rsx: Fix program generation and compact referenced data blocks

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -800,11 +800,14 @@ void GLGSRender::load_program_env()
 	if (update_transform_constants)
 	{
 		// Vertex constants
-		auto mapping = m_transform_constants_buffer->alloc_from_heap(8192, m_uniform_buffer_offset_align);
+		const std::vector<u16>& constant_ids = m_vertex_prog ? m_vertex_prog->constant_ids : std::vector<u16>{};
+		const usz transform_constants_size = constant_ids.empty() ? 8192 : constant_ids.size() * 16;
+
+		auto mapping = m_transform_constants_buffer->alloc_from_heap(transform_constants_size, m_uniform_buffer_offset_align);
 		auto buf = static_cast<u8*>(mapping.first);
 		fill_vertex_program_constants_data(buf, m_vertex_prog ? m_vertex_prog->constant_ids : std::vector<u16>{});
 
-		m_transform_constants_buffer->bind_range(GL_VERTEX_CONSTANT_BUFFERS_BIND_SLOT, mapping.second, 8192);
+		m_transform_constants_buffer->bind_range(GL_VERTEX_CONSTANT_BUFFERS_BIND_SLOT, mapping.second, transform_constants_size);
 	}
 
 	if (update_fragment_constants && !update_instruction_buffers)

rpcs3/Emu/RSX/GL/GLVertexProgram.cpp

@@ -58,17 +58,21 @@ void GLVertexDecompilerThread::insertInputs(std::stringstream& OS, const std::ve
 
 void GLVertexDecompilerThread::insertConstants(std::stringstream& OS, const std::vector<ParamType>& constants)
 {
-	OS << "layout(std140, binding = 2) uniform VertexConstantsBuffer\n";
+
-	OS << "{\n";
-	OS << "	vec4 vc[468];\n";
-	OS << "};\n\n";
 
 	for (const ParamType &PT: constants)
 	{
 		for (const ParamItem &PI : PT.items)
 		{
-			if (PI.name == "vc[468]")
+			if (PI.name.starts_with("vc["))
+			{
+				OS << "layout(std140, binding = 2) uniform VertexConstantsBuffer\n";
+				OS << "{\n";
+				OS << "	vec4 " << PI.name << ";\n";
+				OS << "};\n\n";
+
 				continue;
+			}
 
 			OS << "uniform " << PT.type << " " << PI.name << ";\n";
 		}
@@ -272,10 +276,14 @@ void GLVertexProgram::Decompile(const RSXVertexProgram& prog)
 	GLVertexDecompilerThread decompiler(prog, source, parr);
 	decompiler.Task();
 
+	if (has_indexed_constants = decompiler.properties.has_indexed_constants;
+		!has_indexed_constants)
+	{
+		constant_ids = std::vector<u16>(decompiler.m_constant_ids.begin(), decompiler.m_constant_ids.end());
+	}
+
 	shader.create(::glsl::program_domain::glsl_vertex_program, source);
 	id = shader.id();
-	has_indexed_constants = decompiler.properties.has_indexed_constants;
-	constant_ids = std::move(decompiler.m_constant_ids);
 }
 
 void GLVertexProgram::Delete()

rpcs3/Emu/RSX/Program/ShaderParam.h

@@ -145,7 +145,7 @@ struct ParamType
 			std::vector<ParamItem> new_list;
 			for (const auto& it : items)
 			{
-				if (it.name != item.name)
+				if (it.name != name)
 				{
 					new_list.emplace_back(it.name, it.location, it.value);
 				}
@@ -156,6 +156,7 @@ struct ParamType
 			}
 
 			std::swap(items, new_list);
+			return true;
 		}
 
 		items.push_back(item);

rpcs3/Emu/RSX/Program/VertexProgramDecompiler.cpp

@@ -127,7 +127,7 @@ std::string VertexProgramDecompiler::GetSRC(const u32 n)
 	case RSX_VP_REGISTER_TYPE_CONSTANT:
 		m_parr.AddParam(PF_PARAM_UNIFORM, getFloatTypeName(4), std::string("vc[468]"));
 		properties.has_indexed_constants |= !!d3.index_const;
-		m_constant_ids.push_back(d1.const_src);
+		m_constant_ids.insert(static_cast<u16>(d1.const_src));
 		ret += std::string("vc[") + std::to_string(d1.const_src) + (d3.index_const ? " + " + AddAddrReg() : "") + "]";
 		break;
 
@@ -399,20 +399,20 @@ std::string VertexProgramDecompiler::BuildCode()
 		std::vector<std::pair<std::string, std::string>> reloc_table;
 		reloc_table.reserve(m_constant_ids.size());
 
-		// First sort the data in ascending order
-		std::sort(m_constant_ids.begin(), m_constant_ids.end());
-
 		// Build the string lookup table
+		int offset = 0;
 		for (const auto& index : m_constant_ids)
 		{
-			reloc_table.emplace_back(fmt::format("vc[%d]", index), fmt::format("vc[%llu]", reloc_table.size()));
+			const auto i = offset++;
+			if (i == index) continue; // Replace with self
+			reloc_table.emplace_back(fmt::format("vc[%d]", index), fmt::format("vc[%d]", i));
 		}
 
 		// One-time patch
 		main_body = fmt::replace_all(main_body, reloc_table);
 
 		// Rename the array type
-		auto type_list = ensure(m_parr.SearchParam(PF_PARAM_CONST, getFloatTypeName(4)));
+		auto type_list = ensure(m_parr.SearchParam(PF_PARAM_UNIFORM, getFloatTypeName(4)));
 		const auto item = ParamItem(fmt::format("vc[%llu]", m_constant_ids.size()), -1);
 		type_list->ReplaceOrInsert("vc[468]", item);
 	}

rpcs3/Emu/RSX/Program/VertexProgramDecompiler.h

@@ -63,7 +63,7 @@ struct VertexProgramDecompiler
 	const RSXVertexProgram& m_prog;
 	ParamArray m_parr;
 
-	std::vector<u16> m_constant_ids;
+	std::set<u16> m_constant_ids;
 
 	static std::string NotZeroPositive(const std::string& code);
 	std::string GetMask(bool is_sca) const;

rpcs3/Emu/RSX/RSXThread.cpp

@@ -863,10 +863,6 @@ namespace rsx
 	void thread::fill_vertex_program_constants_data(void* buffer, const std::vector<u16>& reloc_table)
 	{
 		if (!reloc_table.empty()) [[ likely ]]
-		{
-			memcpy(buffer, rsx::method_registers.transform_constants.data(), 468 * 4 * sizeof(float));
-		}
-		else
 		{
 			char* dst = reinterpret_cast<char*>(buffer);
 			for (const auto& index : reloc_table)
@@ -875,6 +871,10 @@ namespace rsx
 				dst += 16;
 			}
 		}
+		else
+		{
+			memcpy(buffer, rsx::method_registers.transform_constants.data(), 468 * 4 * sizeof(float));
+		}
 	}
 
 	void thread::fill_fragment_state_buffer(void* buffer, const RSXFragmentProgram& /*fragment_program*/)

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -1955,12 +1955,15 @@ void VKGSRender::load_program_env()
 		check_heap_status(VK_HEAP_CHECK_TRANSFORM_CONSTANTS_STORAGE);
 
 		// Transform constants
-		auto mem = m_transform_constants_ring_info.alloc<256>(8192);
+		const std::vector<u16>& constant_ids = m_vertex_prog ? m_vertex_prog->constant_ids : std::vector<u16>{};
-		auto buf = m_transform_constants_ring_info.map(mem, 8192);
+		const usz transform_constants_size = constant_ids.empty() ? 8192 : utils::align(constant_ids.size() * 16, m_device->gpu().get_limits().minUniformBufferOffsetAlignment);
 
-		fill_vertex_program_constants_data(buf, m_vertex_prog ? m_vertex_prog->constant_ids : std::vector<u16>{});
+		auto mem = m_transform_constants_ring_info.alloc<1>(transform_constants_size);
+		auto buf = m_transform_constants_ring_info.map(mem, transform_constants_size);
+
+		fill_vertex_program_constants_data(buf, constant_ids);
 		m_transform_constants_ring_info.unmap();
-		m_vertex_constants_buffer_info = { m_transform_constants_ring_info.heap->value, mem, 8192 };
+		m_vertex_constants_buffer_info = { m_transform_constants_ring_info.heap->value, mem, transform_constants_size };
 	}
 
 	if (update_fragment_constants && !update_instruction_buffers)

rpcs3/Emu/RSX/VK/VKVertexProgram.cpp

@@ -101,27 +101,28 @@ void VKVertexDecompilerThread::insertInputs(std::stringstream& OS, const std::ve
 
 void VKVertexDecompilerThread::insertConstants(std::stringstream & OS, const std::vector<ParamType> & constants)
 {
-	OS << "layout(std140, set=0, binding = 1) uniform VertexConstantsBuffer\n";
-	OS << "{\n";
-	OS << "	vec4 vc[468];\n";
-	OS << "};\n\n";
-
 	vk::glsl::program_input in;
-	in.location = m_binding_table.vertex_constant_buffers_bind_slot;
-	in.domain = glsl::glsl_vertex_program;
-	in.name = "VertexConstantsBuffer";
-	in.type = vk::glsl::input_type_uniform_buffer;
-
-	inputs.push_back(in);
-
-
 	u32 location = m_binding_table.vertex_textures_first_bind_slot;
+
 	for (const ParamType &PT : constants)
 	{
 		for (const ParamItem &PI : PT.items)
 		{
-			if (PI.name == "vc[468]")
+			if (PI.name.starts_with("vc["))
+			{
+				OS << "layout(std140, set=0, binding = " << static_cast<int>(m_binding_table.vertex_constant_buffers_bind_slot) << ") uniform VertexConstantsBuffer\n";
+				OS << "{\n";
+				OS << "	vec4 " << PI.name << ";\n";
+				OS << "};\n\n";
+
+				in.location = m_binding_table.vertex_constant_buffers_bind_slot;
+				in.domain = glsl::glsl_vertex_program;
+				in.name = "VertexConstantsBuffer";
+				in.type = vk::glsl::input_type_uniform_buffer;
+
+				inputs.push_back(in);
 				continue;
+			}
 
 			if (PT.type == "sampler2D" ||
 				PT.type == "samplerCube" ||
@@ -349,9 +350,13 @@ void VKVertexProgram::Decompile(const RSXVertexProgram& prog)
 	VKVertexDecompilerThread decompiler(prog, source, parr, *this);
 	decompiler.Task();
 
+	if (has_indexed_constants = decompiler.properties.has_indexed_constants;
+		!has_indexed_constants)
+	{
+		constant_ids = std::vector<u16>(decompiler.m_constant_ids.begin(), decompiler.m_constant_ids.end());
+	}
+
 	shader.create(::glsl::program_domain::glsl_vertex_program, source);
-	has_indexed_constants = decompiler.properties.has_indexed_constants;
-	constant_ids = std::move(decompiler.m_constant_ids);
 }
 
 void VKVertexProgram::Compile()