rsx: Implement deferred transform constants load

rpcs3/Emu/CMakeLists.txt

@@ -486,6 +486,7 @@ target_sources(rpcs3_emu PRIVATE
     RSX/GL/upscalers/fsr1/fsr_pass.cpp
     RSX/GSRender.cpp
     RSX/Null/NullGSRender.cpp
+    RSX/NV47/FW/draw_call.cpp
     RSX/NV47/FW/reg_context.cpp
     RSX/NV47/HW/common.cpp
     RSX/NV47/HW/nv0039.cpp

rpcs3/Emu/RSX/NV47/FW/draw_call.cpp

@@ -0,0 +1,134 @@
+#include "stdafx.h"
+#include "draw_call.hpp"
+
+#include "Emu/RSX/rsx_methods.h" // FIXME
+#include "Emu/RSX/rsx_utils.h"
+#include "Emu/RSX/Common/BufferUtils.h"
+#include "Emu/RSX/NV47/HW/context.h"
+#include "Emu/RSX/NV47/HW/nv4097.h"
+
+// Always import this after other HW definitions
+#include "Emu/RSX/NV47/HW/context_accessors.define.h"
+
+#include <util/serialization.hpp>
+
+namespace rsx
+{
+	void draw_clause::operator()(utils::serial& ar)
+	{
+		ar(draw_command_ranges, draw_command_barriers, current_range_index, primitive, command, is_immediate_draw, is_disjoint_primitive, primitive_barrier_enable, inline_vertex_array);
+	}
+
+	void draw_clause::insert_command_barrier(command_barrier_type type, u32 arg, u32 index)
+	{
+		ensure(!draw_command_ranges.empty());
+
+		auto _do_barrier_insert = [this](barrier_t&& val)
+		{
+			if (draw_command_barriers.empty() || draw_command_barriers.back() < val)
+			{
+				draw_command_barriers.push_back(val);
+				return;
+			}
+
+			for (auto it = draw_command_barriers.begin(); it != draw_command_barriers.end(); it++)
+			{
+				if (*it < val)
+				{
+					continue;
+				}
+
+				draw_command_barriers.insert(it, val);
+				break;
+			}
+		};
+
+		if (type == primitive_restart_barrier)
+		{
+			// Rasterization flow barrier
+			const auto& last = draw_command_ranges[current_range_index];
+			const auto address = last.first + last.count;
+
+			_do_barrier_insert({ current_range_index, 0, address, index, arg, 0, type });
+		}
+		else
+		{
+			// Execution dependency barrier. Requires breaking the current draw call sequence and start another.
+			if (draw_command_ranges.back().count > 0)
+			{
+				append_draw_command({});
+			}
+			else
+			{
+				// In case of back-to-back modifiers, do not add duplicates
+				current_range_index = draw_command_ranges.size() - 1;
+			}
+
+			_do_barrier_insert({ current_range_index, rsx::get_shared_tag(), ~0u, index, arg, 0, type });
+			last_execution_barrier_index = current_range_index;
+		}
+	}
+
+	void draw_clause::reset(primitive_type type)
+	{
+		current_range_index = ~0u;
+		last_execution_barrier_index = 0;
+
+		command = draw_command::none;
+		primitive = type;
+		primitive_barrier_enable = false;
+
+		draw_command_ranges.clear();
+		draw_command_barriers.clear();
+		inline_vertex_array.clear();
+
+		is_disjoint_primitive = is_primitive_disjointed(primitive);
+	}
+
+	u32 draw_clause::execute_pipeline_dependencies(context* ctx) const
+	{
+		u32 result = 0u;
+		for (;
+			current_barrier_it != draw_command_barriers.end() && current_barrier_it->draw_id == current_range_index;
+			current_barrier_it++)
+		{
+			const auto& barrier = *current_barrier_it;
+			switch (barrier.type)
+			{
+			case primitive_restart_barrier:
+				break;
+			case index_base_modifier_barrier:
+				// Change index base offset
+				REGS(ctx)->decode(NV4097_SET_VERTEX_DATA_BASE_INDEX, barrier.arg);
+				result |= index_base_changed;
+				break;
+			case vertex_base_modifier_barrier:
+				// Change vertex base offset
+				REGS(ctx)->decode(NV4097_SET_VERTEX_DATA_BASE_OFFSET, barrier.arg);
+				result |= vertex_base_changed;
+				break;
+			case vertex_array_offset_modifier_barrier:
+				// Change vertex array offset
+				REGS(ctx)->decode(NV4097_SET_VERTEX_DATA_ARRAY_OFFSET + barrier.index, barrier.arg);
+				result |= vertex_arrays_changed;
+				break;
+			case transform_constant_load_modifier_barrier:
+				// Change the transform load target. Does not change result mask.
+				REGS(ctx)->decode(NV4097_SET_TRANSFORM_PROGRAM_LOAD, barrier.arg);
+				break;
+			case transform_constant_update_barrier:
+				// Update transform constants
+				// REGS(ctx)->decode(NV4097_SET_TRANSFORM_CONSTANT + barrier.index, barrier.arg); // This statement technically does the right thing but has no consequence other than wasting perf.
+				// FIXME: Batching
+				nv4097::set_transform_constant::decode_one(ctx, NV4097_SET_TRANSFORM_CONSTANT + barrier.index, barrier.arg);
+				result |= transform_constants_changed;
+				break;
+			default:
+				fmt::throw_exception("Unreachable");
+			}
+		}
+
+		return result;
+	}
+}
+

rpcs3/Emu/RSX/NV47/FW/draw_call.hpp

@@ -21,6 +21,9 @@ namespace rsx
 		// Location of last execution barrier
 		u32 last_execution_barrier_index{};
 
+		// Draw-time iterator to the draw_command_barriers struct
+		mutable simple_array<barrier_t>::iterator current_barrier_it;
+
 		// Helper functions
 		// Add a new draw command
 		void append_draw_command(const draw_range_t& range)
@@ -55,6 +58,7 @@ namespace rsx
 		bool is_immediate_draw{};          // Set if part of the draw is submitted via push registers
 		bool is_disjoint_primitive{};      // Set if primitive type does not rely on adjacency information
 		bool primitive_barrier_enable{};   // Set once to signal that a primitive restart barrier can be inserted
+		bool is_rendering{};               // Set while we're actually pushing the draw calls to host GPU
 
 		simple_array<u32> inline_vertex_array{};
 
@@ -111,7 +115,8 @@ namespace rsx
 					if (draw_command_ranges[index].first == first &&
 						draw_command_ranges[index].count == count)
 					{
-						// Duplicate entry? WTF!
+						// Duplicate entry. Usually this indicates a botched instancing setup.
+						rsx_log.error("Duplicate draw request. Start=%u, Count=%u", first, count);
 						return;
 					}
 
@@ -206,12 +211,13 @@ namespace rsx
 				: primitive_class::non_polygon;
 		}
 
-
 		void reset(rsx::primitive_type type);
 
 		void begin()
 		{
 			current_range_index = 0;
+			current_barrier_it = draw_command_barriers.begin();
+			is_rendering = true;
 		}
 
 		void end()
@@ -225,6 +231,7 @@ namespace rsx
 			if (current_range_index >= draw_command_ranges.size())
 			{
 				current_range_index = 0;
+				is_rendering = false;
 				return false;
 			}
 
@@ -233,6 +240,7 @@ namespace rsx
 				// Dangling execution barrier
 				ensure(current_range_index > 0 && (current_range_index + 1) == draw_command_ranges.size());
 				current_range_index = 0;
+				is_rendering = false;
 				return false;
 			}
 

rpcs3/Emu/RSX/NV47/FW/draw_call.inc.h

@@ -17,7 +17,9 @@ namespace rsx
 		primitive_restart_barrier,
 		vertex_base_modifier_barrier,
 		index_base_modifier_barrier,
-		vertex_array_offset_modifier_barrier
+		vertex_array_offset_modifier_barrier,
+		transform_constant_load_modifier_barrier,
+		transform_constant_update_barrier
 	};
 
 	enum command_execution_flags : u32
@@ -25,6 +27,7 @@ namespace rsx
 		vertex_base_changed = (1 << 0),
 		index_base_changed = (1 << 1),
 		vertex_arrays_changed = (1 << 2),
+		transform_constants_changed = (1 << 3)
 	};
 
 	enum class primitive_class

rpcs3/Emu/RSX/NV47/HW/nv4097.cpp

@@ -20,8 +20,29 @@ namespace rsx
 			RSX(ctx)->m_graphics_state |= rsx::pipeline_state::fragment_program_ucode_dirty;
 		}
 
-		void set_transform_constant::impl(context* ctx, u32 reg, u32 /*arg*/)
+		void set_transform_constant::decode_one(context* ctx, u32 reg, u32 arg)
 		{
+			const u32 index = reg - NV4097_SET_TRANSFORM_CONSTANT;
+			const u32 constant_id = index / 4;
+			const u8 subreg = index % 4;
+			const u32 load = REGS(ctx)->transform_constant_load();
+
+			REGS(ctx)->transform_constants[load + constant_id][subreg] = arg;
+		}
+
+		void set_transform_constant::impl(context* ctx, u32 reg, u32 arg)
+		{
+			if (RSX(ctx)->in_begin_end && !REGS(ctx)->current_draw_clause.empty())
+			{
+				// Updating constants mid-draw is messy. Push attr barrier.
+				REGS(ctx)->current_draw_clause.insert_command_barrier(
+					rsx::transform_constant_update_barrier,
+					arg,
+					reg - NV4097_SET_TRANSFORM_CONSTANT
+				);
+				return;
+			}
+
 			const u32 index = reg - NV4097_SET_TRANSFORM_CONSTANT;
 			const u32 constant_id = index / 4;
 			const u8 subreg = index % 4;
@@ -439,6 +460,11 @@ namespace rsx
 			REGS(ctx)->decode(reg, REGS(ctx)->latch);
 		}
 
+		void set_transform_constant_load(context* ctx, u32 reg, u32 arg)
+		{
+			util::push_draw_parameter_change(ctx, rsx::transform_constant_load_modifier_barrier, reg, arg);
+		}
+
 		///// Reports
 
 		void get_report(context* ctx, u32 /*reg*/, u32 arg)

rpcs3/Emu/RSX/NV47/HW/nv4097.h

@@ -83,6 +83,8 @@ namespace rsx
 
 		void set_blend_factor(context* ctx, u32 reg, u32 arg);
 
+		void set_transform_constant_load(context* ctx, u32 reg, u32 arg);
+
 #define RSX(ctx) ctx->rsxthr
 #define REGS(ctx) (&rsx::method_registers)
 
@@ -197,6 +199,8 @@ namespace rsx
 		struct set_transform_constant
 		{
 			static void impl(context* ctx, u32 reg, u32 arg);
+
+			static void decode_one(context* ctx, u32 reg, u32 arg);
 		};
 
 		struct set_transform_program

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -39,11 +39,6 @@ namespace rsx
 		rsx_log.trace("RSX method 0x%x (arg=0x%x)", reg << 2, arg);
 	}
 
-	namespace nv0039
-	{
-		
-	}
-
 	void flip_command(context* ctx, u32, u32 arg)
 	{
 		ensure(RSX(ctx)->isHLE);
@@ -1217,113 +1212,6 @@ namespace rsx
 		return registers[reg] == value;
 	}
 
-	void draw_clause::operator()(utils::serial& ar)
-	{
-		ar(draw_command_ranges, draw_command_barriers, current_range_index, primitive, command, is_immediate_draw, is_disjoint_primitive, primitive_barrier_enable, inline_vertex_array);
-	}
-
-	void draw_clause::insert_command_barrier(command_barrier_type type, u32 arg, u32 index)
-	{
-		ensure(!draw_command_ranges.empty());
-
-		auto _do_barrier_insert = [this](barrier_t&& val)
-		{
-			if (draw_command_barriers.empty() || draw_command_barriers.back() < val)
-			{
-				draw_command_barriers.push_back(val);
-				return;
-			}
-
-			for (auto it = draw_command_barriers.begin(); it != draw_command_barriers.end(); it++)
-			{
-				if (*it < val)
-				{
-					continue;
-				}
-
-				draw_command_barriers.insert(it, val);
-				break;
-			}
-		};
-
-		if (type == primitive_restart_barrier)
-		{
-			// Rasterization flow barrier
-			const auto& last = draw_command_ranges[current_range_index];
-			const auto address = last.first + last.count;
-
-			_do_barrier_insert({ current_range_index, 0, address, index, arg, 0, type });
-		}
-		else
-		{
-			// Execution dependency barrier. Requires breaking the current draw call sequence and start another.
-			if (draw_command_ranges.back().count > 0)
-			{
-				append_draw_command({});
-			}
-			else
-			{
-				// In case of back-to-back modifiers, do not add duplicates
-				current_range_index = draw_command_ranges.size() - 1;
-			}
-
-			_do_barrier_insert({ current_range_index, rsx::get_shared_tag(), ~0u, index, arg, 0, type });
-			last_execution_barrier_index = current_range_index;
-		}
-	}
-
-	void draw_clause::reset(primitive_type type)
-	{
-		current_range_index = ~0u;
-		last_execution_barrier_index = 0;
-
-		command = draw_command::none;
-		primitive = type;
-		primitive_barrier_enable = false;
-
-		draw_command_ranges.clear();
-		draw_command_barriers.clear();
-		inline_vertex_array.clear();
-
-		is_disjoint_primitive = is_primitive_disjointed(primitive);
-	}
-
-	u32 draw_clause::execute_pipeline_dependencies(context* ctx) const
-	{
-		u32 result = 0;
-
-		for (const auto &barrier : draw_command_barriers)
-		{
-			if (barrier.draw_id != current_range_index)
-				continue;
-
-			switch (barrier.type)
-			{
-			case primitive_restart_barrier:
-				break;
-			case index_base_modifier_barrier:
-				// Change index base offset
-				REGS(ctx)->decode(NV4097_SET_VERTEX_DATA_BASE_INDEX, barrier.arg);
-				result |= index_base_changed;
-				break;
-			case vertex_base_modifier_barrier:
-				// Change vertex base offset
-				REGS(ctx)->decode(NV4097_SET_VERTEX_DATA_BASE_OFFSET, barrier.arg);
-				result |= vertex_base_changed;
-				break;
-			case vertex_array_offset_modifier_barrier:
-				// Change vertex array offset
-				REGS(ctx)->decode(NV4097_SET_VERTEX_DATA_ARRAY_OFFSET + barrier.index, barrier.arg);
-				result |= vertex_arrays_changed;
-				break;
-			default:
-				fmt::throw_exception("Unreachable");
-			}
-		}
-
-		return result;
-	}
-
 	namespace method_detail
 	{
 		template <u32 Id, u32 Step, u32 Count, template<u32> class T, u32 Index = 0>
@@ -1751,6 +1639,7 @@ namespace rsx
 		bind_range<NV4097_SET_VERTEX_DATA4F_M, 1, 64, nv4097::set_vertex_data4f_m>();
 		bind_range<NV4097_SET_VERTEX_DATA2S_M, 1, 16, nv4097::set_vertex_data2s_m>();
 		bind_range<NV4097_SET_VERTEX_DATA4S_M, 1, 32, nv4097::set_vertex_data4s_m>();
+		bind(NV4097_SET_TRANSFORM_CONSTANT_LOAD, nv4097::set_transform_constant_load);
 		bind_array(NV4097_SET_TRANSFORM_CONSTANT, 1, 32, nv4097::set_transform_constant::impl);
 		bind_array(NV4097_SET_TRANSFORM_PROGRAM, 1, 32, nv4097::set_transform_program::impl);
 		bind(NV4097_GET_REPORT, nv4097::get_report);

rpcs3/emucore.vcxproj

@@ -97,6 +97,7 @@
     <ClCompile Include="Emu\perf_monitor.cpp" />
     <ClCompile Include="Emu\RSX\Common\texture_cache.cpp" />
     <ClCompile Include="Emu\RSX\Core\RSXContext.cpp" />
+    <ClCompile Include="Emu\RSX\NV47\FW\draw_call.cpp" />
     <ClCompile Include="Emu\RSX\NV47\FW\reg_context.cpp" />
     <ClCompile Include="Emu\RSX\NV47\HW\common.cpp" />
     <ClCompile Include="Emu\RSX\NV47\HW\nv0039.cpp" />

rpcs3/emucore.vcxproj.filters

@@ -1246,6 +1246,9 @@
     <ClCompile Include="Emu\RSX\NV47\FW\reg_context.cpp">
       <Filter>Emu\GPU\RSX\NV47\FW</Filter>
     </ClCompile>
+    <ClCompile Include="Emu\RSX\NV47\FW\draw_call.cpp">
+      <Filter>Emu\GPU\RSX\NV47\FW</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Crypto\aes.h">