Merge pull request #7720 from FernandoS27/yfc-gc

First Nugget: Reworked Garbage Collection to be smarter [originally from Project YFC]

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -135,6 +135,20 @@ BufferCacheRuntime::BufferCacheRuntime(const Device& device_)
         buffer.Create();
         glNamedBufferData(buffer.handle, 0x10'000, nullptr, GL_STREAM_COPY);
     }
+
+    device_access_memory = [this]() -> u64 {
+        if (device.CanReportMemoryUsage()) {
+            return device.GetCurrentDedicatedVideoMemory() + 512_MiB;
+        }
+        return 2_GiB; // Return minimum requirements
+    }();
+}
+
+u64 BufferCacheRuntime::GetDeviceMemoryUsage() const {
+    if (device.CanReportMemoryUsage()) {
+        return device_access_memory - device.GetCurrentDedicatedVideoMemory();
+    }
+    return 2_GiB;
 }
 
 void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -89,6 +89,8 @@ public:
     void BindImageBuffer(Buffer& buffer, u32 offset, u32 size,
                          VideoCore::Surface::PixelFormat format);
 
+    u64 GetDeviceMemoryUsage() const;
+
     void BindFastUniformBuffer(size_t stage, u32 binding_index, u32 size) {
         const GLuint handle = fast_uniforms[stage][binding_index].handle;
         const GLsizeiptr gl_size = static_cast<GLsizeiptr>(size);
@@ -151,6 +153,14 @@ public:
         use_storage_buffers = use_storage_buffers_;
     }
 
+    u64 GetDeviceLocalMemory() const {
+        return device_access_memory;
+    }
+
+    bool CanReportMemoryUsage() const {
+        return device.CanReportMemoryUsage();
+    }
+
 private:
     static constexpr std::array PABO_LUT{
         GL_VERTEX_PROGRAM_PARAMETER_BUFFER_NV,          GL_TESS_CONTROL_PROGRAM_PARAMETER_BUFFER_NV,
@@ -184,6 +194,8 @@ private:
     std::array<OGLBuffer, VideoCommon::NUM_COMPUTE_UNIFORM_BUFFERS> copy_compute_uniforms;
 
     u32 index_buffer_offset = 0;
+
+    u64 device_access_memory;
 };
 
 struct BufferCacheParams {

src/video_core/renderer_opengl/gl_device.cpp

@@ -13,12 +13,15 @@
 
 #include <glad/glad.h>
 
+#include "common/literals.h"
 #include "common/logging/log.h"
 #include "common/settings.h"
 #include "shader_recompiler/stage.h"
 #include "video_core/renderer_opengl/gl_device.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 
+using namespace Common::Literals;
+
 namespace OpenGL {
 namespace {
 constexpr std::array LIMIT_UBOS = {
@@ -165,6 +168,7 @@ Device::Device() {
     has_sparse_texture_2 = GLAD_GL_ARB_sparse_texture2;
     warp_size_potentially_larger_than_guest = !is_nvidia && !is_intel;
     need_fastmath_off = is_nvidia;
+    can_report_memory = GLAD_GL_NVX_gpu_memory_info;
 
     // At the moment of writing this, only Nvidia's driver optimizes BufferSubData on exclusive
     // uniform buffers as "push constants"
@@ -276,4 +280,10 @@ void main() {
 })");
 }
 
+u64 Device::GetCurrentDedicatedVideoMemory() const {
+    GLint cur_avail_mem_kb = 0;
+    glGetIntegerv(GL_GPU_MEMORY_INFO_DEDICATED_VIDMEM_NVX, &cur_avail_mem_kb);
+    return static_cast<u64>(cur_avail_mem_kb) * 1_KiB;
+}
+
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_device.h

@@ -20,6 +20,8 @@ public:
 
     [[nodiscard]] std::string GetVendorName() const;
 
+    u64 GetCurrentDedicatedVideoMemory() const;
+
     u32 GetMaxUniformBuffers(Shader::Stage stage) const noexcept {
         return max_uniform_buffers[static_cast<size_t>(stage)];
     }
@@ -168,6 +170,10 @@ public:
         return vendor_name == "ATI Technologies Inc.";
     }
 
+    bool CanReportMemoryUsage() const {
+        return can_report_memory;
+    }
+
 private:
     static bool TestVariableAoffi();
     static bool TestPreciseBug();
@@ -210,6 +216,7 @@ private:
     bool need_fastmath_off{};
     bool has_cbuf_ftou_bug{};
     bool has_bool_ref_bug{};
+    bool can_report_memory{};
 
     std::string vendor_name;
 };

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -484,6 +484,13 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, ProgramManager&
             rescale_read_fbos[i].Create();
         }
     }
+
+    device_access_memory = [this]() -> u64 {
+        if (device.CanReportMemoryUsage()) {
+            return device.GetCurrentDedicatedVideoMemory() + 512_MiB;
+        }
+        return 2_GiB; // Return minimum requirements
+    }();
 }
 
 TextureCacheRuntime::~TextureCacheRuntime() = default;
@@ -500,13 +507,11 @@ ImageBufferMap TextureCacheRuntime::DownloadStagingBuffer(size_t size) {
     return download_buffers.RequestMap(size, false);
 }
 
-u64 TextureCacheRuntime::GetDeviceLocalMemory() const {
-    if (GLAD_GL_NVX_gpu_memory_info) {
-        GLint cur_avail_mem_kb = 0;
-        glGetIntegerv(GL_GPU_MEMORY_INFO_CURRENT_AVAILABLE_VIDMEM_NVX, &cur_avail_mem_kb);
-        return static_cast<u64>(cur_avail_mem_kb) * 1_KiB;
+u64 TextureCacheRuntime::GetDeviceMemoryUsage() const {
+    if (device.CanReportMemoryUsage()) {
+        return device_access_memory - device.GetCurrentDedicatedVideoMemory();
     }
-    return 2_GiB; // Return minimum requirements
+    return 2_GiB;
 }
 
 void TextureCacheRuntime::CopyImage(Image& dst_image, Image& src_image,
@@ -686,6 +691,7 @@ Image::Image(TextureCacheRuntime& runtime_, const VideoCommon::ImageInfo& info_,
     }
     if (IsConverted(runtime->device, info.format, info.type)) {
         flags |= ImageFlagBits::Converted;
+        flags |= ImageFlagBits::CostlyLoad;
         gl_internal_format = IsPixelFormatSRGB(info.format) ? GL_SRGB8_ALPHA8 : GL_RGBA8;
         gl_format = GL_RGBA;
         gl_type = GL_UNSIGNED_INT_8_8_8_8_REV;

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -10,6 +10,7 @@
 #include <glad/glad.h>
 
 #include "shader_recompiler/shader_info.h"
+#include "video_core/renderer_opengl/gl_device.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/util_shaders.h"
 #include "video_core/texture_cache/image_view_base.h"
@@ -21,7 +22,6 @@ struct ResolutionScalingInfo;
 
 namespace OpenGL {
 
-class Device;
 class ProgramManager;
 class StateTracker;
 
@@ -83,7 +83,15 @@ public:
 
     ImageBufferMap DownloadStagingBuffer(size_t size);
 
-    u64 GetDeviceLocalMemory() const;
+    u64 GetDeviceLocalMemory() const {
+        return device_access_memory;
+    }
+
+    u64 GetDeviceMemoryUsage() const;
+
+    bool CanReportMemoryUsage() const {
+        return device.CanReportMemoryUsage();
+    }
 
     bool ShouldReinterpret([[maybe_unused]] Image& dst, [[maybe_unused]] Image& src) {
         return true;
@@ -172,6 +180,7 @@ private:
     std::array<OGLFramebuffer, 4> rescale_draw_fbos;
     std::array<OGLFramebuffer, 4> rescale_read_fbos;
     const Settings::ResolutionScalingInfo& resolution;
+    u64 device_access_memory;
 };
 
 class Image : public VideoCommon::ImageBase {