fix staging buffer

cocos · Dec 5, 2024 · 152ec27 · 152ec27
1 parent dd096b0
commit 152ec27
Show file tree

Hide file tree

Showing 4 changed files with 45 additions and 14 deletions.
diff --git a/native/cocos/renderer/gfx-metal/MTLDevice.mm b/native/cocos/renderer/gfx-metal/MTLDevice.mm
@@ -256,6 +256,7 @@ of this software and associated engine source code (the "Software"), a limited,
         CCMTLGPUStagingBufferPool *bufferPool = _gpuStagingBufferPools[index];
         if (bufferPool) {
             bufferPool->reset();
+            bufferPool->shrinkSize();
             CCMTLGPUGarbageCollectionPool::getInstance()->clear(index);
             static_cast<CCMTLCommandBuffer*>(_cmdBuff)->signalFence();
         }

diff --git a/native/cocos/renderer/gfx-metal/MTLGPUObjects.h b/native/cocos/renderer/gfx-metal/MTLGPUObjects.h
@@ -178,6 +178,8 @@ struct CCMTLGPUDescriptorSet {
 
 class CCMTLGPUStagingBufferPool final {
 public:
+    static constexpr uint32_t CHUNK_SIZE = 16 * 1024 * 1024; // 16M per block by default
+
     CCMTLGPUStagingBufferPool(id<MTLDevice> device)
     : _device(device) {}
 
@@ -225,8 +227,14 @@ class CCMTLGPUStagingBufferPool final {
         }
     }
 
-    void shrinkSize() {
+    void shrinkSize(size_t minimalSize = CHUNK_SIZE) {
+        size_t reservedSize = 0;
         for (auto iter = _pool.begin(); iter != _pool.end() && _pool.size() > 1;) {
+            if (reservedSize < minimalSize) {
+                reservedSize += [iter->mtlBuffer length];
+                ++iter;
+                continue;
+            }
             if (iter->curOffset == 0) {
                 [iter->mtlBuffer release];
                 iter = _pool.erase(iter);
@@ -256,7 +264,7 @@ struct CCMTLGPUBufferImageCopy {
     MTLOrigin destinationOrigin = {0, 0, 0};
 };
 
-//destroy GPU resource only, delete the owner object mannually.
+// destroy GPU resource only, delete the owner object mannually.
 class CCMTLGPUGarbageCollectionPool final {
     using GCFunc = std::function<void(void)>;
 
@@ -298,7 +306,7 @@ class CCMTLGPUGarbageCollectionPool final {
     }
 
 protected:
-    //avoid cross-reference with CCMTLDevice
+    // avoid cross-reference with CCMTLDevice
     std::function<uint8_t(void)> _getFrameIndex;
     ccstd::queue<GCFunc> _releaseQueue[MAX_FRAMES_IN_FLIGHT];
 };

diff --git a/native/cocos/renderer/gfx-vulkan/VKDevice.cpp b/native/cocos/renderer/gfx-vulkan/VKDevice.cpp
@@ -62,7 +62,7 @@ CC_DISABLE_WARNINGS()
 #include "vk_mem_alloc.h"
 #define THSVS_ERROR_CHECK_MIXED_IMAGE_LAYOUT
 // remote potential hazard because of programmable blend
-//#define THSVS_ERROR_CHECK_POTENTIAL_HAZARD
+// #define THSVS_ERROR_CHECK_POTENTIAL_HAZARD
 #define THSVS_SIMPLER_VULKAN_SYNCHRONIZATION_IMPLEMENTATION
 #include "thsvs_simpler_vulkan_synchronization.h"
 CC_ENABLE_WARNINGS()
@@ -642,6 +642,10 @@ void CCVKDevice::acquire(Swapchain *const *swapchains, uint32_t count) {
     _gpuDescriptorSetHub->flush();
     _gpuSemaphorePool->reset();
 
+    auto *bufferPool = gpuStagingBufferPool();
+    bufferPool->reset();
+    bufferPool->shrinkSize();
+
     for (uint32_t i = 0; i < vkSwapchains.size(); ++i) {
         VkSemaphore acquireSemaphore = _gpuSemaphorePool->alloc();
         VkResult res = vkAcquireNextImageKHR(_gpuDevice->vkDevice, vkSwapchains[i], ~0ULL,
@@ -773,12 +777,14 @@ void CCVKDevice::initDeviceFeature() {
     _features[toNumber(Feature::RASTERIZATION_ORDER_NOCOHERENT)] = true;
     _features[toNumber(Feature::MULTI_SAMPLE_RESOLVE_DEPTH_STENCIL)] = checkExtension("VK_KHR_depth_stencil_resolve");
 
-    _gpuContext->debugReport = _gpuContext->checkExtension(VK_EXT_DEBUG_REPORT_EXTENSION_NAME) &&
+    _gpuContext->debugReport =
+        _gpuContext->checkExtension(VK_EXT_DEBUG_REPORT_EXTENSION_NAME) &&
         checkExtension(VK_EXT_DEBUG_MARKER_EXTENSION_NAME) &&
         (vkCmdDebugMarkerBeginEXT != nullptr) &&
         (vkCmdDebugMarkerInsertEXT != nullptr) &&
         (vkCmdDebugMarkerEndEXT != nullptr);
-    _gpuContext->debugUtils = _gpuContext->checkExtension(VK_EXT_DEBUG_UTILS_EXTENSION_NAME) &&
+    _gpuContext->debugUtils =
+        _gpuContext->checkExtension(VK_EXT_DEBUG_UTILS_EXTENSION_NAME) &&
         (vkCmdBeginDebugUtilsLabelEXT != nullptr) &&
         (vkCmdInsertDebugUtilsLabelEXT != nullptr) &&
         (vkCmdEndDebugUtilsLabelEXT != nullptr);
@@ -1104,14 +1110,14 @@ SampleCount CCVKDevice::getMaxSampleCount(Format format, TextureUsage usage, Tex
 
     VkImageFormatProperties imageFormatProperties = {};
     vkGetPhysicalDeviceImageFormatProperties(_gpuContext->physicalDevice, vkFormat, VK_IMAGE_TYPE_2D,
-        VK_IMAGE_TILING_OPTIMAL, usages, 0, &imageFormatProperties);
+                                             VK_IMAGE_TILING_OPTIMAL, usages, 0, &imageFormatProperties);
 
     if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_64_BIT) return SampleCount::X64;
     if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_32_BIT) return SampleCount::X32;
     if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_16_BIT) return SampleCount::X16;
-    if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_8_BIT)  return SampleCount::X8;
-    if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_4_BIT)  return SampleCount::X4;
-    if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_2_BIT)  return SampleCount::X2;
+    if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_8_BIT) return SampleCount::X8;
+    if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_4_BIT) return SampleCount::X4;
+    if (imageFormatProperties.sampleCounts & VK_SAMPLE_COUNT_2_BIT) return SampleCount::X2;
 
     return SampleCount::X1;
 }

diff --git a/native/cocos/renderer/gfx-vulkan/VKGPUObjects.h b/native/cocos/renderer/gfx-vulkan/VKGPUObjects.h
@@ -103,7 +103,7 @@ class CCVKGPUDeviceObject : public GFXDeviceObject<CCVKDeviceObjectDeleter> {
     CCVKGPUDeviceObject() = default;
     ~CCVKGPUDeviceObject() = default;
 
-    virtual void shutdown(){};
+    virtual void shutdown() {};
 };
 
 template <typename T>
@@ -128,7 +128,7 @@ class CCVKGPURenderPass final : public CCVKGPUDeviceObject {
     // helper storage
     ccstd::vector<VkClearValue> clearValues;
     ccstd::vector<VkSampleCountFlagBits> sampleCounts; // per subpass
-    ccstd::vector<bool> hasSelfDependency; // per subpass
+    ccstd::vector<bool> hasSelfDependency;             // per subpass
 
     const CCVKGPUGeneralBarrier *getBarrier(size_t index, CCVKGPUDevice *gpuDevice) const;
     bool hasShadingAttachment(uint32_t subPassId) const;
@@ -697,8 +697,8 @@ class CCVKGPUCommandBufferPool final {
                 vkDestroyCommandPool(_device->vkDevice, pool.vkCommandPool, nullptr);
                 pool.vkCommandPool = VK_NULL_HANDLE;
             }
-            for (auto &item: pool.usedCommandBuffers)item.clear();
-            for (auto &item: pool.commandBuffers)item.clear();
+            for (auto &item : pool.usedCommandBuffers) item.clear();
+            for (auto &item : pool.commandBuffers) item.clear();
         }
         _pools.clear();
     }
@@ -839,6 +839,22 @@ class CCVKGPUStagingBufferPool final {
         }
     }
 
+    void shrinkSize(size_t minimalSize = CHUNK_SIZE) {
+        size_t reservedSize = 0;
+        for (auto iter = _pool.begin(); iter != _pool.end() && _pool.size() > 1;) {
+            if (reservedSize < minimalSize) {
+                reservedSize += iter->gpuBuffer->size;
+                ++iter;
+                continue;
+            }
+            if (iter->curOffset == 0) {
+                iter = _pool.erase(iter);
+            } else {
+                ++iter;
+            }
+        }
+    }
+
 private:
     struct Buffer {
         IntrusivePtr<CCVKGPUBuffer> gpuBuffer;