X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=resample_effect.cpp;h=50b7c6bc5c4730fc6bb25cd18c96a446c8f02986;hp=10dc59c5a2dbab0a0533629b801357636ad31ef0;hb=refs%2Fheads%2Fepoxy;hpb=aa3e630eb0c455921c721f9ccd84e872bc9757bc

diff --git a/resample_effect.cpp b/resample_effect.cpp
index 10dc59c..50b7c6b 100644
--- a/resample_effect.cpp
+++ b/resample_effect.cpp
@@ -1,7 +1,7 @@
 // Three-lobed Lanczos, the most common choice.
 #define LANCZOS_RADIUS 3.0
 
-#include <GL/glew.h>
+#include <epoxy/gl.h>
 #include <assert.h>
 #include <limits.h>
 #include <math.h>
@@ -10,11 +10,14 @@
 
 #include "effect_chain.h"
 #include "effect_util.h"
+#include "fp16.h"
 #include "resample_effect.h"
 #include "util.h"
 
 using namespace std;
 
+namespace movit {
+
 namespace {
 
 float sinc(float x)
@@ -140,7 +143,7 @@ void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsig
 	input_height = height;
 	update_size();
 }
-		
+
 void ResampleEffect::update_size()
 {
 	bool ok = true;
@@ -229,7 +232,6 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
 		assert(false);
 	}
 
-
 	// For many resamplings (e.g. 640 -> 1280), we will end up with the same
 	// set of samples over and over again in a loop. Thus, we can compute only
 	// the first such loop, and then ask the card to repeat the texture for us.
@@ -299,19 +301,13 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
 		int base_src_y = lrintf(center_src_y);
 
 		// Now sample <int_radius> pixels on each side around that point.
-		double sum = 0.0;
 		for (int i = 0; i < src_samples; ++i) {
 			int src_y = base_src_y + i - int_radius;
 			float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y), LANCZOS_RADIUS);
 			weights[(y * src_samples + i) * 2 + 0] = weight * radius_scaling_factor;
 			weights[(y * src_samples + i) * 2 + 1] = (src_y + 0.5) / float(src_size);
-			sum += weights[(y * src_samples + i) * 2 + 0];
 		}
 
-		// Normalize so that the sum becomes one.
-		for (int i = 0; i < src_samples; ++i) {
-			weights[(y * src_samples + i) * 2 + 0] /= sum;
-		}
 	}
 
 	// Now make use of the bilinear filtering in the GPU to reduce the number of samples
@@ -330,14 +326,36 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
 
 	// Now that we know the right width, actually combine the samples.
 	float *bilinear_weights = new float[dst_samples * src_bilinear_samples * 2];
+	fp16_int_t *bilinear_weights_fp16 = new fp16_int_t[dst_samples * src_bilinear_samples * 2];
 	for (unsigned y = 0; y < dst_samples; ++y) {
+		float *bilinear_weights_ptr = bilinear_weights + (y * src_bilinear_samples) * 2;
+		fp16_int_t *bilinear_weights_fp16_ptr = bilinear_weights_fp16 + (y * src_bilinear_samples) * 2;
 		unsigned num_samples_saved = combine_samples(
 			weights + (y * src_samples) * 2,
-			bilinear_weights + (y * src_bilinear_samples) * 2,
+			bilinear_weights_ptr,
 			src_samples,
 			src_samples - src_bilinear_samples);
 		assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
-	}	
+
+		// Convert to fp16.
+		for (int i = 0; i < src_bilinear_samples; ++i) {
+			bilinear_weights_fp16_ptr[i * 2 + 0] = fp64_to_fp16(bilinear_weights_ptr[i * 2 + 0]);
+			bilinear_weights_fp16_ptr[i * 2 + 1] = fp64_to_fp16(bilinear_weights_ptr[i * 2 + 1]);
+		}
+
+		// 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 += fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]);
+			}
+			for (int i = 0; i < src_bilinear_samples; ++i) {
+				bilinear_weights_fp16_ptr[i * 2 + 0] = fp64_to_fp16(
+					fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]) / sum);
+			}
+		}
+	}
 
 	// Encode as a two-component texture. Note the GL_REPEAT.
 	glActiveTexture(GL_TEXTURE0 + *sampler_num);
@@ -350,11 +368,12 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
 	check_error();
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
 	check_error();
-	glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_FLOAT, bilinear_weights);
+	glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_HALF_FLOAT, bilinear_weights_fp16);
 	check_error();
 
 	delete[] weights;
 	delete[] bilinear_weights;
+	delete[] bilinear_weights_fp16;
 }
 
 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
@@ -394,8 +413,11 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const strin
 
 	// We specifically do not want mipmaps on the input texture;
 	// they break minification.
-	glActiveTexture(GL_TEXTURE0);
+	Node *self = chain->find_node_for_effect(this);
+	glActiveTexture(chain->get_input_sampler(self, 0));
 	check_error();
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 	check_error();
 }
+
+}  // namespace movit