aboutsummaryrefslogtreecommitdiff
path: root/weights.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-10-08 06:49:33 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-10-08 07:21:37 +0200
commita5af9c3d351085f4005a949f8c382a3aad4e0b29 (patch)
tree0ad133d931553e68a68588e983976a66b3309083 /weights.h
downloadsvad-a5af9c3d351085f4005a949f8c382a3aad4e0b29.tar.gz
svad-a5af9c3d351085f4005a949f8c382a3aad4e0b29.zip
initv0.1.0
Diffstat (limited to 'weights.h')
-rw-r--r--weights.h73
1 files changed, 73 insertions, 0 deletions
diff --git a/weights.h b/weights.h
new file mode 100644
index 0000000..524fd4e
--- /dev/null
+++ b/weights.h
@@ -0,0 +1,73 @@
+/* Weight manifest shared by the compiled-in tables and the blob loader.
+ *
+ * Every pointer refers to a tensor already in the layout the kernels want; see
+ * tools/export_weights.py for the offline transforms. The two rate variants
+ * differ only in the transform size and in conv1's input channel count, so one
+ * struct describes both. */
+
+#ifndef SVAD_WEIGHTS_H
+#define SVAD_WEIGHTS_H
+
+#include <stddef.h>
+
+#if defined(__GNUC__) || defined(__clang__)
+# define SVAD_ALIGN64 __attribute__((aligned(64)))
+#else
+# define SVAD_ALIGN64 _Alignas(64)
+#endif
+
+/* Shapes that do not vary with sample rate. */
+#define SVAD_C1 128 /* conv1 output channels */
+#define SVAD_C2 64 /* conv2 output channels */
+#define SVAD_C3 64 /* conv3 output channels */
+#define SVAD_C4 128 /* conv4 output channels, and the LSTM input width */
+#define SVAD_HID 128 /* LSTM hidden width */
+#define SVAD_GATES 512 /* 4 * SVAD_HID, gate order i,f,g,o */
+#define SVAD_FRAMES 4 /* transform frames per chunk */
+
+typedef struct {
+ int sample_rate; /* 8000 or 16000 */
+ int chunk; /* new samples per call */
+ int context; /* samples carried from the previous call */
+ int n_fft; /* transform window */
+ int hop; /* n_fft / 2 */
+ int pad; /* right-side reflect pad */
+ int frames; /* == SVAD_FRAMES */
+ int bins; /* n_fft / 2 + 1 magnitude bins */
+ int bins_pad; /* bins rounded up to the kernel's channel granularity */
+
+ const float *stft_basis; /* [2*bins][n_fft], dense transform path */
+ const float *stft_window; /* [n_fft], periodic Hann, rFFT path */
+ const float *conv1_w; /* [128][3*bins_pad] */
+ const float *conv1_b; /* [128] */
+ const float *conv2_w; /* [64][3*128] */
+ const float *conv2_b; /* [64] */
+ const float *conv3_w; /* [64][3*64] */
+ const float *conv3_b; /* [64] */
+ const float *conv4_w; /* [128][3*64] */
+ const float *conv4_b; /* [128] */
+ const float *lstm_w; /* [512][256] over concat(feat, h) */
+ const float *lstm_b; /* [512], bias_ih + bias_hh merged */
+ const float *head_w; /* [128] */
+ float head_b;
+} svad_weights;
+
+/* Largest values across both rate variants, for static scratch sizing. */
+#define SVAD_MAX_CHUNK 512
+#define SVAD_MAX_CONTEXT 64
+#define SVAD_MAX_PAD 64
+#define SVAD_MAX_NFFT 256
+#define SVAD_MAX_BINS 129
+#define SVAD_MAX_BINS_PAD 132
+#define SVAD_MAX_XP (SVAD_MAX_CONTEXT + SVAD_MAX_CHUNK + SVAD_MAX_PAD)
+
+/* Compiled-in tables. Returns NULL for an unsupported rate. */
+const svad_weights *svad_weights_builtin(int sample_rate);
+
+/* Maps an SVAD blob and fills out with pointers into it. On success stores the
+ * mapping in *map / *map_len for svad_weights_unmap. Returns 0 on success. */
+int svad_weights_map(
+ const char *path, int sample_rate, int verify, svad_weights *out, void **map, size_t *map_len);
+void svad_weights_unmap(void *map, size_t map_len);
+
+#endif /* SVAD_WEIGHTS_H */