Merge pull request #11544 from Kelebek1/reduce_stream_buffer_renderdoc

Allow GPUs without rebar to open multiple RenderDoc captures

src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp

@@ -24,25 +24,38 @@ using namespace Common::Literals;
 
 // Maximum potential alignment of a Vulkan buffer
 constexpr VkDeviceSize MAX_ALIGNMENT = 256;
-// Maximum size to put elements in the stream buffer
-constexpr VkDeviceSize MAX_STREAM_BUFFER_REQUEST_SIZE = 8_MiB;
 // Stream buffer size in bytes
-constexpr VkDeviceSize STREAM_BUFFER_SIZE = 128_MiB;
+constexpr VkDeviceSize MAX_STREAM_BUFFER_SIZE = 128_MiB;
-constexpr VkDeviceSize REGION_SIZE = STREAM_BUFFER_SIZE / StagingBufferPool::NUM_SYNCS;
 
-size_t Region(size_t iterator) noexcept {
+size_t GetStreamBufferSize(const Device& device) {
-    return iterator / REGION_SIZE;
+    VkDeviceSize size{0};
+    if (device.HasDebuggingToolAttached()) {
+        ForEachDeviceLocalHostVisibleHeap(device, [&size](size_t index, VkMemoryHeap& heap) {
+            size = std::max(size, heap.size);
+        });
+        // If rebar is not supported, cut the max heap size to 40%. This will allow 2 captures to be
+        // loaded at the same time in RenderDoc. If rebar is supported, this shouldn't be an issue
+        // as the heap will be much larger.
+        if (size <= 256_MiB) {
+            size = size * 40 / 100;
+        }
+    } else {
+        size = MAX_STREAM_BUFFER_SIZE;
+    }
+    return std::min(Common::AlignUp(size, MAX_ALIGNMENT), MAX_STREAM_BUFFER_SIZE);
 }
 } // Anonymous namespace
 
 StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& memory_allocator_,
                                      Scheduler& scheduler_)
-    : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_} {
+    : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_},
+      stream_buffer_size{GetStreamBufferSize(device)}, region_size{stream_buffer_size /
+                                                                   StagingBufferPool::NUM_SYNCS} {
     VkBufferCreateInfo stream_ci = {
         .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
         .pNext = nullptr,
         .flags = 0,
-        .size = STREAM_BUFFER_SIZE,
+        .size = stream_buffer_size,
         .usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
                  VK_BUFFER_USAGE_INDEX_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
         .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
@@ -63,7 +76,7 @@ StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& mem
 StagingBufferPool::~StagingBufferPool() = default;
 
 StagingBufferRef StagingBufferPool::Request(size_t size, MemoryUsage usage, bool deferred) {
-    if (!deferred && usage == MemoryUsage::Upload && size <= MAX_STREAM_BUFFER_REQUEST_SIZE) {
+    if (!deferred && usage == MemoryUsage::Upload && size <= region_size) {
         return GetStreamBuffer(size);
     }
     return GetStagingBuffer(size, usage, deferred);
@@ -101,7 +114,7 @@ StagingBufferRef StagingBufferPool::GetStreamBuffer(size_t size) {
     used_iterator = iterator;
     free_iterator = std::max(free_iterator, iterator + size);
 
-    if (iterator + size >= STREAM_BUFFER_SIZE) {
+    if (iterator + size >= stream_buffer_size) {
         std::fill(sync_ticks.begin() + Region(used_iterator), sync_ticks.begin() + NUM_SYNCS,
                   current_tick);
         used_iterator = 0;

src/video_core/renderer_vulkan/vk_staging_buffer_pool.h

@@ -90,6 +90,9 @@ private:
     void ReleaseCache(MemoryUsage usage);
 
     void ReleaseLevel(StagingBuffersCache& cache, size_t log2);
+    size_t Region(size_t iter) const noexcept {
+        return iter / region_size;
+    }
 
     const Device& device;
     MemoryAllocator& memory_allocator;
@@ -97,6 +100,8 @@ private:
 
     vk::Buffer stream_buffer;
     std::span<u8> stream_pointer;
+    VkDeviceSize stream_buffer_size;
+    VkDeviceSize region_size;
 
     size_t iterator = 0;
     size_t used_iterator = 0;

src/video_core/vulkan_common/vulkan_memory_allocator.cpp

@@ -9,6 +9,7 @@
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/common_types.h"
+#include "common/literals.h"
 #include "common/logging/log.h"
 #include "common/polyfill_ranges.h"
 #include "video_core/vulkan_common/vma.h"
@@ -69,8 +70,7 @@ struct Range {
     case MemoryUsage::Download:
         return VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;
     case MemoryUsage::DeviceLocal:
-        return VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT |
+        return {};
-               VMA_ALLOCATION_CREATE_HOST_ACCESS_ALLOW_TRANSFER_INSTEAD_BIT;
     }
     return {};
 }
@@ -212,7 +212,20 @@ MemoryAllocator::MemoryAllocator(const Device& device_)
     : device{device_}, allocator{device.GetAllocator()},
       properties{device_.GetPhysical().GetMemoryProperties().memoryProperties},
       buffer_image_granularity{
-          device_.GetPhysical().GetProperties().limits.bufferImageGranularity} {}
+          device_.GetPhysical().GetProperties().limits.bufferImageGranularity} {
+    // GPUs not supporting rebar may only have a region with less than 256MB host visible/device
+    // local memory. In that case, opening 2 RenderDoc captures side-by-side is not possible due to
+    // the heap running out of memory. With RenderDoc attached and only a small host/device region,
+    // only allow the stream buffer in this memory heap.
+    if (device.HasDebuggingToolAttached()) {
+        using namespace Common::Literals;
+        ForEachDeviceLocalHostVisibleHeap(device, [this](size_t index, VkMemoryHeap& heap) {
+            if (heap.size <= 256_MiB) {
+                valid_memory_types &= ~(1u << index);
+            }
+        });
+    }
+}
 
 MemoryAllocator::~MemoryAllocator() = default;
 
@@ -244,7 +257,7 @@ vk::Buffer MemoryAllocator::CreateBuffer(const VkBufferCreateInfo& ci, MemoryUsa
         .usage = MemoryUsageVma(usage),
         .requiredFlags = 0,
         .preferredFlags = MemoryUsagePreferedVmaFlags(usage),
-        .memoryTypeBits = 0,
+        .memoryTypeBits = usage == MemoryUsage::Stream ? 0u : valid_memory_types,
         .pool = VK_NULL_HANDLE,
         .pUserData = nullptr,
         .priority = 0.f,

src/video_core/vulkan_common/vulkan_memory_allocator.h

@@ -7,6 +7,7 @@
 #include <span>
 #include <vector>
 #include "common/common_types.h"
+#include "video_core/vulkan_common/vulkan_device.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
 VK_DEFINE_HANDLE(VmaAllocator)
@@ -26,6 +27,18 @@ enum class MemoryUsage {
     Stream,      ///< Requests device local host visible buffer, falling back host memory.
 };
 
+template <typename F>
+void ForEachDeviceLocalHostVisibleHeap(const Device& device, F&& f) {
+    auto memory_props = device.GetPhysical().GetMemoryProperties().memoryProperties;
+    for (size_t i = 0; i < memory_props.memoryTypeCount; i++) {
+        auto& memory_type = memory_props.memoryTypes[i];
+        if ((memory_type.propertyFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT) &&
+            (memory_type.propertyFlags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)) {
+            f(memory_type.heapIndex, memory_props.memoryHeaps[memory_type.heapIndex]);
+        }
+    }
+}
+
 /// Ownership handle of a memory commitment.
 /// Points to a subregion of a memory allocation.
 class MemoryCommit {
@@ -124,6 +137,7 @@ private:
     std::vector<std::unique_ptr<MemoryAllocation>> allocations; ///< Current allocations.
     VkDeviceSize buffer_image_granularity; // The granularity for adjacent offsets between buffers
                                            // and optimal images
+    u32 valid_memory_types{~0u};
 };
 
 } // namespace Vulkan