145 {
147 const auto n = static_cast<uint32_t>(chunks);
148 context->upload_shared_raw(0,
Audio,
reinterpret_cast<const uint8_t*
>(audio.data()), audio.size_bytes());
149
150 std::vector<DependencyStage> stages;
152
153 uint32_t t = k_frame;
154 const auto convolve = [&](
const TensorKernels::Conv1d& layer, uint64_t src,
size_t weight, std::optional<size_t> bias, uint64_t dst) {
157 };
158
159 convolve({ .c_in = 1, .c_out = 258, .kernel = 256, .stride = 128 },
address(Frames), StftWeight, std::nullopt,
address(Transform));
161
163 convolve({ .c_in = 129, .c_out = 128, .kernel = 3, .pad = 1, .activation = relu },
address(Magnitude), Conv1Weight, Conv1Bias,
address(Encoded1));
164 convolve({ .c_in = 128, .c_out = 64, .kernel = 3, .stride = 2, .pad = 1, .activation = relu },
address(Encoded1), Conv2Weight, Conv2Bias,
address(Encoded2));
165 convolve({ .c_in = 64, .c_out = 64, .kernel = 3, .stride = 2, .pad = 1, .activation = relu },
address(Encoded2), Conv3Weight, Conv3Bias,
address(Encoded3));
166 convolve({ .c_in = 64, .c_out = k_hidden, .kernel = 3, .pad = 1, .activation = relu },
address(Encoded3), Conv4Weight, Conv4Bias,
address(Encoded4));
167
170
171 t = 1;
174
175 context->record_stages(stages);
177 if (!read) {
178 return {};
179 }
180
181 std::vector<float> probabilities(chunks);
182 context->download_shared(0, Probability, probabilities.data(), probabilities.size() *
sizeof(
float));
186 context->download_shared(0, Hidden,
kept_hidden.data() + held, chunks * k_hidden *
sizeof(
float));
187 }
188 return probabilities;
189 }
static constexpr uint32_t context_samples
static constexpr uint32_t chunk_samples
std::vector< double > convolve(std::span< const double > src, std::span< const double > ir)
Linear FFT convolution.
DependencyStage frame_audio_stage(uint64_t samples, uint64_t frames, uint32_t chunks, uint32_t chunk_size, uint32_t context, uint32_t pad)
One nn_frame_audio dispatch: cut a sample stream into overlapping frames, reflect padded on the right...
DependencyStage lstm_sequence_stage(uint64_t x, uint64_t w_ih, uint64_t w_hh, uint64_t b_ih, uint64_t b_hh, uint64_t state, uint64_t hs, uint32_t steps, uint32_t hidden)
One nn_lstm_sequence dispatch: an LSTM cell over a whole sequence in one workgroup.
uint32_t conv1d_length(const Conv1d &layer, uint32_t t_in)
Output length of a convolution, 0 when the window does not fit the padded input.
DependencyStage magnitude_stage(uint64_t src, uint64_t dst, uint32_t batch, uint32_t channels, uint32_t time)
One nn_magnitude dispatch: sqrt(re^2 + im^2) of a complex signal stored as two channel halves.
DependencyStage conv1d_stage(const Conv1d &layer, uint64_t src, uint64_t weights, uint64_t bias, uint64_t dst, uint32_t batch, uint32_t t_in)
One nn_conv1d dispatch: [batch, c_in, t_in] to [batch, c_out, conv1d_length].
static constexpr DomainSpec Audio
Domain constant for Audio domain.
std::vector< float > kept_hidden
uint64_t weight_address(size_t slot) const
void ensure_capacity(size_t chunks)
uint64_t address(uint32_t buffer) const
std::unique_ptr< ShaderExecutionContext<> > context