4 #include "effect_chain.h"
5 #include "effect_util.h"
7 #include "fft_pass_effect.h"
14 FFTPassEffect::FFTPassEffect()
17 direction(HORIZONTAL),
19 last_direction(INVALID),
24 register_int("fft_size", &fft_size);
25 register_int("direction", (int *)&direction);
26 register_int("pass_number", &pass_number);
27 register_int("inverse", &inverse);
28 register_uniform_float("num_repeats", &uniform_num_repeats);
29 register_uniform_sampler2d("support_tex", &uniform_support_tex);
30 glGenTextures(1, &tex);
33 FFTPassEffect::~FFTPassEffect()
35 glDeleteTextures(1, &tex);
38 string FFTPassEffect::output_fragment_shader()
41 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
42 return buf + read_file("fft_pass_effect.frag");
45 void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
47 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
49 // This is needed because it counteracts the precision issues we get
50 // because we sample the input texture with normalized coordinates
51 // (especially when the repeat count along the axis is not a power of
52 // two); we very rapidly end up in narrowly missing a texel center,
53 // which causes precision loss to propagate throughout the FFT.
54 Node *self = chain->find_node_for_effect(this);
55 glActiveTexture(chain->get_input_sampler(self, 0));
57 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
59 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
62 // Because of the memory layout (see below) and because we use offsets,
63 // the support texture values for many consecutive values will be
64 // the same. Thus, we can store a smaller texture (giving a small
65 // performance boost) and just sample it with NEAREST. Also, this
66 // counteracts any precision issues we might get from linear
68 glActiveTexture(GL_TEXTURE0 + *sampler_num);
70 glBindTexture(GL_TEXTURE_2D, tex);
72 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
74 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
76 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
79 int input_size = (direction == VERTICAL) ? input_height : input_width;
80 if (last_fft_size != fft_size ||
81 last_direction != direction ||
82 last_pass_number != pass_number ||
83 last_inverse != inverse ||
84 last_input_size != input_size) {
85 generate_support_texture();
88 uniform_support_tex = *sampler_num;
91 assert(input_size % fft_size == 0);
92 uniform_num_repeats = input_size / fft_size;
95 void FFTPassEffect::generate_support_texture()
97 int input_size = (direction == VERTICAL) ? input_height : input_width;
99 // The memory layout follows figure 5.2 on page 25 of
100 // http://gpuwave.sesse.net/gpuwave.pdf -- it can be a bit confusing
101 // at first, but is classically explained more or less as follows:
103 // The classic Cooley-Tukey decimation-in-time FFT algorithm works
104 // by first splitting input data into odd and even elements
105 // (e.g. bit-wise xxxxx0 and xxxxx1 for a size-32 FFT), then FFTing
106 // them separately and combining them using twiddle factors.
107 // So the outer pass (done _last_) looks only at the last bit,
108 // and does one such merge pass of sub-size N/2 (FFT size N).
110 // FFT of the first part must then necessarily be split into xxxx00 and
111 // xxxx10, and similarly xxxx01 and xxxx11 for the other part. Since
112 // these two FFTs are handled identically, it means we split into xxxx0x
113 // and xxxx1x, so that the second-outer pass (done second-to-last)
114 // looks only at the second last bit, and so on. We do two such merge
115 // passes of sub-size N/4 (sub-FFT size N/2).
117 // Thus, the inner, Nth pass (done first) splits at the first bit,
118 // so 0 is paired with 16, 1 with 17 and so on, doing N/2 such merge
119 // passes of sub-size 1 (sub-FFT size 2). We say that the stride is 16.
120 // The second-inner, (N-1)th pass (done second) splits at the second
121 // bit, so the stride is 8, and so on.
123 assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
124 int subfft_size = 1 << pass_number;
125 fp16_int_t *tmp = new fp16_int_t[subfft_size * 4];
130 mulfac = -2.0 * M_PI;
133 assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
134 assert(fft_size % subfft_size == 0);
135 int stride = fft_size / subfft_size;
136 for (int i = 0; i < subfft_size; i++) {
138 double twiddle_real, twiddle_imag;
140 if (k < subfft_size / 2) {
141 twiddle_real = cos(mulfac * (k / double(subfft_size)));
142 twiddle_imag = sin(mulfac * (k / double(subfft_size)));
144 // This is mathematically equivalent to the twiddle factor calculations
145 // in the other branch of the if, but not numerically; the range
146 // reductions on x87 are not all that precise, and this keeps us within
148 k -= subfft_size / 2;
149 twiddle_real = -cos(mulfac * (k / double(subfft_size)));
150 twiddle_imag = -sin(mulfac * (k / double(subfft_size)));
153 // The support texture contains everything we need for the FFT:
154 // Obviously, the twiddle factor (in the Z and W components), but also
155 // which two samples to fetch. These are stored as normalized
156 // X coordinate offsets (Y coordinate for a vertical FFT); the reason
157 // for using offsets and not direct coordinates as in GPUwave
158 // is that we can have multiple FFTs along the same line,
159 // and want to reuse the support texture by repeating it.
160 int base = k * stride * 2;
161 int support_texture_index = i;
163 int src2 = base + stride;
165 if (direction == FFTPassEffect::VERTICAL) {
166 // Compensate for OpenGL's bottom-left convention.
167 support_texture_index = subfft_size - support_texture_index - 1;
170 tmp[support_texture_index * 4 + 0] = fp32_to_fp16(sign * (src1 - i * stride) / double(input_size));
171 tmp[support_texture_index * 4 + 1] = fp32_to_fp16(sign * (src2 - i * stride) / double(input_size));
172 tmp[support_texture_index * 4 + 2] = fp32_to_fp16(twiddle_real);
173 tmp[support_texture_index * 4 + 3] = fp32_to_fp16(twiddle_imag);
176 // Supposedly FFTs are very sensitive to inaccuracies in the twiddle factors,
177 // at least according to a paper by Schatzman (see gpuwave.pdf reference [30]
178 // for the full reference); however, practical testing indicates that it's
179 // not a problem to keep the twiddle factors at 16-bit, at least as long as
180 // we round them properly--it would seem that Schatzman were mainly talking
181 // about poor sin()/cos() approximations. Thus, we store them in 16-bit,
182 // which gives a nice speed boost.
184 // Note that the source coordinates become somewhat less accurate too, though.
185 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F, subfft_size, 1, 0, GL_RGBA, GL_HALF_FLOAT, tmp);
190 last_fft_size = fft_size;
191 last_direction = direction;
192 last_pass_number = pass_number;
193 last_inverse = inverse;
194 last_input_size = input_size;