reorg, get parsimonious paralellism working, to go into squiggle.c

This commit is contained in:
NunoSempere 2023-11-18 23:25:39 +00:00
parent 9a56a63c61
commit a48b15f171
10 changed files with 429 additions and 11 deletions

View File

@ -0,0 +1,107 @@
# Interface:
# make
# make build
# make format
# make run
# Compiler
CC=gcc
# CC=tcc # <= faster compilation
# Main file
SRC=samples.c
OUTPUT=out/samples
SRC_ONE_THREAD=./samples-one-thread.c
OUTPUT_ONE_THREAD=out/samples-one-thread
## Dependencies
# Has no dependencies
MATH=-lm
## Flags
DEBUG= #'-g'
STANDARD=-std=c99
WARNINGS=-Wall
OPTIMIZED=-O3 #-O3 actually gives better performance than -Ofast, at least for this version
OPENMP=-fopenmp
## Formatter
STYLE_BLUEPRINT=webkit
FORMATTER=clang-format -i -style=$(STYLE_BLUEPRINT)
## make build
build: $(SRC)
$(CC) $(OPTIMIZED) $(DEBUG) $(SRC) $(OPENMP) $(MATH) -o $(OUTPUT)
static:
$(CC) $(OPTIMIZED) $(DEBUG) $(SRC) $(OPENMP) $(MATH) -o $(OUTPUT)
format: $(SRC)
$(FORMATTER) $(SRC)
run: $(SRC) $(OUTPUT)
OMP_NUM_THREADS=1 ./$(OUTPUT) && echo
multi:
OMP_NUM_THREADS=1 ./$(OUTPUT) && echo
OMP_NUM_THREADS=2 ./$(OUTPUT) && echo
OMP_NUM_THREADS=4 ./$(OUTPUT) && echo
OMP_NUM_THREADS=8 ./$(OUTPUT) && echo
OMP_NUM_THREADS=16 ./$(OUTPUT) && echo
## Timing
time-linux:
@echo "Requires /bin/time, found on GNU/Linux systems" && echo
@echo "Running 100x and taking avg time: OMP_NUM_THREADS=1 $(OUTPUT)"
@t=$$(/usr/bin/time -f "%e" -p bash -c 'for i in {1..100}; do OMP_NUM_THREADS=1 $(OUTPUT); done' 2>&1 >/dev/null | grep real | awk '{print $$2}' ); echo "scale=2; 1000 * $$t / 100" | bc | sed "s|^|Time using 1 thread: |" | sed 's|$$|ms|' && echo
@echo "Running 100x and taking avg time: OMP_NUM_THREADS=2 $(OUTPUT)"
@t=$$(/usr/bin/time -f "%e" -p bash -c 'for i in {1..100}; do OMP_NUM_THREADS=2 $(OUTPUT); done' 2>&1 >/dev/null | grep real | awk '{print $$2}' ); echo "scale=2; 1000 * $$t / 100" | bc | sed "s|^|Time using 2 threads: |" | sed 's|$$|ms|' && echo
@echo "Running 100x and taking avg time: OMP_NUM_THREADS=4 $(OUTPUT)"
@t=$$(/usr/bin/time -f "%e" -p bash -c 'for i in {1..100}; do OMP_NUM_THREADS=4 $(OUTPUT); done' 2>&1 >/dev/null | grep real | awk '{print $$2}' ); echo "scale=2; 1000 * $$t / 100" | bc | sed "s|^|Time for 4 threads: |" | sed 's|$$|ms|' && echo
@echo "Running 100x and taking avg time: OMP_NUM_THREADS=8 $(OUTPUT)"
@t=$$(/usr/bin/time -f "%e" -p bash -c 'for i in {1..100}; do OMP_NUM_THREADS=8 $(OUTPUT); done' 2>&1 >/dev/null | grep real | awk '{print $$2}' ); echo "scale=2; 1000 * $$t / 100" | bc | sed "s|^|Time using 8 threads: |" | sed 's|$$|ms|' && echo
@echo "Running 100x and taking avg time: OMP_NUM_THREADS=16 $(OUTPUT)"
@t=$$(/usr/bin/time -f "%e" -p bash -c 'for i in {1..100}; do OMP_NUM_THREADS=16 $(OUTPUT); done' 2>&1 >/dev/null | grep real | awk '{print $$2}' ); echo "scale=2; 1000 * $$t / 100" | bc | sed "s|^|Time using 16 threads: |" | sed 's|$$|ms|' && echo
time-linux-fastest:
@echo "Running 100x and taking avg time: OMP_NUM_THREADS=32 $(OUTPUT)"
@t=$$(/usr/bin/time -f "%e" -p bash -c 'for i in {1..100}; do OMP_NUM_THREADS=16 $(OUTPUT); done' 2>&1 >/dev/null | grep real | awk '{print $$2}' ); echo "scale=2; 1000 * $$t / 100" | bc | sed "s|^|Time using 16 threads: |" | sed 's|$$|ms|' && echo
time-linux-simple:
@echo "Requires /bin/time, found on GNU/Linux systems" && echo
OMP_NUM_THREADS=1 /bin/time -f "Time: %es" ./$(OUTPUT) && echo
OMP_NUM_THREADS=2 /bin/time -f "Time: %es" ./$(OUTPUT) && echo
OMP_NUM_THREADS=4 /bin/time -f "Time: %es" ./$(OUTPUT) && echo
OMP_NUM_THREADS=8 /bin/time -f "Time: %es" ./$(OUTPUT) && echo
OMP_NUM_THREADS=16 /bin/time -f "Time: %es" ./$(OUTPUT) && echo
OMP_NUM_THREADS=32 /bin/time -f "Time: %es" ./$(OUTPUT) && echo
## Profiling
profile-linux:
echo "Requires perf, which depends on the kernel version, and might be in linux-tools package or similar"
echo "Must be run as sudo"
$(CC) $(SRC) $(OPENMP) $(MATH) -o $(OUTPUT)
# ./$(OUTPUT)
# gprof:
# gprof $(OUTPUT) gmon.out > analysis.txt
# rm gmon.out
# vim analysis.txt
# rm analysis.txt
# perf:
OMP_NUM_THREADS=16 sudo perf record $(OUTPUT)
sudo perf report
rm perf.data
## Install
debian-install-dependencies:
sudo apt-get install libomp-dev

View File

@ -6,7 +6,7 @@
const float PI = 3.14159265358979323846;
#define N_SAMPLES 1024 * 1000
#define N_SAMPLES (1024 * 1000 )
//Array helpers
@ -180,36 +180,54 @@ float mixture(float (*samplers[])(uint32_t*), float* weights, int n_dists, uint3
// Parallization function
void paralellize(float (*sampler)(uint32_t* seed), float* results, int n_threads, int n_samples){
if((N_SAMPLES % n_threads) != 0){
fprintf(stderr, "Number of samples isn't divisible by number of threads, aborting\n");
exit(1);
}
// int n_samples_per_thread = N_SAMPLES / n_thread;
int sample_index, i, split_array_length;
int n_samples_per_thread = N_SAMPLES / n_threads;
float** split_results = malloc(n_threads * sizeof(float*));
for(int i=0; i<n_threads; i++){
split_results[i] = malloc(n_samples_per_thread * sizeof(float));
}
uint32_t** seeds = malloc(n_threads * sizeof(uint32_t*));
for (uint32_t i = 0; i < n_threads; i++) {
seeds[i] = malloc(sizeof(uint32_t));
*seeds[i] = i + 1; // xorshift can't start with 0
}
#pragma omp parallel private(i, sample_index, split_array_length)
int i;
#pragma omp parallel private(i)
{
#pragma omp for
for (i = 0; i < n_threads; i++) {
// split_array_length = split_array_get_length(i, N_SAMPLES, n_threads);
int lower_bound = i * (n_samples / n_threads);
int upper_bound = ((i+1) * (n_samples / n_threads)) - 1;
// printf("Lower bound: %d, upper bound: %d\n", lower_bound, upper_bound);
for (int j = lower_bound; j < upper_bound; j++) {
results[j] = sampler(seeds[i]);
for (int j = 0; j < n_samples_per_thread; j++) {
split_results[i][j] = sampler(seeds[i]);
}
}
}
for(int i=0; i<n_threads; i++){
int lower_bound = i * (n_samples / n_threads);
int upper_bound = ((i+1) * (n_samples / n_threads)) - 1;
// printf("Lower bound: %d, upper bound: %d\n", lower_bound, upper_bound);
for(int j=lower_bound; j<upper_bound; j++){
results[j] = split_results[i][j-lower_bound];
}
}
for (uint32_t i = 0; i < n_threads; i++) {
free(seeds[i]);
}
free(seeds);
for(int i=0; i<n_threads; i++){
free(split_results[i]);
}
free(split_results);
}
// Functions used for the BOTEC.
@ -257,7 +275,8 @@ int main()
float* split_array_results = malloc(N_SAMPLES * sizeof(float));
paralellize(sample_mixture, split_array_results, n_threads, N_SAMPLES);
printf("Sum(split_array_results, N_SAMPLES)/N_SAMPLES = %f\n", array_sum(split_array_results, N_SAMPLES) / N_SAMPLES);
printf("Sum(split_array_results, N_SAMPLES)/N_SAMPLES = %f\n",
array_sum(split_array_results, N_SAMPLES) / N_SAMPLES);
free(split_array_results);
return 0;

View File

@ -0,0 +1,17 @@
Requires /bin/time, found on GNU/Linux systems
Running 100x and taking avg time: OMP_NUM_THREADS=1 out/samples
Time using 1 thread: 35.60ms
Running 100x and taking avg time: OMP_NUM_THREADS=2 out/samples
Time using 2 threads: 38.40ms
Running 100x and taking avg time: OMP_NUM_THREADS=4 out/samples
Time for 4 threads: 26.30ms
Running 100x and taking avg time: OMP_NUM_THREADS=8 out/samples
Time using 8 threads: 15.30ms
Running 100x and taking avg time: OMP_NUM_THREADS=16 out/samples
Time using 16 threads: 10.40ms

View File

@ -0,0 +1,37 @@
Instead of
```C
#pragma omp parallel private(i, sample_index, split_array_length)
{
#pragma omp for
for (i = 0; i < n_threads; i++) {
split_array_length = split_array_get_length(i, N, n_threads);
for (int j = 0; j < split_array_length; j++) {
results[i][j] = sampler(seeds[i]);
}
}
}
```
Algo como
```C
#pragma omp parallel private(i, sample_index, bounds)
{
#pragma omp for
for (i = 0; i < n_threads; i++) {
int bounds[2] = split_array_get_bounds(i, N, n_threads);
for (int j = bound[0]; j < bounds[1] + 1; j++) {
// o j < bounds[1], no se si el +1 va a ser más elegante
// dentro o fuera, aunque algo me dice que dentro
results[j] = sampler(seeds[i]);
}
}
}
```
Por qué? Porque esto hace que la estructura subyacente sea un solo array,
lo cual implica que no *tengo* que utilizar split_array_functions especializadas
si no quiero

View File

@ -6,6 +6,7 @@
# Compiler
CC=gcc
# CC=clang
# CC=tcc # <= faster compilation
# Main file

Binary file not shown.

View File

@ -0,0 +1,220 @@
#include <math.h>
#include <omp.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
const float PI = 3.14159265358979323846;
#define N_SAMPLES (1024 * 1000 * 1000)
//Array helpers
void array_print(float* array, int length)
{
for (int i = 0; i < length; i++) {
printf("item[%d] = %f\n", i, array[i]);
}
printf("\n");
}
float array_sum(float* array, int length)
{
float output = 0.0;
for (int i = 0; i < length; i++) {
output += array[i];
}
return output;
}
void array_cumsum(float* array_to_sum, float* array_cumsummed, int length)
{
array_cumsummed[0] = array_to_sum[0];
for (int i = 1; i < length; i++) {
array_cumsummed[i] = array_cumsummed[i - 1] + array_to_sum[i];
}
}
// Pseudo Random number generator
uint32_t xorshift32(uint32_t* seed)
{
// Algorithm "xor" from p. 4 of Marsaglia, "Xorshift RN_SAMPLESGs"
// 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);
/*
uint32_t x = *seed;
x ^= x << 13;
x ^= x >> 17;
x ^= x << 5;
return ((float)(*seed = x))/((float) UIN_SAMPLEST32_MAX);
*/
// previously:
// ((float)rand_r(seed) / (float)RAN_SAMPLESD_MAX)
// and before that: rand, but it wasn't thread-safe.
// See: <https://stackoverflow.com/questions/43151361/how-to-create-thread-safe-random-number-generator-in-c-using-rand-r> for why to use rand_r:
// rand() is not thread-safe, as it relies on (shared) hidden seed.
}
float rand_float(float max, uint32_t* seed)
{
return rand_0_to_1(seed) * max;
}
float ur_normal(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;
}
float random_uniform(float from, float to, uint32_t* seed)
{
return rand_0_to_1(seed) * (to - from) + from;
}
float random_normal(float mean, float sigma, uint32_t* seed)
{
return (mean + sigma * ur_normal(seed));
}
float random_lognormal(float logmean, float logsigma, uint32_t* seed)
{
return expf(random_normal(logmean, logsigma, seed));
}
float random_to(float low, float high, uint32_t* seed)
{
const float N_SAMPLESORMAL95CON_SAMPLESFIDEN_SAMPLESCE = 1.6448536269514722;
float loglow = logf(low);
float loghigh = logf(high);
float logmean = (loglow + loghigh) / 2;
float logsigma = (loghigh - loglow) / (2.0 * N_SAMPLESORMAL95CON_SAMPLESFIDEN_SAMPLESCE);
return random_lognormal(logmean, logsigma, seed);
}
// Mixture function
float mixture(float (*samplers[])(uint32_t*), float* weights, int n_dists, uint32_t* seed)
{
// You can see a slightly simpler version of this function in the git history
// or in alt/C-02-better-algorithm-one-thread/
float sum_weights = array_sum(weights, n_dists);
float* cumsummed_normalized_weights = malloc(n_dists * sizeof(float));
cumsummed_normalized_weights[0] = weights[0] / sum_weights;
for (int i = 1; i < n_dists; i++) {
cumsummed_normalized_weights[i] = cumsummed_normalized_weights[i - 1] + weights[i] / sum_weights;
}
//create var holders
float p1, result;
int sample_index, i, own_length;
p1 = random_uniform(0, 1, seed);
for (int i = 0; i < n_dists; i++) {
if (p1 < cumsummed_normalized_weights[i]) {
result = samplers[i](seed);
break;
}
}
free(cumsummed_normalized_weights);
return result;
}
// Parallization function
void paralellize(float (*sampler)(uint32_t* seed), float* results, int n_threads, int n_samples){
if((N_SAMPLES % n_threads) != 0){
fprintf(stderr, "Number of samples isn't divisible by number of threads, aborting\n");
exit(1);
}
// int n_samples_per_thread = N_SAMPLES / n_thread;
int sample_index, i, split_array_length;
uint32_t** seeds = malloc(n_threads * sizeof(uint32_t*));
for (uint32_t i = 0; i < n_threads; i++) {
seeds[i] = malloc(sizeof(uint32_t));
*seeds[i] = i + 1; // xorshift can't start with 0
}
#pragma omp parallel private(i, sample_index, split_array_length)
{
#pragma omp for
for (i = 0; i < n_threads; i++) {
int lower_bound = i * (n_samples / n_threads);
int upper_bound = ((i+1) * (n_samples / n_threads)) - 1;
// printf("Lower bound: %d, upper bound: %d\n", lower_bound, upper_bound);
for (int j = lower_bound; j < upper_bound; j++) {
results[j] = sampler(seeds[i]);
}
}
}
for (uint32_t i = 0; i < n_threads; i++) {
free(seeds[i]);
}
free(seeds);
}
// Functions used for the BOTEC.
// Their type has to be the same, as we will be passing them around.
float sample_0(uint32_t* seed)
{
return 0;
}
float sample_1(uint32_t* seed)
{
return 1;
}
float sample_few(uint32_t* seed)
{
return random_to(1, 3, seed);
}
float sample_many(uint32_t* seed)
{
return random_to(2, 10, seed);
}
float sample_mixture(uint32_t* seed){
float p_a, p_b, p_c;
// Initialize variables
p_a = 0.8;
p_b = 0.5;
p_c = p_a * p_b;
// Generate mixture
int n_dists = 4;
float weights[] = { 1 - p_c, p_c / 2, p_c / 4, p_c / 4 };
float (*samplers[])(uint32_t*) = { sample_0, sample_1, sample_few, sample_many };
return mixture(samplers, weights, n_dists, seed);
}
int main()
{
int n_threads = omp_get_max_threads();
float* split_array_results = malloc(N_SAMPLES * sizeof(float));
paralellize(sample_mixture, split_array_results, n_threads, N_SAMPLES);
printf("Sum(split_array_results, N_SAMPLES)/N_SAMPLES = %f\n", array_sum(split_array_results, N_SAMPLES) / N_SAMPLES);
free(split_array_results);
return 0;
}

View File

@ -0,0 +1,17 @@
Requires /bin/time, found on GNU/Linux systems
Running 100x and taking avg time: OMP_NUM_THREADS=1 out/samples
Time using 1 thread: 32.30ms
Running 100x and taking avg time: OMP_NUM_THREADS=2 out/samples
Time using 2 threads: 30.80ms
Running 100x and taking avg time: OMP_NUM_THREADS=4 out/samples
Time for 4 threads: 30.50ms
Running 100x and taking avg time: OMP_NUM_THREADS=8 out/samples
Time using 8 threads: 12.90ms
Running 100x and taking avg time: OMP_NUM_THREADS=16 out/samples
Time using 16 threads: 9.10ms