graphics: add experimental Pro mode and linear PSSR

Co-authored-by: Joshua Jayasingh <joshuajayasingh08@gmail.com>

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -3,6 +3,7 @@
 
 #include "common/assert.h"
 #include "graphics/guest_gpu/hardwareContext.h"
+#include "graphics/guest_gpu/pssr.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 
@@ -92,6 +93,7 @@ public:
 	                    uint32_t mode);
 	void DispatchIndirect(uint64_t args_addr, uint32_t mode);
 	void WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index);
+	void ProcessPssr(const Pssr::Request& request);
 	void TriggerEvent(uint32_t event_type, uint32_t event_index, uint64_t event_address = 0);
 
 	void WaitForRewind(bool valid);

src/graphics/guest_gpu/command_processor/pm4Dispatch.cpp

@@ -226,6 +226,7 @@ constexpr auto MakeCustomOpcodeDispatchTable() {
 	g_cp_op_custom_func[Pm4::R_POP_MARKER]     = CpOpPopMarker;
 	g_cp_op_custom_func[Pm4::R_FLIP]           = CpOpFlip;
 	g_cp_op_custom_func[Pm4::R_CONTEXT_STATE]  = CpOpContextState;
+	g_cp_op_custom_func[Pm4::R_PSSR]           = CpOpPssr;
 
 	return g_cp_op_custom_func;
 }

src/graphics/guest_gpu/command_processor/pm4Dispatch.h

@@ -151,6 +151,7 @@ uint32_t CpOpPushMarker(CommandProcessor&, uint32_t, const uint32_t*, uint32_t,
 uint32_t CpOpPopMarker(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpWaitRegMem64(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpFlip(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
+uint32_t CpOpPssr(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 
 } // namespace Libs::Graphics
 

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -1706,6 +1706,17 @@ KYTY_CP_OP_PARSER(CpOpFlip) {
 	return 5;
 }
 
+KYTY_CP_OP_PARSER(CpOpPssr) {
+	KYTY_PROFILER_FUNCTION();
+
+	const auto packet_dw = KYTY_PM4_LEN(cmd_id) - 1;
+	EXIT_NOT_IMPLEMENTED(packet_dw < Pssr::RequestDwords || packet_dw > dw);
+	Pssr::Request request {};
+	std::memcpy(&request, buffer, sizeof(request));
+	cp.ProcessPssr(request);
+	return packet_dw;
+}
+
 KYTY_CP_OP_PARSER(CpOpIndexType) {
 	KYTY_PROFILER_FUNCTION();
 

src/graphics/guest_gpu/graphicsRun.cpp

@@ -929,6 +929,11 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 	}
 }
 
+void CommandProcessor::ProcessPssr(const Pssr::Request& request) {
+	EXIT_NOT_IMPLEMENTED(IsAsyncComputeQueue());
+	m_renderer.GetRenderExecutor().ProcessPssr(CurrentBuffer(), request);
+}
+
 void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {
 	EXIT_NOT_IMPLEMENTED(args_addr == 0 || (args_addr & 3u) != 0);
 	if ((mode & Pm4::COMPUTE_DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS) != 0) {

src/graphics/guest_gpu/pm4.h

@@ -81,6 +81,7 @@ constexpr uint32_t R_WRITE_DATA     = 0x15;
 constexpr uint32_t R_FLIP           = 0x17;
 constexpr uint32_t R_DMA_DATA       = 0x19;
 constexpr uint32_t R_CONTEXT_STATE  = 0x1A;
+constexpr uint32_t R_PSSR           = 0x1B;
 
 constexpr uint32_t R_NUM = 0x3F + 1;
 

src/graphics/guest_gpu/pssr.h

@@ -0,0 +1,20 @@
+#ifndef EMULATOR_SRC_GRAPHICS_GUEST_GPU_PSSR_H_
+#define EMULATOR_SRC_GRAPHICS_GUEST_GPU_PSSR_H_
+
+#include <cstdint>
+
+namespace Libs::Graphics::Pssr {
+
+struct Request {
+	uint32_t output[8];
+	uint32_t input[8];
+	uint32_t width;
+	uint32_t height;
+};
+
+constexpr uint32_t RequestDwords = sizeof(Request) / sizeof(uint32_t);
+static_assert(sizeof(Request) == 72);
+
+} // namespace Libs::Graphics::Pssr
+
+#endif // EMULATOR_SRC_GRAPHICS_GUEST_GPU_PSSR_H_