Port UMA part 3

src/video_core/buffer_cache/buffer_cache.h

@@ -327,6 +327,7 @@ std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainCPUBuffer(
     default:
         break;
     }
+    buffer.MarkUsage(buffer.Offset(device_addr), size);
 
     switch (post_op) {
     case ObtainBufferOperation::MarkAsWritten:
@@ -394,8 +395,9 @@ void BufferCache<P>::CommitDrawWrites() {
         }
         const u64 pass = runtime.RenderPassSerial();
         for (const DrawWrite& write : draw_writes) {
-            slot_buffers[write.buffer_id].MarkDrawWrite(pass, draw_wfi, write.device_addr,
-                                                        write.size, write.feedback);
+            Buffer& buffer = slot_buffers[write.buffer_id];
+            buffer.setWriteTick(runtime.CurrentTick());
+            buffer.MarkDrawWrite(pass, draw_wfi, write.device_addr, write.size, write.feedback);
         }
         runtime.MarkRenderPassWrites();
     }
@@ -956,6 +958,7 @@ void BufferCache<P>::BindHostDrawIndirectBuffers() {
         Buffer& buffer = slot_buffers[binding.buffer_id];
         TouchBuffer(buffer, binding.buffer_id);
         SynchronizeBuffer(buffer, binding.device_addr, binding.size);
+        buffer.MarkUsage(buffer.Offset(binding.device_addr), binding.size);
     };
     if (current_draw_indirect->include_count) {
         bind_buffer(channel_state->count_buffer_binding);
@@ -1989,6 +1992,16 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, DAddr device_addr, u64
     }
 
     if constexpr (USE_MEMORY_MAPS) {
+        if constexpr (requires { runtime.DirectDownloadSpan(buffer); }) {
+            const std::span<const u8> direct = runtime.DirectDownloadSpan(buffer);
+            if (!direct.empty()) {
+                for (const BufferCopy& copy : copies) {
+                    device_memory.WriteBlockUnsafe(buffer.CpuAddr() + copy.src_offset,
+                                                   direct.data() + copy.src_offset, copy.size);
+                }
+                return;
+            }
+        }
         auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
         const u8* const mapped_memory = download_staging.mapped_span.data();
         const std::span<BufferCopy> copies_span(copies.data(), copies.data() + copies.size());

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -511,6 +511,14 @@ bool BufferCacheRuntime::CanReorderUpload(const Buffer& buffer,
     return can_use_upload_cmdbuf;
 }
 
+std::span<const u8> BufferCacheRuntime::DirectDownloadSpan(Buffer& buffer) {
+    const std::span<u8> mapping = buffer.CoherentMapping();
+    if (!mapping.empty()) {
+        scheduler.Wait((std::max)(buffer.getWriteTick(), buffer.LastUploadTick()));
+    }
+    return mapping;
+}
+
 std::span<u8> BufferCacheRuntime::DirectUploadSpan(
     Buffer& buffer, std::span<const VideoCommon::BufferCopy> copies) {
     const std::span<u8> mapping = buffer.CoherentMapping();

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -149,6 +149,8 @@ public:
 
     bool CanReorderUpload(const Buffer& buffer, std::span<const VideoCommon::BufferCopy> copies);
 
+    [[nodiscard]] std::span<const u8> DirectDownloadSpan(Buffer& buffer);
+
     [[nodiscard]] std::span<u8> DirectUploadSpan(Buffer& buffer,
                                                  std::span<const VideoCommon::BufferCopy> copies);
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -32,6 +32,15 @@ namespace Vulkan {
 namespace {
 constexpr size_t DEPTH_VIEW = 8;
 
+constexpr VkMemoryBarrier2 HOST_READ_BARRIER{
+    .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER_2,
+    .pNext = nullptr,
+    .srcStageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
+    .srcAccessMask = VK_ACCESS_2_MEMORY_WRITE_BIT,
+    .dstStageMask = VK_PIPELINE_STAGE_2_HOST_BIT,
+    .dstAccessMask = VK_ACCESS_2_HOST_READ_BIT,
+};
+
 struct BeginRenderingCommand {
     void operator()(vk::CommandBuffer cmdbuf, vk::CommandBuffer) const {
         BeginRendering(cmdbuf, attachments);
@@ -477,6 +486,9 @@ u64 Scheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_se
                             this](vk::CommandBuffer cmdbuf, vk::CommandBuffer upload_cmdbuf) {
         upload_cmdbuf.PipelineBarrier(upload_write_barrier);
         upload_cmdbuf.End();
+        if (device.IsUMA()) {
+            cmdbuf.PipelineBarrier(HOST_READ_BARRIER);
+        }
         cmdbuf.End();
 
         if (on_submit) {