diff options
Diffstat (limited to 'src/kernels.h')
| -rw-r--r-- | src/kernels.h | 67 |
1 files changed, 67 insertions, 0 deletions
diff --git a/src/kernels.h b/src/kernels.h new file mode 100644 index 0000000..678b844 --- /dev/null +++ b/src/kernels.h @@ -0,0 +1,67 @@ +/* The single dense primitive behind the whole forward pass. + * + * Every layer of this model computes, for a run of output positions, the dot + * products of a window of activations against a fixed set of weight rows. The + * convolutions reach that form because activations are stored time-major: the + * taps a kernel-5 convolution needs for one output step are five consecutive + * frames, which is one contiguous span, and the next output step starts one + * frame later. The sinc filterbank is the same thing with a span of 251 + * samples advancing ten samples at a time. So the only parameter that changes + * between layers is how far apart consecutive windows start, which is what + * lda is; it is unrelated to the window length k. + * + * Taking every output position in one call is what makes this fast. A single + * position would re-read the whole weight matrix; taking four at a time reuses + * each loaded weight vector four times, and the outer panel loop keeps the + * activations that those four rows share resident in L1 while the weights + * stream past once. + * + * Both implementations are always compiled and chosen through this table, so + * one binary can compare them. + */ + +#ifndef PYSEG_KERNELS_H +#define PYSEG_KERNELS_H + +/* Output rows processed against the full weight matrix before moving on. Sized + * so the activation window and the output panel together stay inside L1. */ +#define PYSEG_MPANEL 32 + +typedef struct { + const char *name; + + /* For s in 0..m-1 and i in 0..n-1: + * y[s*ldy + i] = (bias ? bias[i] : 0) + dot(a + s*lda, w + i*k, k) + * k must be a multiple of four. */ + void (*gemm_nt)(float *y, + int ldy, + const float *a, + int lda, + const float *w, + const float *bias, + int m, + int n, + int k); + + /* One LSTM cell update in place. The preactivations are the sum of rec, + * the hidden-to-hidden product, and in, the projected input for this + * time step; both hold the four gate blocks in PyTorch's order, input then + * forget then candidate then output, each hid long: + * c = sigmoid(f) * c + sigmoid(i) * tanh(g) + * h = sigmoid(o) * tanh(c) + * + * This is here rather than in the forward pass because it is not a + * negligible epilogue. A window drives 4712 cell updates, which is three + * million logistic and hyperbolic evaluations; called one at a time + * through libm that costs more than either convolution stage. hid is a + * multiple of four. */ + void (*lstm_cell)(float *h, float *c, const float *rec, const float *in, int hid); +} pyseg_kernels; + +const pyseg_kernels *pyseg_kernels_scalar(void); + +/* The best set this build supports; equals the scalar set on architectures + * without an implementation here. */ +const pyseg_kernels *pyseg_kernels_simd(void); + +#endif /* PYSEG_KERNELS_H */ |
