]> git.sesse.net Git - movit/blobdiff - fft_pass_effect.cpp
Remove some unneeded conversions from ResampleEffect. Speeds up texture generation...
[movit] / fft_pass_effect.cpp
index deb0a1cd94db7373cddbe40abe326ed891fc29f3..b0006087cd2779282af2ec3750d5ba7761d084f5 100644 (file)
@@ -1,4 +1,4 @@
-#include <GL/glew.h>
+#include <epoxy/gl.h>
 #include <math.h>
 
 #include "effect_chain.h"
@@ -14,12 +14,19 @@ namespace movit {
 FFTPassEffect::FFTPassEffect()
        : input_width(1280),
          input_height(720),
-         direction(HORIZONTAL)
+         direction(HORIZONTAL),
+         last_fft_size(-1),
+         last_direction(INVALID),
+         last_pass_number(-1),
+         last_inverse(-1),
+         last_input_size(-1)
 {
        register_int("fft_size", &fft_size);
        register_int("direction", (int *)&direction);
        register_int("pass_number", &pass_number);
        register_int("inverse", &inverse);
+       register_uniform_float("num_repeats", &uniform_num_repeats);
+       register_uniform_sampler2d("support_tex", &uniform_support_tex);
        glGenTextures(1, &tex);
 }
 
@@ -39,7 +46,18 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
 {
        Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
 
-       int input_size = (direction == VERTICAL) ? input_height : input_width;
+       // This is needed because it counteracts the precision issues we get
+       // because we sample the input texture with normalized coordinates
+       // (especially when the repeat count along the axis is not a power of
+       // two); we very rapidly end up in narrowly missing a texel center,
+       // which causes precision loss to propagate throughout the FFT.
+       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_NEAREST);
+       check_error();
+       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+       check_error();
 
        // Because of the memory layout (see below) and because we use offsets,
        // the support texture values for many consecutive values will be
@@ -47,13 +65,36 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
        // performance boost) and just sample it with NEAREST. Also, this
        // counteracts any precision issues we might get from linear
        // interpolation.
-       Node *self = chain->find_node_for_effect(this);
-       glActiveTexture(chain->get_input_sampler(self, 0));
+       glActiveTexture(GL_TEXTURE0 + *sampler_num);
+       check_error();
+       glBindTexture(GL_TEXTURE_2D, tex);
        check_error();
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
        check_error();
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
        check_error();
+       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
+       check_error();
+
+       int input_size = (direction == VERTICAL) ? input_height : input_width;
+       if (last_fft_size != fft_size ||
+           last_direction != direction ||
+           last_pass_number != pass_number ||
+           last_inverse != inverse ||
+           last_input_size != input_size) {
+               generate_support_texture();
+       }
+
+       uniform_support_tex = *sampler_num;
+       ++*sampler_num;
+
+       assert(input_size % fft_size == 0);
+       uniform_num_repeats = input_size / fft_size;
+}
+
+void FFTPassEffect::generate_support_texture()
+{
+       int input_size = (direction == VERTICAL) ? input_height : input_width;
 
        // The memory layout follows figure 5.2 on page 25 of
        // http://gpuwave.sesse.net/gpuwave.pdf -- it can be a bit confusing
@@ -80,8 +121,8 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
        // bit, so the stride is 8, and so on.
 
        assert((fft_size & (fft_size - 1)) == 0);  // Must be power of two.
-       fp16_int_t *tmp = new fp16_int_t[fft_size * 4];
        int subfft_size = 1 << pass_number;
+       fp16_int_t *tmp = new fp16_int_t[subfft_size * 4];
        double mulfac;
        if (inverse) {
                mulfac = 2.0 * M_PI;
@@ -126,23 +167,12 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
                        support_texture_index = subfft_size - support_texture_index - 1;
                        sign = -1.0;
                }
-               tmp[support_texture_index * 4 + 0] = fp64_to_fp16(sign * (src1 - i * stride) / double(input_size));
-               tmp[support_texture_index * 4 + 1] = fp64_to_fp16(sign * (src2 - i * stride) / double(input_size));
-               tmp[support_texture_index * 4 + 2] = fp64_to_fp16(twiddle_real);
-               tmp[support_texture_index * 4 + 3] = fp64_to_fp16(twiddle_imag);
+               tmp[support_texture_index * 4 + 0] = fp32_to_fp16(sign * (src1 - i * stride) / double(input_size));
+               tmp[support_texture_index * 4 + 1] = fp32_to_fp16(sign * (src2 - i * stride) / double(input_size));
+               tmp[support_texture_index * 4 + 2] = fp32_to_fp16(twiddle_real);
+               tmp[support_texture_index * 4 + 3] = fp32_to_fp16(twiddle_imag);
        }
 
-       glActiveTexture(GL_TEXTURE0 + *sampler_num);
-       check_error();
-       glBindTexture(GL_TEXTURE_2D, tex);
-       check_error();
-       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-       check_error();
-       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-       check_error();
-       glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
-       check_error();
-
        // Supposedly FFTs are very sensitive to inaccuracies in the twiddle factors,
        // at least according to a paper by Schatzman (see gpuwave.pdf reference [30]
        // for the full reference); however, practical testing indicates that it's
@@ -157,11 +187,11 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
 
        delete[] tmp;
 
-       set_uniform_int(glsl_program_num, prefix, "support_tex", *sampler_num);
-       ++*sampler_num;
-
-       assert(input_size % fft_size == 0);
-       set_uniform_float(glsl_program_num, prefix, "num_repeats", input_size / fft_size);
+       last_fft_size = fft_size;
+       last_direction = direction;
+       last_pass_number = pass_number;
+       last_inverse = inverse;
+       last_input_size = input_size;
 }
 
 }  // namespace movit