+// the GPU's bilinear hardware to sample w1 * x[pos1] + w2 * x[pos2],
+// where pos1 and pos2 must be normalized coordinates describing neighboring
+// texels in the mipmap level at which you sample. <num_subtexels> is the
+// number of distinct accessible subtexels in the given mipmap level,
+// calculated by num_texels / movit_texel_subpixel_precision. It is a float
+// for performance reasons, even though it is expected to be a whole number.
+// <inv_num_subtexels> is simply its inverse (1/x). <pos1_pos2_diff> is
+// (pos2-pos1) and <inv_pos1_pos2_diff> is 1/(pos2-pos1).
+//
+// Note that since the GPU might have limited precision in its linear
+// interpolation, the effective weights might be different from the ones you
+// asked for. sum_sq_error, if not nullptr, will contain the sum of the
+// (estimated) squared errors of the two weights.
+//
+// The answer, in "offset", comes as a normalized coordinate,
+// so if e.g. w2 = 0, you have simply offset = pos1. If <rounding_behavior>
+// is COMBINE_ROUND_TO_FP16, the coordinate is assumed to be stored as a
+// rounded fp16 value. This enables more precise calculation of total_weight
+// and sum_sq_error.
+template<class DestFloat>
+void combine_two_samples(float w1, float w2, float pos1, float pos1_pos2_diff, float inv_pos1_pos2_diff, float num_subtexels, float inv_num_subtexels,
+ DestFloat *offset, DestFloat *total_weight, float *sum_sq_error)