Render multithreaded locally by default

src/main/fart.cc

@@ -9,6 +9,10 @@
 #include <iostream>
 #include <string>
 #include <map>
+#include <vector>
+
+#include <pthread.h>
+#include <atomic>
 
 #include "Scene.h"
 #include "distrib/distrib.h"
@@ -16,6 +20,106 @@
 
 using namespace std;
 
+/*
+ * Number of contiguous pixels each worker thread claims at a time.  Pulling
+ * work in small chunks (rather than whole rows) keeps the load balanced even
+ * when individual rows vary wildly in render cost, while keeping the claimed
+ * pixels contiguous so cache locality and dispatch overhead stay close to the
+ * row-based baseline.  See benchmarks: 64 was the sweet spot.
+ */
+#define RENDER_CHUNK_SIZE 64
+
+/*
+ * State shared by the local multithreaded renderer.  Worker threads pull
+ * fixed-size chunks of pixels off a lock-free atomic cursor until the image
+ * is complete.
+ */
+typedef struct
+{
+    Scene * scene;
+    unsigned char * data;
+    int width;
+    int total_pixels;
+    std::atomic<int> next_pixel;    /* dispatch cursor: next unclaimed pixel */
+    std::atomic<int> pixels_done;   /* pixels finished, for progress */
+    std::atomic<int> last_permille; /* last tenth-of-a-percent printed */
+} render_thread_state_t;
+
+static void * render_thread(void * varg)
+{
+    render_thread_state_t * state = (render_thread_state_t *) varg;
+    const int width = state->width;
+    const int total_pixels = state->total_pixels;
+
+    for (;;)
+    {
+        /* atomically claim the next chunk of contiguous pixels */
+        int start = state->next_pixel.fetch_add(RENDER_CHUNK_SIZE,
+                                                std::memory_order_relaxed);
+        if (start >= total_pixels)
+            break;
+        int end = start + RENDER_CHUNK_SIZE;
+        if (end > total_pixels)
+            end = total_pixels;
+
+        for (int pixel = start; pixel < end; pixel++)
+        {
+            int y = pixel / width;
+            int x = pixel % width;
+            state->scene->renderPixel(x, y, &state->data[3 * pixel]);
+        }
+
+        /* report progress without locking: of the threads that advance the
+         * tenth-of-a-percent counter, only the one that wins the update prints,
+         * so the display stays monotonic and free of interleaved output */
+        int done = state->pixels_done.fetch_add(end - start,
+                                                std::memory_order_relaxed)
+                   + (end - start);
+        int permille = (int) (1000L * done / total_pixels);
+        int prev = state->last_permille.load(std::memory_order_relaxed);
+        while (permille > prev)
+        {
+            if (state->last_permille.compare_exchange_weak(
+                    prev, permille, std::memory_order_relaxed))
+            {
+                printf("\e[8D%2.1f%%", permille / 10.0);
+                fflush(stdout);
+                break;
+            }
+            /* prev reloaded by compare_exchange_weak; re-test permille > prev */
+        }
+    }
+
+    return NULL;
+}
+
+static void renderThreaded(Scene & scene,
+                           unsigned char * data,
+                           int width,
+                           int height)
+{
+    int num_threads = sysconf(_SC_NPROCESSORS_ONLN);
+    if (num_threads < 1)
+        num_threads = 1;
+
+    render_thread_state_t state;
+    state.scene = &scene;
+    state.data = data;
+    state.width = width;
+    state.total_pixels = width * height;
+    state.next_pixel.store(0, std::memory_order_relaxed);
+    state.pixels_done.store(0, std::memory_order_relaxed);
+    state.last_permille.store(0, std::memory_order_relaxed);
+
+    vector<pthread_t> threads(num_threads);
+    for (int i = 0; i < num_threads; i++)
+        pthread_create(&threads[i], NULL, render_thread, &state);
+    for (int i = 0; i < num_threads; i++)
+        pthread_join(threads[i], NULL);
+
+    printf("\e[8D");
+}
+
 void usage(const char * progname)
 {
     cout << "Usage: " << progname << " [options] <scene-file>" << endl;
@@ -274,28 +378,8 @@ int main(int argc, char * argv[])
     }
     else
     {
-        int total_pixels = height * width;
+        /* local multithreaded render using all available cores */
-        int total_pixels_1000 = total_pixels / 1000;
+        renderThreaded(scene, data, width, height);
-        if (total_pixels_1000 < 1)
-            total_pixels_1000 = 1;
-        int pixel_num = 0;
-        /* "sequential" version */
-        for (int i = 0; i < height; i++)
-        {
-            for (int j = 0; j < width; j++)
-            {
-                int pixel = i * width + j;
-                scene.renderPixel(j, i, &data[3 * pixel]);
-                pixel_num++;
-                if (pixel_num % total_pixels_1000 == 0)
-                {
-                    double pct = 100.0 * pixel_num / (double) total_pixels;
-                    printf("\e[8D%2.1f%%", pct);
-                    fflush(stdout);
-                }
-            }
-        }
-        printf("\e[8D");
     }
 
     gettimeofday(&after, NULL);         /* stop timing */

src/util/refptr.h

@@ -3,6 +3,7 @@
 #define REFPTR_H REFPTR_H
 
 #include <stdlib.h>             /* NULL */
+#include <atomic>               /* std::atomic */
 
 template <typename T>
 class refptr
@@ -25,7 +26,10 @@ class refptr
         void destroy();
 
         T * m_ptr;
-        int * m_refCount;
+        /* reference count is atomic so that refptr copies may be made
+         * concurrently from multiple threads (e.g. the multithreaded
+         * renderer) without corrupting the count */
+        std::atomic<int> * m_refCount;
 };
 
 template <typename T> refptr<T>::refptr()
@@ -37,8 +41,7 @@ template <typename T> refptr<T>::refptr()
 template <typename T> refptr<T>::refptr(T * ptr)
 {
     m_ptr = ptr;
-    m_refCount = new int;
+    m_refCount = new std::atomic<int>(1);
-    *m_refCount = 1;
 }
 
 template <typename T> refptr<T>::refptr(const refptr<T> & orig)
@@ -57,8 +60,7 @@ template <typename T> refptr<T> & refptr<T>::operator=(T * ptr)
 {
     destroy();
     m_ptr = ptr;
-    m_refCount = new int;
+    m_refCount = new std::atomic<int>(1);
-    *m_refCount = 1;
     return *this;
 }
 
@@ -67,7 +69,7 @@ template <typename T> void refptr<T>::cloneFrom(const refptr<T> & orig)
     this->m_ptr = orig.m_ptr;
     this->m_refCount = orig.m_refCount;
     if (m_refCount != NULL)
-        (*m_refCount)++;
+        m_refCount->fetch_add(1, std::memory_order_relaxed);
 }
 
 template <typename T> refptr<T>::~refptr()
@@ -79,15 +81,13 @@ template <typename T> void refptr<T>::destroy()
 {
     if (m_refCount != NULL)
     {
-        if (*m_refCount <= 1)
+        /* fetch_sub returns the value prior to the decrement; if it was 1
+         * then this was the last reference and we own the cleanup */
+        if (m_refCount->fetch_sub(1, std::memory_order_acq_rel) == 1)
         {
             delete m_ptr;
             delete m_refCount;
         }
-        else
-        {
-            (*m_refCount)--;
-        }
     }
 }