slang-shaders/bezel/Mega_Bezel/shaders/base/common/royale-geometry-functions.inc
2022-08-24 22:32:58 -04:00

813 lines
48 KiB
C++

///////////////////////////// GPL LICENSE NOTICE /////////////////////////////
// royale-geometry-functions, extracted from crt-royale
// crt-royale: A full-featured CRT shader, with cheese.
// GPL & Copyright (C) 2014 TroggleMonkey <trogglemonkey@gmx.com>
//////////////////////////// MACROS AND CONSTANTS ////////////////////////////
// Curvature-related constants:
#define HRG_MAX_POINT_CLOUD_SIZE 9
///////////////////////////// CURVATURE FUNCTIONS /////////////////////////////
vec2 hrg_quadratic_solve(float a, float b_over_2, float c)
{
// Requires: 1.) a, b, and c are quadratic formula coefficients
// 2.) b_over_2 = b/2.0 (simplifies terms to factor 2 out)
// 3.) b_over_2 must be guaranteed < 0 (avoids a branch)
// Returns: Returns vec2(first_solution, discriminant), so the caller
// can choose how to handle the "no intersection" case. The
// Kahan or Citardauq formula is used for numerical robustness.
float discriminant = b_over_2 * b_over_2 - a * c;
float solution0 = c / (-b_over_2 + sqrt(discriminant));
return vec2(solution0, discriminant);
}
vec2 hrg_intersect_sphere(vec3 view_vec, vec3 eye_pos_vec, float in_geom_radius)
{
// Requires: 1.) view_vec and eye_pos_vec are 3D vectors in the sphere's
// local coordinate frame (eye_pos_vec is a position, i.e.
// a vector from the origin to the eye/camera)
// 2.) in_geom_radius is a global containing the sphere's radius
// Returns: Cast a ray of direction view_vec from eye_pos_vec at a
// sphere of radius in_geom_radius, and return the distance to
// the first intersection in units of length(view_vec).
// http://wiki.cgsociety.org/index.php/Ray_Sphere_Intersection
// Quadratic formula coefficients (b_over_2 is guaranteed negative):
float a = dot(view_vec, view_vec);
float b_over_2 = dot(view_vec, eye_pos_vec); // * 2.0 factored out
float c = dot(eye_pos_vec, eye_pos_vec) - in_geom_radius * in_geom_radius;
return hrg_quadratic_solve(a, b_over_2, c);
}
vec2 hrg_intersect_cylinder(vec3 view_vec, vec3 eye_pos_vec, float in_geom_radius)
{
// Requires: 1.) view_vec and eye_pos_vec are 3D vectors in the sphere's
// local coordinate frame (eye_pos_vec is a position, i.e.
// a vector from the origin to the eye/camera)
// 2.) in_geom_radius is a global containing the cylinder's radius
// Returns: Cast a ray of direction view_vec from eye_pos_vec at a
// cylinder of radius in_geom_radius, and return the distance to
// the first intersection in units of length(view_vec). The
// derivation of the coefficients is in Christer Ericson's
// Real-Time Collision Detection, p. 195-196, and this version
// uses LaGrange's identity to reduce operations.
// Arbitrary "cylinder top" reference point for an infinite cylinder:
vec3 cylinder_top_vec = vec3(0, in_geom_radius, 0);
vec3 cylinder_axis_vec = vec3(0, 1, 0);//vec3(0, 2.0*in_geom_radius, 0);
vec3 top_to_eye_vec = eye_pos_vec - cylinder_top_vec;
vec3 axis_x_view = cross(cylinder_axis_vec, view_vec);
vec3 axis_x_top_to_eye = cross(cylinder_axis_vec, top_to_eye_vec);
// Quadratic formula coefficients (b_over_2 is guaranteed negative):
float a = dot(axis_x_view, axis_x_view);
float b_over_2 = dot(axis_x_top_to_eye, axis_x_view);
float c = dot(axis_x_top_to_eye, axis_x_top_to_eye) - in_geom_radius * in_geom_radius; //*dot(cylinder_axis_vec, cylinder_axis_vec);
return hrg_quadratic_solve(a, b_over_2, c);
}
vec2 hrg_cylinder_xyz_to_uv( vec3 intersection_pos_local, vec2 output_aspect, float in_geom_radius)
{
// Requires: An xyz intersection position on a cylinder.
// Returns: video_uv coords mapped to range [-0.5, 0.5]
// Mapping: Define square_uv.x to be the signed arc length in xz-space,
// and define square_uv.y = -intersection_pos_local.y (+v = -y).
// Start with a numerically robust arc length calculation.
float angle_from_image_center = atan(intersection_pos_local.x, intersection_pos_local.z);
float signed_arc_len = angle_from_image_center * in_geom_radius;
// Get a uv-mapping where [-0.5, 0.5] maps to a "square" area, then divide
// by the aspect ratio to stretch the mapping appropriately:
vec2 square_uv = vec2(signed_arc_len, -intersection_pos_local.y);
vec2 video_uv = square_uv / output_aspect;
return video_uv;
}
vec3 hrg_cylinder_uv_to_xyz(vec2 video_uv, vec2 output_aspect, float in_geom_radius)
{
// Requires: video_uv coords mapped to range [-0.5, 0.5]
// Returns: An xyz intersection position on a cylinder. This is the
// inverse of hrg_cylinder_xyz_to_uv().
// Expand video_uv by the aspect ratio to get proportionate x/y lengths,
// then calculate an xyz position for the cylindrical mapping above.
vec2 square_uv = video_uv * output_aspect;
float arc_len = square_uv.x;
float angle_from_image_center = arc_len / in_geom_radius;
float x_pos = sin(angle_from_image_center) * in_geom_radius;
float z_pos = cos(angle_from_image_center) * in_geom_radius;
// Or: z = sqrt(in_geom_radius**2 - x**2)
// Or: z = in_geom_radius/sqrt(1 + tan(angle)**2), x = z * tan(angle)
vec3 intersection_pos_local = vec3(x_pos, -square_uv.y, z_pos);
return intersection_pos_local;
}
vec2 hrg_sphere_xyz_to_uv(vec3 intersection_pos_local, vec2 output_aspect, float in_geom_radius)
{
// Requires: An xyz intersection position on a sphere.
// Returns: video_uv coords mapped to range [-0.5, 0.5]
// Mapping: First define square_uv.x/square_uv.y ==
// intersection_pos_local.x/intersection_pos_local.y. Then,
// length(square_uv) is the arc length from the image center
// at (0, 0, in_geom_radius) along the tangent great circle.
// Credit for this mapping goes to cgwg: I never managed to
// understand his code, but he told me his mapping was based on
// great circle distances when I asked him about it, which
// informed this very similar (almost identical) mapping.
// Start with a numerically robust arc length calculation between the ray-
// sphere intersection point and the image center using a method posted by
// Roger Stafford on comp.soft-sys.matlab:
// https://groups.google.com/d/msg/comp.soft-sys.matlab/zNbUui3bjcA/c0HV_bHSx9cJ
vec3 image_center_pos_local = vec3(0, 0, in_geom_radius);
float cp_len =
length(cross(intersection_pos_local, image_center_pos_local));
float dp = dot(intersection_pos_local, image_center_pos_local);
float angle_from_image_center = atan(cp_len, dp);
float arc_len = angle_from_image_center * in_geom_radius;
// Get a uv-mapping where [-0.5, 0.5] maps to a "square" area, then divide
// by the aspect ratio to stretch the mapping appropriately:
vec2 square_uv_unit = normalize(vec2(intersection_pos_local.x, -intersection_pos_local.y));
vec2 square_uv = arc_len * square_uv_unit;
vec2 video_uv = square_uv / output_aspect;
return video_uv;
}
vec3 hrg_sphere_uv_to_xyz(vec2 video_uv, vec2 output_aspect, float in_geom_radius)
{
// Requires: video_uv coords mapped to range [-0.5, 0.5]
// Returns: An xyz intersection position on a sphere. This is the
// inverse of hrg_sphere_xyz_to_uv().
// Expand video_uv by the aspect ratio to get proportionate x/y lengths,
// then calculate an xyz position for the spherical mapping above.
vec2 square_uv = video_uv * output_aspect;
// Using length or sqrt here butchers the framerate on my 8800GTS if
// this function is called too many times, and so does taking the max
// component of square_uv/square_uv_unit (program length threshold?).
//float arc_len = length(square_uv);
vec2 square_uv_unit = normalize(square_uv);
float arc_len = square_uv.y/square_uv_unit.y;
float angle_from_image_center = arc_len / in_geom_radius;
float xy_dist_from_sphere_center = sin(angle_from_image_center) * in_geom_radius;
//vec2 xy_pos = xy_dist_from_sphere_center * (square_uv/FIX_ZERO(arc_len));
vec2 xy_pos = xy_dist_from_sphere_center * square_uv_unit;
float z_pos = cos(angle_from_image_center) * in_geom_radius;
vec3 intersection_pos_local = vec3(xy_pos.x, -xy_pos.y, z_pos);
return intersection_pos_local;
}
vec2 hrg_sphere_alt_xyz_to_uv(vec3 intersection_pos_local, vec2 output_aspect, float in_geom_radius)
{
// Requires: An xyz intersection position on a cylinder.
// Returns: video_uv coords mapped to range [-0.5, 0.5]
// Mapping: Define square_uv.x to be the signed arc length in xz-space,
// and define square_uv.y == signed arc length in yz-space.
// See hrg_cylinder_xyz_to_uv() for implementation details (very similar).
vec2 angle_from_image_center = atan( vec2(intersection_pos_local.x, -intersection_pos_local.y),
intersection_pos_local.zz);
vec2 signed_arc_len = angle_from_image_center * in_geom_radius;
vec2 video_uv = signed_arc_len / output_aspect;
return video_uv;
}
vec3 hrg_sphere_alt_uv_to_xyz(vec2 video_uv, vec2 output_aspect, float in_geom_radius)
{
// Requires: video_uv coords mapped to range [-0.5, 0.5]
// Returns: An xyz intersection position on a sphere. This is the
// inverse of hrg_sphere_alt_xyz_to_uv().
// See hrg_cylinder_uv_to_xyz() for implementation details (very similar).
vec2 square_uv = video_uv * output_aspect;
vec2 arc_len = square_uv;
vec2 angle_from_image_center = arc_len / in_geom_radius;
vec2 xy_pos = sin(angle_from_image_center) * in_geom_radius;
float z_pos = sqrt(in_geom_radius * in_geom_radius - dot(xy_pos, xy_pos));
return vec3(xy_pos.x, -xy_pos.y, z_pos);
}
vec2 hrg_intersect(vec3 view_vec_local, vec3 eye_pos_local, float in_geom_mode, float in_geom_radius)
{
return in_geom_mode < 2.5 ? hrg_intersect_sphere(view_vec_local, eye_pos_local, in_geom_radius) :
hrg_intersect_cylinder(view_vec_local, eye_pos_local, in_geom_radius);
}
vec2 hrg_xyz_to_uv( vec3 intersection_pos_local, vec2 output_aspect, float in_geom_mode, float in_geom_radius)
{
return in_geom_mode < 1.5 ? hrg_sphere_xyz_to_uv(intersection_pos_local, output_aspect, in_geom_radius) :
in_geom_mode < 2.5 ? hrg_sphere_alt_xyz_to_uv(intersection_pos_local, output_aspect, in_geom_radius) :
hrg_cylinder_xyz_to_uv(intersection_pos_local, output_aspect, in_geom_radius);
}
vec3 hrg_uv_to_xyz(vec2 uv, vec2 output_aspect, float in_geom_mode, float in_geom_radius)
{
return in_geom_mode < 1.5 ? hrg_sphere_uv_to_xyz(uv, output_aspect, in_geom_radius) :
in_geom_mode < 2.5 ? hrg_sphere_alt_uv_to_xyz(uv, output_aspect, in_geom_radius) :
hrg_cylinder_uv_to_xyz(uv, output_aspect, in_geom_radius);
}
vec2 hrg_view_vec_to_uv(vec3 view_vec_local,
vec3 eye_pos_local,
vec2 output_aspect,
float in_geom_mode,
float in_geom_radius,
out vec3 intersection_pos)
{
// Get the intersection point on the primitive, given an eye position
// and view vector already in its local coordinate frame:
vec2 intersect_dist_and_discriminant = hrg_intersect(view_vec_local, eye_pos_local, in_geom_mode, in_geom_radius);
vec3 intersection_pos_local = eye_pos_local + view_vec_local * intersect_dist_and_discriminant.x;
// Save the intersection position to an output parameter:
intersection_pos = intersection_pos_local;
// Transform into uv coords, but give out-of-range coords if the
// view ray doesn't hrg_intersect the primitive in the first place:
return intersect_dist_and_discriminant.y > 0.005 ? hrg_xyz_to_uv(intersection_pos_local, output_aspect, in_geom_mode, in_geom_radius) :
vec2(1);
}
vec3 hrg_get_ideal_global_eye_pos_for_points( vec3 eye_pos,
vec2 output_aspect,
vec3 global_coords[HRG_MAX_POINT_CLOUD_SIZE],
int num_points,
float in_geom_radius,
float in_geom_view_dist)
{
// Requires: Parameters:
// 1.) Starting eye_pos is a global 3D position at which the
// camera contains all points in global_coords[] in its FOV
// 2.) output_aspect = hrg_get_aspect_vector(
// IN.OutputSize.xy.x / IN.OutputSize.xy.y);
// 3.) global_coords is a point cloud containing global xyz
// coords of extreme points on the simulated CRT screen.
// Globals:
// 1.) in_geom_view_dist must be > 0. It controls the "near
// plane" used to interpret flat_video_uv as a view
// vector, which controls the field of view (FOV).
// Eyespace coordinate frame: +x = right, +y = up, +z = back
// Returns: Return an eye position at which the point cloud spans as
// much of the screen as possible (given the FOV controlled by
// in_geom_view_dist) without being cropped or sheared.
// Algorithm:
// 1.) Move the eye laterally to a point which attempts to maximize the
// the amount we can move forward without clipping the CRT screen.
// 2.) Move forward by as much as possible without clipping the CRT.
// Get the allowed movement range by solving for the eye_pos offsets
// that result in each point being projected to a screen edge/corner in
// pseudo-normalized device coords (where xy ranges from [-0.5, 0.5]
// and z = eyespace z):
// pndc_coord = vec3(vec2(eyespace_xyz.x, -eyespace_xyz.y)*
// in_geom_view_dist / (output_aspect * -eyespace_xyz.z), eyespace_xyz.z);
// Notes:
// The field of view is controlled by in_geom_view_dist's magnitude relative to
// the view vector's x and y components:
// view_vec.xy ranges from [-0.5, 0.5] * output_aspect
// view_vec.z = -in_geom_view_dist
// But for the purposes of perspective divide, it should be considered:
// view_vec.xy ranges from [-0.5, 0.5] * output_aspect / in_geom_view_dist
// view_vec.z = -1
int max_centering_iters = 1; // Keep for easy testing.
for(int iter = 0; iter < max_centering_iters; iter++)
{
// 0.) Get the eyespace coordinates of our point cloud:
vec3 eyespace_coords[HRG_MAX_POINT_CLOUD_SIZE];
for(int i = 0; i < num_points; i++)
{
eyespace_coords[i] = global_coords[i] - eye_pos;
}
// 1a.)For each point, find out how far we can move eye_pos in each
// lateral direction without the point clipping the frustum.
// Eyespace +y = up, screenspace +y = down, so flip y after
// applying the eyespace offset (on the way to "clip space").
// Solve for two offsets per point based on:
// (eyespace_xyz.xy - offset_dr) * vec2(1, -1) *
// in_geom_view_dist / (output_aspect * -eyespace_xyz.z) = vec2(-0.5)
// (eyespace_xyz.xy - offset_dr) * vec2(1, -1) *
// in_geom_view_dist / (output_aspect * -eyespace_xyz.z) = vec2(0.5)
// offset_ul and offset_dr represent the farthest we can move the
// eye_pos up-left and down-right. Save the min of all offset_dr's
// and the max of all offset_ul's (since it's negative).
float abs_radius = abs(in_geom_radius); // In case anyone gets ideas. ;)
vec2 offset_dr_min = vec2(10.0 * abs_radius, 10.0 * abs_radius);
vec2 offset_ul_max = vec2(-10.0 * abs_radius, -10.0 * abs_radius);
for(int i = 0; i < num_points; i++)
{
vec2 flipy = vec2(1, -1);
vec3 eyespace_xyz = eyespace_coords[i];
vec2 offset_dr = eyespace_xyz.xy - vec2(-0.5) *
(output_aspect * -eyespace_xyz.z) /
(in_geom_view_dist * flipy);
vec2 offset_ul = eyespace_xyz.xy - vec2(0.5) *
(output_aspect * -eyespace_xyz.z) /
(in_geom_view_dist * flipy);
offset_dr_min = min(offset_dr_min, offset_dr);
offset_ul_max = max(offset_ul_max, offset_ul);
}
// 1b.)Update eye_pos: Adding the average of offset_ul_max and
// offset_dr_min gives it equal leeway on the top vs. bottom
// and left vs. right. Recalculate eyespace_coords accordingly.
vec2 center_offset = 0.5 * (offset_ul_max + offset_dr_min);
eye_pos.xy += center_offset;
for(int i = 0; i < num_points; i++)
{
eyespace_coords[i] = global_coords[i] - eye_pos;
}
// 2a.)For each point, find out how far we can move eye_pos forward
// without the point clipping the frustum. Flip the y
// direction in advance (matters for a later step, not here).
// Solve for four offsets per point based on:
// eyespace_xyz_flipy.x * in_geom_view_dist /
// (output_aspect.x * (offset_z - eyespace_xyz_flipy.z)) =-0.5
// eyespace_xyz_flipy.y * in_geom_view_dist /
// (output_aspect.y * (offset_z - eyespace_xyz_flipy.z)) =-0.5
// eyespace_xyz_flipy.x * in_geom_view_dist /
// (output_aspect.x * (offset_z - eyespace_xyz_flipy.z)) = 0.5
// eyespace_xyz_flipy.y * in_geom_view_dist /
// (output_aspect.y * (offset_z - eyespace_xyz_flipy.z)) = 0.5
// We'll vectorize the actual computation. Take the maximum of
// these four for a single offset, and continue taking the max
// for every point (use max because offset.z is negative).
float offset_z_max = -10.0 * in_geom_radius * in_geom_view_dist;
for(int i = 0; i < num_points; i++)
{
vec3 eyespace_xyz_flipy = eyespace_coords[i] * vec3(1, -1, 1);
vec4 offset_zzzz = eyespace_xyz_flipy.zzzz +
(eyespace_xyz_flipy.xyxy * in_geom_view_dist) /
(vec4(-0.5, -0.5, 0.5, 0.5) * vec4(output_aspect, output_aspect));
// Ignore offsets that push positive x/y values to opposite
// boundaries, and vice versa, and don't let the camera move
// past a point in the dead center of the screen:
offset_z_max = (eyespace_xyz_flipy.x < 0) ? max(offset_z_max, offset_zzzz.x) : offset_z_max;
offset_z_max = (eyespace_xyz_flipy.y < 0) ? max(offset_z_max, offset_zzzz.y) : offset_z_max;
offset_z_max = (eyespace_xyz_flipy.x > 0) ? max(offset_z_max, offset_zzzz.z) : offset_z_max;
offset_z_max = (eyespace_xyz_flipy.y > 0) ? max(offset_z_max, offset_zzzz.w) : offset_z_max;
offset_z_max = max(offset_z_max, eyespace_xyz_flipy.z);
}
// 2b.)Update eye_pos: Add the maximum (smallest negative) z offset.
eye_pos.z += offset_z_max;
}
return eye_pos;
}
vec3 hrg_get_ideal_global_eye_pos( mat3x3 local_to_global,
vec2 output_aspect,
float in_geom_mode,
float in_geom_radius,
float in_geom_view_dist)
{
// Start with an initial eye_pos that includes the entire primitive
// (sphere or cylinder) in its field-of-view:
vec3 high_view = vec3(0, output_aspect.y, -in_geom_view_dist);
vec3 low_view = high_view * vec3(1, -1, 1);
float len_sq = dot(high_view, high_view);
float fov = abs(acos(dot(high_view, low_view)/len_sq));
// Trigonometry/similar triangles say distance = in_geom_radius/sin(fov/2):
float eye_z_spherical = in_geom_radius / sin(fov*0.5);
vec3 eye_pos = in_geom_mode < 2.5 ? vec3(0, 0, eye_z_spherical) :
vec3(0, 0, max(in_geom_view_dist, eye_z_spherical));
// Get global xyz coords of extreme sample points on the simulated CRT
// screen. Start with the center, edge centers, and corners of the
// video image. We can't ignore backfacing points: They're occluded
// by closer points on the primitive, but they may NOT be occluded by
// the convex hull of the remaining samples (i.e. the remaining convex
// hull might not envelope points that do occlude a back-facing point.)
int num_points = HRG_MAX_POINT_CLOUD_SIZE;
vec3 global_coords[HRG_MAX_POINT_CLOUD_SIZE];
global_coords[0] = hrg_uv_to_xyz(vec2(0, 0), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
global_coords[1] = hrg_uv_to_xyz(vec2(0, -0.5), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
global_coords[2] = hrg_uv_to_xyz(vec2(0, 0.5), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
global_coords[3] = hrg_uv_to_xyz(vec2(-0.5, 0), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
global_coords[4] = hrg_uv_to_xyz(vec2(0.5, 0), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
global_coords[5] = hrg_uv_to_xyz(vec2(-0.5, -0.5), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
global_coords[6] = hrg_uv_to_xyz(vec2(0.5, -0.5), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
global_coords[7] = hrg_uv_to_xyz(vec2(-0.5, 0.5), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
global_coords[8] = hrg_uv_to_xyz(vec2(0.5, 0.5), output_aspect, in_geom_mode, in_geom_radius) * local_to_global;
// Adding more inner image points could help in extreme cases, but too many
// points will kille the framerate. For safety, default to the initial
// eye_pos if any z coords are negative:
float num_negative_z_coords = 0;
for(int i = 0; i < num_points; i++)
{
num_negative_z_coords += float(global_coords[0].z < 0);
}
// Outsource the optimized eye_pos calculation:
return num_negative_z_coords > 0.5 ? eye_pos :
hrg_get_ideal_global_eye_pos_for_points(eye_pos,
output_aspect,
global_coords,
num_points,
in_geom_radius,
in_geom_view_dist);
}
mat3x3 hrg_get_pixel_to_object_matrix( mat3x3 global_to_local,
vec3 eye_pos_local,
vec3 view_vec_global,
vec3 intersection_pos_local,
vec3 normal,
vec2 output_pixel_size)
{
// Requires: See hrg_get_curved_video_uv_coords_and_tangent_matrix for
// descriptions of each parameter.
// Returns: Return a transformation matrix from 2D pixel-space vectors
// (where (+1, +1) is a vector to one pixel down-right,
// i.e. same directionality as uv texels) to 3D object-space
// vectors in the CRT's local coordinate frame (right-handed)
// ***which are tangent to the CRT surface at the intersection
// position.*** (Basically, we want to convert pixel-space
// vectors to 3D vectors along the CRT's surface, for later
// conversion to uv vectors.)
// Shorthand inputs:
vec3 pos = intersection_pos_local;
vec3 eye_pos = eye_pos_local;
// Get a piecewise-linear matrix transforming from "pixelspace" offset
// vectors (1 = one pixel) to object space vectors in the tangent
// plane (faster than finding 3 view-object intersections).
// 1.) Get the local view vecs for the pixels to the right and down:
vec3 view_vec_right_global = view_vec_global + vec3(output_pixel_size.x, 0, 0);
vec3 view_vec_down_global = view_vec_global + vec3(0, -output_pixel_size.y, 0);
vec3 view_vec_right_local = view_vec_right_global * global_to_local;
vec3 view_vec_down_local = view_vec_down_global * global_to_local;
// 2.) Using the true intersection point, hrg_intersect the neighboring
// view vectors with the tangent plane:
vec3 intersection_vec_dot_normal = vec3(dot(pos - eye_pos, normal));
vec3 right_pos = eye_pos +
(intersection_vec_dot_normal / dot(view_vec_right_local, normal)) *
view_vec_right_local;
vec3 down_pos = eye_pos +
(intersection_vec_dot_normal / dot(view_vec_down_local, normal)) *
view_vec_down_local;
// 3.) Subtract the original intersection pos from its neighbors; the
// resulting vectors are object-space vectors tangent to the plane.
// These vectors are the object-space transformations of (1, 0)
// and (0, 1) pixel offsets, so they form the first two basis
// vectors of a pixelspace to object space transformation. This
// transformation is 2D to 3D, so use (0, 0, 0) for the third vector.
vec3 object_right_vec = right_pos - pos;
vec3 object_down_vec = down_pos - pos;
mat3x3 pixel_to_object = mat3x3(
object_right_vec.x, object_down_vec.x, 0,
object_right_vec.y, object_down_vec.y, 0,
object_right_vec.z, object_down_vec.z, 0
);
return pixel_to_object;
}
mat3x3 hrg_get_object_to_tangent_matrix(vec3 intersection_pos_local,
vec3 normal,
vec2 output_aspect,
float in_geom_mode)
{
// Requires: See hrg_get_curved_video_uv_coords_and_tangent_matrix for
// descriptions of each parameter.
// Returns: Return a transformation matrix from 3D object-space vectors
// in the CRT's local coordinate frame (right-handed, +y = up)
// to 2D video_uv vectors (+v = down).
// Description:
// The TBN matrix formed by the [tangent, bitangent, normal] basis
// vectors transforms ordinary vectors from tangent->object space.
// The cotangent matrix formed by the [cotangent, cobitangent, normal]
// basis vectors transforms normal vectors (covectors) from
// tangent->object space. It's the inverse-transpose of the TBN matrix.
// We want the inverse of the TBN matrix (transpose of the cotangent
// matrix), which transforms ordinary vectors from object->tangent space.
// Start by calculating the relevant basis vectors in accordance with
// Christian Schüler's blog post "Followup: Normal Mapping Without
// Precomputed Tangents": http://www.thetenthplanet.de/archives/1180
// With our particular uv mapping, the scale of the u and v directions
// is determined entirely by the aspect ratio for cylindrical and ordinary
// spherical mappings, and so tangent and bitangent lengths are also
// determined by it (the alternate mapping is more complex). Therefore, we
// must ensure appropriate cotangent and cobitangent lengths as well.
// Base these off the uv<=>xyz mappings for each primitive.
vec3 pos = intersection_pos_local;
vec3 x_vec = vec3(1, 0, 0);
vec3 y_vec = vec3(0, 1, 0);
// The tangent and bitangent vectors correspond with increasing u and v,
// respectively. Mathematically we'd base the cotangent/cobitangent on
// those, but we'll compute the cotangent/cobitangent directly when we can.
vec3 cotangent_unscaled;
vec3 cobitangent_unscaled;
// in_geom_mode should be constant-folded without RUNTIME_GEOMETRY_MODE.
if(in_geom_mode < 1.5)
{
// Sphere:
// tangent = normalize(cross(normal, cross(x_vec, pos))) * output_aspect.x
// bitangent = normalize(cross(cross(y_vec, pos), normal)) * output_aspect.y
// inv_determinant = 1/length(cross(bitangent, tangent))
// cotangent = cross(normal, bitangent) * inv_determinant
// == normalize(cross(y_vec, pos)) * output_aspect.y * inv_determinant
// cobitangent = cross(tangent, normal) * inv_determinant
// == normalize(cross(x_vec, pos)) * output_aspect.x * inv_determinant
// Simplified (scale by inv_determinant below):
cotangent_unscaled = normalize(cross(y_vec, pos)) * output_aspect.y;
cobitangent_unscaled = normalize(cross(x_vec, pos)) * output_aspect.x;
}
else if(in_geom_mode < 2.5)
{
// Sphere, alternate mapping:
// This mapping works a bit like the cylindrical mapping in two
// directions, which makes the lengths and directions more complex.
// Unfortunately, I can't find much of a shortcut:
vec3 tangent = normalize(cross(y_vec, vec3(pos.x, 0, pos.z))) * output_aspect.x;
vec3 bitangent = normalize(cross(x_vec, vec3(0, pos.yz))) * output_aspect.y;
cotangent_unscaled = cross(normal, bitangent);
cobitangent_unscaled = cross(tangent, normal);
}
else
{
// Cylinder:
// tangent = normalize(cross(y_vec, normal)) * output_aspect.x;
// bitangent = vec3(0, -output_aspect.y, 0);
// inv_determinant = 1/length(cross(bitangent, tangent))
// cotangent = cross(normal, bitangent) * inv_determinant
// == normalize(cross(y_vec, pos)) * output_aspect.y * inv_determinant
// cobitangent = cross(tangent, normal) * inv_determinant
// == vec3(0, -output_aspect.x, 0) * inv_determinant
cotangent_unscaled = cross(y_vec, normal) * output_aspect.y;
cobitangent_unscaled = vec3(0, -output_aspect.x, 0);
}
vec3 computed_normal = cross(cobitangent_unscaled, cotangent_unscaled);
float inv_determinant = inversesqrt(dot(computed_normal, computed_normal));
vec3 cotangent = cotangent_unscaled * inv_determinant;
vec3 cobitangent = cobitangent_unscaled * inv_determinant;
// The [cotangent, cobitangent, normal] column vecs form the cotangent
// frame, i.e. the inverse-transpose TBN matrix. Get its transpose:
mat3x3 object_to_tangent = mat3x3( cotangent,
cobitangent,
normal);
return object_to_tangent;
}
vec2 hrg_get_curved_video_uv_coords_and_tangent_matrix( vec2 flat_video_uv,
vec3 eye_pos_local,
vec2 output_pixel_size,
vec2 output_aspect,
float in_geom_mode,
float in_geom_radius,
float in_geom_view_dist,
mat3x3 global_to_local,
out mat2x2 pixel_to_tangent_video_uv)
{
// Requires: Parameters:
// 1.) flat_video_uv coords are in range [0, 1], where
// (0, 0) is the top-left corner of the screen and
// (1, 1) is the bottom-right corner.
// 2.) eye_pos_local is the 3D camera position in the simulated
// CRT's local coordinate frame. For best results, it must
// be computed based on the same in_geom_view_dist used here.
// 3.) output_pixel_size = vec2(1)/IN.OutputSize.xy
// 4.) output_aspect = hrg_get_aspect_vector(
// IN.OutputSize.xy.x / IN.OutputSize.xy.y);
// 5.) in_geom_mode is a static or runtime mode setting:
// 0 = off, 1 = sphere, 2 = sphere alt., 3 = cylinder
// 6.) global_to_local is a 3x3 matrix transforming (ordinary)
// worldspace vectors to the CRT's local coordinate frame
// Globals:
// 1.) in_geom_view_dist must be > 0. It controls the "near
// plane" used to interpret flat_video_uv as a view
// vector, which controls the field of view (FOV).
// Returns: Return final uv coords in [0, 1], and return a pixel-
// space to video_uv tangent-space matrix in the out parameter.
// (This matrix assumes pixel-space +y = down, like +v = down.)
// We'll transform flat_video_uv into a view vector, project
// the view vector from the camera/eye, hrg_intersect with a sphere
// or cylinder representing the simulated CRT, and convert the
// intersection position into final uv coords and a local
// transformation matrix.
// First get the 3D view vector (output_aspect and in_geom_view_dist are globals):
// 1.) Center uv around (0, 0) and make (-0.5, -0.5) and (0.5, 0.5)
// correspond to the top-left/bottom-right output screen corners.
// 2.) Multiply by output_aspect to preemptively "undo" Retroarch's screen-
// space 2D aspect correction. We'll reapply it in uv-space.
// 3.) (x, y) = (u, -v), because +v is down in 2D screenspace, but +y
// is up in 3D worldspace (enforce a right-handed system).
// 4.) The view vector z controls the "near plane" distance and FOV.
// For the effect of "looking through a window" at a CRT, it should be
// set equal to the user's distance from their physical screen, in
// units of the viewport's physical diagonal size.
vec2 view_uv = (flat_video_uv - vec2(0.5)) * output_aspect;
vec3 view_vec_global = vec3(view_uv.x, -view_uv.y, -in_geom_view_dist);
// Transform the view vector into the CRT's local coordinate frame, convert
// to video_uv coords, and get the local 3D intersection position:
vec3 view_vec_local = view_vec_global * global_to_local;
vec3 pos;
vec2 centered_uv = hrg_view_vec_to_uv( view_vec_local,
eye_pos_local,
output_aspect,
in_geom_mode,
in_geom_radius,
pos);
vec2 video_uv = centered_uv + vec2(0.5);
// Get a pixel-to-tangent-video-uv matrix. The caller could deal with
// all but one of these cases, but that would be more complicated.
//#ifdef DRIVERS_ALLOW_DERIVATIVES
// Derivatives obtain a matrix very fast, but the direction of pixel-
// space +y seems to depend on the pass. Enforce the correct direction
// on a best-effort basis (but it shouldn't matter for antialiasing).
// vec2 duv_dx = dFdx(video_uv);
// vec2 duv_dy = dFdy(video_uv);
// // #ifdef LAST_PASS
// pixel_to_tangent_video_uv = mat2x2( duv_dx.x, duv_dy.x,
// -duv_dx.y, -duv_dy.y);
// #else
// pixel_to_tangent_video_uv = mat2x2( duv_dx.x, duv_dy.x,
// duv_dx.y, duv_dy.y);
// #endif
// #else
// Manually define a transformation matrix. We'll assume pixel-space
// +y = down, just like +v = down.
bool geom_force_correct_tangent_matrix = true;
if(geom_force_correct_tangent_matrix)
{
// Get the surface normal based on the local intersection position:
vec3 normal_base = in_geom_mode < 2.5 ? pos :
vec3(pos.x, 0, pos.z);
vec3 normal = normalize(normal_base);
// Get pixel-to-object and object-to-tangent matrices and combine
// them into a 2x2 pixel-to-tangent matrix for video_uv offsets:
mat3x3 pixel_to_object = hrg_get_pixel_to_object_matrix(global_to_local,
eye_pos_local,
view_vec_global,
pos,
normal,
output_pixel_size);
mat3x3 object_to_tangent = hrg_get_object_to_tangent_matrix(pos, normal, output_aspect, in_geom_mode);
mat3x3 pixel_to_tangent3x3 = pixel_to_object * object_to_tangent;
pixel_to_tangent_video_uv = mat2x2( pixel_to_tangent3x3[0][0], pixel_to_tangent3x3[0][1],
pixel_to_tangent3x3[1][0], pixel_to_tangent3x3[1][1]);//._m00_m01_m10_m11);
}
else
{
// Ignore curvature, and just consider flat scaling. The
// difference is only apparent with strong curvature:
pixel_to_tangent_video_uv = mat2x2( output_pixel_size.x, 0,
0, output_pixel_size.y);
}
//#endif
return video_uv;
}
float HRG_GetBorderDimFactor(vec2 video_uv, vec2 output_aspect, float in_border_size, float in_border_darkness, float in_border_compress)
{
// COPYRIGHT NOTE FOR THIS FUNCTION:
// Copyright (C) 2010-2012 cgwg, 2014 TroggleMonkey
// This function uses an algorithm first coded in several of cgwg's GPL-
// licensed lines in crt-geom-curved.cg and its ancestors.
// Calculate border_dim_factor from the proximity to uv-space image
// borders; output_aspect/in_border_size/border/darkness/in_border_compress are globals:
vec2 edge_dists = min(video_uv, vec2(1) - video_uv) * output_aspect;
vec2 border_penetration = max(vec2(in_border_size) - edge_dists, vec2(0));
float penetration_ratio = length(border_penetration)/in_border_size;
float border_escape_ratio = max(1 - penetration_ratio, 0);
float border_dim_factor = pow(border_escape_ratio, in_border_darkness) * max(1, in_border_compress);
return min(border_dim_factor, 1);
}
// Provide accessors for vector constants that pack scalar uniforms:
vec2 hrg_get_aspect_vector(float geom_aspect_ratio)
{
// Get an aspect ratio vector. Enforce geom_max_aspect_ratio, and prevent
// the absolute scale from affecting the uv-mapping for curvature:
float geom_max_aspect_ratio = 4/3;
float geom_clamped_aspect_ratio = min(geom_aspect_ratio, geom_max_aspect_ratio);
vec2 output_aspect = normalize(vec2(geom_clamped_aspect_ratio, 1));
return output_aspect;
}
vec2 HRG_GetGeomCurvedCoord( vec2 in_coord,
float in_geom_mode,
float in_geom_radius,
float in_geom_view_dist,
float in_geom_tilt_angle_x,
float in_geom_tilt_angle_y,
float in_screen_aspect,
float pin_inner_edge,
vec2 in_source_size,
vec2 in_output_size,
out mat2x2 pixel_to_video_uv)
{
vec2 output_pixel_size = vec2(1.0, 1.0) / in_output_size;
float geom_radius_scaled = in_geom_radius;
vec2 output_aspect = hrg_get_aspect_vector(in_screen_aspect);
// Create a local-to-global rotation matrix for the CRT's coordinate
// frame and its global-to-local inverse. Rotate around the x axis
// first (pitch) and then the y axis (yaw) with yucky Euler angles.
// Positive angles go clockwise around the right-vec and up-vec.
vec2 geom_tilt_angle = vec2(in_geom_tilt_angle_x, in_geom_tilt_angle_y);
vec2 sin_tilt = sin(geom_tilt_angle);
vec2 cos_tilt = cos(geom_tilt_angle);
// Conceptual breakdown:
mat3x3 rot_x_matrix = mat3x3( 1, 0, 0,
0, cos_tilt.y, -sin_tilt.y,
0, sin_tilt.y, cos_tilt.y);
mat3x3 rot_y_matrix = mat3x3( cos_tilt.x, 0, sin_tilt.x,
0, 1, 0,
-sin_tilt.x, 0, cos_tilt.x);
mat3x3 local_to_global = rot_x_matrix * rot_y_matrix;
// This is a pure rotation, so transpose = inverse:
mat3x3 global_to_local = transpose(local_to_global);
// Get an optimal eye position based on in_geom_view_dist, viewport_aspect,
// and CRT radius/rotation:
vec3 eye_pos_global = hrg_get_ideal_global_eye_pos( local_to_global,
output_aspect,
in_geom_mode,
geom_radius_scaled,
in_geom_view_dist);
vec3 eye_pos_local = eye_pos_global * global_to_local;
vec2 curved_coord;
if(in_geom_mode > 0.5)
{
// Put in a test for the projection with a flat plane to compare
// with the distorted coordinate to scale out to the edges of the flat plane
// Also helps with cyndrilical projection where the sides shift in towards the center
vec2 ctr_curved_coord = hrg_get_curved_video_uv_coords_and_tangent_matrix( in_coord,
eye_pos_local,
output_pixel_size,
output_aspect,
in_geom_mode,
geom_radius_scaled,
in_geom_view_dist,
global_to_local,
pixel_to_video_uv) - 0.5;
// Curvature can cause the screen to shrink so we want to scale it back out so it is the same width & height
// Especially helps with cylindrical projection which shrinks a lot
// Right Edge should end up at 1, we scale it back out so it hits 1
// Only do this when not using tilt so we don't mess up what the perspective is doing
if (in_geom_tilt_angle_x == 0 && in_geom_tilt_angle_y == 0)
{
vec2 right_edge_curved_ctr_coord = hrg_get_curved_video_uv_coords_and_tangent_matrix(vec2(1, 0.5),
eye_pos_local,
output_pixel_size,
output_aspect,
in_geom_mode,
geom_radius_scaled,
in_geom_view_dist,
global_to_local,
pixel_to_video_uv) - 0.5;
vec2 bottom_edge_curved_ctr_coord = hrg_get_curved_video_uv_coords_and_tangent_matrix(vec2(0.5, 1),
eye_pos_local,
output_pixel_size,
output_aspect,
in_geom_mode,
geom_radius_scaled,
in_geom_view_dist,
global_to_local,
pixel_to_video_uv) - 0.5;
ctr_curved_coord.x = ctr_curved_coord.x * 0.5 / right_edge_curved_ctr_coord.x;
ctr_curved_coord.y = ctr_curved_coord.y * 0.5 / bottom_edge_curved_ctr_coord.y;
}
if (pin_inner_edge == 1)
{
if (in_geom_tilt_angle_y != 0)
{
vec2 top_edge_curved_ctr_coord = hrg_get_curved_video_uv_coords_and_tangent_matrix(vec2(0.5, 0),
eye_pos_local,
output_pixel_size,
output_aspect,
in_geom_mode,
geom_radius_scaled,
in_geom_view_dist,
global_to_local,
pixel_to_video_uv);
ctr_curved_coord.y = ctr_curved_coord.y - top_edge_curved_ctr_coord.y;
}
if (in_geom_tilt_angle_x != 0)
{
vec2 left_edge_curved_ctr_coord = hrg_get_curved_video_uv_coords_and_tangent_matrix(vec2(0, 0.5),
eye_pos_local,
output_pixel_size,
output_aspect,
in_geom_mode,
geom_radius_scaled,
in_geom_view_dist,
global_to_local,
pixel_to_video_uv);
ctr_curved_coord.x = ctr_curved_coord.x - left_edge_curved_ctr_coord.x;
}
}
curved_coord = ctr_curved_coord + 0.5;
}
else
{
curved_coord = in_coord;
pixel_to_video_uv = mat2x2( output_pixel_size.x, 0,
0, output_pixel_size.y);
}
return curved_coord;
}