3 #include "fft_pass_effect.h"
4 #include "effect_util.h"
7 FFTPassEffect::FFTPassEffect()
12 register_int("fft_size", &fft_size);
13 register_int("direction", (int *)&direction);
14 register_int("pass_number", &pass_number);
15 register_int("inverse", &inverse);
16 glGenTextures(1, &tex);
19 FFTPassEffect::~FFTPassEffect()
21 glDeleteTextures(1, &tex);
24 std::string FFTPassEffect::output_fragment_shader()
27 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
28 return buf + read_file("fft_pass_effect.frag");
31 void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
33 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
35 int input_size = (direction == VERTICAL) ? input_height : input_width;
37 // See the comments on changes_output_size() in the .h file to see
38 // why this is legal. It is _needed_ because it counteracts the
39 // precision issues we get because we sample the input texture with
40 // normalized coordinates (especially when the repeat count along
41 // the axis is not a power of two); we very rapidly end up in narrowly
42 // missing a texel center, which causes precision loss to propagate
43 // throughout the FFT.
44 assert(*sampler_num == 1);
45 glActiveTexture(GL_TEXTURE0);
47 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
49 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
52 // The memory layout follows figure 5.2 on page 25 of
53 // http://gpuwave.sesse.net/gpuwave.pdf -- it can be a bit confusing
54 // at first, but is classically explained more or less as follows:
56 // The classic Cooley-Tukey decimation-in-time FFT algorithm works
57 // by first splitting input data into odd and even elements
58 // (e.g. bit-wise xxxxx0 and xxxxx1 for a size-32 FFT), then FFTing
59 // them separately and combining them using twiddle factors.
60 // So the outer pass (done _last_) looks only at the last bit,
61 // and does one such merge pass of sub-size N/2 (FFT size N).
63 // FFT of the first part must then necessarily be split into xxxx00 and
64 // xxxx10, and similarly xxxx01 and xxxx11 for the other part. Since
65 // these two FFTs are handled identically, it means we split into xxxx0x
66 // and xxxx1x, so that the second-outer pass (done second-to-last)
67 // looks only at the second last bit, and so on. We do two such merge
68 // passes of sub-size N/4 (sub-FFT size N/2).
70 // Thus, the inner, Nth pass (done first) splits at the first bit,
71 // so 0 is paired with 16, 1 with 17 and so on, doing N/2 such merge
72 // passes of sub-size 1 (sub-FFT size 2). We say that the stride is 16.
73 // The second-inner, (N-1)th pass (done second) splits at the second
74 // bit, so the stride is 8, and so on.
76 assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
77 float *tmp = new float[fft_size * 4];
78 int subfft_size = 1 << pass_number;
86 assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
87 assert(fft_size % subfft_size == 0);
88 int stride = fft_size / subfft_size;
89 for (int i = 0; i < fft_size; ++i) {
90 int k = i / stride; // Element number within this sub-FFT.
91 int offset = i % stride; // Sub-FFT number.
92 double twiddle_real, twiddle_imag;
94 if (k < subfft_size / 2) {
95 twiddle_real = cos(mulfac * (k / double(subfft_size)));
96 twiddle_imag = sin(mulfac * (k / double(subfft_size)));
98 // This is mathematically equivalent to the twiddle factor calculations
99 // in the other branch of the if, but not numerically; the range
100 // reductions on x87 are not all that precise, and this keeps us within
102 k -= subfft_size / 2;
103 twiddle_real = -cos(mulfac * (k / double(subfft_size)));
104 twiddle_imag = -sin(mulfac * (k / double(subfft_size)));
107 // The support texture contains everything we need for the FFT:
108 // Obviously, the twiddle factor (in the Z and W components), but also
109 // which two samples to fetch. These are stored as normalized
110 // X coordinate offsets (Y coordinate for a vertical FFT); the reason
111 // for using offsets and not direct coordinates as in GPUwave
112 // is that we can have multiple FFTs along the same line,
113 // and want to reuse the support texture by repeating it.
114 int base = k * stride * 2 + offset;
115 int support_texture_index;
116 if (direction == FFTPassEffect::VERTICAL) {
117 // Compensate for OpenGL's bottom-left convention.
118 support_texture_index = fft_size - i - 1;
120 support_texture_index = i;
122 tmp[support_texture_index * 4 + 0] = (base - support_texture_index) / double(input_size);
123 tmp[support_texture_index * 4 + 1] = (base + stride - support_texture_index) / double(input_size);
124 tmp[support_texture_index * 4 + 2] = twiddle_real;
125 tmp[support_texture_index * 4 + 3] = twiddle_imag;
128 glActiveTexture(GL_TEXTURE0 + *sampler_num);
130 glBindTexture(GL_TEXTURE_1D, tex);
132 glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
134 glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
136 glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
139 // Supposedly FFTs are very sensitive to inaccuracies in the twiddle factors,
140 // at least according to a paper by Schatzman (see gpuwave.pdf reference [30]
141 // for the full reference), so we keep them at 32-bit. However, for
142 // small sizes, all components are exact anyway, so we can cheat there
143 // (although noting that the source coordinates become somewhat less
144 // accurate then, too).
145 glTexImage1D(GL_TEXTURE_1D, 0, (subfft_size <= 4) ? GL_RGBA16F : GL_RGBA32F, fft_size, 0, GL_RGBA, GL_FLOAT, tmp);
150 set_uniform_int(glsl_program_num, prefix, "support_tex", *sampler_num);
153 assert(input_size % fft_size == 0);
154 set_uniform_float(glsl_program_num, prefix, "num_repeats", input_size / fft_size);