Merge pull request #3753 from ReinUsesLisp/ac-vulkan

{gl,vk}_rasterizer: Add lazy default buffer maker and use it for empty buffers

src/video_core/engines/maxwell_3d.h

@@ -1259,7 +1259,8 @@ public:
 
                     GPUVAddr LimitAddress() const {
                         return static_cast<GPUVAddr>((static_cast<GPUVAddr>(limit_high) << 32) |
-                                                     limit_low);
+                                                     limit_low) +
+                               1;
                     }
                 } vertex_array_limit[NumVertexArrays];
 

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -186,8 +186,12 @@ void RasterizerOpenGL::SetupVertexBuffer() {
         const GPUVAddr start = vertex_array.StartAddress();
         const GPUVAddr end = regs.vertex_array_limit[index].LimitAddress();
 
-        ASSERT(end > start);
-        const u64 size = end - start + 1;
+        ASSERT(end >= start);
+        const u64 size = end - start;
+        if (size == 0) {
+            glBindVertexBuffer(static_cast<GLuint>(index), 0, 0, vertex_array.stride);
+            continue;
+        }
         const auto [vertex_buffer, vertex_buffer_offset] = buffer_cache.UploadMemory(start, size);
         glBindVertexBuffer(static_cast<GLuint>(index), vertex_buffer, vertex_buffer_offset,
                            vertex_array.stride);
@@ -311,8 +315,8 @@ std::size_t RasterizerOpenGL::CalculateVertexArraysSize() const {
         const GPUVAddr start = regs.vertex_array[index].StartAddress();
         const GPUVAddr end = regs.vertex_array_limit[index].LimitAddress();
 
-        ASSERT(end > start);
-        size += end - start + 1;
+        size += end - start;
+        ASSERT(end >= start);
     }
 
     return size;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -877,8 +877,12 @@ void RasterizerVulkan::SetupVertexArrays(FixedPipelineState::VertexInput& vertex
         const GPUVAddr start{vertex_array.StartAddress()};
         const GPUVAddr end{regs.vertex_array_limit[index].LimitAddress()};
 
-        ASSERT(end > start);
-        const std::size_t size{end - start + 1};
+        ASSERT(end >= start);
+        const std::size_t size{end - start};
+        if (size == 0) {
+            buffer_bindings.AddVertexBinding(DefaultBuffer(), 0);
+            continue;
+        }
         const auto [buffer, offset] = buffer_cache.UploadMemory(start, size);
         buffer_bindings.AddVertexBinding(buffer, offset);
     }
@@ -1033,8 +1037,7 @@ void RasterizerVulkan::SetupConstBuffer(const ConstBufferEntry& entry,
                                         const Tegra::Engines::ConstBufferInfo& buffer) {
     if (!buffer.enabled) {
         // Set values to zero to unbind buffers
-        update_descriptor_queue.AddBuffer(buffer_cache.GetEmptyBuffer(sizeof(float)), 0,
-                                          sizeof(float));
+        update_descriptor_queue.AddBuffer(DefaultBuffer(), 0, DEFAULT_BUFFER_SIZE);
         return;
     }
 
@@ -1057,7 +1060,9 @@ void RasterizerVulkan::SetupGlobalBuffer(const GlobalBufferEntry& entry, GPUVAdd
     if (size == 0) {
         // Sometimes global memory pointers don't have a proper size. Upload a dummy entry
         // because Vulkan doesn't like empty buffers.
-        constexpr std::size_t dummy_size = 4;
+        // 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);
         return;
@@ -1222,7 +1227,7 @@ std::size_t RasterizerVulkan::CalculateVertexArraysSize() const {
         const GPUVAddr end{regs.vertex_array_limit[index].LimitAddress()};
         DEBUG_ASSERT(end >= start);
 
-        size += (end - start + 1) * regs.vertex_array[index].enable;
+        size += (end - start) * regs.vertex_array[index].enable;
     }
     return size;
 }
@@ -1269,4 +1274,29 @@ RenderPassParams RasterizerVulkan::GetRenderPassParams(Texceptions texceptions)
     return renderpass_params;
 }
 
+VkBuffer RasterizerVulkan::DefaultBuffer() {
+    if (default_buffer) {
+        return *default_buffer;
+    }
+
+    VkBufferCreateInfo ci;
+    ci.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
+    ci.pNext = nullptr;
+    ci.flags = 0;
+    ci.size = DEFAULT_BUFFER_SIZE;
+    ci.usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT | VK_BUFFER_USAGE_VERTEX_BUFFER_BIT |
+               VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT;
+    ci.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+    ci.queueFamilyIndexCount = 0;
+    ci.pQueueFamilyIndices = nullptr;
+    default_buffer = device.GetLogical().CreateBuffer(ci);
+    default_buffer_commit = memory_manager.Commit(default_buffer, false);
+
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([buffer = *default_buffer](vk::CommandBuffer cmdbuf) {
+        cmdbuf.FillBuffer(buffer, 0, DEFAULT_BUFFER_SIZE, 0);
+    });
+    return *default_buffer;
+}
+
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -155,6 +155,7 @@ private:
     using Texceptions = std::bitset<Maxwell::NumRenderTargets + 1>;
 
     static constexpr std::size_t ZETA_TEXCEPTION_INDEX = 8;
+    static constexpr VkDeviceSize DEFAULT_BUFFER_SIZE = 4 * sizeof(float);
 
     void FlushWork();
 
@@ -247,6 +248,8 @@ private:
 
     RenderPassParams GetRenderPassParams(Texceptions texceptions) const;
 
+    VkBuffer DefaultBuffer();
+
     Core::System& system;
     Core::Frontend::EmuWindow& render_window;
     VKScreenInfo& screen_info;
@@ -271,6 +274,9 @@ private:
     VKFenceManager fence_manager;
     VKQueryCache query_cache;
 
+    vk::Buffer default_buffer;
+    VKMemoryCommit default_buffer_commit;
+
     std::array<View, Maxwell::NumRenderTargets> color_attachments;
     View zeta_attachment;
 

src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp

@@ -81,7 +81,7 @@ VKBuffer& VKStagingBufferPool::CreateStagingBuffer(std::size_t size, bool host_v
     ci.size = 1ULL << log2;
     ci.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT |
                VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
-               VK_BUFFER_USAGE_INDEX_BUFFER_BIT;
+               VK_BUFFER_USAGE_INDEX_BUFFER_BIT | VK_BUFFER_USAGE_VERTEX_BUFFER_BIT;
     ci.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
     ci.queueFamilyIndexCount = 0;
     ci.pQueueFamilyIndices = nullptr;