nx2: apply GPU accuracy changes mid-game by stopping and restarting the GPU split

The split starts only with unsafe command reads, decided once at start-up. Switching to
accurate afterwards made the front thread flush caches the back thread owns, which raced on
Scheduler::chunk and crashed. MaybeStart now re-checks the settings between commands: it
drains and stops the split when they no longer allow it and starts it again when they do.
DmaPusher keeps unsafe reads while the split is active to cover the gap until then.

1 files changed, 4 insertions, 1 deletions

diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index 62f7e6bfdb..88d571ae5f 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -151,7 +151,10 @@ bool DmaPusher::Step() {
             }
         }
         // A resumed entry is read again: the words after the acquire may be new by now.
-        const bool use_safe = Settings::IsDMALevelDefault() ? Settings::IsGPULevelHigh() : Settings::IsDMALevelSafe();
+        // A safe read flushes the caches, which belong to the split's back thread while it runs
+        // (the split only starts with unsafe reads; the settings can change mid-game).
+        const bool use_safe = !VideoCommon::GpuSplit::Get().Active() &&
+                              (Settings::IsDMALevelDefault() ? Settings::IsGPULevelHigh() : Settings::IsDMALevelSafe());
         if (use_safe) {
             Tegra::Memory::GpuGuestMemory<Tegra::CommandHeader, Tegra::Memory::GuestMemoryFlags::SafeRead>headers(memory_manager, dma_state.dma_get, header.size, &command_headers);
             const std::span<const CommandHeader> commands{headers.data(), headers.size()};