video_core: Crucial buffer cache fixes + proper GPU clears (#414)

* translator: Use templates for stronger type guarantees

* spirv: Define buffer offsets upfront

* Saves a lot of shader instructions

* buffer_cache: Use dynamic vertex input when available

* Fixes issues when games like dark souls rebind vertex buffers with different stride

* externals: Update boost

* spirv: Use runtime array for ssbos

* ssbos can be large and typically their size will vary, especially in generic copy/clear cs shaders

* fs: Lock when doing case insensitive search

* Dark Souls does fs lookups from different threads

* texture_cache: More precise invalidation from compute

* Fixes unrelated render targets being cleared

* texture_cache: Use hashes for protect gpu modified images from reupload

* translator: Treat V_CNDMASK as float

* Sometimes it can have input modifiers. Worst this will cause is some extra calls to uintBitsToFloat and opposite. But most often this is used as float anyway

* translator: Small optimization for V_SAD_U32

* Fix review

* clang format

src/video_core/renderer_vulkan/renderer_vulkan.h

@@ -47,7 +47,7 @@ public:
     Frame* PrepareFrame(const Libraries::VideoOut::BufferAttributeGroup& attribute,
                         VAddr cpu_address, bool is_eop) {
         const auto info = VideoCore::ImageInfo{attribute, cpu_address};
-        const auto image_id = texture_cache.FindImage(info, false);
+        const auto image_id = texture_cache.FindImage(info);
         auto& image = texture_cache.GetImage(image_id);
         return PrepareFrameInternal(image, is_eop);
     }
@@ -61,7 +61,7 @@ public:
         const Libraries::VideoOut::BufferAttributeGroup& attribute, VAddr cpu_address) {
         vo_buffers_addr.emplace_back(cpu_address);
         const auto info = VideoCore::ImageInfo{attribute, cpu_address};
-        const auto image_id = texture_cache.FindImage(info, false);
+        const auto image_id = texture_cache.FindImage(info);
         return texture_cache.GetImage(image_id);
     }
 

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -96,7 +96,7 @@ bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
     Shader::PushData push_data{};
     u32 binding{};
 
-    for (u32 i = 0; const auto& buffer : info.buffers) {
+    for (const auto& buffer : info.buffers) {
         const auto vsharp = buffer.GetVsharp(info);
         const VAddr address = vsharp.base_address;
         // Most of the time when a metadata is updated with a shader it gets cleared. It means we
@@ -115,7 +115,7 @@ bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
         }
         const u32 size = vsharp.GetSize();
         if (buffer.is_written) {
-            texture_cache.InvalidateMemory(address, size);
+            texture_cache.InvalidateMemory(address, size, true);
         }
         const u32 alignment =
             buffer.is_storage ? instance.StorageMinAlignment() : instance.UniformMinAlignment();
@@ -137,7 +137,6 @@ bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
                                                 : vk::DescriptorType::eUniformBuffer,
             .pBufferInfo = &buffer_infos.back(),
         });
-        i++;
     }
 
     for (const auto& image_desc : info.images) {

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -145,6 +145,9 @@ GraphicsPipeline::GraphicsPipeline(const Instance& instance_, Scheduler& schedul
         dynamic_states.push_back(vk::DynamicState::eColorWriteEnableEXT);
         dynamic_states.push_back(vk::DynamicState::eColorWriteMaskEXT);
     }
+    if (instance.IsVertexInputDynamicState()) {
+        dynamic_states.push_back(vk::DynamicState::eVertexInputEXT);
+    }
 
     const vk::PipelineDynamicStateCreateInfo dynamic_info = {
         .dynamicStateCount = static_cast<u32>(dynamic_states.size()),

src/video_core/renderer_vulkan/vk_instance.cpp

@@ -202,6 +202,8 @@ bool Instance::CreateDevice() {
     add_extension(VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME);
     workgroup_memory_explicit_layout =
         add_extension(VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME);
+    vertex_input_dynamic_state = add_extension(VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
+
     // The next two extensions are required to be available together in order to support write masks
     color_write_en = add_extension(VK_EXT_COLOR_WRITE_ENABLE_EXTENSION_NAME);
     color_write_en &= add_extension(VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME);
@@ -319,6 +321,9 @@ bool Instance::CreateDevice() {
         vk::PhysicalDeviceSynchronization2Features{
             .synchronization2 = true,
         },
+        vk::PhysicalDeviceVertexInputDynamicStateFeaturesEXT{
+            .vertexInputDynamicState = true,
+        },
     };
 
     if (!color_write_en) {
@@ -331,8 +336,8 @@ bool Instance::CreateDevice() {
     } else {
         device_chain.unlink<vk::PhysicalDeviceRobustness2FeaturesEXT>();
     }
-    if (!has_sync2) {
-        device_chain.unlink<vk::PhysicalDeviceSynchronization2Features>();
+    if (!vertex_input_dynamic_state) {
+        device_chain.unlink<vk::PhysicalDeviceVertexInputDynamicStateFeaturesEXT>();
     }
 
     try {

src/video_core/renderer_vulkan/vk_instance.h

@@ -132,6 +132,11 @@ public:
         return color_write_en;
     }
 
+    /// Returns true when VK_EXT_vertex_input_dynamic_state is supported.
+    bool IsVertexInputDynamicState() const {
+        return vertex_input_dynamic_state;
+    }
+
     /// Returns the vendor ID of the physical device
     u32 GetVendorID() const {
         return properties.vendorID;
@@ -257,6 +262,7 @@ private:
     bool external_memory_host{};
     bool workgroup_memory_explicit_layout{};
     bool color_write_en{};
+    bool vertex_input_dynamic_state{};
     u64 min_imported_host_pointer_alignment{};
     u32 subgroup_size{};
     bool tooling_info{};

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -209,6 +209,10 @@ void PipelineCache::RefreshGraphicsKey() {
             continue;
         }
         const auto* bininfo = Liverpool::GetBinaryInfo(*pgm);
+        if (!bininfo->Valid()) {
+            key.stage_hashes[i] = 0;
+            continue;
+        }
         key.stage_hashes[i] = bininfo->shader_hash;
     }
 }