diff options
author | Alexander Gavrilov <angavrilov@gmail.com> | 2019-09-04 10:10:27 +0300 |
---|---|---|
committer | Alexander Gavrilov <angavrilov@gmail.com> | 2019-09-04 10:34:21 +0300 |
commit | fcf2a712eca278f198e8d61845f2078088af1f9a (patch) | |
tree | 77c6c3e0f60cba4f1fb3669029487e922514d7dd /source/blender/blenlib | |
parent | 64f8f7db7cce0c8923afcabb523f7400f744387e (diff) |
Armature: add Inherit Scale options to remove shear or average the scale.
As an inherent property of matrix-based transformation math, non-
uniform scaling of a parent bone induces shear into the transform
matrix of any rotated child. Such matrices cannot be cleanly
decomposed into a combination of location/rotation/scale, which
causes issues for rigging and animation tools.
Blender bones have options to exclude rotation and/or scale from the
inherited transformation, but don't have any support for removing the
often undesired shear component. That goal requires replacing simple
parenting with a combination of multiple bones and constraints. The
same is true about the goal of inheriting some scale, but completely
avoiding shear.
This patch replaces the old Inherit Scale checkbox with a enum that
supports multiple options:
* Full: inherit all effects of scale, like with enabled Inherit Scale.
* Fix Shear: removes shear from the final inherited transformation.
The cleanup math is specifically designed to preserve the main
axis of the bone, its length and total volume, and minimally
affect roll on average. It however will not prevent reappearance
of shear due to local rotation of the child or its children.
* Average: inherit uniform scale that represents the parent volume.
This is the simplest foolproof solution that will inherit some
scale without ever causing shear.
* None: completely remove scale and shear.
* None (Legacy): old disabled Inherit Scale checkbox.
This mode does not handle parent shear in any way, so the child
is likely to end up having both scale and shear. It is retained
for backward compatibility.
Since many rigging-related addons access the use_inherit_scale
property from Python, it is retained as a backward compatibility
stub that provides the old functionality.
As a side effect of reworking the code, this also fixes a matrix
multiplication order bug in the Inherit Rotation code, which caused
the parent local scale to be applied in world space. In rigger
opinion this option is useless in production rigs, so this fix
should not be a problem.
Reviewers: brecht
Differential Revision: https://developer.blender.org/D5588
Diffstat (limited to 'source/blender/blenlib')
-rw-r--r-- | source/blender/blenlib/BLI_math_matrix.h | 6 | ||||
-rw-r--r-- | source/blender/blenlib/intern/math_matrix.c | 126 |
2 files changed, 132 insertions, 0 deletions
diff --git a/source/blender/blenlib/BLI_math_matrix.h b/source/blender/blenlib/BLI_math_matrix.h index f5d87667b73..00c301a01bc 100644 --- a/source/blender/blenlib/BLI_math_matrix.h +++ b/source/blender/blenlib/BLI_math_matrix.h @@ -252,6 +252,9 @@ void normalize_m4_m4(float R[4][4], const float A[4][4]) ATTR_NONNULL(); void orthogonalize_m3(float R[3][3], int axis); void orthogonalize_m4(float R[4][4], int axis); +void orthogonalize_m3_stable(float R[3][3], int axis, bool normalize); +void orthogonalize_m4_stable(float R[4][4], int axis, bool normalize); + bool is_orthogonal_m3(const float mat[3][3]); bool is_orthogonal_m4(const float mat[4][4]); bool is_orthonormal_m3(const float mat[3][3]); @@ -303,8 +306,11 @@ void size_to_mat4(float R[4][4], const float size[3]); void mat3_to_size(float r[3], const float M[3][3]); void mat4_to_size(float r[3], const float M[4][4]); +void mat4_to_size_fix_shear(float r[3], const float M[4][4]); + void translate_m4(float mat[4][4], float tx, float ty, float tz); void rotate_m4(float mat[4][4], const char axis, const float angle); +void rescale_m4(float mat[4][4], float scale[3]); void transform_pivot_set_m4(float mat[4][4], const float pivot[3]); void mat3_to_rot_size(float rot[3][3], float size[3], const float mat3[3][3]); diff --git a/source/blender/blenlib/intern/math_matrix.c b/source/blender/blenlib/intern/math_matrix.c index 2568d42566c..7c1b44978e2 100644 --- a/source/blender/blenlib/intern/math_matrix.c +++ b/source/blender/blenlib/intern/math_matrix.c @@ -1516,6 +1516,111 @@ void orthogonalize_m4(float mat[4][4], int axis) mul_v3_fl(mat[2], size[2]); } +/** Make an orthonormal basis around v1 in a way that is stable and symmetric. */ +static void orthogonalize_stable(float v1[3], float v2[3], float v3[3], bool normalize) +{ + /* Make secondary axis vectors orthogonal to the primary via + * plane projection, which preserves the determinant. */ + float len_sq_v1 = len_squared_v3(v1); + + if (len_sq_v1 > 0.0f) { + madd_v3_v3fl(v2, v1, -dot_v3v3(v2, v1) / len_sq_v1); + madd_v3_v3fl(v3, v1, -dot_v3v3(v3, v1) / len_sq_v1); + + if (normalize) { + mul_v3_fl(v1, 1.0f / sqrtf(len_sq_v1)); + } + } + + /* Make secondary axis vectors orthogonal relative to each other. */ + float norm_v2[3], norm_v3[3], tmp[3]; + float length_v2 = normalize_v3_v3(norm_v2, v2); + float length_v3 = normalize_v3_v3(norm_v3, v3); + float cos_angle = dot_v3v3(norm_v2, norm_v3); + float abs_cos_angle = fabsf(cos_angle); + + /* Apply correction if the shear angle is significant, and not degenerate. */ + if (abs_cos_angle > 1e-4f && abs_cos_angle < 1.0f - FLT_EPSILON) { + /* Adjust v2 by half of the necessary angle correction. + * Thus the angle change is the same for both axis directions. */ + float angle = acosf(cos_angle); + float target_angle = angle + ((float)M_PI_2 - angle) / 2; + + madd_v3_v3fl(norm_v2, norm_v3, -cos_angle); + mul_v3_fl(norm_v2, sinf(target_angle) / len_v3(norm_v2)); + madd_v3_v3fl(norm_v2, norm_v3, cosf(target_angle)); + + /* Make v3 orthogonal. */ + cross_v3_v3v3(tmp, norm_v2, norm_v3); + cross_v3_v3v3(norm_v3, tmp, norm_v2); + normalize_v3(norm_v3); + + /* Re-apply scale, preserving area and proportion. */ + if (!normalize) { + float scale_fac = sqrtf(sinf(angle)); + mul_v3_v3fl(v2, norm_v2, length_v2 * scale_fac); + mul_v3_v3fl(v3, norm_v3, length_v3 * scale_fac); + } + } + + if (normalize) { + copy_v3_v3(v2, norm_v2); + copy_v3_v3(v3, norm_v3); + } +} + +/** + * Make an orthonormal matrix around the selected axis of the given matrix, + * in a way that is symmetric and stable to variations in the input, and + * preserving the value of the determinant, i.e. the overall volume change. + * + * \param axis: Axis to build the orthonormal basis around. + * \param normalize: Normalize the matrix instead of preserving volume. + */ +void orthogonalize_m3_stable(float R[3][3], int axis, bool normalize) +{ + switch (axis) { + case 0: + orthogonalize_stable(R[0], R[1], R[2], normalize); + break; + case 1: + orthogonalize_stable(R[1], R[0], R[2], normalize); + break; + case 2: + orthogonalize_stable(R[2], R[0], R[1], normalize); + break; + default: + BLI_assert(0); + break; + } +} + +/** + * Make an orthonormal matrix around the selected axis of the given matrix, + * in a way that is symmetric and stable to variations in the input, and + * preserving the value of the determinant, i.e. the overall volume change. + * + * \param axis: Axis to build the orthonormal basis around. + * \param normalize: Normalize the matrix instead of preserving volume. + */ +void orthogonalize_m4_stable(float R[4][4], int axis, bool normalize) +{ + switch (axis) { + case 0: + orthogonalize_stable(R[0], R[1], R[2], normalize); + break; + case 1: + orthogonalize_stable(R[1], R[0], R[2], normalize); + break; + case 2: + orthogonalize_stable(R[2], R[0], R[1], normalize); + break; + default: + BLI_assert(0); + break; + } +} + bool is_orthogonal_m3(const float m[3][3]) { int i, j; @@ -1851,6 +1956,19 @@ void mat4_to_size(float size[3], const float mat[4][4]) size[2] = len_v3(mat[2]); } +/** Extract scale factors from the matrix, with correction to ensure + * exact volume in case of a sheared matrix. */ +void mat4_to_size_fix_shear(float size[3], const float mat[4][4]) +{ + mat4_to_size(size, mat); + + float volume = size[0] * size[1] * size[2]; + + if (volume != 0.0f) { + mul_v3_fl(size, cbrtf(fabsf(mat4_to_volume_scale(mat) / volume))); + } +} + /** * This computes the overall volume scale factor of a transformation matrix. * For an orthogonal matrix, it is the product of all three scale values. @@ -2045,6 +2163,14 @@ void rotate_m4(float mat[4][4], const char axis, const float angle) } } +/** Scale a matrix in-place. */ +void rescale_m4(float mat[4][4], float scale[3]) +{ + mul_v3_fl(mat[0], scale[0]); + mul_v3_fl(mat[1], scale[1]); + mul_v3_fl(mat[2], scale[2]); +} + /** * Scale or rotate around a pivot point, * a convenience function to avoid having to do inline. |