Merge pull request 'perf(ui): flush only the written descriptor range, not the whole heap (#61)' (#99) from claude/issue-61 into master

interfaces/Crafter.Graphics-Device.cppm

@@ -161,6 +161,14 @@ export namespace Crafter {
 
         inline static VkPhysicalDeviceMemoryProperties memoryProperties;
 
+        // VkPhysicalDeviceLimits::nonCoherentAtomSize — the alignment (in bytes,
+        // a power of two) that vkFlushMappedMemoryRanges / vkInvalidateMappedMemoryRanges
+        // require for the offset and size of a sub-buffer range. Whole-buffer
+        // (VK_WHOLE_SIZE) flushes sidestep it; ranged flushes must round outward
+        // to it (see VulkanBuffer::FlushDevice(offset,bytes)). Populated at device
+        // creation; defaults to 1 so the rounding math is well-defined before then.
+        inline static VkDeviceSize nonCoherentAtomSize = 1;
+
         // Core physical-device properties, captured once at Initialize from the
         // VkPhysicalDeviceProperties2 query. Kept because the pipeline-cache
         // persistence path needs vendorID / deviceID / pipelineCacheUUID to

interfaces/Crafter.Graphics-VulkanBuffer.cppm

@@ -29,10 +29,35 @@ import std;
 import :Device;
 
 namespace Crafter {
+    // Round a host-write flush range outward to nonCoherentAtomSize boundaries —
+    // the alignment vkFlushMappedMemoryRanges demands for a sub-buffer range
+    // (whole-buffer VK_WHOLE_SIZE calls sidestep it). `atom` is
+    // VkPhysicalDeviceLimits::nonCoherentAtomSize (a power of two ≥ 1) and
+    // `mappingSize` is the byte size the memory was mapped with (the allocation
+    // size). The start rounds down and the end rounds up to atom multiples; the
+    // end is then clamped to `mappingSize`. Clamping is what keeps the range
+    // valid even when `mappingSize` itself is not atom-aligned: the spec permits
+    // a non-atom-multiple size only when offset+size equals the mapping size, and
+    // the clamp lands exactly on that exception. Pure math, so it is unit-tested
+    // without a device.
+    export struct MappedFlushRange { VkDeviceSize offset; VkDeviceSize size; };
+    export constexpr MappedFlushRange AlignMappedFlushRange(
+            VkDeviceSize offset, VkDeviceSize bytes,
+            VkDeviceSize atom, VkDeviceSize mappingSize) {
+        VkDeviceSize begin = (offset / atom) * atom;
+        VkDeviceSize end   = ((offset + bytes + atom - 1) / atom) * atom;
+        if (end > mappingSize) end = mappingSize;
+        return { begin, end - begin };
+    }
+
     export class VulkanBufferBase {
         public:
         VkDeviceAddress address;
         std::uint32_t size;
+        // Byte size the memory was mapped with — equal to the allocation's
+        // VkMemoryRequirements::size, which is ≥ `size` and what ranged flushes
+        // clamp their upper bound to. Only meaningful for mapped buffers.
+        VkDeviceSize mappedSize = 0;
         VkBuffer buffer = VK_NULL_HANDLE;
         VkDeviceMemory memory;
         // Property flags of the memory type actually chosen by GetMemoryType —
@@ -116,6 +141,7 @@ namespace Crafter {
             address = vkGetBufferDeviceAddress(Device::device, &addressInfo);
 
             if constexpr(Mapped) {
+                mappedSize = memReqs.size;
                 Device::CheckVkResult(vkMapMemory(Device::device, memory, 0, memReqs.size, 0, reinterpret_cast<void**>(&(VulkanBufferMappedConditional<T, true>::value))));
             }
         }
@@ -208,6 +234,27 @@ namespace Crafter {
             vkFlushMappedMemoryRanges(Device::device, 1, &range);
         }
 
+        // Flush only the host writes in [offset, offset+bytes) to the device,
+        // instead of the whole buffer. Use after touching a small sub-range
+        // (e.g. one descriptor in a multi-KB descriptor heap) so cache
+        // maintenance scales with the bytes actually written. The range is
+        // rounded outward to nonCoherentAtomSize as the Vulkan spec requires.
+        // No-op on coherent memory, same gate as the whole-buffer FlushDevice().
+        void FlushDevice(VkDeviceSize offset, VkDeviceSize bytes) requires(Mapped) {
+            if (memoryPropertyFlagsChosen & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) {
+                return;
+            }
+            MappedFlushRange r = AlignMappedFlushRange(
+                offset, bytes, Device::nonCoherentAtomSize, mappedSize);
+            VkMappedMemoryRange range = {
+                .sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
+                .memory = memory,
+                .offset = r.offset,
+                .size   = r.size
+            };
+            vkFlushMappedMemoryRanges(Device::device, 1, &range);
+        }
+
         void FlushDevice(VkCommandBuffer cmd, VkAccessFlags dstAccessMask, VkPipelineStageFlags dstStageMask) requires(Mapped) {
             FlushDevice();
             VkBufferMemoryBarrier barrier = {
@@ -251,6 +298,7 @@ namespace Crafter {
             buffer = other.buffer;
             memory = other.memory;
             size = other.size;
+            mappedSize = other.mappedSize;
             capacity = other.capacity;
             usageFlagsCreated = other.usageFlagsCreated;
             memoryPropertyFlagsChosen = other.memoryPropertyFlagsChosen;