add multithreading

.clangd

@@ -1,6 +1,8 @@
 CompileFlags:
   Add:
-    - "-std=c23"
+    - "-std=gnu23"
     - "-xc"
+    - "-march=native"
+    - "-mavx2"
   Remove:
     - "-std=c++*"

.gitignore

@@ -0,0 +1,4 @@
+*.png
+.*
+!/.gitignore
+main

main.c

@@ -1,9 +1,13 @@
+#include <errno.h>
 #include <math.h>
 #include <png.h>
+#include <pthread.h>
+#include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 #include <sys/random.h>
+#include <time.h>
 
 float noise(float x, float y);
 float linear_interpolation(float a, float b, float t);
@@ -13,14 +17,132 @@ float *pseudo_random_gradient_vector_get(float x, float y);
 
 static unsigned char random_bytes[1024];
 
-int main() {
+enum threaded { THRD_MULTI, THRD_SINGLE };
+
+struct single_threaded_arg {
+  enum threaded type;
+  unsigned char *buffer;
+  int height;
+  int width;
+};
+
+struct multi_threaded_arg {
+  enum threaded type;
+  unsigned char *buffer;
+  int height_start;
+  int height_end;
+  int width;
+};
+
+void *calculate(void *args) { // TODO: reorder functions by call order
+  int x = 0;
+  int height = 0, width = 0;
+  unsigned char *buffer;
+
+  if (args != NULL) {
+    enum threaded *threaded = (enum threaded *)args;
+    if (*threaded == THRD_SINGLE) {
+      struct single_threaded_arg *arg = (struct single_threaded_arg *)args;
+      height = arg->height;
+      width = arg->width;
+      buffer = arg->buffer;
+    } else if (*threaded == THRD_MULTI) {
+      struct multi_threaded_arg *arg = (struct multi_threaded_arg *)args;
+
+      x = arg->height_start;
+      height = arg->height_end;
+      width = arg->width;
+      buffer = arg->buffer;
+    }
+  }
+
+  if (height == 0 || width == 0) {
+    printf("things have got seriously wrong!");
+    if (height == 0) {
+      printf(" height is 0!");
+    } else {
+      printf(" width is 0!");
+    }
+    printf(" aborting\n");
+    exit(1);
+  }
+
+  for (; x < height; ++x) {
+    for (int y = 0; y < width; ++y) {
+      float noise_val = noise(x * 0.05, y * 0.05);
+      unsigned char pixel = (unsigned char)((noise_val + 1.0f) * 127.5f);
+
+      buffer[y + x * width] = pixel;
+    }
+  }
+
+  return NULL;
+}
+
+void single_threaded(unsigned char *buffer, int height, int width) {
+  struct single_threaded_arg arg = {
+      .type = THRD_SINGLE, .buffer = buffer, .height = height, .width = width};
+  calculate(&arg);
+}
+
+void multi_threaded(unsigned char *buffer, int height, int width,
+                    int thread_no) {
+  if (thread_no == 1) {
+    single_threaded(buffer, height, width);
+    return;
+  }
+
+  pthread_t threads[thread_no];
+  const unsigned int rows_per_thread = height / thread_no;
+  printf("rows per thread: %u\n", rows_per_thread);
+  struct multi_threaded_arg *args_arr[thread_no];
+  for (int i = 0; i < thread_no; i++) {
+    const int height_start = i * rows_per_thread;
+    int height_end = (i + 1) * rows_per_thread;
+    if (i == thread_no - 1) {
+      height_end = height;
+    }
+
+    struct multi_threaded_arg *arg;
+    arg = malloc(sizeof *arg);
+
+    *arg = (struct multi_threaded_arg){.type = THRD_MULTI,
+                                       .buffer = buffer,
+                                       .height_start = height_start,
+                                       .height_end = height_end,
+                                       .width = width};
+    pthread_create(&threads[i], NULL, calculate, arg);
+    args_arr[i] = arg;
+  }
+
+  for (int i = 0; i < thread_no; ++i) {
+    pthread_join(threads[i], NULL);
+    free(args_arr[i]);
+  }
+}
+
+int main(int argc, char *argv[]) {
   int n = getrandom(random_bytes, 1024, 0);
   if (n != 1024) {
     fprintf(stderr, "err: getrandom didn't work out as planned\n");
     return 1;
   }
 
-  int width = 1024, height = 1024;
+  if (argc < 3) {
+    fprintf(stderr, "err: must supply 3 params!\n");
+    return 1;
+  }
+  int width = atoi(argv[1]), height = atoi(argv[2]);
+  int threads = 1;
+  if (argc == 4) {
+    threads = atoi(argv[3]);
+
+    if (height < threads) {
+      fprintf(stderr,
+              "err: rows < threads. expected: rows >= threads. aborting\n");
+      return 1;
+    }
+  }
 
   png_image img;
   memset(&img, 0, sizeof(img));
@@ -30,19 +152,26 @@ int main() {
   img.version = PNG_IMAGE_VERSION;
 
   unsigned char *buffer;
-  printf("PNG IMAGE SIZE: %u\n", PNG_IMAGE_SIZE(img));
+  printf("PNG IMAGE SIZE: %u bytes\n", PNG_IMAGE_SIZE(img));
   buffer = malloc(PNG_IMAGE_SIZE(img));
 
-  for (int x = 0; x < height; ++x) {
+  struct timespec start, end;
-    for (int y = 0; y < width; ++y) {
-      float noise_val = noise(x * 0.05, y * 0.05);
-      unsigned char pixel = (unsigned char)((noise_val + 1.0f) * 127.5f);
 
-      buffer[y + x * width] = pixel;
+  clock_gettime(CLOCK_MONOTONIC, &start);
-    }
+  if (threads < 2)
-  }
+    single_threaded(buffer, height, width);
+  else
+    multi_threaded(buffer, height, width, threads);
+  clock_gettime(CLOCK_MONOTONIC, &end);
+  printf("calculation time: %fs.\n", (double)(end.tv_sec - start.tv_sec) +
+                                         (end.tv_nsec - start.tv_nsec) / 1e9);
 
+  clock_gettime(CLOCK_MONOTONIC, &start);
   png_image_write_to_file(&img, "noise.png", 0, buffer, 0, NULL);
+  clock_gettime(CLOCK_MONOTONIC, &end);
+  printf("write to file time (single-threaded): %fs.\n",
+         (double)(end.tv_sec - start.tv_sec) +
+             (end.tv_nsec - start.tv_nsec) / 1e9);
 
   free(buffer);
 }