3 files changed, 199 insertions, 6 deletions

diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index cd176dec27..40369b16ef 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -10,11 +10,13 @@
 #include <memory>
 #include <numeric>
 #include <stdexcept>
+#include <utility>
 #include <vector>
 #include <fmt/ranges.h>
 
 #include "common/range_sets.inc"
 #include "video_core/buffer_cache/buffer_cache_base.h"
+#include "video_core/buffer_cache/uniform_stream_reuse.h"
 #include "video_core/guest_memory.h"
 #include "video_core/host1x/gpu_device_memory_manager.h"
 #include "video_core/nx2.h"
@@ -28,6 +30,9 @@ using Core::DEVICE_PAGESIZE;
 template <class P>
 BufferCache<P>::BufferCache(Tegra::MaxwellDeviceMemoryManager& device_memory_, Runtime& runtime_)
     : runtime{runtime_}, device_memory{device_memory_}, memory_tracker{device_memory} {
+    // A new runtime has a new scheduler (its ticks restart) and a new stream buffer: whatever the
+    // stream uniform reuse remembers refers to the old ones. Before the early return below.
+    VideoCommon::ResetUniformStreamReuse<Runtime>();
     // Ensure the first slot is used for the null buffer
     void(slot_buffers.insert(runtime, NullBufferParams{}));
     gpu_modified_ranges.Clear();
@@ -180,6 +185,7 @@ void BufferCache<P>::WriteMemory(DAddr device_addr, u64 size) {
         gpu_modified_ranges.Subtract(device_addr, size);
     }
     memory_tracker.MarkRegionAsCpuModified(device_addr, size);
+    MarkVertexBuffersStale(device_addr, size);
 }
 
 template <class P>
@@ -230,8 +236,17 @@ std::optional<VideoCore::RasterizerDownloadArea> BufferCache<P>::GetFlushArea(DA
     };
     area->preemtive = !IsRegionGpuModified(device_addr_start_aligned,
                                            device_addr_end_aligned - device_addr_start_aligned);
-    memory_tracker.MarkRegionAsPreflushable(device_addr_start_aligned,
-                                            device_addr_end_aligned - device_addr_start_aligned);
+    // Mark only data that the GPU wrote. A mark stays for the session and sends later GPU writes
+    // in the range to a download at each fence.
+    if (!area->preemtive) {
+        memory_tracker.MarkRegionAsPreflushable(device_addr_start_aligned,
+                                                device_addr_end_aligned - device_addr_start_aligned);
+    } else {
+        // The core keeps this area and has no mark. The next GPU write here must arm the range
+        // again, which ends the area and sets the trap.
+        read_trapped_ranges.Subtract(device_addr_start_aligned,
+                                     device_addr_end_aligned - device_addr_start_aligned);
+    }
     return area;
 }
 
@@ -265,6 +280,7 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     }
 
     ClearDownload(*cpu_dest_address, amount);
+    MarkVertexBuffersStale(*cpu_dest_address, amount);
 
     BufferId buffer_a;
     BufferId buffer_b;
@@ -408,7 +424,8 @@ void BufferCache<P>::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr
 template <class P>
 void BufferCache<P>::DisableGraphicsUniformBuffer(size_t stage, u32 index) {
     auto& current = channel_state->uniform_buffers[stage][index];
-    if (current.device_addr == 0 && current.size == 0 && !current.buffer_id) [[likely]] {
+    if (current.device_addr == 0 && current.size == 0 && current.buffer_id == NULL_BUFFER_ID)
+        [[likely]] {
         return;
     }
     current = NULL_BINDING;
@@ -893,7 +910,39 @@ void BufferCache<P>::UpdateVertexBufferSlot(u32 index, const Binding& binding) {
 }
 
 template <class P>
+void BufferCache<P>::MarkVertexBuffersStale(DAddr device_addr, u64 size) {
+    if constexpr (requires { runtime.IsVertexRepacked(u32{}); }) {
+        u32 mask = enabled_vertex_buffers_mask;
+        while (mask != 0) {
+            const u32 index = std::countr_zero(mask);
+            mask &= mask - 1;
+            const Binding& binding = v_buffer[index];
+            if (device_addr < binding.device_addr + binding.size &&
+                binding.device_addr < device_addr + size) {
+                stale_vertex_buffers_mask |= 1U << index;
+            }
+        }
+    }
+}
+
+template <class P>
+void BufferCache<P>::RefreshRepackedVertexBuffers() {
+    if constexpr (requires { runtime.IsVertexRepacked(u32{}); }) {
+        u32 mask = std::exchange(stale_vertex_buffers_mask, 0) & enabled_vertex_buffers_mask;
+        auto& flags = maxwell3d->dirty.flags;
+        while (mask != 0) {
+            const u32 index = std::countr_zero(mask);
+            mask &= mask - 1;
+            if (runtime.IsVertexRepacked(index)) {
+                flags[Dirty::VertexBuffer0 + index] = true;
+            }
+        }
+    }
+}
+
+template <class P>
 void BufferCache<P>::BindHostVertexBuffers() {
+    RefreshRepackedVertexBuffers();
 
 #ifdef __ANDROID__
     const bool use_optimized_vertex_buffers = Settings::values.use_optimized_vertex_buffers.GetValue();
@@ -1078,8 +1127,20 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
         channel_state->fast_bound_uniform_buffers[stage] |= 1u << binding_index;
         channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
         // Stream buffer path to avoid stalling on non-Nvidia drivers or Vulkan
-        const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
-        device_memory.ReadBlockUnsafe(device_addr, span.data(), size);
+        bool streamed = false;
+        if constexpr (!IS_OPENGL && VideoCommon::StreamUniformReusable<Runtime>) {
+            // The same bytes as this slot's previous bind, in the same command buffer: bind that
+            // stream allocation again instead of allocating and copying (uniform_stream_reuse.h).
+            // EDEN_NX2_UBOREUSE=0 turns it off.
+            if (VideoCommon::UniformStreamReuseEnabled()) {
+                streamed = VideoCommon::BindSmallUniformReusing(
+                    runtime, device_memory, stage, index, binding_index, device_addr, size);
+            }
+        }
+        if (!streamed) {
+            const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
+            device_memory.ReadBlockUnsafe(device_addr, span.data(), size);
+        }
         return;
     }
     // Classic cached path
@@ -1655,6 +1716,7 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, DAddr device_addr, u3
         }
     }
     memory_tracker.MarkRegionAsGpuModified(device_addr, size);
+    MarkVertexBuffersStale(device_addr, size);
     gpu_modified_ranges.Add(device_addr, size);
     if (DownloadsAtFence(device_addr, size)) {
         uncommitted_gpu_modified_ranges.Add(device_addr, size);
@@ -2205,6 +2267,7 @@ void BufferCache<P>::InlineMemoryImplementation(DAddr dest_address, size_t copy_
                                                 std::span<const u8> inlined_buffer) {
     ClearDownload(dest_address, copy_size);
     gpu_modified_ranges.Subtract(dest_address, copy_size);
+    MarkVertexBuffersStale(dest_address, copy_size);
 
     BufferId buffer_id = FindBuffer(dest_address, static_cast<u32>(copy_size), false);
     auto& buffer = slot_buffers[buffer_id];
diff --git a/src/video_core/buffer_cache/buffer_cache_base.h b/src/video_core/buffer_cache/buffer_cache_base.h
index 1f0c6cf10f..86c05c8f22 100644
--- a/src/video_core/buffer_cache/buffer_cache_base.h
+++ b/src/video_core/buffer_cache/buffer_cache_base.h
@@ -59,7 +59,8 @@ constexpr u32 NUM_VERTEX_BUFFERS = 32;
 constexpr u32 NUM_TRANSFORM_FEEDBACK_BUFFERS = 4;
 constexpr u32 NUM_GRAPHICS_UNIFORM_BUFFERS = 18;
 constexpr u32 NUM_COMPUTE_UNIFORM_BUFFERS = 8;
-constexpr u32 NUM_STORAGE_BUFFERS = 16;
+// Native compute shaders bind up to 18 global memory ranges; the enable masks are 32 bits wide.
+constexpr u32 NUM_STORAGE_BUFFERS = 32;
 constexpr u32 NUM_TEXTURE_BUFFERS = 32;
 constexpr u32 NUM_STAGES = 5;
 
@@ -525,6 +526,13 @@ private:
 
     void UpdateVertexBufferSlot(u32 index, const Binding& binding);
 
+    /// Flags the vertex buffers overlapping a written range so host copies derived from them are
+    /// rebuilt.
+    void MarkVertexBuffersStale(DAddr device_addr, u64 size);
+
+    /// Re-dirties repacked vertex buffers whose source data was written since they were bound.
+    void RefreshRepackedVertexBuffers();
+
     void ClearDownload(DAddr base_addr, u64 size);
 
     void InlineMemoryImplementation(DAddr dest_address, size_t copy_size,
@@ -545,6 +553,7 @@ private:
     u32 last_index_count = 0;
 
     u32 enabled_vertex_buffers_mask = 0;
+    u32 stale_vertex_buffers_mask = 0;
     u64 vertex_buffers_serial = 0;
     std::array<Binding, 32> v_buffer{};
 
diff --git a/src/video_core/buffer_cache/uniform_stream_reuse.h b/src/video_core/buffer_cache/uniform_stream_reuse.h
new file mode 100644
index 0000000000..7ba31c4717
--- /dev/null
+++ b/src/video_core/buffer_cache/uniform_stream_reuse.h
@@ -0,0 +1,121 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <array>
+#include <concepts>
+#include <cstddef>
+#include <cstdlib>
+#include <cstring>
+#include <span>
+#include <type_traits>
+
+#include "common/common_types.h"
+
+namespace VideoCommon {
+
+/// What a buffer cache runtime offers for stream uniform reuse.
+///  - StreamUniformRef: a small trivially copyable description of one stream allocation (the
+///    host buffer, its offset and whatever else the runtime needs to add it to a descriptor
+///    payload again, e.g. the device address of the descriptor buffer path).
+///  - LastStreamUniformRef(): the allocation made by the latest BindMappedUniformBuffer call.
+///  - RebindStreamUniform(ref, size): adds `ref` to the descriptor payload exactly as
+///    BindMappedUniformBuffer added the allocation the first time.
+template <class R>
+concept StreamUniformReusable = requires(R& runtime, u32 size) {
+    typename R::StreamUniformRef;
+    requires std::is_trivially_copyable_v<typename R::StreamUniformRef>;
+    { runtime.LastStreamUniformRef() } -> std::same_as<typename R::StreamUniformRef>;
+    runtime.RebindStreamUniform(runtime.LastStreamUniformRef(), size);
+    { runtime.CurrentTick() } -> std::convertible_to<u64>;
+};
+
+/// EDEN_NX2_UBOREUSE=0 turns the reuse off (every small uniform buffer is copied, as before).
+[[nodiscard]] inline bool UniformStreamReuseEnabled() noexcept {
+    static const bool enabled = [] {
+        const char* const value = std::getenv("EDEN_NX2_UBOREUSE");
+        return value == nullptr || value[0] != '0';
+    }();
+    return enabled;
+}
+
+template <class Ref>
+class UniformStreamReuse {
+public:
+    static constexpr u32 MaxBytes = 4096; // BufferCache's DEFAULT_SKIP_CACHE_SIZE
+    static constexpr size_t NumStages = 5;
+    static constexpr size_t NumSlots = 18; // Maxwell constant buffers per stage
+
+    struct Slot {
+        Ref ref{};
+        u64 tick{};
+        u32 size{}; // 0: nothing streamed yet
+        std::array<u8, MaxBytes> bytes{};
+    };
+
+    void Reset() noexcept {
+        for (auto& stage : slots) {
+            for (Slot& slot : stage) {
+                slot.size = 0;
+                slot.tick = 0;
+            }
+        }
+    }
+
+    std::array<std::array<Slot, NumSlots>, NumStages> slots{};
+    std::array<u8, MaxBytes> scratch{};
+};
+
+template <class Ref>
+[[nodiscard]] inline UniformStreamReuse<Ref>& UniformStreamReuseState() noexcept {
+    static UniformStreamReuse<Ref> state;
+    return state;
+}
+
+/// Forgets every slot. Call it when a buffer cache (and so its runtime, scheduler and stream
+/// buffer) is created: ticks restart with a new scheduler and the old references are dead.
+template <class Runtime>
+inline void ResetUniformStreamReuse() noexcept {
+    if constexpr (StreamUniformReusable<Runtime>) {
+        UniformStreamReuseState<typename Runtime::StreamUniformRef>().Reset();
+    }
+}
+
+/// Binds the small uniform buffer at `device_addr` for (stage, cbuf `index`) through the stream
+/// buffer, reusing the slot's previous allocation when the bytes and the command buffer are the
+/// same. Returns false (nothing bound) when the buffer does not qualify, so the caller binds it
+/// the classic way.
+template <class Runtime, class DeviceMemory>
+[[nodiscard]] bool BindSmallUniformReusing(Runtime& runtime, DeviceMemory& device_memory,
+                                           size_t stage, u32 index, u32 binding_index,
+                                           u64 device_addr, u32 size) {
+    using Reuse = UniformStreamReuse<typename Runtime::StreamUniformRef>;
+    if (size == 0 || size > Reuse::MaxBytes || stage >= Reuse::NumStages ||
+        index >= Reuse::NumSlots) {
+        return false;
+    }
+    Reuse& reuse = UniformStreamReuseState<typename Runtime::StreamUniformRef>();
+    typename Reuse::Slot& slot = reuse.slots[stage][index];
+    u8* const scratch = reuse.scratch.data();
+    device_memory.ReadBlockUnsafe(device_addr, scratch, size);
+
+    const bool same_bytes =
+        slot.size == size && std::memcmp(slot.bytes.data(), scratch, size) == 0;
+    if (same_bytes && slot.tick == runtime.CurrentTick()) {
+        runtime.RebindStreamUniform(slot.ref, size);
+        return true;
+    }
+    const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
+    std::memcpy(span.data(), scratch, size);
+    slot.ref = runtime.LastStreamUniformRef();
+    // After the allocation: the tick the stream region is tagged with is the one current now.
+    slot.tick = runtime.CurrentTick();
+    if (!same_bytes) {
+        slot.size = size;
+        std::memcpy(slot.bytes.data(), scratch, size);
+    }
+    return true;
+}
+
+} // namespace VideoCommon