X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=resample_effect.cpp;h=aa3a4033ee8e5dec5b2f9826de529dfb98c22567;hp=f4808c4560437288a327156f10a2f0e794e1f991;hb=641053e9fc86b2166e361a983075febc3bb69acd;hpb=c62391987241f1482a99b6f6417fbec1d0ef2344

diff --git a/resample_effect.cpp b/resample_effect.cpp
index f4808c4..aa3a403 100644
--- a/resample_effect.cpp
+++ b/resample_effect.cpp
@@ -7,6 +7,9 @@
 #include <math.h>
 #include <stdio.h>
 #include <algorithm>
+#include <Eigen/Sparse>
+#include <Eigen/SparseQR>
+#include <Eigen/OrderingMethods>
 
 #include "effect_chain.h"
 #include "effect_util.h"
@@ -15,6 +18,7 @@
 #include "resample_effect.h"
 #include "util.h"
 
+using namespace Eigen;
 using namespace std;
 
 namespace movit {
@@ -113,6 +117,22 @@ unsigned combine_samples(const Tap<float> *src, Tap<DestFloat> *dst, unsigned sr
 	return num_samples_saved;
 }
 
+// Normalize so that the sum becomes one. Note that we do it twice;
+// this sometimes helps a tiny little bit when we have many samples.
+template<class T>
+void normalize_sum(Tap<T>* vals, unsigned num)
+{
+	for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {
+		double sum = 0.0;
+		for (unsigned i = 0; i < num; ++i) {
+			sum += to_fp64(vals[i].weight);
+		}
+		for (unsigned i = 0; i < num; ++i) {
+			vals[i].weight = from_fp64<T>(to_fp64(vals[i].weight) / sum);
+		}
+	}
+}
+
 // Make use of the bilinear filtering in the GPU to reduce the number of samples
 // we need to make. This is a bit more complex than BlurEffect since we cannot combine
 // two neighboring samples if their weights have differing signs, so we first need to
@@ -141,19 +161,7 @@ unsigned combine_many_samples(const Tap<float> *weights, unsigned src_size, unsi
 			src_samples,
 			src_samples - src_bilinear_samples);
 		assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
-
-		// Normalize so that the sum becomes one. Note that we do it twice;
-		// this sometimes helps a tiny little bit when we have many samples.
-		for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {
-			double sum = 0.0;
-			for (int i = 0; i < src_bilinear_samples; ++i) {
-				sum += to_fp64(bilinear_weights_ptr[i].weight);
-			}
-			for (int i = 0; i < src_bilinear_samples; ++i) {
-				bilinear_weights_ptr[i].weight = from_fp64<DestFloat>(
-					to_fp64(bilinear_weights_ptr[i].weight) / sum);
-			}
-		}
+		normalize_sum(bilinear_weights_ptr, src_bilinear_samples);
 	}
 	return src_bilinear_samples;
 }
@@ -502,6 +510,9 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
 
 	// Now make use of the bilinear filtering in the GPU to reduce the number of samples
 	// we need to make. Try fp16 first; if it's not accurate enough, we go to fp32.
+	// Our tolerance level for total error is a bit higher than the one for invididual
+	// samples, since one would assume overall errors in the shape don't matter as much.
+	const float max_error = 2.0f / (255.0f * 255.0f);
 	Tap<fp16_int_t> *bilinear_weights_fp16;
 	src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp16);
 	Tap<float> *bilinear_weights_fp32 = NULL;
@@ -513,11 +524,12 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
 			bilinear_weights_fp16 + y * src_bilinear_samples, src_bilinear_samples,
 			src_size);
 		max_sum_sq_error_fp16 = std::max(max_sum_sq_error_fp16, sum_sq_error_fp16);
+		if (max_sum_sq_error_fp16 > max_error) {
+			break;
+		}
 	}
 
-	// Our tolerance level for total error is a bit higher than the one for invididual
-	// samples, since one would assume overall errors in the shape don't matter as much.
-	if (max_sum_sq_error_fp16 > 2.0f / (255.0f * 255.0f)) {
+	if (max_sum_sq_error_fp16 > max_error) {
 		fallback_to_fp32 = true;
 		src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp32);
 	}