Merge pull request #4105 from ReinUsesLisp/resident-buffers

gl_rasterizer: Use NV_vertex_buffer_unified_memory for vertex buffer robustness

4 files changed, 85 insertions, 70 deletions

diff --git a/src/video_core/renderer_vulkan/vk_buffer_cache.cpp b/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
index 1fde38328..f10f96cd8 100644
--- a/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
@@ -37,9 +37,9 @@ std::unique_ptr<VKStreamBuffer> CreateStreamBuffer(const VKDevice& device, VKSch
 
 } // Anonymous namespace
 
-Buffer::Buffer(const VKDevice& device, VKMemoryManager& memory_manager, VAddr cpu_addr,
-               std::size_t size)
-    : VideoCommon::BufferBlock{cpu_addr, size} {
+Buffer::Buffer(const VKDevice& device, VKMemoryManager& memory_manager, VKScheduler& scheduler_,
+               VKStagingBufferPool& staging_pool_, VAddr cpu_addr, std::size_t size)
+    : VideoCommon::BufferBlock{cpu_addr, size}, scheduler{scheduler_}, staging_pool{staging_pool_} {
     VkBufferCreateInfo ci;
     ci.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
     ci.pNext = nullptr;
@@ -56,40 +56,15 @@ Buffer::Buffer(const VKDevice& device, VKMemoryManager& memory_manager, VAddr cp
 
 Buffer::~Buffer() = default;
 
-VKBufferCache::VKBufferCache(VideoCore::RasterizerInterface& rasterizer, Core::System& system,
-                             const VKDevice& device, VKMemoryManager& memory_manager,
-                             VKScheduler& scheduler, VKStagingBufferPool& staging_pool)
-    : VideoCommon::BufferCache<Buffer, VkBuffer, VKStreamBuffer>{rasterizer, system,
-                                                                 CreateStreamBuffer(device,
-                                                                                    scheduler)},
-      device{device}, memory_manager{memory_manager}, scheduler{scheduler}, staging_pool{
-                                                                                staging_pool} {}
-
-VKBufferCache::~VKBufferCache() = default;
-
-std::shared_ptr<Buffer> VKBufferCache::CreateBlock(VAddr cpu_addr, std::size_t size) {
-    return std::make_shared<Buffer>(device, memory_manager, cpu_addr, size);
-}
-
-VkBuffer VKBufferCache::GetEmptyBuffer(std::size_t size) {
-    size = std::max(size, std::size_t(4));
-    const auto& empty = staging_pool.GetUnusedBuffer(size, false);
-    scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([size, buffer = *empty.handle](vk::CommandBuffer cmdbuf) {
-        cmdbuf.FillBuffer(buffer, 0, size, 0);
-    });
-    return *empty.handle;
-}
-
-void VKBufferCache::UploadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
-                                    const u8* data) {
+void Buffer::Upload(std::size_t offset, std::size_t size, const u8* data) const {
     const auto& staging = staging_pool.GetUnusedBuffer(size, true);
     std::memcpy(staging.commit->Map(size), data, size);
 
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([staging = *staging.handle, buffer = buffer.Handle(), offset,
-                      size](vk::CommandBuffer cmdbuf) {
-        cmdbuf.CopyBuffer(staging, buffer, VkBufferCopy{0, offset, size});
+
+    const VkBuffer handle = Handle();
+    scheduler.Record([staging = *staging.handle, handle, offset, size](vk::CommandBuffer cmdbuf) {
+        cmdbuf.CopyBuffer(staging, handle, VkBufferCopy{0, offset, size});
 
         VkBufferMemoryBarrier barrier;
         barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
@@ -98,7 +73,7 @@ void VKBufferCache::UploadBlockData(const Buffer& buffer, std::size_t offset, st
         barrier.dstAccessMask = UPLOAD_ACCESS_BARRIERS;
         barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
         barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barrier.buffer = buffer;
+        barrier.buffer = handle;
         barrier.offset = offset;
         barrier.size = size;
         cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, UPLOAD_PIPELINE_STAGE, 0, {},
@@ -106,12 +81,12 @@ void VKBufferCache::UploadBlockData(const Buffer& buffer, std::size_t offset, st
     });
 }
 
-void VKBufferCache::DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
-                                      u8* data) {
+void Buffer::Download(std::size_t offset, std::size_t size, u8* data) const {
     const auto& staging = staging_pool.GetUnusedBuffer(size, true);
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([staging = *staging.handle, buffer = buffer.Handle(), offset,
-                      size](vk::CommandBuffer cmdbuf) {
+
+    const VkBuffer handle = Handle();
+    scheduler.Record([staging = *staging.handle, handle, offset, size](vk::CommandBuffer cmdbuf) {
         VkBufferMemoryBarrier barrier;
         barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
         barrier.pNext = nullptr;
@@ -119,7 +94,7 @@ void VKBufferCache::DownloadBlockData(const Buffer& buffer, std::size_t offset,
         barrier.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
         barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
         barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
-        barrier.buffer = buffer;
+        barrier.buffer = handle;
         barrier.offset = offset;
         barrier.size = size;
 
@@ -127,17 +102,19 @@ void VKBufferCache::DownloadBlockData(const Buffer& buffer, std::size_t offset,
                                    VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
                                    VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
                                VK_PIPELINE_STAGE_TRANSFER_BIT, 0, {}, barrier, {});
-        cmdbuf.CopyBuffer(buffer, staging, VkBufferCopy{offset, 0, size});
+        cmdbuf.CopyBuffer(handle, staging, VkBufferCopy{offset, 0, size});
     });
     scheduler.Finish();
 
     std::memcpy(data, staging.commit->Map(size), size);
 }
 
-void VKBufferCache::CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
-                              std::size_t dst_offset, std::size_t size) {
+void Buffer::CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst_offset,
+                      std::size_t size) const {
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([src_buffer = src.Handle(), dst_buffer = dst.Handle(), src_offset, dst_offset,
+
+    const VkBuffer dst_buffer = Handle();
+    scheduler.Record([src_buffer = src.Handle(), dst_buffer, src_offset, dst_offset,
                       size](vk::CommandBuffer cmdbuf) {
         cmdbuf.CopyBuffer(src_buffer, dst_buffer, VkBufferCopy{src_offset, dst_offset, size});
 
@@ -165,4 +142,30 @@ void VKBufferCache::CopyBlock(const Buffer& src, const Buffer& dst, std::size_t
     });
 }
 
+VKBufferCache::VKBufferCache(VideoCore::RasterizerInterface& rasterizer, Core::System& system,
+                             const VKDevice& device, VKMemoryManager& memory_manager,
+                             VKScheduler& scheduler, VKStagingBufferPool& staging_pool)
+    : VideoCommon::BufferCache<Buffer, VkBuffer, VKStreamBuffer>{rasterizer, system,
+                                                                 CreateStreamBuffer(device,
+                                                                                    scheduler)},
+      device{device}, memory_manager{memory_manager}, scheduler{scheduler}, staging_pool{
+                                                                                staging_pool} {}
+
+VKBufferCache::~VKBufferCache() = default;
+
+std::shared_ptr<Buffer> VKBufferCache::CreateBlock(VAddr cpu_addr, std::size_t size) {
+    return std::make_shared<Buffer>(device, memory_manager, scheduler, staging_pool, cpu_addr,
+                                    size);
+}
+
+VKBufferCache::BufferInfo VKBufferCache::GetEmptyBuffer(std::size_t size) {
+    size = std::max(size, std::size_t(4));
+    const auto& empty = staging_pool.GetUnusedBuffer(size, false);
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([size, buffer = *empty.handle](vk::CommandBuffer cmdbuf) {
+        cmdbuf.FillBuffer(buffer, 0, size, 0);
+    });
+    return {*empty.handle, 0, 0};
+}
+
 } // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/vk_buffer_cache.h b/src/video_core/renderer_vulkan/vk_buffer_cache.h
index 9ebbef835..3630aca77 100644
--- a/src/video_core/renderer_vulkan/vk_buffer_cache.h
+++ b/src/video_core/renderer_vulkan/vk_buffer_cache.h
@@ -25,15 +25,29 @@ class VKScheduler;
 
 class Buffer final : public VideoCommon::BufferBlock {
 public:
-    explicit Buffer(const VKDevice& device, VKMemoryManager& memory_manager, VAddr cpu_addr,
-                    std::size_t size);
+    explicit Buffer(const VKDevice& device, VKMemoryManager& memory_manager, VKScheduler& scheduler,
+                    VKStagingBufferPool& staging_pool, VAddr cpu_addr, std::size_t size);
     ~Buffer();
 
+    void Upload(std::size_t offset, std::size_t size, const u8* data) const;
+
+    void Download(std::size_t offset, std::size_t size, u8* data) const;
+
+    void CopyFrom(const Buffer& src, std::size_t src_offset, std::size_t dst_offset,
+                  std::size_t size) const;
+
     VkBuffer Handle() const {
         return *buffer.handle;
     }
 
+    u64 Address() const {
+        return 0;
+    }
+
 private:
+    VKScheduler& scheduler;
+    VKStagingBufferPool& staging_pool;
+
     VKBuffer buffer;
 };
 
@@ -44,20 +58,11 @@ public:
                            VKScheduler& scheduler, VKStagingBufferPool& staging_pool);
     ~VKBufferCache();
 
-    VkBuffer GetEmptyBuffer(std::size_t size) override;
+    BufferInfo GetEmptyBuffer(std::size_t size) override;
 
 protected:
     std::shared_ptr<Buffer> CreateBlock(VAddr cpu_addr, std::size_t size) override;
 
-    void UploadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
-                         const u8* data) override;
-
-    void DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
-                           u8* data) override;
-
-    void CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
-                   std::size_t dst_offset, std::size_t size) override;
-
 private:
     const VKDevice& device;
     VKMemoryManager& memory_manager;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 29001953c..e3714ee6d 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -870,10 +870,10 @@ void RasterizerVulkan::BeginTransformFeedback() {
     UNIMPLEMENTED_IF(binding.buffer_offset != 0);
 
     const GPUVAddr gpu_addr = binding.Address();
-    const auto size = static_cast<VkDeviceSize>(binding.buffer_size);
-    const auto [buffer, offset] = buffer_cache.UploadMemory(gpu_addr, size, 4, true);
+    const VkDeviceSize size = static_cast<VkDeviceSize>(binding.buffer_size);
+    const auto info = buffer_cache.UploadMemory(gpu_addr, size, 4, true);
 
-    scheduler.Record([buffer = buffer, offset = offset, size](vk::CommandBuffer cmdbuf) {
+    scheduler.Record([buffer = info.handle, offset = info.offset, size](vk::CommandBuffer cmdbuf) {
         cmdbuf.BindTransformFeedbackBuffersEXT(0, 1, &buffer, &offset, &size);
         cmdbuf.BeginTransformFeedbackEXT(0, 0, nullptr, nullptr);
     });
@@ -925,8 +925,8 @@ void RasterizerVulkan::SetupVertexArrays(FixedPipelineState::VertexInput& vertex
             buffer_bindings.AddVertexBinding(DefaultBuffer(), 0);
             continue;
         }
-        const auto [buffer, offset] = buffer_cache.UploadMemory(start, size);
-        buffer_bindings.AddVertexBinding(buffer, offset);
+        const auto info = buffer_cache.UploadMemory(start, size);
+        buffer_bindings.AddVertexBinding(info.handle, info.offset);
     }
 }
 
@@ -948,7 +948,9 @@ void RasterizerVulkan::SetupIndexBuffer(BufferBindings& buffer_bindings, DrawPar
             break;
         }
         const GPUVAddr gpu_addr = regs.index_array.IndexStart();
-        auto [buffer, offset] = buffer_cache.UploadMemory(gpu_addr, CalculateIndexBufferSize());
+        const auto info = buffer_cache.UploadMemory(gpu_addr, CalculateIndexBufferSize());
+        VkBuffer buffer = info.handle;
+        u64 offset = info.offset;
         std::tie(buffer, offset) = quad_indexed_pass.Assemble(
             regs.index_array.format, params.num_vertices, params.base_vertex, buffer, offset);
 
@@ -962,7 +964,9 @@ void RasterizerVulkan::SetupIndexBuffer(BufferBindings& buffer_bindings, DrawPar
             break;
         }
         const GPUVAddr gpu_addr = regs.index_array.IndexStart();
-        auto [buffer, offset] = buffer_cache.UploadMemory(gpu_addr, CalculateIndexBufferSize());
+        const auto info = buffer_cache.UploadMemory(gpu_addr, CalculateIndexBufferSize());
+        VkBuffer buffer = info.handle;
+        u64 offset = info.offset;
 
         auto format = regs.index_array.format;
         const bool is_uint8 = format == Maxwell::IndexFormat::UnsignedByte;
@@ -1109,10 +1113,9 @@ void RasterizerVulkan::SetupConstBuffer(const ConstBufferEntry& entry,
         Common::AlignUp(CalculateConstBufferSize(entry, buffer), 4 * sizeof(float));
     ASSERT(size <= MaxConstbufferSize);
 
-    const auto [buffer_handle, offset] =
+    const auto info =
         buffer_cache.UploadMemory(buffer.address, size, device.GetUniformBufferAlignment());
-
-    update_descriptor_queue.AddBuffer(buffer_handle, offset, size);
+    update_descriptor_queue.AddBuffer(info.handle, info.offset, size);
 }
 
 void RasterizerVulkan::SetupGlobalBuffer(const GlobalBufferEntry& entry, GPUVAddr address) {
@@ -1126,14 +1129,14 @@ void RasterizerVulkan::SetupGlobalBuffer(const GlobalBufferEntry& entry, GPUVAdd
         // Note: Do *not* use DefaultBuffer() here, storage buffers can be written breaking the
         // default buffer.
         static constexpr std::size_t dummy_size = 4;
-        const auto buffer = buffer_cache.GetEmptyBuffer(dummy_size);
-        update_descriptor_queue.AddBuffer(buffer, 0, dummy_size);
+        const auto info = buffer_cache.GetEmptyBuffer(dummy_size);
+        update_descriptor_queue.AddBuffer(info.handle, info.offset, dummy_size);
         return;
     }
 
-    const auto [buffer, offset] = buffer_cache.UploadMemory(
+    const auto info = buffer_cache.UploadMemory(
         actual_addr, size, device.GetStorageBufferAlignment(), entry.IsWritten());
-    update_descriptor_queue.AddBuffer(buffer, offset, size);
+    update_descriptor_queue.AddBuffer(info.handle, info.offset, size);
 }
 
 void RasterizerVulkan::SetupUniformTexels(const Tegra::Texture::TICEntry& tic,
diff --git a/src/video_core/renderer_vulkan/vk_stream_buffer.h b/src/video_core/renderer_vulkan/vk_stream_buffer.h
index c765c60a0..689f0d276 100644
--- a/src/video_core/renderer_vulkan/vk_stream_buffer.h
+++ b/src/video_core/renderer_vulkan/vk_stream_buffer.h
@@ -35,10 +35,14 @@ public:
     /// Ensures that "size" bytes of memory are available to the GPU, potentially recording a copy.
     void Unmap(u64 size);
 
-    VkBuffer Handle() const {
+    VkBuffer Handle() const noexcept {
         return *buffer;
     }
 
+    u64 Address() const noexcept {
+        return 0;
+    }
+
 private:
     struct Watch final {
         VKFenceWatch fence;