Track errors to make octree reduction more sophisticated

This commit is contained in:
Kovid Goyal 2016-05-09 11:18:42 +05:30
parent e9e4ace6db
commit b6f3368f03

View File

@ -6,6 +6,9 @@
* See https://www.microsoft.com/msj/archive/S3F1.aspx for a simple to follow
* writeup on this algorithm
*
* The implementation below is more sophisticated than the writeup. In particular, it tracks
* total error on each leaf node and uses a memory pool to improve performance.
*
* Distributed under terms of the GPL3 license.
*/
@ -39,6 +42,7 @@ template <typename T> static inline T euclidean_distance(T r1, T g1, T b1, T r2,
}
struct SumPixel { uint64_t red; uint64_t green; uint64_t blue; };
struct DoublePixel { double red; double green; double blue; };
template <typename T> static inline void iadd(T &self, T &other) { self.red += other.red; self.green += other.green; self.blue += other.blue; }
template <class T> class Pool { // {{{
private:
@ -79,6 +83,7 @@ private:
uint64_t pixel_count;
SumPixel sum;
DoublePixel avg;
SumPixel error_sum;
Node* next_reducible_node;
Node *next_available_in_pool;
Node* children[MAX_DEPTH];
@ -88,7 +93,7 @@ public:
// Disable the new behavior warning caused by children() below
#pragma warning( push )
#pragma warning (disable: 4351)
Node() : is_leaf(false), index(0), pixel_count(0), sum(), avg(), next_reducible_node(NULL), next_available_in_pool(NULL), children() {}
Node() : is_leaf(false), index(0), pixel_count(0), sum(), avg(), error_sum(), next_reducible_node(NULL), next_available_in_pool(NULL), children() {}
#pragma warning ( pop )
#endif
@ -97,12 +102,15 @@ public:
this->pixel_count = 0;
this->sum.red = 0; this->sum.green = 0; this->sum.blue = 0;
this->avg.red = 0; this->avg.green = 0; this->avg.blue = 0;
this->error_sum.red = 0; this->error_sum.green = 0; this->error_sum.blue = 0;
this->next_reducible_node = NULL;
for (size_t i = 0; i < MAX_DEPTH; i++) this->children[i] = NULL;
}
void check_compiler() {
if (this->children[0] != NULL) throw std::runtime_error("Compiler failed to default initialize children");
if (this->sum.red != 0) throw std::runtime_error("Compiler failed to default initialize sum");
if (this->avg.red != 0) throw std::runtime_error("Compiler failed to default initialize avg");
}
inline Node* create_child(const size_t level, const size_t depth, unsigned int *leaf_count, Node **reducible_nodes, Pool<Node> &node_pool) {
@ -117,12 +125,22 @@ public:
return c;
}
inline void update_average() {
this->avg.red = (double)this->sum.red / (double)this->pixel_count;
this->avg.green = (double)this->sum.green / (double)this->pixel_count;
this->avg.blue = (double)this->sum.blue / (double)this->pixel_count;
}
void add_color(const uint32_t r, const uint32_t g, const uint32_t b, const size_t depth, const size_t level, unsigned int *leaf_count, Node **reducible_nodes, Pool<Node> &node_pool) {
if (this->is_leaf) {
this->pixel_count++;
this->sum.red += r;
this->sum.green += g;
this->sum.blue += b;
this->update_average();
this->error_sum.red += (r > this->avg.red) ? r - this->avg.red : this->avg.red - r;
this->error_sum.green += (g > this->avg.green) ? g - this->avg.green : this->avg.green - g;
this->error_sum.blue += (b > this->avg.blue) ? b - this->avg.blue : this->avg.blue - b;
} else {
size_t index = get_index(r, g, b, level);
if (this->children[index] == NULL) this->children[index] = this->create_child(level, depth, leaf_count, reducible_nodes, node_pool);
@ -130,30 +148,54 @@ public:
}
}
inline uint64_t total_error() const {
Node *child = NULL;
uint64_t ans = 0;
for (int i = 0; i < MAX_DEPTH; i++) {
if ((child = this->children[i]) != NULL)
ans += child->error_sum.red + child->error_sum.green + child->error_sum.blue;
}
return ans;
}
inline Node* find_best_reducible_node(Node *head) {
uint64_t err = UINT64_MAX,e = 0;
Node *q = head, *ans = head;
while (q != NULL) {
if ((e = q->total_error()) < err) { ans = q; err = e; }
q = q->next_reducible_node;
}
return ans;
}
void reduce(const size_t depth, unsigned int *leaf_count, Node **reducible_nodes, Pool<Node> &node_pool) {
size_t i = 0;
Node *node = NULL;
Node *node = NULL, *child = NULL, *q = NULL;
// Find the deepest level containing at least one reducible node
for (i=depth - 1; i > 0 && reducible_nodes[i] == NULL; i--);
// Reduce the node most recently added to the list at level i
// Could make this smarter by walking the linked list and choosing a
// node that has the least number of pixels or by storing error info
// on the nodes and using that
node = reducible_nodes[i];
reducible_nodes[i] = node->next_reducible_node;
// Find the reducible node at this level that has the least total error
node = find_best_reducible_node(reducible_nodes[i]);
// Remove the found node from the linked list
if (node == reducible_nodes[i]) reducible_nodes[i] = node->next_reducible_node;
else {
q = reducible_nodes[i];
while (q != NULL) {
if (q->next_reducible_node == node) { q->next_reducible_node = node->next_reducible_node; break; }
q = q->next_reducible_node;
}
}
for (i = 0; i < MAX_DEPTH; i++) {
if (node->children[i] != NULL) {
node->sum.red += node->children[i]->sum.red;
node->sum.green += node->children[i]->sum.green;
node->sum.blue += node->children[i]->sum.blue;
if ((child = node->children[i]) != NULL) {
iadd<SumPixel>(node->sum, child->sum);
iadd<SumPixel>(node->error_sum, child->error_sum);
node->pixel_count += node->children[i]->pixel_count;
node_pool.relinquish(node->children[i]); node->children[i] = NULL;
(*leaf_count)--;
}
}
node->update_average();
node->is_leaf = true; *leaf_count += 1;
}