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);
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 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);
}
}
public void Sort()
{
int n = value.Length;
if (n <= 1)
{
return;
}
Assert(__builtin_popcount(n) == 1);
for (int i = 1; i < n; i++)
{
(left[i], rigt[i]) = ((i << 1) | 1) > n - 1 ? (-1, -1) : (i << 1, (i << 1) | 1);
}
int root = 1, spare = 0;
BiSort(root, spare, true);
Flatten(root, 0, n - 2);
output[n - 1] = value[spare];
value.Swap(output);
}
private void PMerge(int n)
{
Assert(__builtin_popcount(n) == 1);
for (int thread_id = 0; thread_id < n - 1; thread_id++)
{
int i = thread_id;
int cto = __builtin_ctz(~i);
int shift = (1 << (cto - 1));
(left[i], rigt[i]) = cto == 0 ? /*leaf*/ (-1, -1) : (i - shift, i + shift);
roots[i] = i;
spares[i] = i + (1 << cto);
}
for (int start_level = 1; start_level <= __builtin_ctz(n); start_level++)
{
//PrintTree(n / 2 - 1, n - 1);
for (int level = start_level; level >= 1; level--)
{
int id_max = n >> level;
for (int thread_id = 0; thread_id < id_max; thread_id++)
{
int i = (thread_id << level) + (1 << (level - 1)) - 1;
int origin_id = i; // origin location on binary tree (before any node swaps)
int root = roots[origin_id];
int spare = spares[origin_id]; // 10101001111 -> 101010
//Print("i",i,"root", root, "spare", spare);
bool ascend = __builtin_popcount(i >> start_level) % 2 == 0;
if ((value[root] < value[spare]) != ascend) // exchange
{
(value[root], value[spare]) = (value[spare], value[root]);
(left[root], rigt[root]) = (rigt[root], left[root]);
}
if (level > 1)
{
(int p, int q) = (left[root], rigt[root]);
while (p != -1)
{
if ((value[p] < value[q]) != ascend) // swap left tree of p & q, go right
{
(value[p], value[q]) = (value[q], value[p]);
(left[p], left[q]) = (left[q], left[p]);
(p, q) = (rigt[p], rigt[q]);
}
else // go left
{
(p, q) = (left[p], left[q]);
}
}
int shift = 1 << (level - 2);
(roots[origin_id - shift], spares[origin_id - shift]) = (left[root], root);
(roots[origin_id + shift], spares[origin_id + shift]) = (rigt[root], spare);
}
}
}
}
for (int level = __builtin_ctz(n); level >= 1; level--)
{
int id_max = n >> level;
for (int thread_id = 0; thread_id < id_max; thread_id++)
{
int i = (thread_id << level) + (1 << (level - 1)) - 1;
int origin_id = i; // origin location on binary tree (before any node swaps)
int root = roots[origin_id];
spares[i] = value[root];
}
}
spares[n - 1] = value[n - 1];
value.Swap(spares);
}