nx2: map GPU memory before later command lists are read

837133792d queues the GMMU maps on the GPU thread (GPU::RunInOrder) so the
service thread never waits for the texture cache. With the GPU split
(Low GPU accuracy) the GPU thread defers queued functions to its back, but
its front reads the command lists: it read the lists submitted after a map
through GPU addresses that weren't mapped yet. Pokopia stopped after its
first frame.

Add GPU::MapInOrder for the GMMU maps and the sparse tile remaps. It runs
in order like RunInOrder, and with the split it runs on the front after
draining the back, before the front reads further command lists. The
service thread still never takes the texture cache lock, and the back is
idle while the map changes the caches. Unmaps and the other queued work
still run on the back.

Assisted-by: Claude Opus 5.5 (bisect, fix, testing and commit message)

5 files changed, 27 insertions, 7 deletions

diff --git a/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp b/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp
index 843d31caf1..c5b23bef3e 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp
@@ -176,7 +176,7 @@ void nvhost_as_gpu::Nx2Map(u64 gpu_address, u64 cpu_address, u64 size) {
     Nx2GpuMemoryRegistry::Get().RecordDeviceRange(cpu_address, size, device_address);
     // On the GPU thread, like unmaps: a GMMU map takes the texture cache lock, and the callers hold
     // nx2:gpu locks that the GPU thread takes under that lock (sparse blocks backed on write).
-    system.GPU().RunInOrder([gmmu = gmmu.get(), gpu_address, device_address, size] {
+    system.GPU().MapInOrder([gmmu = gmmu.get(), gpu_address, device_address, size] {
         gmmu->Map(gpu_address, device_address, size, Tegra::PTEKind::PITCH, false);
     });
     LOG_DEBUG(Service_NVDRV, "Switch 2 GPU mapping {:#x}+{:#x} -> CPU {:#x}", gpu_address, size,
@@ -253,7 +253,7 @@ void nvhost_as_gpu::Nx2Sparse(u64 gpu_address, u64 size, bool create) {
     }
     nx2_sparse_ranges[gpu_address] = {size, device_address};
     // On the GPU thread, as in Nx2Map.
-    system.GPU().RunInOrder([gmmu = gmmu.get(), gpu_address, device_address, size] {
+    system.GPU().MapInOrder([gmmu = gmmu.get(), gpu_address, device_address, size] {
         gmmu->Map(gpu_address, device_address, size, Tegra::PTEKind::PITCH, false);
     });
     LOG_DEBUG(Service_NVDRV, "Switch 2 sparse region {:#x}+{:#x} -> device {:#x}", gpu_address,
@@ -295,7 +295,7 @@ void nvhost_as_gpu::Nx2SparseTile(u64 gpu_address, u64 cpu_address, u64 size) {
     // whose contents are now what the game's memory holds there.
     const DAddr device = device_address + (gpu_address - start);
     auto& smmu = system.Host1x().MemoryManager();
-    system.GPU().RunInOrder([&smmu, device, cpu_address, size, asid = session->asid] {
+    system.GPU().MapInOrder([&smmu, device, cpu_address, size, asid = session->asid] {
         smmu.Unmap(device, size);
         if (cpu_address != 0) {
             smmu.Map(device, cpu_address, size, asid, true);
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index a6f31359a3..444dc6d015 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -210,6 +210,10 @@ struct GPU::Impl {
         gpu_thread.Run(std::move(func), is_async);
     }
 
+    void MapInOrder(std::function<void()>&& func) {
+        gpu_thread.Run(std::move(func), is_async, true);
+    }
+
     /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
     void FlushRegion(DAddr addr, u64 size) {
         gpu_thread.FlushRegion(addr, size, is_async);
@@ -553,6 +557,10 @@ void GPU::RunInOrder(std::function<void()>&& func) {
     impl->RunInOrder(std::move(func));
 }
 
+void GPU::MapInOrder(std::function<void()>&& func) {
+    impl->MapInOrder(std::move(func));
+}
+
 void GPU::InvalidateRegion(DAddr addr, u64 size, bool preserve_gpu_writes) {
     impl->InvalidateRegion(addr, size, preserve_gpu_writes);
 }
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index 4afc89f0e6..d37dfa5990 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -257,6 +257,10 @@ public:
     /// Runs func on the GPU thread after the command lists submitted so far.
     void RunInOrder(std::function<void()>&& func);
 
+    /// RunInOrder for GPU memory maps: also done before the GPU thread reads the command lists
+    /// submitted later (with the GPU split, on its front rather than its back).
+    void MapInOrder(std::function<void()>&& func);
+
     /// Notify rasterizer that any caches of the specified region should be invalidated
     void InvalidateRegion(DAddr addr, u64 size, bool preserve_gpu_writes = false);
 
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index 55200b0647..3af8cb65a8 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -99,7 +99,12 @@ void ThreadManager::StartThread(VideoCore::RendererBase& renderer, Core::Fronten
                 } else if (std::holds_alternative<GPUTickCommand>(next.data)) {
                     split.Defer([gpu = &system.GPU()] { gpu->TickWork(); });
                 } else if (auto* run = std::get_if<RunCommand>(&next.data)) {
-                    split.DeferFunction(std::move(run->func));
+                    if (run->before_later_reads) {
+                        split.Drain();
+                        run->func();
+                    } else {
+                        split.DeferFunction(std::move(run->func));
+                    }
                 } else if (const auto* flush = std::get_if<FlushRegionCommand>(&next.data)) {
                     split.Defer([rasterizer_ptr, addr = flush->addr, size = flush->size] {
                         rasterizer_ptr->FlushRegion(addr, size);
@@ -155,8 +160,8 @@ void ThreadManager::TickGPU(bool is_async) {
     PushCommand(GPUTickCommand(), false, is_async);
 }
 
-void ThreadManager::Run(std::function<void()>&& func, bool is_async) {
-    PushCommand(RunCommand{std::move(func)}, false, is_async);
+void ThreadManager::Run(std::function<void()>&& func, bool is_async, bool before_later_reads) {
+    PushCommand(RunCommand{std::move(func), before_later_reads}, false, is_async);
 }
 
 void ThreadManager::InvalidateRegion(DAddr addr, u64 size) {
diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
index 796c0d89f6..927453dab0 100644
--- a/src/video_core/gpu_thread.h
+++ b/src/video_core/gpu_thread.h
@@ -80,6 +80,9 @@ struct GPUTickCommand final {};
 /// Command to run a function on the GPU thread after the commands queued before it
 struct RunCommand final {
     std::function<void()> func;
+    /// With the GPU split, run on its front (after draining the back) instead of deferring to the
+    /// back, so command lists read after this one see what func did (GMMU maps).
+    bool before_later_reads{};
 };
 
 using CommandData =
@@ -132,7 +135,7 @@ public:
     void TickGPU(bool is_async);
 
     /// Runs func on the GPU thread once the commands queued before it have been processed.
-    void Run(std::function<void()>&& func, bool is_async);
+    void Run(std::function<void()>&& func, bool is_async, bool before_later_reads = false);
 
     void NotifyShutdown();