X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=resample_effect.cpp;h=244a3e2a6081187f19962b531c7fa86dde46203b;hb=6c954b4f0bff0743e13ce6ddcee8bda15b3af234;hp=4dd3ea87e261086939a91e2d26d51ef52c813168;hpb=9c12e38b7cd88a77ef297d080b7c41e6bd6326fb;p=movit

diff --git a/resample_effect.cpp b/resample_effect.cpp
index 4dd3ea8..244a3e2 100644
--- a/resample_effect.cpp
+++ b/resample_effect.cpp
@@ -61,7 +61,7 @@ unsigned gcd(unsigned a, unsigned b)
 }
 
 template<class DestFloat>
-unsigned combine_samples(const Tap<float> *src, Tap<DestFloat> *dst, unsigned src_size, unsigned num_src_samples, unsigned max_samples_saved)
+unsigned combine_samples(const Tap<float> *src, Tap<DestFloat> *dst, float num_subtexels, float inv_num_subtexels, unsigned num_src_samples, unsigned max_samples_saved)
 {
 	// Cut off near-zero values at both sides.
 	unsigned num_samples_saved = 0;
@@ -107,9 +107,9 @@ unsigned combine_samples(const Tap<float> *src, Tap<DestFloat> *dst, unsigned sr
 		float pos2 = src[i + 1].pos;
 		assert(pos2 > pos1);
 
-		fp16_int_t pos, total_weight;
+		DestFloat pos, total_weight;
 		float sum_sq_error;
-		combine_two_samples(w1, w2, pos1, pos2, src_size, &pos, &total_weight, &sum_sq_error);
+		combine_two_samples(w1, w2, pos1, pos2, num_subtexels, inv_num_subtexels, &pos, &total_weight, &sum_sq_error);
 
 		// If the interpolation error is larger than that of about sqrt(2) of
 		// a level at 8-bit precision, don't combine. (You'd think 1.0 was enough,
@@ -142,8 +142,9 @@ void normalize_sum(Tap<T>* vals, unsigned num)
 		for (unsigned i = 0; i < num; ++i) {
 			sum += to_fp64(vals[i].weight);
 		}
+		double inv_sum = 1.0 / sum;
 		for (unsigned i = 0; i < num; ++i) {
-			vals[i].weight = from_fp64<T>(to_fp64(vals[i].weight) / sum);
+			vals[i].weight = from_fp64<T>(to_fp64(vals[i].weight) * inv_sum);
 		}
 	}
 }
@@ -159,23 +160,28 @@ void normalize_sum(Tap<T>* vals, unsigned num)
 template<class DestFloat>
 unsigned combine_many_samples(const Tap<float> *weights, unsigned src_size, unsigned src_samples, unsigned dst_samples, Tap<DestFloat> **bilinear_weights)
 {
-	int src_bilinear_samples = 0;
-	for (unsigned y = 0; y < dst_samples; ++y) {
-		unsigned num_samples_saved = combine_samples<DestFloat>(weights + y * src_samples, NULL, src_size, src_samples, UINT_MAX);
-		src_bilinear_samples = max<int>(src_bilinear_samples, src_samples - num_samples_saved);
+	float num_subtexels = src_size / movit_texel_subpixel_precision;
+	float inv_num_subtexels = movit_texel_subpixel_precision / src_size;
+
+	unsigned max_samples_saved = UINT_MAX;
+	for (unsigned y = 0; y < dst_samples && max_samples_saved > 0; ++y) {
+		unsigned num_samples_saved = combine_samples<DestFloat>(weights + y * src_samples, NULL, num_subtexels, inv_num_subtexels, src_samples, max_samples_saved);
+		max_samples_saved = min(max_samples_saved, num_samples_saved);
 	}
 
 	// Now that we know the right width, actually combine the samples.
+	unsigned src_bilinear_samples = src_samples - max_samples_saved;
 	*bilinear_weights = new Tap<DestFloat>[dst_samples * src_bilinear_samples];
 	for (unsigned y = 0; y < dst_samples; ++y) {
 		Tap<DestFloat> *bilinear_weights_ptr = *bilinear_weights + y * src_bilinear_samples;
 		unsigned num_samples_saved = combine_samples(
 			weights + y * src_samples,
 			bilinear_weights_ptr,
-			src_size,
+			num_subtexels,
+			inv_num_subtexels,
 			src_samples,
-			src_samples - src_bilinear_samples);
-		assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
+			max_samples_saved);
+		assert(num_samples_saved == max_samples_saved);
 		normalize_sum(bilinear_weights_ptr, src_bilinear_samples);
 	}
 	return src_bilinear_samples;