Step back + keeping fixes from unified download + shared mem refactor

src/shader_recompiler/backend/spirv/emit_spirv_atomic.cpp

@@ -13,6 +13,7 @@
 namespace Shader::Backend::SPIRV {
 namespace {
 Id SharedPointer(EmitContext& ctx, Id offset, u32 index_offset = 0) {
+    offset = ctx.BoundSharedOffset(offset, 4 + index_offset * 4);
     const Id shift_id{ctx.Const(2U)};
     Id index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
     if (index_offset > 0) {
@@ -160,7 +161,8 @@ Id EmitSharedAtomicExchange32(EmitContext& ctx, Id offset, Id value) {
 Id EmitSharedAtomicExchange64(EmitContext& ctx, Id offset, Id value) {
     if (ctx.profile.support_shared_int64_atomics && ctx.uses_explicit_workgroup_layout) {
         const Id shift_id{ctx.Const(3U)};
-        const Id index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
+        const Id index{
+            ctx.OpShiftRightArithmetic(ctx.U32[1], ctx.BoundSharedOffset(offset, 8), shift_id)};
         const Id pointer{
             ctx.OpAccessChain(ctx.shared_u64, ctx.shared_memory_u64, ctx.u32_zero_value, index)};
         const auto [scope, semantics]{AtomicArgs(ctx)};

src/shader_recompiler/backend/spirv/emit_spirv_shared_memory.cpp

@@ -31,6 +31,7 @@ std::pair<Id, Id> ExtractArgs(EmitContext& ctx, Id offset, u32 mask, u32 count)
 } // Anonymous namespace
 
 Id EmitLoadSharedU8(EmitContext& ctx, Id offset) {
+    offset = ctx.BoundSharedOffset(offset, 1);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{
             ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
@@ -42,6 +43,7 @@ Id EmitLoadSharedU8(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedS8(EmitContext& ctx, Id offset) {
+    offset = ctx.BoundSharedOffset(offset, 1);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{
             ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
@@ -53,6 +55,7 @@ Id EmitLoadSharedS8(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedU16(EmitContext& ctx, Id offset) {
+    offset = ctx.BoundSharedOffset(offset, 2);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
         return ctx.OpUConvert(ctx.U32[1], ctx.OpLoad(ctx.U16, pointer));
@@ -63,6 +66,7 @@ Id EmitLoadSharedU16(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedS16(EmitContext& ctx, Id offset) {
+    offset = ctx.BoundSharedOffset(offset, 2);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
         return ctx.OpSConvert(ctx.U32[1], ctx.OpLoad(ctx.U16, pointer));
@@ -73,6 +77,7 @@ Id EmitLoadSharedS16(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedU32(EmitContext& ctx, Id offset) {
+    offset = ctx.BoundSharedOffset(offset, 4);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32, ctx.shared_memory_u32, offset, 2)};
         return ctx.OpLoad(ctx.U32[1], pointer);
@@ -82,6 +87,7 @@ Id EmitLoadSharedU32(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedU64(EmitContext& ctx, Id offset) {
+    offset = ctx.BoundSharedOffset(offset, 8);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32x2, ctx.shared_memory_u32x2, offset, 3)};
         return ctx.OpLoad(ctx.U32[2], pointer);
@@ -97,6 +103,7 @@ Id EmitLoadSharedU64(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
+    offset = ctx.BoundSharedOffset(offset, 16);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32x4, ctx.shared_memory_u32x4, offset, 4)};
         return ctx.OpLoad(ctx.U32[4], pointer);
@@ -113,6 +120,7 @@ Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
 }
 
 void EmitWriteSharedU8(EmitContext& ctx, Id offset, Id value) {
+    offset = ctx.BoundSharedOffset(offset, 1);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{
             ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
@@ -123,6 +131,7 @@ void EmitWriteSharedU8(EmitContext& ctx, Id offset, Id value) {
 }
 
 void EmitWriteSharedU16(EmitContext& ctx, Id offset, Id value) {
+    offset = ctx.BoundSharedOffset(offset, 2);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
         ctx.OpStore(pointer, ctx.OpUConvert(ctx.U16, value));
@@ -132,6 +141,7 @@ void EmitWriteSharedU16(EmitContext& ctx, Id offset, Id value) {
 }
 
 void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value) {
+    offset = ctx.BoundSharedOffset(offset, 4);
     Id pointer{};
     if (ctx.uses_explicit_workgroup_layout) {
         pointer = Pointer(ctx, ctx.shared_u32, ctx.shared_memory_u32, offset, 2);
@@ -144,6 +154,7 @@ void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value) {
 }
 
 void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
+    offset = ctx.BoundSharedOffset(offset, 8);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32x2, ctx.shared_memory_u32x2, offset, 3)};
         ctx.OpStore(pointer, value);
@@ -159,6 +170,7 @@ void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
 }
 
 void EmitWriteSharedU128(EmitContext& ctx, Id offset, Id value) {
+    offset = ctx.BoundSharedOffset(offset, 16);
     if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32x4, ctx.shared_memory_u32x4, offset, 4)};
         ctx.OpStore(pointer, value);

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -600,6 +600,16 @@ void EmitContext::DefineLocalMemory(const IR::Program& program) {
     }
 }
 
+Id EmitContext::BoundSharedOffset(Id offset, u32 access_bytes) {
+    if (shared_memory_declared_bytes == 0) {
+        return offset;
+    }
+    const u32 last_valid{shared_memory_declared_bytes > access_bytes
+                             ? shared_memory_declared_bytes - access_bytes
+                             : 0U};
+    return OpUMin(U32[1], offset, Const(last_valid));
+}
+
 void EmitContext::DefineSharedMemory(const IR::Program& program) {
     uses_explicit_workgroup_layout =
         profile.support_explicit_workgroup_layout &&
@@ -608,8 +618,15 @@ void EmitContext::DefineSharedMemory(const IR::Program& program) {
     if (program.shared_memory_size == 0) {
         return;
     }
+    const u32 device_limit{profile.max_shared_memory_size};
+    const u32 shared_memory_size{device_limit != 0 && program.shared_memory_size > device_limit
+                                     ? device_limit
+                                     : program.shared_memory_size};
+    if (shared_memory_size != program.shared_memory_size) {
+        shared_memory_declared_bytes = shared_memory_size;
+    }
     const auto make{[&](Id element_type, u32 element_size) {
-        const u32 num_elements{Common::DivCeil(program.shared_memory_size, element_size)};
+        const u32 num_elements{Common::DivCeil(shared_memory_size, element_size)};
         const Id array_type{TypeArray(element_type, Const(num_elements))};
         Decorate(array_type, spv::Decoration::ArrayStride, element_size);
 
@@ -644,7 +661,7 @@ void EmitContext::DefineSharedMemory(const IR::Program& program) {
         std::tie(shared_memory_u32x4, shared_u32x4, std::ignore) = make(U32[4], 16);
         return;
     }
-    const u32 num_elements{Common::DivCeil(program.shared_memory_size, 4U)};
+    const u32 num_elements{Common::DivCeil(shared_memory_size, 4U)};
     const Id type{TypeArray(U32[1], Const(num_elements))};
     shared_memory_u32_type = TypePointer(spv::StorageClass::Workgroup, type);
 

src/shader_recompiler/backend/spirv/spirv_emit_context.h

@@ -312,6 +312,8 @@ public:
     Id local_memory{};
 
     bool uses_explicit_workgroup_layout{};
+    u32 shared_memory_declared_bytes{};
+    [[nodiscard]] Id BoundSharedOffset(Id offset, u32 access_bytes);
     Id shared_memory_u8{};
     Id shared_memory_u16{};
     Id shared_memory_u32{};

src/shader_recompiler/profile.h

@@ -101,6 +101,9 @@ struct Profile {
 
     u32 gl_max_compute_smem_size{};
 
+    /// Largest workgroup shared memory allocation the device accepts, 0 when unconstrained
+    u32 max_shared_memory_size{};
+
     /// Maxwell and earlier nVidia architectures have broken robust support
     bool has_broken_robust{};
 

src/video_core/buffer_cache/buffer_cache.h

@@ -714,6 +714,7 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
             runtime.CopyToUnifiedMemory(queued.window, slot_buffers[queued.buffer_id], group_span);
         }
         if (!unified_copy_queue.empty()) {
+            runtime.FlushUnifiedMemoryCopies();
             runtime.UnifiedMemoryHostBarrier();
         }
     }
@@ -1741,9 +1742,6 @@ bool BufferCache<P>::SynchronizeBuffer(Buffer& buffer, DAddr device_addr, u32 si
 template <class P>
 void BufferCache<P>::UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
                                   std::span<BufferCopy> copies) {
-    if (TryUnifiedUploadMemory(buffer, copies)) {
-        return;
-    }
     if constexpr (USE_MEMORY_MAPS_FOR_UPLOADS) {
         MappedUploadMemory(buffer, total_size_bytes, copies);
     } else {
@@ -1864,6 +1862,7 @@ bool BufferCache<P>::TryUnifiedDownloadMemory([[maybe_unused]] Buffer& buffer,
             const std::span<const BufferCopy> group_span(groups[i].data(), groups[i].size());
             runtime.CopyToUnifiedMemory(window_ids[i], buffer, group_span);
         }
+        runtime.FlushUnifiedMemoryCopies();
         runtime.UnifiedMemoryHostBarrier();
         runtime.Finish();
         return true;
@@ -1872,34 +1871,6 @@ bool BufferCache<P>::TryUnifiedDownloadMemory([[maybe_unused]] Buffer& buffer,
     }
 }
 
-template <class P>
-bool BufferCache<P>::TryUnifiedUploadMemory([[maybe_unused]] Buffer& buffer,
-                                            [[maybe_unused]] std::span<BufferCopy> copies) {
-    if constexpr (USE_UNIFIED_UPLOADS) {
-        if (Settings::values.enable_gpu_buffer_readback.GetValue()) {
-            return false;
-        }
-        boost::container::small_vector<u64, 4> window_ids;
-        UnifiedWindowGroups groups;
-        for (const BufferCopy& copy : copies) {
-            if (!ResolveUnifiedWindows(buffer.CpuAddr() + copy.dst_offset, copy.dst_offset,
-                                       copy.size, window_ids, groups)) {
-                return false;
-            }
-        }
-        runtime.UnifiedMemoryUploadBarrier();
-        runtime.PreCopyBarrier();
-        for (size_t i = 0; i < window_ids.size(); ++i) {
-            const std::span<const BufferCopy> group_span(groups[i].data(), groups[i].size());
-            runtime.CopyFromUnifiedMemory(window_ids[i], buffer, group_span);
-        }
-        runtime.PostCopyBarrier();
-        return true;
-    } else {
-        return false;
-    }
-}
-
 template <class P>
 void BufferCache<P>::MappedUploadMemory([[maybe_unused]] Buffer& buffer,
                                         [[maybe_unused]] u64 total_size_bytes,

src/video_core/buffer_cache/buffer_cache_base.h

@@ -181,7 +181,6 @@ class BufferCache : public VideoCommon::ChannelSetupCaches<BufferCacheChannelInf
     static constexpr bool SEPARATE_IMAGE_BUFFERS_BINDINGS = P::SEPARATE_IMAGE_BUFFER_BINDINGS;
     static constexpr bool USE_MEMORY_MAPS_FOR_UPLOADS = P::USE_MEMORY_MAPS_FOR_UPLOADS;
     static constexpr bool USE_UNIFIED_MEMORY = P::USE_UNIFIED_MEMORY;
-    static constexpr bool USE_UNIFIED_UPLOADS = P::USE_UNIFIED_UPLOADS;
 
 #ifdef YUZU_LEGACY
     static constexpr s64 TARGET_THRESHOLD = 3_GiB;
@@ -447,8 +446,6 @@ private:
 
     bool TryUnifiedDownloadMemory(Buffer& buffer, std::span<BufferCopy> copies);
 
-    bool TryUnifiedUploadMemory(Buffer& buffer, std::span<BufferCopy> copies);
-
     using UnifiedWindowGroups =
         boost::container::small_vector<boost::container::small_vector<BufferCopy, 16>, 4>;
 

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -262,7 +262,6 @@ struct BufferCacheParams {
     // TODO: Investigate why OpenGL seems to perform worse with persistently mapped buffer uploads
     static constexpr bool USE_MEMORY_MAPS_FOR_UPLOADS = false;
     static constexpr bool USE_UNIFIED_MEMORY = false;
-    static constexpr bool USE_UNIFIED_UPLOADS = false;
 };
 
 using BufferCache = VideoCommon::BufferCache<BufferCacheParams>;

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -237,6 +237,7 @@ ShaderCache::ShaderCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
           .has_gl_bool_ref_bug = device.HasBoolRefBug(),
           .ignore_nan_fp_comparisons = true,
           .gl_max_compute_smem_size = device.GetMaxComputeSharedMemorySize(),
+          .max_shared_memory_size = device.GetMaxComputeSharedMemorySize(),
           .min_ssbo_alignment = device.GetShaderStorageBufferAlignment(),
           // Use the host limit, but never more than the guest can produce. Maxwell exposes 8 clip
           // distances and the SPIR-V output array is sized for at most 8, so clamping here keeps a

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -8,6 +8,7 @@
 #include <array>
 #include <cstring>
 #include <span>
+#include <utility>
 #include <vector>
 
 #include "video_core/buffer_cache/buffer_cache_base.h"
@@ -32,12 +33,30 @@ VkBufferCopy MakeBufferCopy(const VideoCommon::BufferCopy& copy) {
     };
 }
 
-VkBufferCopy MakeUnifiedUploadCopy(const VideoCommon::BufferCopy& copy) {
-    return VkBufferCopy{
-        .srcOffset = copy.dst_offset,
-        .dstOffset = copy.src_offset,
-        .size = copy.size,
-    };
+constexpr size_t MAX_WINDOW_BARRIER_RANGES = 8;
+
+using WindowRange = std::pair<VkDeviceSize, VkDeviceSize>;
+using WindowRanges = boost::container::small_vector<WindowRange, MAX_WINDOW_BARRIER_RANGES>;
+
+void CoalesceWindowRanges(WindowRanges& ranges) {
+    if (ranges.size() < 2) {
+        return;
+    }
+    std::sort(ranges.begin(), ranges.end());
+    size_t merged = 0;
+    for (size_t index = 1; index < ranges.size(); ++index) {
+        if (ranges[index].first <= ranges[merged].second) {
+            ranges[merged].second = (std::max)(ranges[merged].second, ranges[index].second);
+        } else {
+            ranges[++merged] = ranges[index];
+        }
+    }
+    ranges.resize(merged + 1);
+    if (ranges.size() > MAX_WINDOW_BARRIER_RANGES) {
+        const WindowRange bounding{ranges.front().first, ranges.back().second};
+        ranges.clear();
+        ranges.push_back(bounding);
+    }
 }
 
 VkIndexType IndexTypeFromNumElements(const Device& device, u32 num_elements) {
@@ -388,138 +407,110 @@ void BufferCacheRuntime::CopyToUnifiedMemory(
     size_t window_index, VkBuffer src_buffer,
     std::span<const VideoCommon::BufferCopy> copies) {
     if (!unified_memory || src_buffer == VK_NULL_HANDLE || copies.empty() ||
-        window_index >= unified_memory->GetWindowCount()) {
+        window_index >= unified_memory->GetWindowCount() ||
+        unified_memory->GetWindowBuffer(window_index) == VK_NULL_HANDLE) {
         return;
     }
-    const VkBuffer dst_buffer = unified_memory->GetWindowBuffer(window_index);
-    if (dst_buffer == VK_NULL_HANDLE) {
-        return;
-    }
-    VkDeviceSize covered_begin = std::numeric_limits<VkDeviceSize>::max();
-    VkDeviceSize covered_end = 0;
-    for (const VideoCommon::BufferCopy& copy : copies) {
-        covered_begin = (std::min)(covered_begin, static_cast<VkDeviceSize>(copy.dst_offset));
-        covered_end = (std::max)(covered_end,
-                                 static_cast<VkDeviceSize>(copy.dst_offset + copy.size));
-    }
-
-    boost::container::small_vector<VkBufferCopy, 8> vk_copies(copies.size());
-    std::ranges::transform(copies, vk_copies.begin(), MakeBufferCopy);
-
-    const bool foreign = unified_memory->NeedsForeignOwnershipTransfer();
-    const u32 queue_family = device.GetGraphicsFamily();
-
-    scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([src_buffer, dst_buffer, vk_copies, foreign, queue_family, covered_begin,
-                      covered_end](vk::CommandBuffer cmdbuf) {
-        if (foreign) {
-            const VkBufferMemoryBarrier acquire{
-                .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-                .pNext = nullptr,
-                .srcAccessMask = 0,
-                .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
-                .dstQueueFamilyIndex = queue_family,
-                .buffer = dst_buffer,
-                .offset = covered_begin,
-                .size = covered_end - covered_begin,
-            };
-            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
-                                   VK_PIPELINE_STAGE_TRANSFER_BIT, 0, acquire);
-        }
-        cmdbuf.CopyBuffer(src_buffer, dst_buffer, VideoCommon::FixSmallVectorADL(vk_copies));
-        if (foreign) {
-            const VkBufferMemoryBarrier release{
-                .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-                .pNext = nullptr,
-                .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-                .dstAccessMask = 0,
-                .srcQueueFamilyIndex = queue_family,
-                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
-                .buffer = dst_buffer,
-                .offset = covered_begin,
-                .size = covered_end - covered_begin,
-            };
-            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT,
-                                   VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, release);
-        }
-    });
+    PendingUnifiedCopy& pending = pending_unified_copies.emplace_back();
+    pending.window = window_index;
+    pending.buffer = src_buffer;
+    pending.copies.resize(copies.size());
+    std::ranges::transform(copies, pending.copies.begin(), MakeBufferCopy);
 }
 
-void BufferCacheRuntime::CopyFromUnifiedMemory(
-    size_t window_index, VkBuffer dst_buffer,
-    std::span<const VideoCommon::BufferCopy> copies) {
-    if (!unified_memory || dst_buffer == VK_NULL_HANDLE || copies.empty() ||
-        window_index >= unified_memory->GetWindowCount()) {
+void BufferCacheRuntime::FlushUnifiedMemoryCopies() {
+    if (pending_unified_copies.empty()) {
         return;
     }
-    const VkBuffer src_buffer = unified_memory->GetWindowBuffer(window_index);
-    if (src_buffer == VK_NULL_HANDLE) {
-        return;
-    }
-    VkDeviceSize covered_begin = std::numeric_limits<VkDeviceSize>::max();
-    VkDeviceSize covered_end = 0;
-    for (const VideoCommon::BufferCopy& copy : copies) {
-        covered_begin = (std::min)(covered_begin, static_cast<VkDeviceSize>(copy.dst_offset));
-        covered_end = (std::max)(covered_end,
-                                 static_cast<VkDeviceSize>(copy.dst_offset + copy.size));
-    }
-
-    boost::container::small_vector<VkBufferCopy, 8> vk_copies(copies.size());
-    std::ranges::transform(copies, vk_copies.begin(), MakeUnifiedUploadCopy);
-
-    const bool foreign = unified_memory->NeedsForeignOwnershipTransfer();
-    const u32 queue_family = device.GetGraphicsFamily();
-
-    scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([src_buffer, dst_buffer, vk_copies, foreign, queue_family, covered_begin,
-                      covered_end](vk::CommandBuffer cmdbuf) {
-        if (foreign) {
-            const VkBufferMemoryBarrier acquire{
-                .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-                .pNext = nullptr,
-                .srcAccessMask = 0,
-                .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
-                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
-                .dstQueueFamilyIndex = queue_family,
-                .buffer = src_buffer,
-                .offset = covered_begin,
-                .size = covered_end - covered_begin,
-            };
-            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
-                                   VK_PIPELINE_STAGE_TRANSFER_BIT, 0, acquire);
-        }
-        cmdbuf.CopyBuffer(src_buffer, dst_buffer, VideoCommon::FixSmallVectorADL(vk_copies));
-        if (foreign) {
-            const VkBufferMemoryBarrier release{
-                .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
-                .pNext = nullptr,
-                .srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
-                .dstAccessMask = 0,
-                .srcQueueFamilyIndex = queue_family,
-                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
-                .buffer = src_buffer,
-                .offset = covered_begin,
-                .size = covered_end - covered_begin,
-            };
-            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT,
-                                   VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, release);
-        }
-    });
-}
-
-void BufferCacheRuntime::UnifiedMemoryUploadBarrier() {
-    static constexpr VkMemoryBarrier HOST_WRITE_BARRIER{
-        .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
-        .pNext = nullptr,
-        .srcAccessMask = VK_ACCESS_HOST_WRITE_BIT,
-        .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
+    struct UnifiedCopyCommand {
+        VkBuffer buffer;
+        boost::container::small_vector<VkBufferCopy, 8> copies;
     };
-    scheduler.RequestOutsideRenderPassOperationContext();
-    scheduler.Record([](vk::CommandBuffer cmdbuf) {
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_HOST_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
-                               HOST_WRITE_BARRIER);
-    });
+
+    std::stable_sort(pending_unified_copies.begin(), pending_unified_copies.end(),
+                     [](const PendingUnifiedCopy& lhs, const PendingUnifiedCopy& rhs) {
+                         return lhs.window < rhs.window;
+                     });
+
+    const bool foreign = unified_memory->NeedsForeignOwnershipTransfer();
+    const u32 queue_family = device.GetGraphicsFamily();
+
+    size_t group_begin = 0;
+    while (group_begin < pending_unified_copies.size()) {
+        const size_t window = pending_unified_copies[group_begin].window;
+        size_t group_end = group_begin;
+        while (group_end < pending_unified_copies.size() &&
+               pending_unified_copies[group_end].window == window) {
+            ++group_end;
+        }
+        const VkBuffer window_buffer = unified_memory->GetWindowBuffer(window);
+
+        WindowRanges ranges;
+        for (size_t index = group_begin; index < group_end; ++index) {
+            for (const VkBufferCopy& copy : pending_unified_copies[index].copies) {
+                ranges.emplace_back(copy.dstOffset, copy.dstOffset + copy.size);
+            }
+        }
+        CoalesceWindowRanges(ranges);
+
+        boost::container::small_vector<VkBufferMemoryBarrier, MAX_WINDOW_BARRIER_RANGES> acquire;
+        boost::container::small_vector<VkBufferMemoryBarrier, MAX_WINDOW_BARRIER_RANGES> release;
+        if (foreign) {
+            for (const WindowRange& range : ranges) {
+                acquire.push_back(VkBufferMemoryBarrier{
+                    .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
+                    .pNext = nullptr,
+                    .srcAccessMask = 0,
+                    .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+                    .srcQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
+                    .dstQueueFamilyIndex = queue_family,
+                    .buffer = window_buffer,
+                    .offset = range.first,
+                    .size = range.second - range.first,
+                });
+                release.push_back(VkBufferMemoryBarrier{
+                    .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
+                    .pNext = nullptr,
+                    .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+                    .dstAccessMask = 0,
+                    .srcQueueFamilyIndex = queue_family,
+                    .dstQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
+                    .buffer = window_buffer,
+                    .offset = range.first,
+                    .size = range.second - range.first,
+                });
+            }
+        }
+
+        boost::container::small_vector<UnifiedCopyCommand, 4> commands;
+        commands.reserve(group_end - group_begin);
+        for (size_t index = group_begin; index < group_end; ++index) {
+            PendingUnifiedCopy& pending = pending_unified_copies[index];
+            commands.push_back(UnifiedCopyCommand{pending.buffer, std::move(pending.copies)});
+        }
+
+        scheduler.RequestOutsideRenderPassOperationContext();
+        scheduler.Record([window_buffer, acquire = std::move(acquire), release = std::move(release),
+                          commands = std::move(commands)](vk::CommandBuffer cmdbuf) {
+            if (!acquire.empty()) {
+                cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                                       VK_PIPELINE_STAGE_TRANSFER_BIT, 0, {},
+                                       VideoCommon::FixSmallVectorADL(acquire), {});
+            }
+            for (const UnifiedCopyCommand& command : commands) {
+                cmdbuf.CopyBuffer(command.buffer, window_buffer,
+                                  VideoCommon::FixSmallVectorADL(command.copies));
+            }
+            if (!release.empty()) {
+                cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT,
+                                       VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, {},
+                                       VideoCommon::FixSmallVectorADL(release), {});
+            }
+        });
+
+        group_begin = group_end;
+    }
+    pending_unified_copies.clear();
 }
 
 void BufferCacheRuntime::UnifiedMemoryHostBarrier() {
@@ -573,6 +564,7 @@ u32 BufferCacheRuntime::GetStorageBufferAlignment() const {
 }
 
 void BufferCacheRuntime::TickFrame(Common::SlotVector<Buffer>& slot_buffers) noexcept {
+    FlushUnifiedMemoryCopies();
     for (auto it = slot_buffers.begin(); it != slot_buffers.end(); it++) {
         if (scheduler.IsFree(it->LastUsageTick())) {
             it->ResetUsageTracking();

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -10,6 +10,8 @@
 #include <memory>
 #include <span>
 
+#include <boost/container/small_vector.hpp>
+
 #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"
@@ -122,13 +124,10 @@ public:
     void CopyToUnifiedMemory(size_t window_index, VkBuffer src_buffer,
                              std::span<const VideoCommon::BufferCopy> copies);
 
-    void CopyFromUnifiedMemory(size_t window_index, VkBuffer dst_buffer,
-                               std::span<const VideoCommon::BufferCopy> copies);
+    void FlushUnifiedMemoryCopies();
 
     void UnifiedMemoryHostBarrier();
 
-    void UnifiedMemoryUploadBarrier();
-
     u64 CurrentTick();
 
     u64 KnownGpuTick();
@@ -212,6 +211,12 @@ public:
     }
 
 private:
+    struct PendingUnifiedCopy {
+        size_t window;
+        VkBuffer buffer;
+        boost::container::small_vector<VkBufferCopy, 8> copies;
+    };
+
     void BindBuffer(const Buffer& buffer, u32 offset, u32 size) {
         const VkBuffer handle = buffer.Handle();
         if (handle == VK_NULL_HANDLE) {
@@ -237,6 +242,7 @@ private:
 
     vk::Buffer null_buffer;
     std::unique_ptr<HostMemoryImport> unified_memory;
+    boost::container::small_vector<PendingUnifiedCopy, 8> pending_unified_copies;
 
     std::unique_ptr<Uint8Pass> uint8_pass;
     QuadIndexedPass quad_index_pass;
@@ -260,7 +266,6 @@ struct BufferCacheParams {
     static constexpr bool SEPARATE_IMAGE_BUFFER_BINDINGS = false;
     static constexpr bool USE_MEMORY_MAPS_FOR_UPLOADS = true;
     static constexpr bool USE_UNIFIED_MEMORY = true;
-    static constexpr bool USE_UNIFIED_UPLOADS = true;
 };
 
 using BufferCache = VideoCommon::BufferCache<BufferCacheParams>;

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -446,6 +446,7 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .has_broken_fp16_float_controls = driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY,
         .ignore_nan_fp_comparisons = false,
         .has_broken_spirv_subgroup_mask_vector_extract_dynamic = false,
+        .max_shared_memory_size = device.GetMaxComputeSharedMemorySize(),
         .has_broken_robust =
             device.IsNvidia() && device.GetNvidiaArch() <= NvidiaArchitecture::Arch_Pascal,
         .min_ssbo_alignment = device.GetStorageBufferAlignment(),
@@ -921,19 +922,6 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline(
     }
 
     auto program{TranslateProgram(pools.inst, pools.block, env, cfg, host_info)};
-    const VkDriverIdKHR driver_id = device.GetDriverID();
-    const bool needs_shared_mem_clamp =
-        driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY ||
-        driver_id == VK_DRIVER_ID_ARM_PROPRIETARY;
-    const u32 max_shared_memory = device.GetMaxComputeSharedMemorySize();
-    if (needs_shared_mem_clamp && program.shared_memory_size > max_shared_memory) {
-        LOG_WARNING(Render_Vulkan,
-                    "Compute shader {:#016x} requests {}KB shared memory but device max is {}KB - clamping",
-                    key.unique_hash,
-                    program.shared_memory_size / 1024,
-                    max_shared_memory / 1024);
-        program.shared_memory_size = max_shared_memory;
-    }
     const std::vector<u32> code{EmitSPIRV(profile, program)};
     device.SaveShader(code);
     vk::ShaderModule spv_module{BuildShader(device, code)};

src/video_core/vulkan_common/vulkan_memory_allocator.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
 // SPDX-FileCopyrightText: Copyright 2019 yuzu Emulator Project