Add C++ API for streaming zipformer ASR on RK NPU (#1908)
This commit is contained in:
Fangjun Kuang
19
sherpa-onnx/csrc/rknn/macros.h
Normal file
19
sherpa-onnx/csrc/rknn/macros.h
Normal file
@@ -0,0 +1,19 @@
|
||||
// sherpa-onnx/csrc/macros.h
|
||||
//
|
||||
// Copyright 2025 Xiaomi Corporation
|
||||
|
||||
#ifndef SHERPA_ONNX_CSRC_RKNN_MACROS_H_
|
||||
#define SHERPA_ONNX_CSRC_RKNN_MACROS_H_
|
||||
|
||||
#include "sherpa-onnx/csrc/macros.h"
|
||||
|
||||
#define SHERPA_ONNX_RKNN_CHECK(ret, msg, ...) \
|
||||
do { \
|
||||
if (ret != RKNN_SUCC) { \
|
||||
SHERPA_ONNX_LOGE("Return code is: %d", ret); \
|
||||
SHERPA_ONNX_LOGE(msg, ##__VA_ARGS__); \
|
||||
SHERPA_ONNX_EXIT(-1); \
|
||||
} \
|
||||
} while (0)
|
||||
|
||||
#endif // SHERPA_ONNX_CSRC_RKNN_MACROS_H_
|
||||
231
sherpa-onnx/csrc/rknn/online-recognizer-transducer-rknn-impl.h
Normal file
231
sherpa-onnx/csrc/rknn/online-recognizer-transducer-rknn-impl.h
Normal file
@@ -0,0 +1,231 @@
|
||||
// sherpa-onnx/csrc/rknn/online-recognizer-transducer-rknn-impl.h
|
||||
//
|
||||
// Copyright (c) 2025 Xiaomi Corporation
|
||||
|
||||
#ifndef SHERPA_ONNX_CSRC_RKNN_ONLINE_RECOGNIZER_TRANSDUCER_RKNN_IMPL_H_
|
||||
#define SHERPA_ONNX_CSRC_RKNN_ONLINE_RECOGNIZER_TRANSDUCER_RKNN_IMPL_H_
|
||||
|
||||
#include <algorithm>
|
||||
#include <memory>
|
||||
#include <sstream>
|
||||
#include <string>
|
||||
#include <utility>
|
||||
#include <vector>
|
||||
|
||||
#include "sherpa-onnx/csrc/macros.h"
|
||||
#include "sherpa-onnx/csrc/online-recognizer-impl.h"
|
||||
#include "sherpa-onnx/csrc/online-recognizer.h"
|
||||
#include "sherpa-onnx/csrc/rknn/online-stream-rknn.h"
|
||||
#include "sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.h"
|
||||
#include "sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.h"
|
||||
#include "sherpa-onnx/csrc/symbol-table.h"
|
||||
|
||||
namespace sherpa_onnx {
|
||||
|
||||
OnlineRecognizerResult Convert(const OnlineTransducerDecoderResultRknn &src,
|
||||
const SymbolTable &sym_table,
|
||||
float frame_shift_ms, int32_t subsampling_factor,
|
||||
int32_t segment, int32_t frames_since_start) {
|
||||
OnlineRecognizerResult r;
|
||||
r.tokens.reserve(src.tokens.size());
|
||||
r.timestamps.reserve(src.tokens.size());
|
||||
|
||||
std::string text;
|
||||
for (auto i : src.tokens) {
|
||||
auto sym = sym_table[i];
|
||||
|
||||
text.append(sym);
|
||||
|
||||
if (sym.size() == 1 && (sym[0] < 0x20 || sym[0] > 0x7e)) {
|
||||
// for bpe models with byte_fallback
|
||||
// (but don't rewrite printable characters 0x20..0x7e,
|
||||
// which collide with standard BPE units)
|
||||
std::ostringstream os;
|
||||
os << "<0x" << std::hex << std::uppercase
|
||||
<< (static_cast<int32_t>(sym[0]) & 0xff) << ">";
|
||||
sym = os.str();
|
||||
}
|
||||
|
||||
r.tokens.push_back(std::move(sym));
|
||||
}
|
||||
|
||||
if (sym_table.IsByteBpe()) {
|
||||
text = sym_table.DecodeByteBpe(text);
|
||||
}
|
||||
|
||||
r.text = std::move(text);
|
||||
|
||||
float frame_shift_s = frame_shift_ms / 1000. * subsampling_factor;
|
||||
for (auto t : src.timestamps) {
|
||||
float time = frame_shift_s * t;
|
||||
r.timestamps.push_back(time);
|
||||
}
|
||||
|
||||
r.segment = segment;
|
||||
r.start_time = frames_since_start * frame_shift_ms / 1000.;
|
||||
|
||||
return r;
|
||||
}
|
||||
|
||||
class OnlineRecognizerTransducerRknnImpl : public OnlineRecognizerImpl {
|
||||
public:
|
||||
explicit OnlineRecognizerTransducerRknnImpl(
|
||||
const OnlineRecognizerConfig &config)
|
||||
: OnlineRecognizerImpl(config),
|
||||
config_(config),
|
||||
endpoint_(config_.endpoint_config),
|
||||
model_(std::make_unique<OnlineZipformerTransducerModelRknn>(
|
||||
config.model_config)) {
|
||||
if (!config.model_config.tokens_buf.empty()) {
|
||||
sym_ = SymbolTable(config.model_config.tokens_buf, false);
|
||||
} else {
|
||||
/// assuming tokens_buf and tokens are guaranteed not being both empty
|
||||
sym_ = SymbolTable(config.model_config.tokens, true);
|
||||
}
|
||||
|
||||
if (sym_.Contains("<unk>")) {
|
||||
unk_id_ = sym_["<unk>"];
|
||||
}
|
||||
|
||||
decoder_ = std::make_unique<OnlineTransducerGreedySearchDecoderRknn>(
|
||||
model_.get(), unk_id_);
|
||||
}
|
||||
|
||||
template <typename Manager>
|
||||
explicit OnlineRecognizerTransducerRknnImpl(
|
||||
Manager *mgr, const OnlineRecognizerConfig &config)
|
||||
: OnlineRecognizerImpl(mgr, config),
|
||||
config_(config),
|
||||
endpoint_(config_.endpoint_config),
|
||||
model_(
|
||||
std::make_unique<OnlineZipformerTransducerModelRknn>(mgr, config)) {
|
||||
// TODO(fangjun): Support Android
|
||||
}
|
||||
|
||||
std::unique_ptr<OnlineStream> CreateStream() const override {
|
||||
auto stream = std::make_unique<OnlineStreamRknn>(config_.feat_config);
|
||||
auto r = decoder_->GetEmptyResult();
|
||||
stream->SetZipformerResult(std::move(r));
|
||||
stream->SetZipformerEncoderStates(model_->GetEncoderInitStates());
|
||||
return stream;
|
||||
}
|
||||
|
||||
std::unique_ptr<OnlineStream> CreateStream(
|
||||
const std::string &hotwords) const override {
|
||||
SHERPA_ONNX_LOGE("Hotwords for RKNN is not supported now.");
|
||||
return CreateStream();
|
||||
}
|
||||
|
||||
bool IsReady(OnlineStream *s) const override {
|
||||
return s->GetNumProcessedFrames() + model_->ChunkSize() <
|
||||
s->NumFramesReady();
|
||||
}
|
||||
|
||||
// Warmping up engine with wp: warm_up count and max-batch-size
|
||||
|
||||
void DecodeStreams(OnlineStream **ss, int32_t n) const override {
|
||||
for (int32_t i = 0; i < n; ++i) {
|
||||
DecodeStream(reinterpret_cast<OnlineStreamRknn *>(ss[i]));
|
||||
}
|
||||
}
|
||||
|
||||
OnlineRecognizerResult GetResult(OnlineStream *s) const override {
|
||||
OnlineTransducerDecoderResultRknn decoder_result =
|
||||
reinterpret_cast<OnlineStreamRknn *>(s)->GetZipformerResult();
|
||||
decoder_->StripLeadingBlanks(&decoder_result);
|
||||
// TODO(fangjun): Remember to change these constants if needed
|
||||
int32_t frame_shift_ms = 10;
|
||||
int32_t subsampling_factor = 4;
|
||||
auto r = Convert(decoder_result, sym_, frame_shift_ms, subsampling_factor,
|
||||
s->GetCurrentSegment(), s->GetNumFramesSinceStart());
|
||||
r.text = ApplyInverseTextNormalization(std::move(r.text));
|
||||
return r;
|
||||
}
|
||||
|
||||
bool IsEndpoint(OnlineStream *s) const override {
|
||||
if (!config_.enable_endpoint) {
|
||||
return false;
|
||||
}
|
||||
|
||||
int32_t num_processed_frames = s->GetNumProcessedFrames();
|
||||
|
||||
// frame shift is 10 milliseconds
|
||||
float frame_shift_in_seconds = 0.01;
|
||||
|
||||
// subsampling factor is 4
|
||||
int32_t trailing_silence_frames = reinterpret_cast<OnlineStreamRknn *>(s)
|
||||
->GetZipformerResult()
|
||||
.num_trailing_blanks *
|
||||
4;
|
||||
|
||||
return endpoint_.IsEndpoint(num_processed_frames, trailing_silence_frames,
|
||||
frame_shift_in_seconds);
|
||||
}
|
||||
|
||||
void Reset(OnlineStream *s) const override {
|
||||
int32_t context_size = model_->ContextSize();
|
||||
|
||||
{
|
||||
// segment is incremented only when the last
|
||||
// result is not empty, contains non-blanks and longer than context_size)
|
||||
const auto &r =
|
||||
reinterpret_cast<OnlineStreamRknn *>(s)->GetZipformerResult();
|
||||
if (!r.tokens.empty() && r.tokens.back() != 0 &&
|
||||
r.tokens.size() > context_size) {
|
||||
s->GetCurrentSegment() += 1;
|
||||
}
|
||||
}
|
||||
|
||||
// reset encoder states
|
||||
// reinterpret_cast<OnlineStreamRknn*>(s)->SetZipformerEncoderStates(model_->GetEncoderInitStates());
|
||||
auto r = decoder_->GetEmptyResult();
|
||||
auto last_result =
|
||||
reinterpret_cast<OnlineStreamRknn *>(s)->GetZipformerResult();
|
||||
|
||||
// if last result is not empty, then
|
||||
// preserve last tokens as the context for next result
|
||||
if (static_cast<int32_t>(last_result.tokens.size()) > context_size) {
|
||||
r.tokens = {last_result.tokens.end() - context_size,
|
||||
last_result.tokens.end()};
|
||||
}
|
||||
reinterpret_cast<OnlineStreamRknn *>(s)->SetZipformerResult(std::move(r));
|
||||
|
||||
// Note: We only update counters. The underlying audio samples
|
||||
// are not discarded.
|
||||
s->Reset();
|
||||
}
|
||||
|
||||
private:
|
||||
void DecodeStream(OnlineStreamRknn *s) const {
|
||||
int32_t chunk_size = model_->ChunkSize();
|
||||
int32_t chunk_shift = model_->ChunkShift();
|
||||
|
||||
int32_t feature_dim = s->FeatureDim();
|
||||
|
||||
const auto num_processed_frames = s->GetNumProcessedFrames();
|
||||
|
||||
std::vector<float> features =
|
||||
s->GetFrames(num_processed_frames, chunk_size);
|
||||
s->GetNumProcessedFrames() += chunk_shift;
|
||||
|
||||
auto &states = s->GetZipformerEncoderStates();
|
||||
|
||||
auto p = model_->RunEncoder(features, std::move(states));
|
||||
states = std::move(p.second);
|
||||
|
||||
auto &r = s->GetZipformerResult();
|
||||
decoder_->Decode(std::move(p.first), &r);
|
||||
}
|
||||
|
||||
private:
|
||||
OnlineRecognizerConfig config_;
|
||||
SymbolTable sym_;
|
||||
Endpoint endpoint_;
|
||||
int32_t unk_id_ = -1;
|
||||
std::unique_ptr<OnlineZipformerTransducerModelRknn> model_;
|
||||
std::unique_ptr<OnlineTransducerGreedySearchDecoderRknn> decoder_;
|
||||
};
|
||||
|
||||
} // namespace sherpa_onnx
|
||||
|
||||
#endif // SHERPA_ONNX_CSRC_RKNN_ONLINE_RECOGNIZER_TRANSDUCER_RKNN_IMPL_H_
|
||||
60
sherpa-onnx/csrc/rknn/online-stream-rknn.cc
Normal file
60
sherpa-onnx/csrc/rknn/online-stream-rknn.cc
Normal file
@@ -0,0 +1,60 @@
|
||||
// sherpa-onnx/csrc/rknn/online-stream-rknn.cc
|
||||
//
|
||||
// Copyright (c) 2025 Xiaomi Corporation
|
||||
|
||||
#include "sherpa-onnx/csrc/rknn/online-stream-rknn.h"
|
||||
|
||||
#include <utility>
|
||||
#include <vector>
|
||||
|
||||
namespace sherpa_onnx {
|
||||
|
||||
class OnlineStreamRknn::Impl {
|
||||
public:
|
||||
void SetZipformerEncoderStates(std::vector<std::vector<uint8_t>> states) {
|
||||
states_ = std::move(states);
|
||||
}
|
||||
|
||||
std::vector<std::vector<uint8_t>> &GetZipformerEncoderStates() {
|
||||
return states_;
|
||||
}
|
||||
|
||||
void SetZipformerResult(OnlineTransducerDecoderResultRknn r) {
|
||||
result_ = std::move(r);
|
||||
}
|
||||
|
||||
OnlineTransducerDecoderResultRknn &GetZipformerResult() { return result_; }
|
||||
|
||||
private:
|
||||
std::vector<std::vector<uint8_t>> states_;
|
||||
OnlineTransducerDecoderResultRknn result_;
|
||||
};
|
||||
|
||||
OnlineStreamRknn::OnlineStreamRknn(
|
||||
const FeatureExtractorConfig &config /*= {}*/,
|
||||
ContextGraphPtr context_graph /*= nullptr*/)
|
||||
: OnlineStream(config, context_graph), impl_(std::make_unique<Impl>()) {}
|
||||
|
||||
OnlineStreamRknn::~OnlineStreamRknn() = default;
|
||||
|
||||
void OnlineStreamRknn::SetZipformerEncoderStates(
|
||||
std::vector<std::vector<uint8_t>> states) const {
|
||||
impl_->SetZipformerEncoderStates(std::move(states));
|
||||
}
|
||||
|
||||
std::vector<std::vector<uint8_t>> &OnlineStreamRknn::GetZipformerEncoderStates()
|
||||
const {
|
||||
return impl_->GetZipformerEncoderStates();
|
||||
}
|
||||
|
||||
void OnlineStreamRknn::SetZipformerResult(
|
||||
OnlineTransducerDecoderResultRknn r) const {
|
||||
impl_->SetZipformerResult(std::move(r));
|
||||
}
|
||||
|
||||
OnlineTransducerDecoderResultRknn &OnlineStreamRknn::GetZipformerResult()
|
||||
const {
|
||||
return impl_->GetZipformerResult();
|
||||
}
|
||||
|
||||
} // namespace sherpa_onnx
|
||||
38
sherpa-onnx/csrc/rknn/online-stream-rknn.h
Normal file
38
sherpa-onnx/csrc/rknn/online-stream-rknn.h
Normal file
@@ -0,0 +1,38 @@
|
||||
// sherpa-onnx/csrc/rknn/online-stream-rknn.h
|
||||
//
|
||||
// Copyright (c) 2025 Xiaomi Corporation
|
||||
#ifndef SHERPA_ONNX_CSRC_RKNN_ONLINE_STREAM_RKNN_H_
|
||||
#define SHERPA_ONNX_CSRC_RKNN_ONLINE_STREAM_RKNN_H_
|
||||
#include <memory>
|
||||
#include <vector>
|
||||
|
||||
#include "rknn_api.h" // NOLINT
|
||||
#include "sherpa-onnx/csrc/online-stream.h"
|
||||
#include "sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.h"
|
||||
|
||||
namespace sherpa_onnx {
|
||||
|
||||
class OnlineStreamRknn : public OnlineStream {
|
||||
public:
|
||||
explicit OnlineStreamRknn(const FeatureExtractorConfig &config = {},
|
||||
ContextGraphPtr context_graph = nullptr);
|
||||
|
||||
~OnlineStreamRknn();
|
||||
|
||||
void SetZipformerEncoderStates(
|
||||
std::vector<std::vector<uint8_t>> states) const;
|
||||
|
||||
std::vector<std::vector<uint8_t>> &GetZipformerEncoderStates() const;
|
||||
|
||||
void SetZipformerResult(OnlineTransducerDecoderResultRknn r) const;
|
||||
|
||||
OnlineTransducerDecoderResultRknn &GetZipformerResult() const;
|
||||
|
||||
private:
|
||||
class Impl;
|
||||
std::unique_ptr<Impl> impl_;
|
||||
};
|
||||
|
||||
} // namespace sherpa_onnx
|
||||
|
||||
#endif // SHERPA_ONNX_CSRC_RKNN_ONLINE_STREAM_RKNN_H_
|
||||
@@ -0,0 +1,94 @@
|
||||
// sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.cc
|
||||
//
|
||||
// Copyright (c) 2025 Xiaomi Corporation
|
||||
|
||||
#include "sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.h"
|
||||
|
||||
#include <algorithm>
|
||||
#include <utility>
|
||||
#include <vector>
|
||||
|
||||
#include "sherpa-onnx/csrc/macros.h"
|
||||
|
||||
namespace sherpa_onnx {
|
||||
|
||||
OnlineTransducerDecoderResultRknn
|
||||
OnlineTransducerGreedySearchDecoderRknn::GetEmptyResult() const {
|
||||
int32_t context_size = model_->ContextSize();
|
||||
int32_t blank_id = 0; // always 0
|
||||
OnlineTransducerDecoderResultRknn r;
|
||||
r.tokens.resize(context_size, -1);
|
||||
r.tokens.back() = blank_id;
|
||||
|
||||
return r;
|
||||
}
|
||||
|
||||
void OnlineTransducerGreedySearchDecoderRknn::StripLeadingBlanks(
|
||||
OnlineTransducerDecoderResultRknn *r) const {
|
||||
int32_t context_size = model_->ContextSize();
|
||||
|
||||
auto start = r->tokens.begin() + context_size;
|
||||
auto end = r->tokens.end();
|
||||
|
||||
r->tokens = std::vector<int64_t>(start, end);
|
||||
}
|
||||
|
||||
void OnlineTransducerGreedySearchDecoderRknn::Decode(
|
||||
std::vector<float> encoder_out,
|
||||
OnlineTransducerDecoderResultRknn *result) const {
|
||||
auto &r = result[0];
|
||||
auto attr = model_->GetEncoderOutAttr();
|
||||
int32_t num_frames = attr.dims[1];
|
||||
int32_t encoder_out_dim = attr.dims[2];
|
||||
|
||||
int32_t vocab_size = model_->VocabSize();
|
||||
int32_t context_size = model_->ContextSize();
|
||||
|
||||
std::vector<int64_t> decoder_input;
|
||||
std::vector<float> decoder_out;
|
||||
|
||||
if (r.previous_decoder_out.empty()) {
|
||||
decoder_input = {r.tokens.begin() + (r.tokens.size() - context_size),
|
||||
r.tokens.end()};
|
||||
decoder_out = model_->RunDecoder(std::move(decoder_input));
|
||||
|
||||
} else {
|
||||
decoder_out = std::move(r.previous_decoder_out);
|
||||
}
|
||||
|
||||
const float *p_encoder_out = encoder_out.data();
|
||||
for (int32_t t = 0; t != num_frames; ++t) {
|
||||
auto logit = model_->RunJoiner(p_encoder_out, decoder_out.data());
|
||||
p_encoder_out += encoder_out_dim;
|
||||
|
||||
bool emitted = false;
|
||||
if (blank_penalty_ > 0.0) {
|
||||
logit[0] -= blank_penalty_; // assuming blank id is 0
|
||||
}
|
||||
|
||||
auto y = static_cast<int32_t>(std::distance(
|
||||
logit.data(),
|
||||
std::max_element(logit.data(), logit.data() + vocab_size)));
|
||||
// blank id is hardcoded to 0
|
||||
// also, it treats unk as blank
|
||||
if (y != 0 && y != unk_id_) {
|
||||
emitted = true;
|
||||
r.tokens.push_back(y);
|
||||
r.timestamps.push_back(t + r.frame_offset);
|
||||
r.num_trailing_blanks = 0;
|
||||
} else {
|
||||
++r.num_trailing_blanks;
|
||||
}
|
||||
|
||||
if (emitted) {
|
||||
decoder_input = {r.tokens.begin() + (r.tokens.size() - context_size),
|
||||
r.tokens.end()};
|
||||
decoder_out = model_->RunDecoder(std::move(decoder_input));
|
||||
}
|
||||
}
|
||||
|
||||
r.frame_offset += num_frames;
|
||||
r.previous_decoder_out = std::move(decoder_out);
|
||||
}
|
||||
|
||||
} // namespace sherpa_onnx
|
||||
@@ -0,0 +1,52 @@
|
||||
// sherpa-onnx/csrc/rknn/online-transducer-greedy-search-decoder-rknn.h
|
||||
//
|
||||
// Copyright (c) 2025 Xiaomi Corporation
|
||||
|
||||
#ifndef SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_RKNN_H_
|
||||
#define SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_RKNN_H_
|
||||
|
||||
#include <vector>
|
||||
|
||||
#include "sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.h"
|
||||
|
||||
namespace sherpa_onnx {
|
||||
|
||||
struct OnlineTransducerDecoderResultRknn {
|
||||
/// Number of frames after subsampling we have decoded so far
|
||||
int32_t frame_offset = 0;
|
||||
|
||||
/// The decoded token IDs so far
|
||||
std::vector<int64_t> tokens;
|
||||
|
||||
/// number of trailing blank frames decoded so far
|
||||
int32_t num_trailing_blanks = 0;
|
||||
|
||||
/// timestamps[i] contains the output frame index where tokens[i] is decoded.
|
||||
std::vector<int32_t> timestamps;
|
||||
|
||||
std::vector<float> previous_decoder_out;
|
||||
};
|
||||
|
||||
class OnlineTransducerGreedySearchDecoderRknn {
|
||||
public:
|
||||
explicit OnlineTransducerGreedySearchDecoderRknn(
|
||||
OnlineZipformerTransducerModelRknn *model, int32_t unk_id = 2,
|
||||
float blank_penalty = 0.0)
|
||||
: model_(model), unk_id_(unk_id), blank_penalty_(blank_penalty) {}
|
||||
|
||||
OnlineTransducerDecoderResultRknn GetEmptyResult() const;
|
||||
|
||||
void StripLeadingBlanks(OnlineTransducerDecoderResultRknn *r) const;
|
||||
|
||||
void Decode(std::vector<float> encoder_out,
|
||||
OnlineTransducerDecoderResultRknn *result) const;
|
||||
|
||||
private:
|
||||
OnlineZipformerTransducerModelRknn *model_; // Not owned
|
||||
int32_t unk_id_;
|
||||
float blank_penalty_;
|
||||
};
|
||||
|
||||
} // namespace sherpa_onnx
|
||||
|
||||
#endif // SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_RKNN_H_
|
||||
781
sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.cc
Normal file
781
sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.cc
Normal file
@@ -0,0 +1,781 @@
|
||||
// sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.cc
|
||||
//
|
||||
// Copyright (c) 2023 Xiaomi Corporation
|
||||
|
||||
#include "sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.h"
|
||||
|
||||
#include <memory>
|
||||
#include <sstream>
|
||||
#include <string>
|
||||
#include <unordered_map>
|
||||
#include <utility>
|
||||
#include <vector>
|
||||
|
||||
#if __ANDROID_API__ >= 9
|
||||
#include "android/asset_manager.h"
|
||||
#include "android/asset_manager_jni.h"
|
||||
#endif
|
||||
|
||||
#if __OHOS__
|
||||
#include "rawfile/raw_file_manager.h"
|
||||
#endif
|
||||
|
||||
#include "sherpa-onnx/csrc/file-utils.h"
|
||||
#include "sherpa-onnx/csrc/rknn/macros.h"
|
||||
#include "sherpa-onnx/csrc/text-utils.h"
|
||||
|
||||
namespace sherpa_onnx {
|
||||
|
||||
// chw -> hwc
|
||||
static void Transpose(const float *src, int32_t n, int32_t channel,
|
||||
int32_t height, int32_t width, float *dst) {
|
||||
for (int32_t i = 0; i < n; ++i) {
|
||||
for (int32_t h = 0; h < height; ++h) {
|
||||
for (int32_t w = 0; w < width; ++w) {
|
||||
for (int32_t c = 0; c < channel; ++c) {
|
||||
// dst[h, w, c] = src[c, h, w]
|
||||
dst[i * height * width * channel + h * width * channel + w * channel +
|
||||
c] = src[i * height * width * channel + c * height * width +
|
||||
h * width + w];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static std::string ToString(const rknn_tensor_attr &attr) {
|
||||
std::ostringstream os;
|
||||
os << "{";
|
||||
os << attr.index;
|
||||
os << ", name: " << attr.name;
|
||||
os << ", shape: (";
|
||||
std::string sep;
|
||||
for (int32_t i = 0; i < static_cast<int32_t>(attr.n_dims); ++i) {
|
||||
os << sep << attr.dims[i];
|
||||
sep = ",";
|
||||
}
|
||||
os << ")";
|
||||
os << ", n_elems: " << attr.n_elems;
|
||||
os << ", size: " << attr.size;
|
||||
os << ", fmt: " << get_format_string(attr.fmt);
|
||||
os << ", type: " << get_type_string(attr.type);
|
||||
os << ", pass_through: " << (attr.pass_through ? "true" : "false");
|
||||
os << "}";
|
||||
return os.str();
|
||||
}
|
||||
|
||||
static std::unordered_map<std::string, std::string> Parse(
|
||||
const rknn_custom_string &custom_string) {
|
||||
std::unordered_map<std::string, std::string> ans;
|
||||
std::vector<std::string> fields;
|
||||
SplitStringToVector(custom_string.string, ";", false, &fields);
|
||||
|
||||
std::vector<std::string> tmp;
|
||||
for (const auto &f : fields) {
|
||||
SplitStringToVector(f, "=", false, &tmp);
|
||||
if (tmp.size() != 2) {
|
||||
SHERPA_ONNX_LOGE("Invalid custom string %s for %s", custom_string.string,
|
||||
f.c_str());
|
||||
SHERPA_ONNX_EXIT(-1);
|
||||
}
|
||||
ans[std::move(tmp[0])] = std::move(tmp[1]);
|
||||
}
|
||||
|
||||
return ans;
|
||||
}
|
||||
|
||||
class OnlineZipformerTransducerModelRknn::Impl {
|
||||
public:
|
||||
~Impl() {
|
||||
auto ret = rknn_destroy(encoder_ctx_);
|
||||
if (ret != RKNN_SUCC) {
|
||||
SHERPA_ONNX_LOGE("Failed to destroy the encoder context");
|
||||
}
|
||||
|
||||
ret = rknn_destroy(decoder_ctx_);
|
||||
if (ret != RKNN_SUCC) {
|
||||
SHERPA_ONNX_LOGE("Failed to destroy the decoder context");
|
||||
}
|
||||
|
||||
ret = rknn_destroy(joiner_ctx_);
|
||||
if (ret != RKNN_SUCC) {
|
||||
SHERPA_ONNX_LOGE("Failed to destroy the joiner context");
|
||||
}
|
||||
}
|
||||
|
||||
explicit Impl(const OnlineModelConfig &config) : config_(config) {
|
||||
{
|
||||
auto buf = ReadFile(config.transducer.encoder);
|
||||
InitEncoder(buf.data(), buf.size());
|
||||
}
|
||||
|
||||
{
|
||||
auto buf = ReadFile(config.transducer.decoder);
|
||||
InitDecoder(buf.data(), buf.size());
|
||||
}
|
||||
|
||||
{
|
||||
auto buf = ReadFile(config.transducer.joiner);
|
||||
InitJoiner(buf.data(), buf.size());
|
||||
}
|
||||
|
||||
// Now select which core to run for RK3588
|
||||
int32_t ret_encoder = RKNN_SUCC;
|
||||
int32_t ret_decoder = RKNN_SUCC;
|
||||
int32_t ret_joiner = RKNN_SUCC;
|
||||
switch (config_.num_threads) {
|
||||
case 1:
|
||||
ret_encoder = rknn_set_core_mask(encoder_ctx_, RKNN_NPU_CORE_AUTO);
|
||||
ret_decoder = rknn_set_core_mask(decoder_ctx_, RKNN_NPU_CORE_AUTO);
|
||||
ret_joiner = rknn_set_core_mask(joiner_ctx_, RKNN_NPU_CORE_AUTO);
|
||||
break;
|
||||
case 0:
|
||||
ret_encoder = rknn_set_core_mask(encoder_ctx_, RKNN_NPU_CORE_0);
|
||||
ret_decoder = rknn_set_core_mask(decoder_ctx_, RKNN_NPU_CORE_0);
|
||||
ret_joiner = rknn_set_core_mask(joiner_ctx_, RKNN_NPU_CORE_0);
|
||||
break;
|
||||
case -1:
|
||||
ret_encoder = rknn_set_core_mask(encoder_ctx_, RKNN_NPU_CORE_1);
|
||||
ret_decoder = rknn_set_core_mask(decoder_ctx_, RKNN_NPU_CORE_1);
|
||||
ret_joiner = rknn_set_core_mask(joiner_ctx_, RKNN_NPU_CORE_1);
|
||||
break;
|
||||
case -2:
|
||||
ret_encoder = rknn_set_core_mask(encoder_ctx_, RKNN_NPU_CORE_2);
|
||||
ret_decoder = rknn_set_core_mask(decoder_ctx_, RKNN_NPU_CORE_2);
|
||||
ret_joiner = rknn_set_core_mask(joiner_ctx_, RKNN_NPU_CORE_2);
|
||||
break;
|
||||
case -3:
|
||||
ret_encoder = rknn_set_core_mask(encoder_ctx_, RKNN_NPU_CORE_0_1);
|
||||
ret_decoder = rknn_set_core_mask(decoder_ctx_, RKNN_NPU_CORE_0_1);
|
||||
ret_joiner = rknn_set_core_mask(joiner_ctx_, RKNN_NPU_CORE_0_1);
|
||||
break;
|
||||
case -4:
|
||||
ret_encoder = rknn_set_core_mask(encoder_ctx_, RKNN_NPU_CORE_0_1_2);
|
||||
ret_decoder = rknn_set_core_mask(decoder_ctx_, RKNN_NPU_CORE_0_1_2);
|
||||
ret_joiner = rknn_set_core_mask(joiner_ctx_, RKNN_NPU_CORE_0_1_2);
|
||||
break;
|
||||
default:
|
||||
SHERPA_ONNX_LOGE(
|
||||
"Valid num_threads for rk npu is 1 (auto), 0 (core 0), -1 (core "
|
||||
"1), -2 (core 2), -3 (core 0_1), -4 (core 0_1_2). Given: %d",
|
||||
config_.num_threads);
|
||||
break;
|
||||
}
|
||||
if (ret_encoder != RKNN_SUCC) {
|
||||
SHERPA_ONNX_LOGE(
|
||||
"Failed to select npu core to run encoder (You can ignore it if you "
|
||||
"are not using RK3588.");
|
||||
}
|
||||
|
||||
if (ret_decoder != RKNN_SUCC) {
|
||||
SHERPA_ONNX_LOGE(
|
||||
"Failed to select npu core to run decoder (You can ignore it if you "
|
||||
"are not using RK3588.");
|
||||
}
|
||||
|
||||
if (ret_decoder != RKNN_SUCC) {
|
||||
SHERPA_ONNX_LOGE(
|
||||
"Failed to select npu core to run joiner (You can ignore it if you "
|
||||
"are not using RK3588.");
|
||||
}
|
||||
}
|
||||
|
||||
// TODO(fangjun): Support Android
|
||||
|
||||
std::vector<std::vector<uint8_t>> GetEncoderInitStates() const {
|
||||
// encoder_input_attrs_[0] is for the feature
|
||||
// encoder_input_attrs_[1:] is for states
|
||||
// so we use -1 here
|
||||
std::vector<std::vector<uint8_t>> states(encoder_input_attrs_.size() - 1);
|
||||
|
||||
int32_t i = -1;
|
||||
for (auto &attr : encoder_input_attrs_) {
|
||||
i += 1;
|
||||
if (i == 0) {
|
||||
// skip processing the attr for features.
|
||||
continue;
|
||||
}
|
||||
|
||||
if (attr.type == RKNN_TENSOR_FLOAT16) {
|
||||
states[i - 1].resize(attr.n_elems * sizeof(float));
|
||||
} else if (attr.type == RKNN_TENSOR_INT64) {
|
||||
states[i - 1].resize(attr.n_elems * sizeof(int64_t));
|
||||
} else {
|
||||
SHERPA_ONNX_LOGE("Unsupported tensor type: %d, %s", attr.type,
|
||||
get_type_string(attr.type));
|
||||
SHERPA_ONNX_EXIT(-1);
|
||||
}
|
||||
}
|
||||
|
||||
return states;
|
||||
}
|
||||
|
||||
std::pair<std::vector<float>, std::vector<std::vector<uint8_t>>> RunEncoder(
|
||||
std::vector<float> features,
|
||||
std::vector<std::vector<uint8_t>> states) const {
|
||||
std::vector<rknn_input> inputs(encoder_input_attrs_.size());
|
||||
|
||||
for (int32_t i = 0; i < static_cast<int32_t>(inputs.size()); ++i) {
|
||||
auto &input = inputs[i];
|
||||
auto &attr = encoder_input_attrs_[i];
|
||||
input.index = attr.index;
|
||||
|
||||
if (attr.type == RKNN_TENSOR_FLOAT16) {
|
||||
input.type = RKNN_TENSOR_FLOAT32;
|
||||
} else if (attr.type == RKNN_TENSOR_INT64) {
|
||||
input.type = RKNN_TENSOR_INT64;
|
||||
} else {
|
||||
SHERPA_ONNX_LOGE("Unsupported tensor type %d, %s", attr.type,
|
||||
get_type_string(attr.type));
|
||||
SHERPA_ONNX_EXIT(-1);
|
||||
}
|
||||
|
||||
input.fmt = attr.fmt;
|
||||
if (i == 0) {
|
||||
input.buf = reinterpret_cast<void *>(features.data());
|
||||
input.size = features.size() * sizeof(float);
|
||||
} else {
|
||||
input.buf = reinterpret_cast<void *>(states[i - 1].data());
|
||||
input.size = states[i - 1].size();
|
||||
}
|
||||
}
|
||||
|
||||
std::vector<float> encoder_out(encoder_output_attrs_[0].n_elems);
|
||||
|
||||
// Note(fangjun): We can reuse the memory from input argument `states`
|
||||
// auto next_states = GetEncoderInitStates();
|
||||
auto &next_states = states;
|
||||
|
||||
std::vector<rknn_output> outputs(encoder_output_attrs_.size());
|
||||
for (int32_t i = 0; i < outputs.size(); ++i) {
|
||||
auto &output = outputs[i];
|
||||
auto &attr = encoder_output_attrs_[i];
|
||||
output.index = attr.index;
|
||||
output.is_prealloc = 1;
|
||||
|
||||
if (attr.type == RKNN_TENSOR_FLOAT16) {
|
||||
output.want_float = 1;
|
||||
} else if (attr.type == RKNN_TENSOR_INT64) {
|
||||
output.want_float = 0;
|
||||
} else {
|
||||
SHERPA_ONNX_LOGE("Unsupported tensor type %d, %s", attr.type,
|
||||
get_type_string(attr.type));
|
||||
SHERPA_ONNX_EXIT(-1);
|
||||
}
|
||||
|
||||
if (i == 0) {
|
||||
output.size = encoder_out.size() * sizeof(float);
|
||||
output.buf = reinterpret_cast<void *>(encoder_out.data());
|
||||
} else {
|
||||
output.size = next_states[i - 1].size();
|
||||
output.buf = reinterpret_cast<void *>(next_states[i - 1].data());
|
||||
}
|
||||
}
|
||||
|
||||
auto ret = rknn_inputs_set(encoder_ctx_, inputs.size(), inputs.data());
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to set encoder inputs");
|
||||
|
||||
ret = rknn_run(encoder_ctx_, nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to run encoder");
|
||||
|
||||
ret =
|
||||
rknn_outputs_get(encoder_ctx_, outputs.size(), outputs.data(), nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get encoder output");
|
||||
|
||||
for (int32_t i = 0; i < next_states.size(); ++i) {
|
||||
const auto &attr = encoder_input_attrs_[i + 1];
|
||||
if (attr.n_dims == 4) {
|
||||
// TODO(fangjun): The transpose is copied from
|
||||
// https://github.com/airockchip/rknn_model_zoo/blob/main/examples/zipformer/cpp/process.cc#L22
|
||||
// I don't understand why we need to do that.
|
||||
std::vector<uint8_t> dst(next_states[i].size());
|
||||
int32_t n = attr.dims[0];
|
||||
int32_t h = attr.dims[1];
|
||||
int32_t w = attr.dims[2];
|
||||
int32_t c = attr.dims[3];
|
||||
Transpose(reinterpret_cast<const float *>(next_states[i].data()), n, c,
|
||||
h, w, reinterpret_cast<float *>(dst.data()));
|
||||
next_states[i] = std::move(dst);
|
||||
}
|
||||
}
|
||||
|
||||
return {std::move(encoder_out), std::move(next_states)};
|
||||
}
|
||||
|
||||
std::vector<float> RunDecoder(std::vector<int64_t> decoder_input) const {
|
||||
auto &attr = decoder_input_attrs_[0];
|
||||
rknn_input input;
|
||||
|
||||
input.index = 0;
|
||||
input.type = RKNN_TENSOR_INT64;
|
||||
input.fmt = attr.fmt;
|
||||
input.buf = decoder_input.data();
|
||||
input.size = decoder_input.size() * sizeof(int64_t);
|
||||
|
||||
std::vector<float> decoder_out(decoder_output_attrs_[0].n_elems);
|
||||
rknn_output output;
|
||||
output.index = decoder_output_attrs_[0].index;
|
||||
output.is_prealloc = 1;
|
||||
output.want_float = 1;
|
||||
output.size = decoder_out.size() * sizeof(float);
|
||||
output.buf = decoder_out.data();
|
||||
|
||||
auto ret = rknn_inputs_set(decoder_ctx_, 1, &input);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to set decoder inputs");
|
||||
|
||||
ret = rknn_run(decoder_ctx_, nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to run decoder");
|
||||
|
||||
ret = rknn_outputs_get(decoder_ctx_, 1, &output, nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get decoder output");
|
||||
|
||||
return decoder_out;
|
||||
}
|
||||
|
||||
std::vector<float> RunJoiner(const float *encoder_out,
|
||||
const float *decoder_out) const {
|
||||
std::vector<rknn_input> inputs(2);
|
||||
inputs[0].index = 0;
|
||||
inputs[0].type = RKNN_TENSOR_FLOAT32;
|
||||
inputs[0].fmt = joiner_input_attrs_[0].fmt;
|
||||
inputs[0].buf = const_cast<float *>(encoder_out);
|
||||
inputs[0].size = joiner_input_attrs_[0].n_elems * sizeof(float);
|
||||
|
||||
inputs[1].index = 1;
|
||||
inputs[1].type = RKNN_TENSOR_FLOAT32;
|
||||
inputs[1].fmt = joiner_input_attrs_[1].fmt;
|
||||
inputs[1].buf = const_cast<float *>(decoder_out);
|
||||
inputs[1].size = joiner_input_attrs_[1].n_elems * sizeof(float);
|
||||
|
||||
std::vector<float> joiner_out(joiner_output_attrs_[0].n_elems);
|
||||
rknn_output output;
|
||||
output.index = joiner_output_attrs_[0].index;
|
||||
output.is_prealloc = 1;
|
||||
output.want_float = 1;
|
||||
output.size = joiner_out.size() * sizeof(float);
|
||||
output.buf = joiner_out.data();
|
||||
|
||||
auto ret = rknn_inputs_set(joiner_ctx_, inputs.size(), inputs.data());
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to set joiner inputs");
|
||||
|
||||
ret = rknn_run(joiner_ctx_, nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to run joiner");
|
||||
|
||||
ret = rknn_outputs_get(joiner_ctx_, 1, &output, nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get joiner output");
|
||||
|
||||
return joiner_out;
|
||||
}
|
||||
|
||||
int32_t ContextSize() const { return context_size_; }
|
||||
|
||||
int32_t ChunkSize() const { return T_; }
|
||||
|
||||
int32_t ChunkShift() const { return decode_chunk_len_; }
|
||||
|
||||
int32_t VocabSize() const { return vocab_size_; }
|
||||
|
||||
rknn_tensor_attr GetEncoderOutAttr() const {
|
||||
return encoder_output_attrs_[0];
|
||||
}
|
||||
|
||||
private:
|
||||
void InitEncoder(void *model_data, size_t model_data_length) {
|
||||
auto ret =
|
||||
rknn_init(&encoder_ctx_, model_data, model_data_length, 0, nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to init encoder '%s'",
|
||||
config_.transducer.encoder.c_str());
|
||||
|
||||
if (config_.debug) {
|
||||
rknn_sdk_version v;
|
||||
ret = rknn_query(encoder_ctx_, RKNN_QUERY_SDK_VERSION, &v, sizeof(v));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get rknn sdk version");
|
||||
|
||||
SHERPA_ONNX_LOGE("sdk api version: %s, driver version: %s", v.api_version,
|
||||
v.drv_version);
|
||||
}
|
||||
|
||||
rknn_input_output_num io_num;
|
||||
ret = rknn_query(encoder_ctx_, RKNN_QUERY_IN_OUT_NUM, &io_num,
|
||||
sizeof(io_num));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret,
|
||||
"Failed to get I/O information for the encoder");
|
||||
|
||||
if (config_.debug) {
|
||||
SHERPA_ONNX_LOGE("encoder: %d inputs, %d outputs",
|
||||
static_cast<int32_t>(io_num.n_input),
|
||||
static_cast<int32_t>(io_num.n_output));
|
||||
}
|
||||
|
||||
encoder_input_attrs_.resize(io_num.n_input);
|
||||
encoder_output_attrs_.resize(io_num.n_output);
|
||||
|
||||
int32_t i = 0;
|
||||
for (auto &attr : encoder_input_attrs_) {
|
||||
memset(&attr, 0, sizeof(attr));
|
||||
attr.index = i;
|
||||
ret =
|
||||
rknn_query(encoder_ctx_, RKNN_QUERY_INPUT_ATTR, &attr, sizeof(attr));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get attr for encoder input %d", i);
|
||||
i += 1;
|
||||
}
|
||||
|
||||
if (config_.debug) {
|
||||
std::ostringstream os;
|
||||
std::string sep;
|
||||
for (auto &attr : encoder_input_attrs_) {
|
||||
os << sep << ToString(attr);
|
||||
sep = "\n";
|
||||
}
|
||||
SHERPA_ONNX_LOGE("\n----------Encoder inputs info----------\n%s",
|
||||
os.str().c_str());
|
||||
}
|
||||
|
||||
i = 0;
|
||||
for (auto &attr : encoder_output_attrs_) {
|
||||
memset(&attr, 0, sizeof(attr));
|
||||
attr.index = i;
|
||||
ret =
|
||||
rknn_query(encoder_ctx_, RKNN_QUERY_OUTPUT_ATTR, &attr, sizeof(attr));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get attr for encoder output %d",
|
||||
i);
|
||||
i += 1;
|
||||
}
|
||||
|
||||
if (config_.debug) {
|
||||
std::ostringstream os;
|
||||
std::string sep;
|
||||
for (auto &attr : encoder_output_attrs_) {
|
||||
os << sep << ToString(attr);
|
||||
sep = "\n";
|
||||
}
|
||||
SHERPA_ONNX_LOGE("\n----------Encoder outputs info----------\n%s",
|
||||
os.str().c_str());
|
||||
}
|
||||
|
||||
rknn_custom_string custom_string;
|
||||
ret = rknn_query(encoder_ctx_, RKNN_QUERY_CUSTOM_STRING, &custom_string,
|
||||
sizeof(custom_string));
|
||||
SHERPA_ONNX_RKNN_CHECK(
|
||||
ret, "Failed to read custom string from the encoder model");
|
||||
if (config_.debug) {
|
||||
SHERPA_ONNX_LOGE("customs string: %s", custom_string.string);
|
||||
}
|
||||
auto meta = Parse(custom_string);
|
||||
|
||||
for (const auto &p : meta) {
|
||||
SHERPA_ONNX_LOGE("%s: %s", p.first.c_str(), p.second.c_str());
|
||||
}
|
||||
|
||||
if (meta.count("encoder_dims")) {
|
||||
SplitStringToIntegers(meta.at("encoder_dims"), ",", false,
|
||||
&encoder_dims_);
|
||||
}
|
||||
|
||||
if (meta.count("attention_dims")) {
|
||||
SplitStringToIntegers(meta.at("attention_dims"), ",", false,
|
||||
&attention_dims_);
|
||||
}
|
||||
|
||||
if (meta.count("num_encoder_layers")) {
|
||||
SplitStringToIntegers(meta.at("num_encoder_layers"), ",", false,
|
||||
&num_encoder_layers_);
|
||||
}
|
||||
|
||||
if (meta.count("cnn_module_kernels")) {
|
||||
SplitStringToIntegers(meta.at("cnn_module_kernels"), ",", false,
|
||||
&cnn_module_kernels_);
|
||||
}
|
||||
|
||||
if (meta.count("left_context_len")) {
|
||||
SplitStringToIntegers(meta.at("left_context_len"), ",", false,
|
||||
&left_context_len_);
|
||||
}
|
||||
|
||||
if (meta.count("T")) {
|
||||
T_ = atoi(meta.at("T").c_str());
|
||||
}
|
||||
|
||||
if (meta.count("decode_chunk_len")) {
|
||||
decode_chunk_len_ = atoi(meta.at("decode_chunk_len").c_str());
|
||||
}
|
||||
|
||||
if (meta.count("context_size")) {
|
||||
context_size_ = atoi(meta.at("context_size").c_str());
|
||||
}
|
||||
|
||||
if (config_.debug) {
|
||||
auto print = [](const std::vector<int32_t> &v, const char *name) {
|
||||
std::ostringstream os;
|
||||
os << name << ": ";
|
||||
for (auto i : v) {
|
||||
os << i << " ";
|
||||
}
|
||||
#if __OHOS__
|
||||
SHERPA_ONNX_LOGE("%{public}s\n", os.str().c_str());
|
||||
#else
|
||||
SHERPA_ONNX_LOGE("%s\n", os.str().c_str());
|
||||
#endif
|
||||
};
|
||||
print(encoder_dims_, "encoder_dims");
|
||||
print(attention_dims_, "attention_dims");
|
||||
print(num_encoder_layers_, "num_encoder_layers");
|
||||
print(cnn_module_kernels_, "cnn_module_kernels");
|
||||
print(left_context_len_, "left_context_len");
|
||||
#if __OHOS__
|
||||
SHERPA_ONNX_LOGE("T: %{public}d", T_);
|
||||
SHERPA_ONNX_LOGE("decode_chunk_len_: %{public}d", decode_chunk_len_);
|
||||
SHERPA_ONNX_LOGE("context_size: %{public}d", context_size_);
|
||||
#else
|
||||
SHERPA_ONNX_LOGE("T: %d", T_);
|
||||
SHERPA_ONNX_LOGE("decode_chunk_len_: %d", decode_chunk_len_);
|
||||
SHERPA_ONNX_LOGE("context_size: %d", context_size_);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
void InitDecoder(void *model_data, size_t model_data_length) {
|
||||
auto ret =
|
||||
rknn_init(&decoder_ctx_, model_data, model_data_length, 0, nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to init decoder '%s'",
|
||||
config_.transducer.decoder.c_str());
|
||||
|
||||
rknn_input_output_num io_num;
|
||||
ret = rknn_query(decoder_ctx_, RKNN_QUERY_IN_OUT_NUM, &io_num,
|
||||
sizeof(io_num));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret,
|
||||
"Failed to get I/O information for the decoder");
|
||||
|
||||
if (io_num.n_input != 1) {
|
||||
SHERPA_ONNX_LOGE("Expect only 1 decoder input. Given %d",
|
||||
static_cast<int32_t>(io_num.n_input));
|
||||
SHERPA_ONNX_EXIT(-1);
|
||||
}
|
||||
|
||||
if (io_num.n_output != 1) {
|
||||
SHERPA_ONNX_LOGE("Expect only 1 decoder output. Given %d",
|
||||
static_cast<int32_t>(io_num.n_output));
|
||||
SHERPA_ONNX_EXIT(-1);
|
||||
}
|
||||
|
||||
if (config_.debug) {
|
||||
SHERPA_ONNX_LOGE("decoder: %d inputs, %d outputs",
|
||||
static_cast<int32_t>(io_num.n_input),
|
||||
static_cast<int32_t>(io_num.n_output));
|
||||
}
|
||||
|
||||
decoder_input_attrs_.resize(io_num.n_input);
|
||||
decoder_output_attrs_.resize(io_num.n_output);
|
||||
|
||||
int32_t i = 0;
|
||||
for (auto &attr : decoder_input_attrs_) {
|
||||
memset(&attr, 0, sizeof(attr));
|
||||
attr.index = i;
|
||||
ret =
|
||||
rknn_query(decoder_ctx_, RKNN_QUERY_INPUT_ATTR, &attr, sizeof(attr));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get attr for decoder input %d", i);
|
||||
i += 1;
|
||||
}
|
||||
|
||||
if (config_.debug) {
|
||||
std::ostringstream os;
|
||||
std::string sep;
|
||||
for (auto &attr : decoder_input_attrs_) {
|
||||
os << sep << ToString(attr);
|
||||
sep = "\n";
|
||||
}
|
||||
SHERPA_ONNX_LOGE("\n----------Decoder inputs info----------\n%s",
|
||||
os.str().c_str());
|
||||
}
|
||||
|
||||
if (decoder_input_attrs_[0].type != RKNN_TENSOR_INT64) {
|
||||
SHERPA_ONNX_LOGE("Expect int64 for decoder input. Given: %d, %s",
|
||||
decoder_input_attrs_[0].type,
|
||||
get_type_string(decoder_input_attrs_[0].type));
|
||||
SHERPA_ONNX_EXIT(-1);
|
||||
}
|
||||
|
||||
i = 0;
|
||||
for (auto &attr : decoder_output_attrs_) {
|
||||
memset(&attr, 0, sizeof(attr));
|
||||
attr.index = i;
|
||||
ret =
|
||||
rknn_query(decoder_ctx_, RKNN_QUERY_OUTPUT_ATTR, &attr, sizeof(attr));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get attr for decoder output %d",
|
||||
i);
|
||||
i += 1;
|
||||
}
|
||||
|
||||
if (config_.debug) {
|
||||
std::ostringstream os;
|
||||
std::string sep;
|
||||
for (auto &attr : decoder_output_attrs_) {
|
||||
os << sep << ToString(attr);
|
||||
sep = "\n";
|
||||
}
|
||||
SHERPA_ONNX_LOGE("\n----------Decoder outputs info----------\n%s",
|
||||
os.str().c_str());
|
||||
}
|
||||
}
|
||||
|
||||
void InitJoiner(void *model_data, size_t model_data_length) {
|
||||
auto ret =
|
||||
rknn_init(&joiner_ctx_, model_data, model_data_length, 0, nullptr);
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to init joiner '%s'",
|
||||
config_.transducer.joiner.c_str());
|
||||
|
||||
rknn_input_output_num io_num;
|
||||
ret =
|
||||
rknn_query(joiner_ctx_, RKNN_QUERY_IN_OUT_NUM, &io_num, sizeof(io_num));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get I/O information for the joiner");
|
||||
|
||||
if (config_.debug) {
|
||||
SHERPA_ONNX_LOGE("joiner: %d inputs, %d outputs",
|
||||
static_cast<int32_t>(io_num.n_input),
|
||||
static_cast<int32_t>(io_num.n_output));
|
||||
}
|
||||
|
||||
joiner_input_attrs_.resize(io_num.n_input);
|
||||
joiner_output_attrs_.resize(io_num.n_output);
|
||||
|
||||
int32_t i = 0;
|
||||
for (auto &attr : joiner_input_attrs_) {
|
||||
memset(&attr, 0, sizeof(attr));
|
||||
attr.index = i;
|
||||
ret = rknn_query(joiner_ctx_, RKNN_QUERY_INPUT_ATTR, &attr, sizeof(attr));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get attr for joiner input %d", i);
|
||||
i += 1;
|
||||
}
|
||||
|
||||
if (config_.debug) {
|
||||
std::ostringstream os;
|
||||
std::string sep;
|
||||
for (auto &attr : joiner_input_attrs_) {
|
||||
os << sep << ToString(attr);
|
||||
sep = "\n";
|
||||
}
|
||||
SHERPA_ONNX_LOGE("\n----------Joiner inputs info----------\n%s",
|
||||
os.str().c_str());
|
||||
}
|
||||
|
||||
i = 0;
|
||||
for (auto &attr : joiner_output_attrs_) {
|
||||
memset(&attr, 0, sizeof(attr));
|
||||
attr.index = i;
|
||||
ret =
|
||||
rknn_query(joiner_ctx_, RKNN_QUERY_OUTPUT_ATTR, &attr, sizeof(attr));
|
||||
SHERPA_ONNX_RKNN_CHECK(ret, "Failed to get attr for joiner output %d", i);
|
||||
i += 1;
|
||||
}
|
||||
|
||||
if (config_.debug) {
|
||||
std::ostringstream os;
|
||||
std::string sep;
|
||||
for (auto &attr : joiner_output_attrs_) {
|
||||
os << sep << ToString(attr);
|
||||
sep = "\n";
|
||||
}
|
||||
SHERPA_ONNX_LOGE("\n----------Joiner outputs info----------\n%s",
|
||||
os.str().c_str());
|
||||
}
|
||||
|
||||
vocab_size_ = joiner_output_attrs_[0].dims[1];
|
||||
if (config_.debug) {
|
||||
SHERPA_ONNX_LOGE("vocab_size: %d", vocab_size_);
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
OnlineModelConfig config_;
|
||||
rknn_context encoder_ctx_ = 0;
|
||||
rknn_context decoder_ctx_ = 0;
|
||||
rknn_context joiner_ctx_ = 0;
|
||||
|
||||
std::vector<rknn_tensor_attr> encoder_input_attrs_;
|
||||
std::vector<rknn_tensor_attr> encoder_output_attrs_;
|
||||
|
||||
std::vector<rknn_tensor_attr> decoder_input_attrs_;
|
||||
std::vector<rknn_tensor_attr> decoder_output_attrs_;
|
||||
|
||||
std::vector<rknn_tensor_attr> joiner_input_attrs_;
|
||||
std::vector<rknn_tensor_attr> joiner_output_attrs_;
|
||||
|
||||
std::vector<int32_t> encoder_dims_;
|
||||
std::vector<int32_t> attention_dims_;
|
||||
std::vector<int32_t> num_encoder_layers_;
|
||||
std::vector<int32_t> cnn_module_kernels_;
|
||||
std::vector<int32_t> left_context_len_;
|
||||
|
||||
int32_t T_ = 0;
|
||||
int32_t decode_chunk_len_ = 0;
|
||||
|
||||
int32_t context_size_ = 2;
|
||||
int32_t vocab_size_ = 0;
|
||||
};
|
||||
|
||||
OnlineZipformerTransducerModelRknn::~OnlineZipformerTransducerModelRknn() =
|
||||
default;
|
||||
|
||||
OnlineZipformerTransducerModelRknn::OnlineZipformerTransducerModelRknn(
|
||||
const OnlineModelConfig &config)
|
||||
: impl_(std::make_unique<Impl>(config)) {}
|
||||
|
||||
template <typename Manager>
|
||||
OnlineZipformerTransducerModelRknn::OnlineZipformerTransducerModelRknn(
|
||||
Manager *mgr, const OnlineModelConfig &config)
|
||||
: impl_(std::make_unique<OnlineZipformerTransducerModelRknn>(mgr, config)) {
|
||||
}
|
||||
|
||||
std::vector<std::vector<uint8_t>>
|
||||
OnlineZipformerTransducerModelRknn::GetEncoderInitStates() const {
|
||||
return impl_->GetEncoderInitStates();
|
||||
}
|
||||
|
||||
std::pair<std::vector<float>, std::vector<std::vector<uint8_t>>>
|
||||
OnlineZipformerTransducerModelRknn::RunEncoder(
|
||||
std::vector<float> features,
|
||||
std::vector<std::vector<uint8_t>> states) const {
|
||||
return impl_->RunEncoder(std::move(features), std::move(states));
|
||||
}
|
||||
|
||||
std::vector<float> OnlineZipformerTransducerModelRknn::RunDecoder(
|
||||
std::vector<int64_t> decoder_input) const {
|
||||
return impl_->RunDecoder(std::move(decoder_input));
|
||||
}
|
||||
|
||||
std::vector<float> OnlineZipformerTransducerModelRknn::RunJoiner(
|
||||
const float *encoder_out, const float *decoder_out) const {
|
||||
return impl_->RunJoiner(encoder_out, decoder_out);
|
||||
}
|
||||
|
||||
int32_t OnlineZipformerTransducerModelRknn::ContextSize() const {
|
||||
return impl_->ContextSize();
|
||||
}
|
||||
|
||||
int32_t OnlineZipformerTransducerModelRknn::ChunkSize() const {
|
||||
return impl_->ChunkSize();
|
||||
}
|
||||
|
||||
int32_t OnlineZipformerTransducerModelRknn::ChunkShift() const {
|
||||
return impl_->ChunkShift();
|
||||
}
|
||||
|
||||
int32_t OnlineZipformerTransducerModelRknn::VocabSize() const {
|
||||
return impl_->VocabSize();
|
||||
}
|
||||
|
||||
rknn_tensor_attr OnlineZipformerTransducerModelRknn::GetEncoderOutAttr() const {
|
||||
return impl_->GetEncoderOutAttr();
|
||||
}
|
||||
|
||||
#if __ANDROID_API__ >= 9
|
||||
template OnlineZipformerTransducerModelRknn::OnlineZipformerTransducerModelRknn(
|
||||
AAssetManager *mgr, const OnlineModelConfig &config);
|
||||
#endif
|
||||
|
||||
#if __OHOS__
|
||||
template OnlineZipformerTransducerModelRknn::OnlineZipformerTransducerModelRknn(
|
||||
NativeResourceManager *mgr, const OnlineModelConfig &config);
|
||||
#endif
|
||||
|
||||
} // namespace sherpa_onnx
|
||||
@@ -0,0 +1,57 @@
|
||||
// sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.h
|
||||
//
|
||||
// Copyright (c) 2025 Xiaomi Corporation
|
||||
#ifndef SHERPA_ONNX_CSRC_RKNN_ONLINE_ZIPFORMER_TRANSDUCER_MODEL_RKNN_H_
|
||||
#define SHERPA_ONNX_CSRC_RKNN_ONLINE_ZIPFORMER_TRANSDUCER_MODEL_RKNN_H_
|
||||
|
||||
#include <memory>
|
||||
#include <utility>
|
||||
#include <vector>
|
||||
|
||||
#include "rknn_api.h" // NOLINT
|
||||
#include "sherpa-onnx/csrc/online-model-config.h"
|
||||
#include "sherpa-onnx/csrc/online-transducer-model.h"
|
||||
|
||||
namespace sherpa_onnx {
|
||||
|
||||
// this is for zipformer v1, i.e., the folder
|
||||
// pruned_transducer_statelss7_streaming from icefall
|
||||
class OnlineZipformerTransducerModelRknn {
|
||||
public:
|
||||
~OnlineZipformerTransducerModelRknn();
|
||||
|
||||
explicit OnlineZipformerTransducerModelRknn(const OnlineModelConfig &config);
|
||||
|
||||
template <typename Manager>
|
||||
OnlineZipformerTransducerModelRknn(Manager *mgr,
|
||||
const OnlineModelConfig &config);
|
||||
|
||||
std::vector<std::vector<uint8_t>> GetEncoderInitStates() const;
|
||||
|
||||
std::pair<std::vector<float>, std::vector<std::vector<uint8_t>>> RunEncoder(
|
||||
std::vector<float> features,
|
||||
std::vector<std::vector<uint8_t>> states) const;
|
||||
|
||||
std::vector<float> RunDecoder(std::vector<int64_t> decoder_input) const;
|
||||
|
||||
std::vector<float> RunJoiner(const float *encoder_out,
|
||||
const float *decoder_out) const;
|
||||
|
||||
int32_t ContextSize() const;
|
||||
|
||||
int32_t ChunkSize() const;
|
||||
|
||||
int32_t ChunkShift() const;
|
||||
|
||||
int32_t VocabSize() const;
|
||||
|
||||
rknn_tensor_attr GetEncoderOutAttr() const;
|
||||
|
||||
private:
|
||||
class Impl;
|
||||
std::unique_ptr<Impl> impl_;
|
||||
};
|
||||
|
||||
} // namespace sherpa_onnx
|
||||
|
||||
#endif // SHERPA_ONNX_CSRC_RKNN_ONLINE_ZIPFORMER_TRANSDUCER_MODEL_RKNN_H_
|
||||
Reference in New Issue
Block a user