buffer_cache: Handle obtaining buffer views partially within buffers.

Author: squidbus

src/video_core/buffer_cache/buffer_cache.cpp

@@ -360,7 +360,8 @@ std::pair<Buffer*, u32> BufferCache::ObtainBuffer(VAddr device_addr, u32 size, b
     return {&buffer, buffer.Offset(device_addr)};
 }
 
-std::pair<Buffer*, u32> BufferCache::ObtainViewBuffer(VAddr gpu_addr, u32 size) {
+std::pair<Buffer*, u32> BufferCache::ObtainViewBuffer(VAddr gpu_addr, u32 size, bool prefer_gpu) {
+    // Check if any buffer contains the full requested range.
     const u64 page = gpu_addr >> CACHING_PAGEBITS;
     const BufferId buffer_id = page_table[page];
     if (buffer_id) {
@@ -370,6 +371,13 @@ std::pair<Buffer*, u32> BufferCache::ObtainViewBuffer(VAddr gpu_addr, u32 size)
             return {&buffer, buffer.Offset(gpu_addr)};
         }
     }
+    // If no buffer contains the full requested range but some buffer within was GPU-modified,
+    // fall back to ObtainBuffer to create a full buffer and avoid losing GPU modifications.
+    // This is only done if the request prefers to use GPU memory, otherwise we can skip it.
+    if (prefer_gpu && memory_tracker.IsRegionGpuModified(gpu_addr, size)) {
+        return ObtainBuffer(gpu_addr, size, false, false);
+    }
+    // In all other cases, just do a CPU copy to the staging buffer.
     const u32 offset = staging_buffer.Copy(gpu_addr, size, 16);
     return {&staging_buffer, offset};
 }

src/video_core/buffer_cache/buffer_cache.h

@@ -96,7 +96,8 @@ class BufferCache {
                                                        BufferId buffer_id = {});
 
     /// Attempts to obtain a buffer without modifying the cache contents.
-    [[nodiscard]] std::pair<Buffer*, u32> ObtainViewBuffer(VAddr gpu_addr, u32 size);
+    [[nodiscard]] std::pair<Buffer*, u32> ObtainViewBuffer(VAddr gpu_addr, u32 size,
+                                                           bool prefer_gpu);
 
     /// Return true when a region is registered on the cache
     [[nodiscard]] bool IsRegionRegistered(VAddr addr, size_t size);

src/video_core/texture_cache/texture_cache.cpp

@@ -466,6 +466,9 @@ void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_schedule
     const auto& num_mips = image.info.resources.levels;
     ASSERT(num_mips == image.info.mips_layout.size());
 
+    const bool is_gpu_modified = True(image.flags & ImageFlagBits::GpuModified);
+    const bool is_gpu_dirty = True(image.flags & ImageFlagBits::GpuDirty);
+
     boost::container::small_vector<vk::BufferImageCopy, 14> image_copy{};
     for (u32 m = 0; m < num_mips; m++) {
         const u32 width = std::max(image.info.size.width >> m, 1u);
@@ -475,8 +478,6 @@ void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_schedule
         const auto& mip = image.info.mips_layout[m];
 
         // Protect GPU modified resources from accidental CPU reuploads.
-        const bool is_gpu_modified = True(image.flags & ImageFlagBits::GpuModified);
-        const bool is_gpu_dirty = True(image.flags & ImageFlagBits::GpuDirty);
         if (is_gpu_modified && !is_gpu_dirty) {
             const u8* addr = std::bit_cast<u8*>(image.info.guest_address);
             const u64 hash = XXH3_64bits(addr + mip.offset, mip.size);
@@ -515,7 +516,8 @@ void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_schedule
 
     const VAddr image_addr = image.info.guest_address;
     const size_t image_size = image.info.guest_size_bytes;
-    const auto [vk_buffer, buf_offset] = buffer_cache.ObtainViewBuffer(image_addr, image_size);
+    const auto [vk_buffer, buf_offset] =
+        buffer_cache.ObtainViewBuffer(image_addr, image_size, is_gpu_dirty);
     // The obtained buffer may be written by a shader so we need to emit a barrier to prevent RAW
     // hazard
     if (auto barrier = vk_buffer->GetBarrier(vk::AccessFlagBits2::eTransferRead,