Remove some unneeded conversions from ResampleEffect. Speeds up texture generation...

[movit] / fft_pass_effect.cpp

diff --git a/fft_pass_effect.cpp b/fft_pass_effect.cpp
index b46e5e589fc9bb4f548dafc234745c801506beb4..b0006087cd2779282af2ec3750d5ba7761d084f5 100644 (file)
--- a/fft_pass_effect.cpp
+++ b/fft_pass_effect.cpp
@@ -1,4 +1,4 @@
-#include <GL/glew.h>
+#include <epoxy/gl.h>
 #include <math.h>
 
 #include "effect_chain.h"
@@ -25,6 +25,8 @@ FFTPassEffect::FFTPassEffect()
        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);
 }
 
@@ -83,11 +85,11 @@ void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix,
                generate_support_texture();
        }
 
-       set_uniform_int(glsl_program_num, prefix, "support_tex", *sampler_num);
+       uniform_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);
+       uniform_num_repeats = input_size / fft_size;
 }
 
 void FFTPassEffect::generate_support_texture()
@@ -165,10 +167,10 @@ void FFTPassEffect::generate_support_texture()
                        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);
        }
 
        // Supposedly FFTs are very sensitive to inaccuracies in the twiddle factors,