1 // Unit tests for FFTPassEffect.
5 #include "effect_chain.h"
6 #include "gtest/gtest.h"
7 #include "image_format.h"
8 #include "fft_pass_effect.h"
9 #include "multiply_effect.h"
10 #include "test_util.h"
14 // Generate a random number uniformly distributed between [-1.0, 1.0].
15 float uniform_random()
17 return 2.0 * ((float)rand() / RAND_MAX - 0.5);
20 void setup_fft(EffectChain *chain, int fft_size, bool inverse,
21 bool add_normalizer = false,
22 FFTPassEffect::Direction direction = FFTPassEffect::HORIZONTAL)
24 assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
25 for (int i = 1, subsize = 2; subsize <= fft_size; ++i, subsize *= 2) {
26 Effect *fft_effect = chain->add_effect(new FFTPassEffect());
27 bool ok = fft_effect->set_int("fft_size", fft_size);
28 ok |= fft_effect->set_int("pass_number", i);
29 ok |= fft_effect->set_int("inverse", inverse);
30 ok |= fft_effect->set_int("direction", direction);
35 float factor[4] = { 1.0f / fft_size, 1.0f / fft_size, 1.0f / fft_size, 1.0f / fft_size };
36 Effect *multiply_effect = chain->add_effect(new MultiplyEffect());
37 bool ok = multiply_effect->set_vec4("factor", factor);
42 void run_fft(const float *in, float *out, int fft_size, bool inverse,
43 bool add_normalizer = false,
44 FFTPassEffect::Direction direction = FFTPassEffect::HORIZONTAL)
47 if (direction == FFTPassEffect::HORIZONTAL) {
54 EffectChainTester tester(in, width, height, FORMAT_RGBA_PREMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_LINEAR);
55 setup_fft(tester.get_chain(), fft_size, inverse, add_normalizer, direction);
56 tester.run(out, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR, OUTPUT_ALPHA_FORMAT_PREMULTIPLIED);
61 TEST(FFTPassEffectTest, ZeroStaysZero) {
62 const int fft_size = 64;
63 float data[fft_size * 4] = { 0 };
64 float out_data[fft_size * 4];
66 run_fft(data, out_data, fft_size, false);
67 expect_equal(data, out_data, 4, fft_size);
69 run_fft(data, out_data, fft_size, true);
70 expect_equal(data, out_data, 4, fft_size);
73 TEST(FFTPassEffectTest, Impulse) {
74 const int fft_size = 64;
75 float data[fft_size * 4] = { 0 };
76 float expected_data[fft_size * 4], out_data[fft_size * 4];
82 for (int i = 0; i < fft_size; ++i) {
83 expected_data[i * 4 + 0] = data[0];
84 expected_data[i * 4 + 1] = data[1];
85 expected_data[i * 4 + 2] = data[2];
86 expected_data[i * 4 + 3] = data[3];
89 run_fft(data, out_data, fft_size, false);
90 expect_equal(expected_data, out_data, 4, fft_size);
92 run_fft(data, out_data, fft_size, true);
93 expect_equal(expected_data, out_data, 4, fft_size);
96 TEST(FFTPassEffectTest, SingleFrequency) {
97 const int fft_size = 16;
98 float data[fft_size * 4] = { 0 };
99 float expected_data[fft_size * 4], out_data[fft_size * 4];
100 for (int i = 0; i < fft_size; ++i) {
101 data[i * 4 + 0] = sin(2.0 * M_PI * (4.0 * i) / fft_size);
102 data[i * 4 + 1] = 0.0;
103 data[i * 4 + 2] = 0.0;
104 data[i * 4 + 3] = 0.0;
106 for (int i = 0; i < fft_size; ++i) {
107 expected_data[i * 4 + 0] = 0.0;
108 expected_data[i * 4 + 1] = 0.0;
109 expected_data[i * 4 + 2] = 0.0;
110 expected_data[i * 4 + 3] = 0.0;
112 expected_data[4 * 4 + 1] = -8.0;
113 expected_data[12 * 4 + 1] = 8.0;
115 run_fft(data, out_data, fft_size, false, false, FFTPassEffect::HORIZONTAL);
116 expect_equal(expected_data, out_data, 4, fft_size);
118 run_fft(data, out_data, fft_size, false, false, FFTPassEffect::VERTICAL);
119 expect_equal(expected_data, out_data, 4, fft_size);
122 TEST(FFTPassEffectTest, Repeat) {
123 const int fft_size = 64;
124 const int num_repeats = 31; // Prime, to make things more challenging.
125 float data[num_repeats * fft_size * 4] = { 0 };
126 float expected_data[num_repeats * fft_size * 4], out_data[num_repeats * fft_size * 4];
129 for (int i = 0; i < num_repeats * fft_size * 4; ++i) {
130 data[i] = uniform_random();
133 for (int i = 0; i < num_repeats; ++i) {
134 run_fft(data + i * fft_size * 4, expected_data + i * fft_size * 4, fft_size, false);
139 EffectChainTester tester(data, num_repeats * fft_size, 1, FORMAT_RGBA_PREMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_LINEAR);
140 setup_fft(tester.get_chain(), fft_size, false);
141 tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR, OUTPUT_ALPHA_FORMAT_PREMULTIPLIED);
143 expect_equal(expected_data, out_data, 4, num_repeats * fft_size);
147 EffectChainTester tester(data, 1, num_repeats * fft_size, FORMAT_RGBA_PREMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_LINEAR);
148 setup_fft(tester.get_chain(), fft_size, false, false, FFTPassEffect::VERTICAL);
149 tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR, OUTPUT_ALPHA_FORMAT_PREMULTIPLIED);
151 expect_equal(expected_data, out_data, 4, num_repeats * fft_size);
155 TEST(FFTPassEffectTest, TwoDimensional) { // Implicitly tests vertical.
157 const int fft_size = 16;
158 float in[fft_size * fft_size * 4], out[fft_size * fft_size * 4], expected_out[fft_size * fft_size * 4];
159 for (int y = 0; y < fft_size; ++y) {
160 for (int x = 0; x < fft_size; ++x) {
161 in[(y * fft_size + x) * 4 + 0] =
162 sin(2.0 * M_PI * (2 * x + 3 * y) / fft_size);
163 in[(y * fft_size + x) * 4 + 1] = 0.0;
164 in[(y * fft_size + x) * 4 + 2] = 0.0;
165 in[(y * fft_size + x) * 4 + 3] = 0.0;
168 memset(expected_out, 0, sizeof(expected_out));
170 // This result has been verified using the fft2() function in Octave,
172 expected_out[(3 * fft_size + 2) * 4 + 1] = -128.0;
173 expected_out[(13 * fft_size + 14) * 4 + 1] = 128.0;
175 EffectChainTester tester(in, fft_size, fft_size, FORMAT_RGBA_PREMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_LINEAR);
176 setup_fft(tester.get_chain(), fft_size, false, false, FFTPassEffect::HORIZONTAL);
177 setup_fft(tester.get_chain(), fft_size, false, false, FFTPassEffect::VERTICAL);
178 tester.run(out, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR, OUTPUT_ALPHA_FORMAT_PREMULTIPLIED);
180 expect_equal(expected_out, out, 4 * fft_size, fft_size, 0.25, 0.0005);
183 // The classic paper for FFT correctness testing is Funda Ergün:
184 // “Testing Multivariate Linear Functions: Overcoming the Generator Bottleneck”
185 // (http://www.cs.sfu.ca/~funda/PUBLICATIONS/stoc95.ps), which proves that
186 // testing three basic properties of FFTs guarantees that the function is
187 // correct (at least under the assumption that errors are random).
189 // We don't follow the paper directly, though, for a few reasons: First,
190 // Ergün's paper really considers _self-correcting_ systems, which may
191 // be stochastically faulty, and thus uses various relatively complicated
192 // bounds and tests we don't really need. Second, the FFTs it considers
193 // are all about polynomials over finite fields, which means that results
194 // are exact and thus easy to test; we work with floats (half-floats!),
195 // and thus need some error tolerance.
197 // So instead, we follow the implementation of FFTW, which is really the
198 // gold standard when it comes to FFTs these days. They hard-code 20
199 // testing rounds as opposed to the more complicated bounds in the paper,
200 // and have a simpler version of the third test.
202 // The error bounds are set somewhat empirically, but remember that these
203 // inputs will give frequency values as large as ~16, where 0.025 is
204 // within the 9th bit (of 11 total mantissa bits in fp16).
205 const int ergun_rounds = 20;
207 // Test 1: Test that FFT(a + b) = FFT(a) + FFT(b).
208 TEST(FFTPassEffectTest, ErgunLinearityTest) {
210 const int max_fft_size = 64;
211 float a[max_fft_size * 4], b[max_fft_size * 4], sum[max_fft_size * 4];
212 float a_out[max_fft_size * 4], b_out[max_fft_size * 4], sum_out[max_fft_size * 4], expected_sum_out[max_fft_size * 4];
213 for (int fft_size = 2; fft_size <= max_fft_size; fft_size *= 2) {
214 for (int inverse = 0; inverse <= 1; ++inverse) {
215 for (int i = 0; i < ergun_rounds; ++i) {
216 for (int j = 0; j < fft_size * 4; ++j) {
217 a[j] = uniform_random();
218 b[j] = uniform_random();
220 run_fft(a, a_out, fft_size, inverse);
221 run_fft(b, b_out, fft_size, inverse);
223 for (int j = 0; j < fft_size * 4; ++j) {
224 sum[j] = a[j] + b[j];
225 expected_sum_out[j] = a_out[j] + b_out[j];
228 run_fft(sum, sum_out, fft_size, inverse);
229 expect_equal(expected_sum_out, sum_out, 4, fft_size, 0.03, 0.0005);
235 // Test 2: Test that FFT(delta(i)) = 1 (where delta(i) = [1 0 0 0 ...]),
236 // or more specifically, test that FFT(a + delta(i)) - FFT(a) = 1.
237 TEST(FFTPassEffectTest, ErgunImpulseTransform) {
239 const int max_fft_size = 64;
240 float a[max_fft_size * 4], b[max_fft_size * 4];
241 float a_out[max_fft_size * 4], b_out[max_fft_size * 4], sum_out[max_fft_size * 4], expected_sum_out[max_fft_size * 4];
242 for (int fft_size = 2; fft_size <= max_fft_size; fft_size *= 2) {
243 for (int inverse = 0; inverse <= 1; ++inverse) {
244 for (int i = 0; i < ergun_rounds; ++i) {
245 for (int j = 0; j < fft_size * 4; ++j) {
246 a[j] = uniform_random();
248 // Compute delta(j) - a.
255 run_fft(a, a_out, fft_size, inverse);
256 run_fft(b, b_out, fft_size, inverse);
258 for (int j = 0; j < fft_size * 4; ++j) {
259 sum_out[j] = a_out[j] + b_out[j];
260 expected_sum_out[j] = 1.0;
262 expect_equal(expected_sum_out, sum_out, 4, fft_size, 0.025, 0.0005);
268 // Test 3: Test the time-shift property of the FFT, in that a circular left-shift
269 // multiplies the result by e^(j 2pi k/N) (linear phase adjustment).
270 // As fftw_test.c says, “The paper performs more tests, but this code should be
272 TEST(FFTPassEffectTest, ErgunShiftProperty) {
274 const int max_fft_size = 64;
275 float a[max_fft_size * 4], b[max_fft_size * 4];
276 float a_out[max_fft_size * 4], b_out[max_fft_size * 4], expected_a_out[max_fft_size * 4];
277 for (int fft_size = 2; fft_size <= max_fft_size; fft_size *= 2) {
278 for (int inverse = 0; inverse <= 1; ++inverse) {
279 for (int direction = 0; direction <= 1; ++direction) {
280 for (int i = 0; i < ergun_rounds; ++i) {
281 for (int j = 0; j < fft_size * 4; ++j) {
282 a[j] = uniform_random();
285 // Circular shift left by one step.
286 for (int j = 0; j < fft_size * 4; ++j) {
287 b[j] = a[(j + 4) % (fft_size * 4)];
289 run_fft(a, a_out, fft_size, inverse, false, FFTPassEffect::Direction(direction));
290 run_fft(b, b_out, fft_size, inverse, false, FFTPassEffect::Direction(direction));
292 for (int j = 0; j < fft_size; ++j) {
293 double s = -sin(j * 2.0 * M_PI / fft_size);
294 double c = cos(j * 2.0 * M_PI / fft_size);
299 expected_a_out[j * 4 + 0] = b_out[j * 4 + 0] * c - b_out[j * 4 + 1] * s;
300 expected_a_out[j * 4 + 1] = b_out[j * 4 + 0] * s + b_out[j * 4 + 1] * c;
302 expected_a_out[j * 4 + 2] = b_out[j * 4 + 2] * c - b_out[j * 4 + 3] * s;
303 expected_a_out[j * 4 + 3] = b_out[j * 4 + 2] * s + b_out[j * 4 + 3] * c;
305 expect_equal(expected_a_out, a_out, 4, fft_size, 0.025, 0.0005);
312 TEST(FFTPassEffectTest, BigFFTAccuracy) {
314 const int max_fft_size = 2048;
315 float in[max_fft_size * 4], out[max_fft_size * 4], out2[max_fft_size * 4];
316 for (int fft_size = 2; fft_size <= max_fft_size; fft_size *= 2) {
317 for (int j = 0; j < fft_size * 4; ++j) {
318 in[j] = uniform_random();
320 run_fft(in, out, fft_size, false, true); // Forward, with normalization.
321 run_fft(out, out2, fft_size, true); // Reverse.
323 // These error bounds come from
324 // http://en.wikipedia.org/wiki/Fast_Fourier_transform#Accuracy_and_approximations,
325 // with empirically estimated epsilons. Note that the calculated
326 // rms in expect_equal() is divided by sqrt(N), so we compensate
328 double max_error = 0.0009 * log2(fft_size);
329 double rms_limit = 0.0007 * sqrt(log2(fft_size)) / sqrt(fft_size);
330 expect_equal(in, out2, 4, fft_size, max_error, rms_limit);