D3D12: forget cached command list state after a driver callback

A driver callback hands the command buffer to native code the driver knows
nothing about. Streamline, running DLSS, rebinds descriptor heaps, the pipeline
state and the root signature inside that callback. D3D12 binds descriptor heaps
to the command list rather than to a draw, so the driver caches what it believes
is bound for the whole command buffer and sets the heaps only once. After the
callback that cache is a lie: the driver kept feeding handles from its own heap
to SetComputeRootDescriptorTable while Streamline's heap was bound, and the
usermode driver dereferenced out of bounds and crashed.

Only DLSS combined with SSIL reached it. SSIL copies the internal colour buffer
into its last frame texture, and nothing consumes that copy until the next
frame, so the graph is free to schedule those compute passes after the upscaling
callback. SSAO and SSR are pinned ahead of the upscale by their dependency
chains and never land behind it.

Add RenderingDeviceDriver::command_buffer_forget_cached_state(), a no-op for
drivers that cache nothing, and call it from the graph once every driver
callback returns. The D3D12 override clears the same state command_buffer_end()
already clears. Vulkan and FSR2 were never affected.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017KCbSdUNssPYFx8HD2TMD3

Author: Critsium-xy

drivers/d3d12/rendering_device_driver_d3d12.cpp

@@ -2823,6 +2823,18 @@ void RenderingDeviceDriverD3D12::command_buffer_end(CommandBufferID p_cmd_buffer
 	cmd_buf_info->descriptor_heaps_set = false;
 }
 
+void RenderingDeviceDriverD3D12::command_buffer_forget_cached_state(CommandBufferID p_cmd_buffer) {
+	CommandBufferInfo *cmd_buf_info = (CommandBufferInfo *)p_cmd_buffer.id;
+	cmd_buf_info->graphics_pso = nullptr;
+	cmd_buf_info->graphics_root_signature_crc = 0;
+	cmd_buf_info->compute_pso = nullptr;
+	cmd_buf_info->compute_root_signature_crc = 0;
+	cmd_buf_info->nir_graphics_runtime_data_root_param_idx = UINT32_MAX;
+	cmd_buf_info->nir_compute_runtime_data_root_param_idx = UINT32_MAX;
+	cmd_buf_info->pending_dyn_params = true;
+	cmd_buf_info->descriptor_heaps_set = false;
+}
+
 void RenderingDeviceDriverD3D12::command_buffer_execute_secondary(CommandBufferID p_cmd_buffer, VectorView<CommandBufferID> p_secondary_cmd_buffers) {
 	const CommandBufferInfo *cmd_buf_info = (const CommandBufferInfo *)p_cmd_buffer.id;
 	for (uint32_t i = 0; i < p_secondary_cmd_buffers.size(); i++) {

drivers/d3d12/rendering_device_driver_d3d12.h

@@ -540,6 +540,7 @@ class RenderingDeviceDriverD3D12 : public RenderingDeviceDriver {
 	virtual bool command_buffer_begin_secondary(CommandBufferID p_cmd_buffer, RenderPassID p_render_pass, uint32_t p_subpass, FramebufferID p_framebuffer) override final;
 	virtual void command_buffer_end(CommandBufferID p_cmd_buffer) override final;
 	virtual void command_buffer_execute_secondary(CommandBufferID p_cmd_buffer, VectorView<CommandBufferID> p_secondary_cmd_buffers) override final;
+	virtual void command_buffer_forget_cached_state(CommandBufferID p_cmd_buffer) override final;
 
 private:
 	/********************/

servers/rendering/rendering_device_driver.h

@@ -459,6 +459,10 @@ class RenderingDeviceDriver : public RenderingDeviceCommons {
 	virtual bool command_buffer_begin_secondary(CommandBufferID p_cmd_buffer, RenderPassID p_render_pass, uint32_t p_subpass, FramebufferID p_framebuffer) = 0;
 	virtual void command_buffer_end(CommandBufferID p_cmd_buffer) = 0;
 	virtual void command_buffer_execute_secondary(CommandBufferID p_cmd_buffer, VectorView<CommandBufferID> p_secondary_cmd_buffers) = 0;
+	// Called after native code outside the driver has recorded commands into the command buffer.
+	// Such code is free to rebind pipelines, root signatures or descriptor heaps, so a driver that
+	// caches what it believes is currently bound has to forget it and set it again next time.
+	virtual void command_buffer_forget_cached_state(CommandBufferID p_cmd_buffer) {}
 
 	/********************/
 	/**** SWAP CHAIN ****/

servers/rendering/rendering_device_graph.cpp

@@ -1142,6 +1142,7 @@ void RenderingDeviceGraph::_run_render_commands(int32_t p_level, const RecordedC
 			case RecordedCommand::TYPE_DRIVER_CALLBACK: {
 				const RecordedDriverCallbackCommand *driver_callback_command = reinterpret_cast<const RecordedDriverCallbackCommand *>(command);
 				driver_callback_command->callback(driver, r_command_buffer, driver_callback_command->userdata);
+				driver->command_buffer_forget_cached_state(r_command_buffer);
 			} break;
 			case RecordedCommand::TYPE_RAYTRACING_LIST: {
 				const RecordedRaytracingListCommand *raytracing_list_command = reinterpret_cast<const RecordedRaytracingListCommand *>(command);