unsigned long long) __builtin_crypto_vpmsumw(
(__vector unsigned int)vdata0, (__vector unsigned int)vconst1);
v0 = vec_xor(v0, vdata0);
for (i = 16; i < len; i += 16) {
vconst1 = vec_ld(offset + i, vcrc_short_const);
vdata0 = vec_ld(i, (__vector unsigned long long*) p);
VEC_PERM(vdata0, vdata0, vconst1, vperm_const);
vdata0 = (__vector unsigned long long) __builtin_crypto_vpmsumw(
(__vector unsigned int)vdata0, (__vector unsigned int)vconst1);
v0 = vec_xor(v0, vdata0);
}
} else {
/* Load initial values. */
vdata0 = vec_ld(0, (__vector unsigned long long*) p);
vdata1 = vec_ld(16, (__vector unsigned long long*) p);
VEC_PERM(vdata0, vdata0, vdata0, vperm_const);
VEC_PERM(vdata1, vdata1, vdata1, vperm_const);
vdata2 = vec_ld(32, (__vector unsigned long long*) p);
vdata3 = vec_ld(48, (__vector unsigned long long*) p);
VEC_PERM(vdata2, vdata2, vdata2, vperm_const);
VEC_PERM(vdata3, vdata3, vdata3, vperm_const);
vdata4 = vec_ld(64, (__vector unsigned long long*) p);
vdata5 = vec_ld(80, (__vector unsigned long long*) p);
VEC_PERM(vdata4, vdata4, vdata4, vperm_const);
VEC_PERM(vdata5, vdata5, vdata5, vperm_const);
vdata6 = vec_ld(96, (__vector unsigned long long*) p);
vdata7 = vec_ld(112, (__vector unsigned long long*) p);
VEC_PERM(vdata6, vdata6, vdata6, vperm_const);
VEC_PERM(vdata7, vdata7, vdata7, vperm_const);
/* xor in initial value */
vdata0 = vec_xor(vdata0, vcrc);
p = (char *)p + 128;
do {
/* Checksum in blocks of MAX_SIZE. */
block_size = length;
if (block_size > MAX_SIZE) {
block_size = MAX_SIZE;
}
length = length - block_size;
/*
* Work out the offset into the constants table to start at. Each
* constant is 16 bytes, and it is used against 128 bytes of input
* data - 128 / 16 = 8
*/
offset = (MAX_SIZE/8) - (block_size/8);
/* We reduce our final 128 bytes in a separate step */
chunks = (block_size/128)-1;
vconst1 = vec_ld(offset, vcrc_const);
va0 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata0,
(__vector unsigned long long)vconst1);
va1 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata1,
(__vector unsigned long long)vconst1);
va2 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata2,
(__vector unsigned long long)vconst1);
va3 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata3,
(__vector unsigned long long)vconst1);
va4 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata4,
(__vector unsigned long long)vconst1);
va5 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata5,
(__vector unsigned long long)vconst1);
va6 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata6,
(__vector unsigned long long)vconst1);
va7 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata7,
(__vector unsigned long long)vconst1);
if (chunks > 1) {
offset += 16;
vconst2 = vec_ld(offset, vcrc_const);
GROUP_ENDING_NOP;
vdata0 = vec_ld(0, (__vector unsigned long long*) p);
VEC_PERM(vdata0, vdata0, vdata0, vperm_const);
vdata1 = vec_ld(16, (__vector unsigned long long*) p);
VEC_PERM(vdata1, vdata1, vdata1, vperm_const);
vdata2 = vec_ld(32, (__vector unsigned long long*) p);
VEC_PERM(vdata2, vdata2, vdata2, vperm_const);
vdata3 = vec_ld(48, (__vector unsigned long long*) p);
VEC_PERM(vdata3, vdata3, vdata3, vperm_const);
vdata4 = vec_ld(64, (__vector unsigned long long*) p);
VEC_PERM(vdata4, vdata4, vdata4, vperm_const);
vdata5 = vec_ld(80, (__vector unsigned long long*) p);
VEC_PERM(vdata5, vdata5, vdata5, vperm_const);
vdata6 = vec_ld(96, (__vector unsigned long long*) p);
VEC_PERM(vdata6, vdata6, vdata6, vperm_const);
vdata7 = vec_ld(112, (__vector unsigned long long*) p);
VEC_PERM(vdata7, vdata7, vdata7, vperm_const);
p = (char *)p + 128;
/*
* main loop. Each iteration calculates the CRC for a 128-byte
* block.
*/
for (i = 0; i < chunks-2; i++) {
vconst1 = vec_ld(offset, vcrc_const);
offset += 16;
GROUP_ENDING_NOP;
v0 = vec_xor(v0, va0);
va0 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata0,
(__vector unsigned long long)vconst2);
vdata0 = vec_ld(0, (__vector unsigned long long*) p);
VEC_PERM(vdata0, vdata0, vdata0, vperm_const);
GROUP_ENDING_NOP;
v1 = vec_xor(v1, va1);
va1 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata1,
(__vector unsigned long long)vconst2);
vdata1 = vec_ld(16, (__vector unsigned long long*) p);
VEC_PERM(vdata1, vdata1, vdata1, vperm_const);
GROUP_ENDING_NOP;
v2 = vec_xor(v2, va2);
va2 = __builtin_crypto_vpmsumd((__vector unsigned long long)
vdata2, (__vector unsigned long long)vconst2);
vdata2 = vec_ld(32, (__vector unsigned long long*) p);
VEC_PERM(vdata2, vdata2, vdata2, vperm_const);
GROUP_ENDING_NOP;
v3 = vec_xor(v3, va3);
va3 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata3,
(__vector unsigned long long)vconst2);
vdata3 = vec_ld(48, (__vector unsigned long long*) p);
VEC_PERM(vdata3, vdata3, vdata3, vperm_const);
vconst2 = vec_ld(offset, vcrc_const);
GROUP_ENDING_NOP;
v4 = vec_xor(v4, va4);
va4 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata4,
(__vector unsigned long long)vconst1);
vdata4 = vec_ld(64, (__vector unsigned long long*) p);
VEC_PERM(vdata4, vdata4, vdata4, vperm_const);
GROUP_ENDING_NOP;
v5 = vec_xor(v5, va5);
va5 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata5,
(__vector unsigned long long)vconst1);
vdata5 = vec_ld(80, (__vector unsigned long long*) p);
VEC_PERM(vdata5, vdata5, vdata5, vperm_const);
GROUP_ENDING_NOP;
v6 = vec_xor(v6, va6);
va6 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata6,
(__vector unsigned long long)vconst1);
vdata6 = vec_ld(96, (__vector unsigned long long*) p);
VEC_PERM(vdata6, vdata6, vdata6, vperm_const);
GROUP_ENDING_NOP;
v7 = vec_xor(v7, va7);
va7 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata7,
(__vector unsigned long long)vconst1);
vdata7 = vec_ld(112, (__vector unsigned long long*) p);
VEC_PERM(vdata7, vdata7, vdata7, vperm_const);
p = (char *)p + 128;
}
/* First cool down */
vconst1 = vec_ld(offset, vcrc_const);
offset += 16;
v0 = vec_xor(v0, va0);
va0 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata0,
(__vector unsigned long long)vconst1);
GROUP_ENDING_NOP;
v1 = vec_xor(v1, va1);
va1 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata1,
(__vector unsigned long long)vconst1);
GROUP_ENDING_NOP;
v2 = vec_xor(v2, va2);
va2 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata2,
(__vector unsigned long long)vconst1);
GROUP_ENDING_NOP;
v3 = vec_xor(v3, va3);
va3 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata3,
(__vector unsigned long long)vconst1);
GROUP_ENDING_NOP;
v4 = vec_xor(v4, va4);
va4 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata4,
(__vector unsigned long long)vconst1);
GROUP_ENDING_NOP;
v5 = vec_xor(v5, va5);
va5 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata5,
(__vector unsigned long long)vconst1);
GROUP_ENDING_NOP;
v6 = vec_xor(v6, va6);
va6 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata6,
(__vector unsigned long long)vconst1);
GROUP_ENDING_NOP;
v7 = vec_xor(v7, va7);
va7 = __builtin_crypto_vpmsumd((__vector unsigned long long)vdata7,
(__vector unsigned long long)vconst1);
}/* else */
/* Second cool down. */
v0 = vec_xor(v0, va0);
v1 = vec_xor(v1, va1);
v2 = vec_xor(v2, va2);
v3 = vec_xor(v3, va3);
v4 = vec_xor(v4, va4);
v5 = vec_xor(v5, va5);
v6 = vec_xor(v6, va6);
v7 = vec_xor(v7, va7);
/*
* vpmsumd produces a 96 bit result in the least significant bits
* of the register. Since we are bit reflected we have to shift it
* left 32 bits so it occupies the least significant bits in the
* bit reflected domain.
*/
v0 = (__vector unsigned long long)vec_sld((__vector unsigned char)v0,
(__vector unsigned char)vzero, 4);
v1 = (__vector unsigned long long)vec_sld((__vector unsigned char)v1,
(__vector unsigned char)vzero, 4);
v2 = (__vector unsigned long long)vec_sld((__vector unsigned char)v2,
(__vector unsigned char)vzero, 4);
v3 = (__vector unsigned long long)vec_sld((__vector unsigned char)v3,
(__vector unsigned char)vzero, 4);
v4 = (__vector unsigned long long)vec_sld((__vector unsigned char)v4,
(__vector unsigned char)vzero, 4);
v5 = (__vector unsigned long long)vec_sld((__vector unsigned char)v5,
(__vector unsigned char)vzero, 4);
v6 = (__vector unsigned long long)vec_sld((__vector unsigned char)v6,
(__vector unsigned char)vzero, 4);
v7 = (__vector unsigned long long)vec_sld((__vector unsigned char)v7,
(__vector unsigned char)vzero, 4);
/* xor with the last 1024 bits. */
va0 = vec_ld(0, (__vector unsigned long long*) p);
VEC_PERM(va0, va0, va0, vperm_const);
va1 = vec_ld(16, (__vector unsigned long long*) p);
VEC_PERM(va1, va1, va1, vperm_const);
va2 = vec_ld(32, (__vector unsigned long long*) p);
VEC_PERM(va2, va2, va2, vperm_const);
va3 = vec_ld(48, (__vector unsigned long long*) p);
VEC_PERM(va3, va3, va3, vperm_const);
va4 = vec_ld(64, (__vector unsigned long long*) p);
VEC_PERM(va4, va4, va4, vperm_const);
va5 = vec_ld(80, (__vector unsigned long long*) p);
VEC_PERM(va5, va5, va5, vperm_const);
va6 = vec_ld(96, (__vector unsigned long long*) p);
VEC_PERM(va6, va6, va6, vperm_const);
va7 = vec_ld(112, (__vector unsigned long long*) p);
VEC_PERM(va7, va7, va7, vperm_const);
p = (char *)p + 128;
vdata0 = vec_xor(v0, va0);
vdata1 = vec_xor(v1, va1);
vdata2 = vec_xor(v2, va2);
vdata3 = vec_xor(v3, va3);
vdata4 = vec_xor(v4, va4);
vdata5 = vec_xor(v5, va5);
vdata6 = vec_xor(v6, va6);
vdata7 = vec_xor(v7, va7);
/* Check if we have more blocks to process */
next_block = 0;
if (length != 0) {
next_block = 1;
/* zero v0-v7 */
v0 = vec_xor(v0, v0);
v1 = vec_xor(v1, v1);
v2 = vec_xor(v2, v2);
v3 = vec_xor(v3, v3);
v4 = vec_xor(v4, v4);
v5 = vec_xor(v5, v5);
v6 = vec_xor(v6, v6);
v7 = vec_xor(v7, v7);
}
length = length + 128;
} while (next_block);
/* Calculate how many bytes we have left. */
length = (len & 127);
/* Calculate where in (short) constant table we need to start. */
offset = 128 - length;
v0 = vec_ld(offset, vcrc_short_const);
v1 = vec_ld(offset + 16, vcrc_short_const);
v2 = vec_ld(offset + 32, vcrc_short_const);
v3 = vec_ld(offset + 48, vcrc_short_const);
v4 = vec_ld(offset + 64, vcrc_short_const);
v5 = vec_ld(offset + 80, vcrc_short_const);
v6 = vec_ld(offset + 96, vcrc_short_const);
v7 = vec_ld(offset + 112, vcrc_short_const);
offset += 128;
v0 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata0, (__vector unsigned int)v0);
v1 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata1, (__vector unsigned int)v1);
v2 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata2, (__vector unsigned int)v2);
v3 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata3, (__vector unsigned int)v3);
v4 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata4, (__vector unsigned int)v4);
v5 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata5, (__vector unsigned int)v5);
v6 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata6, (__vector unsigned int)v6);
v7 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata7, (__vector unsigned int)v7);
/* Now reduce the tail (0-112 bytes). */
for (i = 0; i < length; i+=16) {
vdata0 = vec_ld(i,(__vector unsigned long long*)p);
VEC_PERM(vdata0, vdata0, vdata0, vperm_const);
va0 = vec_ld(offset + i,vcrc_short_const);
va0 = (__vector unsigned long long)__builtin_crypto_vpmsumw(
(__vector unsigned int)vdata0, (__vector unsigned int)va0);
v0 = vec_xor(v0, va0);
}
/* xor all parallel chunks together. */
v0 = vec_xor(v0, v1);
v2 = vec_xor(v2, v3);
v4 = vec_xor(v4, v5);
v6 = vec_xor(v6, v7);
v0 = vec_xor(v0, v2);
v4 = vec_xor(v4, v6);
v0 = vec_xor(v0, v4);
}
/* Barrett Reduction */
vconst1 = vec_ld(0, v_Barrett_const);
vconst2 = vec_ld(16, v_Barrett_const);
v1 = (__vector unsigned long long)vec_sld((__vector unsigned char)v0,
(__vector unsigned char)v0, 8);
v0 = vec_xor(v1,v0);
/* shift left one bit */
__vector unsigned char vsht_splat = vec_splat_u8 (1);
v0 = (__vector unsigned long long)vec_sll((__vector unsigned char)v0, vsht_splat);
v0 = vec_and(v0, vmask_64bit);
/*
* The reflected version of Barrett reduction. Instead of bit
* reflecting our data (which is expensive to do), we bit reflect our
* constants and our algorithm, which means the intermediate data in
* our vector registers goes from 0-63 instead of 63-0. We can reflect
* the algorithm because we don't carry in mod 2 arithmetic.
*/
/* bottom 32 bits of a */
v1 = vec_and(v0, vmask_32bit);
/* ma */
v1 = __builtin_crypto_vpmsumd((__vector unsigned long long)v1,
(__vector unsigned long long)vconst1);
/* bottom 32bits of ma */
v1 = vec_and(v1, vmask_32bit);
/* qn */
v1 = __builtin_crypto_vpmsumd((__vector unsigned long long)v1,
(__vector unsigned long long)vconst2);
/* a - qn, subtraction is xor in GF(2) */
v0 = vec_xor (v0, v1);
/*
* Since we are bit reflected, the result (ie the low 32 bits) is in
* the high 32 bits. We just need to shift it left 4 bytes
* V0 [ 0 1 X 3 ]
* V0 [ 0 X 2 3 ]
*/
/* shift result into top 64 bits of */
v0 = (__vector unsigned long long)vec_sld((__vector unsigned char)v0,
(__vector unsigned char)vzero, 4);
#if BYTE_ORDER == BIG_ENDIAN
return v0[0];
#else
return v0[1];
#endif
}
Reference in New Issue View Git Blame Copy Permalink