Cache vulkan buffers, remove staging buffers

At this point, runtime is almost the same as the OpenCL implementation.

src/blue_noise.cpp

@@ -157,12 +157,40 @@ void dither::internal::vulkan_copy_buffer(VkDevice device,
   vkFreeCommandBuffers(device, command_pool, 1, &command_buf);
 }
 
+void dither::internal::vulkan_flush_buffer(VkDevice device,
+                                           VkDeviceMemory memory) {
+  VkMappedMemoryRange range{};
+  range.sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
+  range.pNext = nullptr;
+  range.memory = memory;
+  range.offset = 0;
+  range.size = VK_WHOLE_SIZE;
+
+  if (vkFlushMappedMemoryRanges(device, 1, &range) != VK_SUCCESS) {
+    std::clog << "WARNING: vulkan_flush_buffer failed!\n";
+  }
+}
+
+void dither::internal::vulkan_invalidate_buffer(VkDevice device,
+                                                VkDeviceMemory memory) {
+  VkMappedMemoryRange range{};
+  range.sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
+  range.pNext = nullptr;
+  range.memory = memory;
+  range.offset = 0;
+  range.size = VK_WHOLE_SIZE;
+
+  if (vkInvalidateMappedMemoryRanges(device, 1, &range) != VK_SUCCESS) {
+    std::clog << "WARNING: vulkan_invalidate_buffer failed!\n";
+  }
+}
+
 std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
-    VkDevice device, VkPhysicalDevice phys_device,
+    VkDevice device, VkCommandBuffer command_buffer, VkQueue queue,
-    VkCommandBuffer command_buffer, VkCommandPool command_pool, VkQueue queue,
+    VkPipeline pipeline, VkPipelineLayout pipeline_layout,
-    VkBuffer pbp_buf, VkPipeline pipeline, VkPipelineLayout pipeline_layout,
+    VkDescriptorSet descriptor_set, float *filter_out_mapped, int *pbp_mapped,
-    VkDescriptorSet descriptor_set, VkBuffer filter_out_buf, const int width,
+    VkDeviceMemory filter_out_buf_mem, VkDeviceMemory pbp_buf_mem,
-    const int height) {
+    const int width, const int height) {
   const int size = width * height;
   const int pixel_count = size * 4 / 10;
   const int local_size = 256;
@@ -170,59 +198,24 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
       (std::size_t)std::ceil((float)size / (float)local_size);
 
   std::vector<bool> pbp = random_noise(size, pixel_count);
-  std::vector<int> pbp_i(pbp.size());
-  std::vector<float> filter(size);
   bool reversed_pbp = false;
 
-  const auto get_filter = [device, phys_device, command_buffer, command_pool,
+  const auto get_filter = [device, command_buffer, queue, pipeline,
-                           queue, pbp_buf, pipeline, pipeline_layout,
+                           pipeline_layout, descriptor_set, &pbp, &reversed_pbp,
-                           descriptor_set, filter_out_buf, size, &pbp, &pbp_i,
+                           global_size, filter_out_buf_mem, pbp_buf_mem,
-                           &reversed_pbp, global_size, &filter]() -> bool {
+                           pbp_mapped]() -> bool {
-    for (unsigned int i = 0; i < pbp.size(); ++i) {
-      if (reversed_pbp) {
-        pbp_i[i] = pbp[i] ? 0 : 1;
-      } else {
-        pbp_i[i] = pbp[i] ? 1 : 0;
-      }
-    }
-
     vkResetCommandBuffer(command_buffer, 0);
 
-    // Copy pbp buffer.
+    for (unsigned int i = 0; i < pbp.size(); ++i) {
-    {
+      if (reversed_pbp) {
-      VkBuffer staging_buffer;
+        pbp_mapped[i] = pbp[i] ? 0 : 1;
-      VkDeviceMemory staging_buffer_mem;
+      } else {
-
+        pbp_mapped[i] = pbp[i] ? 1 : 0;
-      if (!internal::vulkan_create_buffer(
-              device, phys_device, size * sizeof(int),
-              VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
-              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                  VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
-              staging_buffer, staging_buffer_mem)) {
-        std::clog << "get_filter ERROR: Failed to create staging buffer!\n";
-        return false;
       }
-      utility::Cleanup cleanup_staging_buf(
-          [device](void *ptr) {
-            vkDestroyBuffer(device, *((VkBuffer *)ptr), nullptr);
-          },
-          &staging_buffer);
-      utility::Cleanup cleanup_staging_buf_mem(
-          [device](void *ptr) {
-            vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr);
-          },
-          &staging_buffer_mem);
-
-      void *data_ptr;
-      vkMapMemory(device, staging_buffer_mem, 0, size * sizeof(int), 0,
-                  &data_ptr);
-      std::memcpy(data_ptr, pbp_i.data(), size * sizeof(int));
-      vkUnmapMemory(device, staging_buffer_mem);
-
-      vulkan_copy_buffer(device, command_pool, queue, staging_buffer, pbp_buf,
-                         size * sizeof(int));
     }
 
+    vulkan_flush_buffer(device, pbp_buf_mem);
+
     VkCommandBufferBeginInfo begin_info{};
     begin_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
 
@@ -262,40 +255,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
       return false;
     }
 
-    // Copy back filter_out buffer.
+    vulkan_flush_buffer(device, filter_out_buf_mem);
-    {
-      VkBuffer staging_buffer;
-      VkDeviceMemory staging_buffer_mem;
-
-      if (!internal::vulkan_create_buffer(
-              device, phys_device, size * sizeof(float),
-              VK_BUFFER_USAGE_TRANSFER_DST_BIT,
-              VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                  VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
-              staging_buffer, staging_buffer_mem)) {
-        std::clog << "get_filter ERROR: Failed to create staging buffer!\n";
-        return false;
-      }
-      utility::Cleanup cleanup_staging_buf(
-          [device](void *ptr) {
-            vkDestroyBuffer(device, *((VkBuffer *)ptr), nullptr);
-          },
-          &staging_buffer);
-      utility::Cleanup cleanup_staging_buf_mem(
-          [device](void *ptr) {
-            vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr);
-          },
-          &staging_buffer_mem);
-
-      vulkan_copy_buffer(device, command_pool, queue, filter_out_buf,
-                         staging_buffer, size * sizeof(float));
-
-      void *data_ptr;
-      vkMapMemory(device, staging_buffer_mem, 0, size * sizeof(float), 0,
-                  &data_ptr);
-      std::memcpy(filter.data(), data_ptr, size * sizeof(float));
-      vkUnmapMemory(device, staging_buffer_mem);
-    }
 
     return true;
   };
@@ -322,7 +282,8 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
     std::cerr << "Vulkan: Failed to execute get_filter at start!\n";
   } else {
 #ifndef NDEBUG
-    internal::write_filter(filter, width, "filter_out_start.pgm");
+    internal::write_filter(vulkan_buf_to_vec(filter_out_mapped, size), width,
+                           "filter_out_start.pgm");
 #endif
   }
 
@@ -340,7 +301,8 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
     }
 
     int min, max;
-    std::tie(min, max) = internal::filter_minmax(filter, pbp);
+    std::tie(min, max) =
+        internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
 
     pbp[max] = false;
 
@@ -351,7 +313,8 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 
     // get second buffer's min
     int second_min;
-    std::tie(second_min, std::ignore) = internal::filter_minmax(filter, pbp);
+    std::tie(second_min, std::ignore) =
+        internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
 
     if (second_min == max) {
       pbp[max] = true;
@@ -383,7 +346,8 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
     std::cerr << "Vulkan: Failed to execute do_filter (at end)\n";
   } else {
 #ifndef NDEBUG
-    internal::write_filter(filter, width, "filter_out_final.pgm");
+    internal::write_filter(vulkan_buf_to_vec(filter_out_mapped, size), width,
+                           "filter_out_final.pgm");
     FILE *blue_noise_image = fopen("blue_noise.pbm", "w");
     fprintf(blue_noise_image, "P1\n%d %d\n", width, height);
     for (int y = 0; y < height; ++y) {
@@ -418,7 +382,8 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
       std::cout << i << ' ';
 #endif
       get_filter();
-      std::tie(std::ignore, max) = internal::filter_minmax(filter, pbp);
+      std::tie(std::ignore, max) =
+          internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
       pbp.at(max) = false;
       dither_array.at(max) = i;
 #ifndef NDEBUG
@@ -441,7 +406,8 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
     std::cout << i << ' ';
 #endif
     get_filter();
-    std::tie(min, std::ignore) = internal::filter_minmax(filter, pbp);
+    std::tie(min, std::ignore) =
+        internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
     pbp.at(min) = true;
     dither_array.at(min) = i;
 #ifndef NDEBUG
@@ -457,7 +423,8 @@ 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");
     get_filter();
-    internal::write_filter(filter, width, "filter_mid.pgm");
+    internal::write_filter(vulkan_buf_to_vec(filter_out_mapped, size), width,
+                           "filter_mid.pgm");
     image::Bl pbp_image = toBl(pbp, width);
     pbp_image.writeToFile(image::file_type::PNG, true, "debug_pbp_mid.png");
   }
@@ -469,7 +436,8 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
     std::cout << i << ' ';
 #endif
     get_filter();
-    std::tie(std::ignore, max) = internal::filter_minmax(filter, pbp);
+    std::tie(std::ignore, max) =
+        internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
     pbp.at(max) = true;
     dither_array.at(max) = i;
 #ifndef NDEBUG
@@ -485,7 +453,8 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
 #ifndef NDEBUG
   {
     get_filter();
-    internal::write_filter(filter, width, "filter_after.pgm");
+    internal::write_filter(vulkan_buf_to_vec(filter_out_mapped, size), width,
+                           "filter_after.pgm");
     image::Bl pbp_image = toBl(pbp, width);
     pbp_image.writeToFile(image::file_type::PNG, true, "debug_pbp_after.png");
   }
@@ -494,6 +463,15 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
   return dither_array;
 }
 
+std::vector<float> dither::internal::vulkan_buf_to_vec(float *mapped,
+                                                       unsigned int size) {
+  std::vector<float> v(size);
+
+  std::memcpy(v.data(), mapped, size * sizeof(float));
+
+  return v;
+}
+
 #endif  // DITHERING_VULKAN_ENABLED == 1
 
 #include "image.hpp"
@@ -964,11 +942,13 @@ image::Bl dither::blue_noise(int width, int height, int threads,
 
     VkBuffer filter_out_buf;
     VkDeviceMemory filter_out_buf_mem;
-    if (!internal::vulkan_create_buffer(device, phys_device, filter_out_size,
+    if (!internal::vulkan_create_buffer(
-                                        VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
+            device, phys_device, filter_out_size,
-                                            VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
+            VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
-                                        VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
+            VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
-                                        filter_out_buf, filter_out_buf_mem)) {
+                VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+                VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
+            filter_out_buf, filter_out_buf_mem)) {
       std::clog << "WARNING: Failed to create filter_out buffer!\n";
       goto ENDOF_VULKAN;
     }
@@ -982,14 +962,24 @@ image::Bl dither::blue_noise(int width, int height, int threads,
           vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr);
         },
         &filter_out_buf_mem);
+    void *filter_out_mapped;
+    vkMapMemory(device, filter_out_buf_mem, 0, filter_out_size, 0,
+                &filter_out_mapped);
+    utility::Cleanup cleanup_filter_out_mapped(
+        [device](void *ptr) {
+          vkUnmapMemory(device, *((VkDeviceMemory *)ptr));
+        },
+        &filter_out_buf_mem);
+    float *filter_out_mapped_float = (float *)filter_out_mapped;
 
     VkBuffer pbp_buf;
     VkDeviceMemory pbp_buf_mem;
-    if (!internal::vulkan_create_buffer(device, phys_device, pbp_size,
+    if (!internal::vulkan_create_buffer(
-                                        VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
+            device, phys_device, pbp_size, VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
-                                            VK_BUFFER_USAGE_TRANSFER_DST_BIT,
+            VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
-                                        VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
+                VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                                        pbp_buf, pbp_buf_mem)) {
+                VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
+            pbp_buf, pbp_buf_mem)) {
       std::clog << "WARNING: Failed to create pbp buffer!\n";
       goto ENDOF_VULKAN;
     }
@@ -1003,6 +993,14 @@ image::Bl dither::blue_noise(int width, int height, int threads,
           vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr);
         },
         &pbp_buf_mem);
+    void *pbp_mapped;
+    vkMapMemory(device, pbp_buf_mem, 0, pbp_size, 0, &pbp_mapped);
+    utility::Cleanup cleanup_pbp_mapped(
+        [device](void *ptr) {
+          vkUnmapMemory(device, *((VkDeviceMemory *)ptr));
+        },
+        &pbp_buf_mem);
+    int *pbp_mapped_int = (int *)pbp_mapped;
 
     VkBuffer other_buf;
     VkDeviceMemory other_buf_mem;
@@ -1179,9 +1177,10 @@ image::Bl dither::blue_noise(int width, int height, int threads,
     }
 
     auto result = dither::internal::blue_noise_vulkan_impl(
-        device, phys_device, command_buffer, command_pool, compute_queue,
+        device, command_buffer, compute_queue, compute_pipeline,
-        pbp_buf, compute_pipeline, compute_pipeline_layout,
+        compute_pipeline_layout, compute_descriptor_set,
-        compute_descriptor_set, filter_out_buf, width, height);
+        filter_out_mapped_float, pbp_mapped_int, filter_out_buf_mem,
+        pbp_buf_mem, width, height);
     if (!result.empty()) {
       return internal::rangeToBl(result, width);
     }

src/blue_noise.hpp

@@ -60,12 +60,17 @@ void vulkan_copy_buffer(VkDevice device, VkCommandPool command_pool,
                         VkQueue queue, VkBuffer src_buf, VkBuffer dst_buf,
                         VkDeviceSize size);
 
+void vulkan_flush_buffer(VkDevice device, VkDeviceMemory memory);
+void vulkan_invalidate_buffer(VkDevice device, VkDeviceMemory memory);
+
 std::vector<unsigned int> blue_noise_vulkan_impl(
-    VkDevice device, VkPhysicalDevice phys_device,
+    VkDevice device, VkCommandBuffer command_buffer, VkQueue queue,
-    VkCommandBuffer command_buffer, VkCommandPool command_pool, VkQueue queue,
+    VkPipeline pipeline, VkPipelineLayout pipeline_layout,
-    VkBuffer pbp_buf, VkPipeline pipeline, VkPipelineLayout pipeline_layout,
+    VkDescriptorSet descriptor_set, float *filter_out_mapped, int *pbp_mapped,
-    VkDescriptorSet descriptor_set, VkBuffer filter_out_buf, const int width,
+    VkDeviceMemory filter_out_buf_mem, VkDeviceMemory pbp_buf_mem,
-    const int height);
+    const int width, const int height);
+
+std::vector<float> vulkan_buf_to_vec(float *mapped, unsigned int size);
 #endif
 
 #if DITHERING_OPENCL_ENABLED == 1
@@ -300,6 +305,42 @@ inline std::pair<int, int> filter_minmax(const std::vector<float> &filter,
   return {min_index, max_index};
 }
 
+inline std::pair<int, int> filter_minmax_raw_array(const float *const filter,
+                                                   unsigned int size,
+                                                   std::vector<bool> pbp) {
+  // ensure minority pixel is "true"
+  unsigned int count = 0;
+  for (bool value : pbp) {
+    if (value) {
+      ++count;
+    }
+  }
+  if (count * 2 >= pbp.size()) {
+    // std::cout << "MINMAX flip\n"; // DEBUG
+    for (unsigned int i = 0; i < pbp.size(); ++i) {
+      pbp[i] = !pbp[i];
+    }
+  }
+
+  float min = std::numeric_limits<float>::infinity();
+  float max = -std::numeric_limits<float>::infinity();
+  int min_index = -1;
+  int max_index = -1;
+
+  for (unsigned int i = 0; i < size; ++i) {
+    if (!pbp[i] && filter[i] < min) {
+      min_index = i;
+      min = filter[i];
+    }
+    if (pbp[i] && filter[i] > max) {
+      max_index = i;
+      max = filter[i];
+    }
+  }
+
+  return {min_index, max_index};
+}
+
 inline std::pair<int, int> filter_abs_minmax(const std::vector<float> &filter) {
   float min = std::numeric_limits<float>::infinity();
   float max = -std::numeric_limits<float>::infinity();