fix(replay): match capture GPU on multi-GPU systems

Two fixes for multi-GPU replay crashes:

1. _is_arm_studio_dir() falsely detected upstream renderdoc builds as
   ARM Performance Studio, triggering RTLD_GLOBAL preload of
   librenderdoc.so which corrupted Vulkan replay. Now requires
   "arm-performance-studio" in the path.

2. ReplayOptions now pins forceGPUVendor/forceGPUDeviceID to the GPU
   that was used for capture, by matching the deviceName from structured
   data against GetAvailableGPUs(). Prevents replay from picking a
   different GPU than what captured the frame.

Also resolves RENDERDOC_PYTHON_PATH relative path handling.

Author: BANANASJIM

src/rdc/daemon_server.py

@@ -166,6 +166,35 @@ def _cleanup_temp(state: DaemonState) -> None:
 _log = logging.getLogger("rdc.daemon")
 
 
+def _match_capture_gpu(cap: Any, sd: Any = None) -> Any | None:
+    """Find the GPU used for capture by matching structured data against available GPUs."""
+    try:
+        gpus = cap.GetAvailableGPUs()
+        if not gpus:
+            return None
+        if len(gpus) == 1:
+            return gpus[0]
+        if sd is None:
+            return gpus[0]
+        for i in range(len(sd.chunks)):
+            c = sd.chunks[i]
+            if c.name == "vkEnumeratePhysicalDevices":
+                for j in range(c.NumChildren()):
+                    child = c.GetChild(j)
+                    if child.name == "physProps":
+                        for k in range(child.NumChildren()):
+                            prop = child.GetChild(k)
+                            if prop.name == "deviceName":
+                                name = prop.AsString()
+                                for g in gpus:
+                                    if g.name == name:
+                                        return g
+                break
+    except Exception:  # noqa: BLE001
+        pass
+    return gpus[0] if gpus else None
+
+
 def _load_replay(state: DaemonState) -> str | None:
     """Load renderdoc module and open capture. Returns error string or None."""
     from rdc.discover import find_renderdoc
@@ -189,7 +218,13 @@ def _load_replay(state: DaemonState) -> str | None:
         cap.Shutdown()
         return "local replay not supported on this platform"
 
-    result, controller = cap.OpenCapture(rd.ReplayOptions(), None)
+    opts = rd.ReplayOptions()
+    gpu = _match_capture_gpu(cap, cap.GetStructuredData())
+    if gpu is not None:
+        opts.forceGPUVendor = gpu.vendor
+        opts.forceGPUDeviceID = gpu.deviceID
+        _log.info("replay GPU: %s (vendor=%d id=%d)", gpu.name, gpu.vendor, gpu.deviceID)
+    result, controller = cap.OpenCapture(opts, None)
     if result != rd.ResultCode.Succeeded:
         cap.Shutdown()
         return f"OpenCapture failed: {result}"
@@ -321,8 +356,18 @@ def _load_remote_replay(state: DaemonState, remote_url: str) -> str | None:
             state.local_capture_path = str(local_tmp)
             state.local_capture_is_temp = True
 
+        remote_opts = rd.ReplayOptions()
+        if state.local_capture_path:
+            tmp_cap = rd.OpenCaptureFile()
+            if tmp_cap.OpenFile(state.local_capture_path, "", None) == rd.ResultCode.Succeeded:
+                gpu = _match_capture_gpu(tmp_cap)
+                if gpu is not None:
+                    remote_opts.forceGPUVendor = gpu.vendor
+                    remote_opts.forceGPUDeviceID = gpu.deviceID
+                tmp_cap.Shutdown()
+
         result, controller = remote.OpenCapture(
-            rd.RemoteServer.NoPreference, remote_path, rd.ReplayOptions(), None
+            rd.RemoteServer.NoPreference, remote_path, remote_opts, None
         )
         if result != rd.ResultCode.Succeeded:
             _cleanup_temp_capture(state)

src/rdc/discover.py

@@ -48,9 +48,12 @@ def _get_diagnostic() -> ProbeOutcome | None:
 
 
 def _is_arm_studio_dir(directory: str) -> bool:
-    """Return True if directory contains ARM PS patched renderdoc.so + librenderdoc.so."""
+    """Return True if directory is an ARM Performance Studio renderdoc install."""
     d = Path(directory)
-    return (d / "librenderdoc.so").is_file() and (d / "renderdoc.so").is_file()
+    if not ((d / "librenderdoc.so").is_file() and (d / "renderdoc.so").is_file()):
+        return False
+    parts = d.resolve().parts
+    return any("arm-performance-studio" in p.lower() for p in parts)
 
 
 def _preload_librenderdoc(directory: str) -> None:
@@ -136,7 +139,7 @@ def find_renderdoc() -> ModuleType | None:
 
     env_path = os.environ.get("RENDERDOC_PYTHON_PATH")
     if env_path:
-        candidates.append(env_path)
+        candidates.append(os.path.abspath(env_path))
 
     try:
         candidates.extend(_platform.renderdoc_search_paths())
@@ -203,7 +206,7 @@ def find_renderdoccmd() -> Path | None:
     env_path = os.environ.get("RENDERDOC_PYTHON_PATH")
     if env_path:
         name = "renderdoccmd.exe" if sys.platform == "win32" else "renderdoccmd"
-        candidate = Path(env_path) / name
+        candidate = Path(os.path.abspath(env_path)) / name
         if candidate.exists():
             return candidate
 

tests/mocks/mock_renderdoc.py

@@ -1824,7 +1824,9 @@ def CreateHeadlessWindowingData(width: int, height: int) -> Any:
 
 
 class ReplayOptions:
-    pass
+    forceGPUVendor: int = 0
+    forceGPUDeviceID: int = 0
+    forceGPUDriverName: str = ""
 
 
 def ExecuteAndInject(

tests/unit/test_discover.py

@@ -239,9 +239,16 @@ class TestArmStudioDir:
     """_is_arm_studio_dir detects ARM PS directory layout."""
 
     def test_both_files_present(self, tmp_path: Path) -> None:
+        arm_dir = tmp_path / "arm-performance-studio" / "renderdoc" / "lib"
+        arm_dir.mkdir(parents=True)
+        (arm_dir / "librenderdoc.so").write_text("fake")
+        (arm_dir / "renderdoc.so").write_text("fake")
+        assert _is_arm_studio_dir(str(arm_dir)) is True
+
+    def test_non_arm_dir_with_both_files(self, tmp_path: Path) -> None:
         (tmp_path / "librenderdoc.so").write_text("fake")
         (tmp_path / "renderdoc.so").write_text("fake")
-        assert _is_arm_studio_dir(str(tmp_path)) is True
+        assert _is_arm_studio_dir(str(tmp_path)) is False
 
     def test_missing_librenderdoc(self, tmp_path: Path) -> None:
         (tmp_path / "renderdoc.so").write_text("fake")