Revive voronoi multi-threaded demo.

native_client_sdk/src/examples/demo/voronoi/voronoi.cc

Index: native_client_sdk/src/examples/demo/voronoi/voronoi.cc
===================================================================
--- native_client_sdk/src/examples/demo/voronoi/voronoi.cc	(revision 0)
+++ native_client_sdk/src/examples/demo/voronoi/voronoi.cc	(revision 0)
@@ -0,0 +1,584 @@
+// Copyright 2013 The Native Client SDK Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can
+// be found in the LICENSE file.
+
+#include <assert.h>
+#include <math.h>
+#include <ppapi/cpp/completion_callback.h>
+#include <ppapi/cpp/graphics_2d.h>
+#include <ppapi/cpp/image_data.h>
+#include <ppapi/cpp/input_event.h>
+#include <ppapi/cpp/instance.h>
+#include <ppapi/cpp/module.h>
+#include <ppapi/cpp/rect.h>
+#include <ppapi/cpp/size.h>
+#include <ppapi/cpp/var.h>
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/time.h>
+#include <unistd.h>
+
+#include <string>
+
+#include "threadpool.h"
+
+// Global properties used to setup Voronoi demo.
+namespace {
+const int kMinRectSize = 4;
+const int kStartRecurseSize = 32;

[binji, 2013/05/24 21:47:55]

must be power of 2

[nfullagar1, 2013/05/28 23:11:50]

comment added

+const float kHugeZ = 1.0e38f;
+const float kPI = M_PI;
+const float kTwoPI = kPI * 2.0f;
+const int kFramesToBenchmark = 100;
+const unsigned int kRandomStartSeed = 0xC0DE533D;
+const int kMaxPointCount = 1024;
+const int kStartPointCount = 256;
+
+unsigned int g_rand_state = kRandomStartSeed;
+
+// random number helper
+inline unsigned char rand255() {
+  return static_cast<unsigned char>(rand_r(&g_rand_state) & 255);
+}
+
+// random number helper
+inline float frand() {
+  return (static_cast<float>(rand_r(&g_rand_state)) /
+      static_cast<float>(RAND_MAX));
+}
+
+// reset random seed
+inline void rand_reset(unsigned int seed) {
+  g_rand_state = seed;
+}
+
+inline double getseconds() {
+  const double usec_to_sec = 0.000001;
+  timeval tv;
+  if (0 == gettimeofday(&tv, NULL))
+    return tv.tv_sec + tv.tv_usec * usec_to_sec;
+  return 0.0;
+}
+
+inline uint32_t MakeRGBA(uint32_t r, uint32_t g, uint32_t b, uint32_t a) {
+  return (((a) << 24) | ((r) << 16) | ((g) << 8) | (b));
+}
+}  // namespace
+
+// Vec2, simple 2D vector
+struct Vec2 {
+  float x, y;
+  Vec2() { ; }

[binji, 2013/05/24 21:47:55]

init to 0?

[noelallen1, 2013/05/25 00:21:11]

We typically use {} not {;}

[nfullagar1, 2013/05/28 23:11:50]

This struct is closer to an int or a float (in my mind) which aren't set to zero
by default.

[nfullagar1, 2013/05/28 23:11:50]

Done.

+  Vec2(float px, float py) {
+    x = px;
+    y = py;
+  }
+  void Set(float px, float py) {
+    x = px;
+    y = py;
+  }
+};
+
+// The main object that runs Voronoi simulation.
+class Voronoi : public pp::Instance {
+ public:
+  explicit Voronoi(PP_Instance instance);
+  virtual ~Voronoi();
+
+  virtual bool Init(uint32_t argc, const char* argn[], const char* argv[]) {
+    return true;
+  }
+
+  virtual void DidChangeView(const pp::Rect& position, const pp::Rect& clip);
+
+  // Catch events.
+  virtual bool HandleInputEvent(const pp::InputEvent& event);
+
+  // Catch messages posted from Javascript.
+  virtual void HandleMessage(const pp::Var& message);
+
+ private:
+  // Methods prefixed with 'w' are run on worker threads.
+  int wCell(float x, float y);
+  inline void wFillSpan(uint32_t *pixels, uint32_t color, int width);
+  void wRenderTile(int x, int y, int w, int h);
+  void wProcessTile(int x, int y, int w, int h);
+  void wSubdivide(int x, int y, int w, int h);
+  void wMakeRect(int region, int *x, int *y, int *w, int *h);
+  bool wTestRect(int *m, int x, int y, int w, int h);
+  void wFillRect(int x, int y, int w, int h, uint32_t color);
+  void wRenderRect(int x0, int y0, int x1, int y1);
+  void wRenderRegion(int region);
+  static void wRenderRegionEntry(int region, void *thiz);
+
+  // These methods are only called by the main thread.
+  void Reset();
+  void UpdateSim();
+  void RenderDot(float x, float y, uint32_t color1, uint32_t color2);
+  void SuperimposePositions();
+  void Render();
+  void Draw();
+  void StartBenchmark();
+  void EndBenchmark();
+
+  // Runs a tick of the simulations, updating all buffers.  Flushes the
+  // contents of |image_data_| to the 2D graphics context.
+  void Update();
+
+  // Create and initialize the 2D context used for drawing.
+  void CreateContext(const pp::Size& size);
+  // Destroy the 2D drawing context.
+  void DestroyContext();
+  // Push the pixels to the browser, then attempt to flush the 2D context.
+  void FlushPixelBuffer();
+  static void FlushCallback(void* data, int32_t result);
+
+
+  pp::Graphics2D* graphics_2d_context_;
+  pp::ImageData* image_data_;
+  Vec2 positions_[kMaxPointCount];
+  Vec2 screen_positions_[kMaxPointCount];
+  Vec2 velocities_[kMaxPointCount];
+  uint32_t colors_[kMaxPointCount];
+  float ang_;
+  int point_count_;
+  int num_threads_;
+  int num_regions_;
+  bool draw_points_;
+  bool draw_interiors_;
+  ThreadPool *workers_;
+  int width_;
+  int height_;
+  uint32_t stride_in_pixels_;
+  uint32_t *pixel_buffer_;
+  int benchmark_frame_counter_;
+  bool benchmarking_;
+  double benchmark_start_time_;
+  double benchmark_end_time_;
+};
+
+
+
+void Voronoi::Reset() {
+  rand_reset(kRandomStartSeed);
+  ang_ = 0.0f;
+  for (int i = 0; i < kMaxPointCount; i++) {
+    // random initial start position
+    const float x = frand();
+    const float y = frand();
+    positions_[i].Set(x, y);
+    // random directional velocity ( -1..1, -1..1 )
+    const float speed = 0.0005f;
+    const float u = (frand() * 2.0f - 1.0f) * speed;
+    const float v = (frand() * 2.0f - 1.0f) * speed;
+    velocities_[i].Set(u, v);
+    // 'unique' color (well... unique enough for our purposes)
+    colors_[i] = MakeRGBA(rand255(), rand255(), rand255(), 255);
+  }
+}
+
+Voronoi::Voronoi(PP_Instance instance) : pp::Instance(instance),
+                                         graphics_2d_context_(NULL),
+                                         image_data_(NULL) {
+  draw_points_ = true;
+  draw_interiors_ = true;
+  width_ = 0;
+  height_ = 0;
+  stride_in_pixels_ = 0;
+  pixel_buffer_ = NULL;
+  benchmark_frame_counter_ = 0;
+  benchmarking_ = false;
+
+  point_count_ = kStartPointCount;
+  Reset();
+
+  // By default, do single threaded rendering.
+  num_regions_ = 256;
+  num_threads_ = 1;
+  workers_ = new ThreadPool(num_threads_);
+
+  // Request PPAPI input events for mouse & keyboard.
+  RequestInputEvents(PP_INPUTEVENT_CLASS_MOUSE);

[noelallen1, 2013/05/25 00:21:11]

You aren't looking at mouse events.

[nfullagar1, 2013/05/28 23:11:50]

Not now.  The thought being if a user makes local changes to handle mouse
events, they won't be filtered by default.

+  RequestInputEvents(PP_INPUTEVENT_CLASS_KEYBOARD);
+}
+
+Voronoi::~Voronoi() {
+  delete workers_;
+  DestroyContext();
+}
+
+// This is the core of the Voronoi calculation.  At a given point on the
+// screen, iterate through all voronoi positions and render them as 3D cones.
+// We're looking for the voronoi cell that generates the closest z value.
+// (not really cones - since it is all realative, we avoid doing the

[binji, 2013/05/24 21:47:55]

sp: relative

[nfullagar1, 2013/05/28 23:11:50]

Done.

+// expensive sqrt and test against z*z instead)
+// If multithreading, this function is only called by the worker threads.
+int Voronoi::wCell(float x, float y) {
+  int n = 0;

[binji, 2013/05/24 21:47:55]

maybe a better var name here... closest_index?

[nfullagar1, 2013/05/28 23:11:50]

s/n/closest_cell, removed comment on return statement

+  float zz = kHugeZ;

[binji, 2013/05/24 21:47:55]

maybe minz?

[nfullagar1, 2013/05/28 23:11:50]

Is comment above sufficient?

+  Vec2 *pos = screen_positions_;
+  for (int i = 0; i < point_count_; ++i) {
+    // measured 5.18 cycles per iteration on a core2
+    float dx = x - pos[i].x;
+    float dy = y - pos[i].y;
+    float dd = (dx * dx + dy * dy);
+    if (dd < zz) {
+      zz = dd;
+      n = i;
+    }
+  }
+  // Return the closest cell to point x,y.
+  return n;
+}
+
+// Given a region r, derive a non-overlapping rectangle for a thread to
+// work on.
+// If multithreading, this function is only called by the worker threads.
+void Voronoi::wMakeRect(int r, int *x, int *y, int *w, int *h) {
+  const int parts = 16;
+  assert(parts * parts == num_regions_);

[binji, 2013/05/24 21:47:55]

num_regions should be const too, I think.

[nfullagar1, 2013/05/28 23:11:50]

Done.

[binji, 2013/05/28 23:37:50]

Ah, I just meant it seemed strange to have parts defined here, but have it rely
on num_regions_ being a specific value which is defined in Voronoi -- I see it
is because Render uses it below.

+  *w = width_ / parts;
+  *h = height_ / parts;
+  *x = *w * (r % parts);
+  *y = *h * ((r / parts) % parts);
+}
+
+// Test 4 corners of a rectangle to see if they all belong to the same
+// voronoi cell.  Each test is expensive so bail asap.  Returns true
+// if all 4 corners match.
+// If multithreading, this function is only called by the worker threads.
+bool Voronoi::wTestRect(int *m, int x, int y, int w, int h) {
+  // each test is expensive, so exit ASAP
+  const int m0 = wCell(x, y);
+  const int m1 = wCell(x + w - 1, y);
+  if (m0 != m1) return false;
+  const int m2 = wCell(x, y + h - 1);
+  if (m0 != m2) return false;
+  const int m3 = wCell(x + w - 1, y + h - 1);
+  if (m0 != m3) return false;
+  // all 4 corners belong to the same cell
+  *m = m0;
+  return true;
+}
+
+// Quickly fill a span of pixels with a solid color.  Assumes
+// span width is divisible by 4.
+// If multithreading, this function is only called by the worker threads.
+inline void Voronoi::wFillSpan(uint32_t *pixels, uint32_t color, int width) {
+  if (!draw_interiors_) {
+    const uint32_t gray = MakeRGBA(128, 128, 128, 255);
+    color = gray;
+  }
+  for (int i = 0; i < width; i += 4) {
+    *pixels++ = color;
+    *pixels++ = color;
+    *pixels++ = color;
+    *pixels++ = color;
+  }
+}
+
+// Quickly fill a rectangle with a solid color.  Assumes
+// the width w parameter is evenly divisible by 4.
+// If multithreading, this function is only called by the worker threads.
+void Voronoi::wFillRect(int x, int y, int w, int h, uint32_t color) {
+  const uint32_t pitch = width_;
+  uint32_t *pixels = pixel_buffer_ + y * pitch + x;
+  for (int j = 0; j < h; j++) {
+    wFillSpan(pixels, color, w);
+    pixels += pitch;
+  }
+}
+
+// When recursive subdivision reaches a certain minimum without finding a
+// rectangle that has four matching corners belonging to the same voronoi
+// cell, this function will break the retangular 'tile' into smaller scanlines
+//  and look for opportunities to quick fill at the scanline level.  If the
+// scanline can't be quick filled, it will slow down even further and compute
+// voronoi membership per pixel.
+void Voronoi::wRenderTile(int x, int y, int w, int h) {
+  // rip through a tile
+  uint32_t *pixels = pixel_buffer_ + y * stride_in_pixels_ + x;
+  for (int j = 0; j < h; j++) {
+    // get start and end cell values
+    int ms = wCell(x + 0, y + j);
+    int me = wCell(x + w - 1, y + j);
+    // if the end points are the same, quick fill the span
+    if (ms == me) {
+      wFillSpan(pixels, colors_[ms], w);
+    } else {
+      // else compute each pixel in the span... this is the slow part!
+      uint32_t *p = pixels;
+      *p++ = colors_[ms];
+      for (int i = 1; i < (w - 1); i++) {
+        int m = wCell(x + i, y + j);
+        *p++ = colors_[m];
+      }
+      *p++ = colors_[me];
+    }
+    pixels += stride_in_pixels_;
+  }
+}
+
+// Take a rectangular region and do one of -
+//   If all four corners below to the same voronoi cell, stop recursion and
+// quick fill the rectangle.
+//   If the minimum rectangle size has been reached, break out of recursion
+// and process the rectangle.  This small rectangle is called a tile.
+//   Otherwise, keep recursively subdividing the rectangle into 4 equally
+// sized smaller rectangles.
+// Note: at the moment, these will always be squares, not rectangles.
+// If multithreading, this function is only called by the worker threads.
+void Voronoi::wSubdivide(int x, int y, int w, int h) {
+  int m;
+  // if all 4 corners are equal, quick fill interior
+  if (wTestRect(&m, x, y, w, h)) {
+    wFillRect(x, y, w, h, colors_[m]);
+  } else {
+    // did we reach the minimum rectangle size?
+    if ((w <= kMinRectSize) || (h <= kMinRectSize)) {
+      wRenderTile(x, y, w, h);
+    } else {
+      // else recurse into smaller rectangles
+      const int half_w = w / 2;
+      const int half_h = h / 2;
+      wSubdivide(x, y, half_w, half_h);
+      wSubdivide(x + half_w, y, half_w, half_h);
+      wSubdivide(x, y + half_h, half_w, half_h);
+      wSubdivide(x + half_w, y + half_h, half_w, half_h);
+    }
+  }
+}
+
+// This function cuts up the rectangle into power of 2 sized squares.  It
+// assumes the input rectangle w & h are evenly divisible by
+// kStartRecurseSize.
+// If multithreading, this function is only called by the worker threads.
+void Voronoi::wRenderRect(int x, int y, int w, int h) {
+  for (int iy = y; iy < (y + h); iy += kStartRecurseSize) {
+    for (int ix = x; ix < (x + w); ix += kStartRecurseSize) {
+      wSubdivide(ix, iy, kStartRecurseSize, kStartRecurseSize);
+    }
+  }
+}
+
+// If multithreading, this function is only called by the worker threads.
+void Voronoi::wRenderRegion(int region) {
+  // convert region # into x0, y0, x1, y1 rectangle
+  int x, y, w, h;
+  wMakeRect(region, &x, &y, &w, &h);
+  // render this rectangle
+  wRenderRect(x, y, w, h);

[binji, 2013/05/24 21:47:55]

assert that w, h are power of two?

[nfullagar1, 2013/05/28 23:11:50]

done (but assert is down in CreateContext)

+}
+
+// Entry point for worker thread.  Can't pass a member function around, so we
+// have to do this little round-about.
+void Voronoi::wRenderRegionEntry(int region, void *thiz) {

[binji, 2013/05/24 21:47:55]

I've been calling this a "thunk"

[nfullagar1, 2013/05/28 23:11:50]

hmm... if you don't feel strongly about this, it feels better (imo) for the
Dispatch() function to invoke an Entry function as opposed to a Thunk function.

[binji, 2013/05/28 23:37:50]

Nope, Entry is fine.

+  static_cast<Voronoi*>(thiz)->wRenderRegion(region);
+}
+
+// Function Voronoi::UpdateSim()
+// Run a simple sim to move the voronoi positions.  This update loop
+// is run once per frame.  Called from the main thread only, and only
+// when the worker threads are idle.
+void Voronoi::UpdateSim() {
+  ang_ += 0.002f;
+  if (ang_ > kTwoPI) {
+    ang_ = ang_ - kTwoPI;
+  }
+  float z = cosf(ang_) * 3.0f;
+  // push the points around on the screen for animation
+  for (int j = 0; j < kMaxPointCount; j++) {
+    positions_[j].x += (velocities_[j].x) * z;
+    positions_[j].y += (velocities_[j].y) * z;
+    screen_positions_[j].x = positions_[j].x * width_;
+    screen_positions_[j].y = positions_[j].y * height_;
+  }
+}
+
+// Renders a small diamond shaped dot at x, y clipped against the window
+void Voronoi::RenderDot(float x, float y, uint32_t color1, uint32_t color2) {
+  const int ix = static_cast<int>(x);
+  const int iy = static_cast<int>(y);
+  // clip it against window
+  if (ix < 1) return;
+  if (ix >= (width_ - 1)) return;
+  if (iy < 1) return;
+  if (iy >= (height_ - 1)) return;
+  uint32_t *pixel = pixel_buffer_ + iy * stride_in_pixels_ + ix;
+  // render dot as a small diamond
+  *pixel = color1;
+  *(pixel - 1) = color2;
+  *(pixel + 1) = color2;
+  *(pixel - stride_in_pixels_) = color2;
+  *(pixel + stride_in_pixels_) = color2;
+}
+
+// Superimposes dots on the positions.
+void Voronoi::SuperimposePositions() {
+  const uint32_t white = MakeRGBA(255, 255, 255, 255);
+  const uint32_t gray = MakeRGBA(192, 192, 192, 255);
+  for (int i = 0; i < point_count_; i++) {
+    RenderDot(
+        screen_positions_[i].x, screen_positions_[i].y, white, gray);
+  }
+}
+
+// Renders the Voronoi diagram, dispatching the work to multiple threads.
+// Note: This Dispatch() is from the main PPAPI thread, so care must be taken
+// not to attempt PPAPI calls from the worker threads, since Dispatch() will
+// block here until all work is complete.  The worker threads are compute only
+// and do not make any PPAPI calls.
+void Voronoi::Render() {
+  workers_->Dispatch(num_regions_, wRenderRegionEntry, this);
+  if (draw_points_)
+    SuperimposePositions();
+}
+
+void Voronoi::DidChangeView(const pp::Rect& position, const pp::Rect& clip) {
+  if (position.size().width() == width_ &&
+      position.size().height() == height_)
+    return;  // Size didn't change, no need to update anything.
+
+  // Create a new device context with the new size.
+  DestroyContext();
+  CreateContext(position.size());
+  Update();
+}
+
+void Voronoi::StartBenchmark() {
+  Reset();
+  printf("Benchmark started...\n");
+  benchmark_frame_counter_ = kFramesToBenchmark;
+  benchmarking_ = true;
+  benchmark_start_time_ = getseconds();
+}
+
+void Voronoi::EndBenchmark() {
+  benchmark_end_time_ = getseconds();
+  printf("Benchmark ended... time: %2.5f\n",
+      benchmark_end_time_ - benchmark_start_time_);
+  benchmarking_ = false;
+  benchmark_frame_counter_ = 0;
+  pp::Var result(benchmark_end_time_ - benchmark_start_time_);
+  PostMessage(result);
+}
+
+// Handle input events from the user.
+bool Voronoi::HandleInputEvent(const pp::InputEvent& event) {
+  switch (event.GetType()) {
+    case PP_INPUTEVENT_TYPE_KEYDOWN: {
+      pp::KeyboardInputEvent key = pp::KeyboardInputEvent(event);
+      uint32_t key_code = key.GetKeyCode();
+      if (key_code == 84)  // 't' key
+        if (!benchmarking_)
+          StartBenchmark();
+      break;
+    }
+    default:
+      return false;
+  }
+  return true;
+}
+
+// Handle messages sent from Javascript.
+void Voronoi::HandleMessage(const pp::Var& message) {
+  if (message.is_string()) {
+    std::string message_string = message.AsString();
+    if (message_string == "run benchmark" && !benchmarking_)
+      StartBenchmark();
+    else if (message_string == "with points")
+      draw_points_ = true;
+    else if (message_string == "without points")
+      draw_points_ = false;
+    else if (message_string == "with interiors")
+      draw_interiors_ = true;
+    else if (message_string == "without interiors")
+      draw_interiors_ = false;
+    else if (strstr(message_string.c_str(), "points:")) {
+      int num_points = atoi(strstr(message_string.c_str(), " "));
+      point_count_ = num_points < kMaxPointCount ? num_points : kMaxPointCount;

[binji, 2013/05/24 21:47:55]

check for negative?

[nfullagar1, 2013/05/28 23:11:50]

Done.

+    } else if (strstr(message_string.c_str(), "threads:")) {
+      int thread_count = atoi(strstr(message_string.c_str(), " "));
+      delete workers_;
+      workers_ = new ThreadPool(thread_count);
+    }
+  }
+}
+
+void Voronoi::FlushCallback(void* thiz, int32_t result) {
+  static_cast<Voronoi*>(thiz)->Update();
+}
+
+// Update the 2d region and flush to make it visible on the page.
+void Voronoi::FlushPixelBuffer() {
+  graphics_2d_context_->PaintImageData(*image_data_, pp::Point(0, 0));
+  graphics_2d_context_->Flush(pp::CompletionCallback(&FlushCallback, this));
+}
+
+void Voronoi::Update() {
+  // Don't call FlushPixelBuffer() when benchmarking - vsync is enabled by
+  // default, and will throttle the benchmark results.
+  do {
+    UpdateSim();
+    Render();
+    if (!benchmarking_) break;
+    --benchmark_frame_counter_;
+  } while (benchmark_frame_counter_ > 0);
+  if (benchmarking_)
+    EndBenchmark();
+  FlushPixelBuffer();
+}
+
+void Voronoi::CreateContext(const pp::Size& size) {
+  const size_t bytes = size.width() * size.height();

[binji, 2013/05/24 21:47:55]

unused

[nfullagar1, 2013/05/28 23:11:50]

good catch

+  graphics_2d_context_ = new pp::Graphics2D(this, size, false);
+  if (!BindGraphics(*graphics_2d_context_))
+    printf("Couldn't bind the device context\n");
+  image_data_ = new pp::ImageData(this,
+                                  PP_IMAGEDATAFORMAT_BGRA_PREMUL,
+                                  size,
+                                  false);
+  width_ = image_data_->size().width();
+  height_ = image_data_->size().height();
+  stride_in_pixels_ = static_cast<uint32_t>(image_data_->stride() / 4);
+  pixel_buffer_ = static_cast<uint32_t*>(image_data_->data());
+}
+
+void Voronoi::DestroyContext() {
+  delete graphics_2d_context_;
+  delete image_data_;
+  graphics_2d_context_ = NULL;
+  image_data_ = NULL;
+  width_ = 0;
+  height_ = 0;
+  stride_in_pixels_ = 0;
+  pixel_buffer_ = NULL;
+}
+
+// The Module class.  The browser calls the CreateInstance() method to create

[binji, 2013/05/24 21:47:55]

kill comment

[nfullagar1, 2013/05/28 23:11:50]

Done.

+// an instance of you NaCl module on the web page.  The browser creates a new
+// instance for each <embed> tag with type="application/x-nacl".
+class VoronoiModule : public pp::Module {
+ public:
+  VoronoiModule() : pp::Module() {}
+  virtual ~VoronoiModule() {}
+
+  // Create and return a Voronoi instance.
+  virtual pp::Instance* CreateInstance(PP_Instance instance) {
+    return new Voronoi(instance);
+  }
+};
+
+// Factory function called by the browser when the module is first loaded.

[binji, 2013/05/24 21:47:55]

kill comment

[nfullagar1, 2013/05/28 23:11:50]

Done.

+// The browser keeps a singleton of this module.  It calls the
+// CreateInstance() method on the object you return to make instances.  There
+// is one instance per <embed> tag on the page.  This is the main binding
+// point for your NaCl module with the browser.
+namespace pp {
+Module* CreateModule() {
+  return new VoronoiModule();
+}
+}  // namespace pp
+
Property changes on: native_client_sdk/src/examples/demo/voronoi/voronoi.cc
___________________________________________________________________
Added: svn:executable
   + *