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.cpp

@@ -161,6 +161,49 @@ void dither::internal::vulkan_copy_buffer(VkDevice device,
   vkFreeCommandBuffers(device, command_pool, 1, &command_buf);
 }
 
+void dither::internal::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) {
+  VkCommandBufferAllocateInfo alloc_info{};
+  alloc_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
+  alloc_info.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
+  alloc_info.commandPool = command_pool;
+  alloc_info.commandBufferCount = 1;
+
+  VkCommandBuffer command_buf;
+  vkAllocateCommandBuffers(device, &alloc_info, &command_buf);
+
+  VkCommandBufferBeginInfo begin_info{};
+  begin_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
+  begin_info.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
+
+  vkBeginCommandBuffer(command_buf, &begin_info);
+
+  std::vector<VkBufferCopy> regions;
+  for (auto tuple : pieces) {
+    VkBufferCopy copy_region{};
+    copy_region.size = std::get<0>(tuple);
+    copy_region.srcOffset = std::get<1>(tuple);
+    copy_region.dstOffset = std::get<1>(tuple);
+    regions.push_back(copy_region);
+  }
+  vkCmdCopyBuffer(command_buf, src_buf, dst_buf, regions.size(),
+                  regions.data());
+
+  vkEndCommandBuffer(command_buf);
+
+  VkSubmitInfo submit_info{};
+  submit_info.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
+  submit_info.commandBufferCount = 1;
+  submit_info.pCommandBuffers = &command_buf;
+
+  vkQueueSubmit(queue, 1, &submit_info, VK_NULL_HANDLE);
+  vkQueueWaitIdle(queue);
+
+  vkFreeCommandBuffers(device, command_pool, 1, &command_buf);
+}
+
 void dither::internal::vulkan_flush_buffer(VkDevice device,
                                            VkDeviceMemory memory) {
   VkMappedMemoryRange range{};
@@ -175,6 +218,40 @@ void dither::internal::vulkan_flush_buffer(VkDevice device,
   }
 }
 
+void dither::internal::vulkan_flush_buffer_pieces(
+    VkDevice device, const VkDeviceSize phys_atom_size, VkDeviceMemory memory,
+    const std::vector<std::tuple<VkDeviceSize, VkDeviceSize>> &pieces) {
+  std::vector<VkMappedMemoryRange> ranges;
+  for (auto tuple : pieces) {
+    VkMappedMemoryRange range{};
+    range.sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
+    range.pNext = nullptr;
+    range.memory = memory;
+    range.offset = std::get<1>(tuple);
+    range.size = std::get<0>(tuple);
+
+    // TODO dynamically handle multiple pieces for more efficient flushes.
+    // This may not be necessary if pieces is always size 1.
+
+    if (range.offset % phys_atom_size != 0) {
+      range.offset = (range.offset / phys_atom_size) * phys_atom_size;
+    }
+
+    if (range.size < phys_atom_size) {
+      range.size = phys_atom_size;
+    } else if (range.size % phys_atom_size != 0) {
+      range.size = (range.size / phys_atom_size) * phys_atom_size;
+    }
+
+    ranges.push_back(range);
+  }
+
+  if (vkFlushMappedMemoryRanges(device, ranges.size(), ranges.data()) !=
+      VK_SUCCESS) {
+    std::clog << "WARNING: vulkan_flush_buffer failed!\n";
+  }
+}
+
 void dither::internal::vulkan_invalidate_buffer(VkDevice device,
                                                 VkDeviceMemory memory) {
   VkMappedMemoryRange range{};
@@ -267,6 +344,15 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
       &staging_filter_buffer_mem);
   float *filter_mapped_float = (float *)filter_mapped;
 
+  std::vector<std::size_t> changed_indices;
+
+  VkDeviceSize phys_atom_size;
+  {
+    VkPhysicalDeviceProperties props;
+    vkGetPhysicalDeviceProperties(phys_device, &props);
+    phys_atom_size = props.limits.nonCoherentAtomSize;
+  }
+
   {
 #ifndef NDEBUG
     printf("Inserting %d pixels into image of max count %d\n", pixel_count,
@@ -285,12 +371,12 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 #endif
   }
 
-  if (!vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+  if (!vulkan_get_filter(
-                         pipeline, pipeline_layout, descriptor_set,
+          device, phys_atom_size, command_buffer, command_pool, queue, pbp_buf,
-                         filter_out_buf, size, pbp, reversed_pbp, global_size,
+          pipeline, pipeline_layout, descriptor_set, filter_out_buf, size, pbp,
-                         pbp_mapped_int, staging_pbp_buffer,
+          reversed_pbp, global_size, pbp_mapped_int, staging_pbp_buffer,
           staging_pbp_buffer_mem, staging_filter_buffer_mem,
-                         staging_filter_buffer)) {
+          staging_filter_buffer, nullptr)) {
     std::cerr << "Vulkan: Failed to execute get_filter at start!\n";
   } else {
 #ifndef NDEBUG
@@ -309,12 +395,13 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
     printf("Iteration %d\n", ++iterations);
 #endif
 
-    if (!vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+    if (!vulkan_get_filter(device, phys_atom_size, command_buffer, command_pool,
-                           pipeline, pipeline_layout, descriptor_set,
+                           queue, pbp_buf, pipeline, pipeline_layout,
-                           filter_out_buf, size, pbp, reversed_pbp, global_size,
+                           descriptor_set, filter_out_buf, size, pbp,
-                           pbp_mapped_int, staging_pbp_buffer,
+                           reversed_pbp, global_size, pbp_mapped_int,
-                           staging_pbp_buffer_mem, staging_filter_buffer_mem,
+                           staging_pbp_buffer, staging_pbp_buffer_mem,
-                           staging_filter_buffer)) {
+                           staging_filter_buffer_mem, staging_filter_buffer,
+                           &changed_indices)) {
       std::cerr << "Vulkan: Failed to execute do_filter\n";
       break;
     }
@@ -343,12 +430,15 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 
     pbp[max] = false;
 
-    if (!vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+    changed_indices.push_back(max);
-                           pipeline, pipeline_layout, descriptor_set,
+
-                           filter_out_buf, size, pbp, reversed_pbp, global_size,
+    if (!vulkan_get_filter(device, phys_atom_size, command_buffer, command_pool,
-                           pbp_mapped_int, staging_pbp_buffer,
+                           queue, pbp_buf, pipeline, pipeline_layout,
-                           staging_pbp_buffer_mem, staging_filter_buffer_mem,
+                           descriptor_set, filter_out_buf, size, pbp,
-                           staging_filter_buffer)) {
+                           reversed_pbp, global_size, pbp_mapped_int,
+                           staging_pbp_buffer, staging_pbp_buffer_mem,
+                           staging_filter_buffer_mem, staging_filter_buffer,
+                           &changed_indices)) {
       std::cerr << "Vulkan: Failed to execute do_filter\n";
       break;
     }
@@ -369,9 +459,11 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 
     if (second_min == max) {
       pbp[max] = true;
+      changed_indices.push_back(max);
       break;
     } else {
       pbp[second_min] = true;
+      changed_indices.push_back(second_min);
     }
 
 #ifndef NDEBUG
@@ -393,12 +485,12 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 #endif
   }
 
-  if (!vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+  if (!vulkan_get_filter(
-                         pipeline, pipeline_layout, descriptor_set,
+          device, phys_atom_size, command_buffer, command_pool, queue, pbp_buf,
-                         filter_out_buf, size, pbp, reversed_pbp, global_size,
+          pipeline, pipeline_layout, descriptor_set, filter_out_buf, size, pbp,
-                         pbp_mapped_int, staging_pbp_buffer,
+          reversed_pbp, global_size, pbp_mapped_int, staging_pbp_buffer,
           staging_pbp_buffer_mem, staging_filter_buffer_mem,
-                         staging_filter_buffer)) {
+          staging_filter_buffer, &changed_indices)) {
     std::cerr << "Vulkan: Failed to execute do_filter (at end)\n";
   } else {
 #ifndef NDEBUG
@@ -437,12 +529,12 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 #ifndef NDEBUG
       std::cout << i << ' ';
 #endif
-      vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+      vulkan_get_filter(device, phys_atom_size, command_buffer, command_pool,
-                        pipeline, pipeline_layout, descriptor_set,
+                        queue, pbp_buf, pipeline, pipeline_layout,
-                        filter_out_buf, size, pbp, reversed_pbp, global_size,
+                        descriptor_set, filter_out_buf, size, pbp, reversed_pbp,
-                        pbp_mapped_int, staging_pbp_buffer,
+                        global_size, pbp_mapped_int, staging_pbp_buffer,
                         staging_pbp_buffer_mem, staging_filter_buffer_mem,
-                        staging_filter_buffer);
+                        staging_filter_buffer, &changed_indices);
       auto vulkan_minmax_opt = vulkan_minmax(
           device, phys_device, command_buffer, command_pool, queue,
           minmax_pipeline, minmax_pipeline_layout, minmax_desc_sets, max_in_buf,
@@ -456,6 +548,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
       std::tie(std::ignore, max) = vulkan_minmax_opt.value();
       pbp.at(max) = false;
       dither_array.at(max) = i;
+      changed_indices.push_back(max);
 #ifndef NDEBUG
       if (set.find(max) != set.end()) {
         std::cout << "\nWARNING: Reusing index " << max << '\n';
@@ -475,11 +568,12 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 #ifndef NDEBUG
     std::cout << i << ' ';
 #endif
-    vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+    vulkan_get_filter(device, phys_atom_size, command_buffer, command_pool,
-                      pipeline, pipeline_layout, descriptor_set, filter_out_buf,
+                      queue, pbp_buf, pipeline, pipeline_layout, descriptor_set,
-                      size, pbp, reversed_pbp, global_size, pbp_mapped_int,
+                      filter_out_buf, size, pbp, reversed_pbp, global_size,
-                      staging_pbp_buffer, staging_pbp_buffer_mem,
+                      pbp_mapped_int, staging_pbp_buffer,
-                      staging_filter_buffer_mem, staging_filter_buffer);
+                      staging_pbp_buffer_mem, staging_filter_buffer_mem,
+                      staging_filter_buffer, &changed_indices);
     auto vulkan_minmax_opt = vulkan_minmax(
         device, phys_device, command_buffer, command_pool, queue,
         minmax_pipeline, minmax_pipeline_layout, minmax_desc_sets, max_in_buf,
@@ -493,6 +587,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
     std::tie(min, std::ignore) = vulkan_minmax_opt.value();
     pbp.at(min) = true;
     dither_array.at(min) = i;
+    changed_indices.push_back(min);
 #ifndef NDEBUG
     if (set.find(min) != set.end()) {
       std::cout << "\nWARNING: Reusing index " << min << '\n';
@@ -505,11 +600,12 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
   {
     image::Bl min_pixels = internal::rangeToBl(dither_array, width);
     min_pixels.writeToFile(image::file_type::PNG, true, "da_mid_pixels.png");
-    vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+    vulkan_get_filter(device, phys_atom_size, command_buffer, command_pool,
-                      pipeline, pipeline_layout, descriptor_set, filter_out_buf,
+                      queue, pbp_buf, pipeline, pipeline_layout, descriptor_set,
-                      size, pbp, reversed_pbp, global_size, pbp_mapped_int,
+                      filter_out_buf, size, pbp, reversed_pbp, global_size,
-                      staging_pbp_buffer, staging_pbp_buffer_mem,
+                      pbp_mapped_int, staging_pbp_buffer,
-                      staging_filter_buffer_mem, staging_filter_buffer);
+                      staging_pbp_buffer_mem, staging_filter_buffer_mem,
+                      staging_filter_buffer, &changed_indices);
     internal::write_filter(vulkan_buf_to_vec(filter_mapped_float, size), width,
                            "filter_mid.pgm");
     image::Bl pbp_image = toBl(pbp, width);
@@ -518,15 +614,21 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 #endif
   std::cout << "\nRanking last half of pixels...\n";
   reversed_pbp = true;
+  bool first_reversed_run = true;
   for (unsigned int i = (size + 1) / 2; i < (unsigned int)size; ++i) {
 #ifndef NDEBUG
     std::cout << i << ' ';
 #endif
-    vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+    if (first_reversed_run) {
-                      pipeline, pipeline_layout, descriptor_set, filter_out_buf,
+      changed_indices.clear();
-                      size, pbp, reversed_pbp, global_size, pbp_mapped_int,
+      first_reversed_run = false;
-                      staging_pbp_buffer, staging_pbp_buffer_mem,
+    }
-                      staging_filter_buffer_mem, staging_filter_buffer);
+    vulkan_get_filter(device, phys_atom_size, command_buffer, command_pool,
+                      queue, pbp_buf, pipeline, pipeline_layout, descriptor_set,
+                      filter_out_buf, size, pbp, reversed_pbp, global_size,
+                      pbp_mapped_int, staging_pbp_buffer,
+                      staging_pbp_buffer_mem, staging_filter_buffer_mem,
+                      staging_filter_buffer, &changed_indices);
     auto vulkan_minmax_opt = vulkan_minmax(
         device, phys_device, command_buffer, command_pool, queue,
         minmax_pipeline, minmax_pipeline_layout, minmax_desc_sets, max_in_buf,
@@ -540,6 +642,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
     std::tie(std::ignore, max) = vulkan_minmax_opt.value();
     pbp.at(max) = true;
     dither_array.at(max) = i;
+    changed_indices.push_back(max);
 #ifndef NDEBUG
     if (set.find(max) != set.end()) {
       std::cout << "\nWARNING: Reusing index " << max << '\n';
@@ -552,11 +655,12 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 
 #ifndef NDEBUG
   {
-    vulkan_get_filter(device, command_buffer, command_pool, queue, pbp_buf,
+    vulkan_get_filter(device, phys_atom_size, command_buffer, command_pool,
-                      pipeline, pipeline_layout, descriptor_set, filter_out_buf,
+                      queue, pbp_buf, pipeline, pipeline_layout, descriptor_set,
-                      size, pbp, reversed_pbp, global_size, pbp_mapped_int,
+                      filter_out_buf, size, pbp, reversed_pbp, global_size,
-                      staging_pbp_buffer, staging_pbp_buffer_mem,
+                      pbp_mapped_int, staging_pbp_buffer,
-                      staging_filter_buffer_mem, staging_filter_buffer);
+                      staging_pbp_buffer_mem, staging_filter_buffer_mem,
+                      staging_filter_buffer, nullptr);
     internal::write_filter(vulkan_buf_to_vec(filter_mapped_float, size), width,
                            "filter_after.pgm");
     image::Bl pbp_image = toBl(pbp, width);

src/blue_noise.hpp

@@ -21,6 +21,7 @@
 #include <random>
 #include <stdexcept>
 #include <thread>
+#include <tuple>
 #include <unordered_set>
 #include <vector>
 
@@ -66,8 +67,15 @@ void vulkan_copy_buffer(VkDevice device, VkCommandPool command_pool,
                         VkQueue queue, VkBuffer src_buf, VkBuffer dst_buf,
                         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(
@@ -85,15 +93,28 @@ std::vector<unsigned int> blue_noise_vulkan_impl(
 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,
+    VkDevice device, const VkDeviceSize phys_atom_size,
-    VkQueue queue, VkBuffer pbp_buf, VkPipeline pipeline,
+    VkCommandBuffer command_buffer, VkCommandPool command_pool, VkQueue queue,
-    VkPipelineLayout pipeline_layout, VkDescriptorSet descriptor_set,
+    VkBuffer pbp_buf, VkPipeline pipeline, VkPipelineLayout pipeline_layout,
-    VkBuffer filter_out_buf, const int size, std::vector<bool> &pbp,
+    VkDescriptorSet descriptor_set, VkBuffer filter_out_buf, const int size,
-    bool reversed_pbp, const std::size_t global_size, int *pbp_mapped_int,
+    std::vector<bool> &pbp, bool reversed_pbp, const std::size_t global_size,
-    VkBuffer staging_pbp_buffer, VkDeviceMemory staging_pbp_buffer_mem,
+    int *pbp_mapped_int, VkBuffer staging_pbp_buffer,
-    VkDeviceMemory staging_filter_buffer_mem, VkBuffer staging_filter_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 (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 {
     if (reversed_pbp) {
       for (unsigned int i = 0; i < pbp.size(); ++i) {
         pbp_mapped_int[i] = pbp[i] ? 0 : 1;
@@ -103,12 +124,26 @@ inline bool vulkan_get_filter(
         pbp_mapped_int[i] = pbp[i] ? 1 : 0;
       }
     }
-
+  }
-  vulkan_flush_buffer(device, staging_pbp_buffer_mem);
 
   // Copy pbp buffer.
+  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;