Merge pull request #66178 from clayjohn/double-precision-rendering

Emulate double precision for regular rendering operation when REAL_T_IS_DOUBLE

servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp

@@ -714,6 +714,14 @@ void RenderForwardClustered::_fill_instance_data(RenderListType p_render_list, i
 		instance_data.lightmap_uv_scale[2] = inst->lightmap_uv_scale.size.x;
 		instance_data.lightmap_uv_scale[3] = inst->lightmap_uv_scale.size.y;
 
+#ifdef REAL_T_IS_DOUBLE
+		// Split the origin into two components, the float approximation and the missing precision
+		// In the shader we will combine these back together to restore the lost precision.
+		RendererRD::MaterialStorage::split_double(inst->transform.origin.x, &instance_data.transform[12], &instance_data.transform[3]);
+		RendererRD::MaterialStorage::split_double(inst->transform.origin.y, &instance_data.transform[13], &instance_data.transform[7]);
+		RendererRD::MaterialStorage::split_double(inst->transform.origin.z, &instance_data.transform[14], &instance_data.transform[11]);
+#endif
+
 		bool cant_repeat = instance_data.flags & INSTANCE_DATA_FLAG_MULTIMESH || inst->mesh_instance.is_valid();
 
 		if (prev_surface != nullptr && !cant_repeat && prev_surface->sort.sort_key1 == surface->sort.sort_key1 && prev_surface->sort.sort_key2 == surface->sort.sort_key2 && repeats < RenderElementInfo::MAX_REPEATS) {
@@ -3039,6 +3047,11 @@ RenderForwardClustered::RenderForwardClustered() {
 		{
 			defines += "\n#define MATERIAL_UNIFORM_SET " + itos(MATERIAL_UNIFORM_SET) + "\n";
 		}
+#ifdef REAL_T_IS_DOUBLE
+		{
+			defines += "\n#define USE_DOUBLE_PRECISION \n";
+		}
+#endif
 
 		scene_shader.init(defines);
 	}

servers/rendering/renderer_rd/forward_clustered/scene_shader_forward_clustered.cpp

@@ -621,10 +621,10 @@ void SceneShaderForwardClustered::init(const String p_defines) {
 		//shader compiler
 		ShaderCompiler::DefaultIdentifierActions actions;
 
-		actions.renames["MODEL_MATRIX"] = "model_matrix";
+		actions.renames["MODEL_MATRIX"] = "read_model_matrix";
 		actions.renames["MODEL_NORMAL_MATRIX"] = "model_normal_matrix";
 		actions.renames["VIEW_MATRIX"] = "scene_data.view_matrix";
-		actions.renames["INV_VIEW_MATRIX"] = "scene_data.inv_view_matrix";
+		actions.renames["INV_VIEW_MATRIX"] = "inv_view_matrix";
 		actions.renames["PROJECTION_MATRIX"] = "projection_matrix";
 		actions.renames["INV_PROJECTION_MATRIX"] = "inv_projection_matrix";
 		actions.renames["MODELVIEW_MATRIX"] = "modelview";
@@ -757,6 +757,8 @@ void SceneShaderForwardClustered::init(const String p_defines) {
 		actions.usage_defines["RADIANCE"] = "#define CUSTOM_RADIANCE_USED\n";
 		actions.usage_defines["IRRADIANCE"] = "#define CUSTOM_IRRADIANCE_USED\n";
 
+		actions.usage_defines["MODEL_MATRIX"] = "#define MODEL_MATRIX_USED\n";
+
 		actions.render_mode_defines["skip_vertex_transform"] = "#define SKIP_TRANSFORM_USED\n";
 		actions.render_mode_defines["world_vertex_coords"] = "#define VERTEX_WORLD_COORDS_USED\n";
 		actions.render_mode_defines["ensure_correct_normals"] = "#define ENSURE_CORRECT_NORMALS\n";

servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp

@@ -1757,6 +1757,14 @@ void RenderForwardMobile::_render_list_template(RenderingDevice::DrawListID p_dr
 			RendererRD::MaterialStorage::store_transform(Transform3D(), push_constant.transform);
 		}
 
+#ifdef REAL_T_IS_DOUBLE
+		// Split the origin into two components, the float approximation and the missing precision
+		// In the shader we will combine these back together to restore the lost precision.
+		RendererRD::MaterialStorage::split_double(inst->transform.origin.x, &push_constant.transform[12], &push_constant.transform[3]);
+		RendererRD::MaterialStorage::split_double(inst->transform.origin.y, &push_constant.transform[13], &push_constant.transform[7]);
+		RendererRD::MaterialStorage::split_double(inst->transform.origin.z, &push_constant.transform[14], &push_constant.transform[11]);
+#endif
+
 		push_constant.flags = inst->flags_cache;
 		push_constant.gi_offset = inst->gi_offset_cache;
 		push_constant.layer_mask = inst->layer_mask;
@@ -2472,6 +2480,11 @@ RenderForwardMobile::RenderForwardMobile() {
 	{
 		defines += "\n#define MATERIAL_UNIFORM_SET " + itos(MATERIAL_UNIFORM_SET) + "\n";
 	}
+#ifdef REAL_T_IS_DOUBLE
+	{
+		defines += "\n#define USE_DOUBLE_PRECISION \n";
+	}
+#endif
 
 	scene_shader.init(defines);
 

servers/rendering/renderer_rd/forward_mobile/scene_shader_forward_mobile.cpp

@@ -521,10 +521,10 @@ void SceneShaderForwardMobile::init(const String p_defines) {
 		//shader compiler
 		ShaderCompiler::DefaultIdentifierActions actions;
 
-		actions.renames["MODEL_MATRIX"] = "model_matrix";
+		actions.renames["MODEL_MATRIX"] = "read_model_matrix";
 		actions.renames["MODEL_NORMAL_MATRIX"] = "model_normal_matrix";
 		actions.renames["VIEW_MATRIX"] = "scene_data.view_matrix";
-		actions.renames["INV_VIEW_MATRIX"] = "scene_data.inv_view_matrix";
+		actions.renames["INV_VIEW_MATRIX"] = "inv_view_matrix";
 		actions.renames["PROJECTION_MATRIX"] = "projection_matrix";
 		actions.renames["INV_PROJECTION_MATRIX"] = "inv_projection_matrix";
 		actions.renames["MODELVIEW_MATRIX"] = "modelview";
@@ -657,6 +657,8 @@ void SceneShaderForwardMobile::init(const String p_defines) {
 		actions.usage_defines["RADIANCE"] = "#define CUSTOM_RADIANCE_USED\n";
 		actions.usage_defines["IRRADIANCE"] = "#define CUSTOM_IRRADIANCE_USED\n";
 
+		actions.usage_defines["MODEL_MATRIX"] = "#define MODEL_MATRIX_USED\n";
+
 		actions.render_mode_defines["skip_vertex_transform"] = "#define SKIP_TRANSFORM_USED\n";
 		actions.render_mode_defines["world_vertex_coords"] = "#define VERTEX_WORLD_COORDS_USED\n";
 		actions.render_mode_defines["ensure_correct_normals"] = "#define ENSURE_CORRECT_NORMALS\n";

servers/rendering/renderer_rd/shaders/forward_clustered/scene_forward_clustered.glsl

@@ -129,12 +129,52 @@ invariant gl_Position;
 
 #GLOBALS
 
+#ifdef USE_DOUBLE_PRECISION
+// Helper functions for emulating double precision when adding floats.
+vec3 quick_two_sum(vec3 a, vec3 b, out vec3 out_p) {
+	vec3 s = a + b;
+	out_p = b - (s - a);
+	return s;
+}
+
+vec3 two_sum(vec3 a, vec3 b, out vec3 out_p) {
+	vec3 s = a + b;
+	vec3 v = s - a;
+	out_p = (a - (s - v)) + (b - v);
+	return s;
+}
+
+vec3 double_add_vec3(vec3 base_a, vec3 prec_a, vec3 base_b, vec3 prec_b, out vec3 out_precision) {
+	vec3 s, t, se, te;
+	s = two_sum(base_a, base_b, se);
+	t = two_sum(prec_a, prec_b, te);
+	se += t;
+	s = quick_two_sum(s, se, se);
+	se += te;
+	s = quick_two_sum(s, se, out_precision);
+	return s;
+}
+#endif
+
 void vertex_shader(in uint instance_index, in bool is_multimesh, in uint multimesh_offset, in SceneData scene_data, in mat4 model_matrix, out vec4 screen_pos) {
 	vec4 instance_custom = vec4(0.0);
 #if defined(COLOR_USED)
 	color_interp = color_attrib;
 #endif
 
+	mat4 inv_view_matrix = scene_data.inv_view_matrix;
+
+#ifdef USE_DOUBLE_PRECISION
+	vec3 model_precision = vec3(model_matrix[0][3], model_matrix[1][3], model_matrix[2][3]);
+	model_matrix[0][3] = 0.0;
+	model_matrix[1][3] = 0.0;
+	model_matrix[2][3] = 0.0;
+	vec3 view_precision = vec3(inv_view_matrix[0][3], inv_view_matrix[1][3], inv_view_matrix[2][3]);
+	inv_view_matrix[0][3] = 0.0;
+	inv_view_matrix[1][3] = 0.0;
+	inv_view_matrix[2][3] = 0.0;
+#endif
+
 	mat3 model_normal_matrix;
 	if (bool(instances.data[instance_index].flags & INSTANCE_FLAGS_NON_UNIFORM_SCALE)) {
 		model_normal_matrix = transpose(inverse(mat3(model_matrix)));
@@ -142,11 +182,12 @@ void vertex_shader(in uint instance_index, in bool is_multimesh, in uint multime
 		model_normal_matrix = mat3(model_matrix);
 	}
 
+	mat4 matrix;
+	mat4 read_model_matrix = model_matrix;
+
 	if (is_multimesh) {
 		//multimesh, instances are for it
 
-		mat4 matrix;
-
 #ifdef USE_PARTICLE_TRAILS
 		uint trail_size = (instances.data[instance_index].flags >> INSTANCE_FLAGS_PARTICLE_TRAIL_SHIFT) & INSTANCE_FLAGS_PARTICLE_TRAIL_MASK;
 		uint stride = 3 + 1 + 1; //particles always uses this format
@@ -232,7 +273,14 @@ void vertex_shader(in uint instance_index, in bool is_multimesh, in uint multime
 #endif
 		//transpose
 		matrix = transpose(matrix);
-		model_matrix = model_matrix * matrix;
+#if !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED) || defined(MODEL_MATRIX_USED)
+		// Normally we can bake the multimesh transform into the model matrix, but when using double precision
+		// we avoid baking it in so we can emulate high precision.
+		read_model_matrix = model_matrix * matrix;
+#if !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED)
+		model_matrix = read_model_matrix;
+#endif // !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED)
+#endif // !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED) || defined(MODEL_MATRIX_USED)
 		model_normal_matrix = model_normal_matrix * mat3(matrix);
 	}
 
@@ -297,7 +345,22 @@ void vertex_shader(in uint instance_index, in bool is_multimesh, in uint multime
 // using local coordinates (default)
 #if !defined(SKIP_TRANSFORM_USED) && !defined(VERTEX_WORLD_COORDS_USED)
 
+#ifdef USE_DOUBLE_PRECISION
+	// We separate the basis from the origin becasue the basis is fine with single point precision.
+	// Then we combine the translations from the model matrix and the view matrix using emulated doubles.
+	// We add the result to the vertex and ignore the final lost precision.
+	vec3 model_origin = model_matrix[3].xyz;
+	if (is_multimesh) {
+		vertex = mat3(matrix) * vertex;
+		model_origin = double_add_vec3(model_origin, model_precision, matrix[3].xyz, vec3(0.0), model_precision);
+	}
+	vertex = mat3(model_matrix) * vertex;
+	vec3 temp_precision; // Will be ignored.
+	vertex += double_add_vec3(model_origin, model_precision, scene_data.inv_view_matrix[3].xyz, view_precision, temp_precision);
+	vertex = mat3(scene_data.view_matrix) * vertex;
+#else
 	vertex = (modelview * vec4(vertex, 1.0)).xyz;
+#endif
 #ifdef NORMAL_USED
 	normal = modelview_normal * normal;
 #endif
@@ -490,7 +553,6 @@ layout(location = 10) in flat uint instance_index_interp;
 
 //defines to keep compatibility with vertex
 
-#define model_matrix instances.data[draw_call.instance_index].transform
 #ifdef USE_MULTIVIEW
 #define projection_matrix scene_data.projection_matrix_view[ViewIndex]
 #define inv_projection_matrix scene_data.inv_projection_matrix_view[ViewIndex]
@@ -737,6 +799,17 @@ void fragment_shader(in SceneData scene_data) {
 	vec2 alpha_texture_coordinate = vec2(0.0, 0.0);
 #endif // ALPHA_ANTIALIASING_EDGE_USED
 
+	mat4 inv_view_matrix = scene_data.inv_view_matrix;
+	mat4 read_model_matrix = instances.data[instance_index].transform;
+#ifdef USE_DOUBLE_PRECISION
+	read_model_matrix[0][3] = 0.0;
+	read_model_matrix[1][3] = 0.0;
+	read_model_matrix[2][3] = 0.0;
+	inv_view_matrix[0][3] = 0.0;
+	inv_view_matrix[1][3] = 0.0;
+	inv_view_matrix[2][3] = 0.0;
+#endif
+
 	{
 #CODE : FRAGMENT
 	}

servers/rendering/renderer_rd/shaders/forward_mobile/scene_forward_mobile.glsl

@@ -123,6 +123,33 @@ invariant gl_Position;
 
 #define scene_data scene_data_block.data
 
+#ifdef USE_DOUBLE_PRECISION
+// Helper functions for emulating double precision when adding floats.
+vec3 quick_two_sum(vec3 a, vec3 b, out vec3 out_p) {
+	vec3 s = a + b;
+	out_p = b - (s - a);
+	return s;
+}
+
+vec3 two_sum(vec3 a, vec3 b, out vec3 out_p) {
+	vec3 s = a + b;
+	vec3 v = s - a;
+	out_p = (a - (s - v)) + (b - v);
+	return s;
+}
+
+vec3 double_add_vec3(vec3 base_a, vec3 prec_a, vec3 base_b, vec3 prec_b, out vec3 out_precision) {
+	vec3 s, t, se, te;
+	s = two_sum(base_a, base_b, se);
+	t = two_sum(prec_a, prec_b, te);
+	se += t;
+	s = quick_two_sum(s, se, se);
+	se += te;
+	s = quick_two_sum(s, se, out_precision);
+	return s;
+}
+#endif
+
 void main() {
 	vec4 instance_custom = vec4(0.0);
 #if defined(COLOR_USED)
@@ -132,6 +159,17 @@ void main() {
 	bool is_multimesh = bool(draw_call.flags & INSTANCE_FLAGS_MULTIMESH);
 
 	mat4 model_matrix = draw_call.transform;
+	mat4 inv_view_matrix = scene_data.inv_view_matrix;
+#ifdef USE_DOUBLE_PRECISION
+	vec3 model_precision = vec3(model_matrix[0][3], model_matrix[1][3], model_matrix[2][3]);
+	model_matrix[0][3] = 0.0;
+	model_matrix[1][3] = 0.0;
+	model_matrix[2][3] = 0.0;
+	vec3 view_precision = vec3(inv_view_matrix[0][3], inv_view_matrix[1][3], inv_view_matrix[2][3]);
+	inv_view_matrix[0][3] = 0.0;
+	inv_view_matrix[1][3] = 0.0;
+	inv_view_matrix[2][3] = 0.0;
+#endif
 
 	mat3 model_normal_matrix;
 	if (bool(draw_call.flags & INSTANCE_FLAGS_NON_UNIFORM_SCALE)) {
@@ -140,11 +178,12 @@ void main() {
 		model_normal_matrix = mat3(model_matrix);
 	}
 
+	mat4 matrix;
+	mat4 read_model_matrix = model_matrix;
+
 	if (is_multimesh) {
 		//multimesh, instances are for it
 
-		mat4 matrix;
-
 #ifdef USE_PARTICLE_TRAILS
 		uint trail_size = (draw_call.flags >> INSTANCE_FLAGS_PARTICLE_TRAIL_SHIFT) & INSTANCE_FLAGS_PARTICLE_TRAIL_MASK;
 		uint stride = 3 + 1 + 1; //particles always uses this format
@@ -230,7 +269,15 @@ void main() {
 #endif
 		//transpose
 		matrix = transpose(matrix);
-		model_matrix = model_matrix * matrix;
+
+#if !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED) || defined(MODEL_MATRIX_USED)
+		// Normally we can bake the multimesh transform into the model matrix, but when using double precision
+		// we avoid baking it in so we can emulate high precision.
+		read_model_matrix = model_matrix * matrix;
+#if !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED)
+		model_matrix = read_model_matrix;
+#endif // !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED)
+#endif // !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED) || defined(MODEL_MATRIX_USED)
 		model_normal_matrix = model_normal_matrix * mat3(matrix);
 	}
 
@@ -297,7 +344,22 @@ void main() {
 // using local coordinates (default)
 #if !defined(SKIP_TRANSFORM_USED) && !defined(VERTEX_WORLD_COORDS_USED)
 
+#ifdef USE_DOUBLE_PRECISION
+	// We separate the basis from the origin becasue the basis is fine with single point precision.
+	// Then we combine the translations from the model matrix and the view matrix using emulated doubles.
+	// We add the result to the vertex and ignore the final lost precision.
+	vec3 model_origin = model_matrix[3].xyz;
+	if (is_multimesh) {
+		vertex = mat3(matrix) * vertex;
+		model_origin = double_add_vec3(model_origin, model_precision, matrix[3].xyz, vec3(0.0), model_precision);
+	}
+	vertex = mat3(model_matrix) * vertex;
+	vec3 temp_precision;
+	vertex += double_add_vec3(model_origin, model_precision, scene_data.inv_view_matrix[3].xyz, view_precision, temp_precision);
+	vertex = mat3(scene_data.view_matrix) * vertex;
+#else
 	vertex = (modelview * vec4(vertex, 1.0)).xyz;
+#endif
 #ifdef NORMAL_USED
 	normal = modelview_normal * normal;
 #endif
@@ -468,7 +530,6 @@ layout(location = 9) highp in float dp_clip;
 
 //defines to keep compatibility with vertex
 
-#define model_matrix draw_call.transform
 #ifdef USE_MULTIVIEW
 #define projection_matrix scene_data.projection_matrix_view[ViewIndex]
 #define inv_projection_matrix scene_data.inv_projection_matrix_view[ViewIndex]
@@ -685,6 +746,17 @@ void main() {
 	vec2 alpha_texture_coordinate = vec2(0.0, 0.0);
 #endif // ALPHA_ANTIALIASING_EDGE_USED
 
+	mat4 inv_view_matrix = scene_data.inv_view_matrix;
+	mat4 read_model_matrix = draw_call.transform;
+#ifdef USE_DOUBLE_PRECISION
+	read_model_matrix[0][3] = 0.0;
+	read_model_matrix[1][3] = 0.0;
+	read_model_matrix[2][3] = 0.0;
+	inv_view_matrix[0][3] = 0.0;
+	inv_view_matrix[1][3] = 0.0;
+	inv_view_matrix[2][3] = 0.0;
+#endif
+
 	{
 #CODE : FRAGMENT
 	}

servers/rendering/renderer_rd/storage_rd/material_storage.h

@@ -311,6 +311,18 @@ public:
 		}
 	}
 
+	// http://andrewthall.org/papers/df64_qf128.pdf
+#ifdef REAL_T_IS_DOUBLE
+	static _FORCE_INLINE_ void split_double(double a, float *ahi, float *alo) {
+		const double SPLITTER = (1 << 29) + 1;
+		double t = a * SPLITTER;
+		double thi = t - (t - a);
+		double tlo = a - thi;
+		*ahi = (float)thi;
+		*alo = (float)tlo;
+	}
+#endif
+
 	/* Samplers */
 
 	_FORCE_INLINE_ RID sampler_rd_get_default(RS::CanvasItemTextureFilter p_filter, RS::CanvasItemTextureRepeat p_repeat) {

servers/rendering/renderer_rd/storage_rd/render_scene_data_rd.cpp

@@ -59,6 +59,12 @@ void RenderSceneDataRD::update_ubo(RID p_uniform_buffer, RS::ViewportDebugDraw p
 	RendererRD::MaterialStorage::store_transform(cam_transform, ubo.inv_view_matrix);
 	RendererRD::MaterialStorage::store_transform(cam_transform.affine_inverse(), ubo.view_matrix);
 
+#ifdef REAL_T_IS_DOUBLE
+	RendererRD::MaterialStorage::split_double(-cam_transform.origin.x, &ubo.inv_view_matrix[12], &ubo.inv_view_matrix[3]);
+	RendererRD::MaterialStorage::split_double(-cam_transform.origin.y, &ubo.inv_view_matrix[13], &ubo.inv_view_matrix[7]);
+	RendererRD::MaterialStorage::split_double(-cam_transform.origin.z, &ubo.inv_view_matrix[14], &ubo.inv_view_matrix[11]);
+#endif
+
 	for (uint32_t v = 0; v < view_count; v++) {
 		projection = correction * view_projection[v];
 		RendererRD::MaterialStorage::store_camera(projection, ubo.projection_matrix_view[v]);

servers/rendering/shader_types.cpp

@@ -86,8 +86,8 @@ ShaderTypes::ShaderTypes() {
 	shader_modes[RS::SHADER_SPATIAL].functions["vertex"].main_function = true;
 
 	//builtins
-	shader_modes[RS::SHADER_SPATIAL].functions["vertex"].built_ins["MODEL_MATRIX"] = ShaderLanguage::TYPE_MAT4;
-	shader_modes[RS::SHADER_SPATIAL].functions["vertex"].built_ins["MODEL_NORMAL_MATRIX"] = ShaderLanguage::TYPE_MAT3;
+	shader_modes[RS::SHADER_SPATIAL].functions["vertex"].built_ins["MODEL_MATRIX"] = constt(ShaderLanguage::TYPE_MAT4);
+	shader_modes[RS::SHADER_SPATIAL].functions["vertex"].built_ins["MODEL_NORMAL_MATRIX"] = constt(ShaderLanguage::TYPE_MAT3);
 	shader_modes[RS::SHADER_SPATIAL].functions["vertex"].built_ins["VIEW_MATRIX"] = constt(ShaderLanguage::TYPE_MAT4);
 	shader_modes[RS::SHADER_SPATIAL].functions["vertex"].built_ins["INV_VIEW_MATRIX"] = constt(ShaderLanguage::TYPE_MAT4);
 	shader_modes[RS::SHADER_SPATIAL].functions["vertex"].built_ins["PROJECTION_MATRIX"] = ShaderLanguage::TYPE_MAT4;