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 glGenTextures(1, &tex);
31 FFTPassEffect::~FFTPassEffect()
33 glDeleteTextures(1, &tex);
36 string FFTPassEffect::output_fragment_shader()
39 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
40 return buf + read_file("fft_pass_effect.frag");
43 void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
45 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
47 // This is needed because it counteracts the precision issues we get
48 // because we sample the input texture with normalized coordinates
49 // (especially when the repeat count along the axis is not a power of
50 // two); we very rapidly end up in narrowly missing a texel center,
51 // which causes precision loss to propagate throughout the FFT.
52 Node *self = chain->find_node_for_effect(this);
53 glActiveTexture(chain->get_input_sampler(self, 0));
55 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
57 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
60 // Because of the memory layout (see below) and because we use offsets,
61 // the support texture values for many consecutive values will be
62 // the same. Thus, we can store a smaller texture (giving a small
63 // performance boost) and just sample it with NEAREST. Also, this
64 // counteracts any precision issues we might get from linear
66 glActiveTexture(GL_TEXTURE0 + *sampler_num);
68 glBindTexture(GL_TEXTURE_2D, tex);
70 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
72 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
74 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
77 int input_size = (direction == VERTICAL) ? input_height : input_width;
78 if (last_fft_size != fft_size ||
79 last_direction != direction ||
80 last_pass_number != pass_number ||
81 last_inverse != inverse ||
82 last_input_size != input_size) {
83 generate_support_texture();
86 set_uniform_int(glsl_program_num, prefix, "support_tex", *sampler_num);
89 assert(input_size % fft_size == 0);
90 set_uniform_float(glsl_program_num, prefix, "num_repeats", input_size / fft_size);
93 void FFTPassEffect::generate_support_texture()
95 int input_size = (direction == VERTICAL) ? input_height : input_width;
97 // The memory layout follows figure 5.2 on page 25 of
98 // http://gpuwave.sesse.net/gpuwave.pdf -- it can be a bit confusing
99 // at first, but is classically explained more or less as follows:
101 // The classic Cooley-Tukey decimation-in-time FFT algorithm works
102 // by first splitting input data into odd and even elements
103 // (e.g. bit-wise xxxxx0 and xxxxx1 for a size-32 FFT), then FFTing
104 // them separately and combining them using twiddle factors.
105 // So the outer pass (done _last_) looks only at the last bit,
106 // and does one such merge pass of sub-size N/2 (FFT size N).
108 // FFT of the first part must then necessarily be split into xxxx00 and
109 // xxxx10, and similarly xxxx01 and xxxx11 for the other part. Since
110 // these two FFTs are handled identically, it means we split into xxxx0x
111 // and xxxx1x, so that the second-outer pass (done second-to-last)
112 // looks only at the second last bit, and so on. We do two such merge
113 // passes of sub-size N/4 (sub-FFT size N/2).
115 // Thus, the inner, Nth pass (done first) splits at the first bit,
116 // so 0 is paired with 16, 1 with 17 and so on, doing N/2 such merge
117 // passes of sub-size 1 (sub-FFT size 2). We say that the stride is 16.
118 // The second-inner, (N-1)th pass (done second) splits at the second
119 // bit, so the stride is 8, and so on.
121 assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
122 int subfft_size = 1 << pass_number;
123 fp16_int_t *tmp = new fp16_int_t[subfft_size * 4];
128 mulfac = -2.0 * M_PI;
131 assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
132 assert(fft_size % subfft_size == 0);
133 int stride = fft_size / subfft_size;
134 for (int i = 0; i < subfft_size; i++) {
136 double twiddle_real, twiddle_imag;
138 if (k < subfft_size / 2) {
139 twiddle_real = cos(mulfac * (k / double(subfft_size)));
140 twiddle_imag = sin(mulfac * (k / double(subfft_size)));
142 // This is mathematically equivalent to the twiddle factor calculations
143 // in the other branch of the if, but not numerically; the range
144 // reductions on x87 are not all that precise, and this keeps us within
146 k -= subfft_size / 2;
147 twiddle_real = -cos(mulfac * (k / double(subfft_size)));
148 twiddle_imag = -sin(mulfac * (k / double(subfft_size)));
151 // The support texture contains everything we need for the FFT:
152 // Obviously, the twiddle factor (in the Z and W components), but also
153 // which two samples to fetch. These are stored as normalized
154 // X coordinate offsets (Y coordinate for a vertical FFT); the reason
155 // for using offsets and not direct coordinates as in GPUwave
156 // is that we can have multiple FFTs along the same line,
157 // and want to reuse the support texture by repeating it.
158 int base = k * stride * 2;
159 int support_texture_index = i;
161 int src2 = base + stride;
163 if (direction == FFTPassEffect::VERTICAL) {
164 // Compensate for OpenGL's bottom-left convention.
165 support_texture_index = subfft_size - support_texture_index - 1;
168 tmp[support_texture_index * 4 + 0] = fp64_to_fp16(sign * (src1 - i * stride) / double(input_size));
169 tmp[support_texture_index * 4 + 1] = fp64_to_fp16(sign * (src2 - i * stride) / double(input_size));
170 tmp[support_texture_index * 4 + 2] = fp64_to_fp16(twiddle_real);
171 tmp[support_texture_index * 4 + 3] = fp64_to_fp16(twiddle_imag);
174 // Supposedly FFTs are very sensitive to inaccuracies in the twiddle factors,
175 // at least according to a paper by Schatzman (see gpuwave.pdf reference [30]
176 // for the full reference); however, practical testing indicates that it's
177 // not a problem to keep the twiddle factors at 16-bit, at least as long as
178 // we round them properly--it would seem that Schatzman were mainly talking
179 // about poor sin()/cos() approximations. Thus, we store them in 16-bit,
180 // which gives a nice speed boost.
182 // Note that the source coordinates become somewhat less accurate too, though.
183 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F, subfft_size, 1, 0, GL_RGBA, GL_HALF_FLOAT, tmp);
188 last_fft_size = fft_size;
189 last_direction = direction;
190 last_pass_number = pass_number;
191 last_inverse = inverse;
192 last_input_size = input_size;