Vulkan compute: move buffer init to before fn call

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

README.md

@@ -1,4 +1,5 @@
 # Blue_Noise_Generation
 
 Currently, this project only generates blue-noise, which can be used for dithering.  
-The blue-noise can be generated via OpenCL or with threads on the CPU.
+The blue-noise can be generated via OpenCL, Vulkan compute, or with threads on
+the CPU.

src/blue_noise.hpp

@@ -47,6 +47,12 @@ 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);
@@ -58,7 +64,8 @@ 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_flush_buffer(VkDevice device, VkDeviceMemory memory);
 void vulkan_invalidate_buffer(VkDevice device, VkDeviceMemory memory);
@@ -67,8 +74,13 @@ 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,
+    VkDescriptorSet descriptor_set, VkBuffer filter_out_buf,
-    const int height);
+    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);
 
@@ -145,6 +157,16 @@ 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;
+  }
+}