graphics: execute valid PM4 rewind packets

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -112,6 +112,7 @@ public:
 	void SetCeComplete(bool complete) { m_ce_complete = complete; }
 	void WaitCe();
 	void WaitDeDiff(uint32_t diff);
+	void WaitForRewind(bool valid);
 	void IncrementDe();
 	void IncrementCe();
 

src/graphics/guest_gpu/command_processor/pm4Dispatch.cpp

@@ -224,6 +224,7 @@ constexpr auto MakeOpcodeDispatchTable() {
 	g_cp_op_func[Pm4::IT_EVENT_WRITE_EOS]           = CpOpEventWriteEos;
 	g_cp_op_func[Pm4::IT_DMA_DATA]                  = CpOpDmaData;
 	g_cp_op_func[Pm4::IT_ACQUIRE_MEM]               = CpOpAcquireMem;
+	g_cp_op_func[Pm4::IT_REWIND]                    = CpOpRewind;
 	g_cp_op_func[Pm4::IT_SET_CONTEXT_REG]           = CpOpSetContextReg;
 	g_cp_op_func[Pm4::IT_SET_SH_REG]                = CpOpSetShaderReg;
 	g_cp_op_func[Pm4::IT_SET_UCONFIG_REG]           = CpOpSetUconfigReg;

src/graphics/guest_gpu/command_processor/pm4Dispatch.h

@@ -133,6 +133,7 @@ uint32_t CpOpDrawIndirectMulti(CommandProcessor&, uint32_t, const uint32_t*, uin
 uint32_t CpOpDrawIndex(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpDrawIndexOffset(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpPfpSyncMe(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
+uint32_t CpOpRewind(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpIndexType(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpNumInstances(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpDrawIndexAuto(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -1830,6 +1830,17 @@ KYTY_CP_OP_PARSER(CpOpPfpSyncMe) {
 	return 1;
 }
 
+KYTY_CP_OP_PARSER(CpOpRewind) {
+	KYTY_PROFILER_FUNCTION();
+
+	EXIT_NOT_IMPLEMENTED(cmd_id != 0xc0005900);
+	EXIT_NOT_IMPLEMENTED((buffer[0] & ~0x81000000u) != 0);
+
+	cp.WaitForRewind((buffer[0] & 0x80000000u) != 0);
+
+	return 1;
+}
+
 KYTY_CP_OP_PARSER(CpOpSetPredication) {
 	KYTY_PROFILER_FUNCTION();
 

src/graphics/guest_gpu/graphicsRun.cpp

@@ -374,6 +374,12 @@ void CommandProcessor::WaitDeDiff(uint32_t diff) {
 	}
 }
 
+void CommandProcessor::WaitForRewind(bool valid) {
+	if (!valid) {
+		SuspendPm4();
+	}
+}
+
 void CommandProcessor::IncrementDe() {
 	m_de_count++;
 }

tests/ShaderRecompilerComputeTests.cpp

@@ -21096,6 +21096,39 @@ void CheckPm4WaitResume(RenderContext &renderer) {
   std::printf("[host]    %-32s ok\n", "Pm4WaitResume");
 }
 
+void CheckPm4RewindResume(RenderContext &renderer) {
+  GraphicsInitJmpTables();
+  CommandProcessor processor(renderer);
+
+  uint32_t suffix = 0;
+  const auto address = reinterpret_cast<uint64_t>(&suffix);
+  std::array<uint32_t, 7> commands{};
+  commands[0] = KYTY_PM4(2, Pm4::IT_REWIND, 0);
+  commands[1] = 0;
+  commands[2] = KYTY_PM4(5, Pm4::IT_WRITE_DATA, 0);
+  commands[3] = 0;
+  commands[4] = static_cast<uint32_t>(address);
+  commands[5] = static_cast<uint32_t>(address >> 32u);
+  commands[6] = 33;
+
+  Pm4Execution execution;
+  Require("Pm4RewindResume", "pending",
+          processor.Process(execution, commands.data(), commands.size()) ==
+                  Pm4ProcessResult::Blocked &&
+              suffix == 0,
+          "pending rewind did not preserve its command position");
+  Require("Pm4RewindResume", "patch valid",
+          Gen5::AgcRewindPatchSetRewindState(commands.data(), 1) == 0 &&
+              commands[1] == 0x80000000u,
+          "rewind state patch did not set the valid bit");
+  Require("Pm4RewindResume", "resume",
+          processor.Process(execution, commands.data(), commands.size()) ==
+                  Pm4ProcessResult::Complete &&
+              suffix == 33,
+          "valid rewind did not resume at the following packet");
+  std::printf("[host]    %-32s ok\n", "Pm4RewindResume");
+}
+
 void CheckPm4CeCompletion(RenderContext &renderer) {
   GraphicsInitJmpTables();
   CommandProcessor processor(renderer);
@@ -21223,6 +21256,11 @@ int main(int argc, char **argv) {
     CheckPm4ContextStateOperations(vulkan.RuntimeRenderer());
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--rewind-only") == 0) {
+    VulkanHarness vulkan;
+    CheckPm4RewindResume(vulkan.RuntimeRenderer());
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--image-overlap-only") == 0) {
     CheckDepthAttachmentWrites();
     CheckDynamicRenderingState();
@@ -21381,6 +21419,7 @@ int main(int argc, char **argv) {
   CheckPm4PolygonOffsetRegisters(vulkan.RuntimeRenderer());
   CheckPm4ContextStateOperations(vulkan.RuntimeRenderer());
   CheckPm4WaitResume(vulkan.RuntimeRenderer());
+  CheckPm4RewindResume(vulkan.RuntimeRenderer());
   CheckPm4CeCompletion(vulkan.RuntimeRenderer());
   CheckEmbeddedFetchVertexOffset();
   CheckEmbeddedFetchLaneSpill();