493 lines
15 KiB
C++
493 lines
15 KiB
C++
// sherpa-onnx/csrc/rknn/online-zipformer-transducer-model-rknn.cc
|
|
//
|
|
// Copyright (c) 2025 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/rknn/utils.h"
|
|
#include "sherpa-onnx/csrc/text-utils.h"
|
|
|
|
namespace sherpa_onnx {
|
|
|
|
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());
|
|
}
|
|
|
|
SetCoreMask(encoder_ctx_, config_.num_threads);
|
|
SetCoreMask(decoder_ctx_, config_.num_threads);
|
|
SetCoreMask(joiner_ctx_, config_.num_threads);
|
|
}
|
|
|
|
template <typename Manager>
|
|
Impl(Manager *mgr, const OnlineModelConfig &config) : config_(config) {
|
|
{
|
|
auto buf = ReadFile(mgr, config.transducer.encoder);
|
|
InitEncoder(buf.data(), buf.size());
|
|
}
|
|
|
|
{
|
|
auto buf = ReadFile(mgr, config.transducer.decoder);
|
|
InitDecoder(buf.data(), buf.size());
|
|
}
|
|
|
|
{
|
|
auto buf = ReadFile(mgr, config.transducer.joiner);
|
|
InitJoiner(buf.data(), buf.size());
|
|
}
|
|
|
|
SetCoreMask(encoder_ctx_, config_.num_threads);
|
|
SetCoreMask(decoder_ctx_, config_.num_threads);
|
|
SetCoreMask(joiner_ctx_, config_.num_threads);
|
|
}
|
|
|
|
// 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 ConvertNCHWtoNHWC 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];
|
|
ConvertNCHWtoNHWC(
|
|
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) {
|
|
InitContext(model_data, model_data_length, config_.debug, &encoder_ctx_);
|
|
|
|
InitInputOutputAttrs(encoder_ctx_, config_.debug, &encoder_input_attrs_,
|
|
&encoder_output_attrs_);
|
|
|
|
rknn_custom_string custom_string =
|
|
GetCustomString(encoder_ctx_, config_.debug);
|
|
|
|
auto meta = Parse(custom_string, config_.debug);
|
|
|
|
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_);
|
|
#else
|
|
SHERPA_ONNX_LOGE("T: %d", T_);
|
|
SHERPA_ONNX_LOGE("decode_chunk_len_: %d", decode_chunk_len_);
|
|
#endif
|
|
}
|
|
}
|
|
|
|
void InitDecoder(void *model_data, size_t model_data_length) {
|
|
InitContext(model_data, model_data_length, config_.debug, &decoder_ctx_);
|
|
|
|
InitInputOutputAttrs(decoder_ctx_, config_.debug, &decoder_input_attrs_,
|
|
&decoder_output_attrs_);
|
|
|
|
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);
|
|
}
|
|
|
|
context_size_ = decoder_input_attrs_[0].dims[1];
|
|
if (config_.debug) {
|
|
SHERPA_ONNX_LOGE("context_size: %d", context_size_);
|
|
}
|
|
}
|
|
|
|
void InitJoiner(void *model_data, size_t model_data_length) {
|
|
InitContext(model_data, model_data_length, config_.debug, &joiner_ctx_);
|
|
|
|
InitInputOutputAttrs(joiner_ctx_, config_.debug, &joiner_input_attrs_,
|
|
&joiner_output_attrs_);
|
|
|
|
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<Impl>(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
|