diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-09-13 20:03:39 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-09-13 20:13:28 +0200 |
| commit | 64c3a1b528a88b4ae21bf3cbcff2537200477806 (patch) | |
| tree | 85da00748d7bbb5036aeb5a7a4b60b5abf75aeea /tests | |
| download | pyseg-64c3a1b528a88b4ae21bf3cbcff2537200477806.tar.gz pyseg-64c3a1b528a88b4ae21bf3cbcff2537200477806.zip | |
Init
Diffstat (limited to 'tests')
| -rw-r--r-- | tests/test_kernels.c | 151 | ||||
| -rw-r--r-- | tests/test_parity.c | 214 | ||||
| -rw-r--r-- | tests/test_robust.c | 133 | ||||
| -rw-r--r-- | tests/test_slide.c | 212 | ||||
| -rw-r--r-- | tests/test_util.h | 29 |
5 files changed, 739 insertions, 0 deletions
diff --git a/tests/test_kernels.c b/tests/test_kernels.c new file mode 100644 index 0000000..d4f25d6 --- /dev/null +++ b/tests/test_kernels.c @@ -0,0 +1,151 @@ +/* The dense primitive against a naive triple loop, over the shapes the model + * actually uses plus enough odd ones to exercise every edge path. + * + * The reference here is deliberately the most obvious implementation possible: + * the point is to catch a blocking or tail-handling mistake, and a second + * clever implementation would be liable to the same mistakes. */ + +#include "kernels.h" +#include "test_util.h" + +#include <math.h> +#include <stdlib.h> + +static unsigned rng_state = 12345u; + +static float next_float(void) { + rng_state = rng_state * 1664525u + 1013904223u; + return (float)((rng_state >> 8) & 0xFFFF) / 32768.0f - 1.0f; +} + +static void gemm_ref(float *y, + int ldy, + const float *a, + int lda, + const float *w, + const float *bias, + int m, + int n, + int k) { + for (int s = 0; s < m; ++s) + for (int i = 0; i < n; ++i) { + double acc = bias ? bias[i] : 0.0; + for (int j = 0; j < k; ++j) + acc += (double)a[(size_t)s * lda + j] * w[(size_t)i * k + j]; + y[(size_t)s * ldy + i] = (float)acc; + } +} + +/* Tolerance scales with the window length because that is how many terms the + * two orders can disagree over. */ +static int +check(const char *label, const pyseg_kernels *kern, int m, int n, int k, int lda, int with_bias) { + const size_t a_len = (size_t)(m - 1) * lda + k + 8; + float *a = malloc(a_len * sizeof(float)); + float *w = malloc((size_t)n * k * sizeof(float)); + float *bias = malloc((size_t)n * sizeof(float)); + float *got = malloc((size_t)m * n * sizeof(float)); + float *want = malloc((size_t)m * n * sizeof(float)); + + for (size_t i = 0; i < a_len; ++i) + a[i] = next_float(); + for (size_t i = 0; i < (size_t)n * k; ++i) + w[i] = next_float(); + for (int i = 0; i < n; ++i) + bias[i] = next_float(); + + kern->gemm_nt(got, n, a, lda, w, with_bias ? bias : NULL, m, n, k); + gemm_ref(want, n, a, lda, w, with_bias ? bias : NULL, m, n, k); + + double worst = 0.0; + for (int i = 0; i < m * n; ++i) { + const double d = fabs((double)got[i] - want[i]); + if (d > worst) + worst = d; + } + const double limit = 2e-6 * k; + TEST_ASSERT(worst <= limit, "%s %s m=%d n=%d k=%d lda=%d bias=%d: worst %.3e > %.3e", label, + kern->name, m, n, k, lda, with_bias, worst, limit); + + free(a); + free(w); + free(bias); + free(got); + free(want); + return 0; +} + +/* The SIMD cell update replaces libm's exponential and hyperbolic tangent with + * polynomial approximations, so it is checked against libm directly rather + * than only through the model. The gate arguments are swept well past the + * range where both saturate, because that range is where a range reduction + * built on assembling an exponent by hand goes wrong. */ +static void check_lstm_cell(void) { + enum { HID = 128 }; + float rec[4 * HID], in[4 * HID]; + float h_scalar[HID], c_scalar[HID], h_simd[HID], c_simd[HID]; + + for (int trial = 0; trial < 64; ++trial) { + const float span = 1.0f + (float)trial * 3.0f; /* up to about 190 */ + for (int i = 0; i < 4 * HID; ++i) { + rec[i] = next_float() * span; + in[i] = next_float() * span; + } + for (int i = 0; i < HID; ++i) + c_scalar[i] = c_simd[i] = next_float() * 8.0f; + + pyseg_kernels_scalar()->lstm_cell(h_scalar, c_scalar, rec, in, HID); + pyseg_kernels_simd()->lstm_cell(h_simd, c_simd, rec, in, HID); + + double worst = 0.0; + for (int i = 0; i < HID; ++i) { + const double dh = fabs((double)h_scalar[i] - h_simd[i]); + const double dc = fabs((double)c_scalar[i] - c_simd[i]); + if (dh > worst) + worst = dh; + if (dc > worst) + worst = dc; + if (!isfinite(h_simd[i]) || !isfinite(c_simd[i])) + worst = 1.0; + } + TEST_ASSERT(worst <= 1e-6, "lstm_cell over +/-%.0f: worst %.3e", span, worst); + } +} + +int main(void) { + const pyseg_kernels *sets[] = {pyseg_kernels_scalar(), pyseg_kernels_simd()}; + + /* The shapes the forward pass issues, in order. */ + static const struct { + const char *label; + int m, n, k, lda, bias; + } model_shapes[] = { + {"sinc", 384, 80, 256, 10, 0}, {"sinc-tail", 231, 80, 256, 10, 0}, + {"conv2", 384, 60, 400, 80, 1}, {"conv3", 231, 60, 300, 60, 1}, + {"lstm-in-l0", 589, 512, 60, 60, 1}, {"lstm-in-l1", 589, 512, 256, 256, 1}, + {"lstm-hh", 1, 512, 128, 0, 1}, {"linear0", 589, 128, 256, 256, 1}, + {"linear1", 589, 128, 128, 128, 1}, {"classifier", 589, 7, 128, 128, 1}, + }; + + for (size_t s = 0; s < sizeof sets / sizeof *sets; ++s) + for (size_t i = 0; i < sizeof model_shapes / sizeof *model_shapes; ++i) + check(model_shapes[i].label, sets[s], model_shapes[i].m, model_shapes[i].n, + model_shapes[i].k, model_shapes[i].lda ? model_shapes[i].lda : model_shapes[i].k, + model_shapes[i].bias); + + /* Every combination of leftover rows and leftover weight rows. */ + for (size_t s = 0; s < sizeof sets / sizeof *sets; ++s) + for (int m = 1; m <= 9; ++m) + for (int n = 1; n <= 13; ++n) + for (int k = 4; k <= 20; k += 4) + check("edge", sets[s], m, n, k, k + 3, m & 1); + + /* Panels: a shape larger than one panel and one that straddles a boundary. */ + for (size_t s = 0; s < sizeof sets / sizeof *sets; ++s) { + check("panel-exact", sets[s], 2 * PYSEG_MPANEL, 8, 16, 16, 1); + check("panel-split", sets[s], 2 * PYSEG_MPANEL + 3, 9, 12, 17, 1); + } + + check_lstm_cell(); + return TEST_REPORT(); +} diff --git a/tests/test_parity.c b/tests/test_parity.c new file mode 100644 index 0000000..5002185 --- /dev/null +++ b/tests/test_parity.c @@ -0,0 +1,214 @@ +/* The engine against the PyTorch model on real speech. + * + * Two thresholds, because they answer different questions. The probability + * difference says the arithmetic agrees; the decision agreement says nothing + * downstream changes, which is what actually matters, since every consumer of + * this model takes the most likely class per frame and throws the rest away. + * + * Bit-exactness is not a goal and is not achievable: the reference sums in a + * different order, and float32 addition is not associative. */ + +#include "pyseg.h" +#include "test_util.h" + +#include <math.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> + +#define PROB_TOLERANCE 2e-3 + +typedef struct { + unsigned windows; + float *inputs; /* [windows][PYSEG_WINDOW_SAMPLES] */ + float *probs; /* [windows][PYSEG_FRAMES][PYSEG_CLASSES] */ +} fixture; + +static int load(const char *path, fixture *f) { + FILE *fp = fopen(path, "rb"); + if (!fp) { + fprintf(stderr, "cannot open %s\n", path); + return -1; + } + char magic[4]; + unsigned n, samples, frames; + if (fread(magic, 1, 4, fp) != 4 || memcmp(magic, "PSGG", 4) != 0 || fread(&n, 4, 1, fp) != 1 || + fread(&samples, 4, 1, fp) != 1 || fread(&frames, 4, 1, fp) != 1 || + samples != PYSEG_WINDOW_SAMPLES || frames != PYSEG_FRAMES) { + fprintf(stderr, "%s: not a matching fixture\n", path); + fclose(fp); + return -1; + } + const size_t in_len = (size_t)n * PYSEG_WINDOW_SAMPLES; + const size_t out_len = (size_t)n * PYSEG_FRAMES * PYSEG_CLASSES; + f->windows = n; + f->inputs = malloc(in_len * sizeof(float)); + f->probs = malloc(out_len * sizeof(float)); + const int ok = f->inputs && f->probs && fread(f->inputs, sizeof(float), in_len, fp) == in_len && + fread(f->probs, sizeof(float), out_len, fp) == out_len; + fclose(fp); + if (!ok) { + fprintf(stderr, "%s: truncated\n", path); + return -1; + } + return 0; +} + +static int argmax(const float *row) { + int best = 0; + for (int i = 1; i < PYSEG_CLASSES; ++i) + if (row[i] > row[best]) + best = i; + return best; +} + +static int run(const fixture *f, unsigned flags, const char *label, float *first_window_out) { + pyseg_t *v = pyseg_create_ex(NULL, flags); + TEST_ASSERT(v != NULL, "%s: create failed", label); + if (!v) + return -1; + + float *probs = malloc((size_t)PYSEG_FRAMES * PYSEG_CLASSES * sizeof(float)); + double worst_prob = 0.0, worst_sum = 0.0; + size_t disagreements = 0, total = 0; + + for (unsigned c = 0; c < f->windows; ++c) { + pyseg_process_f32(v, f->inputs + (size_t)c * PYSEG_WINDOW_SAMPLES, probs); + if (c == 0 && first_window_out) + memcpy(first_window_out, probs, (size_t)PYSEG_FRAMES * PYSEG_CLASSES * sizeof(float)); + + const float *want = f->probs + (size_t)c * PYSEG_FRAMES * PYSEG_CLASSES; + for (int t = 0; t < PYSEG_FRAMES; ++t) { + const float *g = probs + (size_t)t * PYSEG_CLASSES; + const float *e = want + (size_t)t * PYSEG_CLASSES; + double sum = 0.0; + for (int i = 0; i < PYSEG_CLASSES; ++i) { + const double d = fabs((double)g[i] - e[i]); + if (d > worst_prob) + worst_prob = d; + sum += g[i]; + } + if (fabs(sum - 1.0) > worst_sum) + worst_sum = fabs(sum - 1.0); + if (argmax(g) != argmax(e)) + ++disagreements; + ++total; + } + } + + fprintf(stderr, "%-8s %-6s %u windows: max |prob diff| %.3e, %zu/%zu frames disagree\n", label, + pyseg_backend_name(v), f->windows, worst_prob, disagreements, total); + + TEST_ASSERT(worst_prob <= PROB_TOLERANCE, "%s: max |prob diff| %.3e > %.3e", label, worst_prob, + PROB_TOLERANCE); + TEST_ASSERT(worst_sum <= 1e-5, "%s: probabilities sum to 1 +/- %.3e", label, worst_sum); + TEST_ASSERT(disagreements == 0, "%s: %zu of %zu frames pick a different class", label, + disagreements, total); + + free(probs); + pyseg_destroy(v); + return 0; +} + +int main(int argc, char **argv) { + if (argc < 2) { + fprintf(stderr, "usage: test_parity <golden.bin>\n"); + return 2; + } + fixture f; + if (load(argv[1], &f) != 0) + return 2; + + const size_t out_len = (size_t)PYSEG_FRAMES * PYSEG_CLASSES; + float *simd_first = malloc(out_len * sizeof(float)); + float *scalar_first = malloc(out_len * sizeof(float)); + + run(&f, 0, "simd", simd_first); + run(&f, PYSEG_SCALAR_KERNELS, "scalar", scalar_first); + + /* The two kernel sets accumulate in different orders, so they are held to + * each other more tightly than either is held to the reference. */ + double worst = 0.0; + for (size_t i = 0; i < out_len; ++i) { + const double d = fabs((double)simd_first[i] - scalar_first[i]); + if (d > worst) + worst = d; + } + fprintf(stderr, "simd vs scalar: max |prob diff| %.3e\n", worst); + TEST_ASSERT(worst <= 1e-4, "simd vs scalar diverge by %.3e", worst); + + /* Batching is a scheduling decision, not a numerical one: the windows of a + * batch share weight loads but never touch each other's activations, so + * grouping must not move a single bit. */ + if (f.windows >= 4) { + pyseg_t *batched = pyseg_create_batch(NULL, 0, 4); + pyseg_t *single = pyseg_create(); + TEST_ASSERT(batched && pyseg_batch(batched) == 4, "batched create failed"); + if (batched && single) { + float *grouped = malloc(4 * out_len * sizeof(float)); + float *apart = malloc(4 * out_len * sizeof(float)); + pyseg_process_batch_f32(batched, f.inputs, 4, grouped); + for (int i = 0; i < 4; ++i) + pyseg_process_f32(single, f.inputs + (size_t)i * PYSEG_WINDOW_SAMPLES, + apart + (size_t)i * out_len); + TEST_ASSERT(memcmp(grouped, apart, 4 * out_len * sizeof(float)) == 0, + "a batch of four disagrees with four separate windows"); + + /* A partial batch is the tail of every recording. */ + pyseg_process_batch_f32(batched, f.inputs, 3, grouped); + TEST_ASSERT(memcmp(grouped, apart, 3 * out_len * sizeof(float)) == 0, + "a partial batch disagrees with separate windows"); + free(grouped); + free(apart); + } + pyseg_destroy(batched); + pyseg_destroy(single); + } + + TEST_ASSERT(pyseg_create_batch(NULL, 0, 0) == NULL, "a batch of zero was accepted"); + TEST_ASSERT(pyseg_create_batch(NULL, 0, PYSEG_MAX_BATCH + 1) == NULL, + "an oversized batch was accepted"); + + /* Determinism: the engine carries no state between windows, so re-running + * a window must reproduce it bit for bit. */ + pyseg_t *v = pyseg_create(); + float *again = malloc(out_len * sizeof(float)); + if (v) { + pyseg_process_f32(v, f.inputs + (size_t)(f.windows - 1) * PYSEG_WINDOW_SAMPLES, again); + float *repeat = malloc(out_len * sizeof(float)); + pyseg_process_f32(v, f.inputs, repeat); + pyseg_process_f32(v, f.inputs + (size_t)(f.windows - 1) * PYSEG_WINDOW_SAMPLES, repeat); + TEST_ASSERT(memcmp(again, repeat, out_len * sizeof(float)) == 0, + "a repeated window produced different bits"); + free(repeat); + pyseg_destroy(v); + } + + /* A weight file must parse to the same numbers as the embedded copy. */ + pyseg_t *fromfile = pyseg_create_from_file("pyseg_weights.bin"); + if (fromfile) { + float *filed = malloc(out_len * sizeof(float)); + pyseg_process_f32(fromfile, f.inputs, filed); + pyseg_t *embedded = pyseg_create(); + float *builtin = malloc(out_len * sizeof(float)); + pyseg_process_f32(embedded, f.inputs, builtin); + TEST_ASSERT(memcmp(filed, builtin, out_len * sizeof(float)) == 0, + "mapped weights disagree with the embedded copy"); + free(filed); + free(builtin); + pyseg_destroy(embedded); + pyseg_destroy(fromfile); + } else { + fprintf(stderr, "note: pyseg_weights.bin not readable, file path untested\n"); + } + + TEST_ASSERT(pyseg_create_from_file("/nonexistent/pyseg.bin") == NULL, + "a missing weight file was accepted"); + + free(again); + free(simd_first); + free(scalar_first); + free(f.inputs); + free(f.probs); + return TEST_REPORT(); +} diff --git a/tests/test_robust.c b/tests/test_robust.c new file mode 100644 index 0000000..7dd43a7 --- /dev/null +++ b/tests/test_robust.c @@ -0,0 +1,133 @@ +/* Inputs the model will never see in training but will see in production. + * + * The model normalises its input by the window's own standard deviation, so + * digital silence divides by roughly the epsilon in the normalisation and a + * constant offset does the same. Those are the cases that produce infinities + * if the epsilon is dropped or applied in the wrong place, and a whole window + * of silence is entirely ordinary in a real recording. */ + +#include "pyseg.h" +#include "test_util.h" + +#include <math.h> +#include <stdlib.h> +#include <string.h> + +static unsigned rng_state = 987654321u; + +static float next_float(void) { + rng_state = rng_state * 1664525u + 1013904223u; + return (float)((rng_state >> 8) & 0xFFFF) / 32768.0f - 1.0f; +} + +static void check_output(const char *label, const float *probs) { + for (int t = 0; t < PYSEG_FRAMES; ++t) { + const float *row = probs + (size_t)t * PYSEG_CLASSES; + double sum = 0.0; + for (int i = 0; i < PYSEG_CLASSES; ++i) { + if (!isfinite(row[i]) || row[i] < 0.0f || row[i] > 1.0f) { + TEST_ASSERT(0, "%s: frame %d class %d is %g", label, t, i, row[i]); + return; + } + sum += row[i]; + } + if (fabs(sum - 1.0) > 1e-5) { + TEST_ASSERT(0, "%s: frame %d sums to %.9f", label, t, sum); + return; + } + } + TEST_ASSERT(1, "%s", label); +} + +int main(void) { + pyseg_t *v = pyseg_create(); + TEST_ASSERT(v != NULL, "create failed"); + if (!v) + return TEST_REPORT(); + + float *wav = malloc(PYSEG_WINDOW_SAMPLES * sizeof(float)); + float *probs = malloc((size_t)PYSEG_FRAMES * PYSEG_CLASSES * sizeof(float)); + float *active = malloc((size_t)PYSEG_FRAMES * PYSEG_SPEAKERS * sizeof(float)); + int16_t *pcm = malloc(PYSEG_WINDOW_SAMPLES * sizeof(int16_t)); + + memset(wav, 0, PYSEG_WINDOW_SAMPLES * sizeof(float)); + pyseg_process_f32(v, wav, probs); + check_output("digital silence", probs); + + for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) + wav[i] = 0.5f; + pyseg_process_f32(v, wav, probs); + check_output("constant offset", probs); + + for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) + wav[i] = (i & 1) ? 1.0f : -1.0f; + pyseg_process_f32(v, wav, probs); + check_output("full-scale square wave", probs); + + for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) + wav[i] = next_float() * 4.0f; + pyseg_process_f32(v, wav, probs); + check_output("out-of-range noise", probs); + + for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) + wav[i] = next_float() * 1e-8f; + pyseg_process_f32(v, wav, probs); + check_output("near-denormal noise", probs); + + /* A single impulse in an otherwise silent window: the normalisation + * amplifies the silence around it by the full dynamic range. */ + memset(wav, 0, PYSEG_WINDOW_SAMPLES * sizeof(float)); + wav[PYSEG_WINDOW_SAMPLES / 2] = 1.0f; + pyseg_process_f32(v, wav, probs); + check_output("lone impulse", probs); + + for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) + pcm[i] = (int16_t)(i & 1 ? 32767 : -32768); + pyseg_process_s16(v, pcm, probs); + check_output("s16 full scale", probs); + + /* The s16 entry point must be the f32 one with a fixed scaling. */ + for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) { + pcm[i] = (int16_t)((int)(next_float() * 16000.0f)); + wav[i] = (float)pcm[i] * (1.0f / 32768.0f); + } + float *from_s16 = malloc((size_t)PYSEG_FRAMES * PYSEG_CLASSES * sizeof(float)); + pyseg_process_s16(v, pcm, from_s16); + pyseg_process_f32(v, wav, probs); + TEST_ASSERT(memcmp(from_s16, probs, (size_t)PYSEG_FRAMES * PYSEG_CLASSES * sizeof(float)) == 0, + "s16 and f32 entry points disagree on identical audio"); + + /* Decoding must land inside the powerset and stay consistent with it. */ + pyseg_to_multilabel(probs, active); + for (int t = 0; t < PYSEG_FRAMES; ++t) { + const float *a = active + (size_t)t * PYSEG_SPEAKERS; + int count = 0; + for (int s = 0; s < PYSEG_SPEAKERS; ++s) { + if (a[s] != 0.0f && a[s] != 1.0f) { + TEST_ASSERT(0, "frame %d speaker %d is %g, not a hard decision", t, s, a[s]); + break; + } + count += a[s] != 0.0f; + } + if (count > 2) { + TEST_ASSERT(0, "frame %d has %d speakers, more than the powerset allows", t, count); + break; + } + } + TEST_ASSERT(1, "multilabel decoding stays inside the powerset"); + + TEST_ASSERT(pyseg_powerset_class(-1, -1) == 0, "empty set is not class 0"); + TEST_ASSERT(pyseg_powerset_class(0, -1) == 1, "speaker 0 alone is not class 1"); + TEST_ASSERT(pyseg_powerset_class(2, -1) == 3, "speaker 2 alone is not class 3"); + TEST_ASSERT(pyseg_powerset_class(0, 1) == 4, "speakers 0 and 1 are not class 4"); + TEST_ASSERT(pyseg_powerset_class(1, 2) == 6, "speakers 1 and 2 are not class 6"); + + free(from_s16); + free(pcm); + free(active); + free(probs); + free(wav); + pyseg_destroy(v); + pyseg_destroy(NULL); + return TEST_REPORT(); +} diff --git a/tests/test_slide.c b/tests/test_slide.c new file mode 100644 index 0000000..a0fc8ca --- /dev/null +++ b/tests/test_slide.c @@ -0,0 +1,212 @@ +/* The sliding-window driver against the aggregation pyannote performs, and the + * binarizer against its own definition. + * + * The aggregate is not just the model run repeatedly: window placement, + * rounding onto the frame grid, the Hamming weights and the dropped tail all + * have to match, and each is a place where an off-by-one produces output that + * looks entirely reasonable and is shifted by a frame. */ + +#include "pyseg.h" +#include "test_util.h" + +#include <math.h> +#include <stdio.h> +#include <stdlib.h> +#include <string.h> + +typedef struct { + unsigned num_samples; + unsigned frames; + double step; + float *samples; + float *tracks; /* [frames][PYSEG_TRACKS] */ +} fixture; + +static int load(const char *path, fixture *f) { + FILE *fp = fopen(path, "rb"); + if (!fp) { + fprintf(stderr, "cannot open %s\n", path); + return -1; + } + char magic[4]; + if (fread(magic, 1, 4, fp) != 4 || memcmp(magic, "PSGS", 4) != 0 || + fread(&f->num_samples, 4, 1, fp) != 1 || fread(&f->frames, 4, 1, fp) != 1 || + fread(&f->step, 8, 1, fp) != 1) { + fprintf(stderr, "%s: not a matching fixture\n", path); + fclose(fp); + return -1; + } + f->samples = malloc((size_t)f->num_samples * sizeof(float)); + f->tracks = malloc((size_t)f->frames * PYSEG_TRACKS * sizeof(float)); + const int ok = f->samples && f->tracks && + fread(f->samples, sizeof(float), f->num_samples, fp) == f->num_samples && + fread(f->tracks, sizeof(float), (size_t)f->frames * PYSEG_TRACKS, fp) == + (size_t)f->frames * PYSEG_TRACKS; + fclose(fp); + if (!ok) { + fprintf(stderr, "%s: truncated\n", path); + return -1; + } + return 0; +} + +static double compare(const float *got, const float *want, size_t n) { + double worst = 0.0; + for (size_t i = 0; i < n; ++i) { + const double d = fabs((double)got[i] - want[i]); + if (d > worst) + worst = d; + } + return worst; +} + +static void check_slide(const fixture *f, int threads) { + pyseg_slide_params p; + pyseg_slide_params_default(&p); + p.step = f->step; + p.threads = threads; + + const size_t frames = pyseg_slide_frames(f->num_samples, p.step); + TEST_ASSERT(frames == f->frames, "threads=%d: %zu frames, reference has %u", threads, frames, + f->frames); + if (frames != f->frames) + return; + + float *tracks = malloc(frames * PYSEG_TRACKS * sizeof(float)); + const size_t got = pyseg_slide(f->samples, f->num_samples, &p, tracks); + TEST_ASSERT(got == frames, "threads=%d: slide returned %zu", threads, got); + + if (got == frames) { + const double worst = compare(tracks, f->tracks, frames * PYSEG_TRACKS); + fprintf(stderr, "slide threads=%d: %zu frames, max |track diff| %.3e\n", threads, frames, + worst); + TEST_ASSERT(worst <= 1e-5, "threads=%d: max |track diff| %.3e", threads, worst); + } + free(tracks); +} + +/* A hand-built score sequence, so the expected segments can be reasoned about + * rather than captured. */ +static void check_binarize(void) { + enum { N = 40 }; + float scores[N * PYSEG_TRACKS]; + memset(scores, 0, sizeof scores); + for (int f = 10; f < 20; ++f) + scores[f * PYSEG_TRACKS] = 1.0f; + for (int f = 22; f < 25; ++f) + scores[f * PYSEG_TRACKS] = 1.0f; + + pyseg_binarize_params p; + pyseg_binarize_params_default(&p); + + pyseg_segment out[8]; + size_t n = pyseg_binarize(scores, N, PYSEG_TRACK_SPEECH, &p, out, 8); + TEST_ASSERT(n == 2, "two runs of ones gave %zu segments", n); + if (n == 2) { + const double half = 0.5 * PYSEG_FRAME_SIZE / PYSEG_SAMPLE_RATE; + TEST_ASSERT(fabs(out[0].start - (pyseg_frame_time(10) + half)) < 1e-9, + "first segment starts at %.6f", out[0].start); + TEST_ASSERT(fabs(out[0].end - (pyseg_frame_time(20) + half)) < 1e-9, + "first segment ends at %.6f", out[0].end); + TEST_ASSERT(fabs(out[1].start - (pyseg_frame_time(22) + half)) < 1e-9, + "second segment starts at %.6f", out[1].start); + } + + /* The gap between the two runs is two frames, about 34 ms. */ + p.min_duration_off = 0.1; + n = pyseg_binarize(scores, N, PYSEG_TRACK_SPEECH, &p, out, 8); + TEST_ASSERT(n == 1, "a 34 ms gap survived a 100 ms fill: %zu segments", n); + + p.min_duration_off = 0.0; + p.min_duration_on = 0.1; + n = pyseg_binarize(scores, N, PYSEG_TRACK_SPEECH, &p, out, 8); + TEST_ASSERT(n == 1, "a 51 ms segment survived a 100 ms minimum: %zu segments", n); + + /* Counting without writing must agree with writing. */ + pyseg_binarize_params_default(&p); + TEST_ASSERT(pyseg_binarize(scores, N, PYSEG_TRACK_SPEECH, &p, NULL, 0) == 2, + "counting pass disagrees with the writing pass"); + TEST_ASSERT(pyseg_binarize(scores, N, PYSEG_TRACK_SPEECH, &p, out, 1) == 2, + "an undersized buffer did not report the full count"); + + /* A track that is high from the first frame to the last is one segment. */ + for (int f = 0; f < N; ++f) + scores[f * PYSEG_TRACKS + PYSEG_TRACK_OVERLAP] = 1.0f; + n = pyseg_binarize(scores, N, PYSEG_TRACK_OVERLAP, &p, out, 8); + TEST_ASSERT(n == 1, "an always-active track gave %zu segments", n); + + TEST_ASSERT(pyseg_binarize(scores, N, PYSEG_TRACKS, &p, out, 8) == 0, + "an out-of-range track was accepted"); +} + +/* Frame counts read off pyannote's Inference for the same inputs. The formula + * involves two roundings and a tail-window rule, so a table of answers taken + * from the reference is worth more than a second derivation of the formula. */ +static void check_frame_counts(void) { + static const struct { + size_t samples; + double step; + size_t frames; + } cases[] = { + {720000, 1.0, 2668}, {640000, 2.0, 2371}, {592000, 0.5, 2194}, + {400000, 3.0, 1482}, {144000, 1.0, 534}, {160000, 1.0, 594}, + }; + for (size_t i = 0; i < sizeof cases / sizeof *cases; ++i) { + const size_t got = pyseg_slide_frames(cases[i].samples, cases[i].step); + TEST_ASSERT(got == cases[i].frames, "%zu samples at step %g: %zu frames, expected %zu", + cases[i].samples, cases[i].step, got, cases[i].frames); + } + TEST_ASSERT(pyseg_slide_frames(0, 1.0) == 0, "an empty recording gave frames"); + TEST_ASSERT(pyseg_slide_frames(160000, 0.0) == 0, "a zero step gave frames"); +} + +int main(int argc, char **argv) { + if (argc < 2) { + fprintf(stderr, "usage: test_slide <slide_golden.bin>\n"); + return 2; + } + fixture f; + if (load(argv[1], &f) != 0) + return 2; + + check_frame_counts(); + + check_slide(&f, 1); + check_slide(&f, 4); + check_binarize(); + + /* The 16-bit entry point must agree with the float one on audio that came + * from 16-bit samples in the first place, which the fixture did. */ + { + int16_t *pcm = malloc(f.num_samples * sizeof(int16_t)); + float *from_s16 = malloc((size_t)f.frames * PYSEG_TRACKS * sizeof(float)); + for (unsigned i = 0; i < f.num_samples; ++i) + pcm[i] = (int16_t)lrintf(f.samples[i] * 32768.0f); + pyseg_slide_params p; + pyseg_slide_params_default(&p); + p.step = f.step; + TEST_ASSERT(pyseg_slide_s16(pcm, f.num_samples, &p, from_s16) == f.frames, + "the 16-bit slide failed"); + const double worst = compare(from_s16, f.tracks, (size_t)f.frames * PYSEG_TRACKS); + TEST_ASSERT(worst <= 1e-5, "16-bit slide differs by %.3e", worst); + free(from_s16); + free(pcm); + } + + /* Recordings shorter than one window are padded, not rejected. */ + float *tiny = calloc(PYSEG_SAMPLE_RATE, sizeof(float)); + for (int i = 0; i < PYSEG_SAMPLE_RATE; ++i) + tiny[i] = 0.1f * sinf((float)i * 0.01f); + const size_t tiny_frames = pyseg_slide_frames(PYSEG_SAMPLE_RATE, 1.0); + TEST_ASSERT(tiny_frames == PYSEG_SAMPLE_RATE / PYSEG_FRAME_STEP + 1, + "a one-second recording gave %zu frames", tiny_frames); + float *tiny_tracks = malloc(tiny_frames * PYSEG_TRACKS * sizeof(float)); + TEST_ASSERT(pyseg_slide(tiny, PYSEG_SAMPLE_RATE, NULL, tiny_tracks) == tiny_frames, + "a one-second recording failed"); + free(tiny_tracks); + free(tiny); + + free(f.samples); + free(f.tracks); + return TEST_REPORT(); +} diff --git a/tests/test_util.h b/tests/test_util.h new file mode 100644 index 0000000..1d8c80e --- /dev/null +++ b/tests/test_util.h @@ -0,0 +1,29 @@ +/* Assertion harness for the C tests: every check reports, nothing aborts, and + * the process exit status is the verdict. Keeping going after a failure is + * deliberate; when a numeric change breaks several gates at once, the full + * list localises it faster than the first failure alone. */ + +#ifndef PYSEG_TEST_UTIL_H +#define PYSEG_TEST_UTIL_H + +#include <stdio.h> + +static int test_failures; +static int test_checks; + +#define TEST_ASSERT(cond, ...) \ + do { \ + ++test_checks; \ + if (!(cond)) { \ + ++test_failures; \ + fprintf(stderr, "FAIL %s:%d: ", __FILE__, __LINE__); \ + fprintf(stderr, __VA_ARGS__); \ + fputc('\n', stderr); \ + } \ + } while (0) + +#define TEST_REPORT() \ + (fprintf(stderr, "%s: %d checks, %d failed\n", __FILE__, test_checks, test_failures), \ + test_failures ? 1 : 0) + +#endif /* PYSEG_TEST_UTIL_H */ |
