Merge branch 'master' into ssbo-align

46 files changed, 1168 insertions, 596 deletions

diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index cf9266d54..c22c7631c 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -4,7 +4,7 @@
 add_subdirectory(host_shaders)
 
 if(LIBVA_FOUND)
-    set_source_files_properties(host1x/codecs/codec.cpp
+    set_source_files_properties(host1x/ffmpeg/ffmpeg.cpp
         PROPERTIES COMPILE_DEFINITIONS LIBVA_FOUND=1)
     list(APPEND FFmpeg_LIBRARIES ${LIBVA_LIBRARIES})
 endif()
@@ -15,6 +15,7 @@ add_library(video_core STATIC
     buffer_cache/buffer_cache.cpp
     buffer_cache/buffer_cache.h
     buffer_cache/memory_tracker_base.h
+    buffer_cache/usage_tracker.h
     buffer_cache/word_manager.h
     cache_types.h
     cdma_pusher.cpp
@@ -66,6 +67,8 @@ add_library(video_core STATIC
     host1x/codecs/vp9.cpp
     host1x/codecs/vp9.h
     host1x/codecs/vp9_types.h
+    host1x/ffmpeg/ffmpeg.cpp
+    host1x/ffmpeg/ffmpeg.h
     host1x/control.cpp
     host1x/control.h
     host1x/host1x.cpp
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 2648970b6..6d1fc3887 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -67,6 +67,7 @@ void BufferCache<P>::TickFrame() {
     if (!channel_state) {
         return;
     }
+    runtime.TickFrame(slot_buffers);
 
     // Calculate hits and shots and move hit bits to the right
     const u32 hits = std::reduce(channel_state->uniform_cache_hits.begin(),
@@ -230,7 +231,10 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     for (const IntervalType& add_interval : tmp_intervals) {
         common_ranges.add(add_interval);
     }
-    runtime.CopyBuffer(dest_buffer, src_buffer, copies);
+    const auto& copy = copies[0];
+    src_buffer.MarkUsage(copy.src_offset, copy.size);
+    dest_buffer.MarkUsage(copy.dst_offset, copy.size);
+    runtime.CopyBuffer(dest_buffer, src_buffer, copies, true);
     if (has_new_downloads) {
         memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount);
     }
@@ -258,9 +262,10 @@ bool BufferCache<P>::DMAClear(GPUVAddr dst_address, u64 amount, u32 value) {
     common_ranges.subtract(subtract_interval);
 
     const BufferId buffer = FindBuffer(*cpu_dst_address, static_cast<u32>(size));
-    auto& dest_buffer = slot_buffers[buffer];
+    Buffer& dest_buffer = slot_buffers[buffer];
     const u32 offset = dest_buffer.Offset(*cpu_dst_address);
     runtime.ClearBuffer(dest_buffer, offset, size, value);
+    dest_buffer.MarkUsage(offset, size);
     return true;
 }
 
@@ -603,6 +608,7 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
             VAddr orig_cpu_addr = static_cast<VAddr>(second_copy.src_offset);
             const IntervalType base_interval{orig_cpu_addr, orig_cpu_addr + copy.size};
             async_downloads += std::make_pair(base_interval, 1);
+            buffer.MarkUsage(copy.src_offset, copy.size);
             runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
             normalized_copies.push_back(second_copy);
         }
@@ -621,8 +627,9 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
                     // Have in mind the staging buffer offset for the copy
                     copy.dst_offset += download_staging.offset;
                     const std::array copies{copy};
-                    runtime.CopyBuffer(download_staging.buffer, slot_buffers[buffer_id], copies,
-                                       false);
+                    Buffer& buffer = slot_buffers[buffer_id];
+                    buffer.MarkUsage(copy.src_offset, copy.size);
+                    runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
                 }
                 runtime.PostCopyBarrier();
                 runtime.Finish();
@@ -742,7 +749,7 @@ void BufferCache<P>::BindHostIndexBuffer() {
                 {BufferCopy{.src_offset = upload_staging.offset, .dst_offset = 0, .size = size}}};
             std::memcpy(upload_staging.mapped_span.data(),
                         draw_state.inline_index_draw_indexes.data(), size);
-            runtime.CopyBuffer(buffer, upload_staging.buffer, copies);
+            runtime.CopyBuffer(buffer, upload_staging.buffer, copies, true);
         } else {
             buffer.ImmediateUpload(0, draw_state.inline_index_draw_indexes);
         }
@@ -754,6 +761,7 @@ void BufferCache<P>::BindHostIndexBuffer() {
             offset + draw_state.index_buffer.first * draw_state.index_buffer.FormatSizeInBytes();
         runtime.BindIndexBuffer(buffer, new_offset, size);
     } else {
+        buffer.MarkUsage(offset, size);
         runtime.BindIndexBuffer(draw_state.topology, draw_state.index_buffer.format,
                                 draw_state.index_buffer.first, draw_state.index_buffer.count,
                                 buffer, offset, size);
@@ -790,6 +798,7 @@ void BufferCache<P>::BindHostVertexBuffers() {
 
             const u32 stride = maxwell3d->regs.vertex_streams[index].stride;
             const u32 offset = buffer.Offset(binding.cpu_addr);
+            buffer.MarkUsage(offset, binding.size);
 
             host_bindings.buffers.push_back(&buffer);
             host_bindings.offsets.push_back(offset);
@@ -895,6 +904,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
     if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
         channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
     }
+    buffer.MarkUsage(offset, size);
     if constexpr (NEEDS_BIND_UNIFORM_INDEX) {
         runtime.BindUniformBuffer(stage, binding_index, buffer, offset, size);
     } else {
@@ -913,6 +923,7 @@ void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) {
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
+        buffer.MarkUsage(offset, size);
         const bool is_written = ((channel_state->written_storage_buffers[stage] >> index) & 1) != 0;
 
         if (is_written) {
@@ -943,6 +954,7 @@ void BufferCache<P>::BindHostGraphicsTextureBuffers(size_t stage) {
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
         const PixelFormat format = binding.format;
+        buffer.MarkUsage(offset, size);
         if constexpr (SEPARATE_IMAGE_BUFFERS_BINDINGS) {
             if (((channel_state->image_texture_buffers[stage] >> index) & 1) != 0) {
                 runtime.BindImageBuffer(buffer, offset, size, format);
@@ -975,9 +987,10 @@ void BufferCache<P>::BindHostTransformFeedbackBuffers() {
         MarkWrittenBuffer(binding.buffer_id, binding.cpu_addr, size);
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
+        buffer.MarkUsage(offset, size);
         host_bindings.buffers.push_back(&buffer);
         host_bindings.offsets.push_back(offset);
-        host_bindings.sizes.push_back(binding.size);
+        host_bindings.sizes.push_back(size);
     }
     if (host_bindings.buffers.size() > 0) {
         runtime.BindTransformFeedbackBuffers(host_bindings);
@@ -1001,6 +1014,7 @@ void BufferCache<P>::BindHostComputeUniformBuffers() {
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
+        buffer.MarkUsage(offset, size);
         if constexpr (NEEDS_BIND_UNIFORM_INDEX) {
             runtime.BindComputeUniformBuffer(binding_index, buffer, offset, size);
             ++binding_index;
@@ -1021,6 +1035,7 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
+        buffer.MarkUsage(offset, size);
         const bool is_written =
             ((channel_state->written_compute_storage_buffers >> index) & 1) != 0;
 
@@ -1053,6 +1068,7 @@ void BufferCache<P>::BindHostComputeTextureBuffers() {
 
         const u32 offset = buffer.Offset(binding.cpu_addr);
         const PixelFormat format = binding.format;
+        buffer.MarkUsage(offset, size);
         if constexpr (SEPARATE_IMAGE_BUFFERS_BINDINGS) {
             if (((channel_state->image_compute_texture_buffers >> index) & 1) != 0) {
                 runtime.BindImageBuffer(buffer, offset, size, format);
@@ -1172,10 +1188,11 @@ void BufferCache<P>::UpdateVertexBuffer(u32 index) {
     if (!gpu_memory->IsWithinGPUAddressRange(gpu_addr_end)) {
         size = static_cast<u32>(gpu_memory->MaxContinuousRange(gpu_addr_begin, size));
     }
+    const BufferId buffer_id = FindBuffer(*cpu_addr, size);
     channel_state->vertex_buffers[index] = Binding{
         .cpu_addr = *cpu_addr,
         .size = size,
-        .buffer_id = FindBuffer(*cpu_addr, size),
+        .buffer_id = buffer_id,
     };
 }
 
@@ -1192,11 +1209,6 @@ void BufferCache<P>::UpdateDrawIndirect() {
             .size = static_cast<u32>(size),
             .buffer_id = FindBuffer(*cpu_addr, static_cast<u32>(size)),
         };
-        VAddr cpu_addr_start = Common::AlignDown(*cpu_addr, 64);
-        VAddr cpu_addr_end = Common::AlignUp(*cpu_addr + size, 64);
-        IntervalType interval{cpu_addr_start, cpu_addr_end};
-        ClearDownload(interval);
-        common_ranges.subtract(interval);
     };
     if (current_draw_indirect->include_count) {
         update(current_draw_indirect->count_start_address, sizeof(u32),
@@ -1406,7 +1418,8 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id,
         .dst_offset = dst_base_offset,
         .size = overlap.SizeBytes(),
     });
-    runtime.CopyBuffer(new_buffer, overlap, copies);
+    new_buffer.MarkUsage(copies[0].dst_offset, copies[0].size);
+    runtime.CopyBuffer(new_buffer, overlap, copies, true);
     DeleteBuffer(overlap_id, true);
 }
 
@@ -1419,7 +1432,9 @@ BufferId BufferCache<P>::CreateBuffer(VAddr cpu_addr, u32 wanted_size) {
     const u32 size = static_cast<u32>(overlap.end - overlap.begin);
     const BufferId new_buffer_id = slot_buffers.insert(runtime, rasterizer, overlap.begin, size);
     auto& new_buffer = slot_buffers[new_buffer_id];
-    runtime.ClearBuffer(new_buffer, 0, new_buffer.SizeBytes(), 0);
+    const size_t size_bytes = new_buffer.SizeBytes();
+    runtime.ClearBuffer(new_buffer, 0, size_bytes, 0);
+    new_buffer.MarkUsage(0, size_bytes);
     for (const BufferId overlap_id : overlap.ids) {
         JoinOverlap(new_buffer_id, overlap_id, !overlap.has_stream_leap);
     }
@@ -1472,11 +1487,6 @@ void BufferCache<P>::TouchBuffer(Buffer& buffer, BufferId buffer_id) noexcept {
 
 template <class P>
 bool BufferCache<P>::SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size) {
-    return SynchronizeBufferImpl(buffer, cpu_addr, size);
-}
-
-template <class P>
-bool BufferCache<P>::SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 size) {
     boost::container::small_vector<BufferCopy, 4> copies;
     u64 total_size_bytes = 0;
     u64 largest_copy = 0;
@@ -1499,51 +1509,6 @@ bool BufferCache<P>::SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 s
 }
 
 template <class P>
-bool BufferCache<P>::SynchronizeBufferNoModified(Buffer& buffer, VAddr cpu_addr, u32 size) {
-    boost::container::small_vector<BufferCopy, 4> copies;
-    u64 total_size_bytes = 0;
-    u64 largest_copy = 0;
-    IntervalSet found_sets{};
-    auto make_copies = [&] {
-        for (auto& interval : found_sets) {
-            const std::size_t sub_size = interval.upper() - interval.lower();
-            const VAddr cpu_addr_ = interval.lower();
-            copies.push_back(BufferCopy{
-                .src_offset = total_size_bytes,
-                .dst_offset = cpu_addr_ - buffer.CpuAddr(),
-                .size = sub_size,
-            });
-            total_size_bytes += sub_size;
-            largest_copy = std::max<u64>(largest_copy, sub_size);
-        }
-        const std::span<BufferCopy> copies_span(copies.data(), copies.size());
-        UploadMemory(buffer, total_size_bytes, largest_copy, copies_span);
-    };
-    memory_tracker.ForEachUploadRange(cpu_addr, size, [&](u64 cpu_addr_out, u64 range_size) {
-        const VAddr base_adr = cpu_addr_out;
-        const VAddr end_adr = base_adr + range_size;
-        const IntervalType add_interval{base_adr, end_adr};
-        found_sets.add(add_interval);
-    });
-    if (found_sets.empty()) {
-        return true;
-    }
-    const IntervalType search_interval{cpu_addr, cpu_addr + size};
-    auto it = common_ranges.lower_bound(search_interval);
-    auto it_end = common_ranges.upper_bound(search_interval);
-    if (it == common_ranges.end()) {
-        make_copies();
-        return false;
-    }
-    while (it != it_end) {
-        found_sets.subtract(*it);
-        it++;
-    }
-    make_copies();
-    return false;
-}
-
-template <class P>
 void BufferCache<P>::UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
                                   std::span<BufferCopy> copies) {
     if constexpr (USE_MEMORY_MAPS_FOR_UPLOADS) {
@@ -1591,7 +1556,8 @@ void BufferCache<P>::MappedUploadMemory([[maybe_unused]] Buffer& buffer,
             // Apply the staging offset
             copy.src_offset += upload_staging.offset;
         }
-        runtime.CopyBuffer(buffer, upload_staging.buffer, copies);
+        const bool can_reorder = runtime.CanReorderUpload(buffer, copies);
+        runtime.CopyBuffer(buffer, upload_staging.buffer, copies, true, can_reorder);
     }
 }
 
@@ -1633,7 +1599,8 @@ void BufferCache<P>::InlineMemoryImplementation(VAddr dest_address, size_t copy_
         }};
         u8* const src_pointer = upload_staging.mapped_span.data();
         std::memcpy(src_pointer, inlined_buffer.data(), copy_size);
-        runtime.CopyBuffer(buffer, upload_staging.buffer, copies);
+        const bool can_reorder = runtime.CanReorderUpload(buffer, copies);
+        runtime.CopyBuffer(buffer, upload_staging.buffer, copies, true, can_reorder);
     } else {
         buffer.ImmediateUpload(buffer.Offset(dest_address), inlined_buffer.first(copy_size));
     }
@@ -1686,8 +1653,9 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si
         for (BufferCopy& copy : copies) {
             // Modify copies to have the staging offset in mind
             copy.dst_offset += download_staging.offset;
+            buffer.MarkUsage(copy.src_offset, copy.size);
         }
-        runtime.CopyBuffer(download_staging.buffer, buffer, copies_span);
+        runtime.CopyBuffer(download_staging.buffer, buffer, copies_span, true);
         runtime.Finish();
         for (const BufferCopy& copy : copies) {
             const VAddr copy_cpu_addr = buffer.CpuAddr() + copy.src_offset;
diff --git a/src/video_core/buffer_cache/buffer_cache_base.h b/src/video_core/buffer_cache/buffer_cache_base.h
index eed267361..d6d696d8c 100644
--- a/src/video_core/buffer_cache/buffer_cache_base.h
+++ b/src/video_core/buffer_cache/buffer_cache_base.h
@@ -529,10 +529,6 @@ private:
 
     bool SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size);
 
-    bool SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 size);
-
-    bool SynchronizeBufferNoModified(Buffer& buffer, VAddr cpu_addr, u32 size);
-
     void UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
                       std::span<BufferCopy> copies);
 
diff --git a/src/video_core/buffer_cache/usage_tracker.h b/src/video_core/buffer_cache/usage_tracker.h
new file mode 100644
index 000000000..ab05fe415
--- /dev/null
+++ b/src/video_core/buffer_cache/usage_tracker.h
@@ -0,0 +1,79 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include "common/alignment.h"
+#include "common/common_types.h"
+
+namespace VideoCommon {
+
+class UsageTracker {
+    static constexpr size_t BYTES_PER_BIT_SHIFT = 6;
+    static constexpr size_t PAGE_SHIFT = 6 + BYTES_PER_BIT_SHIFT;
+    static constexpr size_t PAGE_BYTES = 1 << PAGE_SHIFT;
+
+public:
+    explicit UsageTracker(size_t size) {
+        const size_t num_pages = (size >> PAGE_SHIFT) + 1;
+        pages.resize(num_pages, 0ULL);
+    }
+
+    void Reset() noexcept {
+        std::ranges::fill(pages, 0ULL);
+    }
+
+    void Track(u64 offset, u64 size) noexcept {
+        const size_t page = offset >> PAGE_SHIFT;
+        const size_t page_end = (offset + size) >> PAGE_SHIFT;
+        TrackPage(page, offset, size);
+        if (page == page_end) {
+            return;
+        }
+        for (size_t i = page + 1; i < page_end; i++) {
+            pages[i] = ~u64{0};
+        }
+        const size_t offset_end = offset + size;
+        const size_t offset_end_page_aligned = Common::AlignDown(offset_end, PAGE_BYTES);
+        TrackPage(page_end, offset_end_page_aligned, offset_end - offset_end_page_aligned);
+    }
+
+    [[nodiscard]] bool IsUsed(u64 offset, u64 size) const noexcept {
+        const size_t page = offset >> PAGE_SHIFT;
+        const size_t page_end = (offset + size) >> PAGE_SHIFT;
+        if (IsPageUsed(page, offset, size)) {
+            return true;
+        }
+        for (size_t i = page + 1; i < page_end; i++) {
+            if (pages[i] != 0) {
+                return true;
+            }
+        }
+        const size_t offset_end = offset + size;
+        const size_t offset_end_page_aligned = Common::AlignDown(offset_end, PAGE_BYTES);
+        return IsPageUsed(page_end, offset_end_page_aligned, offset_end - offset_end_page_aligned);
+    }
+
+private:
+    void TrackPage(u64 page, u64 offset, u64 size) noexcept {
+        const size_t offset_in_page = offset % PAGE_BYTES;
+        const size_t first_bit = offset_in_page >> BYTES_PER_BIT_SHIFT;
+        const size_t num_bits = std::min(size, PAGE_BYTES) >> BYTES_PER_BIT_SHIFT;
+        const size_t mask = ~u64{0} >> (64 - num_bits);
+        pages[page] |= (~u64{0} & mask) << first_bit;
+    }
+
+    bool IsPageUsed(u64 page, u64 offset, u64 size) const noexcept {
+        const size_t offset_in_page = offset % PAGE_BYTES;
+        const size_t first_bit = offset_in_page >> BYTES_PER_BIT_SHIFT;
+        const size_t num_bits = std::min(size, PAGE_BYTES) >> BYTES_PER_BIT_SHIFT;
+        const size_t mask = ~u64{0} >> (64 - num_bits);
+        const size_t mask2 = (~u64{0} & mask) << first_bit;
+        return (pages[page] & mask2) != 0;
+    }
+
+private:
+    std::vector<u64> pages;
+};
+
+} // namespace VideoCommon
diff --git a/src/video_core/engines/fermi_2d.cpp b/src/video_core/engines/fermi_2d.cpp
index 02e161270..91f10aec2 100644
--- a/src/video_core/engines/fermi_2d.cpp
+++ b/src/video_core/engines/fermi_2d.cpp
@@ -72,7 +72,7 @@ void Fermi2D::Blit() {
     UNIMPLEMENTED_IF_MSG(regs.clip_enable != 0, "Clipped blit enabled");
 
     const auto& args = regs.pixels_from_memory;
-    constexpr s64 null_derivate = 1ULL << 32;
+    constexpr s64 null_derivative = 1ULL << 32;
     Surface src = regs.src;
     const auto bytes_per_pixel = BytesPerBlock(PixelFormatFromRenderTargetFormat(src.format));
     const bool delegate_to_gpu = src.width > 512 && src.height > 512 && bytes_per_pixel <= 8 &&
@@ -89,7 +89,7 @@ void Fermi2D::Blit() {
         .operation = regs.operation,
         .filter = args.sample_mode.filter,
         .must_accelerate =
-            args.du_dx != null_derivate || args.dv_dy != null_derivate || delegate_to_gpu,
+            args.du_dx != null_derivative || args.dv_dy != null_derivative || delegate_to_gpu,
         .dst_x0 = args.dst_x0,
         .dst_y0 = args.dst_y0,
         .dst_x1 = args.dst_x0 + args.dst_width,
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 32d767d85..592c28ba3 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -268,7 +268,7 @@ size_t Maxwell3D::EstimateIndexBufferSize() {
                                                         std::numeric_limits<u32>::max()};
     const size_t byte_size = regs.index_buffer.FormatSizeInBytes();
     const size_t log2_byte_size = Common::Log2Ceil64(byte_size);
-    const size_t cap{GetMaxCurrentVertices() * 3 * byte_size};
+    const size_t cap{GetMaxCurrentVertices() * 4 * byte_size};
     const size_t lower_cap =
         std::min<size_t>(static_cast<size_t>(end_address - start_address), cap);
     return std::min<size_t>(
diff --git a/src/video_core/fence_manager.h b/src/video_core/fence_manager.h
index c0e6471fe..805a89900 100644
--- a/src/video_core/fence_manager.h
+++ b/src/video_core/fence_manager.h
@@ -86,10 +86,7 @@ public:
             uncommitted_operations.emplace_back(std::move(func));
         }
         pending_operations.emplace_back(std::move(uncommitted_operations));
-        {
-            std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
-            QueueFence(new_fence);
-        }
+        QueueFence(new_fence);
         if (!delay_fence) {
             func();
         }
diff --git a/src/video_core/host1x/codecs/codec.cpp b/src/video_core/host1x/codecs/codec.cpp
index dbcf508e5..1030db681 100644
--- a/src/video_core/host1x/codecs/codec.cpp
+++ b/src/video_core/host1x/codecs/codec.cpp
@@ -1,11 +1,7 @@
 // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
-#include <algorithm>
-#include <fstream>
-#include <vector>
 #include "common/assert.h"
-#include "common/scope_exit.h"
 #include "common/settings.h"
 #include "video_core/host1x/codecs/codec.h"
 #include "video_core/host1x/codecs/h264.h"
@@ -14,242 +10,17 @@
 #include "video_core/host1x/host1x.h"
 #include "video_core/memory_manager.h"
 
-extern "C" {
-#include <libavfilter/buffersink.h>
-#include <libavfilter/buffersrc.h>
-#include <libavutil/opt.h>
-#ifdef LIBVA_FOUND
-// for querying VAAPI driver information
-#include <libavutil/hwcontext_vaapi.h>
-#endif
-}
-
 namespace Tegra {
-namespace {
-constexpr AVPixelFormat PREFERRED_GPU_FMT = AV_PIX_FMT_NV12;
-constexpr AVPixelFormat PREFERRED_CPU_FMT = AV_PIX_FMT_YUV420P;
-constexpr std::array PREFERRED_GPU_DECODERS = {
-    AV_HWDEVICE_TYPE_CUDA,
-#ifdef _WIN32
-    AV_HWDEVICE_TYPE_D3D11VA,
-    AV_HWDEVICE_TYPE_DXVA2,
-#elif defined(__unix__)
-    AV_HWDEVICE_TYPE_VAAPI,
-    AV_HWDEVICE_TYPE_VDPAU,
-#endif
-    // last resort for Linux Flatpak (w/ NVIDIA)
-    AV_HWDEVICE_TYPE_VULKAN,
-};
-
-void AVPacketDeleter(AVPacket* ptr) {
-    av_packet_free(&ptr);
-}
-
-using AVPacketPtr = std::unique_ptr<AVPacket, decltype(&AVPacketDeleter)>;
-
-AVPixelFormat GetGpuFormat(AVCodecContext* av_codec_ctx, const AVPixelFormat* pix_fmts) {
-    for (const AVPixelFormat* p = pix_fmts; *p != AV_PIX_FMT_NONE; ++p) {
-        if (*p == av_codec_ctx->pix_fmt) {
-            return av_codec_ctx->pix_fmt;
-        }
-    }
-    LOG_INFO(Service_NVDRV, "Could not find compatible GPU AV format, falling back to CPU");
-    av_buffer_unref(&av_codec_ctx->hw_device_ctx);
-    av_codec_ctx->pix_fmt = PREFERRED_CPU_FMT;
-    return PREFERRED_CPU_FMT;
-}
-
-// List all the currently available hwcontext in ffmpeg
-std::vector<AVHWDeviceType> ListSupportedContexts() {
-    std::vector<AVHWDeviceType> contexts{};
-    AVHWDeviceType current_device_type = AV_HWDEVICE_TYPE_NONE;
-    do {
-        current_device_type = av_hwdevice_iterate_types(current_device_type);
-        contexts.push_back(current_device_type);
-    } while (current_device_type != AV_HWDEVICE_TYPE_NONE);
-    return contexts;
-}
-
-} // namespace
-
-void AVFrameDeleter(AVFrame* ptr) {
-    av_frame_free(&ptr);
-}
 
 Codec::Codec(Host1x::Host1x& host1x_, const Host1x::NvdecCommon::NvdecRegisters& regs)
     : host1x(host1x_), state{regs}, h264_decoder(std::make_unique<Decoder::H264>(host1x)),
       vp8_decoder(std::make_unique<Decoder::VP8>(host1x)),
       vp9_decoder(std::make_unique<Decoder::VP9>(host1x)) {}
 
-Codec::~Codec() {
-    if (!initialized) {
-        return;
-    }
-    // Free libav memory
-    avcodec_free_context(&av_codec_ctx);
-    av_buffer_unref(&av_gpu_decoder);
-
-    if (filters_initialized) {
-        avfilter_graph_free(&av_filter_graph);
-    }
-}
-
-bool Codec::CreateGpuAvDevice() {
-    static constexpr auto HW_CONFIG_METHOD = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX;
-    static const auto supported_contexts = ListSupportedContexts();
-    for (const auto& type : PREFERRED_GPU_DECODERS) {
-        if (std::none_of(supported_contexts.begin(), supported_contexts.end(),
-                         [&type](const auto& context) { return context == type; })) {
-            LOG_DEBUG(Service_NVDRV, "{} explicitly unsupported", av_hwdevice_get_type_name(type));
-            continue;
-        }
-        // Avoid memory leak from not cleaning up after av_hwdevice_ctx_create
-        av_buffer_unref(&av_gpu_decoder);
-        const int hwdevice_res = av_hwdevice_ctx_create(&av_gpu_decoder, type, nullptr, nullptr, 0);
-        if (hwdevice_res < 0) {
-            LOG_DEBUG(Service_NVDRV, "{} av_hwdevice_ctx_create failed {}",
-                      av_hwdevice_get_type_name(type), hwdevice_res);
-            continue;
-        }
-#ifdef LIBVA_FOUND
-        if (type == AV_HWDEVICE_TYPE_VAAPI) {
-            // we need to determine if this is an impersonated VAAPI driver
-            AVHWDeviceContext* hwctx =
-                static_cast<AVHWDeviceContext*>(static_cast<void*>(av_gpu_decoder->data));
-            AVVAAPIDeviceContext* vactx = static_cast<AVVAAPIDeviceContext*>(hwctx->hwctx);
-            const char* vendor_name = vaQueryVendorString(vactx->display);
-            if (strstr(vendor_name, "VDPAU backend")) {
-                // VDPAU impersonated VAAPI impl's are super buggy, we need to skip them
-                LOG_DEBUG(Service_NVDRV, "Skipping vdapu impersonated VAAPI driver");
-                continue;
-            } else {
-                // according to some user testing, certain vaapi driver (Intel?) could be buggy
-                // so let's log the driver name which may help the developers/supporters
-                LOG_DEBUG(Service_NVDRV, "Using VAAPI driver: {}", vendor_name);
-            }
-        }
-#endif
-        for (int i = 0;; i++) {
-            const AVCodecHWConfig* config = avcodec_get_hw_config(av_codec, i);
-            if (!config) {
-                LOG_DEBUG(Service_NVDRV, "{} decoder does not support device type {}.",
-                          av_codec->name, av_hwdevice_get_type_name(type));
-                break;
-            }
-            if ((config->methods & HW_CONFIG_METHOD) != 0 && config->device_type == type) {
-                LOG_INFO(Service_NVDRV, "Using {} GPU decoder", av_hwdevice_get_type_name(type));
-                av_codec_ctx->pix_fmt = config->pix_fmt;
-                return true;
-            }
-        }
-    }
-    return false;
-}
-
-void Codec::InitializeAvCodecContext() {
-    av_codec_ctx = avcodec_alloc_context3(av_codec);
-    av_opt_set(av_codec_ctx->priv_data, "tune", "zerolatency", 0);
-    av_codec_ctx->thread_count = 0;
-    av_codec_ctx->thread_type &= ~FF_THREAD_FRAME;
-}
-
-void Codec::InitializeGpuDecoder() {
-    if (!CreateGpuAvDevice()) {
-        av_buffer_unref(&av_gpu_decoder);
-        return;
-    }
-    auto* hw_device_ctx = av_buffer_ref(av_gpu_decoder);
-    ASSERT_MSG(hw_device_ctx, "av_buffer_ref failed");
-    av_codec_ctx->hw_device_ctx = hw_device_ctx;
-    av_codec_ctx->get_format = GetGpuFormat;
-}
-
-void Codec::InitializeAvFilters(AVFrame* frame) {
-    const AVFilter* buffer_src = avfilter_get_by_name("buffer");
-    const AVFilter* buffer_sink = avfilter_get_by_name("buffersink");
-    AVFilterInOut* inputs = avfilter_inout_alloc();
-    AVFilterInOut* outputs = avfilter_inout_alloc();
-    SCOPE_EXIT({
-        avfilter_inout_free(&inputs);
-        avfilter_inout_free(&outputs);
-    });
-
-    // Don't know how to get the accurate time_base but it doesn't matter for yadif filter
-    // so just use 1/1 to make buffer filter happy
-    std::string args = fmt::format("video_size={}x{}:pix_fmt={}:time_base=1/1", frame->width,
-                                   frame->height, frame->format);
-
-    av_filter_graph = avfilter_graph_alloc();
-    int ret = avfilter_graph_create_filter(&av_filter_src_ctx, buffer_src, "in", args.c_str(),
-                                           nullptr, av_filter_graph);
-    if (ret < 0) {
-        LOG_ERROR(Service_NVDRV, "avfilter_graph_create_filter source error: {}", ret);
-        return;
-    }
-
-    ret = avfilter_graph_create_filter(&av_filter_sink_ctx, buffer_sink, "out", nullptr, nullptr,
-                                       av_filter_graph);
-    if (ret < 0) {
-        LOG_ERROR(Service_NVDRV, "avfilter_graph_create_filter sink error: {}", ret);
-        return;
-    }
-
-    inputs->name = av_strdup("out");
-    inputs->filter_ctx = av_filter_sink_ctx;
-    inputs->pad_idx = 0;
-    inputs->next = nullptr;
-
-    outputs->name = av_strdup("in");
-    outputs->filter_ctx = av_filter_src_ctx;
-    outputs->pad_idx = 0;
-    outputs->next = nullptr;
-
-    const char* description = "yadif=1:-1:0";
-    ret = avfilter_graph_parse_ptr(av_filter_graph, description, &inputs, &outputs, nullptr);
-    if (ret < 0) {
-        LOG_ERROR(Service_NVDRV, "avfilter_graph_parse_ptr error: {}", ret);
-        return;
-    }
-
-    ret = avfilter_graph_config(av_filter_graph, nullptr);
-    if (ret < 0) {
-        LOG_ERROR(Service_NVDRV, "avfilter_graph_config error: {}", ret);
-        return;
-    }
-
-    filters_initialized = true;
-}
+Codec::~Codec() = default;
 
 void Codec::Initialize() {
-    const AVCodecID codec = [&] {
-        switch (current_codec) {
-        case Host1x::NvdecCommon::VideoCodec::H264:
-            return AV_CODEC_ID_H264;
-        case Host1x::NvdecCommon::VideoCodec::VP8:
-            return AV_CODEC_ID_VP8;
-        case Host1x::NvdecCommon::VideoCodec::VP9:
-            return AV_CODEC_ID_VP9;
-        default:
-            UNIMPLEMENTED_MSG("Unknown codec {}", current_codec);
-            return AV_CODEC_ID_NONE;
-        }
-    }();
-    av_codec = avcodec_find_decoder(codec);
-
-    InitializeAvCodecContext();
-    if (Settings::values.nvdec_emulation.GetValue() == Settings::NvdecEmulation::Gpu) {
-        InitializeGpuDecoder();
-    }
-    if (const int res = avcodec_open2(av_codec_ctx, av_codec, nullptr); res < 0) {
-        LOG_ERROR(Service_NVDRV, "avcodec_open2() Failed with result {}", res);
-        avcodec_free_context(&av_codec_ctx);
-        av_buffer_unref(&av_gpu_decoder);
-        return;
-    }
-    if (!av_codec_ctx->hw_device_ctx) {
-        LOG_INFO(Service_NVDRV, "Using FFmpeg software decoding");
-    }
-    initialized = true;
+    initialized = decode_api.Initialize(current_codec);
 }
 
 void Codec::SetTargetCodec(Host1x::NvdecCommon::VideoCodec codec) {
@@ -264,14 +35,18 @@ void Codec::Decode() {
     if (is_first_frame) {
         Initialize();
     }
+
     if (!initialized) {
         return;
     }
+
+    // Assemble bitstream.
     bool vp9_hidden_frame = false;
-    const auto& frame_data = [&]() {
+    size_t configuration_size = 0;
+    const auto packet_data = [&]() {
         switch (current_codec) {
         case Tegra::Host1x::NvdecCommon::VideoCodec::H264:
-            return h264_decoder->ComposeFrame(state, is_first_frame);
+            return h264_decoder->ComposeFrame(state, &configuration_size, is_first_frame);
         case Tegra::Host1x::NvdecCommon::VideoCodec::VP8:
             return vp8_decoder->ComposeFrame(state);
         case Tegra::Host1x::NvdecCommon::VideoCodec::VP9:
@@ -283,89 +58,35 @@ void Codec::Decode() {
             return std::span<const u8>{};
         }
     }();
-    AVPacketPtr packet{av_packet_alloc(), AVPacketDeleter};
-    if (!packet) {
-        LOG_ERROR(Service_NVDRV, "av_packet_alloc failed");
-        return;
-    }
-    packet->data = const_cast<u8*>(frame_data.data());
-    packet->size = static_cast<s32>(frame_data.size());
-    if (const int res = avcodec_send_packet(av_codec_ctx, packet.get()); res != 0) {
-        LOG_DEBUG(Service_NVDRV, "avcodec_send_packet error {}", res);
+
+    // Send assembled bitstream to decoder.
+    if (!decode_api.SendPacket(packet_data, configuration_size)) {
         return;
     }
-    // Only receive/store visible frames
+
+    // Only receive/store visible frames.
     if (vp9_hidden_frame) {
         return;
     }
-    AVFramePtr initial_frame{av_frame_alloc(), AVFrameDeleter};
-    AVFramePtr final_frame{nullptr, AVFrameDeleter};
-    ASSERT_MSG(initial_frame, "av_frame_alloc initial_frame failed");
-    if (const int ret = avcodec_receive_frame(av_codec_ctx, initial_frame.get()); ret) {
-        LOG_DEBUG(Service_NVDRV, "avcodec_receive_frame error {}", ret);
-        return;
-    }
-    if (initial_frame->width == 0 || initial_frame->height == 0) {
-        LOG_WARNING(Service_NVDRV, "Zero width or height in frame");
-        return;
-    }
-    bool is_interlaced = initial_frame->interlaced_frame != 0;
-    if (av_codec_ctx->hw_device_ctx) {
-        final_frame = AVFramePtr{av_frame_alloc(), AVFrameDeleter};
-        ASSERT_MSG(final_frame, "av_frame_alloc final_frame failed");
-        // Can't use AV_PIX_FMT_YUV420P and share code with software decoding in vic.cpp
-        // because Intel drivers crash unless using AV_PIX_FMT_NV12
-        final_frame->format = PREFERRED_GPU_FMT;
-        const int ret = av_hwframe_transfer_data(final_frame.get(), initial_frame.get(), 0);
-        ASSERT_MSG(!ret, "av_hwframe_transfer_data error {}", ret);
-    } else {
-        final_frame = std::move(initial_frame);
-    }
-    if (final_frame->format != PREFERRED_CPU_FMT && final_frame->format != PREFERRED_GPU_FMT) {
-        UNIMPLEMENTED_MSG("Unexpected video format: {}", final_frame->format);
-        return;
-    }
-    if (!is_interlaced) {
-        av_frames.push(std::move(final_frame));
-    } else {
-        if (!filters_initialized) {
-            InitializeAvFilters(final_frame.get());
-        }
-        if (const int ret = av_buffersrc_add_frame_flags(av_filter_src_ctx, final_frame.get(),
-                                                         AV_BUFFERSRC_FLAG_KEEP_REF);
-            ret) {
-            LOG_DEBUG(Service_NVDRV, "av_buffersrc_add_frame_flags error {}", ret);
-            return;
-        }
-        while (true) {
-            auto filter_frame = AVFramePtr{av_frame_alloc(), AVFrameDeleter};
 
-            int ret = av_buffersink_get_frame(av_filter_sink_ctx, filter_frame.get());
+    // Receive output frames from decoder.
+    decode_api.ReceiveFrames(frames);
 
-            if (ret == AVERROR(EAGAIN) || ret == AVERROR(AVERROR_EOF))
-                break;
-            if (ret < 0) {
-                LOG_DEBUG(Service_NVDRV, "av_buffersink_get_frame error {}", ret);
-                return;
-            }
-
-            av_frames.push(std::move(filter_frame));
-        }
-    }
-    while (av_frames.size() > 10) {
-        LOG_TRACE(Service_NVDRV, "av_frames.push overflow dropped frame");
-        av_frames.pop();
+    while (frames.size() > 10) {
+        LOG_DEBUG(HW_GPU, "ReceiveFrames overflow, dropped frame");
+        frames.pop();
     }
 }
 
-AVFramePtr Codec::GetCurrentFrame() {
+std::unique_ptr<FFmpeg::Frame> Codec::GetCurrentFrame() {
     // Sometimes VIC will request more frames than have been decoded.
-    // in this case, return a nullptr and don't overwrite previous frame data
-    if (av_frames.empty()) {
-        return AVFramePtr{nullptr, AVFrameDeleter};
+    // in this case, return a blank frame and don't overwrite previous data.
+    if (frames.empty()) {
+        return {};
     }
-    AVFramePtr frame = std::move(av_frames.front());
-    av_frames.pop();
+
+    auto frame = std::move(frames.front());
+    frames.pop();
     return frame;
 }
 
diff --git a/src/video_core/host1x/codecs/codec.h b/src/video_core/host1x/codecs/codec.h
index 06fe00a4b..f700ae129 100644
--- a/src/video_core/host1x/codecs/codec.h
+++ b/src/video_core/host1x/codecs/codec.h
@@ -4,28 +4,15 @@
 #pragma once
 
 #include <memory>
+#include <optional>
 #include <string_view>
 #include <queue>
 #include "common/common_types.h"
+#include "video_core/host1x/ffmpeg/ffmpeg.h"
 #include "video_core/host1x/nvdec_common.h"
 
-extern "C" {
-#if defined(__GNUC__) || defined(__clang__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wconversion"
-#endif
-#include <libavcodec/avcodec.h>
-#include <libavfilter/avfilter.h>
-#if defined(__GNUC__) || defined(__clang__)
-#pragma GCC diagnostic pop
-#endif
-}
-
 namespace Tegra {
 
-void AVFrameDeleter(AVFrame* ptr);
-using AVFramePtr = std::unique_ptr<AVFrame, decltype(&AVFrameDeleter)>;
-
 namespace Decoder {
 class H264;
 class VP8;
@@ -51,7 +38,7 @@ public:
     void Decode();
 
     /// Returns next decoded frame
-    [[nodiscard]] AVFramePtr GetCurrentFrame();
+    [[nodiscard]] std::unique_ptr<FFmpeg::Frame> GetCurrentFrame();
 
     /// Returns the value of current_codec
     [[nodiscard]] Host1x::NvdecCommon::VideoCodec GetCurrentCodec() const;
@@ -60,25 +47,9 @@ public:
     [[nodiscard]] std::string_view GetCurrentCodecName() const;
 
 private:
-    void InitializeAvCodecContext();
-
-    void InitializeAvFilters(AVFrame* frame);
-
-    void InitializeGpuDecoder();
-
-    bool CreateGpuAvDevice();
-
     bool initialized{};
-    bool filters_initialized{};
     Host1x::NvdecCommon::VideoCodec current_codec{Host1x::NvdecCommon::VideoCodec::None};
-
-    const AVCodec* av_codec{nullptr};
-    AVCodecContext* av_codec_ctx{nullptr};
-    AVBufferRef* av_gpu_decoder{nullptr};
-
-    AVFilterContext* av_filter_src_ctx{nullptr};
-    AVFilterContext* av_filter_sink_ctx{nullptr};
-    AVFilterGraph* av_filter_graph{nullptr};
+    FFmpeg::DecodeApi decode_api;
 
     Host1x::Host1x& host1x;
     const Host1x::NvdecCommon::NvdecRegisters& state;
@@ -86,7 +57,7 @@ private:
     std::unique_ptr<Decoder::VP8> vp8_decoder;
     std::unique_ptr<Decoder::VP9> vp9_decoder;
 
-    std::queue<AVFramePtr> av_frames{};
+    std::queue<std::unique_ptr<FFmpeg::Frame>> frames{};
 };
 
 } // namespace Tegra
diff --git a/src/video_core/host1x/codecs/h264.cpp b/src/video_core/host1x/codecs/h264.cpp
index ece79b1e2..309a7f1d5 100644
--- a/src/video_core/host1x/codecs/h264.cpp
+++ b/src/video_core/host1x/codecs/h264.cpp
@@ -30,7 +30,7 @@ H264::H264(Host1x::Host1x& host1x_) : host1x{host1x_} {}
 H264::~H264() = default;
 
 std::span<const u8> H264::ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters& state,
-                                       bool is_first_frame) {
+                                       size_t* out_configuration_size, bool is_first_frame) {
     H264DecoderContext context;
     host1x.MemoryManager().ReadBlock(state.picture_info_offset, &context,
                                      sizeof(H264DecoderContext));
@@ -39,6 +39,7 @@ std::span<const u8> H264::ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters
     if (!is_first_frame && frame_number != 0) {
         frame.resize_destructive(context.stream_len);
         host1x.MemoryManager().ReadBlock(state.frame_bitstream_offset, frame.data(), frame.size());
+        *out_configuration_size = 0;
         return frame;
     }
 
@@ -157,6 +158,7 @@ std::span<const u8> H264::ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters
     frame.resize(encoded_header.size() + context.stream_len);
     std::memcpy(frame.data(), encoded_header.data(), encoded_header.size());
 
+    *out_configuration_size = encoded_header.size();
     host1x.MemoryManager().ReadBlock(state.frame_bitstream_offset,
                                      frame.data() + encoded_header.size(), context.stream_len);
 
diff --git a/src/video_core/host1x/codecs/h264.h b/src/video_core/host1x/codecs/h264.h
index d6b556322..1deaf4632 100644
--- a/src/video_core/host1x/codecs/h264.h
+++ b/src/video_core/host1x/codecs/h264.h
@@ -67,6 +67,7 @@ public:
 
     /// Compose the H264 frame for FFmpeg decoding
     [[nodiscard]] std::span<const u8> ComposeFrame(const Host1x::NvdecCommon::NvdecRegisters& state,
+                                                   size_t* out_configuration_size,
                                                    bool is_first_frame = false);
 
 private:
diff --git a/src/video_core/host1x/ffmpeg/ffmpeg.cpp b/src/video_core/host1x/ffmpeg/ffmpeg.cpp
new file mode 100644
index 000000000..dcd07e6d2
--- /dev/null
+++ b/src/video_core/host1x/ffmpeg/ffmpeg.cpp
@@ -0,0 +1,419 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "common/assert.h"
+#include "common/logging/log.h"
+#include "common/scope_exit.h"
+#include "common/settings.h"
+#include "video_core/host1x/ffmpeg/ffmpeg.h"
+
+extern "C" {
+#ifdef LIBVA_FOUND
+// for querying VAAPI driver information
+#include <libavutil/hwcontext_vaapi.h>
+#endif
+}
+
+namespace FFmpeg {
+
+namespace {
+
+constexpr AVPixelFormat PreferredGpuFormat = AV_PIX_FMT_NV12;
+constexpr AVPixelFormat PreferredCpuFormat = AV_PIX_FMT_YUV420P;
+constexpr std::array PreferredGpuDecoders = {
+    AV_HWDEVICE_TYPE_CUDA,
+#ifdef _WIN32
+    AV_HWDEVICE_TYPE_D3D11VA,
+    AV_HWDEVICE_TYPE_DXVA2,
+#elif defined(__unix__)
+    AV_HWDEVICE_TYPE_VAAPI,
+    AV_HWDEVICE_TYPE_VDPAU,
+#endif
+    // last resort for Linux Flatpak (w/ NVIDIA)
+    AV_HWDEVICE_TYPE_VULKAN,
+};
+
+AVPixelFormat GetGpuFormat(AVCodecContext* codec_context, const AVPixelFormat* pix_fmts) {
+    for (const AVPixelFormat* p = pix_fmts; *p != AV_PIX_FMT_NONE; ++p) {
+        if (*p == codec_context->pix_fmt) {
+            return codec_context->pix_fmt;
+        }
+    }
+
+    LOG_INFO(HW_GPU, "Could not find compatible GPU AV format, falling back to CPU");
+    av_buffer_unref(&codec_context->hw_device_ctx);
+
+    codec_context->pix_fmt = PreferredCpuFormat;
+    return codec_context->pix_fmt;
+}
+
+std::string AVError(int errnum) {
+    char errbuf[AV_ERROR_MAX_STRING_SIZE] = {};
+    av_make_error_string(errbuf, sizeof(errbuf) - 1, errnum);
+    return errbuf;
+}
+
+} // namespace
+
+Packet::Packet(std::span<const u8> data) {
+    m_packet = av_packet_alloc();
+    m_packet->data = const_cast<u8*>(data.data());
+    m_packet->size = static_cast<s32>(data.size());
+}
+
+Packet::~Packet() {
+    av_packet_free(&m_packet);
+}
+
+Frame::Frame() {
+    m_frame = av_frame_alloc();
+}
+
+Frame::~Frame() {
+    av_frame_free(&m_frame);
+}
+
+Decoder::Decoder(Tegra::Host1x::NvdecCommon::VideoCodec codec) {
+    const AVCodecID av_codec = [&] {
+        switch (codec) {
+        case Tegra::Host1x::NvdecCommon::VideoCodec::H264:
+            return AV_CODEC_ID_H264;
+        case Tegra::Host1x::NvdecCommon::VideoCodec::VP8:
+            return AV_CODEC_ID_VP8;
+        case Tegra::Host1x::NvdecCommon::VideoCodec::VP9:
+            return AV_CODEC_ID_VP9;
+        default:
+            UNIMPLEMENTED_MSG("Unknown codec {}", codec);
+            return AV_CODEC_ID_NONE;
+        }
+    }();
+
+    m_codec = avcodec_find_decoder(av_codec);
+}
+
+bool Decoder::SupportsDecodingOnDevice(AVPixelFormat* out_pix_fmt, AVHWDeviceType type) const {
+    for (int i = 0;; i++) {
+        const AVCodecHWConfig* config = avcodec_get_hw_config(m_codec, i);
+        if (!config) {
+            LOG_DEBUG(HW_GPU, "{} decoder does not support device type {}", m_codec->name,
+                      av_hwdevice_get_type_name(type));
+            break;
+        }
+        if ((config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) != 0 &&
+            config->device_type == type) {
+            LOG_INFO(HW_GPU, "Using {} GPU decoder", av_hwdevice_get_type_name(type));
+            *out_pix_fmt = config->pix_fmt;
+            return true;
+        }
+    }
+
+    return false;
+}
+
+std::vector<AVHWDeviceType> HardwareContext::GetSupportedDeviceTypes() {
+    std::vector<AVHWDeviceType> types;
+    AVHWDeviceType current_device_type = AV_HWDEVICE_TYPE_NONE;
+
+    while (true) {
+        current_device_type = av_hwdevice_iterate_types(current_device_type);
+        if (current_device_type == AV_HWDEVICE_TYPE_NONE) {
+            return types;
+        }
+
+        types.push_back(current_device_type);
+    }
+}
+
+HardwareContext::~HardwareContext() {
+    av_buffer_unref(&m_gpu_decoder);
+}
+
+bool HardwareContext::InitializeForDecoder(DecoderContext& decoder_context,
+                                           const Decoder& decoder) {
+    const auto supported_types = GetSupportedDeviceTypes();
+    for (const auto type : PreferredGpuDecoders) {
+        AVPixelFormat hw_pix_fmt;
+
+        if (std::ranges::find(supported_types, type) == supported_types.end()) {
+            LOG_DEBUG(HW_GPU, "{} explicitly unsupported", av_hwdevice_get_type_name(type));
+            continue;
+        }
+
+        if (!this->InitializeWithType(type)) {
+            continue;
+        }
+
+        if (decoder.SupportsDecodingOnDevice(&hw_pix_fmt, type)) {
+            decoder_context.InitializeHardwareDecoder(*this, hw_pix_fmt);
+            return true;
+        }
+    }
+
+    return false;
+}
+
+bool HardwareContext::InitializeWithType(AVHWDeviceType type) {
+    av_buffer_unref(&m_gpu_decoder);
+
+    if (const int ret = av_hwdevice_ctx_create(&m_gpu_decoder, type, nullptr, nullptr, 0);
+        ret < 0) {
+        LOG_DEBUG(HW_GPU, "av_hwdevice_ctx_create({}) failed: {}", av_hwdevice_get_type_name(type),
+                  AVError(ret));
+        return false;
+    }
+
+#ifdef LIBVA_FOUND
+    if (type == AV_HWDEVICE_TYPE_VAAPI) {
+        // We need to determine if this is an impersonated VAAPI driver.
+        auto* hwctx = reinterpret_cast<AVHWDeviceContext*>(m_gpu_decoder->data);
+        auto* vactx = static_cast<AVVAAPIDeviceContext*>(hwctx->hwctx);
+        const char* vendor_name = vaQueryVendorString(vactx->display);
+        if (strstr(vendor_name, "VDPAU backend")) {
+            // VDPAU impersonated VAAPI impls are super buggy, we need to skip them.
+            LOG_DEBUG(HW_GPU, "Skipping VDPAU impersonated VAAPI driver");
+            return false;
+        } else {
+            // According to some user testing, certain VAAPI drivers (Intel?) could be buggy.
+            // Log the driver name just in case.
+            LOG_DEBUG(HW_GPU, "Using VAAPI driver: {}", vendor_name);
+        }
+    }
+#endif
+
+    return true;
+}
+
+DecoderContext::DecoderContext(const Decoder& decoder) {
+    m_codec_context = avcodec_alloc_context3(decoder.GetCodec());
+    av_opt_set(m_codec_context->priv_data, "tune", "zerolatency", 0);
+    m_codec_context->thread_count = 0;
+    m_codec_context->thread_type &= ~FF_THREAD_FRAME;
+}
+
+DecoderContext::~DecoderContext() {
+    av_buffer_unref(&m_codec_context->hw_device_ctx);
+    avcodec_free_context(&m_codec_context);
+}
+
+void DecoderContext::InitializeHardwareDecoder(const HardwareContext& context,
+                                               AVPixelFormat hw_pix_fmt) {
+    m_codec_context->hw_device_ctx = av_buffer_ref(context.GetBufferRef());
+    m_codec_context->get_format = GetGpuFormat;
+    m_codec_context->pix_fmt = hw_pix_fmt;
+}
+
+bool DecoderContext::OpenContext(const Decoder& decoder) {
+    if (const int ret = avcodec_open2(m_codec_context, decoder.GetCodec(), nullptr); ret < 0) {
+        LOG_ERROR(HW_GPU, "avcodec_open2 error: {}", AVError(ret));
+        return false;
+    }
+
+    if (!m_codec_context->hw_device_ctx) {
+        LOG_INFO(HW_GPU, "Using FFmpeg software decoding");
+    }
+
+    return true;
+}
+
+bool DecoderContext::SendPacket(const Packet& packet) {
+    if (const int ret = avcodec_send_packet(m_codec_context, packet.GetPacket()); ret < 0) {
+        LOG_ERROR(HW_GPU, "avcodec_send_packet error: {}", AVError(ret));
+        return false;
+    }
+
+    return true;
+}
+
+std::unique_ptr<Frame> DecoderContext::ReceiveFrame(bool* out_is_interlaced) {
+    auto dst_frame = std::make_unique<Frame>();
+
+    const auto ReceiveImpl = [&](AVFrame* frame) {
+        if (const int ret = avcodec_receive_frame(m_codec_context, frame); ret < 0) {
+            LOG_ERROR(HW_GPU, "avcodec_receive_frame error: {}", AVError(ret));
+            return false;
+        }
+
+        *out_is_interlaced = frame->interlaced_frame != 0;
+        return true;
+    };
+
+    if (m_codec_context->hw_device_ctx) {
+        // If we have a hardware context, make a separate frame here to receive the
+        // hardware result before sending it to the output.
+        Frame intermediate_frame;
+
+        if (!ReceiveImpl(intermediate_frame.GetFrame())) {
+            return {};
+        }
+
+        dst_frame->SetFormat(PreferredGpuFormat);
+        if (const int ret =
+                av_hwframe_transfer_data(dst_frame->GetFrame(), intermediate_frame.GetFrame(), 0);
+            ret < 0) {
+            LOG_ERROR(HW_GPU, "av_hwframe_transfer_data error: {}", AVError(ret));
+            return {};
+        }
+    } else {
+        // Otherwise, decode the frame as normal.
+        if (!ReceiveImpl(dst_frame->GetFrame())) {
+            return {};
+        }
+    }
+
+    return dst_frame;
+}
+
+DeinterlaceFilter::DeinterlaceFilter(const Frame& frame) {
+    const AVFilter* buffer_src = avfilter_get_by_name("buffer");
+    const AVFilter* buffer_sink = avfilter_get_by_name("buffersink");
+    AVFilterInOut* inputs = avfilter_inout_alloc();
+    AVFilterInOut* outputs = avfilter_inout_alloc();
+    SCOPE_EXIT({
+        avfilter_inout_free(&inputs);
+        avfilter_inout_free(&outputs);
+    });
+
+    // Don't know how to get the accurate time_base but it doesn't matter for yadif filter
+    // so just use 1/1 to make buffer filter happy
+    std::string args = fmt::format("video_size={}x{}:pix_fmt={}:time_base=1/1", frame.GetWidth(),
+                                   frame.GetHeight(), static_cast<int>(frame.GetPixelFormat()));
+
+    m_filter_graph = avfilter_graph_alloc();
+    int ret = avfilter_graph_create_filter(&m_source_context, buffer_src, "in", args.c_str(),
+                                           nullptr, m_filter_graph);
+    if (ret < 0) {
+        LOG_ERROR(HW_GPU, "avfilter_graph_create_filter source error: {}", AVError(ret));
+        return;
+    }
+
+    ret = avfilter_graph_create_filter(&m_sink_context, buffer_sink, "out", nullptr, nullptr,
+                                       m_filter_graph);
+    if (ret < 0) {
+        LOG_ERROR(HW_GPU, "avfilter_graph_create_filter sink error: {}", AVError(ret));
+        return;
+    }
+
+    inputs->name = av_strdup("out");
+    inputs->filter_ctx = m_sink_context;
+    inputs->pad_idx = 0;
+    inputs->next = nullptr;
+
+    outputs->name = av_strdup("in");
+    outputs->filter_ctx = m_source_context;
+    outputs->pad_idx = 0;
+    outputs->next = nullptr;
+
+    const char* description = "yadif=1:-1:0";
+    ret = avfilter_graph_parse_ptr(m_filter_graph, description, &inputs, &outputs, nullptr);
+    if (ret < 0) {
+        LOG_ERROR(HW_GPU, "avfilter_graph_parse_ptr error: {}", AVError(ret));
+        return;
+    }
+
+    ret = avfilter_graph_config(m_filter_graph, nullptr);
+    if (ret < 0) {
+        LOG_ERROR(HW_GPU, "avfilter_graph_config error: {}", AVError(ret));
+        return;
+    }
+
+    m_initialized = true;
+}
+
+bool DeinterlaceFilter::AddSourceFrame(const Frame& frame) {
+    if (const int ret = av_buffersrc_add_frame_flags(m_source_context, frame.GetFrame(),
+                                                     AV_BUFFERSRC_FLAG_KEEP_REF);
+        ret < 0) {
+        LOG_ERROR(HW_GPU, "av_buffersrc_add_frame_flags error: {}", AVError(ret));
+        return false;
+    }
+
+    return true;
+}
+
+std::unique_ptr<Frame> DeinterlaceFilter::DrainSinkFrame() {
+    auto dst_frame = std::make_unique<Frame>();
+    const int ret = av_buffersink_get_frame(m_sink_context, dst_frame->GetFrame());
+
+    if (ret == AVERROR(EAGAIN) || ret == AVERROR(AVERROR_EOF)) {
+        return {};
+    }
+
+    if (ret < 0) {
+        LOG_ERROR(HW_GPU, "av_buffersink_get_frame error: {}", AVError(ret));
+        return {};
+    }
+
+    return dst_frame;
+}
+
+DeinterlaceFilter::~DeinterlaceFilter() {
+    avfilter_graph_free(&m_filter_graph);
+}
+
+void DecodeApi::Reset() {
+    m_deinterlace_filter.reset();
+    m_hardware_context.reset();
+    m_decoder_context.reset();
+    m_decoder.reset();
+}
+
+bool DecodeApi::Initialize(Tegra::Host1x::NvdecCommon::VideoCodec codec) {
+    this->Reset();
+    m_decoder.emplace(codec);
+    m_decoder_context.emplace(*m_decoder);
+
+    // Enable GPU decoding if requested.
+    if (Settings::values.nvdec_emulation.GetValue() == Settings::NvdecEmulation::Gpu) {
+        m_hardware_context.emplace();
+        m_hardware_context->InitializeForDecoder(*m_decoder_context, *m_decoder);
+    }
+
+    // Open the decoder context.
+    if (!m_decoder_context->OpenContext(*m_decoder)) {
+        this->Reset();
+        return false;
+    }
+
+    return true;
+}
+
+bool DecodeApi::SendPacket(std::span<const u8> packet_data, size_t configuration_size) {
+    FFmpeg::Packet packet(packet_data);
+    return m_decoder_context->SendPacket(packet);
+}
+
+void DecodeApi::ReceiveFrames(std::queue<std::unique_ptr<Frame>>& frame_queue) {
+    // Receive raw frame from decoder.
+    bool is_interlaced;
+    auto frame = m_decoder_context->ReceiveFrame(&is_interlaced);
+    if (!frame) {
+        return;
+    }
+
+    if (!is_interlaced) {
+        // If the frame is not interlaced, we can pend it now.
+        frame_queue.push(std::move(frame));
+    } else {
+        // Create the deinterlacer if needed.
+        if (!m_deinterlace_filter) {
+            m_deinterlace_filter.emplace(*frame);
+        }
+
+        // Add the frame we just received.
+        if (!m_deinterlace_filter->AddSourceFrame(*frame)) {
+            return;
+        }
+
+        // Pend output fields.
+        while (true) {
+            auto filter_frame = m_deinterlace_filter->DrainSinkFrame();
+            if (!filter_frame) {
+                break;
+            }
+
+            frame_queue.push(std::move(filter_frame));
+        }
+    }
+}
+
+} // namespace FFmpeg
diff --git a/src/video_core/host1x/ffmpeg/ffmpeg.h b/src/video_core/host1x/ffmpeg/ffmpeg.h
new file mode 100644
index 000000000..1de0bbd83
--- /dev/null
+++ b/src/video_core/host1x/ffmpeg/ffmpeg.h
@@ -0,0 +1,213 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <memory>
+#include <optional>
+#include <span>
+#include <vector>
+#include <queue>
+
+#include "common/common_funcs.h"
+#include "common/common_types.h"
+#include "video_core/host1x/nvdec_common.h"
+
+extern "C" {
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wconversion"
+#endif
+
+#include <libavcodec/avcodec.h>
+#include <libavfilter/avfilter.h>
+#include <libavfilter/buffersink.h>
+#include <libavfilter/buffersrc.h>
+#include <libavutil/avutil.h>
+#include <libavutil/opt.h>
+
+#if defined(__GNUC__) || defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+}
+
+namespace FFmpeg {
+
+class Packet;
+class Frame;
+class Decoder;
+class HardwareContext;
+class DecoderContext;
+class DeinterlaceFilter;
+
+// Wraps an AVPacket, a container for compressed bitstream data.
+class Packet {
+public:
+    YUZU_NON_COPYABLE(Packet);
+    YUZU_NON_MOVEABLE(Packet);
+
+    explicit Packet(std::span<const u8> data);
+    ~Packet();
+
+    AVPacket* GetPacket() const {
+        return m_packet;
+    }
+
+private:
+    AVPacket* m_packet{};
+};
+
+// Wraps an AVFrame, a container for audio and video stream data.
+class Frame {
+public:
+    YUZU_NON_COPYABLE(Frame);
+    YUZU_NON_MOVEABLE(Frame);
+
+    explicit Frame();
+    ~Frame();
+
+    int GetWidth() const {
+        return m_frame->width;
+    }
+
+    int GetHeight() const {
+        return m_frame->height;
+    }
+
+    AVPixelFormat GetPixelFormat() const {
+        return static_cast<AVPixelFormat>(m_frame->format);
+    }
+
+    int GetStride(int plane) const {
+        return m_frame->linesize[plane];
+    }
+
+    int* GetStrides() const {
+        return m_frame->linesize;
+    }
+
+    u8* GetData(int plane) const {
+        return m_frame->data[plane];
+    }
+
+    u8** GetPlanes() const {
+        return m_frame->data;
+    }
+
+    void SetFormat(int format) {
+        m_frame->format = format;
+    }
+
+    AVFrame* GetFrame() const {
+        return m_frame;
+    }
+
+private:
+    AVFrame* m_frame{};
+};
+
+// Wraps an AVCodec, a type containing information about a codec.
+class Decoder {
+public:
+    YUZU_NON_COPYABLE(Decoder);
+    YUZU_NON_MOVEABLE(Decoder);
+
+    explicit Decoder(Tegra::Host1x::NvdecCommon::VideoCodec codec);
+    ~Decoder() = default;
+
+    bool SupportsDecodingOnDevice(AVPixelFormat* out_pix_fmt, AVHWDeviceType type) const;
+
+    const AVCodec* GetCodec() const {
+        return m_codec;
+    }
+
+private:
+    const AVCodec* m_codec{};
+};
+
+// Wraps AVBufferRef for an accelerated decoder.
+class HardwareContext {
+public:
+    YUZU_NON_COPYABLE(HardwareContext);
+    YUZU_NON_MOVEABLE(HardwareContext);
+
+    static std::vector<AVHWDeviceType> GetSupportedDeviceTypes();
+
+    explicit HardwareContext() = default;
+    ~HardwareContext();
+
+    bool InitializeForDecoder(DecoderContext& decoder_context, const Decoder& decoder);
+
+    AVBufferRef* GetBufferRef() const {
+        return m_gpu_decoder;
+    }
+
+private:
+    bool InitializeWithType(AVHWDeviceType type);
+
+    AVBufferRef* m_gpu_decoder{};
+};
+
+// Wraps an AVCodecContext.
+class DecoderContext {
+public:
+    YUZU_NON_COPYABLE(DecoderContext);
+    YUZU_NON_MOVEABLE(DecoderContext);
+
+    explicit DecoderContext(const Decoder& decoder);
+    ~DecoderContext();
+
+    void InitializeHardwareDecoder(const HardwareContext& context, AVPixelFormat hw_pix_fmt);
+    bool OpenContext(const Decoder& decoder);
+    bool SendPacket(const Packet& packet);
+    std::unique_ptr<Frame> ReceiveFrame(bool* out_is_interlaced);
+
+    AVCodecContext* GetCodecContext() const {
+        return m_codec_context;
+    }
+
+private:
+    AVCodecContext* m_codec_context{};
+};
+
+// Wraps an AVFilterGraph.
+class DeinterlaceFilter {
+public:
+    YUZU_NON_COPYABLE(DeinterlaceFilter);
+    YUZU_NON_MOVEABLE(DeinterlaceFilter);
+
+    explicit DeinterlaceFilter(const Frame& frame);
+    ~DeinterlaceFilter();
+
+    bool AddSourceFrame(const Frame& frame);
+    std::unique_ptr<Frame> DrainSinkFrame();
+
+private:
+    AVFilterGraph* m_filter_graph{};
+    AVFilterContext* m_source_context{};
+    AVFilterContext* m_sink_context{};
+    bool m_initialized{};
+};
+
+class DecodeApi {
+public:
+    YUZU_NON_COPYABLE(DecodeApi);
+    YUZU_NON_MOVEABLE(DecodeApi);
+
+    DecodeApi() = default;
+    ~DecodeApi() = default;
+
+    bool Initialize(Tegra::Host1x::NvdecCommon::VideoCodec codec);
+    void Reset();
+
+    bool SendPacket(std::span<const u8> packet_data, size_t configuration_size);
+    void ReceiveFrames(std::queue<std::unique_ptr<Frame>>& frame_queue);
+
+private:
+    std::optional<FFmpeg::Decoder> m_decoder;
+    std::optional<FFmpeg::DecoderContext> m_decoder_context;
+    std::optional<FFmpeg::HardwareContext> m_hardware_context;
+    std::optional<FFmpeg::DeinterlaceFilter> m_deinterlace_filter;
+};
+
+} // namespace FFmpeg
diff --git a/src/video_core/host1x/nvdec.cpp b/src/video_core/host1x/nvdec.cpp
index a4bd5b79f..b8f5866d3 100644
--- a/src/video_core/host1x/nvdec.cpp
+++ b/src/video_core/host1x/nvdec.cpp
@@ -28,7 +28,7 @@ void Nvdec::ProcessMethod(u32 method, u32 argument) {
     }
 }
 
-AVFramePtr Nvdec::GetFrame() {
+std::unique_ptr<FFmpeg::Frame> Nvdec::GetFrame() {
     return codec->GetCurrentFrame();
 }
 
diff --git a/src/video_core/host1x/nvdec.h b/src/video_core/host1x/nvdec.h
index 3949d5181..ddddb8d28 100644
--- a/src/video_core/host1x/nvdec.h
+++ b/src/video_core/host1x/nvdec.h
@@ -23,7 +23,7 @@ public:
     void ProcessMethod(u32 method, u32 argument);
 
     /// Return most recently decoded frame
-    [[nodiscard]] AVFramePtr GetFrame();
+    [[nodiscard]] std::unique_ptr<FFmpeg::Frame> GetFrame();
 
 private:
     /// Invoke codec to decode a frame
diff --git a/src/video_core/host1x/vic.cpp b/src/video_core/host1x/vic.cpp
index 10d7ef884..2a5eba415 100644
--- a/src/video_core/host1x/vic.cpp
+++ b/src/video_core/host1x/vic.cpp
@@ -82,27 +82,26 @@ void Vic::Execute() {
         return;
     }
     const VicConfig config{host1x.MemoryManager().Read<u64>(config_struct_address + 0x20)};
-    const AVFramePtr frame_ptr = nvdec_processor->GetFrame();
-    const auto* frame = frame_ptr.get();
+    auto frame = nvdec_processor->GetFrame();
     if (!frame) {
         return;
     }
     const u64 surface_width = config.surface_width_minus1 + 1;
     const u64 surface_height = config.surface_height_minus1 + 1;
-    if (static_cast<u64>(frame->width) != surface_width ||
-        static_cast<u64>(frame->height) != surface_height) {
+    if (static_cast<u64>(frame->GetWidth()) != surface_width ||
+        static_cast<u64>(frame->GetHeight()) != surface_height) {
         // TODO: Properly support multiple video streams with differing frame dimensions
         LOG_WARNING(Service_NVDRV, "Frame dimensions {}x{} don't match surface dimensions {}x{}",
-                    frame->width, frame->height, surface_width, surface_height);
+                    frame->GetWidth(), frame->GetHeight(), surface_width, surface_height);
     }
     switch (config.pixel_format) {
     case VideoPixelFormat::RGBA8:
     case VideoPixelFormat::BGRA8:
     case VideoPixelFormat::RGBX8:
-        WriteRGBFrame(frame, config);
+        WriteRGBFrame(std::move(frame), config);
         break;
     case VideoPixelFormat::YUV420:
-        WriteYUVFrame(frame, config);
+        WriteYUVFrame(std::move(frame), config);
         break;
     default:
         UNIMPLEMENTED_MSG("Unknown video pixel format {:X}", config.pixel_format.Value());
@@ -110,10 +109,14 @@ void Vic::Execute() {
     }
 }
 
-void Vic::WriteRGBFrame(const AVFrame* frame, const VicConfig& config) {
+void Vic::WriteRGBFrame(std::unique_ptr<FFmpeg::Frame> frame, const VicConfig& config) {
     LOG_TRACE(Service_NVDRV, "Writing RGB Frame");
 
-    if (!scaler_ctx || frame->width != scaler_width || frame->height != scaler_height) {
+    const auto frame_width = frame->GetWidth();
+    const auto frame_height = frame->GetHeight();
+    const auto frame_format = frame->GetPixelFormat();
+
+    if (!scaler_ctx || frame_width != scaler_width || frame_height != scaler_height) {
         const AVPixelFormat target_format = [pixel_format = config.pixel_format]() {
             switch (pixel_format) {
             case VideoPixelFormat::RGBA8:
@@ -129,27 +132,26 @@ void Vic::WriteRGBFrame(const AVFrame* frame, const VicConfig& config) {
 
         sws_freeContext(scaler_ctx);
         // Frames are decoded into either YUV420 or NV12 formats. Convert to desired RGB format
-        scaler_ctx = sws_getContext(frame->width, frame->height,
-                                    static_cast<AVPixelFormat>(frame->format), frame->width,
-                                    frame->height, target_format, 0, nullptr, nullptr, nullptr);
-        scaler_width = frame->width;
-        scaler_height = frame->height;
+        scaler_ctx = sws_getContext(frame_width, frame_height, frame_format, frame_width,
+                                    frame_height, target_format, 0, nullptr, nullptr, nullptr);
+        scaler_width = frame_width;
+        scaler_height = frame_height;
         converted_frame_buffer.reset();
     }
     if (!converted_frame_buffer) {
-        const size_t frame_size = frame->width * frame->height * 4;
+        const size_t frame_size = frame_width * frame_height * 4;
         converted_frame_buffer = AVMallocPtr{static_cast<u8*>(av_malloc(frame_size)), av_free};
     }
-    const std::array<int, 4> converted_stride{frame->width * 4, frame->height * 4, 0, 0};
+    const std::array<int, 4> converted_stride{frame_width * 4, frame_height * 4, 0, 0};
     u8* const converted_frame_buf_addr{converted_frame_buffer.get()};
-    sws_scale(scaler_ctx, frame->data, frame->linesize, 0, frame->height, &converted_frame_buf_addr,
-              converted_stride.data());
+    sws_scale(scaler_ctx, frame->GetPlanes(), frame->GetStrides(), 0, frame_height,
+              &converted_frame_buf_addr, converted_stride.data());
 
     // Use the minimum of surface/frame dimensions to avoid buffer overflow.
     const u32 surface_width = static_cast<u32>(config.surface_width_minus1) + 1;
     const u32 surface_height = static_cast<u32>(config.surface_height_minus1) + 1;
-    const u32 width = std::min(surface_width, static_cast<u32>(frame->width));
-    const u32 height = std::min(surface_height, static_cast<u32>(frame->height));
+    const u32 width = std::min(surface_width, static_cast<u32>(frame_width));
+    const u32 height = std::min(surface_height, static_cast<u32>(frame_height));
     const u32 blk_kind = static_cast<u32>(config.block_linear_kind);
     if (blk_kind != 0) {
         // swizzle pitch linear to block linear
@@ -169,23 +171,23 @@ void Vic::WriteRGBFrame(const AVFrame* frame, const VicConfig& config) {
     }
 }
 
-void Vic::WriteYUVFrame(const AVFrame* frame, const VicConfig& config) {
+void Vic::WriteYUVFrame(std::unique_ptr<FFmpeg::Frame> frame, const VicConfig& config) {
     LOG_TRACE(Service_NVDRV, "Writing YUV420 Frame");
 
     const std::size_t surface_width = config.surface_width_minus1 + 1;
     const std::size_t surface_height = config.surface_height_minus1 + 1;
     const std::size_t aligned_width = (surface_width + 0xff) & ~0xffUL;
     // Use the minimum of surface/frame dimensions to avoid buffer overflow.
-    const auto frame_width = std::min(surface_width, static_cast<size_t>(frame->width));
-    const auto frame_height = std::min(surface_height, static_cast<size_t>(frame->height));
+    const auto frame_width = std::min(surface_width, static_cast<size_t>(frame->GetWidth()));
+    const auto frame_height = std::min(surface_height, static_cast<size_t>(frame->GetHeight()));
 
-    const auto stride = static_cast<size_t>(frame->linesize[0]);
+    const auto stride = static_cast<size_t>(frame->GetStride(0));
 
     luma_buffer.resize_destructive(aligned_width * surface_height);
     chroma_buffer.resize_destructive(aligned_width * surface_height / 2);
 
     // Populate luma buffer
-    const u8* luma_src = frame->data[0];
+    const u8* luma_src = frame->GetData(0);
     for (std::size_t y = 0; y < frame_height; ++y) {
         const std::size_t src = y * stride;
         const std::size_t dst = y * aligned_width;
@@ -196,16 +198,16 @@ void Vic::WriteYUVFrame(const AVFrame* frame, const VicConfig& config) {
 
     // Chroma
     const std::size_t half_height = frame_height / 2;
-    const auto half_stride = static_cast<size_t>(frame->linesize[1]);
+    const auto half_stride = static_cast<size_t>(frame->GetStride(1));
 
-    switch (frame->format) {
+    switch (frame->GetPixelFormat()) {
     case AV_PIX_FMT_YUV420P: {
         // Frame from FFmpeg software
         // Populate chroma buffer from both channels with interleaving.
         const std::size_t half_width = frame_width / 2;
         u8* chroma_buffer_data = chroma_buffer.data();
-        const u8* chroma_b_src = frame->data[1];
-        const u8* chroma_r_src = frame->data[2];
+        const u8* chroma_b_src = frame->GetData(1);
+        const u8* chroma_r_src = frame->GetData(2);
         for (std::size_t y = 0; y < half_height; ++y) {
             const std::size_t src = y * half_stride;
             const std::size_t dst = y * aligned_width;
@@ -219,7 +221,7 @@ void Vic::WriteYUVFrame(const AVFrame* frame, const VicConfig& config) {
     case AV_PIX_FMT_NV12: {
         // Frame from VA-API hardware
         // This is already interleaved so just copy
-        const u8* chroma_src = frame->data[1];
+        const u8* chroma_src = frame->GetData(1);
         for (std::size_t y = 0; y < half_height; ++y) {
             const std::size_t src = y * stride;
             const std::size_t dst = y * aligned_width;
diff --git a/src/video_core/host1x/vic.h b/src/video_core/host1x/vic.h
index 3d9753047..6c868f062 100644
--- a/src/video_core/host1x/vic.h
+++ b/src/video_core/host1x/vic.h
@@ -39,9 +39,9 @@ public:
 private:
     void Execute();
 
-    void WriteRGBFrame(const AVFrame* frame, const VicConfig& config);
+    void WriteRGBFrame(std::unique_ptr<FFmpeg::Frame> frame, const VicConfig& config);
 
-    void WriteYUVFrame(const AVFrame* frame, const VicConfig& config);
+    void WriteYUVFrame(std::unique_ptr<FFmpeg::Frame> frame, const VicConfig& config);
 
     Host1x& host1x;
     std::shared_ptr<Tegra::Host1x::Nvdec> nvdec_processor;
diff --git a/src/video_core/query_cache/query_cache.h b/src/video_core/query_cache/query_cache.h
index 78b42b518..efa9adf7a 100644
--- a/src/video_core/query_cache/query_cache.h
+++ b/src/video_core/query_cache/query_cache.h
@@ -266,7 +266,7 @@ void QueryCacheBase<Traits>::CounterReport(GPUVAddr addr, QueryType counter_type
             return;
         }
         if (False(query_base->flags & QueryFlagBits::IsFinalValueSynced)) [[unlikely]] {
-            UNREACHABLE();
+            ASSERT(false);
             return;
         }
         query_base->value += streamer->GetAmmendValue();
diff --git a/src/video_core/renderer_null/null_rasterizer.cpp b/src/video_core/renderer_null/null_rasterizer.cpp
index 65cd5aa06..4f1d5b548 100644
--- a/src/video_core/renderer_null/null_rasterizer.cpp
+++ b/src/video_core/renderer_null/null_rasterizer.cpp
@@ -3,6 +3,7 @@
 
 #include "common/alignment.h"
 #include "core/memory.h"
+#include "video_core/control/channel_state.h"
 #include "video_core/host1x/host1x.h"
 #include "video_core/memory_manager.h"
 #include "video_core/renderer_null/null_rasterizer.h"
@@ -99,8 +100,14 @@ bool RasterizerNull::AccelerateDisplay(const Tegra::FramebufferConfig& config,
 }
 void RasterizerNull::LoadDiskResources(u64 title_id, std::stop_token stop_loading,
                                        const VideoCore::DiskResourceLoadCallback& callback) {}
-void RasterizerNull::InitializeChannel(Tegra::Control::ChannelState& channel) {}
-void RasterizerNull::BindChannel(Tegra::Control::ChannelState& channel) {}
-void RasterizerNull::ReleaseChannel(s32 channel_id) {}
+void RasterizerNull::InitializeChannel(Tegra::Control::ChannelState& channel) {
+    CreateChannel(channel);
+}
+void RasterizerNull::BindChannel(Tegra::Control::ChannelState& channel) {
+    BindToChannel(channel.bind_id);
+}
+void RasterizerNull::ReleaseChannel(s32 channel_id) {
+    EraseChannel(channel_id);
+}
 
 } // namespace Null
diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.cpp b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
index 38d553d3c..dfd696de6 100644
--- a/src/video_core/renderer_opengl/gl_buffer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
@@ -178,13 +178,14 @@ void BufferCacheRuntime::CopyBuffer(GLuint dst_buffer, Buffer& src_buffer,
 }
 
 void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, GLuint src_buffer,
-                                    std::span<const VideoCommon::BufferCopy> copies, bool barrier) {
+                                    std::span<const VideoCommon::BufferCopy> copies, bool barrier,
+                                    bool) {
     CopyBuffer(dst_buffer.Handle(), src_buffer, copies, barrier);
 }
 
 void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
-                                    std::span<const VideoCommon::BufferCopy> copies) {
-    CopyBuffer(dst_buffer.Handle(), src_buffer.Handle(), copies);
+                                    std::span<const VideoCommon::BufferCopy> copies, bool) {
+    CopyBuffer(dst_buffer.Handle(), src_buffer.Handle(), copies, true);
 }
 
 void BufferCacheRuntime::PreCopyBarrier() {
diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.h b/src/video_core/renderer_opengl/gl_buffer_cache.h
index e8dbbd3a2..000f29a82 100644
--- a/src/video_core/renderer_opengl/gl_buffer_cache.h
+++ b/src/video_core/renderer_opengl/gl_buffer_cache.h
@@ -30,6 +30,8 @@ public:
 
     void MakeResident(GLenum access) noexcept;
 
+    void MarkUsage(u64 offset, u64 size) {}
+
     [[nodiscard]] GLuint View(u32 offset, u32 size, VideoCore::Surface::PixelFormat format);
 
     [[nodiscard]] GLuint64EXT HostGpuAddr() const noexcept {
@@ -66,22 +68,29 @@ public:
 
     [[nodiscard]] StagingBufferMap DownloadStagingBuffer(size_t size);
 
+    bool CanReorderUpload(const Buffer&, std::span<const VideoCommon::BufferCopy>) {
+        return false;
+    }
+
     void CopyBuffer(GLuint dst_buffer, GLuint src_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
+                    std::span<const VideoCommon::BufferCopy> copies, bool barrier);
 
     void CopyBuffer(GLuint dst_buffer, Buffer& src_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
+                    std::span<const VideoCommon::BufferCopy> copies, bool barrier);
 
     void CopyBuffer(Buffer& dst_buffer, GLuint src_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
+                    std::span<const VideoCommon::BufferCopy> copies, bool barrier,
+                    bool can_reorder_upload = false);
 
     void CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies);
+                    std::span<const VideoCommon::BufferCopy> copies, bool);
 
     void PreCopyBarrier();
     void PostCopyBarrier();
     void Finish();
 
+    void TickFrame(VideoCommon::SlotVector<Buffer>&) noexcept {}
+
     void ClearBuffer(Buffer& dest_buffer, u32 offset, size_t size, u32 value);
 
     void BindIndexBuffer(Buffer& buffer, u32 offset, u32 size);
diff --git a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
index 44a771d65..af0a453ee 100644
--- a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
+++ b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
@@ -559,7 +559,9 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
 }
 
 void GraphicsPipeline::ConfigureTransformFeedbackImpl() const {
-    glTransformFeedbackAttribsNV(num_xfb_attribs, xfb_attribs.data(), GL_SEPARATE_ATTRIBS);
+    const GLenum buffer_mode =
+        num_xfb_buffers_active == 1 ? GL_INTERLEAVED_ATTRIBS : GL_SEPARATE_ATTRIBS;
+    glTransformFeedbackAttribsNV(num_xfb_attribs, xfb_attribs.data(), buffer_mode);
 }
 
 void GraphicsPipeline::GenerateTransformFeedbackState() {
@@ -567,12 +569,14 @@ void GraphicsPipeline::GenerateTransformFeedbackState() {
     // when this is required.
     GLint* cursor{xfb_attribs.data()};
 
+    num_xfb_buffers_active = 0;
     for (size_t feedback = 0; feedback < Maxwell::NumTransformFeedbackBuffers; ++feedback) {
         const auto& layout = key.xfb_state.layouts[feedback];
         UNIMPLEMENTED_IF_MSG(layout.stride != layout.varying_count * 4, "Stride padding");
         if (layout.varying_count == 0) {
             continue;
         }
+        num_xfb_buffers_active++;
 
         const auto& locations = key.xfb_state.varyings[feedback];
         std::optional<u32> current_index;
diff --git a/src/video_core/renderer_opengl/gl_graphics_pipeline.h b/src/video_core/renderer_opengl/gl_graphics_pipeline.h
index 74fc9cc3d..2f70c1ae9 100644
--- a/src/video_core/renderer_opengl/gl_graphics_pipeline.h
+++ b/src/video_core/renderer_opengl/gl_graphics_pipeline.h
@@ -154,6 +154,7 @@ private:
 
     static constexpr std::size_t XFB_ENTRY_STRIDE = 3;
     GLsizei num_xfb_attribs{};
+    u32 num_xfb_buffers_active{};
     std::array<GLint, 128 * XFB_ENTRY_STRIDE * Maxwell::NumTransformFeedbackBuffers> xfb_attribs{};
 
     std::mutex built_mutex;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 27e2de1bf..9995b6dd4 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -555,7 +555,7 @@ void RasterizerOpenGL::OnCacheInvalidation(VAddr addr, u64 size) {
     }
     {
         std::scoped_lock lock{buffer_cache.mutex};
-        buffer_cache.CachedWriteMemory(addr, size);
+        buffer_cache.WriteMemory(addr, size);
     }
     shader_cache.InvalidateRegion(addr, size);
 }
diff --git a/src/video_core/renderer_vulkan/renderer_vulkan.cpp b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
index 7e7a80740..c4c30d807 100644
--- a/src/video_core/renderer_vulkan/renderer_vulkan.cpp
+++ b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
@@ -132,16 +132,12 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
     const bool use_accelerated =
         rasterizer.AccelerateDisplay(*framebuffer, framebuffer_addr, framebuffer->stride);
     const bool is_srgb = use_accelerated && screen_info.is_srgb;
+    RenderScreenshot(*framebuffer, use_accelerated);
 
-    {
-        std::scoped_lock lock{rasterizer.LockCaches()};
-        RenderScreenshot(*framebuffer, use_accelerated);
-
-        Frame* frame = present_manager.GetRenderFrame();
-        blit_screen.DrawToSwapchain(frame, *framebuffer, use_accelerated, is_srgb);
-        scheduler.Flush(*frame->render_ready);
-        present_manager.Present(frame);
-    }
+    Frame* frame = present_manager.GetRenderFrame();
+    blit_screen.DrawToSwapchain(frame, *framebuffer, use_accelerated, is_srgb);
+    scheduler.Flush(*frame->render_ready);
+    present_manager.Present(frame);
 
     gpu.RendererFrameEndNotify();
     rasterizer.TickFrame();
diff --git a/src/video_core/renderer_vulkan/vk_blit_screen.cpp b/src/video_core/renderer_vulkan/vk_blit_screen.cpp
index 52fc142d1..66483a900 100644
--- a/src/video_core/renderer_vulkan/vk_blit_screen.cpp
+++ b/src/video_core/renderer_vulkan/vk_blit_screen.cpp
@@ -137,6 +137,56 @@ BlitScreen::BlitScreen(Core::Memory::Memory& cpu_memory_, Core::Frontend::EmuWin
 
 BlitScreen::~BlitScreen() = default;
 
+static Common::Rectangle<f32> NormalizeCrop(const Tegra::FramebufferConfig& framebuffer,
+                                            const ScreenInfo& screen_info) {
+    f32 left, top, right, bottom;
+
+    if (!framebuffer.crop_rect.IsEmpty()) {
+        // If crop rectangle is not empty, apply properties from rectangle.
+        left = static_cast<f32>(framebuffer.crop_rect.left);
+        top = static_cast<f32>(framebuffer.crop_rect.top);
+        right = static_cast<f32>(framebuffer.crop_rect.right);
+        bottom = static_cast<f32>(framebuffer.crop_rect.bottom);
+    } else {
+        // Otherwise, fall back to framebuffer dimensions.
+        left = 0;
+        top = 0;
+        right = static_cast<f32>(framebuffer.width);
+        bottom = static_cast<f32>(framebuffer.height);
+    }
+
+    // Apply transformation flags.
+    auto framebuffer_transform_flags = framebuffer.transform_flags;
+
+    if (True(framebuffer_transform_flags & Service::android::BufferTransformFlags::FlipH)) {
+        // Switch left and right.
+        std::swap(left, right);
+    }
+    if (True(framebuffer_transform_flags & Service::android::BufferTransformFlags::FlipV)) {
+        // Switch top and bottom.
+        std::swap(top, bottom);
+    }
+
+    framebuffer_transform_flags &= ~Service::android::BufferTransformFlags::FlipH;
+    framebuffer_transform_flags &= ~Service::android::BufferTransformFlags::FlipV;
+    if (True(framebuffer_transform_flags)) {
+        UNIMPLEMENTED_MSG("Unsupported framebuffer_transform_flags={}",
+                          static_cast<u32>(framebuffer_transform_flags));
+    }
+
+    // Get the screen properties.
+    const f32 screen_width = static_cast<f32>(screen_info.width);
+    const f32 screen_height = static_cast<f32>(screen_info.height);
+
+    // Normalize coordinate space.
+    left /= screen_width;
+    top /= screen_height;
+    right /= screen_width;
+    bottom /= screen_height;
+
+    return Common::Rectangle<f32>(left, top, right, bottom);
+}
+
 void BlitScreen::Recreate() {
     present_manager.WaitPresent();
     scheduler.Finish();
@@ -354,17 +404,10 @@ void BlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer,
         source_image_view = smaa->Draw(scheduler, image_index, source_image, source_image_view);
     }
     if (fsr) {
-        auto crop_rect = framebuffer.crop_rect;
-        if (crop_rect.GetWidth() == 0) {
-            crop_rect.right = framebuffer.width;
-        }
-        if (crop_rect.GetHeight() == 0) {
-            crop_rect.bottom = framebuffer.height;
-        }
-        crop_rect = crop_rect.Scale(Settings::values.resolution_info.up_factor);
-        VkExtent2D fsr_input_size{
-            .width = Settings::values.resolution_info.ScaleUp(framebuffer.width),
-            .height = Settings::values.resolution_info.ScaleUp(framebuffer.height),
+        const auto crop_rect = NormalizeCrop(framebuffer, screen_info);
+        const VkExtent2D fsr_input_size{
+            .width = Settings::values.resolution_info.ScaleUp(screen_info.width),
+            .height = Settings::values.resolution_info.ScaleUp(screen_info.height),
         };
         VkImageView fsr_image_view =
             fsr->Draw(scheduler, image_index, source_image_view, fsr_input_size, crop_rect);
@@ -1397,61 +1440,37 @@ void BlitScreen::SetUniformData(BufferData& data, const Layout::FramebufferLayou
 
 void BlitScreen::SetVertexData(BufferData& data, const Tegra::FramebufferConfig& framebuffer,
                                const Layout::FramebufferLayout layout) const {
-    const auto& framebuffer_transform_flags = framebuffer.transform_flags;
-    const auto& framebuffer_crop_rect = framebuffer.crop_rect;
-
-    static constexpr Common::Rectangle<f32> texcoords{0.f, 0.f, 1.f, 1.f};
-    auto left = texcoords.left;
-    auto right = texcoords.right;
-
-    switch (framebuffer_transform_flags) {
-    case Service::android::BufferTransformFlags::Unset:
-        break;
-    case Service::android::BufferTransformFlags::FlipV:
-        // Flip the framebuffer vertically
-        left = texcoords.right;
-        right = texcoords.left;
-        break;
-    default:
-        UNIMPLEMENTED_MSG("Unsupported framebuffer_transform_flags={}",
-                          static_cast<u32>(framebuffer_transform_flags));
-        break;
-    }
+    f32 left, top, right, bottom;
 
-    UNIMPLEMENTED_IF(framebuffer_crop_rect.left != 0);
-
-    f32 left_start{};
-    if (framebuffer_crop_rect.Top() > 0) {
-        left_start = static_cast<f32>(framebuffer_crop_rect.Top()) /
-                     static_cast<f32>(framebuffer_crop_rect.Bottom());
-    }
-    f32 scale_u = static_cast<f32>(framebuffer.width) / static_cast<f32>(screen_info.width);
-    f32 scale_v = static_cast<f32>(framebuffer.height) / static_cast<f32>(screen_info.height);
-    // Scale the output by the crop width/height. This is commonly used with 1280x720 rendering
-    // (e.g. handheld mode) on a 1920x1080 framebuffer.
-    if (!fsr) {
-        if (framebuffer_crop_rect.GetWidth() > 0) {
-            scale_u = static_cast<f32>(framebuffer_crop_rect.GetWidth()) /
-                      static_cast<f32>(screen_info.width);
-        }
-        if (framebuffer_crop_rect.GetHeight() > 0) {
-            scale_v = static_cast<f32>(framebuffer_crop_rect.GetHeight()) /
-                      static_cast<f32>(screen_info.height);
-        }
+    if (fsr) {
+        // FSR has already applied the crop, so we just want to render the image
+        // it has produced.
+        left = 0;
+        top = 0;
+        right = 1;
+        bottom = 1;
+    } else {
+        // Get the normalized crop rectangle.
+        const auto crop = NormalizeCrop(framebuffer, screen_info);
+
+        // Apply the crop.
+        left = crop.left;
+        top = crop.top;
+        right = crop.right;
+        bottom = crop.bottom;
     }
 
+    // Map the coordinates to the screen.
     const auto& screen = layout.screen;
     const auto x = static_cast<f32>(screen.left);
     const auto y = static_cast<f32>(screen.top);
     const auto w = static_cast<f32>(screen.GetWidth());
     const auto h = static_cast<f32>(screen.GetHeight());
-    data.vertices[0] = ScreenRectVertex(x, y, texcoords.top * scale_u, left_start + left * scale_v);
-    data.vertices[1] =
-        ScreenRectVertex(x + w, y, texcoords.bottom * scale_u, left_start + left * scale_v);
-    data.vertices[2] =
-        ScreenRectVertex(x, y + h, texcoords.top * scale_u, left_start + right * scale_v);
-    data.vertices[3] =
-        ScreenRectVertex(x + w, y + h, texcoords.bottom * scale_u, left_start + right * scale_v);
+
+    data.vertices[0] = ScreenRectVertex(x, y, left, top);
+    data.vertices[1] = ScreenRectVertex(x + w, y, right, top);
+    data.vertices[2] = ScreenRectVertex(x, y + h, left, bottom);
+    data.vertices[3] = ScreenRectVertex(x + w, y + h, right, bottom);
 }
 
 void BlitScreen::CreateSMAA(VkExtent2D smaa_size) {
diff --git a/src/video_core/renderer_vulkan/vk_buffer_cache.cpp b/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
index 976c3f6a6..5958f52f7 100644
--- a/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
@@ -79,13 +79,13 @@ vk::Buffer CreateBuffer(const Device& device, const MemoryAllocator& memory_allo
 } // Anonymous namespace
 
 Buffer::Buffer(BufferCacheRuntime&, VideoCommon::NullBufferParams null_params)
-    : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(null_params) {}
+    : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(null_params), tracker{4096} {}
 
 Buffer::Buffer(BufferCacheRuntime& runtime, VideoCore::RasterizerInterface& rasterizer_,
                VAddr cpu_addr_, u64 size_bytes_)
     : VideoCommon::BufferBase<VideoCore::RasterizerInterface>(rasterizer_, cpu_addr_, size_bytes_),
-      device{&runtime.device}, buffer{
-                                   CreateBuffer(*device, runtime.memory_allocator, SizeBytes())} {
+      device{&runtime.device}, buffer{CreateBuffer(*device, runtime.memory_allocator, SizeBytes())},
+      tracker{SizeBytes()} {
     if (runtime.device.HasDebuggingToolAttached()) {
         buffer.SetObjectNameEXT(fmt::format("Buffer 0x{:x}", CpuAddr()).c_str());
     }
@@ -359,12 +359,31 @@ u32 BufferCacheRuntime::GetStorageBufferAlignment() const {
     return static_cast<u32>(device.GetStorageBufferAlignment());
 }
 
+void BufferCacheRuntime::TickFrame(VideoCommon::SlotVector<Buffer>& slot_buffers) noexcept {
+    for (auto it = slot_buffers.begin(); it != slot_buffers.end(); it++) {
+        it->ResetUsageTracking();
+    }
+}
+
 void BufferCacheRuntime::Finish() {
     scheduler.Finish();
 }
 
+bool BufferCacheRuntime::CanReorderUpload(const Buffer& buffer,
+                                          std::span<const VideoCommon::BufferCopy> copies) {
+    if (Settings::values.disable_buffer_reorder) {
+        return false;
+    }
+    const bool can_use_upload_cmdbuf =
+        std::ranges::all_of(copies, [&](const VideoCommon::BufferCopy& copy) {
+            return !buffer.IsRegionUsed(copy.dst_offset, copy.size);
+        });
+    return can_use_upload_cmdbuf;
+}
+
 void BufferCacheRuntime::CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
-                                    std::span<const VideoCommon::BufferCopy> copies, bool barrier) {
+                                    std::span<const VideoCommon::BufferCopy> copies, bool barrier,
+                                    bool can_reorder_upload) {
     if (dst_buffer == VK_NULL_HANDLE || src_buffer == VK_NULL_HANDLE) {
         return;
     }
@@ -380,9 +399,18 @@ void BufferCacheRuntime::CopyBuffer(VkBuffer dst_buffer, VkBuffer src_buffer,
         .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
         .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
     };
+
     // Measuring a popular game, this number never exceeds the specified size once data is warmed up
     boost::container::small_vector<VkBufferCopy, 8> vk_copies(copies.size());
     std::ranges::transform(copies, vk_copies.begin(), MakeBufferCopy);
+    if (src_buffer == staging_pool.StreamBuf() && can_reorder_upload) {
+        scheduler.RecordWithUploadBuffer([src_buffer, dst_buffer, vk_copies](
+                                             vk::CommandBuffer, vk::CommandBuffer upload_cmdbuf) {
+            upload_cmdbuf.CopyBuffer(src_buffer, dst_buffer, vk_copies);
+        });
+        return;
+    }
+
     scheduler.RequestOutsideRenderPassOperationContext();
     scheduler.Record([src_buffer, dst_buffer, vk_copies, barrier](vk::CommandBuffer cmdbuf) {
         if (barrier) {
diff --git a/src/video_core/renderer_vulkan/vk_buffer_cache.h b/src/video_core/renderer_vulkan/vk_buffer_cache.h
index 833dfac45..0b3fbd6d0 100644
--- a/src/video_core/renderer_vulkan/vk_buffer_cache.h
+++ b/src/video_core/renderer_vulkan/vk_buffer_cache.h
@@ -5,6 +5,7 @@
 
 #include "video_core/buffer_cache/buffer_cache_base.h"
 #include "video_core/buffer_cache/memory_tracker_base.h"
+#include "video_core/buffer_cache/usage_tracker.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
 #include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
@@ -34,6 +35,18 @@ public:
         return *buffer;
     }
 
+    [[nodiscard]] bool IsRegionUsed(u64 offset, u64 size) const noexcept {
+        return tracker.IsUsed(offset, size);
+    }
+
+    void MarkUsage(u64 offset, u64 size) noexcept {
+        tracker.Track(offset, size);
+    }
+
+    void ResetUsageTracking() noexcept {
+        tracker.Reset();
+    }
+
     operator VkBuffer() const noexcept {
         return *buffer;
     }
@@ -49,6 +62,7 @@ private:
     const Device* device{};
     vk::Buffer buffer;
     std::vector<BufferView> views;
+    VideoCommon::UsageTracker tracker;
 };
 
 class QuadArrayIndexBuffer;
@@ -67,6 +81,8 @@ public:
                                 ComputePassDescriptorQueue& compute_pass_descriptor_queue,
                                 DescriptorPool& descriptor_pool);
 
+    void TickFrame(VideoCommon::SlotVector<Buffer>& slot_buffers) noexcept;
+
     void Finish();
 
     u64 GetDeviceLocalMemory() const;
@@ -81,12 +97,15 @@ public:
 
     [[nodiscard]] StagingBufferRef DownloadStagingBuffer(size_t size, bool deferred = false);
 
+    bool CanReorderUpload(const Buffer& buffer, std::span<const VideoCommon::BufferCopy> copies);
+
     void FreeDeferredStagingBuffer(StagingBufferRef& ref);
 
     void PreCopyBarrier();
 
     void CopyBuffer(VkBuffer src_buffer, VkBuffer dst_buffer,
-                    std::span<const VideoCommon::BufferCopy> copies, bool barrier = true);
+                    std::span<const VideoCommon::BufferCopy> copies, bool barrier,
+                    bool can_reorder_upload = false);
 
     void PostCopyBarrier();
 
diff --git a/src/video_core/renderer_vulkan/vk_fsr.cpp b/src/video_core/renderer_vulkan/vk_fsr.cpp
index ce8f3f3c2..f7a05fbc0 100644
--- a/src/video_core/renderer_vulkan/vk_fsr.cpp
+++ b/src/video_core/renderer_vulkan/vk_fsr.cpp
@@ -34,7 +34,7 @@ FSR::FSR(const Device& device_, MemoryAllocator& memory_allocator_, size_t image
 }
 
 VkImageView FSR::Draw(Scheduler& scheduler, size_t image_index, VkImageView image_view,
-                      VkExtent2D input_image_extent, const Common::Rectangle<int>& crop_rect) {
+                      VkExtent2D input_image_extent, const Common::Rectangle<f32>& crop_rect) {
 
     UpdateDescriptorSet(image_index, image_view);
 
@@ -61,15 +61,21 @@ VkImageView FSR::Draw(Scheduler& scheduler, size_t image_index, VkImageView imag
 
         cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *easu_pipeline);
 
+        const f32 input_image_width = static_cast<f32>(input_image_extent.width);
+        const f32 input_image_height = static_cast<f32>(input_image_extent.height);
+        const f32 output_image_width = static_cast<f32>(output_size.width);
+        const f32 output_image_height = static_cast<f32>(output_size.height);
+        const f32 viewport_width = (crop_rect.right - crop_rect.left) * input_image_width;
+        const f32 viewport_x = crop_rect.left * input_image_width;
+        const f32 viewport_height = (crop_rect.bottom - crop_rect.top) * input_image_height;
+        const f32 viewport_y = crop_rect.top * input_image_height;
+
         std::array<u32, 4 * 4> push_constants;
-        FsrEasuConOffset(
-            push_constants.data() + 0, push_constants.data() + 4, push_constants.data() + 8,
-            push_constants.data() + 12,
-
-            static_cast<f32>(crop_rect.GetWidth()), static_cast<f32>(crop_rect.GetHeight()),
-            static_cast<f32>(input_image_extent.width), static_cast<f32>(input_image_extent.height),
-            static_cast<f32>(output_size.width), static_cast<f32>(output_size.height),
-            static_cast<f32>(crop_rect.left), static_cast<f32>(crop_rect.top));
+        FsrEasuConOffset(push_constants.data() + 0, push_constants.data() + 4,
+                         push_constants.data() + 8, push_constants.data() + 12,
+
+                         viewport_width, viewport_height, input_image_width, input_image_height,
+                         output_image_width, output_image_height, viewport_x, viewport_y);
         cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, push_constants);
 
         {
diff --git a/src/video_core/renderer_vulkan/vk_fsr.h b/src/video_core/renderer_vulkan/vk_fsr.h
index 8bb9fc23a..3505c1416 100644
--- a/src/video_core/renderer_vulkan/vk_fsr.h
+++ b/src/video_core/renderer_vulkan/vk_fsr.h
@@ -17,7 +17,7 @@ public:
     explicit FSR(const Device& device, MemoryAllocator& memory_allocator, size_t image_count,
                  VkExtent2D output_size);
     VkImageView Draw(Scheduler& scheduler, size_t image_index, VkImageView image_view,
-                     VkExtent2D input_image_extent, const Common::Rectangle<int>& crop_rect);
+                     VkExtent2D input_image_extent, const Common::Rectangle<f32>& crop_rect);
 
 private:
     void CreateDescriptorPool();
diff --git a/src/video_core/renderer_vulkan/vk_master_semaphore.cpp b/src/video_core/renderer_vulkan/vk_master_semaphore.cpp
index 6b288b994..ac8b6e838 100644
--- a/src/video_core/renderer_vulkan/vk_master_semaphore.cpp
+++ b/src/video_core/renderer_vulkan/vk_master_semaphore.cpp
@@ -100,12 +100,14 @@ void MasterSemaphore::Wait(u64 tick) {
     Refresh();
 }
 
-VkResult MasterSemaphore::SubmitQueue(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                                      VkSemaphore wait_semaphore, u64 host_tick) {
+VkResult MasterSemaphore::SubmitQueue(vk::CommandBuffer& cmdbuf, vk::CommandBuffer& upload_cmdbuf,
+                                      VkSemaphore signal_semaphore, VkSemaphore wait_semaphore,
+                                      u64 host_tick) {
     if (semaphore) {
-        return SubmitQueueTimeline(cmdbuf, signal_semaphore, wait_semaphore, host_tick);
+        return SubmitQueueTimeline(cmdbuf, upload_cmdbuf, signal_semaphore, wait_semaphore,
+                                   host_tick);
     } else {
-        return SubmitQueueFence(cmdbuf, signal_semaphore, wait_semaphore, host_tick);
+        return SubmitQueueFence(cmdbuf, upload_cmdbuf, signal_semaphore, wait_semaphore, host_tick);
     }
 }
 
@@ -115,6 +117,7 @@ static constexpr std::array<VkPipelineStageFlags, 2> wait_stage_masks{
 };
 
 VkResult MasterSemaphore::SubmitQueueTimeline(vk::CommandBuffer& cmdbuf,
+                                              vk::CommandBuffer& upload_cmdbuf,
                                               VkSemaphore signal_semaphore,
                                               VkSemaphore wait_semaphore, u64 host_tick) {
     const VkSemaphore timeline_semaphore = *semaphore;
@@ -123,6 +126,8 @@ VkResult MasterSemaphore::SubmitQueueTimeline(vk::CommandBuffer& cmdbuf,
     const std::array signal_values{host_tick, u64(0)};
     const std::array signal_semaphores{timeline_semaphore, signal_semaphore};
 
+    const std::array cmdbuffers{*upload_cmdbuf, *cmdbuf};
+
     const u32 num_wait_semaphores = wait_semaphore ? 1 : 0;
     const VkTimelineSemaphoreSubmitInfo timeline_si{
         .sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO,
@@ -138,8 +143,8 @@ VkResult MasterSemaphore::SubmitQueueTimeline(vk::CommandBuffer& cmdbuf,
         .waitSemaphoreCount = num_wait_semaphores,
         .pWaitSemaphores = &wait_semaphore,
         .pWaitDstStageMask = wait_stage_masks.data(),
-        .commandBufferCount = 1,
-        .pCommandBuffers = cmdbuf.address(),
+        .commandBufferCount = static_cast<u32>(cmdbuffers.size()),
+        .pCommandBuffers = cmdbuffers.data(),
         .signalSemaphoreCount = num_signal_semaphores,
         .pSignalSemaphores = signal_semaphores.data(),
     };
@@ -147,19 +152,23 @@ VkResult MasterSemaphore::SubmitQueueTimeline(vk::CommandBuffer& cmdbuf,
     return device.GetGraphicsQueue().Submit(submit_info);
 }
 
-VkResult MasterSemaphore::SubmitQueueFence(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                                           VkSemaphore wait_semaphore, u64 host_tick) {
+VkResult MasterSemaphore::SubmitQueueFence(vk::CommandBuffer& cmdbuf,
+                                           vk::CommandBuffer& upload_cmdbuf,
+                                           VkSemaphore signal_semaphore, VkSemaphore wait_semaphore,
+                                           u64 host_tick) {
     const u32 num_signal_semaphores = signal_semaphore ? 1 : 0;
     const u32 num_wait_semaphores = wait_semaphore ? 1 : 0;
 
+    const std::array cmdbuffers{*upload_cmdbuf, *cmdbuf};
+
     const VkSubmitInfo submit_info{
         .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
         .pNext = nullptr,
         .waitSemaphoreCount = num_wait_semaphores,
         .pWaitSemaphores = &wait_semaphore,
         .pWaitDstStageMask = wait_stage_masks.data(),
-        .commandBufferCount = 1,
-        .pCommandBuffers = cmdbuf.address(),
+        .commandBufferCount = static_cast<u32>(cmdbuffers.size()),
+        .pCommandBuffers = cmdbuffers.data(),
         .signalSemaphoreCount = num_signal_semaphores,
         .pSignalSemaphores = &signal_semaphore,
     };
diff --git a/src/video_core/renderer_vulkan/vk_master_semaphore.h b/src/video_core/renderer_vulkan/vk_master_semaphore.h
index 3f599d7bd..7dfb93ffb 100644
--- a/src/video_core/renderer_vulkan/vk_master_semaphore.h
+++ b/src/video_core/renderer_vulkan/vk_master_semaphore.h
@@ -52,14 +52,16 @@ public:
     void Wait(u64 tick);
 
     /// Submits the device graphics queue, updating the tick as necessary
-    VkResult SubmitQueue(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                         VkSemaphore wait_semaphore, u64 host_tick);
+    VkResult SubmitQueue(vk::CommandBuffer& cmdbuf, vk::CommandBuffer& upload_cmdbuf,
+                         VkSemaphore signal_semaphore, VkSemaphore wait_semaphore, u64 host_tick);
 
 private:
-    VkResult SubmitQueueTimeline(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                                 VkSemaphore wait_semaphore, u64 host_tick);
-    VkResult SubmitQueueFence(vk::CommandBuffer& cmdbuf, VkSemaphore signal_semaphore,
-                              VkSemaphore wait_semaphore, u64 host_tick);
+    VkResult SubmitQueueTimeline(vk::CommandBuffer& cmdbuf, vk::CommandBuffer& upload_cmdbuf,
+                                 VkSemaphore signal_semaphore, VkSemaphore wait_semaphore,
+                                 u64 host_tick);
+    VkResult SubmitQueueFence(vk::CommandBuffer& cmdbuf, vk::CommandBuffer& upload_cmdbuf,
+                              VkSemaphore signal_semaphore, VkSemaphore wait_semaphore,
+                              u64 host_tick);
 
     void WaitThread(std::stop_token token);
 
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index 0d604eee3..2a13b2a72 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -263,6 +263,22 @@ Shader::RuntimeInfo MakeRuntimeInfo(std::span<const Shader::IR::Program> program
     info.y_negate = key.state.y_negate != 0;
     return info;
 }
+
+size_t GetTotalPipelineWorkers() {
+    const size_t max_core_threads =
+        std::max<size_t>(static_cast<size_t>(std::thread::hardware_concurrency()), 2ULL) - 1ULL;
+#ifdef ANDROID
+    // Leave at least a few cores free in android
+    constexpr size_t free_cores = 3ULL;
+    if (max_core_threads <= free_cores) {
+        return 1ULL;
+    }
+    return max_core_threads - free_cores;
+#else
+    return max_core_threads;
+#endif
+}
+
 } // Anonymous namespace
 
 size_t ComputePipelineCacheKey::Hash() const noexcept {
@@ -294,11 +310,8 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, const Device& device
       texture_cache{texture_cache_}, shader_notify{shader_notify_},
       use_asynchronous_shaders{Settings::values.use_asynchronous_shaders.GetValue()},
       use_vulkan_pipeline_cache{Settings::values.use_vulkan_driver_pipeline_cache.GetValue()},
-#ifdef ANDROID
-      workers(1, "VkPipelineBuilder"),
-#else
-      workers(std::max(std::thread::hardware_concurrency(), 2U) - 1, "VkPipelineBuilder"),
-#endif
+      workers(device.HasBrokenParallelShaderCompiling() ? 1ULL : GetTotalPipelineWorkers(),
+              "VkPipelineBuilder"),
       serialization_thread(1, "VkPipelineSerialization") {
     const auto& float_control{device.FloatControlProperties()};
     const VkDriverId driver_id{device.GetDriverID()};
@@ -338,6 +351,7 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, const Device& device
         .support_geometry_shader_passthrough = device.IsNvGeometryShaderPassthroughSupported(),
         .support_native_ndc = device.IsExtDepthClipControlSupported(),
         .support_scaled_attributes = !device.MustEmulateScaledFormats(),
+        .support_multi_viewport = device.SupportsMultiViewport(),
 
         .warp_size_potentially_larger_than_guest = device.IsWarpSizePotentiallyBiggerThanGuest(),
 
diff --git a/src/video_core/renderer_vulkan/vk_query_cache.cpp b/src/video_core/renderer_vulkan/vk_query_cache.cpp
index 66c03bf17..078777cdd 100644
--- a/src/video_core/renderer_vulkan/vk_query_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_query_cache.cpp
@@ -211,6 +211,13 @@ public:
             return;
         }
         PauseCounter();
+        const auto driver_id = device.GetDriverID();
+        if (driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY ||
+            driver_id == VK_DRIVER_ID_ARM_PROPRIETARY || driver_id == VK_DRIVER_ID_MESA_TURNIP) {
+            pending_sync.clear();
+            sync_values_stash.clear();
+            return;
+        }
         sync_values_stash.clear();
         sync_values_stash.emplace_back();
         std::vector<HostSyncValues>* sync_values = &sync_values_stash.back();
@@ -1378,6 +1385,12 @@ bool QueryCacheRuntime::HostConditionalRenderingCompareValues(VideoCommon::Looku
         return true;
     }
 
+    auto driver_id = impl->device.GetDriverID();
+    if (driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY ||
+        driver_id == VK_DRIVER_ID_ARM_PROPRIETARY || driver_id == VK_DRIVER_ID_MESA_TURNIP) {
+        return true;
+    }
+
     for (size_t i = 0; i < 2; i++) {
         is_null[i] = !is_in_ac[i] && check_value(objects[i]->address);
     }
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 059b7cb40..e0ab1eaac 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -82,7 +82,7 @@ VkViewport GetViewportState(const Device& device, const Maxwell& regs, size_t in
     }
 
     if (y_negate) {
-        y += height;
+        y += conv(static_cast<f32>(regs.surface_clip.height));
         height = -height;
     }
 
@@ -199,7 +199,7 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
     if (!pipeline) {
         return;
     }
-    std::scoped_lock lock{LockCaches()};
+    std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
     // update engine as channel may be different.
     pipeline->SetEngine(maxwell3d, gpu_memory);
     pipeline->Configure(is_indexed);
@@ -621,7 +621,7 @@ void RasterizerVulkan::OnCacheInvalidation(VAddr addr, u64 size) {
     }
     {
         std::scoped_lock lock{buffer_cache.mutex};
-        buffer_cache.CachedWriteMemory(addr, size);
+        buffer_cache.WriteMemory(addr, size);
     }
     pipeline_cache.InvalidateRegion(addr, size);
 }
@@ -710,7 +710,6 @@ void RasterizerVulkan::TiledCacheBarrier() {
 }
 
 void RasterizerVulkan::FlushCommands() {
-    std::scoped_lock lock{LockCaches()};
     if (draw_counter == 0) {
         return;
     }
@@ -808,7 +807,6 @@ void RasterizerVulkan::FlushWork() {
     if ((++draw_counter & 7) != 7) {
         return;
     }
-    std::scoped_lock lock{LockCaches()};
     if (draw_counter < DRAWS_TO_DISPATCH) {
         // Send recorded tasks to the worker thread
         scheduler.DispatchWork();
@@ -923,9 +921,13 @@ void RasterizerVulkan::UpdateDynamicStates() {
 }
 
 void RasterizerVulkan::HandleTransformFeedback() {
+    static std::once_flag warn_unsupported;
+
     const auto& regs = maxwell3d->regs;
     if (!device.IsExtTransformFeedbackSupported()) {
-        LOG_ERROR(Render_Vulkan, "Transform feedbacks used but not supported");
+        std::call_once(warn_unsupported, [&] {
+            LOG_ERROR(Render_Vulkan, "Transform feedbacks used but not supported");
+        });
         return;
     }
     query_cache.CounterEnable(VideoCommon::QueryType::StreamingByteCount,
@@ -1503,7 +1505,7 @@ void RasterizerVulkan::UpdateVertexInput(Tegra::Engines::Maxwell3D::Regs& regs)
 void RasterizerVulkan::InitializeChannel(Tegra::Control::ChannelState& channel) {
     CreateChannel(channel);
     {
-        std::scoped_lock lock{LockCaches()};
+        std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
         texture_cache.CreateChannel(channel);
         buffer_cache.CreateChannel(channel);
     }
@@ -1516,7 +1518,7 @@ void RasterizerVulkan::BindChannel(Tegra::Control::ChannelState& channel) {
     const s32 channel_id = channel.bind_id;
     BindToChannel(channel_id);
     {
-        std::scoped_lock lock{LockCaches()};
+        std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
         texture_cache.BindToChannel(channel_id);
         buffer_cache.BindToChannel(channel_id);
     }
@@ -1529,7 +1531,7 @@ void RasterizerVulkan::BindChannel(Tegra::Control::ChannelState& channel) {
 void RasterizerVulkan::ReleaseChannel(s32 channel_id) {
     EraseChannel(channel_id);
     {
-        std::scoped_lock lock{LockCaches()};
+        std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
         texture_cache.EraseChannel(channel_id);
         buffer_cache.EraseChannel(channel_id);
     }
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index ce3dfbaab..ad069556c 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -133,10 +133,6 @@ public:
 
     void ReleaseChannel(s32 channel_id) override;
 
-    std::scoped_lock<std::recursive_mutex, std::recursive_mutex> LockCaches() {
-        return std::scoped_lock{buffer_cache.mutex, texture_cache.mutex};
-    }
-
 private:
     static constexpr size_t MAX_TEXTURES = 192;
     static constexpr size_t MAX_IMAGES = 48;
diff --git a/src/video_core/renderer_vulkan/vk_scheduler.cpp b/src/video_core/renderer_vulkan/vk_scheduler.cpp
index 3be7837f4..146923db4 100644
--- a/src/video_core/renderer_vulkan/vk_scheduler.cpp
+++ b/src/video_core/renderer_vulkan/vk_scheduler.cpp
@@ -22,11 +22,12 @@ namespace Vulkan {
 
 MICROPROFILE_DECLARE(Vulkan_WaitForWorker);
 
-void Scheduler::CommandChunk::ExecuteAll(vk::CommandBuffer cmdbuf) {
+void Scheduler::CommandChunk::ExecuteAll(vk::CommandBuffer cmdbuf,
+                                         vk::CommandBuffer upload_cmdbuf) {
     auto command = first;
     while (command != nullptr) {
         auto next = command->GetNext();
-        command->Execute(cmdbuf);
+        command->Execute(cmdbuf, upload_cmdbuf);
         command->~Command();
         command = next;
     }
@@ -180,7 +181,7 @@ void Scheduler::WorkerThread(std::stop_token stop_token) {
             // Perform the work, tracking whether the chunk was a submission
             // before executing.
             const bool has_submit = work->HasSubmit();
-            work->ExecuteAll(current_cmdbuf);
+            work->ExecuteAll(current_cmdbuf, current_upload_cmdbuf);
 
             // If the chunk was a submission, reallocate the command buffer.
             if (has_submit) {
@@ -205,6 +206,13 @@ void Scheduler::AllocateWorkerCommandBuffer() {
         .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
         .pInheritanceInfo = nullptr,
     });
+    current_upload_cmdbuf = vk::CommandBuffer(command_pool->Commit(), device.GetDispatchLoader());
+    current_upload_cmdbuf.Begin({
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+        .pNext = nullptr,
+        .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
+        .pInheritanceInfo = nullptr,
+    });
 }
 
 u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
@@ -212,7 +220,17 @@ u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_se
     InvalidateState();
 
     const u64 signal_value = master_semaphore->NextTick();
-    Record([signal_semaphore, wait_semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
+    RecordWithUploadBuffer([signal_semaphore, wait_semaphore, signal_value,
+                            this](vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf) {
+        static constexpr VkMemoryBarrier WRITE_BARRIER{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
+        };
+        upload_cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT,
+                                      VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0, WRITE_BARRIER);
+        upload_cmdbuf.End();
         cmdbuf.End();
 
         if (on_submit) {
@@ -221,7 +239,7 @@ u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_se
 
         std::scoped_lock lock{submit_mutex};
         switch (const VkResult result = master_semaphore->SubmitQueue(
-                    cmdbuf, signal_semaphore, wait_semaphore, signal_value)) {
+                    cmdbuf, upload_cmdbuf, signal_semaphore, wait_semaphore, signal_value)) {
         case VK_SUCCESS:
             break;
         case VK_ERROR_DEVICE_LOST:
diff --git a/src/video_core/renderer_vulkan/vk_scheduler.h b/src/video_core/renderer_vulkan/vk_scheduler.h
index da03803aa..f8d8ca80a 100644
--- a/src/video_core/renderer_vulkan/vk_scheduler.h
+++ b/src/video_core/renderer_vulkan/vk_scheduler.h
@@ -80,7 +80,8 @@ public:
 
     /// Send work to a separate thread.
     template <typename T>
-    void Record(T&& command) {
+        requires std::is_invocable_v<T, vk::CommandBuffer, vk::CommandBuffer>
+    void RecordWithUploadBuffer(T&& command) {
         if (chunk->Record(command)) {
             return;
         }
@@ -88,6 +89,15 @@ public:
         (void)chunk->Record(command);
     }
 
+    template <typename T>
+        requires std::is_invocable_v<T, vk::CommandBuffer>
+    void Record(T&& c) {
+        this->RecordWithUploadBuffer(
+            [command = std::move(c)](vk::CommandBuffer cmdbuf, vk::CommandBuffer) {
+                command(cmdbuf);
+            });
+    }
+
     /// Returns the current command buffer tick.
     [[nodiscard]] u64 CurrentTick() const noexcept {
         return master_semaphore->CurrentTick();
@@ -119,7 +129,7 @@ private:
     public:
         virtual ~Command() = default;
 
-        virtual void Execute(vk::CommandBuffer cmdbuf) const = 0;
+        virtual void Execute(vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf) const = 0;
 
         Command* GetNext() const {
             return next;
@@ -142,8 +152,8 @@ private:
         TypedCommand(TypedCommand&&) = delete;
         TypedCommand& operator=(TypedCommand&&) = delete;
 
-        void Execute(vk::CommandBuffer cmdbuf) const override {
-            command(cmdbuf);
+        void Execute(vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf) const override {
+            command(cmdbuf, upload_cmdbuf);
         }
 
     private:
@@ -152,7 +162,7 @@ private:
 
     class CommandChunk final {
     public:
-        void ExecuteAll(vk::CommandBuffer cmdbuf);
+        void ExecuteAll(vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf);
 
         template <typename T>
         bool Record(T& command) {
@@ -228,6 +238,7 @@ private:
     VideoCommon::QueryCacheBase<QueryCacheParams>* query_cache = nullptr;
 
     vk::CommandBuffer current_cmdbuf;
+    vk::CommandBuffer current_upload_cmdbuf;
 
     std::unique_ptr<CommandChunk> chunk;
     std::function<void()> on_submit;
diff --git a/src/video_core/renderer_vulkan/vk_smaa.cpp b/src/video_core/renderer_vulkan/vk_smaa.cpp
index 5efd7d66e..70644ea82 100644
--- a/src/video_core/renderer_vulkan/vk_smaa.cpp
+++ b/src/video_core/renderer_vulkan/vk_smaa.cpp
@@ -672,7 +672,7 @@ void SMAA::UploadImages(Scheduler& scheduler) {
     UploadImage(m_device, m_allocator, scheduler, m_static_images[Search], search_extent,
                 VK_FORMAT_R8_UNORM, ARRAY_TO_SPAN(searchTexBytes));
 
-    scheduler.Record([&](vk::CommandBuffer& cmdbuf) {
+    scheduler.Record([&](vk::CommandBuffer cmdbuf) {
         for (auto& images : m_dynamic_images) {
             for (size_t i = 0; i < MaxDynamicImage; i++) {
                 ClearColorImage(cmdbuf, *images.images[i]);
@@ -707,7 +707,7 @@ VkImageView SMAA::Draw(Scheduler& scheduler, size_t image_index, VkImage source_
     UpdateDescriptorSets(source_image_view, image_index);
 
     scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([=, this](vk::CommandBuffer& cmdbuf) {
+    scheduler.Record([=, this](vk::CommandBuffer cmdbuf) {
         TransitionImageLayout(cmdbuf, source_image, VK_IMAGE_LAYOUT_GENERAL);
         TransitionImageLayout(cmdbuf, edges_image, VK_IMAGE_LAYOUT_GENERAL);
         BeginRenderPass(cmdbuf, m_renderpasses[EdgeDetection], edge_detection_framebuffer,
diff --git a/src/video_core/renderer_vulkan/vk_staging_buffer_pool.h b/src/video_core/renderer_vulkan/vk_staging_buffer_pool.h
index d3deb9072..f63a20327 100644
--- a/src/video_core/renderer_vulkan/vk_staging_buffer_pool.h
+++ b/src/video_core/renderer_vulkan/vk_staging_buffer_pool.h
@@ -36,6 +36,10 @@ public:
     StagingBufferRef Request(size_t size, MemoryUsage usage, bool deferred = false);
     void FreeDeferred(StagingBufferRef& ref);
 
+    [[nodiscard]] VkBuffer StreamBuf() const noexcept {
+        return *stream_buffer;
+    }
+
     void TickFrame();
 
 private:
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index de34f6d49..5dbec2e62 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -1785,8 +1785,22 @@ ImageView::ImageView(TextureCacheRuntime&, const VideoCommon::ImageInfo& info,
     : VideoCommon::ImageViewBase{info, view_info, gpu_addr_},
       buffer_size{VideoCommon::CalculateGuestSizeInBytes(info)} {}
 
-ImageView::ImageView(TextureCacheRuntime&, const VideoCommon::NullImageViewParams& params)
-    : VideoCommon::ImageViewBase{params} {}
+ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::NullImageViewParams& params)
+    : VideoCommon::ImageViewBase{params}, device{&runtime.device} {
+    if (device->HasNullDescriptor()) {
+        return;
+    }
+
+    // Handle fallback for devices without nullDescriptor
+    ImageInfo info{};
+    info.format = PixelFormat::A8B8G8R8_UNORM;
+
+    null_image = MakeImage(*device, runtime.memory_allocator, info, {});
+    image_handle = *null_image;
+    for (u32 i = 0; i < Shader::NUM_TEXTURE_TYPES; i++) {
+        image_views[i] = MakeView(VK_FORMAT_A8B8G8R8_UNORM_PACK32, VK_IMAGE_ASPECT_COLOR_BIT);
+    }
+}
 
 ImageView::~ImageView() = default;
 
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.h b/src/video_core/renderer_vulkan/vk_texture_cache.h
index 7a0807709..edf5d7635 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.h
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.h
@@ -267,6 +267,7 @@ private:
     vk::ImageView depth_view;
     vk::ImageView stencil_view;
     vk::ImageView color_view;
+    vk::Image null_image;
     VkImage image_handle = VK_NULL_HANDLE;
     VkImageView render_target = VK_NULL_HANDLE;
     VkSampleCountFlagBits samples = VK_SAMPLE_COUNT_1_BIT;
diff --git a/src/video_core/texture_cache/slot_vector.h b/src/video_core/texture_cache/slot_vector.h
index 9df6a2903..3ffa2a661 100644
--- a/src/video_core/texture_cache/slot_vector.h
+++ b/src/video_core/texture_cache/slot_vector.h
@@ -138,6 +138,10 @@ public:
         return Iterator(this, SlotId{SlotId::INVALID_INDEX});
     }
 
+    [[nodiscard]] size_t size() const noexcept {
+        return values_capacity - free_list.size();
+    }
+
 private:
     struct NonTrivialDummy {
         NonTrivialDummy() noexcept {}
diff --git a/src/video_core/vulkan_common/vulkan_device.cpp b/src/video_core/vulkan_common/vulkan_device.cpp
index e518756d2..fde36a49c 100644
--- a/src/video_core/vulkan_common/vulkan_device.cpp
+++ b/src/video_core/vulkan_common/vulkan_device.cpp
@@ -635,6 +635,12 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
             has_broken_cube_compatibility = true;
         }
     }
+    if (is_qualcomm) {
+        const u32 version = (properties.properties.driverVersion << 3) >> 3;
+        if (version < VK_MAKE_API_VERSION(0, 255, 615, 512)) {
+            has_broken_parallel_compiling = true;
+        }
+    }
     if (extensions.sampler_filter_minmax && is_amd) {
         // Disable ext_sampler_filter_minmax on AMD GCN4 and lower as it is broken.
         if (!features.shader_float16_int8.shaderFloat16) {
@@ -863,7 +869,8 @@ bool Device::ShouldBoostClocks() const {
         driver_id == VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA ||
         driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY || driver_id == VK_DRIVER_ID_MESA_TURNIP;
 
-    const bool is_steam_deck = vendor_id == 0x1002 && device_id == 0x163F;
+    const bool is_steam_deck = (vendor_id == 0x1002 && device_id == 0x163F) ||
+                               (vendor_id == 0x1002 && device_id == 0x1435);
 
     const bool is_debugging = this->HasDebuggingToolAttached();
 
diff --git a/src/video_core/vulkan_common/vulkan_device.h b/src/video_core/vulkan_common/vulkan_device.h
index b213ed7dd..4f3846345 100644
--- a/src/video_core/vulkan_common/vulkan_device.h
+++ b/src/video_core/vulkan_common/vulkan_device.h
@@ -102,6 +102,7 @@ VK_DEFINE_HANDLE(VmaAllocator)
     EXTENSION_NAME(VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME)                                   \
     EXTENSION_NAME(VK_EXT_EXTENDED_DYNAMIC_STATE_2_EXTENSION_NAME)                                 \
     EXTENSION_NAME(VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME)                                 \
+    EXTENSION_NAME(VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME)                                     \
     EXTENSION_NAME(VK_EXT_4444_FORMATS_EXTENSION_NAME)                                             \
     EXTENSION_NAME(VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME)                                       \
     EXTENSION_NAME(VK_EXT_ROBUSTNESS_2_EXTENSION_NAME)                                             \
@@ -599,6 +600,11 @@ public:
         return has_broken_cube_compatibility;
     }
 
+    /// Returns true if parallel shader compiling has issues with the current driver.
+    bool HasBrokenParallelShaderCompiling() const {
+        return has_broken_parallel_compiling;
+    }
+
     /// Returns the vendor name reported from Vulkan.
     std::string_view GetVendorName() const {
         return properties.driver.driverName;
@@ -663,6 +669,10 @@ public:
         return supports_conditional_barriers;
     }
 
+    bool SupportsMultiViewport() const {
+        return features2.features.multiViewport;
+    }
+
     [[nodiscard]] static constexpr bool CheckBrokenCompute(VkDriverId driver_id,
                                                            u32 driver_version) {
         if (driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
@@ -794,6 +804,7 @@ private:
     bool is_non_gpu{};                         ///< Is SoftwareRasterizer, FPGA, non-GPU device.
     bool has_broken_compute{};                 ///< Compute shaders can cause crashes
     bool has_broken_cube_compatibility{};      ///< Has broken cube compatibility bit
+    bool has_broken_parallel_compiling{};      ///< Has broken parallel shader compiling.
     bool has_renderdoc{};                      ///< Has RenderDoc attached
     bool has_nsight_graphics{};                ///< Has Nsight Graphics attached
     bool supports_d24_depth{};                 ///< Supports D24 depth buffers.
diff --git a/src/video_core/vulkan_common/vulkan_wrapper.h b/src/video_core/vulkan_common/vulkan_wrapper.h
index 0487cd3b6..a0c70797f 100644
--- a/src/video_core/vulkan_common/vulkan_wrapper.h
+++ b/src/video_core/vulkan_common/vulkan_wrapper.h
@@ -1101,6 +1101,10 @@ public:
         return &handle;
     }
 
+    VkCommandBuffer operator*() const noexcept {
+        return handle;
+    }
+
     void Begin(const VkCommandBufferBeginInfo& begin_info) const {
         Check(dld->vkBeginCommandBuffer(handle, &begin_info));
     }