early-access version 3805

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 3804.
+This is the source code for early-access 3805.
 
 ## Legal Notice
 

src/core/hle/service/nvnflinger/buffer_queue_producer.cpp

@@ -449,6 +449,7 @@ Status BufferQueueProducer::QueueBuffer(s32 slot, const QueueBufferInput& input,
     case NativeWindowScalingMode::ScaleToWindow:
     case NativeWindowScalingMode::ScaleCrop:
     case NativeWindowScalingMode::NoScaleCrop:
+    case NativeWindowScalingMode::PreserveAspectRatio:
         break;
     default:
         LOG_ERROR(Service_Nvnflinger, "unknown scaling mode {}", scaling_mode);

src/core/hle/service/nvnflinger/window.h

@@ -41,6 +41,7 @@ enum class NativeWindowScalingMode : s32 {
     ScaleToWindow = 1,
     ScaleCrop = 2,
     NoScaleCrop = 3,
+    PreserveAspectRatio = 4,
 };
 
 /// Transform parameter for QueueBuffer

src/video_core/renderer_opengl/gl_graphics_pipeline.cpp

@@ -220,7 +220,8 @@ GraphicsPipeline::GraphicsPipeline(const Device& device, TextureCache& texture_c
     ASSERT(num_textures <= MAX_TEXTURES);
     ASSERT(num_images <= MAX_IMAGES);
 
-    const bool assembly_shaders{assembly_programs[0].handle != 0};
+    const auto backend = device.GetShaderBackend();
+    const bool assembly_shaders{backend == Settings::ShaderBackend::Glasm};
     use_storage_buffers =
         !assembly_shaders || num_storage_buffers <= device.GetMaxGLASMStorageBufferBlocks();
     writes_global_memory &= !use_storage_buffers;
@@ -230,7 +231,6 @@ GraphicsPipeline::GraphicsPipeline(const Device& device, TextureCache& texture_c
         GenerateTransformFeedbackState();
     }
     const bool in_parallel = thread_worker != nullptr;
-    const auto backend = device.GetShaderBackend();
     auto func{[this, sources_ = std::move(sources), sources_spirv_ = std::move(sources_spirv),
                shader_notify, backend, in_parallel,
                force_context_flush](ShaderContext::Context*) mutable {
@@ -559,15 +559,13 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
 }
 
 void GraphicsPipeline::ConfigureTransformFeedbackImpl() const {
-    glTransformFeedbackStreamAttribsNV(num_xfb_attribs, xfb_attribs.data(), num_xfb_strides,
-                                       xfb_streams.data(), GL_INTERLEAVED_ATTRIBS);
+    glTransformFeedbackAttribsNV(num_xfb_attribs, xfb_attribs.data(), GL_SEPARATE_ATTRIBS);
 }
 
 void GraphicsPipeline::GenerateTransformFeedbackState() {
     // TODO(Rodrigo): Inject SKIP_COMPONENTS*_NV when required. An unimplemented message will signal
     // when this is required.
     GLint* cursor{xfb_attribs.data()};
-    GLint* current_stream{xfb_streams.data()};
 
     for (size_t feedback = 0; feedback < Maxwell::NumTransformFeedbackBuffers; ++feedback) {
         const auto& layout = key.xfb_state.layouts[feedback];
@@ -575,15 +573,6 @@ void GraphicsPipeline::GenerateTransformFeedbackState() {
         if (layout.varying_count == 0) {
             continue;
         }
-        *current_stream = static_cast<GLint>(feedback);
-        if (current_stream != xfb_streams.data()) {
-            // When stepping one stream, push the expected token
-            cursor[0] = GL_NEXT_BUFFER_NV;
-            cursor[1] = 0;
-            cursor[2] = 0;
-            cursor += XFB_ENTRY_STRIDE;
-        }
-        ++current_stream;
 
         const auto& locations = key.xfb_state.varyings[feedback];
         std::optional<u32> current_index;
@@ -619,7 +608,6 @@ void GraphicsPipeline::GenerateTransformFeedbackState() {
         }
     }
     num_xfb_attribs = static_cast<GLsizei>((cursor - xfb_attribs.data()) / XFB_ENTRY_STRIDE);
-    num_xfb_strides = static_cast<GLsizei>(current_stream - xfb_streams.data());
 }
 
 void GraphicsPipeline::WaitForBuild() {

src/video_core/renderer_opengl/gl_graphics_pipeline.h

@@ -154,9 +154,7 @@ private:
 
     static constexpr std::size_t XFB_ENTRY_STRIDE = 3;
     GLsizei num_xfb_attribs{};
-    GLsizei num_xfb_strides{};
     std::array<GLint, 128 * XFB_ENTRY_STRIDE * Maxwell::NumTransformFeedbackBuffers> xfb_attribs{};
-    std::array<GLint, Maxwell::NumTransformFeedbackBuffers> xfb_streams{};
 
     std::mutex built_mutex;
     std::condition_variable built_condvar;

src/video_core/renderer_opengl/util_shaders.cpp

@@ -68,6 +68,7 @@ void UtilShaders::ASTCDecode(Image& image, const StagingBufferMap& map,
                              std::span<const VideoCommon::SwizzleParameters> swizzles) {
     static constexpr GLuint BINDING_INPUT_BUFFER = 0;
     static constexpr GLuint BINDING_OUTPUT_IMAGE = 0;
+    program_manager.LocalMemoryWarmup();
 
     const Extent2D tile_size{
         .width = VideoCore::Surface::DefaultBlockWidth(image.info.format),