rsx/gcm: Implement rsx dma. Refactor gcm/rsx to not be as codependent

rpcs3/Emu/Cell/Modules/cellGcmSys.cpp

@@ -26,16 +26,19 @@ const u32 tiled_pitches[] = {
 	0x00010000
 };
 
+struct CellGcmSysConfig {
+	u32 zculls_addr;
+	vm::ptr<CellGcmDisplayInfo> gcm_buffers{ vm::null };
+	u32 tiles_addr;
+	u32 ctxt_addr;
+	CellGcmConfig current_config;
+	CellGcmContextData current_context;
+	gcmInfo gcm_info;
+};
+
+u64 system_mode = 0;
 u32 local_size = 0;
 u32 local_addr = 0;
-u64 system_mode = 0;
-
-CellGcmConfig current_config;
-CellGcmContextData current_context;
-gcmInfo gcm_info;
-
-u32 map_offset_addr = 0;
-u32 map_offset_pos = 0;
 
 // Auxiliary functions
 
@@ -82,7 +85,12 @@ void InitOffsetTable()
 u32 cellGcmGetLabelAddress(u8 index)
 {
 	cellGcmSys.trace("cellGcmGetLabelAddress(index=%d)", index);
-	return gcm_info.label_addr + 0x10 * index;
+	const auto m_config = fxm::get<CellGcmSysConfig>();
+
+	if (!m_config)
+		return 0;
+
+	return m_config->gcm_info.label_addr + 0x10 * index;
 }
 
 vm::ptr<CellGcmReportData> cellGcmGetReportDataAddressLocation(u32 index, u32 location)
@@ -94,7 +102,7 @@ vm::ptr<CellGcmReportData> cellGcmGetReportDataAddressLocation(u32 index, u32 lo
 			cellGcmSys.error("cellGcmGetReportDataAddressLocation: Wrong local index (%d)", index);
 			return vm::null;
 		}
-		return vm::ptr<CellGcmReportData>::make(0xC0000000 + index * 0x10);
+		return vm::ptr<CellGcmReportData>::make(0x40301400 + index * 0x10);
 	}
 
 	if (location == CELL_GCM_LOCATION_MAIN) {
@@ -117,7 +125,7 @@ u64 cellGcmGetTimeStamp(u32 index)
 		cellGcmSys.error("cellGcmGetTimeStamp: Wrong local index (%d)", index);
 		return 0;
 	}
-	return vm::read64(0xC0000000 + index * 0x10);
+	return vm::read64(0x40301400 + index * 0x10);
 }
 
 u32 cellGcmGetCurrentField()
@@ -144,7 +152,7 @@ u32 cellGcmGetNotifyDataAddress(u32 index)
  */
 vm::ptr<CellGcmReportData> _cellGcmFunc12()
 {
-	return vm::ptr<CellGcmReportData>::make(0xC0000000); // TODO
+	return vm::ptr<CellGcmReportData>::make(0x40301400); // TODO
 }
 
 u32 cellGcmGetReport(u32 type, u32 index)
@@ -172,7 +180,7 @@ u32 cellGcmGetReportDataAddress(u32 index)
 		cellGcmSys.error("cellGcmGetReportDataAddress: Wrong local index (%d)", index);
 		return 0;
 	}
-	return 0xC0000000 + index * 0x10;
+	return 0x40301400 + index * 0x10;
 }
 
 u32 cellGcmGetReportDataLocation(u32 index, u32 location)
@@ -192,7 +200,7 @@ u64 cellGcmGetTimeStampLocation(u32 index, u32 location)
 			cellGcmSys.error("cellGcmGetTimeStampLocation: Wrong local index (%d)", index);
 			return 0;
 		}
-		return vm::read64(0xC0000000 + index * 0x10);
+		return vm::read64(0x40301400 + index * 0x10);
 	}
 
 	if (location == CELL_GCM_LOCATION_MAIN) {
@@ -214,20 +222,31 @@ u64 cellGcmGetTimeStampLocation(u32 index, u32 location)
 u32 cellGcmGetControlRegister()
 {
 	cellGcmSys.trace("cellGcmGetControlRegister()");
+	const auto m_config = fxm::get<CellGcmSysConfig>();
 
-	return gcm_info.control_addr;
+	if (!m_config)
+		return 0;
+	return m_config->gcm_info.control_addr;
 }
 
 u32 cellGcmGetDefaultCommandWordSize()
 {
 	cellGcmSys.trace("cellGcmGetDefaultCommandWordSize()");
-	return gcm_info.command_size;
+	const auto m_config = fxm::get<CellGcmSysConfig>();
+
+	if (!m_config)
+		return 0;
+	return m_config->gcm_info.command_size;
 }
 
 u32 cellGcmGetDefaultSegmentWordSize()
 {
 	cellGcmSys.trace("cellGcmGetDefaultSegmentWordSize()");
-	return gcm_info.segment_size;
+	const auto m_config = fxm::get<CellGcmSysConfig>();
+
+	if (!m_config)
+		return 0;
+	return m_config->gcm_info.segment_size;
 }
 
 s32 cellGcmInitDefaultFifoMode(s32 mode)
@@ -280,8 +299,10 @@ s32 cellGcmBindZcull(u8 index, u32 offset, u32 width, u32 height, u32 cullStart,
 void cellGcmGetConfiguration(vm::ptr<CellGcmConfig> config)
 {
 	cellGcmSys.trace("cellGcmGetConfiguration(config=*0x%x)", config);
+	const auto m_config = fxm::get<CellGcmSysConfig>();
 
-	*config = current_config;
+	if (m_config)
+		*config = m_config->current_config;
 }
 
 u32 cellGcmGetFlipStatus()
@@ -297,9 +318,9 @@ u32 cellGcmGetTiledPitchSize(u32 size)
 {
 	cellGcmSys.trace("cellGcmGetTiledPitchSize(size=%d)", size);
 
-	for (size_t i=0; i < sizeof(tiled_pitches) / sizeof(tiled_pitches[0]) - 1; i++) {
+	for (size_t i = 0; i < sizeof(tiled_pitches) / sizeof(tiled_pitches[0]) - 1; i++) {
-		if (tiled_pitches[i] < size && size <= tiled_pitches[i+1]) {
+		if (tiled_pitches[i] < size && size <= tiled_pitches[i + 1]) {
-			return tiled_pitches[i+1];
+			return tiled_pitches[i + 1];
 		}
 	}
 	return 0;
@@ -324,8 +345,12 @@ s32 _cellGcmInitBody(vm::pptr<CellGcmContextData> context, u32 cmdSize, u32 ioSi
 {
 	cellGcmSys.warning("_cellGcmInitBody(context=**0x%x, cmdSize=0x%x, ioSize=0x%x, ioAddress=0x%x)", context, cmdSize, ioSize, ioAddress);
 
-	current_config.ioAddress = 0;
+	auto m_config = fxm::make<CellGcmSysConfig>();
-	current_config.localAddress = 0;
+	if (!m_config)
+		return CELL_GCM_ERROR_FAILURE;
+
+	m_config->current_config.ioAddress = 0;
+	m_config->current_config.localAddress = 0;
 	local_size = 0;
 	local_addr = 0;
 
@@ -356,33 +381,42 @@ s32 _cellGcmInitBody(vm::pptr<CellGcmContextData> context, u32 cmdSize, u32 ioSi
 		return CELL_GCM_ERROR_FAILURE;
 	}
 
-	map_offset_addr = 0;
+	m_config->current_config.ioSize = ioSize;
-	map_offset_pos = 0;
+	m_config->current_config.ioAddress = ioAddress;
-	current_config.ioSize = ioSize;
+	m_config->current_config.localSize = local_size;
-	current_config.ioAddress = ioAddress;
+	m_config->current_config.localAddress = local_addr;
-	current_config.localSize = local_size;
+	m_config->current_config.memoryFrequency = 650000000;
-	current_config.localAddress = local_addr;
+	m_config->current_config.coreFrequency = 500000000;
-	current_config.memoryFrequency = 650000000;
-	current_config.coreFrequency = 500000000;
 
 	// Create contexts
 
+	u32 addr = vm::falloc(0x40000000, 0x400000);
+	if (addr == 0 || addr != 0x40000000)
+		fmt::throw_exception("Failed to alloc 0x40000000.");
+
 	g_defaultCommandBufferBegin = ioAddress;
 	g_defaultCommandBufferFragmentCount = cmdSize / (32 * 1024);
 
-	gcm_info.context_addr = vm::alloc(0x1000, vm::main);
+	m_config->gcm_info.context_addr = 0x40000000;
-	gcm_info.control_addr = vm::alloc(0x1000, vm::main);
+	m_config->gcm_info.control_addr = 0x40100000;
-	gcm_info.label_addr = vm::alloc(0x1000, vm::main); // ???
+	m_config->gcm_info.label_addr = 0x40300000;
 
-	current_context.begin.set(g_defaultCommandBufferBegin + 4096); // 4 kb reserved at the beginning
+	m_config->current_context.begin.set(g_defaultCommandBufferBegin + 4096); // 4 kb reserved at the beginning
-	current_context.end.set(g_defaultCommandBufferBegin + 32 * 1024 - 4); // 4b at the end for jump
+	m_config->current_context.end.set(g_defaultCommandBufferBegin + 32 * 1024 - 4); // 4b at the end for jump
-	current_context.current = current_context.begin;
+	m_config->current_context.current = m_config->current_context.begin;
-	current_context.callback.set(ppu_function_manager::addr + 8 * FIND_FUNC(cellGcmCallback));
+	m_config->current_context.callback.set(ppu_function_manager::addr + 8 * FIND_FUNC(cellGcmCallback));
 
-	vm::_ref<CellGcmContextData>(gcm_info.context_addr) = current_context;
+	m_config->ctxt_addr = context.addr();
-	context->set(gcm_info.context_addr);
+	m_config->gcm_buffers.set(vm::alloc(sizeof(CellGcmDisplayInfo) * 8, vm::main));
+	m_config->zculls_addr = vm::alloc(sizeof(CellGcmZcullInfo) * 8, vm::main);
+	m_config->tiles_addr = vm::alloc(sizeof(CellGcmTileInfo) * 15, vm::main);
 
-	auto& ctrl = vm::_ref<CellGcmControl>(gcm_info.control_addr);
+	vm::_ref<CellGcmContextData>(m_config->gcm_info.context_addr) = m_config->current_context;
+	context->set(m_config->gcm_info.context_addr);
+
+	// 0x40 is to offset CellGcmControl from RsxDmaControl
+	m_config->gcm_info.control_addr += 0x40;
+	auto& ctrl = vm::_ref<CellGcmControl>(m_config->gcm_info.control_addr);
 	ctrl.put = 0;
 	ctrl.get = 0;
 	ctrl.ref = -1;
@@ -390,15 +424,9 @@ s32 _cellGcmInitBody(vm::pptr<CellGcmContextData> context, u32 cmdSize, u32 ioSi
 	const auto render = fxm::get<GSRender>();
 	render->intr_thread = idm::make_ptr<ppu_thread>("_gcm_intr_thread", 1, 0x4000);
 	render->intr_thread->run();
-	render->ctxt_addr = context.addr();
-	render->gcm_buffers.set(vm::alloc(sizeof(CellGcmDisplayInfo) * 8, vm::main));
-	render->zculls_addr = vm::alloc(sizeof(CellGcmZcullInfo) * 8, vm::main);
-	render->tiles_addr = vm::alloc(sizeof(CellGcmTileInfo) * 15, vm::main);
-	render->gcm_buffers_count = 0;
-	render->gcm_current_buffer = 0;
 	render->main_mem_addr = 0;
-	render->label_addr = gcm_info.label_addr;
+	render->label_addr = m_config->gcm_info.label_addr;
-	render->init(g_defaultCommandBufferBegin, cmdSize, gcm_info.control_addr, local_addr);
+	render->init(ioAddress, ioSize, m_config->gcm_info.control_addr - 0x40, local_addr);
 
 	return CELL_OK;
 }
@@ -431,6 +459,10 @@ s32 cellGcmSetDisplayBuffer(u8 id, u32 offset, u32 pitch, u32 width, u32 height)
 {
 	cellGcmSys.trace("cellGcmSetDisplayBuffer(id=0x%x, offset=0x%x, pitch=%d, width=%d, height=%d)", id, offset, width ? pitch / width : pitch, width, height);
 
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return CELL_GCM_ERROR_FAILURE;
+
 	if (id > 7)
 	{
 		cellGcmSys.error("cellGcmSetDisplayBuffer: CELL_GCM_ERROR_FAILURE");
@@ -439,16 +471,21 @@ s32 cellGcmSetDisplayBuffer(u8 id, u32 offset, u32 pitch, u32 width, u32 height)
 
 	const auto render = fxm::get<GSRender>();
 
-	auto buffers = render->gcm_buffers;
+	auto buffers = render->display_buffers;
 
 	buffers[id].offset = offset;
 	buffers[id].pitch = pitch;
 	buffers[id].width = width;
 	buffers[id].height = height;
 
-	if (id + 1 > render->gcm_buffers_count)
+	m_config->gcm_buffers[id].offset = offset;
+	m_config->gcm_buffers[id].pitch = pitch;
+	m_config->gcm_buffers[id].width = width;
+	m_config->gcm_buffers[id].height = height;
+
+	if (id + 1 > render->display_buffers_count)
 	{
-		render->gcm_buffers_count = id + 1;
+		render->display_buffers_count = id + 1;
 	}
 
 	return CELL_OK;
@@ -465,17 +502,7 @@ void cellGcmSetFlipMode(u32 mode)
 {
 	cellGcmSys.warning("cellGcmSetFlipMode(mode=%d)", mode);
 
-	switch (mode)
+	fxm::get<GSRender>()->requested_vsync.store(mode == CELL_GCM_DISPLAY_VSYNC);
-	{
-	case CELL_GCM_DISPLAY_HSYNC:
-	case CELL_GCM_DISPLAY_VSYNC:
-	case CELL_GCM_DISPLAY_HSYNC_WITH_NOISE:
-		fxm::get<GSRender>()->flip_mode = mode;
-		break;
-
-	default:
-		break;
-	}
 }
 
 void cellGcmSetFlipStatus()
@@ -488,6 +515,9 @@ void cellGcmSetFlipStatus()
 s32 cellGcmSetPrepareFlip(ppu_thread& ppu, vm::ptr<CellGcmContextData> ctxt, u32 id)
 {
 	cellGcmSys.trace("cellGcmSetPrepareFlip(ctxt=*0x%x, id=0x%x)", ctxt, id);
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return CELL_GCM_ERROR_FAILURE;
 
 	if (id > 7)
 	{
@@ -506,9 +536,9 @@ s32 cellGcmSetPrepareFlip(ppu_thread& ppu, vm::ptr<CellGcmContextData> ctxt, u32
 
 	const u32 cmd_size = rsx::make_command(ctxt->current, GCM_FLIP_COMMAND, { id });
 
-	if (ctxt.addr() == gcm_info.context_addr)
+	if (ctxt.addr() == m_config->gcm_info.context_addr)
 	{
-		vm::_ref<CellGcmControl>(gcm_info.control_addr).put += cmd_size;
+		vm::_ref<CellGcmControl>(m_config->gcm_info.control_addr).put += cmd_size;
 	}
 
 	return id;
@@ -535,15 +565,12 @@ void cellGcmSetSecondVFrequency(u32 freq)
 	switch (freq)
 	{
 	case CELL_GCM_DISPLAY_FREQUENCY_59_94HZ:
-		render->frequency_mode = freq;
 		render->fps_limit = 59.94;
 		break;
 	case CELL_GCM_DISPLAY_FREQUENCY_SCANOUT:
-		render->frequency_mode = freq;
 		cellGcmSys.todo("Unimplemented display frequency: Scanout");
 		break;
 	case CELL_GCM_DISPLAY_FREQUENCY_DISABLE:
-		render->frequency_mode = freq;
 		cellGcmSys.todo("Unimplemented display frequency: Disabled");
 		break;
 	default:
@@ -557,6 +584,10 @@ s32 cellGcmSetTileInfo(u8 index, u8 location, u32 offset, u32 size, u32 pitch, u
 	cellGcmSys.warning("cellGcmSetTileInfo(index=%d, location=%d, offset=%d, size=%d, pitch=%d, comp=%d, base=%d, bank=%d)",
 		index, location, offset, size, pitch, comp, base, bank);
 
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return CELL_GCM_ERROR_FAILURE;
+
 	if (index >= rsx::limits::tiles_count || base >= 2048 || bank >= 4)
 	{
 		cellGcmSys.error("cellGcmSetTileInfo: CELL_GCM_ERROR_INVALID_VALUE");
@@ -591,7 +622,7 @@ s32 cellGcmSetTileInfo(u8 index, u8 location, u32 offset, u32 size, u32 pitch, u
 	tile.base = base;
 	tile.bank = bank;
 
-	vm::_ptr<CellGcmTileInfo>(render->tiles_addr)[index] = tile.pack();
+	vm::_ptr<CellGcmTileInfo>(m_config->tiles_addr)[index] = tile.pack();
 	return CELL_OK;
 }
 
@@ -633,6 +664,10 @@ void cellGcmSetZcull(u8 index, u32 offset, u32 width, u32 height, u32 cullStart,
 	cellGcmSys.todo("cellGcmSetZcull(index=%d, offset=0x%x, width=%d, height=%d, cullStart=0x%x, zFormat=0x%x, aaFormat=0x%x, zCullDir=0x%x, zCullFormat=0x%x, sFunc=0x%x, sRef=0x%x, sMask=0x%x)",
 		index, offset, width, height, cullStart, zFormat, aaFormat, zCullDir, zCullFormat, sFunc, sRef, sMask);
 
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return;
+
 	if (index >= rsx::limits::zculls_count)
 	{
 		cellGcmSys.error("cellGcmSetZcull: CELL_GCM_ERROR_INVALID_VALUE");
@@ -654,7 +689,7 @@ void cellGcmSetZcull(u8 index, u32 offset, u32 width, u32 height, u32 cullStart,
 	zcull.sRef = sRef;
 	zcull.sMask = sMask;
 
-	vm::_ptr<CellGcmZcullInfo>(render->zculls_addr)[index] = zcull.pack();
+	vm::_ptr<CellGcmZcullInfo>(m_config->zculls_addr)[index] = zcull.pack();
 }
 
 s32 cellGcmUnbindTile(u8 index)
@@ -690,26 +725,35 @@ s32 cellGcmUnbindZcull(u8 index)
 u32 cellGcmGetTileInfo()
 {
 	cellGcmSys.warning("cellGcmGetTileInfo()");
-	return fxm::get<GSRender>()->tiles_addr;
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return 0;
+	return m_config->tiles_addr;
 }
 
 u32 cellGcmGetZcullInfo()
 {
 	cellGcmSys.warning("cellGcmGetZcullInfo()");
-	return fxm::get<GSRender>()->zculls_addr;
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return 0;
+	return m_config->zculls_addr;
 }
 
 u32 cellGcmGetDisplayInfo()
 {
 	cellGcmSys.warning("cellGcmGetDisplayInfo()");
-	return fxm::get<GSRender>()->gcm_buffers.addr();
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return 0;
+	return m_config->gcm_buffers.addr();
 }
 
 s32 cellGcmGetCurrentDisplayBufferId(vm::ptr<u8> id)
 {
 	cellGcmSys.warning("cellGcmGetCurrentDisplayBufferId(id=*0x%x)", id);
 
-	if ((*id = fxm::get<GSRender>()->gcm_current_buffer) > UINT8_MAX)
+	if ((*id = fxm::get<GSRender>()->current_display_buffer) > UINT8_MAX)
 	{
 		fmt::throw_exception("Unexpected" HERE);
 	}
@@ -898,7 +942,6 @@ s32 gcmMapEaIoAddress(u32 ea, u32 io, u32 size, bool is_strict)
 		{
 			offsetTable.ioAddress[(ea >> 20) + i] = (io >> 20) + i;
 			offsetTable.eaAddress[(io >> 20) + i] = (ea >> 20) + i;
-			render->strict_ordering[(io >> 20) + i] = is_strict;
 		}
 	}
 	else
@@ -948,6 +991,7 @@ s32 cellGcmMapMainMemory(u32 ea, u32 size, vm::ptr<u32> offset)
 {
 	cellGcmSys.warning("cellGcmMapMainMemory(ea=0x%x, size=0x%x, offset=*0x%x)", ea, size, offset);
 
+	if (size == 0) return CELL_OK;
 	if ((ea & 0xFFFFF) || (size & 0xFFFFF)) return CELL_GCM_ERROR_FAILURE;
 
 	u32 io = RSXIOMem.Map(ea, size);
@@ -962,7 +1006,6 @@ s32 cellGcmMapMainMemory(u32 ea, u32 size, vm::ptr<u32> offset)
 		{
 			offsetTable.ioAddress[(ea >> 20) + i] = (u16)((io >> 20) + i);
 			offsetTable.eaAddress[(io >> 20) + i] = (u16)((ea >> 20) + i);
-			render->strict_ordering[(io >> 20) + i] = false;
 		}
 
 		*offset = io;
@@ -1113,23 +1156,27 @@ s32 cellGcmSetCursorImageOffset(u32 offset)
 void cellGcmSetDefaultCommandBuffer()
 {
 	cellGcmSys.warning("cellGcmSetDefaultCommandBuffer()");
-	vm::write32(fxm::get<GSRender>()->ctxt_addr, gcm_info.context_addr);
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (m_config)
+		vm::write32(m_config->ctxt_addr, m_config->gcm_info.context_addr);
 }
 
 s32 cellGcmSetDefaultCommandBufferAndSegmentWordSize(u32 bufferSize, u32 segmentSize)
 {
 	cellGcmSys.warning("cellGcmSetDefaultCommandBufferAndSegmentWordSize(bufferSize=0x%x, segmentSize=0x%x)", bufferSize, segmentSize);
-
+	auto m_config = fxm::get<CellGcmSysConfig>();
-	const auto& put = vm::_ref<CellGcmControl>(gcm_info.control_addr).put;
+	if (!m_config)
-	const auto& get = vm::_ref<CellGcmControl>(gcm_info.control_addr).get;
+		return CELL_GCM_ERROR_FAILURE;
+	const auto& put = vm::_ref<CellGcmControl>(m_config->gcm_info.control_addr).put;
+	const auto& get = vm::_ref<CellGcmControl>(m_config->gcm_info.control_addr).get;
 
 	if (put != 0x1000 || get != 0x1000 || bufferSize < segmentSize * 2)
 	{
 		return CELL_GCM_ERROR_FAILURE;
 	}
 
-	gcm_info.command_size = bufferSize;
+	m_config->gcm_info.command_size = bufferSize;
-	gcm_info.segment_size = segmentSize;
+	m_config->gcm_info.segment_size = segmentSize;
 
 	return CELL_OK;
 }
@@ -1148,9 +1195,11 @@ s32 _cellGcmSetFlipCommand(ppu_thread& ppu, vm::ptr<CellGcmContextData> ctx, u32
 s32 _cellGcmSetFlipCommandWithWaitLabel(ppu_thread& ppu, vm::ptr<CellGcmContextData> ctx, u32 id, u32 label_index, u32 label_value)
 {
 	cellGcmSys.trace("cellGcmSetFlipCommandWithWaitLabel(ctx=*0x%x, id=0x%x, label_index=0x%x, label_value=0x%x)", ctx, id, label_index, label_value);
-
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return CELL_GCM_ERROR_FAILURE;
 	s32 res = cellGcmSetPrepareFlip(ppu, ctx, id);
-	vm::write32(gcm_info.label_addr + 0x10 * label_index, label_value);
+	vm::write32(m_config->gcm_info.label_addr + 0x10 * label_index, label_value);
 	return res < 0 ? CELL_GCM_ERROR_FAILURE : CELL_OK;
 }
 
@@ -1159,6 +1208,10 @@ s32 cellGcmSetTile(u8 index, u8 location, u32 offset, u32 size, u32 pitch, u8 co
 	cellGcmSys.warning("cellGcmSetTile(index=%d, location=%d, offset=%d, size=%d, pitch=%d, comp=%d, base=%d, bank=%d)",
 		index, location, offset, size, pitch, comp, base, bank);
 
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return CELL_GCM_ERROR_FAILURE;
+
 	// Copied form cellGcmSetTileInfo
 	if (index >= rsx::limits::tiles_count || base >= 2048 || bank >= 4)
 	{
@@ -1194,7 +1247,7 @@ s32 cellGcmSetTile(u8 index, u8 location, u32 offset, u32 size, u32 pitch, u8 co
 	tile.base = base;
 	tile.bank = bank;
 
-	vm::_ptr<CellGcmTileInfo>(render->tiles_addr)[index] = tile.pack();
+	vm::_ptr<CellGcmTileInfo>(m_config->tiles_addr)[index] = tile.pack();
 	return CELL_OK;
 }
 
@@ -1286,8 +1339,11 @@ static bool isInCommandBufferExcept(u32 getPos, u32 bufferBegin, u32 bufferEnd)
 s32 cellGcmCallback(ppu_thread& ppu, vm::ptr<CellGcmContextData> context, u32 count)
 {
 	cellGcmSys.trace("cellGcmCallback(context=*0x%x, count=0x%x)", context, count);
+	auto m_config = fxm::get<CellGcmSysConfig>();
+	if (!m_config)
+		return CELL_GCM_ERROR_FAILURE;
 
-	auto& ctrl = vm::_ref<CellGcmControl>(gcm_info.control_addr);
+	auto& ctrl = vm::_ref<CellGcmControl>(m_config->gcm_info.control_addr);
 
 	// Flush command buffer (ie allow RSX to read up to context->current)
 	ctrl.put.exchange(getOffsetFromAddress(context->current.addr()));

rpcs3/Emu/Cell/lv2/sys_rsx.h

@@ -1,5 +1,44 @@
 #pragma once
 
+struct RsxDmaControl {
+	u8 resv[0x40];
+	atomic_be_t<u32> put;
+	atomic_be_t<u32> get;
+	atomic_be_t<u32> ref;
+	be_t<u32> unk[2];
+	be_t<u32> unk1;
+};
+
+struct RsxSemaphore {
+	be_t<u32> val;
+	be_t<u32> pad;
+	be_t<u64> timestamp;
+};
+
+struct RsxNotify {
+	be_t<u64> timestamp;
+	be_t<u64> zero;
+};
+
+struct RsxReport {
+	be_t<u64> timestamp;
+	be_t<u32> val;
+	be_t<u32> pad;
+};
+
+struct RsxReports {
+	RsxSemaphore semaphore[0x100];
+	RsxNotify notify[64];
+	RsxReport report[2048];
+};
+
+struct RsxDisplayInfo {
+	be_t<u32> offset;
+	be_t<u32> pitch;
+	be_t<u32> width;
+	be_t<u32> height;
+};
+
 // SysCalls
 s32 sys_rsx_device_open();
 s32 sys_rsx_device_close();

rpcs3/Emu/Memory/vm.cpp

@@ -786,6 +786,7 @@ namespace vm
 				std::make_shared<block_t>(0xC0000000, 0x10000000), // video
 				std::make_shared<block_t>(0xD0000000, 0x10000000), // stack
 				std::make_shared<block_t>(0xE0000000, 0x20000000), // SPU reserved
+				std::make_shared<block_t>(0x40000000, 0x10000000), // rsx contexts
 			};
 		}
 	}

rpcs3/Emu/RSX/D3D12/D3D12GSRender.cpp

@@ -523,10 +523,9 @@ void D3D12GSRender::flip(int buffer)
 		size_t offset = 0;
 		if (false)
 		{
-			CellGcmDisplayInfo* buffers = nullptr;// = vm::ps3::_ptr<CellGcmDisplayInfo>(m_gcm_buffers_addr);
+			u32 addr = rsx::get_address(display_buffers[current_display_buffer].offset, CELL_GCM_LOCATION_LOCAL);
-			u32 addr = rsx::get_address(gcm_buffers[gcm_current_buffer].offset, CELL_GCM_LOCATION_LOCAL);
+			w = display_buffers[current_display_buffer].width;
-			w = gcm_buffers[gcm_current_buffer].width;
+			h = display_buffers[current_display_buffer].height;
-			h = gcm_buffers[gcm_current_buffer].height;
 			u8 *src_buffer = vm::ps3::_ptr<u8>(addr);
 
 			row_pitch = align(w * 4, 256);

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -985,9 +985,9 @@ void GLGSRender::flip(int buffer)
 		return;
 	}
 
-	u32 buffer_width = gcm_buffers[buffer].width;
+	u32 buffer_width = display_buffers[buffer].width;
-	u32 buffer_height = gcm_buffers[buffer].height;
+	u32 buffer_height = display_buffers[buffer].height;
-	u32 buffer_pitch = gcm_buffers[buffer].pitch;
+	u32 buffer_pitch = display_buffers[buffer].pitch;
 
 	// Calculate blit coordinates
 	coordi aspect_ratio;
@@ -1016,7 +1016,7 @@ void GLGSRender::flip(int buffer)
 	aspect_ratio.size = new_size;
 
 	// Find the source image
-	rsx::tiled_region buffer_region = get_tiled_address(gcm_buffers[buffer].offset, CELL_GCM_LOCATION_LOCAL);
+	rsx::tiled_region buffer_region = get_tiled_address(display_buffers[buffer].offset, CELL_GCM_LOCATION_LOCAL);
 	u32 absolute_address = buffer_region.address + buffer_region.base;
 	gl::texture *render_target_texture = m_rtts.get_texture_from_render_target_if_applicable(absolute_address);
 

rpcs3/Emu/RSX/RSXThread.cpp

@@ -50,34 +50,36 @@ namespace rsx
 			}
 
 			fmt::throw_exception("GetAddress(offset=0x%x, location=0x%x): RSXIO memory not mapped" HERE, offset, location);
-
-				//if (fxm::get<GSRender>()->strict_ordering[offset >> 20])
-				//{
-				//	_mm_mfence(); // probably doesn't have any effect on current implementation
-				//}
 		}
 
-			case CELL_GCM_CONTEXT_DMA_TO_MEMORY_GET_REPORT:
+		case CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_LOCAL:
-				return 0x100000 + offset; // TODO: Properly implement
+			return 0x40301400 + offset;
 
 		case CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN:
-				return 0x800 + offset;	// TODO: Properly implement
+		{
+			if (u32 result = RSXIOMem.RealAddr(0x0e000000 + offset))
+			{
+				return result;
+			}
+
+			fmt::throw_exception("GetAddress(offset=0x%x, location=0x%x): RSXIO memory not mapped" HERE, offset, location);
+		}
 
 		case CELL_GCM_CONTEXT_DMA_TO_MEMORY_GET_NOTIFY0:
-				return 0x40 + offset; // TODO: Properly implement
+			fmt::throw_exception("Unimplemented CELL_GCM_CONTEXT_DMA_TO_MEMORY_GET_NOTIFY0 (offset=0x%x, location=0x%x)" HERE, offset, location);
 
 		case CELL_GCM_CONTEXT_DMA_NOTIFY_MAIN_0:
 			fmt::throw_exception("Unimplemented CELL_GCM_CONTEXT_DMA_NOTIFY_MAIN_0 (offset=0x%x, location=0x%x)" HERE, offset, location);
 
 		case CELL_GCM_CONTEXT_DMA_SEMAPHORE_RW:
 		case CELL_GCM_CONTEXT_DMA_SEMAPHORE_R:
-				return 0x100 + offset; // TODO: Properly implement
+			return 0x40300000 + offset;
 
 		case CELL_GCM_CONTEXT_DMA_DEVICE_RW:
-				fmt::throw_exception("Unimplemented CELL_GCM_CONTEXT_DMA_DEVICE_RW (offset=0x%x, location=0x%x)" HERE, offset, location);
+			return 0x40000000 + offset;
 
 		case CELL_GCM_CONTEXT_DMA_DEVICE_R:
-				fmt::throw_exception("Unimplemented CELL_GCM_CONTEXT_DMA_DEVICE_R (offset=0x%x, location=0x%x)" HERE, offset, location);
+			return 0x40000000 + offset;
 
 		default:
 			fmt::throw_exception("Invalid location (offset=0x%x, location=0x%x)" HERE, offset, location);
@@ -1069,9 +1071,9 @@ namespace rsx
 		rsx::method_registers.reset();
 	}
 
-	void thread::init(const u32 ioAddress, const u32 ioSize, const u32 ctrlAddress, const u32 localAddress)
+	void thread::init(u32 ioAddress, u32 ioSize, u32 ctrlAddress, u32 localAddress)
 	{
-		ctrl = vm::_ptr<CellGcmControl>(ctrlAddress);
+		ctrl = vm::_ptr<RsxDmaControl>(ctrlAddress);
 		this->ioAddress = ioAddress;
 		this->ioSize = ioSize;
 		local_mem_addr = localAddress;

rpcs3/Emu/RSX/RSXThread.h

@@ -22,6 +22,8 @@
 #include "Utilities/variant.hpp"
 #include "define_new_memleakdetect.h"
 
+#include "Emu/Cell/lv2/sys_rsx.h"
+
 extern u64 get_system_time();
 
 extern bool user_asked_for_frame_capture;
@@ -113,7 +115,7 @@ namespace rsx
 		std::vector<u32> element_push_buffer;
 
 	public:
-		CellGcmControl* ctrl = nullptr;
+		RsxDmaControl* ctrl = nullptr;
 
 		Timer timer_sync;
 
@@ -131,20 +133,17 @@ namespace rsx
 
 		u32 ioAddress, ioSize;
 		u32 flip_status;
-		int flip_mode;
 		int debug_level;
-		int frequency_mode;
 
-		u32 tiles_addr;
+		atomic_t<bool> requested_vsync{false};
-		u32 zculls_addr;
+
-		vm::ps3::ptr<CellGcmDisplayInfo> gcm_buffers = vm::null;
+		RsxDisplayInfo display_buffers[8];
-		u32 gcm_buffers_count;
+		u32 display_buffers_count{0};
-		u32 gcm_current_buffer;
+		u32 current_display_buffer{0};
 		u32 ctxt_addr;
 		u32 label_addr;
 
 		u32 local_mem_addr, main_mem_addr;
-		bool strict_ordering[0x1000];
 
 		bool m_rtts_dirty;
 		bool m_transform_constants_dirty;
@@ -345,7 +344,7 @@ namespace rsx
 
 	public:
 		void reset();
-		void init(const u32 ioAddress, const u32 ioSize, const u32 ctrlAddress, const u32 localAddress);
+		void init(u32 ioAddress, u32 ioSize, u32 ctrlAddress, u32 localAddress);
 
 		tiled_region get_tiled_address(u32 offset, u32 location);
 		GcmTileInfo *find_tile(u32 offset, u32 location);

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -2143,9 +2143,9 @@ void VKGSRender::flip(int buffer)
 
 	if (!resize_screen)
 	{
-		u32 buffer_width = gcm_buffers[buffer].width;
+		u32 buffer_width = display_buffers[buffer].width;
-		u32 buffer_height = gcm_buffers[buffer].height;
+		u32 buffer_height = display_buffers[buffer].height;
-		u32 buffer_pitch = gcm_buffers[buffer].pitch;
+		u32 buffer_pitch = display_buffers[buffer].pitch;
 
 		areai screen_area = coordi({}, { (int)buffer_width, (int)buffer_height });
 

rpcs3/Emu/RSX/gcm_enums.cpp

@@ -968,7 +968,7 @@ rsx::blit_engine::context_dma rsx::blit_engine::to_context_dma(u32 in)
 {
 	switch (in)
 	{
-	case CELL_GCM_CONTEXT_DMA_TO_MEMORY_GET_REPORT: return rsx::blit_engine::context_dma::to_memory_get_report;
+	case CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_LOCAL: return rsx::blit_engine::context_dma::to_memory_get_report;
 	case CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN: return rsx::blit_engine::context_dma::report_location_main;
 	case CELL_GCM_CONTEXT_DMA_MEMORY_HOST_BUFFER: return rsx::blit_engine::context_dma::memory_host_buffer;
 	}

rpcs3/Emu/RSX/gcm_enums.h

@@ -704,7 +704,7 @@ enum
 {
 	CELL_GCM_CONTEXT_DMA_MEMORY_FRAME_BUFFER = 0xFEED0000, // Local memory
 	CELL_GCM_CONTEXT_DMA_MEMORY_HOST_BUFFER = 0xFEED0001, // Main memory
-	CELL_GCM_CONTEXT_DMA_TO_MEMORY_GET_REPORT = 0x66626660,
+	CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_LOCAL = 0x66626660,
 	CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN = 0xBAD68000,
 	CELL_GCM_CONTEXT_DMA_NOTIFY_MAIN_0 = 0x6660420F,
 

rpcs3/Emu/RSX/rsx_decode.h

@@ -1988,6 +1988,32 @@ struct registers_decoder<NV4097_SET_TRANSFORM_PROGRAM_START>
 	}
 };
 
+template<>
+struct registers_decoder<NV406E_SET_CONTEXT_DMA_SEMAPHORE>
+{
+	struct decoded_type
+	{
+	private:
+		union
+		{
+			u32 raw_value;
+		} m_data;
+	public:
+		decoded_type(u32 raw_value) { m_data.raw_value = raw_value; }
+
+		u32 context_dma() const
+		{
+			return m_data.raw_value;
+		}
+	};
+
+	static std::string dump(decoded_type &&decoded_values)
+	{
+		return "NV406E semaphore: context = " + std::to_string(decoded_values.context_dma());
+	}
+};
+
+
 template<>
 struct registers_decoder<NV406E_SEMAPHORE_OFFSET>
 {
@@ -4600,7 +4626,7 @@ constexpr std::integer_sequence<u32, NV4097_SET_VIEWPORT_HORIZONTAL, NV4097_SET_
 	   NV4097_SET_VERTEX_ATTRIB_OUTPUT_MASK, NV4097_SET_SHADER_CONTROL,
 	   NV4097_SET_VERTEX_DATA_BASE_OFFSET, NV4097_SET_INDEX_ARRAY_ADDRESS,
 	   NV4097_SET_VERTEX_DATA_BASE_INDEX, NV4097_SET_SHADER_PROGRAM,
-	   NV4097_SET_TRANSFORM_PROGRAM_START, NV406E_SEMAPHORE_OFFSET, NV4097_SET_SEMAPHORE_OFFSET,
+	   NV4097_SET_TRANSFORM_PROGRAM_START, NV406E_SET_CONTEXT_DMA_SEMAPHORE, NV406E_SEMAPHORE_OFFSET, NV4097_SET_SEMAPHORE_OFFSET,
 	   NV3089_IMAGE_IN_OFFSET, NV3062_SET_OFFSET_DESTIN, NV309E_SET_OFFSET, NV3089_DS_DX, NV3089_DT_DY,
 	   NV0039_PITCH_IN, NV0039_PITCH_OUT, NV0039_LINE_LENGTH_IN, NV0039_LINE_COUNT, NV0039_OFFSET_OUT,
 	   NV0039_OFFSET_IN, NV4097_SET_VERTEX_ATTRIB_INPUT_MASK, NV4097_SET_FREQUENCY_DIVIDER_OPERATION,

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -60,9 +60,13 @@ namespace rsx
 
 		void semaphore_acquire(thread* rsx, u32 _reg, u32 arg)
 		{
-			//TODO: dma
+			const u32 addr = get_address(method_registers.semaphore_offset_406e(), method_registers.semaphore_context_dma_406e());
-			while (vm::ps3::read32(rsx->label_addr + method_registers.semaphore_offset_406e()) != arg)
+			while (vm::ps3::read32(addr) != arg)
 			{
+				// todo: LLE: why does this one keep hanging? is it vsh system semaphore? whats actually pushing this to the command buffer?!
+				if (addr == 0x40000030)
+					break;
+
 				if (Emu.IsStopped())
 					break;
 
@@ -72,8 +76,8 @@ namespace rsx
 
 		void semaphore_release(thread* rsx, u32 _reg, u32 arg)
 		{
-			//TODO: dma
+			const u32 addr = get_address(method_registers.semaphore_offset_406e(), method_registers.semaphore_context_dma_406e());
-			vm::ps3::write32(rsx->label_addr + method_registers.semaphore_offset_406e(), arg);
+			vm::ps3::write32(addr, arg);
 		}
 	}
 
@@ -100,8 +104,12 @@ namespace rsx
 				//
 			}
 
-			//TODO: dma
+			const u32 index = method_registers.semaphore_offset_4097() >> 4;
-			vm::ps3::write32(rsx->label_addr + method_registers.semaphore_offset_4097(), arg);
+
+			auto& sema = vm::ps3::_ref<RsxReports>(rsx->label_addr);
+			sema.semaphore[index].val = arg;
+			sema.semaphore[index].pad = 0;
+			sema.semaphore[index].timestamp = rsx->timestamp();
 		}
 
 		void back_end_write_semaphore_release(thread* rsx, u32 _reg, u32 arg)
@@ -111,9 +119,13 @@ namespace rsx
 				//
 			}
 
-			//TODO: dma
+			const u32 index = method_registers.semaphore_offset_4097() >> 4;
-			vm::ps3::write32(rsx->label_addr + method_registers.semaphore_offset_4097(),
+			u32 val = (arg & 0xff00ff00) | ((arg & 0xff) << 16) | ((arg >> 16) & 0xff);
-				(arg & 0xff00ff00) | ((arg & 0xff) << 16) | ((arg >> 16) & 0xff));
+
+			auto& sema = vm::ps3::_ref<RsxReports>(rsx->label_addr);
+			sema.semaphore[index].val = val;
+			sema.semaphore[index].pad = 0;
+			sema.semaphore[index].timestamp = rsx->timestamp();
 		}
 
 		template<u32 id, u32 index, int count, typename type>
@@ -323,10 +335,9 @@ namespace rsx
 
 			switch (report_dma)
 			{
-			case blit_engine::context_dma::to_memory_get_report: location = CELL_GCM_LOCATION_LOCAL; break;
+			case blit_engine::context_dma::to_memory_get_report: location = CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_LOCAL; break;
-			case blit_engine::context_dma::report_location_main:
+			case blit_engine::context_dma::report_location_main: location = CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_MAIN; break;
-			case blit_engine::context_dma::memory_host_buffer: 
+			case blit_engine::context_dma::memory_host_buffer: location = CELL_GCM_CONTEXT_DMA_MEMORY_HOST_BUFFER; break;
-				location = CELL_GCM_LOCATION_MAIN; break;
 			default:
 				LOG_WARNING(RSX, "nv4097::get_report: bad report dma: 0x%x", (u8)report_dma);
 				return;
@@ -339,12 +350,14 @@ namespace rsx
 			switch (type)
 			{
 			case CELL_GCM_ZPASS_PIXEL_CNT:
+				// todo: actual zculling, here we just report max, which seems to be enough for most games, but causes them to render *everything*
+				result->value = 0xFFFFFFFF;
+				break;
 			case CELL_GCM_ZCULL_STATS:
 			case CELL_GCM_ZCULL_STATS1:
 			case CELL_GCM_ZCULL_STATS2:
 			case CELL_GCM_ZCULL_STATS3:
 				result->value = 0;
-				LOG_WARNING(RSX, "NV4097_GET_REPORT: Unimplemented type %d", type);
 				break;
 
 			default:
@@ -352,8 +365,8 @@ namespace rsx
 				LOG_ERROR(RSX, "NV4097_GET_REPORT: Bad type %d", type);
 				break;
 			}
-
+			// This padding is needed to be set to 0, as games may use it for sync
-			//result->padding = 0;
+			result->padding = 0;
 		}
 
 		void clear_report_value(thread* rsx, u32 _reg, u32 arg)
@@ -514,14 +527,14 @@ namespace rsx
 				//HACK: it's extension of the flip-hack. remove this when textures cache would be properly implemented
 				for (int i = 0; i < rsx::limits::color_buffers_count; ++i)
 				{
-					u32 begin = rsx->gcm_buffers[i].offset;
+					u32 begin = rsx->display_buffers[i].offset;
 
 					if (dst_offset < begin || !begin)
 					{
 						continue;
 					}
 
-					if (rsx->gcm_buffers[i].width < 720 || rsx->gcm_buffers[i].height < 480)
+					if (rsx->display_buffers[i].width < 720 || rsx->display_buffers[i].height < 480)
 					{
 						continue;
 					}
@@ -531,7 +544,7 @@ namespace rsx
 						return;
 					}
 
-					u32 end = begin + rsx->gcm_buffers[i].height * rsx->gcm_buffers[i].pitch;
+					u32 end = begin + rsx->display_buffers[i].height * rsx->display_buffers[i].pitch;
 
 					if (dst_offset < end)
 					{
@@ -862,14 +875,13 @@ namespace rsx
 			rsx->timer_sync.Start();
 		}
 		
-		rsx->gcm_current_buffer = arg;
+		rsx->current_display_buffer = arg;
 		rsx->flip(arg);
 		// After each flip PS3 system is executing a routine that changes registers value to some default.
 		// Some game use this default state (SH3).
 		rsx->reset();
 
 		rsx->last_flip_time = get_system_time() - 1000000;
-		rsx->gcm_current_buffer = arg;
 		rsx->flip_status = CELL_GCM_DISPLAY_FLIP_STATUS_DONE;
 
 		if (rsx->flip_handler)
@@ -971,7 +983,7 @@ namespace rsx
 		registers[NV4097_SET_SURFACE_FORMAT] = (8 << 0) | (2 << 5) | (0 << 12) | (1 << 16) | (1 << 24);
 
 		// rsx dma initial values
-		registers[NV4097_SET_CONTEXT_DMA_REPORT] = CELL_GCM_CONTEXT_DMA_TO_MEMORY_GET_REPORT;
+		registers[NV4097_SET_CONTEXT_DMA_REPORT] = CELL_GCM_CONTEXT_DMA_REPORT_LOCATION_LOCAL;
 		registers[NV406E_SET_CONTEXT_DMA_SEMAPHORE] = CELL_GCM_CONTEXT_DMA_SEMAPHORE_RW;
 		registers[NV3062_SET_CONTEXT_DMA_IMAGE_DESTIN] = CELL_GCM_CONTEXT_DMA_MEMORY_FRAME_BUFFER;
 		registers[NV309E_SET_CONTEXT_DMA_IMAGE] = CELL_GCM_CONTEXT_DMA_MEMORY_FRAME_BUFFER;

rpcs3/Emu/RSX/rsx_methods.h

@@ -889,6 +889,11 @@ namespace rsx
 			return decode<NV4097_SET_BEGIN_END>().primitive();
 		}
 
+		u32 semaphore_context_dma_406e() const
+		{
+			return decode<NV406E_SET_CONTEXT_DMA_SEMAPHORE>().context_dma();
+		}
+
 		u32 semaphore_offset_406e() const
 		{
 			return decode<NV406E_SEMAPHORE_OFFSET>().semaphore_offset();

rpcs3/rpcs3qt/rsx_debugger.cpp

@@ -411,9 +411,7 @@ void Buffer::mouseDoubleClickEvent(QMouseEvent* event)
 			return;
 		}
 
-		const auto buffers = render->gcm_buffers;
+		const auto buffers = render->display_buffers;
-		if (!buffers)
-			return;
 
 		// TODO: Is there any better way to choose the color buffers
 #define SHOW_BUFFER(id) \
@@ -705,12 +703,9 @@ void rsx_debugger::GetBuffers()
 
 	// Draw Buffers
 	// TODO: Currently it only supports color buffers
-	for (u32 bufferId=0; bufferId < render->gcm_buffers_count; bufferId++)
+	for (u32 bufferId=0; bufferId < render->display_buffers_count; bufferId++)
 	{
-		if(!vm::check_addr(render->gcm_buffers.addr()))
+		auto buffers = render->display_buffers;
-			continue;
-
-		auto buffers = render->gcm_buffers;
 		u32 RSXbuffer_addr = render->local_mem_addr + buffers[bufferId].offset;
 
 		if(!vm::check_addr(RSXbuffer_addr))