//GPUArray<int> h = new GPUArray<int>(100000000);
public Game(GameWindow window) : base(window)
{
new Shader2D("SIFT.shaders.plain_color.glsl").Run(Window.Width, Window.Height, Window.Canvas);
//object a = new int();
//Console.WriteLine(a.GetType());
////return;
int n = 100000000;
int loop_n = 1; // 60 * 32;
GPUIntArray s_gpu = new GPUIntArray(n);
List <int> s_cpu = new List <int>(n); for (int i = 0; i < n; i++)
{
s_cpu.Add(0);
}
Print(s_gpu.IsValue(7122));
s_gpu.Value(7122);
Print(s_gpu.IsValue(7122));
GL.Finish();
//Print("GPU time:", a=Timing(() =>
//{
// for (int i = 0; i < n * power * 11; i++)
// {
// s_gpu[Rand.Next(n)] = Rand.Next();
// }
// GL.Finish();
//}));
TimeSpan a, b;
Print("GPU time:", a = Timing(() =>
{
for (int i = 0; i < loop_n; i++)
{
//Print("i =", i);
s_gpu.Random();
//s_gpu.Data(Shuffled(Range(n)).ToArray());
s_gpu.Sort();
//s_gpu.Value(0);
//if (!s_gpu.IsSorted())
//{
// Print(s_gpu);
//}
Assert(s_gpu.IsSorted());
}
GL.Finish();
}));
Print("CPU time:", b = Timing(() =>
{
for (int i = 0; i < loop_n; i++)
{
//Print("i =", i);
s_cpu.Random();
s_cpu.Sort();
Assert(s_cpu.IsSorted());
}
}));
Print("speedup =", b.TotalMilliseconds / a.TotalMilliseconds);
Print("finish");
}
public void Sort()
{
int n = value.Length;
if (n <= 1)
{
return;
}
int max_level = (__builtin_popcount(n) == 1 ? 31 : 32) - __builtin_clz(n);
left.Value(0); rigt.Value(n - 1);
//Print(value);
for (int level = 1; level <= max_level; level++)
{
//OpenTK.Graphics.OpenGL.GL.Finish();
//Print(Timing(() =>
//{
new Shader("SIFT.shaders.merge_sort.glsl").QueueForRunInSequence(n,
("level", level),
value, left, rigt, output);
// OpenTK.Graphics.OpenGL.GL.Finish();
//}));
value.Swap(output);
//Print(value);
//Print(debug);
}
}
public void Sort()
{
int n = value.Length;
if (n <= 1)
{
return;
}
int num_levels = (__builtin_popcount(n) == 1 ? 31 : 32) - __builtin_clz(n);
left.Value(0); rigt.Value(n - 1);
for (int start_level = 1; start_level <= num_levels; start_level++)
{
for (int level = start_level; level >= 1; level--)
{
new Shader("SIFT.shaders.bitonic_merge.glsl").QueueForRunInSequence(n,
("start_level", start_level),
("level", level),
("reverse", level == start_level ? 1 : 0),
value, left, rigt, output);
value.Swap(output);
//Print(value);
}
}
}
public void Sort()
{
int n = value.Length;
if (n <= 1)
{
return;
}
int max_level = (__builtin_popcount(n) == 1 ? 31 : 32) - __builtin_clz(n);
left.Value(0); rigt.Value(n - 1);
//Print(value);
int stride = 16;// Math.Max(1, n / Shader.default_group_size_x / gpu_execute_unit_count);
//Print("stride =", stride);
for (int level = 1; level <= max_level; level++)
{
//OpenTK.Graphics.OpenGL.GL.Finish();
//Print(Timing(() =>
//{
new Shader("SIFT.shaders.merge_batch_sort.glsl").QueueForRunInSequence((n + stride - 1) / stride,
("level", level),
("stride", stride),
value, left, rigt, output);
// OpenTK.Graphics.OpenGL.GL.Finish();
//}));
value.Swap(output);
//Print(value);
//Print(debug);
}
}
public void Sort()
{
int n = value.Length;
if (n <= 1)
{
return;
}
int max_level = (__builtin_popcount(n) == 1 ? 31 : 32) - __builtin_clz(n);
left.Value(0); rigt.Value(n - 1);
//Print(value);
for (int level = 1; level <= max_level; level++)
{
int total_execute_cnt = 0;
for (int stride_level = max_level; stride_level >= 0; stride_level--)
{
// (1 << sl) - 1, 3(1 << sl) - 1, 5(1 << sl) - 1, 7(1 << sl) - 1, ...
int offset = (1 << stride_level) - 1;
//int stride = 1 << stride_level << 1;
// offset + ? * stride <= n - 1
// ? * stride <= n - 1 - offset
int execute_cnt = ((n - 1 - offset) >> stride_level >> 1) + 1;
total_execute_cnt += execute_cnt;
//OpenTK.Graphics.OpenGL.GL.Finish();
//Print(Timing(() =>
//{
new Shader("SIFT.shaders.merge_sort_write_a.glsl").QueueForRunInSequence(execute_cnt,
("level", level),
("stride_level", stride_level),
value, left, rigt, a);
// OpenTK.Graphics.OpenGL.GL.Finish();
//}));
//Print(a);
}
//Print();
Assert(total_execute_cnt == n);
new Shader("SIFT.shaders.merge_sort_read_a_write_o.glsl").QueueForRunInSequence(n,
("level", level),
value, left, rigt, a, output);
value.Swap(output);
//Print(value);
//Print(debug);
}
}