Fix T71344: Optix render errors with motion blur and unknown bone constraint relationship

The OptiX SRT motion expects a motion defined by translation,
rotation, shear and scale, but the matrix decomposition code in
Cycles was not able to extract shear information and instead
produced a stretch matrix with the information baked in. This
caused conflicting transforms between traversal and shading
and lead to render artifacts.
This patch changes the matrix decomposition to produce factors
inline with what OptiX expects to fix that.

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D6605

intern/cycles/device/device_optix.cpp

@@ -1461,6 +1461,9 @@ class OptiXDevice : public Device {
             srt_data[i].a = decomp[i].z.x;  // scale.x.y
             srt_data[i].b = decomp[i].z.y;  // scale.x.z
             srt_data[i].c = decomp[i].w.x;  // scale.y.z
+            assert(decomp[i].z.z == 0.0f);  // scale.y.x
+            assert(decomp[i].w.y == 0.0f);  // scale.z.x
+            assert(decomp[i].w.z == 0.0f);  // scale.z.y
 
             // Pivot point
             srt_data[i].pvx = 0.0f;

intern/cycles/util/util_math_float4.h

@@ -179,6 +179,11 @@ ccl_device_inline float4 operator+=(float4 &a, const float4 &b)
   return a = a + b;
 }
 
+ccl_device_inline float4 operator-=(float4 &a, const float4 &b)
+{
+  return a = a - b;
+}
+
 ccl_device_inline float4 operator*=(float4 &a, const float4 &b)
 {
   return a = a * b;

intern/cycles/util/util_transform.cpp

@@ -227,6 +227,7 @@ static void transform_decompose(DecomposedTransform *decomp, const Transform *tf
   M.y.w = 0.0f;
   M.z.w = 0.0f;
 
+#if 0
   Transform R = M;
   float norm;
   int iteration = 0;
@@ -260,6 +261,41 @@ static void transform_decompose(DecomposedTransform *decomp, const Transform *tf
   decomp->y.w = scale.x.x;
   decomp->z = make_float4(scale.x.y, scale.x.z, scale.y.x, scale.y.y);
   decomp->w = make_float4(scale.y.z, scale.z.x, scale.z.y, scale.z.z);
+#else
+  float3 colx = transform_get_column(&M, 0);
+  float3 coly = transform_get_column(&M, 1);
+  float3 colz = transform_get_column(&M, 2);
+
+  /* extract scale and shear first */
+  float3 scale, shear;
+  scale.x = len(colx);
+  colx /= scale.x;
+  shear.z = dot(colx, coly);
+  coly -= shear.z * colx;
+  scale.y = len(coly);
+  coly /= scale.y;
+  shear.y = dot(colx, colz);
+  colz -= shear.y * colx;
+  shear.x = dot(coly, colz);
+  colz -= shear.x * coly;
+  scale.z = len(colz);
+  colz /= scale.z;
+
+  transform_set_column(&M, 0, colx);
+  transform_set_column(&M, 1, coly);
+  transform_set_column(&M, 2, colz);
+
+  if (transform_negative_scale(M)) {
+    scale *= -1.0f;
+    M = M * transform_scale(-1.0f, -1.0f, -1.0f);
+  }
+
+  decomp->x = transform_to_quat(M);
+
+  decomp->y.w = scale.x;
+  decomp->z = make_float4(shear.z, shear.y, 0.0f, scale.y);
+  decomp->w = make_float4(shear.x, 0.0f, 0.0f, scale.z);
+#endif
 }
 
 void transform_motion_decompose(DecomposedTransform *decomp, const Transform *motion, size_t size)