From 9edbbf2af401f821c0be6a266e65975e3de25fb3 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 16 Jan 2022 06:34:43 +0100
Subject: [PATCH] Garbage Collection: Final tuning.

---
 src/video_core/buffer_cache/buffer_cache.h    |  6 ++--
 .../renderer_opengl/gl_buffer_cache.cpp       |  2 +-
 .../renderer_opengl/gl_texture_cache.cpp      |  4 +--
 .../renderer_vulkan/vk_texture_cache.cpp      |  2 +-
 src/video_core/texture_cache/image_base.h     | 10 +++---
 src/video_core/texture_cache/texture_cache.h  | 36 ++++++++++++-------
 6 files changed, 36 insertions(+), 24 deletions(-)

diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 644c6e57e..21bfb76a4 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -464,9 +464,9 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
     const s64 device_memory = static_cast<s64>(runtime.GetDeviceLocalMemory());
     const s64 min_spacing_expected = device_memory - 1_GiB - 512_MiB;
     const s64 min_spacing_critical = device_memory - 1_GiB;
-    const s64 mem_tresshold = std::min(device_memory, TARGET_THRESHOLD);
-    const s64 min_vacancy_expected = (6 * mem_tresshold) / 10;
-    const s64 min_vacancy_critical = (3 * mem_tresshold) / 10;
+    const s64 mem_threshold = std::min(device_memory, TARGET_THRESHOLD);
+    const s64 min_vacancy_expected = (6 * mem_threshold) / 10;
+    const s64 min_vacancy_critical = (3 * mem_threshold) / 10;
     minimum_memory = static_cast<u64>(
         std::max(std::min(device_memory - min_vacancy_expected, min_spacing_expected),
                  DEFAULT_EXPECTED_MEMORY));
diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.cpp b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
index 45db72e6f..f73b0af5f 100644
--- a/src/video_core/renderer_opengl/gl_buffer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
@@ -150,7 +150,7 @@ u64 BufferCacheRuntime::GetDeviceMemoryUsage() 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;
+        return device_access_memory - static_cast<u64>(cur_avail_mem_kb) * 1_KiB;
     }
     return 2_GiB;
 }
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp
index 40f52eacb..63586d9d5 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp
@@ -513,7 +513,7 @@ u64 TextureCacheRuntime::GetDeviceMemoryUsage() 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;
+        return device_access_memory - static_cast<u64>(cur_avail_mem_kb) * 1_KiB;
     }
     return 2_GiB;
 }
@@ -695,7 +695,7 @@ Image::Image(TextureCacheRuntime& runtime_, const VideoCommon::ImageInfo& info_,
     }
     if (IsConverted(runtime->device, info.format, info.type)) {
         flags |= ImageFlagBits::Converted;
-        flags |= ImageFlagBits::GCProtected;
+        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;
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index 884f501ef..f2890d263 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -1211,7 +1211,7 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
         } else {
             flags |= VideoCommon::ImageFlagBits::Converted;
         }
-        flags |= VideoCommon::ImageFlagBits::GCProtected;
+        flags |= VideoCommon::ImageFlagBits::CostlyLoad;
     }
     if (runtime->device.HasDebuggingToolAttached()) {
         original_image.SetObjectNameEXT(VideoCommon::Name(*this).c_str());
diff --git a/src/video_core/texture_cache/image_base.h b/src/video_core/texture_cache/image_base.h
index 279f39269..dd0106432 100644
--- a/src/video_core/texture_cache/image_base.h
+++ b/src/video_core/texture_cache/image_base.h
@@ -29,11 +29,11 @@ enum class ImageFlagBits : u32 {
     Sparse = 1 << 9,      ///< Image has non continous submemory.
 
     // Garbage Collection Flags
-    BadOverlap = 1 << 10,  ///< This image overlaps other but doesn't fit, has higher
-                           ///< garbage collection priority
-    Alias = 1 << 11,       ///< This image has aliases and has priority on garbage
-                           ///< collection
-    GCProtected = 1 << 12, ///< Protected from low-tier GC as they are costy to load back.
+    BadOverlap = 1 << 10, ///< This image overlaps other but doesn't fit, has higher
+                          ///< garbage collection priority
+    Alias = 1 << 11,      ///< This image has aliases and has priority on garbage
+                          ///< collection
+    CostlyLoad = 1 << 12, ///< Protected from low-tier GC as it is costly to load back.
 
     // Rescaler
     Rescaled = 1 << 13,
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 7b6bd8697..efc1c4525 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -53,17 +53,16 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
         const s64 device_memory = static_cast<s64>(runtime.GetDeviceLocalMemory());
         const s64 min_spacing_expected = device_memory - 1_GiB - 512_MiB;
         const s64 min_spacing_critical = device_memory - 1_GiB;
-        const s64 mem_tresshold = std::min(device_memory, TARGET_THRESHOLD);
-        const s64 min_vacancy_expected = (6 * mem_tresshold) / 10;
-        const s64 min_vacancy_critical = (3 * mem_tresshold) / 10;
+        const s64 mem_threshold = std::min(device_memory, TARGET_THRESHOLD);
+        const s64 min_vacancy_expected = (6 * mem_threshold) / 10;
+        const s64 min_vacancy_critical = (3 * mem_threshold) / 10;
         expected_memory = static_cast<u64>(
             std::max(std::min(device_memory - min_vacancy_expected, min_spacing_expected),
                      DEFAULT_EXPECTED_MEMORY));
         critical_memory = static_cast<u64>(
             std::max(std::min(device_memory - min_vacancy_critical, min_spacing_critical),
                      DEFAULT_CRITICAL_MEMORY));
-        minimum_memory = static_cast<u64>((device_memory - mem_tresshold) / 2);
-        LOG_CRITICAL(Debug, "Available Memory: {}", device_memory / 1_MiB);
+        minimum_memory = static_cast<u64>((device_memory - mem_threshold) / 2);
     } else {
         expected_memory = DEFAULT_EXPECTED_MEMORY + 512_MiB;
         critical_memory = DEFAULT_CRITICAL_MEMORY + 1_GiB;
@@ -73,11 +72,12 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
 
 template <class P>
 void TextureCache<P>::RunGarbageCollector() {
-    const bool high_priority_mode = total_used_memory >= expected_memory;
-    const bool aggressive_mode = total_used_memory >= critical_memory;
-    const u64 ticks_to_destroy = aggressive_mode ? 10ULL : high_priority_mode ? 25ULL : 100ULL;
-    size_t num_iterations = aggressive_mode ? 300 : (high_priority_mode ? 50 : 10);
-    const auto clean_up = [this, &num_iterations, high_priority_mode](ImageId image_id) {
+    bool high_priority_mode = total_used_memory >= expected_memory;
+    bool aggressive_mode = total_used_memory >= critical_memory;
+    const u64 ticks_to_destroy = aggressive_mode ? 10ULL : high_priority_mode ? 25ULL : 50ULL;
+    size_t num_iterations = aggressive_mode ? 40 : (high_priority_mode ? 20 : 10);
+    const auto clean_up = [this, &num_iterations, &high_priority_mode,
+                           &aggressive_mode](ImageId image_id) {
         if (num_iterations == 0) {
             return true;
         }
@@ -85,7 +85,8 @@ void TextureCache<P>::RunGarbageCollector() {
         auto& image = slot_images[image_id];
         const bool must_download =
             image.IsSafeDownload() && False(image.flags & ImageFlagBits::BadOverlap);
-        if (!high_priority_mode && must_download) {
+        if (!high_priority_mode &&
+            (must_download || True(image.flags & ImageFlagBits::CostlyLoad))) {
             return false;
         }
         if (must_download) {
@@ -100,6 +101,18 @@ void TextureCache<P>::RunGarbageCollector() {
         }
         UnregisterImage(image_id);
         DeleteImage(image_id, image.scale_tick > frame_tick + 5);
+        if (total_used_memory < critical_memory) {
+            if (aggressive_mode) {
+                // Sink the aggresiveness.
+                num_iterations >>= 2;
+                aggressive_mode = false;
+                return false;
+            }
+            if (high_priority_mode && total_used_memory < expected_memory) {
+                num_iterations >>= 1;
+                high_priority_mode = false;
+            }
+        }
         return false;
     };
     lru_cache.ForEachItemBelow(frame_tick - ticks_to_destroy, clean_up);
@@ -120,7 +133,6 @@ void TextureCache<P>::TickFrame() {
     runtime.TickFrame();
     critical_gc = 0;
     ++frame_tick;
-    LOG_CRITICAL(Debug, "Current memory: {}", total_used_memory / 1_MiB);
 }
 
 template <class P>