points.h

Go to the documentation of this file.

/*
    This file is part of darktable,
    Copyright (C) 2010-2011 johannes hanika.
    Copyright (C) 2012 Jean-Sébastien Pédron.
    Copyright (C) 2012 Richard Wonka.
    Copyright (C) 2013-2014, 2016 Tobias Ellinghaus.
    Copyright (C) 2014-2016 Roman Lebedev.
    Copyright (C) 2017 luzpaz.
    Copyright (C) 2018 Peter Budai.
    Copyright (C) 2018 Robert Bridge.
    Copyright (C) 2019 Andreas Schneider.
    Copyright (C) 2019, 2025-2026 Aurélien PIERRE.
    Copyright (C) 2020 Heiko Bauke.
    Copyright (C) 2020 Pascal Obry.
    Copyright (C) 2022 Martin Bařinka.
    Copyright (C) 2024 Alynx Zhou.
    
    darktable is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation, either version 3 of the License, or
    (at your option) any later version.
    
    darktable is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License for more details.
    
    You should have received a copy of the GNU General Public License
    along with darktable.  If not, see <http://www.gnu.org/licenses/>.
*/
 
#ifndef DT_COMMON_POINTS_H
#define DT_COMMON_POINTS_H
 
#include "system/macros.h"
#include "system/mem_alloc.h"
#include "common/logging.h"
#include "system/openmp.h"   // dt_get_thread_num()
 
/* Process-wide singleton with no per-call context to ride on: this accessor is the
 * intended end state (same category as dt_conf_*), implemented by the orchestrator. */
struct dt_points_t *dt_points_get_global(void);
 
#ifndef __SSE2__
 
#if !defined _XOPEN_SOURCE && !defined(__DragonFly__) && !defined(__FreeBSD__) && !defined(__NetBSD__)       \
    && !defined(__OpenBSD__) && !defined(_WIN32)
#define _XOPEN_SOURCE
#endif
 
#include <stdlib.h>
 
// xorshift128+, period 2^128-1, apparently passes all TestU01 suite tests.
typedef struct dt_points_state_t
{
  uint64_t state0;
  uint64_t state1;
} dt_points_state_t;
 
typedef struct dt_points_t
{
  dt_points_state_t *s;
  unsigned int num;                 // 0 when initialisation failed; see dt_points_init()
} dt_points_t;
 
static inline void dt_points_init(dt_points_t *p, const unsigned int num_threads)
{
  p->s = (dt_points_state_t *)malloc(sizeof(dt_points_state_t) * num_threads);
  p->num = num_threads;
 
  if(IS_NULL_PTR(p->s))
  {
    dt_print(DT_DEBUG_ALWAYS,
             "[points] out of memory allocating the RNG state for %u thread(s)\n", num_threads);
    p->num = 0;
    return;
  }
 
  for(int k = 0; k < (int)num_threads; k++)
  {
    p->s[k].state0 = 1 + k;
    p->s[k].state1 = 2 + k;
  }
}
 
static inline void dt_points_cleanup(dt_points_t *p)
{
  dt_free(p->s);
  p->s = NULL;
  p->num = 0;
}
 
static inline float dt_points_get_for(dt_points_t *p, const unsigned int thread_num)
{
  /* dt_points_init() logs and leaves p->s NULL when it cannot allocate the state, which
   * only happens if the system is genuinely out of memory. Degrade to a fixed value
   * rather than dereferencing NULL: callers use this for dithering and noise, where
   * losing randomness costs a little image quality, whereas crashing costs the session.
   * The failure is already reported once, at init, instead of per pixel. */
  if(IS_NULL_PTR(p) || IS_NULL_PTR(p->s) || thread_num >= p->num) return 0.5f;
 
  uint64_t s1 = p->s[thread_num].state0;
  uint64_t s0 = p->s[thread_num].state1;
  p->s[thread_num].state0 = s0;
  s1 ^= s1 << 23;
  s1 ^= s1 >> 17;
  s1 ^= s0;
  s1 ^= s0 >> 26;
  p->s[thread_num].state1 = s1;
  // return (state0 + state1) / ((double)((uint64_t)-1) + 1.0);
  union {
      float f;
      uint32_t u;
  } v;
  v.u = 0x3f800000 |
      ((p->s[thread_num].state0 + p->s[thread_num].state1) >> 41); // faster than double version.
  return v.f - 1.0f;
}
 
static inline float dt_points_get()
{
  return dt_points_get_for(dt_points_get_global(), dt_get_thread_num());
}
 
#else
 
#include <inttypes.h>
 
#define MEXP 19937
 
#ifndef SFMT_PARAMS_H
#define SFMT_PARAMS_H
 
#if !defined(MEXP)
#ifdef __GNUC__
#warning "MEXP is not defined. I assume MEXP is 19937."
#endif
#define MEXP 19937
#endif
/*-----------------
  BASIC DEFINITIONS
  -----------------*/
#define N (MEXP / 128 + 1)
#define N32 (N * 4)
#define N64 (N * 2)
 
/*----------------------
  the parameters of SFMT
  following definitions are in paramsXXXX.h file.
  ----------------------*/
#if 0
#if MEXP == 607
#include "SFMT-params607.h"
#elif MEXP == 1279
#include "SFMT-params1279.h"
#elif MEXP == 2281
#include "SFMT-params2281.h"
#elif MEXP == 4253
#include "SFMT-params4253.h"
#elif MEXP == 11213
#include "SFMT-params11213.h"
#elif MEXP == 19937
#include "SFMT-params19937.h"
#elif MEXP == 44497
#include "SFMT-params44497.h"
#elif MEXP == 86243
#include "SFMT-params86243.h"
#elif MEXP == 132049
#include "SFMT-params132049.h"
#elif MEXP == 216091
#include "SFMT-params216091.h"
#else
#ifdef __GNUC__
#error "MEXP is not valid."
#undef MEXP
#else
#undef MEXP
#endif
#endif
 
#endif
 
#endif /* SFMT_PARAMS_H */
 
#ifndef SFMT_PARAMS19937_H
#define SFMT_PARAMS19937_H
 
#define POS1 122
#define SL1 18
#define SL2 1
#define SR1 11
#define SR2 1
#define MSK1 0xdfffffefU
#define MSK2 0xddfecb7fU
#define MSK3 0xbffaffffU
#define MSK4 0xbffffff6U
#define PARITY1 0x00000001U
#define PARITY2 0x00000000U
#define PARITY3 0x00000000U
#define PARITY4 0x13c9e684U
 
 
#define ALTI_SL1                                                                                             \
  {                                                                                                          \
    SL1, SL1, SL1, SL1                                                                                       \
  }
#define ALTI_SR1                                                                                             \
  {                                                                                                          \
    SR1, SR1, SR1, SR1                                                                                       \
  }
#define ALTI_MSK                                                                                             \
  {                                                                                                          \
    MSK1, MSK2, MSK3, MSK4                                                                                   \
  }
#define ALTI_MSK64                                                                                           \
  {                                                                                                          \
    MSK2, MSK1, MSK4, MSK3                                                                                   \
  }
#define ALTI_SL2_PERM                                                                                        \
  {                                                                                                          \
    1, 2, 3, 23, 5, 6, 7, 0, 9, 10, 11, 4, 13, 14, 15, 8                                                     \
  }
#define ALTI_SL2_PERM64                                                                                      \
  {                                                                                                          \
    1, 2, 3, 4, 5, 6, 7, 31, 9, 10, 11, 12, 13, 14, 15, 0                                                    \
  }
#define ALTI_SR2_PERM                                                                                        \
  {                                                                                                          \
    7, 0, 1, 2, 11, 4, 5, 6, 15, 8, 9, 10, 17, 12, 13, 14                                                    \
  }
#define ALTI_SR2_PERM64                                                                                      \
  {                                                                                                          \
    15, 0, 1, 2, 3, 4, 5, 6, 17, 8, 9, 10, 11, 12, 13, 14                                                    \
  }
#define IDSTR "SFMT-19937:122-18-1-11-1:dfffffef-ddfecb7f-bffaffff-bffffff6"
 
#endif /* SFMT_PARAMS19937_H */
 
typedef union w128_t
{
  __m128i si;
  uint32_t u[4];
} w128_t;
 
typedef struct sfmt_state_t
{
  w128_t sfmt[N];
  uint32_t *psfmt32;
#if !defined(BIG_ENDIAN64) || defined(ONLY64)
  uint64_t *psfmt64;
#endif
  int idx;
  int initialized;
  uint32_t parity[4];
} sfmt_state_t;
 
#ifndef SFMT_H
#define SFMT_H
 
#include <stdio.h>
 
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)
#include <inttypes.h>
#elif defined(_MSC_VER) || defined(__BORLANDC__)
typedef unsigned int uint32_t;
typedef unsigned __int64 uint64_t;
#define inline __inline
#else
#include <inttypes.h>
#if defined(__GNUC__)
#define inline __inline__
#endif
#endif
 
#ifndef PRIu64
#if defined(_MSC_VER) || defined(__BORLANDC__)
#define PRIu64 "I64u"
#define PRIx64 "I64x"
#else
#define PRIu64 "llu"
#define PRIx64 "llx"
#endif
#endif
 
#if defined(__GNUC__)
#define ALWAYSINLINE __attribute__((always_inline))
#else
#define ALWAYSINLINE
#endif
 
#if defined(_MSC_VER)
#if _MSC_VER >= 1200
#define PRE_ALWAYS __forceinline
#else
#define PRE_ALWAYS inline
#endif
#else
#define PRE_ALWAYS inline
#endif
 
static inline uint32_t gen_rand32(struct sfmt_state_t *s);
static inline uint64_t gen_rand64(struct sfmt_state_t *s);
static inline void init_gen_rand(struct sfmt_state_t *s, uint32_t seed) __attribute__((unused));
static inline void init_by_array(struct sfmt_state_t *s, uint32_t *init_key, int key_length)
    __attribute__((unused));
static inline const char *get_idstring(void) __attribute__((unused));
inline static float to_real2f(uint32_t v)
{
  union {
      float f;
      uint32_t u;
  } x;
  x.u = 0x3f800000 | (v >> 9); // faster than double version.
  return x.f - 1.0f;
  /* divided by 2^32 */
}
inline static float genrand_real2f(struct sfmt_state_t *s)
{
  return to_real2f(gen_rand32(s));
}
 
#endif
#ifndef SFMT_SSE2_H
#define SFMT_SSE2_H
 
PRE_ALWAYS static __m128i mm_recursion(__m128i *a, __m128i *b, __m128i c, __m128i d,
                                       __m128i mask) ALWAYSINLINE;
 
PRE_ALWAYS static __m128i mm_recursion(__m128i *a, __m128i *b, __m128i c, __m128i d, __m128i mask)
{
  __m128i v, x, y, z;
 
  x = _mm_load_si128(a);
  y = _mm_srli_epi32(*b, SR1);
  z = _mm_srli_si128(c, SR2);
  v = _mm_slli_epi32(d, SL1);
  z = _mm_xor_si128(z, x);
  z = _mm_xor_si128(z, v);
  x = _mm_slli_si128(x, SL2);
  y = _mm_and_si128(y, mask);
  z = _mm_xor_si128(z, x);
  z = _mm_xor_si128(z, y);
  return z;
}
 
inline static void gen_rand_all(struct sfmt_state_t *s)
{
  int i;
  __m128i r, r1, r2, mask;
  mask = _mm_set_epi32(MSK4, MSK3, MSK2, MSK1);
 
  r1 = _mm_load_si128(&(s->sfmt[N - 2].si));
  r2 = _mm_load_si128(&(s->sfmt[N - 1].si));
  for(i = 0; i < N - POS1; i++)
  {
    r = mm_recursion(&(s->sfmt[i].si), &(s->sfmt[i + POS1].si), r1, r2, mask);
    _mm_store_si128(&(s->sfmt[i].si), r);
    r1 = r2;
    r2 = r;
  }
  for(; i < N; i++)
  {
    r = mm_recursion(&(s->sfmt[i].si), &(s->sfmt[i + POS1 - N].si), r1, r2, mask);
    _mm_store_si128(&(s->sfmt[i].si), r);
    r1 = r2;
    r2 = r;
  }
}
 
inline static void gen_rand_array(struct sfmt_state_t *s, w128_t *array, int size)
{
  int i, j;
  __m128i r, r1, r2, mask;
  mask = _mm_set_epi32(MSK4, MSK3, MSK2, MSK1);
 
  r1 = _mm_load_si128(&(s->sfmt[N - 2].si));
  r2 = _mm_load_si128(&(s->sfmt[N - 1].si));
  for(i = 0; i < N - POS1; i++)
  {
    r = mm_recursion(&(s->sfmt[i].si), &(s->sfmt[i + POS1].si), r1, r2, mask);
    _mm_store_si128(&array[i].si, r);
    r1 = r2;
    r2 = r;
  }
  for(; i < N; i++)
  {
    r = mm_recursion(&(s->sfmt[i].si), &array[i + POS1 - N].si, r1, r2, mask);
    _mm_store_si128(&array[i].si, r);
    r1 = r2;
    r2 = r;
  }
  /* main loop */
  for(; i < size - N; i++)
  {
    r = mm_recursion(&array[i - N].si, &array[i + POS1 - N].si, r1, r2, mask);
    _mm_store_si128(&array[i].si, r);
    r1 = r2;
    r2 = r;
  }
  for(j = 0; j < 2 * N - size; j++)
  {
    r = _mm_load_si128(&array[j + size - N].si);
    _mm_store_si128(&(s->sfmt[j].si), r);
  }
  for(; i < size; i++)
  {
    r = mm_recursion(&array[i - N].si, &array[i + POS1 - N].si, r1, r2, mask);
    _mm_store_si128(&array[i].si, r);
    _mm_store_si128(&(s->sfmt[j++].si), r);
    r1 = r2;
    r2 = r;
  }
}
 
#endif
#include <assert.h>
#include <string.h>
//#include "SFMT.h"
//#include "SFMT-params.h"
 
#if defined(__BIG_ENDIAN__) && !defined(__amd64) && !defined(BIG_ENDIAN64)
#define BIG_ENDIAN64 1
#endif
#if defined(HAVE_ALTIVEC) && !defined(BIG_ENDIAN64)
#define BIG_ENDIAN64 1
#endif
#if defined(ONLY64) && !defined(BIG_ENDIAN64)
#if defined(__GNUC__)
#error "-DONLY64 must be specified with -DBIG_ENDIAN64"
#endif
#undef ONLY64
#endif
 
 
typedef struct dt_points_t
{
  sfmt_state_t **s;
  unsigned int num;
} dt_points_t;
 
#if 0
/*--------------------------------------
  FILE GLOBAL VARIABLES
  internal state, index counter and flag
  --------------------------------------*/
static w128_t sfmt[N];
static uint32_t *psfmt32 = &sfmt[0].u[0];
#if !defined(BIG_ENDIAN64) || defined(ONLY64)
static uint64_t *psfmt64 = (uint64_t *)&sfmt[0].u[0];
#endif
static int idx;
static int initialized = 0;
static uint32_t parity[4] = {PARITY1, PARITY2, PARITY3, PARITY4};
#endif
 
/*----------------
  STATIC FUNCTIONS
  ----------------*/
inline static int idxof(int i);
inline static void rshift128(w128_t *out, w128_t const *in, int shift);
inline static void lshift128(w128_t *out, w128_t const *in, int shift);
inline static void gen_rand_all(sfmt_state_t *s);
inline static void gen_rand_array(sfmt_state_t *s, w128_t *array, int size);
inline static uint32_t func1(uint32_t x);
inline static uint32_t func2(uint32_t x);
static void period_certification(sfmt_state_t *s);
#if defined(BIG_ENDIAN64) && !defined(ONLY64)
inline static void swap(w128_t *array, int size);
#endif
 
/*#if defined(HAVE_ALTIVEC)
#include "SFMT-alti.h"
#elif defined(HAVE_SSE2)
#include "SFMT-sse2.h"
#endif*/
 
#ifdef ONLY64
inline static int idxof(int i)
{
  return i ^ 1;
}
#else
inline static int idxof(int i)
{
  return i;
}
#endif
#ifdef ONLY64
inline static void rshift128(w128_t *out, w128_t const *in, int shift)
{
  uint64_t th, tl, oh, ol;
 
  th = ((uint64_t)in->u[2] << 32) | ((uint64_t)in->u[3]);
  tl = ((uint64_t)in->u[0] << 32) | ((uint64_t)in->u[1]);
 
  oh = th >> (shift * 8);
  ol = tl >> (shift * 8);
  ol |= th << (64 - shift * 8);
  out->u[0] = (uint32_t)(ol >> 32);
  out->u[1] = (uint32_t)ol;
  out->u[2] = (uint32_t)(oh >> 32);
  out->u[3] = (uint32_t)oh;
}
#else
inline static void rshift128(w128_t *out, w128_t const *in, int shift)
{
  uint64_t th, tl, oh, ol;
 
  th = ((uint64_t)in->u[3] << 32) | ((uint64_t)in->u[2]);
  tl = ((uint64_t)in->u[1] << 32) | ((uint64_t)in->u[0]);
 
  oh = th >> (shift * 8);
  ol = tl >> (shift * 8);
  ol |= th << (64 - shift * 8);
  out->u[1] = (uint32_t)(ol >> 32);
  out->u[0] = (uint32_t)ol;
  out->u[3] = (uint32_t)(oh >> 32);
  out->u[2] = (uint32_t)oh;
}
#endif
#ifdef ONLY64
inline static void lshift128(w128_t *out, w128_t const *in, int shift)
{
  uint64_t th, tl, oh, ol;
 
  th = ((uint64_t)in->u[2] << 32) | ((uint64_t)in->u[3]);
  tl = ((uint64_t)in->u[0] << 32) | ((uint64_t)in->u[1]);
 
  oh = th << (shift * 8);
  ol = tl << (shift * 8);
  oh |= tl >> (64 - shift * 8);
  out->u[0] = (uint32_t)(ol >> 32);
  out->u[1] = (uint32_t)ol;
  out->u[2] = (uint32_t)(oh >> 32);
  out->u[3] = (uint32_t)oh;
}
#else
inline static void lshift128(w128_t *out, w128_t const *in, int shift)
{
  uint64_t th, tl, oh, ol;
 
  th = ((uint64_t)in->u[3] << 32) | ((uint64_t)in->u[2]);
  tl = ((uint64_t)in->u[1] << 32) | ((uint64_t)in->u[0]);
 
  oh = th << (shift * 8);
  ol = tl << (shift * 8);
  oh |= tl >> (64 - shift * 8);
  out->u[1] = (uint32_t)(ol >> 32);
  out->u[0] = (uint32_t)ol;
  out->u[3] = (uint32_t)(oh >> 32);
  out->u[2] = (uint32_t)oh;
}
#endif
 
#if(!defined(HAVE_ALTIVEC)) && (!defined(HAVE_SSE2))
#ifdef ONLY64
inline static void do_recursion(w128_t *r, w128_t *a, w128_t *b, w128_t *c, w128_t *d)
{
  w128_t x;
  w128_t y;
 
  lshift128(&x, a, SL2);
  rshift128(&y, c, SR2);
  r->u[0] = a->u[0] ^ x.u[0] ^ ((b->u[0] >> SR1) & MSK2) ^ y.u[0] ^ (d->u[0] << SL1);
  r->u[1] = a->u[1] ^ x.u[1] ^ ((b->u[1] >> SR1) & MSK1) ^ y.u[1] ^ (d->u[1] << SL1);
  r->u[2] = a->u[2] ^ x.u[2] ^ ((b->u[2] >> SR1) & MSK4) ^ y.u[2] ^ (d->u[2] << SL1);
  r->u[3] = a->u[3] ^ x.u[3] ^ ((b->u[3] >> SR1) & MSK3) ^ y.u[3] ^ (d->u[3] << SL1);
}
#else
inline static void do_recursion(w128_t *r, w128_t *a, w128_t *b, w128_t *c, w128_t *d)
{
  w128_t x;
  w128_t y;
 
  lshift128(&x, a, SL2);
  rshift128(&y, c, SR2);
  r->u[0] = a->u[0] ^ x.u[0] ^ ((b->u[0] >> SR1) & MSK1) ^ y.u[0] ^ (d->u[0] << SL1);
  r->u[1] = a->u[1] ^ x.u[1] ^ ((b->u[1] >> SR1) & MSK2) ^ y.u[1] ^ (d->u[1] << SL1);
  r->u[2] = a->u[2] ^ x.u[2] ^ ((b->u[2] >> SR1) & MSK3) ^ y.u[2] ^ (d->u[2] << SL1);
  r->u[3] = a->u[3] ^ x.u[3] ^ ((b->u[3] >> SR1) & MSK4) ^ y.u[3] ^ (d->u[3] << SL1);
}
#endif
#endif
 
#if defined(BIG_ENDIAN64) && !defined(ONLY64) && !defined(HAVE_ALTIVEC)
inline static void swap(w128_t *array, int size)
{
  for(int i = 0; i < size; i++)
  {
    uint32_t x = array[i].u[0];
    uint32_t y = array[i].u[2];
    array[i].u[0] = array[i].u[1];
    array[i].u[2] = array[i].u[3];
    array[i].u[1] = x;
    array[i].u[3] = y;
  }
}
#endif
static uint32_t func1(uint32_t x)
{
  return (x ^ (x >> 27)) * (uint32_t)1664525UL;
}
 
static uint32_t func2(uint32_t x)
{
  return (x ^ (x >> 27)) * (uint32_t)1566083941UL;
}
 
static void period_certification(sfmt_state_t *s)
{
  int inner = 0;
 
  for(int i = 0; i < 4; i++) inner ^= s->psfmt32[idxof(i)] & s->parity[i];
  for(int i = 16; i > 0; i >>= 1) inner ^= inner >> i;
  inner &= 1;
  /* check OK */
  if(inner == 1)
  {
    return;
  }
  /* check NG, and modification */
  for(int i = 0; i < 4; i++)
  {
    uint32_t work = 1;
    for(int j = 0; j < 32; j++)
    {
      if((work & s->parity[i]) != 0)
      {
        s->psfmt32[idxof(i)] ^= work;
        return;
      }
      work = work << 1;
    }
  }
}
 
/*----------------
  PUBLIC FUNCTIONS
  ----------------*/
const char *get_idstring(void)
{
  return IDSTR;
}
 
#ifndef ONLY64
uint32_t gen_rand32(sfmt_state_t *s)
{
  uint32_t r;
 
  // assert(s->initialized);
  if(s->idx >= N32)
  {
    gen_rand_all(s);
    s->idx = 0;
  }
  r = s->psfmt32[s->idx++];
  return r;
}
#endif
uint64_t gen_rand64(sfmt_state_t *s)
{
#if defined(BIG_ENDIAN64) && !defined(ONLY64)
  uint32_t r1, r2;
#else
  uint64_t r;
#endif
 
  // assert(s->initialized);
  // assert(s->idx % 2 == 0);
 
  if(s->idx >= N32)
  {
    gen_rand_all(s);
    s->idx = 0;
  }
#if defined(BIG_ENDIAN64) && !defined(ONLY64)
  r1 = s->psfmt32[s->idx];
  r2 = s->psfmt32[s->idx + 1];
  s->idx += 2;
  return ((uint64_t)r2 << 32) | r1;
#else
  r = s->psfmt64[s->idx / 2];
  s->idx += 2;
  return r;
#endif
}
 
 
void init_gen_rand(sfmt_state_t *s, uint32_t seed)
{
  int i;
 
  s->psfmt32[idxof(0)] = seed;
  for(i = 1; i < N32; i++)
  {
    s->psfmt32[idxof(i)] = 1812433253UL * (s->psfmt32[idxof(i - 1)] ^ (s->psfmt32[idxof(i - 1)] >> 30)) + i;
  }
  s->idx = N32;
  period_certification(s);
  s->initialized = 1;
}
 
void init_by_array(sfmt_state_t *s, uint32_t *init_key, int key_length)
{
  int i, j, count;
  uint32_t r;
  int lag;
  int mid;
  int size = N * 4;
 
  if(size >= 623)
  {
    lag = 11;
  }
  else if(size >= 68)
  {
    lag = 7;
  }
  else if(size >= 39)
  {
    lag = 5;
  }
  else
  {
    lag = 3;
  }
  mid = (size - lag) / 2;
 
  memset(s->sfmt, 0x8b, sizeof(s->sfmt));
  if(key_length + 1 > N32)
  {
    count = key_length + 1;
  }
  else
  {
    count = N32;
  }
  r = func1(s->psfmt32[idxof(0)] ^ s->psfmt32[idxof(mid)] ^ s->psfmt32[idxof(N32 - 1)]);
  s->psfmt32[idxof(mid)] += r;
  r += key_length;
  s->psfmt32[idxof(mid + lag)] += r;
  s->psfmt32[idxof(0)] = r;
 
  count--;
  for(i = 1, j = 0; (j < count) && (j < key_length); j++)
  {
    r = func1(s->psfmt32[idxof(i)] ^ s->psfmt32[idxof((i + mid) % N32)]
              ^ s->psfmt32[idxof((i + N32 - 1) % N32)]);
    s->psfmt32[idxof((i + mid) % N32)] += r;
    r += init_key[j] + i;
    s->psfmt32[idxof((i + mid + lag) % N32)] += r;
    s->psfmt32[idxof(i)] = r;
    i = (i + 1) % N32;
  }
  for(; j < count; j++)
  {
    r = func1(s->psfmt32[idxof(i)] ^ s->psfmt32[idxof((i + mid) % N32)]
              ^ s->psfmt32[idxof((i + N32 - 1) % N32)]);
    s->psfmt32[idxof((i + mid) % N32)] += r;
    r += i;
    s->psfmt32[idxof((i + mid + lag) % N32)] += r;
    s->psfmt32[idxof(i)] = r;
    i = (i + 1) % N32;
  }
  for(j = 0; j < N32; j++)
  {
    r = func2(s->psfmt32[idxof(i)] + s->psfmt32[idxof((i + mid) % N32)]
              + s->psfmt32[idxof((i + N32 - 1) % N32)]);
    s->psfmt32[idxof((i + mid) % N32)] ^= r;
    r -= i;
    s->psfmt32[idxof((i + mid + lag) % N32)] ^= r;
    s->psfmt32[idxof(i)] = r;
    i = (i + 1) % N32;
  }
 
  s->idx = N32;
  period_certification(s);
  s->initialized = 1;
}
 
 
static inline void dt_points_init(dt_points_t *p, const unsigned int num_threads)
{
  /* NOT the pixelpipe cache arena. This is a small (~20 kB), permanent, process-wide
   * allocation made once during dt_init(), long before any pipeline exists. The cache
   * arena is LRU-managed and refuses allocations when the system is under memory
   * pressure -- by design -- so asking it for the RNG state made startup fail on a
   * machine that was merely low on free RAM:
   *
   *   [pixelpipe_cache] refusing to allocate 0 MiB: the system has only 1638 MiB
   *                     of available RAM left (pressure floor: 5175 MiB)
   *
   * and the unchecked result then segfaulted in the loop below. A permanent startup
   * buffer belongs to the ordinary allocator. */
  sfmt_state_t *states = (sfmt_state_t *)dt_alloc_align(sizeof(sfmt_state_t) * num_threads);
  p->s = (sfmt_state_t **)calloc(num_threads, sizeof(sfmt_state_t *));
  p->num = num_threads;
 
  if(IS_NULL_PTR(states) || IS_NULL_PTR(p->s))
  {
    dt_print(DT_DEBUG_ALWAYS,
             "[points] out of memory allocating the RNG state for %u thread(s)\n", num_threads);
    dt_free_align(states);
    dt_free(p->s);
    p->s = NULL;
    p->num = 0;
    return;
  }
 
  int seed = 0xD71337;
  for(int i = 0; i < (int)num_threads; i++)
  {
    p->s[i] = states + i;
#if !defined(BIG_ENDIAN64) || defined(ONLY64)
    p->s[i]->psfmt64 = (uint64_t *)&(p->s[i]->sfmt[0].u[0]);
#endif
    p->s[i]->psfmt32 = &(p->s[i]->sfmt[0].u[0]);
    p->s[i]->initialized = 0;
    p->s[i]->parity[0] = PARITY1;
    p->s[i]->parity[1] = PARITY2;
    p->s[i]->parity[2] = PARITY3;
    p->s[i]->parity[3] = PARITY4;
    init_gen_rand(p->s[i], seed);
    seed ^= seed << 1;
  }
}
 
static inline void dt_points_cleanup(dt_points_t *p)
{
  if(IS_NULL_PTR(p->s)) return;      // init failed, or already cleaned up
  dt_free_align(p->s[0]);
  dt_free(p->s);
  p->s = NULL;
  p->num = 0;
}
 
static inline float dt_points_get_for(dt_points_t *p, const unsigned int thread_num)
{
  /* dt_points_init() logs and leaves p->s NULL when it cannot allocate the state, which
   * only happens if the system is genuinely out of memory. Degrade to a fixed value
   * rather than dereferencing NULL: callers use this for dithering and noise, where
   * losing randomness costs a little image quality, whereas crashing costs the session.
   * The failure is already reported once, at init, instead of per pixel. */
  if(IS_NULL_PTR(p) || IS_NULL_PTR(p->s) || thread_num >= p->num) return 0.5f;
 
  return genrand_real2f(p->s[thread_num]);
}
 
static inline float dt_points_get()
{
  return dt_points_get_for(dt_points_get_global(), dt_get_thread_num());
}
 
/* The vendored SFMT generator above defines its parameters as bare macros -- `N`, `POS1`,
 * `MSK1`, `inline` -- at file scope, and leaks every one of them into whatever is included
 * after this header. `N` is the dangerous one: sqlite3.h declares
 * `sqlite3_compileoption_get(int N)`, so any translation unit that reaches sqlite3.h *after*
 * points.h fails to compile, with an error pointing at sqlite3.h and no hint as to why.
 *
 * That was latent for as long as something else happened to include sqlite3.h earlier in
 * every affected unit -- which the old gui/gtk.h did, by dragging in history/history.h. Take
 * one god-header out of the include graph and a dozen IOPs stop compiling for reasons that
 * have nothing to do with the change. Undefine them here so the leak cannot come back.
 */
#undef N
#undef N32
#undef N64
#undef MEXP
#undef POS1
#undef SL1
#undef SL2
#undef SR1
#undef SR2
#undef MSK1
#undef MSK2
#undef MSK3
#undef MSK4
#undef PARITY1
#undef PARITY2
#undef PARITY3
#undef PARITY4
#undef IDSTR
#undef ALWAYSINLINE
#undef PRE_ALWAYS
 
#endif
 
#endif // DT_COMMON_POINTS_H
 
// clang-format off
// modelines: These editor modelines have been set for all relevant files by tools/update_modelines.py
// vim: shiftwidth=2 expandtab tabstop=2 cindent
// kate: tab-indents: off; indent-width 2; replace-tabs on; indent-mode cstyle; remove-trailing-spaces modified;
// clang-format on