Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
52 lines
1.9 KiB
C++
52 lines
1.9 KiB
C++
#pragma once
|
|
|
|
#include "engine/community_models/sanotts/config.h"
|
|
#ifdef ENGINE_HAVE_NEON
|
|
#include "engine/community_models/sanotts/neon_weights.h"
|
|
#endif
|
|
|
|
#include <cstdint>
|
|
#include <functional>
|
|
#include <memory>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
namespace engine::models::sanotts {
|
|
|
|
/** Complete nano-lineage decoder (frame-stage
|
|
* acoustic blocks -> mel projection -> noise-fed ConvNeXt-1D -> spectrum
|
|
* head), for CPU sessions selecting neon through sanotts.exec_plan.
|
|
*
|
|
* Same inputs and output as the GGML decoder graph in runtime.cpp, so the
|
|
* rest of the pipeline (durations, token stage, host iSTFT, DC block) is
|
|
* unchanged. Outputs match the GGML path to float rounding, not bit-exact:
|
|
* acoustic.output is folded into decoder.embed, GELU uses an erf
|
|
* approximation with |error| < 1.5e-7, and sums run in a different order.
|
|
*
|
|
*/
|
|
class SanoTtsNeonDecoder {
|
|
public:
|
|
static bool available();
|
|
|
|
#ifdef ENGINE_HAVE_NEON
|
|
SanoTtsNeonDecoder(const SanoTtsConfig &config, const SanoTtsNeonWeights &weights);
|
|
#else
|
|
using WeightReader = std::function<std::vector<float>(const std::string &)>;
|
|
SanoTtsNeonDecoder(const SanoTtsConfig &config, const WeightReader &read_weight, int threads);
|
|
#endif
|
|
~SanoTtsNeonDecoder();
|
|
|
|
/** context: acoustic_hidden x frames, feats: 3 x frames, noise:
|
|
* noise_channels x frames (all channel-major, as the GGML graph's
|
|
* inputs). Returns frames x (n_fft + 2) rows of [log-magnitude | phase],
|
|
* as the GGML graph's spectrum output. Not thread-safe: one call at a
|
|
* time per instance (the runtime serialises synthesize()). */
|
|
std::vector<float> decode(const std::vector<float> &context, const std::vector<float> &feats,
|
|
const std::vector<float> &noise, int64_t frames, int threads);
|
|
|
|
private:
|
|
struct Impl;
|
|
std::unique_ptr<Impl> impl_;
|
|
};
|
|
|
|
} // namespace engine::models::sanotts
|