rsx: Add support for batched multidraw

gl: Fix multidraw [WIP]
rsx: Ignore vertex base when data source is generated using arithmetic
vk: Check pending flag before doing fence poke
vk/gl: Fix for inlined array and immediate draws
rsx: Collapse joined draws when batching

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -30,6 +30,8 @@ GLGSRender::GLGSRender() : GSRender()
 		m_vertex_cache.reset(new gl::null_vertex_cache());
 	else
 		m_vertex_cache.reset(new gl::weak_vertex_cache());
+
+	supports_multidraw = true;
 }
 
 extern CellGcmContextData current_context;
@@ -510,21 +512,85 @@ void GLGSRender::end()
 		m_program->validate();
 	}
 
+	const GLenum draw_mode = gl::draw_mode(rsx::method_registers.current_draw_clause.primitive);
+	bool single_draw = rsx::method_registers.current_draw_clause.first_count_commands.size() <= 1 || rsx::method_registers.current_draw_clause.is_disjoint_primitive;
+
 	if (indexed_draw_info)
 	{
 		const GLenum index_type = std::get<0>(indexed_draw_info.value());
 		const u32 index_offset = std::get<1>(indexed_draw_info.value());
 
-		if (__glcheck gl_state.enable(rsx::method_registers.restart_index_enabled(), GL_PRIMITIVE_RESTART))
+		if (gl_state.enable(rsx::method_registers.restart_index_enabled(), GL_PRIMITIVE_RESTART))
 		{
-			__glcheck glPrimitiveRestartIndex((index_type == GL_UNSIGNED_SHORT)? 0xffff: 0xffffffff);
+			glPrimitiveRestartIndex((index_type == GL_UNSIGNED_SHORT)? 0xffff: 0xffffffff);
 		}
 
-		__glcheck glDrawElements(gl::draw_mode(rsx::method_registers.current_draw_clause.primitive), vertex_draw_count, index_type, (GLvoid *)(uintptr_t)index_offset);
+		if (single_draw)
+		{
+			glDrawElements(draw_mode, vertex_draw_count, index_type, (GLvoid *)(uintptr_t)index_offset);
+		}
+		else
+		{
+			std::vector<GLsizei> counts;
+			std::vector<const GLvoid*> offsets;
+
+			const auto draw_count = rsx::method_registers.current_draw_clause.first_count_commands.size();
+			const u32 type_scale = (index_type == GL_UNSIGNED_SHORT) ? 1 : 2;
+			uintptr_t index_ptr = index_offset;
+
+			counts.reserve(draw_count);
+			offsets.reserve(draw_count);
+
+			for (const auto &range : rsx::method_registers.current_draw_clause.first_count_commands)
+			{
+				const auto index_size = get_index_count(rsx::method_registers.current_draw_clause.primitive, range.second);
+				counts.push_back(index_size);
+				offsets.push_back((const GLvoid*)index_ptr);
+
+				index_ptr += (index_size << type_scale);
+			}
+
+			for (int i = 0; i < draw_count; ++i)
+			{
+				if (counts[i] > 0)
+					glDrawElements(draw_mode, counts[i], index_type, offsets[i]);
+			}
+
+			//glMultiDrawElements(draw_mode, counts.data(), index_type, offsets.data(), (GLsizei)draw_count);
+		}
 	}
 	else
 	{
-		glDrawArrays(gl::draw_mode(rsx::method_registers.current_draw_clause.primitive), 0, vertex_draw_count);
+		if (single_draw)
+		{
+			glDrawArrays(draw_mode, 0, vertex_draw_count);
+		}
+		else
+		{
+			std::vector<GLint> firsts;
+			std::vector<GLsizei> counts;
+			const auto draw_count = rsx::method_registers.current_draw_clause.first_count_commands.size();
+
+			firsts.reserve(draw_count);
+			counts.reserve(draw_count);
+
+			u32 base_index = rsx::method_registers.current_draw_clause.first_count_commands.front().first;
+			for (const auto &range : rsx::method_registers.current_draw_clause.first_count_commands)
+			{
+				firsts.push_back(range.first - base_index);
+				counts.push_back(range.second);
+			}
+
+			///*
+			// TEST FOR DRIVER BUGS - AMD: SHAME, SHAME, SHAME
+			for (int i = 0; i < draw_count; i++)
+			{
+				if (counts[i] > 0)
+					glDrawArrays(draw_mode, firsts[i], counts[i]);
+			}//*/
+
+			//glMultiDrawArrays(draw_mode, firsts.data(), counts.data(), (GLsizei)draw_count);
+		}
 	}
 
 	m_attrib_ring_buffer->notify();
@@ -570,7 +636,7 @@ void GLGSRender::set_viewport()
 
 	//NOTE: window origin does not affect scissor region (probably only affects viewport matrix; already applied)
 	//See LIMBO [NPUB-30373] which uses shader window origin = top
-	__glcheck glScissor(scissor_x, scissor_y, scissor_w, scissor_h);
+	glScissor(scissor_x, scissor_y, scissor_w, scissor_h);
 	glEnable(GL_SCISSOR_TEST);
 }
 

rpcs3/Emu/RSX/GL/GLProcTable.h

@@ -168,6 +168,7 @@ OPENGL_PROC(PFNGLMAPBUFFERRANGEPROC, MapBufferRange);
 OPENGL_PROC(PFNGLBINDBUFFERRANGEPROC, BindBufferRange);
 OPENGL_PROC(PFNGLBINDBUFFERBASEPROC, BindBufferBase);
 
+OPENGL_PROC(PFNGLMULTIDRAWELEMENTSPROC, MultiDrawElements);
 OPENGL_PROC(PFNGLMULTIDRAWARRAYSPROC, MultiDrawArrays);
 
 OPENGL_PROC(PFNGLGETTEXTUREIMAGEEXTPROC, GetTextureImageEXT);

rpcs3/Emu/RSX/RSXThread.cpp

@@ -277,6 +277,21 @@ namespace rsx
 	{
 		rsx::method_registers.current_draw_clause.inline_vertex_array.resize(0);
 		in_begin_end = true;
+
+		switch (rsx::method_registers.current_draw_clause.primitive)
+		{
+		case rsx::primitive_type::line_loop:
+		case rsx::primitive_type::line_strip:
+		case rsx::primitive_type::polygon:
+		case rsx::primitive_type::quad_strip:
+		case rsx::primitive_type::triangle_fan:
+		case rsx::primitive_type::triangle_strip:
+			// Adjacency matters for these types
+			rsx::method_registers.current_draw_clause.is_disjoint_primitive = false;
+			break;
+		default:
+			rsx::method_registers.current_draw_clause.is_disjoint_primitive = true;
+		}
 	}
 
 	void thread::append_to_push_buffer(u32 attribute, u32 size, u32 subreg_index, vertex_base_type type, u32 value)
@@ -376,6 +391,65 @@ namespace rsx
 		// Raise priority above other threads
 		thread_ctrl::set_native_priority(1);
 
+		// Deferred calls are used to batch draws together
+		u32 deferred_primitive_type = 0;
+		u32 deferred_call_size = 0;
+		bool has_deferred_call = false;
+
+		auto flush_command_queue = [&]()
+		{
+			//TODO: Split first-count pairs if not consecutive
+			bool split_command = false;
+			std::vector <std::pair<u32, u32>> split_ranges;
+			auto first_count_cmds = method_registers.current_draw_clause.first_count_commands;
+
+			if (method_registers.current_draw_clause.first_count_commands.size() > 1)
+			{
+				u32 next = method_registers.current_draw_clause.first_count_commands.front().first;
+				u32 last_head = 0;
+
+				for (int n = 0; n < first_count_cmds.size(); ++n)
+				{
+					const auto &v = first_count_cmds[n];
+					if (v.first != next)
+					{
+						split_command = true;
+						split_ranges.push_back(std::make_pair(last_head, n));
+						last_head = n + 1;
+					}
+
+					next = v.first + v.second;
+				}
+
+				if (split_command)
+					split_ranges.push_back(std::make_pair(last_head, first_count_cmds.size() - 1));
+			}
+
+			if (!split_command)
+			{
+				methods[NV4097_SET_BEGIN_END](this, NV4097_SET_BEGIN_END, 0);
+			}
+			else
+			{
+				std::vector<std::pair<u32, u32>> tmp;
+				auto list_head = first_count_cmds.begin();
+
+				for (auto &range : split_ranges)
+				{
+					tmp.resize(range.second - range.first + 1);
+					std::copy(list_head + range.first, list_head + range.second, tmp.begin());
+
+					methods[NV4097_SET_BEGIN_END](this, NV4097_SET_BEGIN_END, deferred_primitive_type);
+					method_registers.current_draw_clause.first_count_commands = tmp;
+					methods[NV4097_SET_BEGIN_END](this, NV4097_SET_BEGIN_END, 0);
+				}
+			}
+
+			deferred_primitive_type = 0;
+			deferred_call_size = 0;
+			has_deferred_call = false;
+		};
+
 		// TODO: exit condition
 		while (!Emu.IsStopped())
 		{
@@ -387,6 +461,9 @@ namespace rsx
 
 			if (put == get || !Emu.IsRunning())
 			{
+				if (has_deferred_call)
+					flush_command_queue();
+
 				do_internal_task();
 				continue;
 			}
@@ -472,7 +549,92 @@ namespace rsx
 				u32 reg = ((cmd & RSX_METHOD_NON_INCREMENT_CMD_MASK) == RSX_METHOD_NON_INCREMENT_CMD) ? first_cmd : first_cmd + i;
 				u32 value = args[i];
 
-				//LOG_NOTICE(RSX, "%s(0x%x) = 0x%x", get_method_name(reg).c_str(), reg, value);
+				bool execute_method_call = true;
+
+				if (supports_multidraw)
+				{
+					//TODO: Make this cleaner
+					bool flush_commands_flag = has_deferred_call;
+
+					switch (reg)
+					{
+					case NV4097_SET_BEGIN_END:
+					{
+						// Hook; Allows begin to go through, but ignores end
+						if (value && value != deferred_primitive_type)
+							deferred_primitive_type = value;
+						else
+						{
+							deferred_call_size++;
+
+							// Combine all calls since the last one
+							auto &first_count = method_registers.current_draw_clause.first_count_commands;
+							if (first_count.size() > deferred_call_size)
+							{
+								const auto &batch_first_count = first_count[deferred_call_size - 1];
+								u32 count = batch_first_count.second;
+								u32 next = batch_first_count.first + count;
+
+								for (int n = deferred_call_size; n < first_count.size(); n++)
+								{
+									if (first_count[n].first != next)
+									{
+										LOG_ERROR(RSX, "Non-continous first-count range passed as one draw; will be split.");
+
+										first_count[deferred_call_size - 1].second = count;
+										deferred_call_size++;
+
+										count = first_count[deferred_call_size - 1].second;
+										next = first_count[deferred_call_size - 1].first + count;
+										continue;
+									}
+
+									count += first_count[n].second;
+									next += first_count[n].second;
+								}
+
+								first_count[deferred_call_size - 1].second = count;
+								first_count.resize(deferred_call_size);
+							}
+
+							has_deferred_call = true;
+							flush_commands_flag = false;
+							execute_method_call = false;
+						}
+
+						break;
+					}
+					// These commands do not alter the pipeline state and deferred calls can still be active
+					// TODO: Add more commands here
+					case NV4097_INVALIDATE_VERTEX_FILE:
+						flush_commands_flag = false;
+						break;
+					case NV4097_DRAW_ARRAYS:
+					{
+						const auto cmd = method_registers.current_draw_clause.command;
+						if (cmd != rsx::draw_command::array && cmd != rsx::draw_command::none)
+							break;
+
+						flush_commands_flag = false;
+						break;
+					}
+					case NV4097_DRAW_INDEX_ARRAY:
+					{
+						const auto cmd = method_registers.current_draw_clause.command;
+						if (cmd != rsx::draw_command::indexed && cmd != rsx::draw_command::none)
+							break;
+
+						flush_commands_flag = false;
+						break;
+					}
+					}
+
+					if (flush_commands_flag)
+					{
+						flush_command_queue();
+					}
+				}
+
 				method_registers.decode(reg, value);
 
 				if (capture_current_frame)
@@ -480,9 +642,12 @@ namespace rsx
 					frame_debug.command_queue.push_back(std::make_pair(reg, value));
 				}
 
-				if (auto method = methods[reg])
+				if (execute_method_call)
 				{
-					method(this, reg, value);
+					if (auto method = methods[reg])
+					{
+						method(this, reg, value);
+					}
 				}
 
 				if (invalid_command_interrupt_raised)
@@ -1534,7 +1699,7 @@ namespace rsx
 			for (const auto &block : layout.interleaved_blocks)
 			{
 				u32 unique_verts;
-				u32 vertex_base = first_vertex * block.attribute_stride;
+				u32 vertex_base = 0;
 
 				if (block.single_vertex)
 				{
@@ -1553,6 +1718,7 @@ namespace rsx
 				else
 				{
 					unique_verts = vertex_count;
+					vertex_base = first_vertex * block.attribute_stride;
 				}
 
 				const u32 data_size = block.attribute_stride * unique_verts;

rpcs3/Emu/RSX/RSXThread.h

@@ -144,6 +144,11 @@ namespace rsx
 		std::array<push_buffer_vertex_info, 16> vertex_push_buffers;
 		std::vector<u32> element_push_buffer;
 
+		s32 m_skip_frame_ctr = 0;
+		bool skip_frame = false;
+
+		bool supports_multidraw = false;
+
 	public:
 		RsxDmaControl* ctrl = nullptr;
 
@@ -183,9 +188,6 @@ namespace rsx
 		bool m_transform_constants_dirty;
 		bool m_textures_dirty[16];
 
-	protected:
-		s32 m_skip_frame_ctr = 0;
-		bool skip_frame = false;
 	protected:
 		std::array<u32, 4> get_color_surface_addresses() const;
 		u32 get_zeta_surface_address() const;

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -637,6 +637,8 @@ VKGSRender::VKGSRender() : GSRender()
 	}
 
 	m_current_frame = &frame_context_storage[0];
+
+	supports_multidraw = true;
 }
 
 VKGSRender::~VKGSRender()
@@ -1166,10 +1168,23 @@ void VKGSRender::end()
 	}
 
 	std::optional<std::tuple<VkDeviceSize, VkIndexType> > index_info = std::get<4>(upload_info);
+	bool single_draw = rsx::method_registers.current_draw_clause.first_count_commands.size() <= 1 || rsx::method_registers.current_draw_clause.is_disjoint_primitive;
+
 	if (!index_info)
 	{
-		const auto vertex_count = std::get<1>(upload_info);
-		vkCmdDraw(*m_current_command_buffer, vertex_count, 1, 0, 0);
+		if (single_draw)
+		{
+			const auto vertex_count = std::get<1>(upload_info);
+			vkCmdDraw(*m_current_command_buffer, vertex_count, 1, 0, 0);
+		}
+		else
+		{
+			const auto base_vertex = rsx::method_registers.current_draw_clause.first_count_commands.front().first;
+			for (const auto &range : rsx::method_registers.current_draw_clause.first_count_commands)
+			{
+				vkCmdDraw(*m_current_command_buffer, range.second, 1, range.first - base_vertex, 0);
+			}
+		}
 	}
 	else
 	{
@@ -1178,9 +1193,22 @@ void VKGSRender::end()
 		VkDeviceSize offset;
 
 		std::tie(offset, index_type) = index_info.value();
-
 		vkCmdBindIndexBuffer(*m_current_command_buffer, m_index_buffer_ring_info.heap->value, offset, index_type);
-		vkCmdDrawIndexed(*m_current_command_buffer, index_count, 1, 0, 0, 0);
+
+		if (single_draw)
+		{
+			vkCmdDrawIndexed(*m_current_command_buffer, index_count, 1, 0, 0, 0);
+		}
+		else
+		{
+			u32 first_vertex = 0;
+			for (const auto &range : rsx::method_registers.current_draw_clause.first_count_commands)
+			{
+				const auto verts = get_index_count(rsx::method_registers.current_draw_clause.primitive, range.second);
+				vkCmdDrawIndexed(*m_current_command_buffer, verts, 1, 0, first_vertex, 0);
+				first_vertex += verts;
+			}
+		}
 	}
 
 	vk::leave_uninterruptible();
@@ -1441,7 +1469,10 @@ void VKGSRender::flush_command_queue(bool hard_sync)
 
 		//Clear all command buffer statuses
 		for (auto &cb : m_primary_cb_list)
-			cb.poke();
+		{
+			if (cb.pending)
+				cb.poke();
+		}
 
 		m_last_flushable_cb = -1;
 		m_flush_commands = false;
@@ -1623,7 +1654,9 @@ void VKGSRender::do_local_task()
 	if (m_last_flushable_cb > -1)
 	{
 		auto cb = &m_primary_cb_list[m_last_flushable_cb];
-		cb->poke();
+
+		if (cb->pending)
+			cb->poke();
 
 		if (!cb->pending)
 			m_last_flushable_cb = -1;

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -300,7 +300,7 @@ namespace rsx
 		{
 			if (arg)
 			{
-				rsx::method_registers.current_draw_clause.first_count_commands.clear();
+				rsx::method_registers.current_draw_clause.first_count_commands.resize(0);
 				rsx::method_registers.current_draw_clause.command = draw_command::none;
 				rsx::method_registers.current_draw_clause.primitive = rsx::method_registers.primitive_mode();
 				rsxthr->begin();

rpcs3/Emu/RSX/rsx_methods.h

@@ -29,6 +29,7 @@ namespace rsx
 		draw_command command;
 
 		bool is_immediate_draw;
+		bool is_disjoint_primitive;
 
 		std::vector<u32> inline_vertex_array;