Merge pull request #1816 from RPCS3/WIP

OpenGL renderer: improved vertex attributes setup

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -307,27 +307,38 @@ void GLGSRender::end()
 
 	//setup textures
 	{
-		int texture_index = 0;
+		//int texture_index = 0;
 		for (int i = 0; i < rsx::limits::textures_count; ++i)
 		{
-			if (!textures[i].enabled())
-			{
-				continue;
-			}
-
 			int location;
 			if (m_program->uniforms.has_location("texture" + std::to_string(i), &location))
 			{
-				__glcheck glProgramUniform1i(m_program->id(), location, texture_index);
+				if (!textures[i].enabled())
-				__glcheck m_gl_textures[i].init(texture_index, textures[i]);
+				{
+					glActiveTexture(GL_TEXTURE0 + i);
+					glBindTexture(GL_TEXTURE_2D, 0);
+					glProgramUniform1i(m_program->id(), location, i);
 
-				texture_index++;
+					continue;
+				}
+
+				m_gl_textures[i].set_target(get_gl_target_for_texture(textures[i]));
+				__glcheck m_gl_texture_cache.upload_texture(i, textures[i], m_gl_textures[i], m_rtts);
+
+				__glcheck glProgramUniform1i(m_program->id(), location, i);
+				//__glcheck m_gl_textures[i].init(i, textures[i]);
+
+				//texture_index++;
 
 				if (m_program->uniforms.has_location("texture" + std::to_string(i) + "_cm", &location))
 				{
 					if (textures[i].format() & CELL_GCM_TEXTURE_UN)
 					{
-						//glProgramUniform4f(m_program->id(), location, textures[i].width(), textures[i].height(), textures[i].depth(), 1.0f);
+						u32 width = std::max<u32>(textures[i].width(), 1);
+						u32 height = std::max<u32>(textures[i].height(), 1);
+						u32 depth = std::max<u32>(textures[i].depth(), 1);
+
+						glProgramUniform4f(m_program->id(), location, 1.f / width, 1.f / height, 1.f / depth, 1.0f);
 					}
 				}
 			}
@@ -349,8 +360,6 @@ void GLGSRender::end()
 		*/
 	}
 
-	__glcheck 0;
-
 	u32 offset_in_index_buffer = set_vertex_buffer();
 	m_vao.bind();
 
@@ -430,8 +439,6 @@ void GLGSRender::set_viewport()
 	}
 
 	glEnable(GL_SCISSOR_TEST);
-
-	__glcheck 0;
 }
 
 void GLGSRender::on_init_thread()
@@ -685,8 +692,6 @@ bool GLGSRender::load_program()
 		auto mapping = m_uniform_ring_buffer.alloc_from_reserve(fragment_constants_size, m_uniform_buffer_offset_align);
 		fragment_constants_offset = mapping.second;
 
-		u32 buffer_offset = 0;
-
 		static const __m128i mask = _mm_set_epi8(
 			0xE, 0xF, 0xC, 0xD,
 			0xA, 0xB, 0x8, 0x9,
@@ -695,20 +700,23 @@ bool GLGSRender::load_program()
 
 		auto ucode = (const rsx::fragment_program::ucode_instr *)info.fragment_shader.decompiled->raw->ucode_ptr;
 
+		auto dst = (const rsx::fragment_program::ucode_instr *)mapping.first;
+
 		for (const auto& constant : info.fragment_shader.decompiled->constants)
 		{
-			const void *data = ucode + u32(constant.id / sizeof(rsx::fragment_program::ucode_instr));
+			const void *src = ucode + u32(constant.id / sizeof(*ucode));
-			const __m128i &vector = _mm_loadu_si128((const __m128i*)data);
-			const __m128i &shuffled_vector = _mm_shuffle_epi8(vector, mask);
-			_mm_stream_si128((__m128i*)((char*)mapping.first + buffer_offset), shuffled_vector);
 
-			//float x = ((float*)((char*)mapping.first + buffer_offset))[0];
+			const __m128i &vector = _mm_loadu_si128((const __m128i*)src);
-			//float y = ((float*)((char*)mapping.first + buffer_offset))[1];
+			const __m128i &shuffled_vector = _mm_shuffle_epi8(vector, mask);
-			//float z = ((float*)((char*)mapping.first + buffer_offset))[2];
+			_mm_stream_si128((__m128i*)dst, shuffled_vector);
-			//float w = ((float*)((char*)mapping.first + buffer_offset))[3];
+
+			float x = ((float*)dst)[0];
+			float y = ((float*)dst)[1];
+			float z = ((float*)dst)[2];
+			float w = ((float*)dst)[3];
 
 			//LOG_WARNING(RSX, "fc%u = {%g, %g, %g, %g}", constant.id, x, y, z, w);
-			buffer_offset += 4 * sizeof(f32);
+			++dst;
 		}
 	}
 

rpcs3/Emu/RSX/GL/rsx_gl_cache.cpp

@@ -259,6 +259,9 @@ rsx::complete_shader glsl_complete_shader(const rsx::decompiled_shader &shader,
 				"	gl_Position.w = o0.w;\n";
 		}
 
+		{
+			std::string code_end;
+
 			for (std::size_t index = 0; index < 16; ++index)
 			{
 				if (shader.input_attributes & (1 << index))
@@ -266,10 +269,55 @@ rsx::complete_shader glsl_complete_shader(const rsx::decompiled_shader &shader,
 					// result.code += "in vec4 " + rsx::vertex_program::input_attrib_names[index] + ";\n";
 
 					// TODO: use actual information about vertex inputs
-				result.code += "layout(location=" + std::to_string(location++) + ") uniform samplerBuffer " + rsx::vertex_program::input_attrib_names[index] + "_buffer" + ";\n";
+					const std::string &attrib_name = rsx::vertex_program::input_attrib_names[index];
-				result.code += "vec4 " + rsx::vertex_program::input_attrib_names[index]
+
-					+ " = texelFetch(" + rsx::vertex_program::input_attrib_names[index] + "_buffer, gl_VertexID).rgba;\n";
+					result.code += "uniform ";
+
+					if (state.is_int & (1 << index))
+					{
+						result.code += "isamplerBuffer ";
+						code_end += "ivec4 ";
 					}
+					else
+					{
+						result.code += "samplerBuffer ";
+						code_end += "vec4 ";
+					}
+
+					result.code += attrib_name + "_buffer" + ";\n";
+
+					code_end += attrib_name + ";\n";
+
+					std::string vertex_id;
+
+					if (state.is_array & (1 << index))
+					{
+						vertex_id = "gl_VertexID";
+
+						if (state.frequency[index] > 1)
+						{
+							if (state.divider_op & (1 << index))
+							{
+								vertex_id += " % ";
+							}
+							else
+							{
+								vertex_id += " / ";
+							}
+
+							vertex_id += std::to_string(state.frequency[index]);
+						}
+					}
+					else
+					{
+						vertex_id = "0";
+					}
+
+					prepare += '\t' + attrib_name + " = texelFetch(" + attrib_name + "_buffer, " + vertex_id + ");\n";
+				}
+			}
+
+			result.code += code_end;
 		}
 
 		{

rpcs3/Emu/RSX/GL/vertex_buffer.cpp

@@ -160,17 +160,9 @@ u32 GLGSRender::set_vertex_buffer()
 	//initialize vertex attributes
 	//merge all vertex arrays
 
-	std::chrono::time_point<std::chrono::system_clock> then = std::chrono::system_clock::now();
+	static const u32 texture_index_offset = rsx::limits::textures_count + rsx::limits::vertex_textures_count;
 
-	const std::string reg_table[] =
+	std::chrono::time_point<std::chrono::system_clock> then = std::chrono::system_clock::now();
-	{
-		"in_pos", "in_weight", "in_normal",
-		"in_diff_color", "in_spec_color",
-		"in_fog",
-		"in_point_size", "in_7",
-		"in_tc0", "in_tc1", "in_tc2", "in_tc3",
-		"in_tc4", "in_tc5", "in_tc6", "in_tc7"
-	};
 
 	u32 input_mask = rsx::method_registers[NV4097_SET_VERTEX_ATTRIB_INPUT_MASK];
 	u32 min_index = 0, max_index = 0;
@@ -185,7 +177,9 @@ u32 GLGSRender::set_vertex_buffer()
 	for (u8 index = 0; index < rsx::limits::vertex_count; ++index)
 	{
 		if (vertex_arrays_info[index].size == 0)
+		{
 			continue;
+		}
 
 		max_vertex_attrib_size += 16;
 	}
@@ -240,9 +234,9 @@ u32 GLGSRender::set_vertex_buffer()
 
 			if (!vertex_info.size) // disabled, bind a null sampler
 			{
-				glActiveTexture(GL_TEXTURE0 + index + rsx::limits::textures_count);
+				glActiveTexture(GL_TEXTURE0 + index + texture_index_offset);
 				glBindTexture(GL_TEXTURE_BUFFER, 0);
-				glProgramUniform1i(m_program->id(), location, index + rsx::limits::textures_count);
+				glProgramUniform1i(m_program->id(), location, index + texture_index_offset);
 				continue;
 			}
 
@@ -279,7 +273,7 @@ u32 GLGSRender::set_vertex_buffer()
 			texture.copy_from(m_attrib_ring_buffer, gl_type, mapping.second, data_size);
 
 			//Link texture to uniform
-			m_program->uniforms.texture(location, index + rsx::limits::textures_count, texture);
+			m_program->uniforms.texture(location, index + texture_index_offset, texture);
 			if (!is_primitive_native(draw_mode))
 			{
 				std::tie(vertex_draw_count, offset_in_index_buffer) = get_index_array_for_emulated_non_indexed_draw({ { 0, vertex_draw_count } }, draw_mode, m_index_ring_buffer);
@@ -309,9 +303,9 @@ u32 GLGSRender::set_vertex_buffer()
 			bool enabled = !!(input_mask & (1 << index));
 			if (!enabled)
 			{
-				glActiveTexture(GL_TEXTURE0 + index + rsx::limits::textures_count);
+				glActiveTexture(GL_TEXTURE0 + index + texture_index_offset);
 				glBindTexture(GL_TEXTURE_BUFFER, 0);
-				glProgramUniform1i(m_program->id(), location, index + rsx::limits::textures_count);
+				glProgramUniform1i(m_program->id(), location, index + texture_index_offset);
 				continue;
 			}
 
@@ -367,7 +361,7 @@ u32 GLGSRender::set_vertex_buffer()
 				texture.copy_from(m_attrib_ring_buffer, gl_type, buffer_offset, data_size);
 
 				//Link texture to uniform
-				m_program->uniforms.texture(location, index + rsx::limits::textures_count, texture);
+				m_program->uniforms.texture(location, index + texture_index_offset, texture);
 			}
 			else if (register_vertex_info[index].size > 0)
 			{
@@ -392,7 +386,7 @@ u32 GLGSRender::set_vertex_buffer()
 					texture.copy_from(m_attrib_ring_buffer, gl_type, mapping.second, data_size);
 
 					//Link texture to uniform
-					m_program->uniforms.texture(location, index + rsx::limits::textures_count, texture);
+					m_program->uniforms.texture(location, index + texture_index_offset, texture);
 					break;
 				}
 				default:
@@ -402,9 +396,9 @@ u32 GLGSRender::set_vertex_buffer()
 			}
 			else
 			{
-				glActiveTexture(GL_TEXTURE0 + index + rsx::limits::textures_count);
+				glActiveTexture(GL_TEXTURE0 + index + texture_index_offset);
 				glBindTexture(GL_TEXTURE_BUFFER, 0);
-				glProgramUniform1i(m_program->id(), location, index + rsx::limits::textures_count);
+				glProgramUniform1i(m_program->id(), location, index + texture_index_offset);
 				continue;
 			}
 		}

rpcs3/Emu/RSX/RSXTexture.cpp

@@ -72,6 +72,8 @@ namespace rsx
 		case rsx::texture_dimension::dimension1d: return rsx::texture_dimension_extended::texture_dimension_1d;
 		case rsx::texture_dimension::dimension3d: return rsx::texture_dimension_extended::texture_dimension_2d;
 		case rsx::texture_dimension::dimension2d: return cubemap() ? rsx::texture_dimension_extended::texture_dimension_cubemap : rsx::texture_dimension_extended::texture_dimension_2d;
+
+		default: ASSUME(0);
 		}
 	}
 

rpcs3/Emu/RSX/RSXThread.cpp

@@ -751,8 +751,39 @@ namespace rsx
 
 		u32 fp_info = rsx::method_registers[NV4097_SET_SHADER_PROGRAM];
 
+		result.state.input_attributes = rsx::method_registers[NV4097_SET_VERTEX_ATTRIB_INPUT_MASK];
 		result.state.output_attributes = rsx::method_registers[NV4097_SET_VERTEX_ATTRIB_OUTPUT_MASK];
 		result.state.ctrl = rsx::method_registers[NV4097_SET_SHADER_CONTROL];
+		result.state.divider_op = rsx::method_registers[NV4097_SET_FREQUENCY_DIVIDER_OPERATION];
+		 
+		result.state.is_array = 0;
+		result.state.is_int = 0;
+
+		for (u8 index = 0; index < rsx::limits::vertex_count; ++index)
+		{
+			bool is_int = false;
+
+			if (vertex_arrays_info[index].size > 0)
+			{
+				result.state.is_array |= 1 << index;
+				is_int = is_int_type(vertex_arrays_info[index].type);
+				result.state.frequency[index] = vertex_arrays_info[index].frequency;
+			}
+			else if (register_vertex_info[index].size > 0)
+			{
+				is_int = is_int_type(register_vertex_info[index].type);
+				result.state.frequency[index] = register_vertex_info[index].frequency;
+			}
+			else
+			{
+				result.state.frequency[index] = 0;
+			}
+
+			if (is_int)
+			{
+				result.state.is_int |= 1 << index;
+			}
+		}
 
 		result.vertex_shader.ucode_ptr = transform_program;
 		result.vertex_shader.offset = rsx::method_registers[NV4097_SET_TRANSFORM_PROGRAM_START];

rpcs3/rpcs3.vcxproj

@@ -67,10 +67,12 @@
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug - LLVM|x64'">
     <OutDir>$(SolutionDir)bin\</OutDir>
     <TargetName>$(ProjectName)-dbg</TargetName>
+    <LibraryPath>$(SolutionDir)lib\Debug-$(Platform)\;$(LibraryPath)</LibraryPath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug - MemLeak|x64'">
     <OutDir>$(SolutionDir)bin\</OutDir>
     <TargetName>$(ProjectName)-dbg</TargetName>
+    <LibraryPath>$(SolutionDir)lib\Debug-$(Platform)\;$(LibraryPath)</LibraryPath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <OutDir>$(SolutionDir)bin\</OutDir>
@@ -81,6 +83,7 @@
     <OutDir>$(SolutionDir)bin\</OutDir>
     <LinkIncremental>false</LinkIncremental>
     <RunCodeAnalysis>false</RunCodeAnalysis>
+    <LibraryPath>$(SolutionDir)lib\Release-$(Platform)\;$(LibraryPath)</LibraryPath>
   </PropertyGroup>
   <ItemDefinitionGroup>
     <PreBuildEvent>

rsx_program_decompiler

@@ -1 +1 @@
-Subproject commit e4d938c76850549acd837326e2fe0890d5b4be03
+Subproject commit 9f8814af57264c82b7e063c1df5d71dc32c7a951