forked from personal/squiggle.c
472 lines
15 KiB
C
472 lines
15 KiB
C
#include <float.h> // FLT_MAX, FLT_MIN
|
|
#include <limits.h> // INT_MAX
|
|
#include <math.h> // erf, sqrt
|
|
#include <stdint.h>
|
|
#include <stdio.h>
|
|
#include <stdlib.h>
|
|
#include <sys/types.h>
|
|
#include <time.h>
|
|
|
|
#define EXIT_ON_ERROR 0
|
|
#define MAX_ERROR_LENGTH 500
|
|
#define PROCESS_ERROR(...) \
|
|
do { \
|
|
if (EXIT_ON_ERROR) { \
|
|
printf("@, in %s (%d)", __FILE__, __LINE__); \
|
|
exit(1); \
|
|
} else { \
|
|
char error_msg[MAX_ERROR_LENGTH]; \
|
|
snprintf(error_msg, MAX_ERROR_LENGTH, "@, in %s (%d)", __FILE__, __LINE__); \
|
|
struct box error = { .empty = 1, .error_msg = error_msg }; \
|
|
return error; \
|
|
} \
|
|
} while (0)
|
|
#define NUM_SAMPLES 1000000
|
|
|
|
struct box {
|
|
int empty;
|
|
float content;
|
|
char* error_msg;
|
|
};
|
|
|
|
// Example cdf
|
|
float cdf_uniform_0_1(float x)
|
|
{
|
|
if (x < 0) {
|
|
return 0;
|
|
} else if (x > 1) {
|
|
return 1;
|
|
} else {
|
|
return x;
|
|
}
|
|
}
|
|
|
|
float cdf_squared_0_1(float x)
|
|
{
|
|
if (x < 0) {
|
|
return 0;
|
|
} else if (x > 1) {
|
|
return 1;
|
|
} else {
|
|
return x * x;
|
|
}
|
|
}
|
|
|
|
float cdf_normal_0_1(float x)
|
|
{
|
|
float mean = 0;
|
|
float std = 1;
|
|
return 0.5 * (1 + erf((x - mean) / (std * sqrt(2)))); // erf from math.h
|
|
}
|
|
|
|
// [x] to do: add beta.
|
|
// [x] for the cdf, use this incomplete beta function implementation, based on continuous fractions:
|
|
// <https://codeplea.com/incomplete-beta-function-c>
|
|
// <https://github.com/codeplea/incbeta>
|
|
|
|
#define STOP_BETA 1.0e-8
|
|
#define TINY_BETA 1.0e-30
|
|
struct box incbeta(float a, float b, float x)
|
|
{
|
|
// Descended from <https://github.com/codeplea/incbeta/blob/master/incbeta.c>,
|
|
// but modified to return a box struct and floats instead of doubles.
|
|
// [ ] to do: add attribution in README
|
|
// Original code under this license:
|
|
/*
|
|
* zlib License
|
|
*
|
|
* Regularized Incomplete Beta Function
|
|
*
|
|
* Copyright (c) 2016, 2017 Lewis Van Winkle
|
|
* http://CodePlea.com
|
|
*
|
|
* This software is provided 'as-is', without any express or implied
|
|
* warranty. In no event will the authors be held liable for any damages
|
|
* arising from the use of this software.
|
|
*
|
|
* Permission is granted to anyone to use this software for any purpose,
|
|
* including commercial applications, and to alter it and redistribute it
|
|
* freely, subject to the following restrictions:
|
|
*
|
|
* 1. The origin of this software must not be misrepresented; you must not
|
|
* claim that you wrote the original software. If you use this software
|
|
* in a product, an acknowledgement in the product documentation would be
|
|
* appreciated but is not required.
|
|
* 2. Altered source versions must be plainly marked as such, and must not be
|
|
* misrepresented as being the original software.
|
|
* 3. This notice may not be removed or altered from any source distribution.
|
|
*/
|
|
if (x < 0.0 || x > 1.0) {
|
|
PROCESS_ERROR("x out of bounds [0, 1], in function incbeta");
|
|
}
|
|
|
|
/*The continued fraction converges nicely for x < (a+1)/(a+b+2)*/
|
|
if (x > (a + 1.0) / (a + b + 2.0)) {
|
|
struct box symmetric_incbeta = incbeta(b, a, 1.0 - x);
|
|
if (symmetric_incbeta.empty) {
|
|
return symmetric_incbeta; // propagate error
|
|
} else {
|
|
struct box result = {
|
|
.empty = 0,
|
|
.content = 1 - symmetric_incbeta.content
|
|
};
|
|
return result;
|
|
}
|
|
}
|
|
|
|
/*Find the first part before the continued fraction.*/
|
|
const float lbeta_ab = lgamma(a) + lgamma(b) - lgamma(a + b);
|
|
const float front = exp(log(x) * a + log(1.0 - x) * b - lbeta_ab) / a;
|
|
|
|
/*Use Lentz's algorithm to evaluate the continued fraction.*/
|
|
float f = 1.0, c = 1.0, d = 0.0;
|
|
|
|
int i, m;
|
|
for (i = 0; i <= 200; ++i) {
|
|
m = i / 2;
|
|
|
|
float numerator;
|
|
if (i == 0) {
|
|
numerator = 1.0; /*First numerator is 1.0.*/
|
|
} else if (i % 2 == 0) {
|
|
numerator = (m * (b - m) * x) / ((a + 2.0 * m - 1.0) * (a + 2.0 * m)); /*Even term.*/
|
|
} else {
|
|
numerator = -((a + m) * (a + b + m) * x) / ((a + 2.0 * m) * (a + 2.0 * m + 1)); /*Odd term.*/
|
|
}
|
|
|
|
/*Do an iteration of Lentz's algorithm.*/
|
|
d = 1.0 + numerator * d;
|
|
if (fabs(d) < TINY_BETA)
|
|
d = TINY_BETA;
|
|
d = 1.0 / d;
|
|
|
|
c = 1.0 + numerator / c;
|
|
if (fabs(c) < TINY_BETA)
|
|
c = TINY_BETA;
|
|
|
|
const float cd = c * d;
|
|
f *= cd;
|
|
|
|
/*Check for stop.*/
|
|
if (fabs(1.0 - cd) < STOP_BETA) {
|
|
struct box result = {
|
|
.empty = 0,
|
|
.content = front * (f - 1.0)
|
|
};
|
|
return result;
|
|
}
|
|
}
|
|
|
|
PROCESS_ERROR("More loops needed, did not converge, in function incbeta");
|
|
}
|
|
|
|
struct box cdf_beta(float x)
|
|
{
|
|
if (x < 0) {
|
|
struct box result = { .empty = 0, .content = 0 };
|
|
return result;
|
|
} else if (x > 1) {
|
|
struct box result = { .empty = 0, .content = 1 };
|
|
return result;
|
|
} else {
|
|
float successes = 1, failures = (2023 - 1945);
|
|
return incbeta(successes, failures, x);
|
|
}
|
|
}
|
|
|
|
// Inverse cdf at point
|
|
// Two versions of this function:
|
|
// - raw, dealing with cdfs that return floats
|
|
// - box, dealing with cdfs that return a box.
|
|
|
|
// Inverse cdf
|
|
struct box inverse_cdf_float(float cdf(float), float p)
|
|
{
|
|
// given a cdf: [-Inf, Inf] => [0,1]
|
|
// returns a box with either
|
|
// x such that cdf(x) = p
|
|
// or an error
|
|
// if EXIT_ON_ERROR is set to 1, it exits instead of providing an error
|
|
|
|
float low = -1.0;
|
|
float high = 1.0;
|
|
|
|
// 1. Make sure that cdf(low) < p < cdf(high)
|
|
int interval_found = 0;
|
|
while ((!interval_found) && (low > -FLT_MAX / 4) && (high < FLT_MAX / 4)) {
|
|
// ^ Using FLT_MIN and FLT_MAX is overkill
|
|
// but it's also the *correct* thing to do.
|
|
|
|
int low_condition = (cdf(low) < p);
|
|
int high_condition = (p < cdf(high));
|
|
if (low_condition && high_condition) {
|
|
interval_found = 1;
|
|
} else if (!low_condition) {
|
|
low = low * 2;
|
|
} else if (!high_condition) {
|
|
high = high * 2;
|
|
}
|
|
}
|
|
|
|
if (!interval_found) {
|
|
PROCESS_ERROR("Interval containing the target value not found, in function inverse_cdf");
|
|
} else {
|
|
|
|
int convergence_condition = 0;
|
|
int count = 0;
|
|
while (!convergence_condition && (count < (INT_MAX / 2))) {
|
|
float mid = (high + low) / 2;
|
|
int mid_not_new = (mid == low) || (mid == high);
|
|
// float width = high - low;
|
|
// if ((width < 1e-8) || mid_not_new){
|
|
if (mid_not_new) {
|
|
convergence_condition = 1;
|
|
} else {
|
|
float mid_sign = cdf(mid) - p;
|
|
if (mid_sign < 0) {
|
|
low = mid;
|
|
} else if (mid_sign > 0) {
|
|
high = mid;
|
|
} else if (mid_sign == 0) {
|
|
low = mid;
|
|
high = mid;
|
|
}
|
|
}
|
|
}
|
|
|
|
if (convergence_condition) {
|
|
struct box result = { .empty = 0, .content = low };
|
|
return result;
|
|
} else {
|
|
PROCESS_ERROR("Search process did not converge, in function inverse_cdf");
|
|
}
|
|
}
|
|
}
|
|
|
|
struct box inverse_cdf_box(struct box cdf_box(float), float p)
|
|
{
|
|
// given a cdf: [-Inf, Inf] => Box([0,1])
|
|
// returns a box with either
|
|
// x such that cdf(x) = p
|
|
// or an error
|
|
// if EXIT_ON_ERROR is set to 1, it exits instead of providing an error
|
|
|
|
float low = -1.0;
|
|
float high = 1.0;
|
|
|
|
// 1. Make sure that cdf(low) < p < cdf(high)
|
|
int interval_found = 0;
|
|
while ((!interval_found) && (low > -FLT_MAX / 4) && (high < FLT_MAX / 4)) {
|
|
// ^ Using FLT_MIN and FLT_MAX is overkill
|
|
// but it's also the *correct* thing to do.
|
|
struct box cdf_low = cdf_box(low);
|
|
if (cdf_low.empty) {
|
|
PROCESS_ERROR(cdf_low.error_msg);
|
|
}
|
|
|
|
struct box cdf_high = cdf_box(high);
|
|
if (cdf_high.empty) {
|
|
PROCESS_ERROR(cdf_low.error_msg);
|
|
}
|
|
|
|
int low_condition = (cdf_low.content < p);
|
|
int high_condition = (p < cdf_high.content);
|
|
if (low_condition && high_condition) {
|
|
interval_found = 1;
|
|
} else if (!low_condition) {
|
|
low = low * 2;
|
|
} else if (!high_condition) {
|
|
high = high * 2;
|
|
}
|
|
}
|
|
|
|
if (!interval_found) {
|
|
PROCESS_ERROR("Interval containing the target value not found, in function inverse_cdf");
|
|
} else {
|
|
|
|
int convergence_condition = 0;
|
|
int count = 0;
|
|
while (!convergence_condition && (count < (INT_MAX / 2))) {
|
|
float mid = (high + low) / 2;
|
|
int mid_not_new = (mid == low) || (mid == high);
|
|
// float width = high - low;
|
|
if (mid_not_new) {
|
|
// if ((width < 1e-8) || mid_not_new){
|
|
convergence_condition = 1;
|
|
} else {
|
|
struct box cdf_mid = cdf_box(mid);
|
|
if (cdf_mid.empty) {
|
|
PROCESS_ERROR(cdf_mid.error_msg);
|
|
}
|
|
float mid_sign = cdf_mid.content - p;
|
|
if (mid_sign < 0) {
|
|
low = mid;
|
|
} else if (mid_sign > 0) {
|
|
high = mid;
|
|
} else if (mid_sign == 0) {
|
|
low = mid;
|
|
high = mid;
|
|
}
|
|
}
|
|
}
|
|
|
|
if (convergence_condition) {
|
|
struct box result = { .empty = 0, .content = low };
|
|
return result;
|
|
} else {
|
|
PROCESS_ERROR("Search process did not converge, in function inverse_cdf");
|
|
}
|
|
}
|
|
}
|
|
|
|
// Some randomness functions for:
|
|
// - Sampling from a cdf
|
|
// - Benchmarking against a previous approach, which will be faster, but less general
|
|
|
|
// Get random number between 0 and 1
|
|
uint32_t xorshift32(uint32_t* seed)
|
|
{
|
|
// Algorithm "xor" from p. 4 of Marsaglia, "Xorshift RNGs"
|
|
// See <https://stackoverflow.com/questions/53886131/how-does-xorshift32-works>
|
|
// https://en.wikipedia.org/wiki/Xorshift
|
|
// Also some drama: <https://www.pcg-random.org/posts/on-vignas-pcg-critique.html>, <https://prng.di.unimi.it/>
|
|
|
|
uint32_t x = *seed;
|
|
x ^= x << 13;
|
|
x ^= x >> 17;
|
|
x ^= x << 5;
|
|
return *seed = x;
|
|
}
|
|
|
|
// Distribution & sampling functions
|
|
float rand_0_to_1(uint32_t* seed)
|
|
{
|
|
return ((float)xorshift32(seed)) / ((float)UINT32_MAX);
|
|
}
|
|
|
|
// Sampler based on inverse cdf and randomness function
|
|
struct box sampler_box_cdf(struct box cdf(float), uint32_t* seed)
|
|
{
|
|
float p = rand_0_to_1(seed);
|
|
struct box result = inverse_cdf_box(cdf, p);
|
|
return result;
|
|
}
|
|
struct box sampler_float_cdf(float cdf(float), uint32_t* seed)
|
|
{
|
|
float p = rand_0_to_1(seed);
|
|
struct box result = inverse_cdf_float(cdf, p);
|
|
return result;
|
|
}
|
|
|
|
// Comparison point with raw normal sampler
|
|
const float PI = 3.14159265358979323846;
|
|
float sampler_normal_0_1(uint32_t* seed)
|
|
{
|
|
float u1 = rand_0_to_1(seed);
|
|
float u2 = rand_0_to_1(seed);
|
|
float z = sqrtf(-2.0 * log(u1)) * sin(2 * PI * u2);
|
|
return z;
|
|
}
|
|
|
|
// Some testers
|
|
void test_inverse_cdf_float(char* cdf_name, float cdf_float(float))
|
|
{
|
|
struct box result = inverse_cdf_float(cdf_float, 0.5);
|
|
if (result.empty) {
|
|
printf("Inverse for %s not calculated\n", cdf_name);
|
|
exit(1);
|
|
} else {
|
|
printf("Inverse of %s at %f is: %f\n", cdf_name, 0.5, result.content);
|
|
}
|
|
}
|
|
void test_inverse_cdf_box(char* cdf_name, struct box cdf_box(float))
|
|
{
|
|
struct box result = inverse_cdf_box(cdf_box, 0.5);
|
|
if (result.empty) {
|
|
printf("Inverse for %s not calculated\n", cdf_name);
|
|
exit(1);
|
|
} else {
|
|
printf("Inverse of %s at %f is: %f\n", cdf_name, 0.5, result.content);
|
|
}
|
|
}
|
|
|
|
void test_and_time_sampler_float(char* cdf_name, float cdf_float(float), uint32_t* seed)
|
|
{
|
|
printf("\nGetting some samples from %s:\n", cdf_name);
|
|
clock_t begin = clock();
|
|
for (int i = 0; i < NUM_SAMPLES; i++) {
|
|
struct box sample = sampler_float_cdf(cdf_float, seed);
|
|
if (sample.empty) {
|
|
printf("Error in sampler function for %s", cdf_name);
|
|
} else {
|
|
// printf("%f\n", sample.content);
|
|
}
|
|
}
|
|
clock_t end = clock();
|
|
float time_spent = (float)(end - begin) / CLOCKS_PER_SEC;
|
|
printf("Time spent: %f\n", time_spent);
|
|
}
|
|
|
|
void test_and_time_sampler_box(char* cdf_name, struct box cdf_box(float), uint32_t* seed)
|
|
{
|
|
printf("\nGetting some samples from %s:\n", cdf_name);
|
|
clock_t begin = clock();
|
|
for (int i = 0; i < NUM_SAMPLES; i++) {
|
|
struct box sample = sampler_box_cdf(cdf_box, seed);
|
|
if (sample.empty) {
|
|
printf("Error in sampler function for %s", cdf_name);
|
|
} else {
|
|
// printf("%f\n", sample.content);
|
|
}
|
|
}
|
|
clock_t end = clock();
|
|
float time_spent = (float)(end - begin) / CLOCKS_PER_SEC;
|
|
printf("Time spent: %f\n", time_spent);
|
|
}
|
|
|
|
int main()
|
|
{
|
|
// Test inverse cdf float
|
|
test_inverse_cdf_float("cdf_uniform_0_1", cdf_uniform_0_1);
|
|
test_inverse_cdf_float("cdf_squared_0_1", cdf_squared_0_1);
|
|
test_inverse_cdf_float("cdf_normal_0_1", cdf_normal_0_1);
|
|
|
|
// Test inverse cdf box
|
|
test_inverse_cdf_box("cdf_beta", cdf_beta);
|
|
|
|
// Testing samplers
|
|
// set randomness seed
|
|
uint32_t* seed = malloc(sizeof(uint32_t));
|
|
*seed = 1000; // xorshift can't start with 0
|
|
|
|
// Test float sampler
|
|
test_and_time_sampler_float("cdf_uniform_0_1", cdf_uniform_0_1, seed);
|
|
test_and_time_sampler_float("cdf_squared_0_1", cdf_squared_0_1, seed);
|
|
test_and_time_sampler_float("cdf_normal_0_1", cdf_normal_0_1, seed);
|
|
|
|
// Get some normal samples using a previous approach
|
|
printf("\nGetting some samples from sampler_normal_0_1\n");
|
|
|
|
clock_t begin_2 = clock();
|
|
|
|
for (int i = 0; i < NUM_SAMPLES; i++) {
|
|
float normal_sample = sampler_normal_0_1(seed);
|
|
// printf("%f\n", normal_sample);
|
|
}
|
|
|
|
clock_t end_2 = clock();
|
|
float time_spent_2 = (float)(end_2 - begin_2) / CLOCKS_PER_SEC;
|
|
printf("Time spent: %f\n", time_spent_2);
|
|
|
|
// Test box sampler
|
|
test_and_time_sampler_box("cdf_beta", cdf_beta, seed);
|
|
// Ok, this is slower than python!!
|
|
// Partly this is because I am using a more general algorithm,
|
|
// which applies to any cdf
|
|
// But I am also using really anal convergence conditions.
|
|
// This could be optimized.
|
|
|
|
free(seed);
|
|
return 0;
|
|
}
|