Vulkan compute: move buffer init to before fn call

Change max/min out buffers to have same size as max/min in buffers.

src/blue_noise.hpp

@@ -21,6 +21,7 @@
 #include <random>
 #include <stdexcept>
 #include <thread>
+#include <tuple>
 #include <unordered_set>
 #include <vector>
 
@@ -47,6 +48,12 @@ struct QueueFamilyIndices {
 
 QueueFamilyIndices vulkan_find_queue_families(VkPhysicalDevice device);
 
+struct FloatAndIndex {
+  float value;
+  int pbp;
+  int idx;
+};
+
 std::optional<uint32_t> vulkan_find_memory_type(VkPhysicalDevice phys_dev,
                                                 uint32_t t_filter,
                                                 VkMemoryPropertyFlags props);
@@ -58,45 +65,85 @@ bool vulkan_create_buffer(VkDevice device, VkPhysicalDevice phys_dev,
 
 void vulkan_copy_buffer(VkDevice device, VkCommandPool command_pool,
                         VkQueue queue, VkBuffer src_buf, VkBuffer dst_buf,
-                        VkDeviceSize size);
+                        VkDeviceSize size, VkDeviceSize src_offset = 0,
+                        VkDeviceSize dst_offset = 0);
+void vulkan_copy_buffer_pieces(
+    VkDevice device, VkCommandPool command_pool, VkQueue queue,
+    VkBuffer src_buf, VkBuffer dst_buf,
+    const std::vector<std::tuple<VkDeviceSize, VkDeviceSize> > &pieces);
 
 void vulkan_flush_buffer(VkDevice device, VkDeviceMemory memory);
+void vulkan_flush_buffer_pieces(
+    VkDevice device, const VkDeviceSize phys_atom_size, VkDeviceMemory memory,
+    const std::vector<std::tuple<VkDeviceSize, VkDeviceSize> > &pieces);
 void vulkan_invalidate_buffer(VkDevice device, VkDeviceMemory memory);
 
 std::vector<unsigned int> blue_noise_vulkan_impl(
     VkDevice device, VkPhysicalDevice phys_device,
     VkCommandBuffer command_buffer, VkCommandPool command_pool, VkQueue queue,
     VkBuffer pbp_buf, VkPipeline pipeline, VkPipelineLayout pipeline_layout,
-    VkDescriptorSet descriptor_set, VkBuffer filter_out_buf, const int width,
-    const int height);
+    VkDescriptorSet descriptor_set, VkBuffer filter_out_buf,
+    VkPipeline minmax_pipeline, VkPipelineLayout minmax_pipeline_layout,
+    std::array<VkDescriptorSet, 2> minmax_descriptor_sets, VkBuffer max_in_buf,
+    VkBuffer min_in_buf, VkBuffer max_out_buf, VkBuffer min_out_buf,
+    VkBuffer state_buf, const int width, const int height,
+    VkBuffer minmax_staging_buf, VkDeviceMemory minmax_staging_buf_mem,
+    void *minmax_mapped);
 
 std::vector<float> vulkan_buf_to_vec(float *mapped, unsigned int size);
 
 inline bool vulkan_get_filter(
-    VkDevice device, VkCommandBuffer command_buffer, VkCommandPool command_pool,
-    VkQueue queue, VkBuffer pbp_buf, VkPipeline pipeline,
-    VkPipelineLayout pipeline_layout, VkDescriptorSet descriptor_set,
-    VkBuffer filter_out_buf, const int size, std::vector<bool> &pbp,
-    bool reversed_pbp, const std::size_t global_size, int *pbp_mapped_int,
-    VkBuffer staging_pbp_buffer, VkDeviceMemory staging_pbp_buffer_mem,
-    VkDeviceMemory staging_filter_buffer_mem, VkBuffer staging_filter_buffer) {
+    VkDevice device, const VkDeviceSize phys_atom_size,
+    VkCommandBuffer command_buffer, VkCommandPool command_pool, VkQueue queue,
+    VkBuffer pbp_buf, VkPipeline pipeline, VkPipelineLayout pipeline_layout,
+    VkDescriptorSet descriptor_set, VkBuffer filter_out_buf, const int size,
+    std::vector<bool> &pbp, bool reversed_pbp, const std::size_t global_size,
+    int *pbp_mapped_int, VkBuffer staging_pbp_buffer,
+    VkDeviceMemory staging_pbp_buffer_mem,
+    VkDeviceMemory staging_filter_buffer_mem, VkBuffer staging_filter_buffer,
+    std::vector<std::size_t> *changed) {
   vkResetCommandBuffer(command_buffer, 0);
 
-  if (reversed_pbp) {
-    for (unsigned int i = 0; i < pbp.size(); ++i) {
-      pbp_mapped_int[i] = pbp[i] ? 0 : 1;
+  if (changed != nullptr && changed->size() > 0) {
+    if (reversed_pbp) {
+      for (auto idx : *changed) {
+        pbp_mapped_int[idx] = pbp[idx] ? 0 : 1;
+      }
+    } else {
+      for (auto idx : *changed) {
+        pbp_mapped_int[idx] = pbp[idx] ? 1 : 0;
+      }
     }
   } else {
-    for (unsigned int i = 0; i < pbp.size(); ++i) {
-      pbp_mapped_int[i] = pbp[i] ? 1 : 0;
+    if (reversed_pbp) {
+      for (unsigned int i = 0; i < pbp.size(); ++i) {
+        pbp_mapped_int[i] = pbp[i] ? 0 : 1;
+      }
+    } else {
+      for (unsigned int i = 0; i < pbp.size(); ++i) {
+        pbp_mapped_int[i] = pbp[i] ? 1 : 0;
+      }
     }
   }
 
-  vulkan_flush_buffer(device, staging_pbp_buffer_mem);
-
   // Copy pbp buffer.
-  vulkan_copy_buffer(device, command_pool, queue, staging_pbp_buffer, pbp_buf,
-                     size * sizeof(int));
+  if (changed != nullptr && changed->size() > 0) {
+    std::vector<std::tuple<VkDeviceSize, VkDeviceSize> > pieces;
+    for (auto idx : *changed) {
+      pieces.emplace_back(std::make_tuple(sizeof(int), idx * sizeof(int)));
+    }
+
+    vulkan_flush_buffer_pieces(device, phys_atom_size, staging_pbp_buffer_mem,
+                               pieces);
+
+    vulkan_copy_buffer_pieces(device, command_pool, queue, staging_pbp_buffer,
+                              pbp_buf, pieces);
+    changed->clear();
+  } else {
+    vulkan_flush_buffer(device, staging_pbp_buffer_mem);
+    vulkan_copy_buffer(device, command_pool, queue, staging_pbp_buffer, pbp_buf,
+                       size * sizeof(int));
+  }
 
   VkCommandBufferBeginInfo begin_info{};
   begin_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
@@ -145,6 +192,16 @@ inline bool vulkan_get_filter(
   return true;
 }
 
+std::optional<std::pair<int, int>> vulkan_minmax(
+    VkDevice device, VkPhysicalDevice phys_dev, VkCommandBuffer command_buffer,
+    VkCommandPool command_pool, VkQueue queue, VkPipeline minmax_pipeline,
+    VkPipelineLayout minmax_pipeline_layout,
+    std::array<VkDescriptorSet, 2> minmax_desc_sets, VkBuffer max_in_buf,
+    VkBuffer min_in_buf, VkBuffer max_out_buf, VkBuffer min_out_buf,
+    VkBuffer state_buf, const int size, const float *const filter_mapped,
+    const std::vector<bool> &pbp, VkBuffer staging_buf,
+    VkDeviceMemory staging_buf_mem, void *staging_mapped);
+
 #endif
 
 #if DITHERING_OPENCL_ENABLED == 1

src/blue_noise_minmax.glsl

@@ -0,0 +1,54 @@
+#version 450
+
+struct FloatAndIndex {
+  float value;
+  int pbp;
+  int idx;
+};
+
+layout(binding = 0) readonly buffer MaxIn { FloatAndIndex max_in[]; };
+layout(binding = 1) readonly buffer MinIn { FloatAndIndex min_in[]; };
+layout(binding = 2) writeonly buffer MaxOut { FloatAndIndex max_out[]; };
+layout(binding = 3) writeonly buffer MinOut { FloatAndIndex min_out[]; };
+
+layout(binding = 4) readonly buffer State { int size; };
+
+layout(local_size_x = 256) in;
+
+void main() {
+  uint index = gl_GlobalInvocationID.x;
+  if (index >= (size + 1) / 2) {
+    return;
+  }
+
+  if (index * 2 + 1 < size) {
+    if (max_in[index * 2].pbp != 0 && max_in[index * 2 + 1].pbp != 0) {
+      if (max_in[index * 2].value > max_in[index * 2 + 1].value) {
+        max_out[index] = max_in[index * 2];
+      } else {
+        max_out[index] = max_in[index * 2 + 1];
+      }
+    } else if (max_in[index * 2].pbp != 0 && max_in[index * 2 + 1].pbp == 0) {
+      max_out[index] = max_in[index * 2];
+    } else {
+      max_out[index] = max_in[index * 2 + 1];
+    }
+
+    if (min_in[index * 2].pbp == 0 && min_in[index * 2 + 1].pbp == 0) {
+      if (min_in[index * 2].value < min_in[index * 2 + 1].value) {
+        min_out[index] = min_in[index * 2];
+      } else {
+        min_out[index] = min_in[index * 2 + 1];
+      }
+    } else if (min_in[index * 2].pbp == 0 && min_in[index * 2 + 1].pbp != 0) {
+      min_out[index] = min_in[index * 2];
+    } else {
+      min_out[index] = min_in[index * 2 + 1];
+    }
+  } else {
+    max_out[index] = max_in[index * 2];
+    max_out[index + 1].pbp = 0;
+    min_out[index] = min_in[index * 2];
+    min_out[index + 1].pbp = 1;
+  }
+}