Optimize vector quantization step

This change improves the compression speed for both DXT and ETC encodings.

Explanation:

The vector quantization algorithm takes floating point vectors as input and performs vector preprocessing right before the quantization. At the same time, selector training vectors are generated directly from integer selector values, packed into a single uint64. It would therefore be more efficient to perform preprocessing of the selector training vectors (which includes sorting and deduplication) while still having them in a packed form. Additional performance boost is achieved by using multiple threads for sorting the training vectors.

DXT Testing:

The modified algorithm has been tested on the Kodak test set using 64-bit build with default settings (running on Windows 10, i7-4790, 3.6GHz). All the decompressed test images are identical to the images being compressed and decompressed using original version of Crunch (revision ea9b8d8).

[Compressing Kodak set without mipmaps using DXT1 encoding]
Original: 1582222 bytes / 28.869 sec
Modified: 1468204 bytes / 5.477 sec
Improvement: 7.21% (compression ratio) / 81.03% (compression time)

[Compressing Kodak set with mipmaps using DXT1 encoding]
Original: 2065243 bytes / 36.961 sec
Modified: 1914805 bytes / 7.322 sec
Improvement: 7.28% (compression ratio) / 80.19% (compression time)

ETC Testing:

The modified algorithm has been tested on the Kodak test set using 64-bit build with default settings (running on Windows 10, i7-4790, 3.6GHz). The ETC1 quantization parameters have been selected in such a way, so that ETC1 compression gives approximately the same average Luma PSNR as the corresponding DXT1 compression (which is equal to 34.044 dB for the Kodak test set compressed without mipmaps using DXT1 encoding and default quality settings).

[Compressing Kodak set without mipmaps using ETC1 encoding]
Total size: 1607858 bytes
Total time: 12.766 sec
Average bitrate: 1.363 bpp
Average Luma PSNR: 34.050 dB

Author: alexander-suvorov

bin/crunch_x64.exe

@@ -530,7 +530,7 @@ void dxt_hc::determine_color_endpoint_codebook_task(uint64 data, void*) {
       uint b = blocks[i];
       uint weight = (uint)(math::clamp<uint>(endpoint_weight * m_block_weights[b], 1, 2048) * encoding_weight[m_block_encodings[b]]);
       uint32 selector = 0;
-      for (uint sh = 0, p = 0; p < 16; p++, sh += 2) {
+      for (uint p = 0; p < 16; p++) {
         uint error_best = cUINT32_MAX;
         uint8 s_best = 0;
         for (uint8 t = 0; t < 4; t++) {
@@ -541,9 +541,9 @@ void dxt_hc::determine_color_endpoint_codebook_task(uint64 data, void*) {
             error_best = error;
           }
         }
-        selector |= s_best << sh;
+        selector = selector << 2 | s_best;
       }
-      m_block_selectors[cColor][b] = selector | (uint64)weight << 32;
+      m_block_selectors[cColor][b] = (uint64)selector << 32 | weight;
     }
 
     dxt_endpoint_refiner::params refinerParams;
@@ -609,7 +609,7 @@ void dxt_hc::determine_color_endpoint_codebook_task_etc(uint64 data, void*) {
         uint b = blocks[i];
         uint weight = (uint)(math::clamp<uint>(0x8000 * endpoint_weight * m_block_weights[b] * (m_block_encodings[b] ? 0.972f : 1.0f), 1, 0xFFFF));
         uint32 selector = 0;
-        for (uint sh = 0, p = 0; p < 8; p++, sh += 2) {
+        for (uint p = 0; p < 8; p++) {
           uint error_best = cUINT32_MAX;
           uint8 s_best = 0;
           for (uint8 s = 0; s < 4; s++) {
@@ -619,9 +619,9 @@ void dxt_hc::determine_color_endpoint_codebook_task_etc(uint64 data, void*) {
               error_best = error;
             }
           }
-          selector |= s_best << sh;
+          selector = selector << 2 | s_best;
         }
-        m_block_selectors[cColor][b] = selector | (uint64)weight << 32;
+        m_block_selectors[cColor][b] = (uint64)selector << ((b & 1) ? 32 : 48) | weight;
       }
     }
   }
@@ -663,13 +663,59 @@ void dxt_hc::determine_color_endpoint_clusters_task(uint64 data, void* pData_ptr
 }
 
 void dxt_hc::determine_color_endpoints() {
-  tree_clusterizer<vec6F> vq;
+  uint num_tasks = m_pTask_pool->get_num_threads() + 1;
+  crnlib::vector<std::pair<vec6F, uint> > endpoints;
   for (uint t = 0; t < m_tiles.size(); t++) {
     if (m_tiles[t].pixels.size())
-      vq.add_training_vec(m_tiles[t].color_endpoint, (uint)(m_tiles[t].pixels.size() * m_tiles[t].weight));
+      endpoints.push_back(std::make_pair(m_tiles[t].color_endpoint, (uint)(m_tiles[t].pixels.size() * m_tiles[t].weight)));
   }
 
-  vq.generate_codebook(math::minimum<uint>(m_num_tiles, m_params.m_color_endpoint_codebook_size), true, m_pTask_pool);
+  struct Node {
+    std::pair<vec6F, uint> *p, *pEnd;
+    Node (std::pair<vec6F, uint>* begin, std::pair<vec6F, uint>* end) : p(begin), pEnd(end) {}
+    bool operator<(const Node& other) const { return *p > *other.p; }
+    static void sort_task(uint64 data, void* ptr) { std::sort(((Node*)ptr)->p, ((Node*)ptr)->pEnd); }
+  };
+
+  crnlib::vector<Node> nodes;
+  Node node(0, endpoints.get_ptr());
+  for (uint i = 0; i < num_tasks; i++) {
+    node.p = node.pEnd;
+    node.pEnd = endpoints.get_ptr() + endpoints.size() * (i + 1) / num_tasks;
+    if (node.p != node.pEnd)
+      nodes.push_back(node);
+  }
+
+  for (uint i = 0; i < nodes.size(); i++)
+    m_pTask_pool->queue_task(&Node::sort_task, i, &nodes[i]);
+  m_pTask_pool->join();
+
+  std::priority_queue<Node> queue;
+  for (uint i = 0; i < nodes.size(); i++)
+    queue.push(nodes[i]);
+
+  crnlib::vector<vec6F> vectors;
+  crnlib::vector<uint> weights;
+  vectors.reserve(endpoints.size());
+  weights.reserve(endpoints.size());
+  while (queue.size()) {
+    Node node = queue.top();
+    std::pair<vec6F, uint>* endpoint = node.p++;
+    queue.pop();
+    if (node.p != node.pEnd)
+      queue.push(node);
+    if (!vectors.size() || endpoint->first != vectors.back()) {
+      vectors.push_back(endpoint->first);
+      weights.push_back(endpoint->second);
+    } else if (weights.back() > UINT_MAX - endpoint->second) {
+      weights.back() = UINT_MAX;
+    } else {
+      weights.back() += endpoint->second;
+    }
+  }
+
+  tree_clusterizer<vec6F> vq;
+  vq.generate_codebook(vectors.get_ptr(), weights.get_ptr(), vectors.size(), math::minimum<uint>(m_num_tiles, m_params.m_color_endpoint_codebook_size), true, m_pTask_pool);
   m_color_clusters.resize(vq.get_codebook_size());
 
   for (uint i = 0; i <= m_pTask_pool->get_num_threads(); i++)
@@ -757,7 +803,7 @@ void dxt_hc::determine_alpha_endpoint_codebook_task(uint64 data, void*) {
         uint b = blocks[i];
         uint weight = encoding_weight[m_block_encodings[b]];
         uint64 selector = 0;
-        for (uint sh = 0, p = 0; p < 16; p++, sh += 3) {
+        for (uint p = 0; p < 16; p++) {
           uint error_best = cUINT32_MAX;
           uint8 s_best = 0;
           for (uint8 t = 0; t < 8; t++) {
@@ -769,9 +815,9 @@ void dxt_hc::determine_alpha_endpoint_codebook_task(uint64 data, void*) {
               error_best = error;
             }
           }
-          selector |= (uint64)s_best << sh;
+          selector = selector << 3 | s_best;
         }
-        m_block_selectors[cAlpha0 + a][b] = selector | (uint64)weight << 48;
+        m_block_selectors[cAlpha0 + a][b] = selector << 16 | weight;
       }
     }
 
@@ -823,18 +869,64 @@ void dxt_hc::determine_alpha_endpoint_clusters_task(uint64 data, void* pData_ptr
 }
 
 void dxt_hc::determine_alpha_endpoints() {
-  tree_clusterizer<vec2F> vq;
+  uint num_tasks = m_pTask_pool->get_num_threads() + 1;
+  crnlib::vector<std::pair<vec2F, uint> > endpoints;
   for (uint a = 0; a < m_num_alpha_blocks; a++) {
     for (uint t = 0; t < m_tiles.size(); t++) {
       if (m_tiles[t].pixels.size())
-        vq.add_training_vec(m_tiles[t].alpha_endpoints[a], m_tiles[t].pixels.size());
+        endpoints.push_back(std::make_pair(m_tiles[t].alpha_endpoints[a], m_tiles[t].pixels.size()));
     }
   }
 
-  vq.generate_codebook(math::minimum<uint>(m_num_tiles, m_params.m_alpha_endpoint_codebook_size), false, m_pTask_pool);
+  struct Node {
+    std::pair<vec2F, uint> *p, *pEnd;
+    Node (std::pair<vec2F, uint>* begin, std::pair<vec2F, uint>* end) : p(begin), pEnd(end) {}
+    bool operator<(const Node& other) const { return *p > *other.p; }
+    static void sort_task(uint64 data, void* ptr) { std::sort(((Node*)ptr)->p, ((Node*)ptr)->pEnd); }
+  };
+
+  crnlib::vector<Node> nodes;
+  Node node(0, endpoints.get_ptr());
+  for (uint i = 0; i < num_tasks; i++) {
+    node.p = node.pEnd;
+    node.pEnd = endpoints.get_ptr() + endpoints.size() * (i + 1) / num_tasks;
+    if (node.p != node.pEnd)
+      nodes.push_back(node);
+  }
+
+  for (uint i = 0; i < nodes.size(); i++)
+    m_pTask_pool->queue_task(&Node::sort_task, i, &nodes[i]);
+  m_pTask_pool->join();
+
+  std::priority_queue<Node> queue;
+  for (uint i = 0; i < nodes.size(); i++)
+    queue.push(nodes[i]);
+
+  crnlib::vector<vec2F> vectors;
+  crnlib::vector<uint> weights;
+  vectors.reserve(endpoints.size());
+  weights.reserve(endpoints.size());
+  while (queue.size()) {
+    Node node = queue.top();
+    std::pair<vec2F, uint>* endpoint = node.p++;
+    queue.pop();
+    if (node.p != node.pEnd)
+      queue.push(node);
+    if (!vectors.size() || endpoint->first != vectors.back()) {
+      vectors.push_back(endpoint->first);
+      weights.push_back(endpoint->second);
+    } else if (weights.back() > UINT_MAX - endpoint->second) {
+      weights.back() = UINT_MAX;
+    } else {
+      weights.back() += endpoint->second;
+    }
+  }
+
+  tree_clusterizer<vec2F> vq;
+  vq.generate_codebook(vectors.get_ptr(), weights.get_ptr(), vectors.size(), math::minimum<uint>(m_num_tiles, m_params.m_alpha_endpoint_codebook_size), false, m_pTask_pool);
   m_alpha_clusters.resize(vq.get_codebook_size());
 
-  for (uint i = 0; i <= m_pTask_pool->get_num_threads(); i++)
+  for (uint i = 0; i < num_tasks; i++)
     m_pTask_pool->queue_object_task(this, &dxt_hc::determine_alpha_endpoint_clusters_task, i, &vq);
   m_pTask_pool->join();
 
@@ -859,7 +951,7 @@ void dxt_hc::determine_alpha_endpoints() {
     }
   }
 
-  for (uint i = 0; i <= m_pTask_pool->get_num_threads(); i++)
+  for (uint i = 0; i < num_tasks; i++)
     m_pTask_pool->queue_object_task(this, &dxt_hc::determine_alpha_endpoint_codebook_task, i, NULL);
   m_pTask_pool->join();
 }
@@ -911,16 +1003,68 @@ void dxt_hc::create_color_selector_codebook_task(uint64 data, void* pData_ptr) {
   }
 }
 
+struct SelectorNode {
+  uint64 *p, *pEnd;
+  SelectorNode (uint64* begin, uint64* end) : p(begin), pEnd(end) {}
+  bool operator<(const SelectorNode& other) const { return *p > *other.p; }
+  static void sort_task(uint64 data, void* ptr) { std::sort(((SelectorNode*)ptr)->p, ((SelectorNode*)ptr)->pEnd); }
+};
+
 void dxt_hc::create_color_selector_codebook() {
-  tree_clusterizer<vec16F> selector_vq;
-  vec16F v;
-  for (uint n = m_has_etc_color_blocks ? m_num_blocks >> 1 : m_num_blocks, b = 0; b < n; b++) {
-    uint64 selector = m_has_etc_color_blocks ? m_block_selectors[cColor][b << 1] | m_block_selectors[cColor][b << 1 | 1] << 16 : m_block_selectors[cColor][b];
-    for (uint8 p = 0; p < 16; p++, selector >>= 2)
-      v[p] = ((selector & 3) + 0.5f) * 0.25f;
-    selector_vq.add_training_vec(v, m_has_etc_color_blocks ? (selector & 0xFFFF) + (selector >> 16) : selector);
+  uint num_tasks = m_pTask_pool->get_num_threads() + 1;
+  crnlib::vector<uint64> selectors(m_has_etc_color_blocks ? m_num_blocks >> 1 : m_num_blocks);
+  for (uint i = 0, b = 0, step = m_has_etc_color_blocks ? 2 : 1; b < m_num_blocks; b += step)
+    selectors[i++] = m_block_selectors[cColor][b] + (m_has_etc_color_blocks ? m_block_selectors[cColor][b + 1] : 0);
+
+  crnlib::vector<SelectorNode> nodes;
+  SelectorNode node(0, selectors.get_ptr());
+  for (uint i = 0; i < num_tasks; i++) {
+    node.p = node.pEnd;
+    node.pEnd = selectors.get_ptr() + selectors.size() * (i + 1) / num_tasks;
+    if (node.p != node.pEnd)
+      nodes.push_back(node);
   }
-  selector_vq.generate_codebook(m_params.m_color_selector_codebook_size, false, m_pTask_pool);
+
+  for (uint i = 0; i < nodes.size(); i++)
+    m_pTask_pool->queue_task(&SelectorNode::sort_task, i, &nodes[i]);
+  m_pTask_pool->join();
+
+  std::priority_queue<SelectorNode> queue;
+  for (uint i = 0; i < nodes.size(); i++)
+    queue.push(nodes[i]);
+
+  float v[4];
+  for (uint s = 0; s < 4; s++)
+    v[s] = (s + 0.5f) * 0.25f;
+
+  crnlib::vector<vec16F> vectors;
+  crnlib::vector<uint> weights;
+  vectors.reserve(selectors.size());
+  weights.reserve(selectors.size());
+  for (uint64 prev_selector = 0; queue.size();) {
+    SelectorNode node = queue.top();
+    uint64 selector = *node.p++;
+    queue.pop();
+    if (node.p != node.pEnd)
+      queue.push(node);
+    uint weight = (uint)selector;
+    selector >>= 32;
+    if (!vectors.size() || selector != prev_selector) {
+      prev_selector = selector;
+      vec16F vector;
+      for (uint p = 0; p < 16; p++, selector >>= 2)
+        vector[15 - p] = v[selector & 3];
+      vectors.push_back(vector);
+      weights.push_back(weight);
+    } else if (weights.back() > UINT_MAX - weight) {
+      weights.back() = UINT_MAX;
+    } else {
+      weights.back() += weight;
+    }
+  }
+
+  tree_clusterizer<vec16F> selector_vq;
+  selector_vq.generate_codebook(vectors.get_ptr(), weights.get_ptr(), vectors.size(), m_params.m_color_selector_codebook_size, false, m_pTask_pool);
   m_color_selectors.resize(selector_vq.get_codebook_size());
   m_color_selectors_used.resize(selector_vq.get_codebook_size());
   for (uint i = 0; i < selector_vq.get_codebook_size(); i++) {
@@ -930,7 +1074,6 @@ void dxt_hc::create_color_selector_codebook() {
       m_color_selectors[i] |= (uint)(v[j] * 4.0f) << sh;
   }
 
-  uint num_tasks = m_pTask_pool->get_num_threads() + 1;
   crnlib::vector<crnlib::vector<color_selector_details> > selector_details(num_tasks);
   for (uint t = 0; t < num_tasks; t++) {
     selector_details[t].resize(m_color_selectors.size());
@@ -1024,17 +1167,62 @@ void dxt_hc::create_alpha_selector_codebook_task(uint64 data, void* pData_ptr) {
 }
 
 void dxt_hc::create_alpha_selector_codebook() {
-  tree_clusterizer<vec16F> selector_vq;
-  vec16F v;
-  for (uint c = cAlpha0; c < cAlpha0 + m_num_alpha_blocks; c++) {
-    for (uint b = 0; b < m_num_blocks; b += m_has_etc_color_blocks ? 2 : 1) {
-      uint64 selector = m_block_selectors[c][b];
-      for (uint8 p = 0; p < 16; p++, selector >>= 3)
-        v[p] = ((selector & 7) + 0.5f) * 0.125f;
-      selector_vq.add_training_vec(v, selector);
+  uint num_tasks = m_pTask_pool->get_num_threads() + 1;
+  crnlib::vector<uint64> selectors(m_num_alpha_blocks * (m_has_etc_color_blocks ? m_num_blocks >> 1 : m_num_blocks));
+  for (uint i = 0, c = cAlpha0; c < cAlpha0 + m_num_alpha_blocks; c++) {
+    for (uint b = 0, step = m_has_etc_color_blocks ? 2 : 1; b < m_num_blocks; b += step)
+      selectors[i++] = m_block_selectors[c][b];
+  }
+
+  crnlib::vector<SelectorNode> nodes;
+  SelectorNode node(0, selectors.get_ptr());
+  for (uint i = 0; i < num_tasks; i++) {
+    node.p = node.pEnd;
+    node.pEnd = selectors.get_ptr() + selectors.size() * (i + 1) / num_tasks;
+    if (node.p != node.pEnd)
+      nodes.push_back(node);
+  }
+
+  for (uint i = 0; i < nodes.size(); i++)
+    m_pTask_pool->queue_task(&SelectorNode::sort_task, i, &nodes[i]);
+  m_pTask_pool->join();
+
+  std::priority_queue<SelectorNode> queue;
+  for (uint i = 0; i < nodes.size(); i++)
+    queue.push(nodes[i]);
+
+  float v[8];
+  for (uint s = 0; s < 8; s++)
+    v[s] = (s + 0.5f) * 0.125f;
+
+  crnlib::vector<vec16F> vectors;
+  crnlib::vector<uint> weights;
+  vectors.reserve(selectors.size());
+  weights.reserve(selectors.size());
+  for (uint64 prev_selector = 0; queue.size();) {
+    SelectorNode node = queue.top();
+    uint64 selector = *node.p++;
+    queue.pop();
+    if (node.p != node.pEnd)
+      queue.push(node);
+    uint weight = (uint16)selector;
+    selector >>= 16;
+    if (!vectors.size() || selector != prev_selector) {
+      prev_selector = selector;
+      vec16F vector;
+      for (uint p = 0; p < 16; p++, selector >>= 3)
+        vector[15 - p] = v[selector & 7];
+      vectors.push_back(vector);
+      weights.push_back(weight);
+    } else if (weights.back() > UINT_MAX - weight) {
+      weights.back() = UINT_MAX;
+    } else {
+      weights.back() += weight;
     }
   }
-  selector_vq.generate_codebook(m_params.m_alpha_selector_codebook_size, false, m_pTask_pool);
+
+  tree_clusterizer<vec16F> selector_vq;
+  selector_vq.generate_codebook(vectors.get_ptr(), weights.get_ptr(), vectors.size(), m_params.m_alpha_selector_codebook_size, false, m_pTask_pool);
   m_alpha_selectors.resize(selector_vq.get_codebook_size());
   m_alpha_selectors_used.resize(selector_vq.get_codebook_size());
   for (uint i = 0; i < selector_vq.get_codebook_size(); i++) {
@@ -1044,7 +1232,6 @@ void dxt_hc::create_alpha_selector_codebook() {
       m_alpha_selectors[i] |= (uint64)(v[j] * 8.0f) << sh;
   }
 
-  uint num_tasks = m_pTask_pool->get_num_threads() + 1;
   crnlib::vector<crnlib::vector<alpha_selector_details> > selector_details(num_tasks);
   for (uint t = 0; t < num_tasks; t++) {
     selector_details[t].resize(m_alpha_selectors.size());