/* 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 #include 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(); }