cleanup: remove legacy cpp sources (now on GitHub)

cpp/.gitattributes

@@ -1,2 +0,0 @@
-*.axmodel filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text

cpp/CMakeLists.txt

@@ -1,53 +0,0 @@
-cmake_minimum_required(VERSION 3.13 FATAL_ERROR)
-project(zipvoice)
-
-if (NOT DEFINED CMAKE_CXX_STANDARD)
-    set(CMAKE_CXX_STANDARD 14)
-endif()
-
-if(CMAKE_BUILD_TYPE MATCHES Debug)
-    set(CMAKE_CXX_FLAGS "-fvisibility=hidden -g -O0")
-elseif(CMAKE_BUILD_TYPE MATCHES Release)
-    set(CMAKE_CXX_FLAGS "-fvisibility=hidden -O2 -fdata-sections -ffunction-sections")
-endif()
-
-# Platform dependencies (AXERA or AXCL)
-include(cmake/msp_dependencies.cmake)
-
-include_directories(${MSP_INC_DIR})
-
-# Project includes
-include_directories(${CMAKE_SOURCE_DIR})
-include_directories(${CMAKE_SOURCE_DIR}/src/)
-
-# Core source files (common to AXERA and AXCL)
-set(SRC
-    src/EngineWrapper.cpp
-    src/tokenizer.cpp
-    src/fbank.cpp
-    src/zipvoice_engine.cpp
-    src/vocoder.cpp
-    third_party/kissfft/kiss_fft.c
-)
-
-# AXCL-only source files (middleware + utilities)
-if(AXCL)
-    aux_source_directory(${CMAKE_SOURCE_DIR}/src/middleware AXCL_SRCS)
-    aux_source_directory(${CMAKE_SOURCE_DIR}/src/utilities  AXCL_SRCS)
-    list(APPEND SRC ${AXCL_SRCS})
-    message(STATUS "AXCL middleware: ${AXCL_SRCS}")
-endif()
-
-add_executable(${PROJECT_NAME} zipvoice.cpp ${SRC})
-target_include_directories(${PROJECT_NAME} PRIVATE third_party/kissfft)
-target_link_libraries(${PROJECT_NAME} ${MSP_LIBS})
-
-install(TARGETS ${PROJECT_NAME}
-        RUNTIME DESTINATION ./)
-set_target_properties(${PROJECT_NAME}
-    PROPERTIES
-    INSTALL_RPATH "$ORIGIN/")
-
-message(STATUS "========================================")
-message(STATUS "ZipVoice C++ Build Configuration")
-message(STATUS "========================================")

cpp/README.md

@@ -1,90 +0,0 @@
-# ZipVoice C++ — AXERA & AXCL
-
-## 性能
-
-### Distill 模型 (num_step=4)
-
-| 场景 | 音频 | Python AX650<br>耗时 RTF | C++ AX650<br>耗时 RTF | C++ AXCL<br>耗时 RTF | Python AX630C<br>耗时 RTF | C++ AX630C<br>耗时 RTF |
-|------|------|-------------------------|----------------------|---------------------|--------------------------|-----------------------|
-| 中文句子 | 6.41s | 1.992s/0.311 | 1.057s/0.165 | 0.918s/**0.143** | 10.296s/1.606 | 4.650s/0.725 |
-| 中文段落 | 44.97s | 13.457s/0.301 | 7.357s/0.164 | 6.443s/**0.143** | 71.574s/1.600 | 32.590s/0.725 |
-| 英文句子 | 6.41s | 2.045s/0.319 | 1.067s/0.166 | 0.919s/**0.143** | 10.686s/1.667 | 4.654s/0.726 |
-| 英文段落 | 59.16s | 19.715s/0.305 | 10.529s/0.178 | 9.207s/**0.156** | 106.183s/1.640 | 46.541s/0.787 |
-
-### 普通模型 (num_step=10)
-
-| 场景 | 音频时长 | Python AX650 (耗时/RTF) | C++ AX650 (耗时/RTF) | C++ AXCL (耗时/RTF) |
-|------|---------|------------------------|---------------------|---------------------|
-| 中文句子 | 6.41s | 5.781s / 0.902 | 5.709s / 0.891 | 5.452s / **0.850** |
-| 中文段落 | 44.97s | 40.292s / 0.901 | 39.714s / 0.883 | 38.138s / **0.848** |
-| 英文句子 | 6.41s | 5.711s / 0.891 | 5.693s / 0.888 | 5.447s / **0.850** |
-| 英文段落 | 59.16s | 62.161s / 0.960 | 56.759s / 0.957 | 54.451s / **0.920** |
-
-## 环境准备
-
-下载交叉编译器、AXERA SDK、AXCL SDK 到指定目录（记为 `<TOOLCHAIN_DIR>`）：
-
-```bash
-bash download_bsp.sh   
-
-# 或手动:
-mkdir -p <TOOLCHAIN_DIR> && cd <TOOLCHAIN_DIR>
-wget https://developer.arm.com/-/media/Files/downloads/gnu-a/9.2-2019.12/binrel/gcc-arm-9.2-2019.12-x86_64-aarch64-none-linux-gnu.tar.xz
-tar -xf gcc-arm-9.2-2019.12-x86_64-aarch64-none-linux-gnu.tar.xz
-git clone https://github.com/AXERA-TECH/ax650n_bsp_sdk.git --depth=1
-git clone https://github.com/AXERA-TECH/ax620e_bsp_sdk.git --depth=1
-git clone https://github.com/Abandon-ht/axcl_bsp_sdk.git --depth=1
-```
-
-修改 `cmake/msp_dependencies.cmake` 和 `toolchains/aarch64-none-linux-gnu.toolchain.cmake` 中的路径。
-
-英文分词需 Python 依赖（运行设备上安装）：
-```bash
-pip install jieba numpy pypinyin piper_phonemize
-```
-
-## 编译
-
-```bash
-bash build_ax650.sh    # → install/ax650/zipvoice
-bash build_ax630c.sh   # → install/ax630c/zipvoice
-bash build_axcl.sh     # → install/axcl/zipvoice
-```
-
-## 运行
-
-在 repo 根目录（`ZipVoice.AXERA/`）执行：
-
-```bash
-# 板端
-bash run_ax650.sh distill zh sentence
-bash run_ax650.sh distill zh paragraph
-bash run_ax650.sh standard zh sentence
-bash run_ax650.sh standard zh paragraph
-bash run_ax650.sh distill en sentence
-bash run_ax650.sh distill en paragraph
-bash run_ax650.sh standard en sentence
-bash run_ax650.sh standard en paragraph
-
-# 算力卡
-bash run_axcl.sh distill zh paragraph
-bash run_axcl.sh standard en sentence
-
-# AX630C (仅 distill)
-bash run_ax630c.sh zh sentence
-bash run_ax630c.sh en paragraph
-...
-```
-
-## 参数
-
-| 参数 | 说明 | 默认值 |
-|------|------|--------|
-| `--model-dir` | 模型目录 | 必填 |
-| `--prompt-wav` | 提示音频 | 必填 |
-| `--prompt-text` | 提示文本 | 必填 |
-| `--text` / `--text-file` | 合成文本 | 二选一 |
-| `--vocoder-model` | vocos_full.axmodel 路径 | 必填 |
-| `--output-wav` | 输出 WAV | output.wav |
-| `--num-step` | 采样步数 | 10 (distill: 4) |
-| `--seed` | 随机种子 | 42 |

cpp/build_ax650.sh

@@ -1,30 +0,0 @@
-#!/bin/bash
-# Cross-compile ZipVoice for AXERA (AX650 demo board)
-#
-# Prerequisites:
-#   - AX650 BSP SDK: cpp/ax650n_bsp_sdk/  (run download_bsp.sh first)
-#   - Cross compiler: /data/shared/huyuan/toolchains/gcc-arm-9.2-2019.12-x86_64-aarch64-none-linux-gnu/
-
-set -e
-
-SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
-cd "$SCRIPT_DIR"
-
-rm -rf build_ax650
-mkdir -p build_ax650 && cd build_ax650
-
-cmake .. \
-  -DAX650=ON \
-  -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-none-linux-gnu.toolchain.cmake \
-  -DCMAKE_INSTALL_PREFIX=../install/ax650 \
-  -DCMAKE_BUILD_TYPE=Release \
-  $@
-
-make -j$(nproc)
-make install
-
-echo ""
-echo "========================================"
-echo "Build complete!"
-echo "Executable: $(pwd)/../install/ax650/zipvoice"
-echo "========================================"

cpp/build_axcl.sh

@@ -1,34 +0,0 @@
-#!/bin/bash
-# Cross-compile ZipVoice for AXCL (AXCL compute card)
-#
-# Prerequisites:
-#   - AXCL BSP SDK: /data/shared/huyuan/toolchains/axcl_bsp_sdk/
-#   - Cross compiler: /data/shared/huyuan/toolchains/gcc-arm-9.2-2019.12-x86_64-aarch64-none-linux-gnu/
-#
-# Clone SDK (one-time):
-#   cd /data/shared/huyuan/toolchains
-#   git clone https://github.com/Abandon-ht/axcl_bsp_sdk.git --depth=1
-
-set -e
-
-SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
-cd "$SCRIPT_DIR"
-
-rm -rf build_axcl
-mkdir -p build_axcl && cd build_axcl
-
-cmake .. \
-  -DAXCL=ON \
-  -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-none-linux-gnu.toolchain.cmake \
-  -DCMAKE_INSTALL_PREFIX=../install/axcl \
-  -DCMAKE_BUILD_TYPE=Release \
-  $@
-
-make -j$(nproc)
-make install
-
-echo ""
-echo "========================================"
-echo "Build complete!"
-echo "Executable: $(pwd)/../install/axcl/zipvoice"
-echo "========================================"

cpp/cmake/msp_dependencies.cmake

@@ -1,68 +0,0 @@
-# BSP / SDK Dependencies Configuration
-# Supports both AXERA (demo board) and AXCL (compute card)
-#
-# AXERA: AX650 / AX630C / AX620Q
-# AXCL:  AXCL
-
-if (AX650)
-    add_definitions(-DAX650)
-    if(NOT BSP_MSP_DIR)
-        set(BSP_MSP_DIR /data/shared/huyuan/toolchains/ax650n_bsp_sdk/msp/out)
-    endif()
-    if(NOT EXISTS ${BSP_MSP_DIR})
-        message(FATAL_ERROR "BSP_MSP_DIR ${BSP_MSP_DIR} not exist. Run download_bsp.sh first.")
-    endif()
-    set(MSP_INC_DIR ${BSP_MSP_DIR}/include)
-    set(MSP_LIB_DIR ${BSP_MSP_DIR}/lib)
-    list(APPEND MSP_LIBS ax_sys ax_engine ax_interpreter)
-
-elseif(AX630C)
-    add_definitions(-DAX630C)
-    if(NOT BSP_MSP_DIR)
-        set(BSP_MSP_DIR /data/shared/huyuan/toolchains/ax620e_bsp_sdk/msp/out/arm64_glibc)
-    endif()
-    set(MSP_INC_DIR ${BSP_MSP_DIR}/include)
-    set(MSP_LIB_DIR ${BSP_MSP_DIR}/lib)
-    list(APPEND MSP_LIBS ax_sys ax_engine ax_interpreter)
-
-elseif(AX620Q)
-    add_definitions(-DAX620Q)
-    if(NOT BSP_MSP_DIR)
-        set(BSP_MSP_DIR ${CMAKE_SOURCE_DIR}/ax620e_bsp_sdk/msp/out/arm_uclibc)
-    endif()
-    if(NOT EXISTS ${BSP_MSP_DIR})
-        message(FATAL_ERROR "BSP_MSP_DIR ${BSP_MSP_DIR} not exist.")
-    endif()
-    set(MSP_INC_DIR ${BSP_MSP_DIR}/include)
-    set(MSP_LIB_DIR ${BSP_MSP_DIR}/lib)
-    list(APPEND MSP_LIBS ax_sys ax_engine ax_interpreter)
-
-elseif(AXCL)
-    add_definitions(-DAXCL)
-    add_definitions(-DENV_AXCL_RUNTIME_API_ENABLE)
-    add_definitions(-DENV_AXCL_NATIVE_API_ENABLE)
-    add_definitions(-DENV_HAS_STD_FILESYSTEM)
-    add_definitions(-DENV_HAS_POSIX_FILE_STAT)
-
-    # AXCL SDK at /data/shared/huyuan/toolchains/axcl_bsp_sdk/out/
-    if(NOT AXCL_SDK_DIR)
-        set(AXCL_SDK_DIR /data/shared/huyuan/toolchains/axcl_bsp_sdk/out)
-    endif()
-    if(NOT EXISTS ${AXCL_SDK_DIR})
-        message(FATAL_ERROR "AXCL_SDK_DIR ${AXCL_SDK_DIR} not exist. Run download_bsp.sh first.")
-    endif()
-    set(MSP_INC_DIR ${AXCL_SDK_DIR}/include ${AXCL_SDK_DIR}/bsp)
-    set(MSP_LIB_DIR ${AXCL_SDK_DIR}/lib)
-    list(APPEND MSP_LIBS
-        axcl_rt axcl_pkg axcl_comm axcl_npu spdlog
-        axcl_token axcl_pcie_msg axcl_pcie_dma)
-    set(CMAKE_CXX_STANDARD 17)
-
-else()
-    message(FATAL_ERROR "Unknown platform. Set -DAX650, -DAX630C, -DAX620Q, or -DAXCL.")
-endif()
-
-link_directories(${MSP_LIB_DIR})
-message(STATUS "MSP_INC_DIR: ${MSP_INC_DIR}")
-message(STATUS "MSP_LIB_DIR: ${MSP_LIB_DIR}")
-message(STATUS "MSP_LIBS:    ${MSP_LIBS}")

cpp/download_bsp.sh

@@ -1,23 +0,0 @@
-#!/bin/bash
-# Download BSP SDKs to /data/shared/huyuan/toolchains/
-
-TARGET=/data/shared/huyuan/toolchains
-mkdir -p $TARGET && cd $TARGET
-
-if [ ! -d ax650n_bsp_sdk ]; then
-  echo "Downloading ax650n_bsp_sdk..."
-  git clone https://github.com/AXERA-TECH/ax650n_bsp_sdk.git --depth=1
-fi
-
-if [ ! -d ax620e_bsp_sdk ]; then
-  echo "Downloading ax620e_bsp_sdk..."
-  git clone https://github.com/AXERA-TECH/ax620e_bsp_sdk.git --depth=1
-fi
-
-if [ ! -d axcl_bsp_sdk ]; then
-  echo "Downloading axcl_bsp_sdk..."
-  git clone https://github.com/Abandon-ht/axcl_bsp_sdk.git --depth=1
-fi
-
-echo "Done. SDKs at $TARGET/"
-ls -d ax650n_bsp_sdk ax620e_bsp_sdk axcl_bsp_sdk 2>/dev/null

cpp/install/ax650/zipvoice

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:bb407d306e74de864ee7966937552da945db72207ccdea7c7b17950805642978
-size 582632

cpp/install/ax650/zipvoice_axera

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0d15168590f21a4255f223055008fb6832babf8a1f096c39f93afe0d78fb4c98
-size 582624

cpp/install/axcl/zipvoice

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f70b9a838f6505d99d2c1cc3131c89565aa25550109336387fbc8a3d49205165
-size 658128

cpp/scripts/compare_vocoder.py

@@ -1,277 +0,0 @@
-#!/usr/bin/env python3
-"""
-Compare PyTorch vocoder vs quantized axmodel output.
-Saves intermediate tensors for board-side comparison.
-
-Usage:
-  # Dev machine: generate test data
-  python3 compare_vocoder.py --save-test-data
-
-  # Board: run axmodel on same input, save output
-  python3 compare_vocoder.py --run-axmodel
-
-  # Dev machine: compare results
-  python3 compare_vocoder.py --compare
-"""
-import sys, os, math, argparse
-import numpy as np
-
-REPO_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-ONNX_DIR = os.path.join(REPO_DIR, 'cpp', 'vocoder_onnx')
-TEST_DIR = os.path.join(ONNX_DIR, 'test_data')
-os.makedirs(TEST_DIR, exist_ok=True)
-
-FEAT_SCALE = 0.1
-N_FFT = 1024
-HOP = 256
-
-
-def irfft_overlap_add(real_spec, imag_spec):
-    """C++ IRFFT + overlap-add, verified to match PyTorch istft (cos_sim=0.999999)."""
-    n_freqs = N_FFT // 2 + 1
-    T = real_spec.shape[1]
-    window = 0.5 * (1.0 - np.cos(2.0 * math.pi * np.arange(N_FFT) / (N_FFT - 1)))
-    window_sq = window ** 2
-
-    # Build IRFFT basis
-    irfft_cos = np.zeros((N_FFT, n_freqs), dtype=np.float32)
-    irfft_sin = np.zeros((N_FFT, n_freqs), dtype=np.float32)
-    for n in range(N_FFT):
-        for k in range(n_freqs):
-            ang = 2.0 * math.pi * k * n / N_FFT
-            if k == 0:
-                irfft_cos[n, k] = 1.0 / N_FFT
-                irfft_sin[n, k] = 0.0
-            elif k == n_freqs - 1:
-                irfft_cos[n, k] = math.cos(ang) / N_FFT
-                irfft_sin[n, k] = 0.0
-            else:
-                irfft_cos[n, k] = math.cos(ang) * (2.0 / N_FFT)
-                irfft_sin[n, k] = math.sin(ang) * (2.0 / N_FFT)
-
-    out_len = (T - 1) * HOP + N_FFT
-    audio = np.zeros(out_len, dtype=np.float32)
-    envelope = np.zeros(out_len, dtype=np.float32)
-
-    for t in range(T):
-        r = real_spec[0, t, :]
-        im = imag_spec[0, t, :]
-        frame = irfft_cos @ r - irfft_sin @ im
-        pos = t * HOP
-        for n in range(N_FFT):
-            p = pos + n
-            if p < out_len:
-                audio[p] += frame[n] * window[n]
-                envelope[p] += window_sq[n]
-
-    audio /= np.maximum(envelope, 1e-10)
-    pad = N_FFT // 2
-    return audio[pad:pad + (T - 1) * HOP]
-
-
-def compare(name, ref, test):
-    ref = np.asarray(ref, dtype=np.float32).flatten()
-    test = np.asarray(test, dtype=np.float32).flatten()
-    if len(ref) != len(test):
-        print(f"  [{name}] SIZE MISMATCH: ref={ref.shape} test={test.shape}")
-        return
-    diff = np.abs(ref - test)
-    sig = np.mean(np.abs(ref)) + 1e-10
-    cos = np.dot(ref, test) / (np.linalg.norm(ref) * np.linalg.norm(test) + 1e-10)
-    print(f"  [{name}] max_err={diff.max():.2e}  rel_err={diff.mean()/sig:.2e}  cos_sim={cos:.6f}")
-
-
-def cmd_save_test_data():
-    """Generate test mel and save PT/ONNX reference outputs."""
-    import torch
-    sys.path.insert(0, REPO_DIR)
-    from scripts.local_vocos import LocalVocos
-    import onnxruntime as ort
-
-    # Load models
-    vocoder = LocalVocos()
-    sd = torch.load(f'{REPO_DIR}/resources/vocos-mel-24khz/pytorch_model.bin',
-                    weights_only=True, map_location='cpu')
-    sd = {k: v for k, v in sd.items() if k.startswith(('backbone.', 'head.'))}
-    vocoder.load_state_dict(sd)
-    vocoder.eval()
-
-    sess_f = ort.InferenceSession(f'{ONNX_DIR}/vocos_full_B1_T620.onnx')
-
-    # Load real mel or generate random
-    mel_bin = os.path.join(REPO_DIR, 'cpp', 'output_mel.bin')
-    if os.path.exists(mel_bin):
-        real_mel = np.fromfile(mel_bin, dtype=np.float32).reshape(-1, 100)  # [T, 100]
-        T = real_mel.shape[0]
-        # Undo feat_scale, transpose to [1, 100, T]
-        mel_input = (real_mel / FEAT_SCALE).T[np.newaxis, :, :].astype(np.float32)  # [1, 100, T]
-        print(f"Using real mel: shape={mel_input.shape}, range=[{mel_input.min():.3f}, {mel_input.max():.3f}]")
-    else:
-        T = 200
-        mel_input = np.random.RandomState(42).randn(1, 100, T).astype(np.float32) * 3.0
-        print(f"Using random mel: shape={mel_input.shape}")
-
-    # Pad to 620 for ONNX
-    T_pad = 620
-    mel_onnx = np.zeros((1, 100, T_pad), dtype=np.float32)
-    mel_onnx[:, :, :T] = mel_input[:, :, :T]
-
-    # PT inference
-    mel_pt = torch.from_numpy(mel_input)
-    with torch.no_grad():
-        features_pt = vocoder.backbone(mel_pt)
-        audio_pt = vocoder.head(features_pt).squeeze().numpy()
-        h = vocoder.head.out(features_pt)
-        mag, phase = h.chunk(2, dim=-1)
-        mag = torch.exp(mag).clamp(max=1e2)
-        real_pt = (mag * torch.cos(phase)).numpy()
-        imag_pt = (mag * torch.sin(phase)).numpy()
-
-    # ONNX inference
-    onnx_out = sess_f.run(None, {'mel': mel_onnx})
-    real_onnx = onnx_out[0][:, :T, :]
-    imag_onnx = onnx_out[1][:, :T, :]
-    audio_onnx = irfft_overlap_add(real_onnx, imag_onnx)
-
-    # Save everything
-    np.save(f'{TEST_DIR}/mel_input.npy', mel_input)
-    np.save(f'{TEST_DIR}/mel_onnx_padded.npy', mel_onnx)
-    np.save(f'{TEST_DIR}/pt_real.npy', real_pt)
-    np.save(f'{TEST_DIR}/pt_imag.npy', imag_pt)
-    np.save(f'{TEST_DIR}/pt_audio.npy', audio_pt)
-    np.save(f'{TEST_DIR}/onnx_real.npy', real_onnx)
-    np.save(f'{TEST_DIR}/onnx_imag.npy', imag_onnx)
-    np.save(f'{TEST_DIR}/onnx_audio.npy', audio_onnx)
-    np.save(f'{TEST_DIR}/T_frames.npy', np.array([T], dtype=np.int32))
-
-    # Meta
-    with open(f'{TEST_DIR}/info.txt', 'w') as f:
-        f.write(f"T={T}\n")
-        f.write(f"feat_scale={FEAT_SCALE}\n")
-        f.write(f"mel_range=[{mel_input.min():.4f}, {mel_input.max():.4f}]\n")
-        f.write(f"pt_real_range=[{real_pt.min():.4f}, {real_pt.max():.4f}]\n")
-        f.write(f"pt_audio_len={len(audio_pt)}\n")
-
-    # Verify PT vs ONNX
-    print("\n=== PT vs ONNX (dev machine) ===")
-    compare('real_spectrum', real_pt, real_onnx)
-    compare('imag_spectrum', imag_pt, imag_onnx)
-    compare('audio', audio_pt, audio_onnx)
-
-    # Also write audio files for listening
-    import soundfile as sf
-    sf.write(f'{TEST_DIR}/pt_audio.wav', audio_pt, 24000)
-    sf.write(f'{TEST_DIR}/onnx_audio.wav', audio_onnx, 24000)
-
-    print(f"\nTest data saved to {TEST_DIR}/")
-    print("Copy to board and run: python3 compare_vocoder.py --run-axmodel")
-
-
-def cmd_run_axmodel():
-    """Run axmodel on board with the same test mel, save output."""
-    import onnxruntime as ort
-    from axengine import InferenceSession
-
-    T = int(np.load(f'{TEST_DIR}/T_frames.npy')[0])
-    mel_onnx = np.load(f'{TEST_DIR}/mel_onnx_padded.npy')
-
-    # Run axmodel
-    model_path = f'{ONNX_DIR}/axmodel/vocos_full.axmodel'
-    if not os.path.exists(model_path):
-        print(f"ERROR: {model_path} not found")
-        return
-
-    print(f"Loading axmodel: {model_path}")
-    sess = InferenceSession(model_path)
-
-    print(f"Running inference (mel shape={mel_onnx.shape})...")
-    outputs = sess.run(None, {'mel': mel_onnx})
-    print(f"Output keys: {list(outputs.keys()) if isinstance(outputs, dict) else type(outputs)}")
-
-    # Extract real/imag
-    if isinstance(outputs, dict):
-        real_ax = outputs['real'][:, :T, :]
-        imag_ax = outputs['imag'][:, :T, :]
-    elif isinstance(outputs, (list, tuple)):
-        real_ax = outputs[0][:, :T, :]
-        imag_ax = outputs[1][:, :T, :]
-    else:
-        real_ax = outputs[:, :T, :]  # guess
-        imag_ax = None
-
-    audio_ax = irfft_overlap_add(real_ax, imag_ax)
-
-    # Save
-    np.save(f'{TEST_DIR}/ax_real.npy', real_ax)
-    np.save(f'{TEST_DIR}/ax_imag.npy', imag_ax)
-    np.save(f'{TEST_DIR}/ax_audio.npy', audio_ax)
-    import soundfile as sf
-    sf.write(f'{TEST_DIR}/ax_audio.wav', audio_ax, 24000)
-
-    # Compare with ONNX reference
-    onnx_real = np.load(f'{TEST_DIR}/onnx_real.npy')
-    onnx_imag = np.load(f'{TEST_DIR}/onnx_imag.npy')
-    onnx_audio = np.load(f'{TEST_DIR}/onnx_audio.npy')
-
-    print("\n=== axmodel vs ONNX (board) ===")
-    compare('real_spectrum', onnx_real, real_ax)
-    compare('imag_spectrum', onnx_imag, imag_ax)
-    compare('audio', onnx_audio, audio_ax)
-
-    print(f"\nResults saved to {TEST_DIR}/")
-    print("Copy back to dev machine and run: python3 compare_vocoder.py --compare")
-
-
-def cmd_compare():
-    """Compare all outputs (dev machine, after copying ax_*.npy from board)."""
-    pt_audio    = np.load(f'{TEST_DIR}/pt_audio.npy')
-    onnx_audio  = np.load(f'{TEST_DIR}/onnx_audio.npy')
-    ax_audio    = np.load(f'{TEST_DIR}/ax_audio.npy')
-
-    pt_real  = np.load(f'{TEST_DIR}/pt_real.npy')
-    onnx_real = np.load(f'{TEST_DIR}/onnx_real.npy')
-    ax_real   = np.load(f'{TEST_DIR}/ax_real.npy')
-
-    pt_imag  = np.load(f'{TEST_DIR}/pt_imag.npy')
-    onnx_imag = np.load(f'{TEST_DIR}/onnx_imag.npy')
-    ax_imag   = np.load(f'{TEST_DIR}/ax_imag.npy')
-
-    print("=== Full Comparison ===")
-    print("\n--- Spectrum ---")
-    compare('real: PT vs ONNX', pt_real, onnx_real)
-    compare('real: PT vs axmodel', pt_real, ax_real)
-    compare('real: ONNX vs axmodel', onnx_real, ax_real)
-    print()
-    compare('imag: PT vs ONNX', pt_imag, onnx_imag)
-    compare('imag: PT vs axmodel', pt_imag, ax_imag)
-    compare('imag: ONNX vs axmodel', onnx_imag, ax_imag)
-
-    print("\n--- Audio ---")
-    compare('audio: PT vs ONNX', pt_audio, onnx_audio)
-    compare('audio: PT vs axmodel', pt_audio, ax_audio)
-    compare('audio: ONNX vs axmodel', onnx_audio, ax_audio)
-
-
-def main():
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--save-test-data', action='store_true')
-    parser.add_argument('--run-axmodel', action='store_true')
-    parser.add_argument('--compare', action='store_true')
-    args = parser.parse_args()
-
-    if args.save_test_data:
-        cmd_save_test_data()
-    elif args.run_axmodel:
-        cmd_run_axmodel()
-    elif args.compare:
-        cmd_compare()
-    else:
-        print("Usage: --save-test-data | --run-axmodel | --compare")
-        print("\nWorkflow:")
-        print("  1. Dev machine:  python3 compare_vocoder.py --save-test-data")
-        print("  2. Copy TEST_DIR to board, run: python3 compare_vocoder.py --run-axmodel")
-        print("  3. Copy ax_*.npy back to dev, run: python3 compare_vocoder.py --compare")
-
-
-if __name__ == '__main__':
-    main()

cpp/scripts/export_vocos_onnx.py

@@ -1,125 +0,0 @@
-#!/usr/bin/env python3
-"""
-Export Vocos ONNX models with onnxsim optimization.
-
-Static shapes (B=1, T=620):
-  vocos_backbone_B1_T620.onnx     mel [1,100,620] → features [1,620,512]
-  vocos_head_linear_B1_T620.onnx  features [1,620,512] → (real, imag) [1,620,513]
-  vocos_full_B1_T620.onnx         mel [1,100,620] → (real, imag) [1,620,513]
-
-IRFFT + window + overlap-add is handled in C++ (simple, ~50 lines with existing FFT).
-"""
-import sys, os, math
-import numpy as np
-import torch
-import torch.nn as nn
-import onnx
-from onnxsim import simplify
-
-REPO_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-sys.path.insert(0, REPO_DIR)
-from scripts.local_vocos import LocalVocos, ISTFTHead
-
-OUT_DIR = os.path.join(REPO_DIR, 'cpp', 'vocoder_onnx')
-os.makedirs(OUT_DIR, exist_ok=True)
-
-
-class HeadLinearOnly(nn.Module):
-    """Linear(512→1026) + mag/phase → real/imag spectrum. No ISTFT."""
-    def __init__(self, head: ISTFTHead):
-        super().__init__()
-        self.out = head.out  # Linear(512, 1026)
-    def forward(self, x):    # x: [B, T, 512]
-        h = self.out(x); mag, phase = h.chunk(2, dim=-1)
-        mag = torch.exp(mag).clamp(max=1e2)
-        return mag * torch.cos(phase), mag * torch.sin(phase)
-
-
-class FullModel(nn.Module):
-    def __init__(self, backbone, head_linear):
-        super().__init__()
-        self.backbone = backbone
-        self.head_linear = head_linear
-    def forward(self, mel):
-        return self.head_linear(self.backbone(mel))
-
-
-def export_onnx(model, dummy, path, in_names, out_names, dyn=None):
-    tmp = path + '.tmp'
-    torch.onnx.export(model, dummy, tmp,
-                      input_names=in_names, output_names=out_names,
-                      dynamic_axes=dyn, opset_version=17, do_constant_folding=True)
-    model_simp, check = simplify(tmp)
-    assert check, f"onnxsim verification failed for {path}"
-    onnx.save(model_simp, path)
-    os.remove(tmp)
-    size_mb = os.path.getsize(path) / 1024 / 1024
-    print(f"  {os.path.basename(path):50s} {size_mb:6.1f} MB")
-
-
-def verify(name, pt_list, onnx_list):
-    for i, (pt, onnx_val) in enumerate(zip(pt_list, onnx_list)):
-        diff = np.abs(pt - onnx_val)
-        sig = max(np.mean(np.abs(pt)), 1e-10)
-        print(f"  [{name}/{i}] max_err={diff.max():.2e}  rel_err={diff.mean()/sig:.2e}  "
-              f"shape={pt.shape}")
-
-
-def main():
-    print("Loading Vocos...")
-    vocoder = LocalVocos()
-    sd = torch.load(f'{REPO_DIR}/resources/vocos-mel-24khz/pytorch_model.bin',
-                    weights_only=True, map_location='cpu')
-    sd = {k: v for k, v in sd.items() if k.startswith(('backbone.', 'head.'))}
-    vocoder.load_state_dict(sd)
-    vocoder.eval()
-
-    backbone = vocoder.backbone
-    head_linear = HeadLinearOnly(vocoder.head).eval()
-
-    B, T = 1, 620
-    dummy_mel = torch.randn(B, 100, T)
-
-    with torch.no_grad():
-        bb_out = backbone(dummy_mel)
-        pt_real, pt_imag = head_linear(bb_out)
-
-    print(f"PT output: real={pt_real.shape}, imag={pt_imag.shape}")
-
-    # --- Export static (B=1, T=620) ---
-    print("\n=== Static models (B=1, T=620) ===")
-    export_onnx(backbone, dummy_mel,
-                f'{OUT_DIR}/vocos_backbone_B1_T620.onnx',
-                ['mel'], ['features'])
-    export_onnx(head_linear, bb_out,
-                f'{OUT_DIR}/vocos_head_linear_B1_T620.onnx',
-                ['features'], ['real', 'imag'])
-    export_onnx(FullModel(backbone, head_linear).eval(), dummy_mel,
-                f'{OUT_DIR}/vocos_full_B1_T620.onnx',
-                ['mel'], ['real', 'imag'])
-
-    # --- Verify with onnxruntime ---
-    print("\n=== Verification ===")
-    import onnxruntime as ort
-
-    for tag, model_file, input_dict, pt_expected in [
-        ('backbone', 'vocos_backbone_B1_T620.onnx',
-         {'mel': dummy_mel.numpy()}, [bb_out.numpy()]),
-        ('head_linear', 'vocos_head_linear_B1_T620.onnx',
-         {'features': bb_out.numpy()}, [pt_real.numpy(), pt_imag.numpy()]),
-        ('full', 'vocos_full_B1_T620.onnx',
-         {'mel': dummy_mel.numpy()}, [pt_real.numpy(), pt_imag.numpy()]),
-    ]:
-        sess = ort.InferenceSession(f'{OUT_DIR}/{model_file}')
-        onnx_out = sess.run(None, input_dict)
-        verify(tag, pt_expected, onnx_out)
-
-    print("\nDone! All exported and verified.")
-    for f in sorted(os.listdir(OUT_DIR)):
-        if 'B1_T620' in f:
-            size = os.path.getsize(f'{OUT_DIR}/{f}') / 1024 / 1024
-            print(f"  {f} ({size:.1f} MB)")
-
-
-if __name__ == '__main__':
-    main()

cpp/scripts/gen_cat_tokens.py

@@ -1,44 +0,0 @@
-#!/usr/bin/env python3
-"""Generate cat_tokens binary for C++ ZipVoice inference. Works for any language.
-
-Usage:
-  python3 gen_cat_tokens.py --prompt "prompt text" --text "target text"
-  python3 gen_cat_tokens.py --prompt-file prompt.txt --text-file target.txt
-"""
-import sys, os, argparse
-import numpy as np
-
-REPO_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))  # cpp/scripts → repo root
-sys.path.insert(0, REPO_DIR)
-
-from scripts.local_tokenizer import LocalEmiliaTokenizer
-from scripts.text_processing import normalize_punctuation
-
-parser = argparse.ArgumentParser()
-parser.add_argument('--prompt', default='')
-parser.add_argument('--text', default='')
-parser.add_argument('--prompt-file', default='')
-parser.add_argument('--text-file', default='')
-parser.add_argument('--output', default='cat_tokens.bin')
-parser.add_argument('--max-tokens', type=int, default=384)
-args = parser.parse_args()
-
-prompt = args.prompt or (open(args.prompt_file).read().strip() if args.prompt_file else '')
-text = args.text or (open(args.text_file).read().strip() if args.text_file else '')
-
-if not prompt or not text:
-    print("ERROR: provide --prompt/--prompt-file and --text/--text-file")
-    sys.exit(1)
-
-token_file = os.path.join(REPO_DIR, 'resources', 'zipvoice_hf', 'zipvoice', 'tokens.txt')
-tokenizer = LocalEmiliaTokenizer(token_file=token_file)
-
-pids = tokenizer.texts_to_token_ids([normalize_punctuation(prompt)])[0]
-tids = tokenizer.texts_to_token_ids([normalize_punctuation(text)])[0]
-
-cat = pids + tids + [tokenizer.pad_id]
-ct = np.full((args.max_tokens,), tokenizer.pad_id, dtype=np.int32)
-ct[:len(cat)] = np.array(cat, dtype=np.int32)
-ct.tofile(args.output)
-
-print(f'prompt_tokens_len={len(pids)}  text_tokens_len={len(tids)}  saved: {args.output}')

cpp/scripts/gen_pinyin_table.py

@@ -1,95 +0,0 @@
-#!/usr/bin/env python3
-"""Generate C++ header with Chinese character → pinyin token mapping."""
-import sys, os
-REPO_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-sys.path.insert(0, REPO_DIR)
-
-from pypinyin.pinyin_dict import pinyin_dict
-from pypinyin.contrib.tone_convert import to_initials, to_finals_tone3
-
-TOKEN_FILE = os.path.join(REPO_DIR, 'resources', 'zipvoice_hf', 'zipvoice', 'tokens.txt')
-OUTPUT = os.path.join(REPO_DIR, 'cpp', 'src', 'pinyin_table.hpp')
-
-# Load token2id
-token2id = {}
-with open(TOKEN_FILE) as f:
-    for line in f:
-        if '\t' in line:
-            t, tid = line.rstrip().split('\t')
-            token2id[t] = int(tid)
-
-# Build char → token IDs
-char_entries = []
-for codepoint, pinyins in pinyin_dict.items():
-    ch = chr(codepoint)
-    py = pinyins.split(',')[0]
-    initial = to_initials(py, strict=False)
-    final = to_finals_tone3(py, strict=False, neutral_tone_with_five=True)
-    if not final: continue
-
-    ini_token = initial + '0' if initial else ''
-    tokens = []
-    if ini_token and ini_token in token2id:
-        tokens.append(token2id[ini_token])
-    if final in token2id:
-        tokens.append(token2id[final])
-
-    if tokens:
-        char_entries.append((codepoint, tokens))
-
-# Sort by codepoint for binary search
-char_entries.sort(key=lambda x: x[0])
-
-# Generate C++ header
-with open(OUTPUT, 'w') as f:
-    f.write('''// Auto-generated Chinese character → token ID mapping
-// Source: pypinyin.pinyin_dict, tokens.txt
-// Generated by gen_pinyin_table.py
-
-#pragma once
-#include <cstdint>
-#include <string>
-#include <vector>
-
-struct PinyinEntry {
-    uint32_t codepoint;
-    int16_t  token1;
-    int16_t  token2;  // -1 if only one token
-};
-
-// Sorted by codepoint, use binary search
-static const PinyinEntry PINYIN_TABLE[] = {
-''')
-
-    for codepoint, tokens in char_entries:
-        t1 = tokens[0]
-        t2 = tokens[1] if len(tokens) > 1 else -1
-        f.write(f'    {{0x{codepoint:04X}, {t1}, {t2}}},  // {chr(codepoint)}\n')
-
-    f.write(f'''}};
-
-static const int PINYIN_TABLE_SIZE = {len(char_entries)};
-
-// Lookup tokens for a Chinese character. Returns empty vector if not found.
-inline std::vector<int> pinyin_lookup(uint32_t codepoint) {{
-    // Binary search
-    int lo = 0, hi = PINYIN_TABLE_SIZE - 1;
-    while (lo <= hi) {{
-        int mid = (lo + hi) / 2;
-        if (PINYIN_TABLE[mid].codepoint == codepoint) {{
-            std::vector<int> result;
-            result.push_back(PINYIN_TABLE[mid].token1);
-            if (PINYIN_TABLE[mid].token2 >= 0)
-                result.push_back(PINYIN_TABLE[mid].token2);
-            return result;
-        }}
-        if (PINYIN_TABLE[mid].codepoint < codepoint)
-            lo = mid + 1;
-        else
-            hi = mid - 1;
-    }}
-    return {{}};
-}}
-''')
-
-print(f"Generated {OUTPUT}: {len(char_entries)} entries")

cpp/scripts/generate_vocoder_calib.py

@@ -1,141 +0,0 @@
-#!/usr/bin/env python3
-"""
-Generate vocoder calibration data on dev machine (no axengine needed).
-
-Collects mel features from:
-  1. Real mel output files (output_mel.bin from previous C++ runs)
-  2. Random mel features in the expected value range
-  3. Zero/silence mel features
-Then runs them through ONNX to get intermediate features for head_linear calibration.
-
-Usage:
-  python3 generate_vocoder_calib.py --output-dir ./calib_data_vocoder
-"""
-import sys, os, argparse, json, glob
-from pathlib import Path
-import numpy as np
-import onnxruntime as ort
-
-SCRIPT_DIR = Path(__file__).resolve().parent
-REPO_DIR = SCRIPT_DIR.parent.parent  # scripts/ → cpp/ → repo root
-ONNX_DIR = SCRIPT_DIR.parent / "vocoder_onnx"  # cpp/vocoder_onnx
-
-
-def main():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--output-dir", default=str(ONNX_DIR / "calib_data"))
-    parser.add_argument("--num-samples", type=int, default=16)
-    parser.add_argument("--seed", type=int, default=42)
-    args = parser.parse_args()
-
-    out_dir = Path(args.output_dir)
-    out_dir.mkdir(parents=True, exist_ok=True)
-    rng = np.random.RandomState(args.seed)
-
-    T = 620
-    mel_samples = []
-
-    # --- Source 1: Real mel features if available (undo feat_scale to match vocoder input) ---
-    feat_scale = 0.1
-    for pattern in ["output_mel.bin", "output_mel_debug.bin"]:
-        for p in [REPO_DIR / pattern, SCRIPT_DIR / pattern]:
-            if p.exists():
-                mel = np.fromfile(str(p), dtype=np.float32).reshape(-1, 100).T[np.newaxis, :, :] / feat_scale  # [1, 100, frames]
-                gen_frames = min(mel.shape[2], T)
-                padded = np.zeros((1, 100, T), dtype=np.float32)
-                padded[0, :, :gen_frames] = mel[0, :, :gen_frames]
-                mel_samples.append(padded)
-                print(f"Real mel from {p}: shape={mel.shape}")
-
-    # --- Source 2: Random mel in vocoder input range (after /feat_scale) ---
-    # Python: features / 0.1 before vocoder → range ~[-8.6, 4.0]
-    # So calibration mel should be in [feat_scaled] range, divided by feat_scale
-    # to match the actual vocoder input distribution
-    feat_scale = 0.1
-    for i in range(max(0, args.num_samples - len(mel_samples))):
-        length = rng.randint(100, T + 1)
-        # Generate in feat_scaled range, then undo feat_scale
-        mel = (rng.randn(1, 100, T).astype(np.float32) * 0.3 - 0.04) / feat_scale
-        mel[0, :, length:] = 0.0
-        mel_samples.append(mel)
-
-    # --- Source 3: Edge cases (with feat_scale undo) ---
-    mel_samples.append(np.zeros((1, 100, T), dtype=np.float32))
-    mel_samples.append(np.ones((1, 100, T), dtype=np.float32) * 0.5 / feat_scale)
-    mel_samples.append(rng.randn(1, 100, T).astype(np.float32) * 0.1 / feat_scale)
-    mel_samples.append(rng.randn(1, 100, T).astype(np.float32) / feat_scale)
-
-    print(f"\nTotal mel samples: {len(mel_samples)}")
-
-    # --- Run ONNX to get backbone outputs (for head_linear calibration) ---
-    bb_path = ONNX_DIR / "vocos_backbone_B1_T620.onnx"
-    hl_path = ONNX_DIR / "vocos_head_linear_B1_T620.onnx"
-
-    if not bb_path.exists():
-        print(f"ERROR: {bb_path} not found. Run export_vocos_onnx.py first.")
-        return
-
-    sess_bb = ort.InferenceSession(str(bb_path))
-    print(f"Loaded backbone: {bb_path}")
-
-    # --- Save backbone calibration data ---
-    bb_dir = out_dir / "vocos_backbone" / "mel"
-    bb_dir.mkdir(parents=True, exist_ok=True)
-    bb_entries = []
-    backbone_outputs = []
-
-    for i, mel in enumerate(mel_samples):
-        np.save(bb_dir / f"{i:04d}.npy", mel)
-        bb_entries.append({"file": f"mel/{i:04d}.npy", "shape": list(mel.shape)})
-
-        # Run ONNX to get head_linear input
-        feat = sess_bb.run(None, {'mel': mel})[0]
-        backbone_outputs.append(feat)
-
-    print(f"Backbone calibration: {len(bb_entries)} samples")
-
-    # --- Save head_linear calibration data ---
-    hl_dir = out_dir / "vocos_head_linear" / "features"
-    hl_dir.mkdir(parents=True, exist_ok=True)
-    hl_entries = []
-
-    for i, feat in enumerate(backbone_outputs):
-        np.save(hl_dir / f"{i:04d}.npy", feat)
-        hl_entries.append({"file": f"features/{i:04d}.npy", "shape": list(feat.shape)})
-
-    print(f"Head_linear calibration: {len(hl_entries)} samples")
-
-    # --- Verify head_linear ONNX with calibration data ---
-    if hl_path.exists():
-        sess_hl = ort.InferenceSession(str(hl_path))
-        for i in range(min(3, len(backbone_outputs))):
-            r, im = sess_hl.run(None, {'features': backbone_outputs[i]})
-            print(f"  Verify head[{i}]: real range=[{r.min():.3f},{r.max():.3f}], "
-                  f"imag range=[{im.min():.3f},{im.max():.3f}]")
-
-    # --- Manifest ---
-    manifest = {
-        "description": "Vocoder ONNX calibration data",
-        "backbone": {
-            "model": "vocos_backbone_B1_T620.onnx",
-            "input": "mel", "shape": [1, 100, 620], "dtype": "float32",
-            "num_samples": len(bb_entries),
-            "files": bb_entries,
-        },
-        "head_linear": {
-            "model": "vocos_head_linear_B1_T620.onnx",
-            "input": "features", "shape": [1, 620, 512], "dtype": "float32",
-            "num_samples": len(hl_entries),
-            "files": hl_entries,
-        },
-    }
-    with open(out_dir / "calib_manifest.json", "w") as f:
-        json.dump(manifest, f, indent=2, ensure_ascii=False)
-
-    print(f"\nDone! Output: {out_dir}")
-    print(f"  vocos_backbone/mel/   : {len(bb_entries)} .npy files")
-    print(f"  vocos_head_linear/features/: {len(hl_entries)} .npy files")
-
-
-if __name__ == "__main__":
-    main()

cpp/scripts/py_daemon.py

@@ -1,57 +0,0 @@
-#!/usr/bin/env python3
-"""
-ZipVoice Python daemon: loads tokenizer once, handles C++ requests via stdin/stdout.
-Protocol (tab-separated lines to preserve spaces in text paths):
-  count\t<prompt_file>\t<text_file>
-    -> prints "COUNT <prompt_len> <text_len>"
-  tokenize\t<prompt_file>\t<text_file>\t<max_tokens>\t<output_bin>
-    -> prints "TOKENS <prompt_len> <text_len>"
-  quit
-    -> exits
-"""
-import sys, os
-import numpy as np
-
-REPO_DIR = sys.argv[1]
-sys.path.insert(0, REPO_DIR)
-
-from scripts.local_tokenizer import LocalEmiliaTokenizer
-from scripts.text_processing import normalize_punctuation
-
-TOKEN_FILE = os.path.join(REPO_DIR, "resources", "zipvoice_hf", "zipvoice", "tokens.txt")
-tokenizer = LocalEmiliaTokenizer(token_file=TOKEN_FILE)
-
-print("READY", flush=True)
-
-for line in sys.stdin:
-    line = line.rstrip("\n")
-    if not line:
-        continue
-    parts = line.split("\t")
-    cmd = parts[0]
-
-    if cmd == "count":
-        prompt_file, text_file = parts[1], parts[2]
-        pt = normalize_punctuation(open(prompt_file).read().strip())
-        tt = normalize_punctuation(open(text_file).read().strip())
-        pids = tokenizer.texts_to_token_ids([pt])[0]
-        tids = tokenizer.texts_to_token_ids([tt])[0]
-        print(f"COUNT {len(pids)} {len(tids)}", flush=True)
-
-    elif cmd == "tokenize":
-        prompt_file, text_file, max_tokens, output_bin = parts[1], parts[2], int(parts[3]), parts[4]
-        pt = normalize_punctuation(open(prompt_file).read().strip())
-        tt = normalize_punctuation(open(text_file).read().strip())
-        pids = tokenizer.texts_to_token_ids([pt])[0]
-        tids = tokenizer.texts_to_token_ids([tt])[0]
-        cat = pids + tids + [tokenizer.pad_id]
-        if len(cat) > max_tokens:
-            print(f"ERROR too_many_tokens {len(cat)}>{max_tokens}", flush=True)
-            continue
-        ct = np.full((max_tokens,), tokenizer.pad_id, dtype=np.int32)
-        ct[:len(cat)] = np.array(cat, dtype=np.int32)
-        ct.tofile(output_bin)
-        print(f"TOKENS {len(pids)} {len(tids)}", flush=True)
-
-    elif cmd == "quit":
-        break

cpp/scripts/quantize_vocoder.sh

@@ -1,123 +0,0 @@
-#!/usr/bin/env bash
-# Quantize vocoder ONNX models to axmodel using pulsar2.
-#
-# Prerequisites:
-#   1. export_vocos_onnx.py (generates vocoder_onnx/*.onnx)
-#   2. generate_vocoder_calib.py (generates vocoder_onnx/calib_data/)
-#   3. source <npu_dev_env> (activates pulsar2, e.g. $HOME/npu-codebase/script/npu_dev)
-#
-# Usage:
-#   bash quantize_vocoder.sh
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-PARENT_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
-cd "${PARENT_DIR}"
-
-ONNX_DIR="${PARENT_DIR}/vocoder_onnx"
-CALIB_DIR="${ONNX_DIR}/calib_data"
-BUILD_DIR="${ONNX_DIR}/pulsar2_build"
-AXMODEL_DIR="${ONNX_DIR}/axmodel"
-
-# --- Config ---
-TARGET_HARDWARE="${TARGET_HARDWARE:-AX650}"
-NPU_MODE="${NPU_MODE:-NPU3}"
-CHECK_LEVEL="${CHECK_LEVEL:-0}"
-DEFAULT_DATA_TYPE="${DEFAULT_DATA_TYPE:-U16}"
-CALIBRATION_SIZE="${CALIBRATION_SIZE:--1}"
-
-FULL_MODEL="${ONNX_DIR}/vocos_full_B1_T620.onnx"
-CALIB_DIR_BB="${CALIB_DIR}/vocos_backbone"
-
-if [ ! -f "${FULL_MODEL}" ]; then
-    echo "ERROR: ${FULL_MODEL} not found. Run export_vocos_onnx.py first."
-    exit 1
-fi
-
-if [ ! -d "${CALIB_DIR_BB}" ]; then
-    echo "ERROR: ${CALIB_DIR_BB} not found. Run generate_vocoder_calib.py first."
-    exit 1
-fi
-
-# Activate pulsar2 if not already available
-if ! command -v pulsar2 >/dev/null 2>&1; then
-    echo "Activating pulsar2 environment..."
-    set +u
-    source "${NPU_DEV_ENV:-$HOME/npu-codebase/script/npu_dev}"
-    set -u
-fi
-
-mkdir -p "${BUILD_DIR}" "${AXMODEL_DIR}"
-
-# --- Step 1: Package calibration data as tar.gz ---
-echo "=== Packaging calibration data ==="
-
-# Calibration for full model: mel -> (real, imag)
-# Input tensor is "mel", calibration data in vocos_backbone/mel/
-FULL_CALIB_TAR="${CALIB_DIR}/vocos_full_mel.tar.gz"
-INPUT_DIR="${CALIB_DIR_BB}/mel"
-
-if [ ! -f "${FULL_CALIB_TAR}" ]; then
-    echo "Creating ${FULL_CALIB_TAR}..."
-    cd "${INPUT_DIR}"
-    tar czf "${FULL_CALIB_TAR}" --transform 's|.*/||' *.npy
-    cd "${SCRIPT_DIR}"
-fi
-echo "  ${FULL_CALIB_TAR} ($(du -h "${FULL_CALIB_TAR}" | cut -f1))"
-
-# --- Step 2: Build pulsar2 config ---
-echo ""
-echo "=== Quantizing vocos_full ==="
-
-CONFIG="${BUILD_DIR}/config_vocos_full.json"
-
-cat > "${CONFIG}" << EOF
-{
-    "model_type": "ONNX",
-    "npu_mode": "${NPU_MODE}",
-    "input": "${FULL_MODEL}",
-    "output_name": "vocos_full.axmodel",
-    "output_dir": "${BUILD_DIR}",
-    "target_hardware": "${TARGET_HARDWARE}",
-    "onnx_opt": {"disable_onnx_optimization": false, "enable_onnxsim": true},
-    "quant": {
-        "input_configs": [
-            {"tensor_name": "mel", "calibration_dataset": "${FULL_CALIB_TAR}", "calibration_format": "Numpy", "calibration_size": ${CALIBRATION_SIZE}}
-        ],
-        "layer_configs": [
-            {"op_type": "Pow", "data_type": "U8"},
-            {"start_tensor_names": ["DEFAULT"], "end_tensor_names": ["DEFAULT"], "data_type": "${DEFAULT_DATA_TYPE}"}
-        ],
-        "calibration_method": "MinMax",
-        "enable_smooth_quant": true,
-        "conv_bias_data_type": "FP32",
-        "precision_analysis": true,
-        "precision_analysis_method": "EndToEnd",
-        "disable_auto_refine_scale": true,
-        "transformer_opt_level": 0
-    },
-    "input_processors": [{"tensor_name": "DEFAULT"}],
-    "compiler": {"check": ${CHECK_LEVEL}, "enable_slice_mode": false}
-}
-EOF
-
-echo "Config: ${CONFIG}"
-echo "Input shapes: mel:1x100x620"
-
-# --- Step 3: Run pulsar2 ---
-echo ""
-echo "=== Running pulsar2 build ==="
-pulsar2 build --config "${CONFIG}" --input_shapes "mel:1x100x620" 2>&1 | tee "${BUILD_DIR}/build_vocos_full.log"
-
-# --- Step 4: Copy output ---
-BUILD_OUTPUT="${BUILD_DIR}/vocos_full.axmodel"
-if [ -f "${BUILD_OUTPUT}" ]; then
-    cp "${BUILD_OUTPUT}" "${AXMODEL_DIR}/"
-    echo ""
-    echo "=== Done! ==="
-    ls -lh "${AXMODEL_DIR}/vocos_full.axmodel"
-else
-    echo "ERROR: Build failed, ${BUILD_OUTPUT} not found"
-    exit 1
-fi

cpp/src/EngineWrapper.cpp

@@ -1,13 +0,0 @@
-/**************************************************************************************************
- * ZipVoice C++ — Unified EngineWrapper implementation (AXERA & AXCL)
- *
- * Conditionally includes the platform-specific implementation.
- **************************************************************************************************/
-
-#if defined(AX650) || defined(AX630C) || defined(AX620Q)
-    #include "EngineWrapper_axera.cpp"
-#elif defined(AXCL)
-    #include "EngineWrapper_axcl.cpp"
-#else
-    #error "Unknown platform. Define AX650, AX630C, AX620Q, or AXCL."
-#endif

cpp/src/EngineWrapper.hpp

@@ -1,81 +0,0 @@
-/**************************************************************************************************
- * ZipVoice C++ — Unified EngineWrapper (AXERA & AXCL)
- *
- * Supports both AXERA (demo board) and AXCL (compute card) via conditional compilation.
- *   AX650 / AX630C / AX620Q → AXERA  path (ax_engine_api)
- *   AXCL                           → AXCL   path (middleware::runner)
- *
- * Interface is identical — callers don't need to know which backend is active.
- **************************************************************************************************/
-
-#pragma once
-
-#include <string>
-#include <vector>
-#include <unordered_map>
-#include <cstring>
-#include <cstdint>
-#include <memory>
-
-#if defined(AX650) || defined(AX630C) || defined(AX620Q)
-    #include "ax_engine_api.h"
-#elif defined(AXCL)
-    #include "middleware/axcl_runtime_runner.hpp"
-#endif
-
-class EngineWrapper {
-public:
-    EngineWrapper();
-    ~EngineWrapper();
-
-    /**
-     * Initialize and load a model.
-     * @param strModelPath  Path to .axmodel file
-     * @param nNpuType      AXERA: VNPU type (0=default); AXCL: device_index (0=first card)
-     * @param axclConfig    AXCL: path to axcl.json (nullptr → /usr/local/axcl/axcl.json). Ignored on AXERA.
-     */
-    int Init(const char* strModelPath, uint32_t nNpuType = 0, const char* axclConfig = nullptr);
-
-    // Index-based I/O
-    int SetInput(void* pInput, int index);
-    int RunSync();
-    int GetOutput(void* pOutput, int index);
-
-    // Name-based I/O
-    int SetInputByName(const char* name, void* pInput);
-    int GetOutputByName(const char* name, void* pOutput);
-
-    int GetInputSize(int index);
-    int GetOutputSize(int index);
-    int GetInputSizeByName(const char* name);
-    int GetOutputSizeByName(const char* name);
-
-    int GetInputIndex(const char* name);
-    int GetOutputIndex(const char* name);
-    int GetInputCount() const { return m_input_num; }
-    int GetOutputCount() const { return m_output_num; }
-    const char* GetInputName(int index) const;
-    const char* GetOutputName(int index) const;
-#if defined(AX650) || defined(AX630C) || defined(AX620Q)
-    const char* GetInputDtypeStr(int index) const;
-    const char* GetOutputDtypeStr(int index) const;
-#endif
-
-    bool HasInit() const { return m_hasInit; }
-    int Release();
-
-private:
-    bool m_hasInit;
-    int m_input_num, m_output_num;
-
-    std::unordered_map<std::string, int> m_input_name_to_idx;
-    std::unordered_map<std::string, int> m_output_name_to_idx;
-
-#if defined(AX650) || defined(AX630C) || defined(AX620Q)
-    AX_ENGINE_HANDLE m_handle;
-    AX_ENGINE_IO_INFO_T *m_io_info;
-    AX_ENGINE_IO_T m_io;
-#elif defined(AXCL)
-    std::unique_ptr<middleware::runner> m_runner;
-#endif
-};

cpp/src/EngineWrapper_axcl.cpp

@@ -1,217 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXCL C++ Port
- *
- * EngineWrapper implementation using middleware::runner (AXCL Runtime API).
- *
- * Key differences from AXERA:
- *   - Uses axclrtMemcpy for host↔device data transfer
- *   - No CMM (contiguous memory) allocation
- *   - Runner manages device-side memory and NPU execution
- **************************************************************************************************/
-
-#include "EngineWrapper.hpp"
-
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-
-#include <axcl.h>
-
-EngineWrapper::EngineWrapper()
-    : m_hasInit(false), m_input_num(0), m_output_num(0) {}
-
-EngineWrapper::~EngineWrapper() {
-    Release();
-}
-
-int EngineWrapper::Init(const char* strModelPath,
-                         uint32_t nNpuType,
-                         const char* axclConfig) {
-    const char* config = axclConfig ? axclConfig : "/usr/local/axcl/axcl.json";
-    uint32_t device_index = nNpuType;      // nNpuType = device_index on AXCL
-    uint32_t npu_kind = 0;                  // AXCL_VNPU_DISABLE (default)
-
-    // 1. Create runtime runner
-    m_runner = std::make_unique<middleware::runtime_runner>();
-
-    // 2. Initialize runner (handles axclrtEngineInit internally)
-    if (!m_runner->init(config, device_index, npu_kind)) {
-        fprintf(stderr, "[ERROR] AXCL runner init failed for model: %s\n", strModelPath);
-        return -1;
-    }
-
-    // 3. Load model from file
-    if (!m_runner->load(strModelPath)) {
-        fprintf(stderr, "[ERROR] AXCL runner load failed for model: %s\n", strModelPath);
-        return -1;
-    }
-
-    // 4. Prepare IO buffers (allocates device memory, sets up IO)
-    if (!m_runner->prepare(true, true, 0, 0)) {
-        fprintf(stderr, "[ERROR] AXCL runner prepare failed for model: %s\n", strModelPath);
-        return -1;
-    }
-
-    // 5. Build name-to-index maps
-    m_input_num = static_cast<int>(m_runner->get_input_count());
-    m_output_num = static_cast<int>(m_runner->get_output_count());
-
-    m_input_name_to_idx.clear();
-    for (int i = 0; i < m_input_num; ++i) {
-        std::string name = m_runner->get_input_name(i);
-        m_input_name_to_idx[name] = i;
-    }
-
-    m_output_name_to_idx.clear();
-    for (int i = 0; i < m_output_num; ++i) {
-        std::string name = m_runner->get_output_name(i);
-        m_output_name_to_idx[name] = i;
-    }
-
-    m_hasInit = true;
-
-    printf("AXCL EngineWrapper loaded: %s (inputs=%d, outputs=%d)\n",
-           strModelPath, m_input_num, m_output_num);
-    for (int i = 0; i < m_input_num; ++i) {
-        printf("  input[%d]: %s (%zu bytes)\n", i,
-               m_runner->get_input_name(i).c_str(),
-               (size_t)m_runner->get_input_size(i));
-    }
-    for (int i = 0; i < m_output_num; ++i) {
-        printf("  output[%d]: %s (%zu bytes)\n", i,
-               m_runner->get_output_name(i).c_str(),
-               (size_t)m_runner->get_output_size(i));
-    }
-
-    return 0;
-}
-
-int EngineWrapper::SetInput(void* pInput, int index) {
-    if (!m_hasInit || index < 0 || index >= m_input_num) return -1;
-
-    uintmax_t size = m_runner->get_input_size(index);
-    void* dev_ptr = m_runner->get_input_pointer(index);
-
-    if (!dev_ptr || size == 0) {
-        fprintf(stderr, "[ERROR] Invalid input pointer/size for index %d\n", index);
-        return -1;
-    }
-
-    // Copy host data to device memory
-    axclError ret = axclrtMemcpy(dev_ptr, pInput, size, AXCL_MEMCPY_HOST_TO_DEVICE);
-    if (ret != 0) {
-        fprintf(stderr, "[ERROR] axclrtMemcpy H2D failed for input %d: 0x%x\n", index, ret);
-        return -1;
-    }
-
-    return 0;
-}
-
-int EngineWrapper::RunSync() {
-    if (!m_hasInit) return -1;
-
-    if (!m_runner->run(false)) {
-        fprintf(stderr, "[ERROR] AXCL runner run failed\n");
-        return -1;
-    }
-
-    return 0;
-}
-
-int EngineWrapper::GetOutput(void* pOutput, int index) {
-    if (!m_hasInit || index < 0 || index >= m_output_num) return -1;
-
-    uintmax_t size = m_runner->get_output_size(index);
-    void* dev_ptr = m_runner->get_output_pointer(index);
-
-    if (!dev_ptr || size == 0) {
-        fprintf(stderr, "[ERROR] Invalid output pointer/size for index %d\n", index);
-        return -1;
-    }
-
-    // Copy device data to host memory
-    axclError ret = axclrtMemcpy(pOutput, dev_ptr, size, AXCL_MEMCPY_DEVICE_TO_HOST);
-    if (ret != 0) {
-        fprintf(stderr, "[ERROR] axclrtMemcpy D2H failed for output %d: 0x%x\n", index, ret);
-        return -1;
-    }
-
-    return 0;
-}
-
-int EngineWrapper::SetInputByName(const char* name, void* pInput) {
-    auto it = m_input_name_to_idx.find(std::string(name));
-    if (it == m_input_name_to_idx.end()) {
-        fprintf(stderr, "[ERROR] Input '%s' not found in model\n", name);
-        return -1;
-    }
-    return SetInput(pInput, it->second);
-}
-
-int EngineWrapper::GetOutputByName(const char* name, void* pOutput) {
-    auto it = m_output_name_to_idx.find(std::string(name));
-    if (it == m_output_name_to_idx.end()) {
-        fprintf(stderr, "[ERROR] Output '%s' not found in model\n", name);
-        return -1;
-    }
-    return GetOutput(pOutput, it->second);
-}
-
-int EngineWrapper::GetInputSize(int index) {
-    if (!m_hasInit || index < 0 || index >= m_input_num) return -1;
-    return static_cast<int>(m_runner->get_input_size(index));
-}
-
-int EngineWrapper::GetOutputSize(int index) {
-    if (!m_hasInit || index < 0 || index >= m_output_num) return -1;
-    return static_cast<int>(m_runner->get_output_size(index));
-}
-
-int EngineWrapper::GetInputSizeByName(const char* name) {
-    int idx = GetInputIndex(name);
-    if (idx < 0) return -1;
-    return GetInputSize(idx);
-}
-
-int EngineWrapper::GetOutputSizeByName(const char* name) {
-    int idx = GetOutputIndex(name);
-    if (idx < 0) return -1;
-    return GetOutputSize(idx);
-}
-
-int EngineWrapper::GetInputIndex(const char* name) {
-    auto it = m_input_name_to_idx.find(std::string(name));
-    return (it != m_input_name_to_idx.end()) ? it->second : -1;
-}
-
-int EngineWrapper::GetOutputIndex(const char* name) {
-    auto it = m_output_name_to_idx.find(std::string(name));
-    return (it != m_output_name_to_idx.end()) ? it->second : -1;
-}
-
-int EngineWrapper::Release() {
-    if (m_runner) {
-        m_runner->final();
-        m_runner.reset();
-    }
-    m_hasInit = false;
-    m_input_name_to_idx.clear();
-    m_output_name_to_idx.clear();
-    m_input_num = 0;
-    m_output_num = 0;
-    return 0;
-}
-
-const char* EngineWrapper::GetInputName(int index) const {
-    static thread_local std::string cached_name;
-    if (!m_hasInit || index < 0 || index >= m_input_num) return nullptr;
-    cached_name = m_runner->get_input_name(index);
-    return cached_name.c_str();
-}
-
-const char* EngineWrapper::GetOutputName(int index) const {
-    static thread_local std::string cached_name;
-    if (!m_hasInit || index < 0 || index >= m_output_num) return nullptr;
-    cached_name = m_runner->get_output_name(index);
-    return cached_name.c_str();
-}

cpp/src/EngineWrapper_axera.cpp

@@ -1,291 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * EngineWrapper implementation
- **************************************************************************************************/
-
-#include "EngineWrapper.hpp"
-#include "utils/io.hpp"
-
-#include <cstdlib>
-
-#if !defined(AX630C)
-static const char *strAlgoModelType[AX_ENGINE_MODEL_TYPE_BUTT] = {"3.6T", "7.2T", "10.8T"};
-#endif
-
-// NPU type enum
-typedef enum axNPU_TYPE_E {
-    AX_NPU_DEFAULT = 0,
-    AX_STD_VNPU_1 = (1 << 0),
-    AX_STD_VNPU_2 = (1 << 1),
-    AX_STD_VNPU_3 = (1 << 2),
-    AX_BL_VNPU_1  = (1 << 3),
-    AX_BL_VNPU_2  = (1 << 4)
-} AX_NPU_TYPE_E;
-
-#if !defined(AX630C)
-static AX_S32 CheckModelVNpu(const std::string &strModel,
-                              const AX_ENGINE_MODEL_TYPE_T &eModelType,
-                              const AX_S32 &nNpuType, AX_U32 &nNpuSet) {
-    AX_ENGINE_NPU_ATTR_T stNpuAttr;
-    memset(&stNpuAttr, 0x00, sizeof(stNpuAttr));
-
-    auto ret = AX_ENGINE_GetVNPUAttr(&stNpuAttr);
-    if (ret == 0) {
-        if (stNpuAttr.eHardMode == AX_ENGINE_VIRTUAL_NPU_DISABLE) {
-            nNpuSet = 0x01;
-        } else if (stNpuAttr.eHardMode == AX_ENGINE_VIRTUAL_NPU_STD) {
-            if (eModelType == AX_ENGINE_MODEL_TYPE1 || eModelType == AX_ENGINE_MODEL_TYPE2)
-                return -1;
-            if (nNpuType == 0) nNpuSet = 0x02;
-            else {
-                if (nNpuType & AX_STD_VNPU_1) nNpuSet |= 0x01;
-                if (nNpuType & AX_STD_VNPU_2) nNpuSet |= 0x02;
-                if (nNpuType & AX_STD_VNPU_3) nNpuSet |= 0x04;
-            }
-        } else if (stNpuAttr.eHardMode == AX_ENGINE_VIRTUAL_NPU_BIG_LITTLE) {
-            if (eModelType == AX_ENGINE_MODEL_TYPE2) return -1;
-            if (nNpuType == 0) {
-                nNpuSet = (eModelType == AX_ENGINE_MODEL_TYPE1) ? 0x01 : 0x02;
-            } else {
-                if (eModelType == AX_ENGINE_MODEL_TYPE1) {
-                    if (nNpuType & AX_BL_VNPU_2) return -1;
-                    if (nNpuType & AX_BL_VNPU_1) nNpuSet |= 0x01;
-                } else {
-                    if (nNpuType & AX_BL_VNPU_1) nNpuSet |= 0x01;
-                    if (nNpuType & AX_BL_VNPU_2) nNpuSet |= 0x02;
-                }
-            }
-        }
-    }
-    return ret;
-}
-#endif
-
-EngineWrapper::EngineWrapper()
-    : m_hasInit(false), m_handle(nullptr), m_io_info(nullptr),
-      m_input_num(0), m_output_num(0) {
-    memset(&m_io, 0, sizeof(m_io));
-}
-
-EngineWrapper::~EngineWrapper() {
-    Release();
-}
-
-int EngineWrapper::Init(const char* strModelPath, uint32_t nNpuType, const char* axclConfig) {
-    (void)axclConfig;  // unused on AXERA
-    // 1. Load model
-    AX_BOOL bLoadModelUseCmm = AX_TRUE;
-    AX_CHAR *pModelBufferVirAddr = nullptr;
-    AX_U64 u64ModelBufferPhyAddr = 0;
-    AX_U32 nModelBufferSize = 0;
-
-    if (bLoadModelUseCmm) {
-        if (!utils::read_file(strModelPath, (AX_VOID **)&pModelBufferVirAddr,
-                               u64ModelBufferPhyAddr, nModelBufferSize)) {
-            printf("Read model(%s) fail\n", strModelPath);
-            return -1;
-        }
-    } else {
-        std::vector<char> model_buffer;
-        if (!utils::read_file(strModelPath, model_buffer)) {
-            printf("Read model(%s) fail\n", strModelPath);
-            return -1;
-        }
-        pModelBufferVirAddr = model_buffer.data();
-        nModelBufferSize = model_buffer.size();
-    }
-
-    auto freeModelBuffer = [&]() {
-        if (bLoadModelUseCmm) {
-            if (u64ModelBufferPhyAddr != 0)
-                AX_SYS_MemFree(u64ModelBufferPhyAddr, &pModelBufferVirAddr);
-        }
-    };
-
-    // 1.1 Get Model Type & Check VNPU (AX650 only)
-#if !defined(AX630C)
-    AX_ENGINE_MODEL_TYPE_T eModelType = AX_ENGINE_MODEL_TYPE0;
-    AX_S32 ret = AX_ENGINE_GetModelType(pModelBufferVirAddr, nModelBufferSize, &eModelType);
-    if (0 != ret || eModelType >= AX_ENGINE_MODEL_TYPE_BUTT) {
-        printf("%s AX_ENGINE_GetModelType fail ret=%x\n", strModelPath, ret);
-        freeModelBuffer();
-        return -1;
-    }
-
-    AX_U32 nNpuSet = 0;
-    ret = CheckModelVNpu(strModelPath, eModelType, nNpuType, nNpuSet);
-    if (0 != ret) {
-        printf("CheckModelVNpu fail\n");
-        freeModelBuffer();
-        return -1;
-    }
-#endif
-#if defined(AX630C)
-    AX_S32 ret = 0;
-#endif
-
-    // 2. Create handle
-    AX_ENGINE_HANDLE handle = nullptr;
-    ret = AX_ENGINE_CreateHandle(&handle, pModelBufferVirAddr, nModelBufferSize);
-    freeModelBuffer();
-
-    auto deinit_handle = [&handle]() {
-        if (handle) { AX_ENGINE_DestroyHandle(handle); }
-        return -1;
-    };
-
-    if (0 != ret || !handle) {
-        printf("Create model(%s) handle fail\n", strModelPath);
-        return deinit_handle();
-    }
-
-    // 3. Create context
-    ret = AX_ENGINE_CreateContext(handle);
-    if (0 != ret) return deinit_handle();
-
-    // 4. Get IO info
-    m_io_info = nullptr;
-    ret = AX_ENGINE_GetIOInfo(handle, &m_io_info);
-    if (0 != ret) return deinit_handle();
-
-    m_input_num = m_io_info->nInputSize;
-    m_output_num = m_io_info->nOutputSize;
-
-    // Build name-to-index maps
-    m_input_name_to_idx.clear();
-    for (int i = 0; i < m_input_num; ++i) {
-        m_input_name_to_idx[std::string(m_io_info->pInputs[i].pName)] = i;
-    }
-    m_output_name_to_idx.clear();
-    for (int i = 0; i < m_output_num; ++i) {
-        m_output_name_to_idx[std::string(m_io_info->pOutputs[i].pName)] = i;
-    }
-
-    // 5. Prepare IO buffers
-    ret = utils::prepare_io("enc", m_io_info, m_io, utils::IO_BUFFER_STRATEGY_DEFAULT);
-    if (0 != ret) {
-        printf("prepare io failed!\n");
-        utils::free_io(m_io);
-        return deinit_handle();
-    }
-
-    m_handle = handle;
-    m_hasInit = true;
-    return 0;
-}
-
-int EngineWrapper::SetInput(void* pInput, int index) {
-    if (!m_hasInit || index < 0 || index >= m_input_num) return -1;
-    return utils::push_io_input(pInput, index, m_io);
-}
-
-int EngineWrapper::RunSync() {
-    if (!m_hasInit) return -1;
-    auto ret = AX_ENGINE_RunSync(m_handle, &m_io);
-    if (0 != ret) {
-        printf("AX_ENGINE_RunSync failed. ret=0x%x\n", ret);
-    }
-    return ret;
-}
-
-int EngineWrapper::GetOutput(void* pOutput, int index) {
-    if (!m_hasInit || index < 0 || index >= m_output_num) return -1;
-    return utils::push_io_output(pOutput, index, m_io);
-}
-
-int EngineWrapper::SetInputByName(const char* name, void* pInput) {
-    auto it = m_input_name_to_idx.find(std::string(name));
-    if (it == m_input_name_to_idx.end()) {
-        printf("Input '%s' not found in model\n", name);
-        return -1;
-    }
-    return SetInput(pInput, it->second);
-}
-
-int EngineWrapper::GetOutputByName(const char* name, void* pOutput) {
-    auto it = m_output_name_to_idx.find(std::string(name));
-    if (it == m_output_name_to_idx.end()) {
-        printf("Output '%s' not found in model\n", name);
-        return -1;
-    }
-    return GetOutput(pOutput, it->second);
-}
-
-int EngineWrapper::GetInputSize(int index) {
-    if (index < 0 || index >= m_input_num) return -1;
-    return m_io.pInputs[index].nSize;
-}
-
-int EngineWrapper::GetOutputSize(int index) {
-    if (index < 0 || index >= m_output_num) return -1;
-    return m_io.pOutputs[index].nSize;
-}
-
-int EngineWrapper::GetInputSizeByName(const char* name) {
-    int idx = GetInputIndex(name);
-    if (idx < 0) return -1;
-    return GetInputSize(idx);
-}
-
-int EngineWrapper::GetOutputSizeByName(const char* name) {
-    int idx = GetOutputIndex(name);
-    if (idx < 0) return -1;
-    return GetOutputSize(idx);
-}
-
-int EngineWrapper::GetInputIndex(const char* name) {
-    auto it = m_input_name_to_idx.find(std::string(name));
-    return (it != m_input_name_to_idx.end()) ? it->second : -1;
-}
-
-int EngineWrapper::GetOutputIndex(const char* name) {
-    auto it = m_output_name_to_idx.find(std::string(name));
-    return (it != m_output_name_to_idx.end()) ? it->second : -1;
-}
-
-const char* EngineWrapper::GetInputName(int index) const {
-    if (index < 0 || index >= m_input_num) return nullptr;
-    return m_io_info->pInputs[index].pName;
-}
-
-const char* EngineWrapper::GetOutputName(int index) const {
-    if (index < 0 || index >= m_output_num) return nullptr;
-    return m_io_info->pOutputs[index].pName;
-}
-
-static const char* dtype_str(AX_ENGINE_DATA_TYPE_T t) {
-    switch (t) {
-        case AX_ENGINE_DT_FLOAT32: return "float32";
-        case AX_ENGINE_DT_FLOAT64: return "float64";
-        case AX_ENGINE_DT_SINT8:   return "sint8";
-        case AX_ENGINE_DT_UINT8:   return "uint8";
-        case AX_ENGINE_DT_SINT16:  return "sint16";
-        case AX_ENGINE_DT_UINT16:  return "uint16";
-        case AX_ENGINE_DT_SINT32:  return "sint32";
-        case AX_ENGINE_DT_UINT32:  return "uint32";
-        default: return "unknown";
-    }
-}
-
-const char* EngineWrapper::GetInputDtypeStr(int index) const {
-    if (index < 0 || index >= m_input_num) return "?";
-    return dtype_str(m_io_info->pInputs[index].eDataType);
-}
-
-const char* EngineWrapper::GetOutputDtypeStr(int index) const {
-    if (index < 0 || index >= m_output_num) return "?";
-    return dtype_str(m_io_info->pOutputs[index].eDataType);
-}
-
-int EngineWrapper::Release() {
-    if (m_handle) {
-        utils::free_io(m_io);
-        AX_ENGINE_DestroyHandle(m_handle);
-        m_handle = nullptr;
-    }
-    m_hasInit = false;
-    m_input_name_to_idx.clear();
-    m_output_name_to_idx.clear();
-    return 0;
-}

cpp/src/cmdline.hpp

@@ -1,732 +0,0 @@
-/*
-  Copyright (c) 2009, Hideyuki Tanaka
-  All rights reserved.
-
-  Redistribution and use in source and binary forms, with or without
-  modification, are permitted provided that the following conditions are met:
-  * Redistributions of source code must retain the above copyright
-  notice, this list of conditions and the following disclaimer.
-  * Redistributions in binary form must reproduce the above copyright
-  notice, this list of conditions and the following disclaimer in the
-  documentation and/or other materials provided with the distribution.
-  * Neither the name of the <organization> nor the
-  names of its contributors may be used to endorse or promote products
-  derived from this software without specific prior written permission.
-
-  THIS SOFTWARE IS PROVIDED BY <copyright holder> ''AS IS'' AND ANY
-  EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
-  WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
-  DISCLAIMED. IN NO EVENT SHALL <copyright holder> BE LIABLE FOR ANY
-  DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
-  (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
-  LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
-  ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-*/
-
-#pragma once
-
-#include <cxxabi.h>
-
-#include <algorithm>
-#include <cstdlib>
-#include <cstring>
-#include <iostream>
-#include <map>
-#include <sstream>
-#include <stdexcept>
-#include <string>
-#include <typeinfo>
-#include <vector>
-
-namespace cmdline {
-
-namespace detail {
-
-template <typename Target, typename Source, bool Same>
-class lexical_cast_t {
- public:
-  static Target cast(const Source &arg) {
-    Target ret;
-    std::stringstream ss;
-    if (!(ss << arg && ss >> ret && ss.eof())) throw std::bad_cast();
-
-    return ret;
-  }
-};
-
-template <typename Target, typename Source>
-class lexical_cast_t<Target, Source, true> {
- public:
-  static Target cast(const Source &arg) { return arg; }
-};
-
-template <typename Source>
-class lexical_cast_t<std::string, Source, false> {
- public:
-  static std::string cast(const Source &arg) {
-    std::ostringstream ss;
-    ss << arg;
-    return ss.str();
-  }
-};
-
-template <typename Target>
-class lexical_cast_t<Target, std::string, false> {
- public:
-  static Target cast(const std::string &arg) {
-    Target ret;
-    std::istringstream ss(arg);
-    if (!(ss >> ret && ss.eof())) throw std::bad_cast();
-    return ret;
-  }
-};
-
-template <typename T1, typename T2>
-struct is_same {
-  static const bool value = false;
-};
-
-template <typename T>
-struct is_same<T, T> {
-  static const bool value = true;
-};
-
-template <typename Target, typename Source>
-Target lexical_cast(const Source &arg) {
-  return lexical_cast_t<Target, Source,
-                        detail::is_same<Target, Source>::value>::cast(arg);
-}
-
-static inline std::string demangle(const std::string &name) {
-  int status = 0;
-  char *p = abi::__cxa_demangle(name.c_str(), 0, 0, &status);
-  std::string ret(p);
-  free(p);
-  return ret;
-}
-
-template <class T>
-std::string readable_typename() {
-  return demangle(typeid(T).name());
-}
-
-template <class T>
-std::string default_value(T def) {
-  return detail::lexical_cast<std::string>(def);
-}
-
-template <>
-inline std::string readable_typename<std::string>() {
-  return "string";
-}
-
-}  // namespace detail
-
-//-----
-
-class cmdline_error : public std::exception {
- public:
-  cmdline_error(const std::string &msg) : msg(msg) {}
-  ~cmdline_error() throw() {}
-  const char *what() const throw() { return msg.c_str(); }
-
- private:
-  std::string msg;
-};
-
-template <class T>
-struct default_reader {
-  T operator()(const std::string &str) { return detail::lexical_cast<T>(str); }
-};
-
-template <class T>
-struct range_reader {
-  range_reader(const T &low, const T &high) : low(low), high(high) {}
-  T operator()(const std::string &s) const {
-    T ret = default_reader<T>()(s);
-    if (!(ret >= low && ret <= high))
-      throw cmdline::cmdline_error("range_error");
-    return ret;
-  }
-
- private:
-  T low, high;
-};
-
-template <class T>
-range_reader<T> range(const T &low, const T &high) {
-  return range_reader<T>(low, high);
-}
-
-template <class T>
-struct oneof_reader {
-  T operator()(const std::string &s) {
-    T ret = default_reader<T>()(s);
-    if (std::find(alt.begin(), alt.end(), ret) == alt.end())
-      throw cmdline_error("");
-    return ret;
-  }
-  void add(const T &v) { alt.push_back(v); }
-
- private:
-  std::vector<T> alt;
-};
-
-template <class T>
-oneof_reader<T> oneof(T a1) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  ret.add(a7);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  ret.add(a7);
-  ret.add(a8);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8, T a9) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  ret.add(a7);
-  ret.add(a8);
-  ret.add(a9);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8, T a9,
-                      T a10) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  ret.add(a7);
-  ret.add(a8);
-  ret.add(a9);
-  ret.add(a10);
-  return ret;
-}
-
-//-----
-
-class parser {
- public:
-  parser() {}
-  ~parser() {
-    for (std::map<std::string, option_base *>::iterator p = options.begin();
-         p != options.end(); p++)
-      delete p->second;
-  }
-
-  void add(const std::string &name, char short_name = 0,
-           const std::string &desc = "") {
-    if (options.count(name))
-      throw cmdline_error("multiple definition: " + name);
-    options[name] = new option_without_value(name, short_name, desc);
-    ordered.push_back(options[name]);
-  }
-
-  template <class T>
-  void add(const std::string &name, char short_name = 0,
-           const std::string &desc = "", bool need = true, const T def = T()) {
-    add(name, short_name, desc, need, def, default_reader<T>());
-  }
-
-  template <class T, class F>
-  void add(const std::string &name, char short_name = 0,
-           const std::string &desc = "", bool need = true, const T def = T(),
-           F reader = F()) {
-    if (options.count(name))
-      throw cmdline_error("multiple definition: " + name);
-    options[name] = new option_with_value_with_reader<T, F>(
-        name, short_name, need, def, desc, reader);
-    ordered.push_back(options[name]);
-  }
-
-  void footer(const std::string &f) { ftr = f; }
-
-  void set_program_name(const std::string &name) { prog_name = name; }
-
-  bool exist(const std::string &name) const {
-    if (options.count(name) == 0)
-      throw cmdline_error("there is no flag: --" + name);
-    return options.find(name)->second->has_set();
-  }
-
-  template <class T>
-  const T &get(const std::string &name) const {
-    if (options.count(name) == 0)
-      throw cmdline_error("there is no flag: --" + name);
-    const option_with_value<T> *p =
-        dynamic_cast<const option_with_value<T> *>(options.find(name)->second);
-    if (p == NULL) throw cmdline_error("type mismatch flag '" + name + "'");
-    return p->get();
-  }
-
-  const std::vector<std::string> &rest() const { return others; }
-
-  bool parse(const std::string &arg) {
-    std::vector<std::string> args;
-
-    std::string buf;
-    bool in_quote = false;
-    for (std::string::size_type i = 0; i < arg.length(); i++) {
-      if (arg[i] == '\"') {
-        in_quote = !in_quote;
-        continue;
-      }
-
-      if (arg[i] == ' ' && !in_quote) {
-        args.push_back(buf);
-        buf = "";
-        continue;
-      }
-
-      if (arg[i] == '\\') {
-        i++;
-        if (i >= arg.length()) {
-          errors.push_back("unexpected occurrence of '\\' at end of string");
-          return false;
-        }
-      }
-
-      buf += arg[i];
-    }
-
-    if (in_quote) {
-      errors.push_back("quote is not closed");
-      return false;
-    }
-
-    if (buf.length() > 0) args.push_back(buf);
-
-    for (size_t i = 0; i < args.size(); i++)
-      std::cout << "\"" << args[i] << "\"" << std::endl;
-
-    return parse(args);
-  }
-
-  bool parse(const std::vector<std::string> &args) {
-    int argc = static_cast<int>(args.size());
-    std::vector<const char *> argv(argc);
-
-    for (int i = 0; i < argc; i++) argv[i] = args[i].c_str();
-
-    return parse(argc, &argv[0]);
-  }
-
-  bool parse(int argc, const char *const argv[]) {
-    errors.clear();
-    others.clear();
-
-    if (argc < 1) {
-      errors.push_back("argument number must be longer than 0");
-      return false;
-    }
-    if (prog_name == "") prog_name = argv[0];
-
-    std::map<char, std::string> lookup;
-    for (std::map<std::string, option_base *>::iterator p = options.begin();
-         p != options.end(); p++) {
-      if (p->first.length() == 0) continue;
-      char initial = p->second->short_name();
-      if (initial) {
-        if (lookup.count(initial) > 0) {
-          lookup[initial] = "";
-          errors.push_back(std::string("short option '") + initial +
-                           "' is ambiguous");
-          return false;
-        } else
-          lookup[initial] = p->first;
-      }
-    }
-
-    for (int i = 1; i < argc; i++) {
-      if (strncmp(argv[i], "--", 2) == 0) {
-        const char *p = strchr(argv[i] + 2, '=');
-        if (p) {
-          std::string name(argv[i] + 2, p);
-          std::string val(p + 1);
-          set_option(name, val);
-        } else {
-          std::string name(argv[i] + 2);
-          if (options.count(name) == 0) {
-            errors.push_back("undefined option: --" + name);
-            continue;
-          }
-          if (options[name]->has_value()) {
-            if (i + 1 >= argc) {
-              errors.push_back("option needs value: --" + name);
-              continue;
-            } else {
-              i++;
-              set_option(name, argv[i]);
-            }
-          } else {
-            set_option(name);
-          }
-        }
-      } else if (strncmp(argv[i], "-", 1) == 0) {
-        if (!argv[i][1]) continue;
-        char last = argv[i][1];
-        for (int j = 2; argv[i][j]; j++) {
-          last = argv[i][j];
-          if (lookup.count(argv[i][j - 1]) == 0) {
-            errors.push_back(std::string("undefined short option: -") +
-                             argv[i][j - 1]);
-            continue;
-          }
-          if (lookup[argv[i][j - 1]] == "") {
-            errors.push_back(std::string("ambiguous short option: -") +
-                             argv[i][j - 1]);
-            continue;
-          }
-          set_option(lookup[argv[i][j - 1]]);
-        }
-
-        if (lookup.count(last) == 0) {
-          errors.push_back(std::string("undefined short option: -") + last);
-          continue;
-        }
-        if (lookup[last] == "") {
-          errors.push_back(std::string("ambiguous short option: -") + last);
-          continue;
-        }
-
-        if (i + 1 < argc && options[lookup[last]]->has_value()) {
-          set_option(lookup[last], argv[i + 1]);
-          i++;
-        } else {
-          set_option(lookup[last]);
-        }
-      } else {
-        others.push_back(argv[i]);
-      }
-    }
-
-    for (std::map<std::string, option_base *>::iterator p = options.begin();
-         p != options.end(); p++)
-      if (!p->second->valid())
-        errors.push_back("need option: --" + std::string(p->first));
-
-    return errors.size() == 0;
-  }
-
-  void parse_check(const std::string &arg) {
-    if (!options.count("help")) add("help", '?', "print this message");
-    check(0, parse(arg));
-  }
-
-  void parse_check(const std::vector<std::string> &args) {
-    if (!options.count("help")) add("help", '?', "print this message");
-    check(args.size(), parse(args));
-  }
-
-  void parse_check(int argc, char *argv[]) {
-    if (!options.count("help")) add("help", '?', "print this message");
-    check(argc, parse(argc, argv));
-  }
-
-  std::string error() const { return errors.size() > 0 ? errors[0] : ""; }
-
-  std::string error_full() const {
-    std::ostringstream oss;
-    for (size_t i = 0; i < errors.size(); i++) oss << errors[i] << std::endl;
-    return oss.str();
-  }
-
-  std::string usage() const {
-    std::ostringstream oss;
-    oss << "usage: " << prog_name << " ";
-    for (size_t i = 0; i < ordered.size(); i++) {
-      if (ordered[i]->must()) oss << ordered[i]->short_description() << " ";
-    }
-
-    oss << "[options] ... " << ftr << std::endl;
-    oss << "options:" << std::endl;
-
-    size_t max_width = 0;
-    for (size_t i = 0; i < ordered.size(); i++) {
-      max_width = std::max(max_width, ordered[i]->name().length());
-    }
-    for (size_t i = 0; i < ordered.size(); i++) {
-      if (ordered[i]->short_name()) {
-        oss << "  -" << ordered[i]->short_name() << ", ";
-      } else {
-        oss << "      ";
-      }
-
-      oss << "--" << ordered[i]->name();
-      for (size_t j = ordered[i]->name().length(); j < max_width + 4; j++)
-        oss << ' ';
-      oss << ordered[i]->description() << std::endl;
-    }
-    return oss.str();
-  }
-
- private:
-  void check(int argc, bool ok) {
-    if ((argc == 1 && !ok) || exist("help")) {
-      std::cerr << usage();
-      exit(0);
-    }
-
-    if (!ok) {
-      std::cerr << error() << std::endl << usage();
-      exit(1);
-    }
-  }
-
-  void set_option(const std::string &name) {
-    if (options.count(name) == 0) {
-      errors.push_back("undefined option: --" + name);
-      return;
-    }
-    if (!options[name]->set()) {
-      errors.push_back("option needs value: --" + name);
-      return;
-    }
-  }
-
-  void set_option(const std::string &name, const std::string &value) {
-    if (options.count(name) == 0) {
-      errors.push_back("undefined option: --" + name);
-      return;
-    }
-    if (!options[name]->set(value)) {
-      errors.push_back("option value is invalid: --" + name + "=" + value);
-      return;
-    }
-  }
-
-  class option_base {
-   public:
-    virtual ~option_base() {}
-
-    virtual bool has_value() const = 0;
-    virtual bool set() = 0;
-    virtual bool set(const std::string &value) = 0;
-    virtual bool has_set() const = 0;
-    virtual bool valid() const = 0;
-    virtual bool must() const = 0;
-
-    virtual const std::string &name() const = 0;
-    virtual char short_name() const = 0;
-    virtual const std::string &description() const = 0;
-    virtual std::string short_description() const = 0;
-  };
-
-  class option_without_value : public option_base {
-   public:
-    option_without_value(const std::string &name, char short_name,
-                         const std::string &desc)
-        : nam(name), snam(short_name), desc(desc), has(false) {}
-    ~option_without_value() {}
-
-    bool has_value() const { return false; }
-
-    bool set() {
-      has = true;
-      return true;
-    }
-
-    bool set(const std::string &) { return false; }
-
-    bool has_set() const { return has; }
-
-    bool valid() const { return true; }
-
-    bool must() const { return false; }
-
-    const std::string &name() const { return nam; }
-
-    char short_name() const { return snam; }
-
-    const std::string &description() const { return desc; }
-
-    std::string short_description() const { return "--" + nam; }
-
-   private:
-    std::string nam;
-    char snam;
-    std::string desc;
-    bool has;
-  };
-
-  template <class T>
-  class option_with_value : public option_base {
-   public:
-    option_with_value(const std::string &name, char short_name, bool need,
-                      const T &def, const std::string &desc)
-        : nam(name),
-          snam(short_name),
-          need(need),
-          has(false),
-          def(def),
-          actual(def) {
-      this->desc = full_description(desc);
-    }
-    ~option_with_value() {}
-
-    const T &get() const { return actual; }
-
-    bool has_value() const { return true; }
-
-    bool set() { return false; }
-
-    bool set(const std::string &value) {
-      try {
-        actual = read(value);
-        has = true;
-      } catch (const std::exception &e) {
-        return false;
-      }
-      return true;
-    }
-
-    bool has_set() const { return has; }
-
-    bool valid() const {
-      if (need && !has) return false;
-      return true;
-    }
-
-    bool must() const { return need; }
-
-    const std::string &name() const { return nam; }
-
-    char short_name() const { return snam; }
-
-    const std::string &description() const { return desc; }
-
-    std::string short_description() const {
-      return "--" + nam + "=" + detail::readable_typename<T>();
-    }
-
-   protected:
-    std::string full_description(const std::string &desc) {
-      return desc + " (" + detail::readable_typename<T>() +
-             (need ? "" : " [=" + detail::default_value<T>(def) + "]") + ")";
-    }
-
-    virtual T read(const std::string &s) = 0;
-
-    std::string nam;
-    char snam;
-    bool need;
-    std::string desc;
-
-    bool has;
-    T def;
-    T actual;
-  };
-
-  template <class T, class F>
-  class option_with_value_with_reader : public option_with_value<T> {
-   public:
-    option_with_value_with_reader(const std::string &name, char short_name,
-                                  bool need, const T def,
-                                  const std::string &desc, F reader)
-        : option_with_value<T>(name, short_name, need, def, desc),
-          reader(reader) {}
-
-   private:
-    T read(const std::string &s) { return reader(s); }
-
-    F reader;
-  };
-
-  std::map<std::string, option_base *> options;
-  std::vector<option_base *> ordered;
-  std::string ftr;
-
-  std::string prog_name;
-  std::vector<std::string> others;
-
-  std::vector<std::string> errors;
-};
-
-}  // namespace cmdline

cpp/src/fbank.cpp

@@ -1,219 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * FBANK implementation with simple DFT-based STFT.
- * For production, use FFTW or an NPU-accelerated implementation.
- **************************************************************************************************/
-
-#include "fbank.hpp"
-
-#include <algorithm>
-#include <cstring>
-#include <cstdio>
-
-#ifndef M_PI
-#define M_PI 3.14159265358979323846
-#endif
-
-static float hz_to_mel(float freq) {
-    return 2595.0f * std::log10(1.0f + freq / 700.0f);
-}
-
-static float mel_to_hz(float mel) {
-    return 700.0f * (std::pow(10.0f, mel / 2595.0f) - 1.0f);
-}
-
-MelFilterBank::MelFilterBank() : m_n_freqs(0) {}
-
-int MelFilterBank::Init(const Config& config) {
-    m_config = config;
-    m_n_freqs = config.n_fft / 2 + 1;
-
-    CreateWindow();
-    CreateMelFilterbank();
-
-    printf("FBANK init: n_mels=%d, n_fft=%d, hop=%d, sr=%d\n",
-           config.n_mels, config.n_fft, config.hop_length, config.sampling_rate);
-    return 0;
-}
-
-void MelFilterBank::CreateWindow() {
-    m_window.resize(m_config.n_fft);
-    for (int i = 0; i < m_config.n_fft; ++i) {
-        m_window[i] = 0.5f * (1.0f - std::cos(2.0f * M_PI * i / (m_config.n_fft - 1)));
-    }
-}
-
-void MelFilterBank::CreateMelFilterbank() {
-    int n_freqs = m_n_freqs;
-    int n_mels = m_config.n_mels;
-
-    // Frequency points for each FFT bin
-    std::vector<float> all_freqs(n_freqs);
-    for (int i = 0; i < n_freqs; ++i) {
-        all_freqs[i] = i * m_config.sampling_rate / (2.0f * (n_freqs - 1));
-    }
-
-    // Mel points
-    float m_min = hz_to_mel(0.0f);
-    float m_max = hz_to_mel(m_config.sampling_rate / 2.0f);
-    std::vector<float> m_pts(n_mels + 2);
-    for (int i = 0; i < n_mels + 2; ++i) {
-        m_pts[i] = m_min + (m_max - m_min) * i / (n_mels + 1);
-    }
-
-    // Convert to Hz
-    std::vector<float> f_pts(n_mels + 2);
-    for (int i = 0; i < n_mels + 2; ++i) {
-        f_pts[i] = mel_to_hz(m_pts[i]);
-    }
-
-    // Build filterbank matrix [n_freqs, n_mels]
-    m_mel_basis.resize(n_freqs * n_mels, 0.0f);
-
-    for (int m = 0; m < n_mels; ++m) {
-        for (int k = 0; k < n_freqs; ++k) {
-            float freq = all_freqs[k];
-            float left = f_pts[m];
-            float center = f_pts[m + 1];
-            float right = f_pts[m + 2];
-
-            float val = 0.0f;
-            if (freq >= left && freq <= center) {
-                val = (freq - left) / (center - left);
-            } else if (freq >= center && freq <= right) {
-                val = (right - freq) / (right - center);
-            }
-            m_mel_basis[k * n_mels + m] = val;
-        }
-    }
-}
-
-int MelFilterBank::ComputeNumFrames(int num_samples, int hop_length) {
-    return (num_samples + hop_length / 2) / hop_length;
-}
-
-// Radix-2 FFT (in-place, complex)
-static void fft(std::vector<float>& real, std::vector<float>& imag, bool inverse = false) {
-    int n = (int)real.size();
-    // Bit-reversal permutation
-    for (int i = 1, j = 0; i < n; ++i) {
-        int bit = n >> 1;
-        for (; j & bit; bit >>= 1) j ^= bit;
-        j ^= bit;
-        if (i < j) {
-            std::swap(real[i], real[j]);
-            std::swap(imag[i], imag[j]);
-        }
-    }
-    // Butterfly
-    for (int len = 2; len <= n; len <<= 1) {
-        float angle = 2.0f * M_PI / len * (inverse ? 1.0f : -1.0f);
-        float w_real = std::cos(angle);
-        float w_imag = std::sin(angle);
-        for (int i = 0; i < n; i += len) {
-            float cur_real = 1.0f, cur_imag = 0.0f;
-            for (int j = 0; j < len / 2; ++j) {
-                int a = i + j;
-                int b = i + j + len / 2;
-                float t_real = cur_real * real[b] - cur_imag * imag[b];
-                float t_imag = cur_real * imag[b] + cur_imag * real[b];
-                real[b] = real[a] - t_real;
-                imag[b] = imag[a] - t_imag;
-                real[a] += t_real;
-                imag[a] += t_imag;
-                float next_real = cur_real * w_real - cur_imag * w_imag;
-                cur_imag = cur_real * w_imag + cur_imag * w_real;
-                cur_real = next_real;
-            }
-        }
-    }
-    if (inverse) {
-        for (int i = 0; i < n; ++i) {
-            real[i] /= n;
-            imag[i] /= n;
-        }
-    }
-}
-
-void MelFilterBank::ComputeSTFT(const std::vector<float>& samples,
-                                 std::vector<float>& spec_real,
-                                 std::vector<float>& spec_imag,
-                                 int& num_frames) {
-    int n_fft = m_config.n_fft;
-    int hop = m_config.hop_length;
-    int num_samples = static_cast<int>(samples.size());
-
-    num_frames = ComputeNumFrames(num_samples, hop);
-    int n_freqs = n_fft / 2 + 1;
-
-    spec_real.assign(num_frames * n_freqs, 0.0f);
-    spec_imag.assign(num_frames * n_freqs, 0.0f);
-
-    // Pad signal (center padding, matching Python torch.stft center=True)
-    int pad_amount = n_fft / 2;
-    std::vector<float> padded(pad_amount + num_samples + pad_amount, 0.0f);
-    std::copy(samples.begin(), samples.end(), padded.begin() + pad_amount);
-
-    // FFT workspace
-    std::vector<float> fft_real(n_fft), fft_imag(n_fft);
-
-    for (int frame = 0; frame < num_frames; ++frame) {
-        int start = frame * hop;
-        // Apply window and copy to FFT buffer
-        for (int n = 0; n < n_fft; ++n) {
-            fft_real[n] = padded[start + n] * m_window[n];
-            fft_imag[n] = 0.0f;
-        }
-        // Forward FFT
-        fft(fft_real, fft_imag, false);
-        // Extract first n_freqs bins
-        for (int k = 0; k < n_freqs; ++k) {
-            spec_real[frame * n_freqs + k] = fft_real[k];
-            spec_imag[frame * n_freqs + k] = fft_imag[k];
-        }
-    }
-}
-
-std::vector<float> MelFilterBank::Extract(const std::vector<float>& samples, int sample_rate) {
-    if (sample_rate != m_config.sampling_rate) {
-        printf("WARNING: sample_rate mismatch: expected %d, got %d\n",
-               m_config.sampling_rate, sample_rate);
-    }
-
-    int num_frames;
-    std::vector<float> spec_real, spec_imag;
-    ComputeSTFT(samples, spec_real, spec_imag, num_frames);
-
-    int n_freqs = m_n_freqs;
-    int n_mels = m_config.n_mels;
-
-    // Compute magnitude spectrogram [num_frames, n_freqs]
-    std::vector<float> spec_mag(num_frames * n_freqs);
-    for (int i = 0; i < num_frames * n_freqs; ++i) {
-        spec_mag[i] = std::sqrt(spec_real[i] * spec_real[i] + spec_imag[i] * spec_imag[i]);
-    }
-
-    // Apply mel filterbank: mel[frame, m] = sum_freq(spec[frame, freq] * basis[freq, m])
-    std::vector<float> mel(num_frames * n_mels, 0.0f);
-    for (int f = 0; f < num_frames; ++f) {
-        for (int m = 0; m < n_mels; ++m) {
-            float sum = 0.0f;
-            for (int k = 0; k < n_freqs; ++k) {
-                sum += spec_mag[f * n_freqs + k] * m_mel_basis[k * n_mels + m];
-            }
-            mel[f * n_mels + m] = std::log(std::max(sum, 1e-7f));
-        }
-    }
-
-    // Expected num_frames
-    int expected_frames = ComputeNumFrames(static_cast<int>(samples.size()), m_config.hop_length);
-    if (num_frames > expected_frames) {
-        // Truncate
-        std::vector<float> truncated(expected_frames * n_mels);
-        std::copy(mel.begin(), mel.begin() + expected_frames * n_mels, truncated.begin());
-        return truncated;
-    }
-
-    return mel;
-}

cpp/src/fbank.hpp

@@ -1,65 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * FBANK: Mel filterbank feature extraction for prompt audio.
- *
- * Extracts log-mel spectrogram features from audio waveform, matching the
- * Python LocalVocosFbank behavior.
- **************************************************************************************************/
-
-#pragma once
-
-#include <vector>
-#include <cmath>
-#include <cstdint>
-
-class MelFilterBank {
-public:
-    struct Config {
-        int sampling_rate;
-        int n_mels;
-        int n_fft;
-        int hop_length;
-
-        Config() : sampling_rate(24000), n_mels(100), n_fft(1024), hop_length(256) {}
-    };
-
-    MelFilterBank();
-
-    /**
-     * Initialize the filterbank with given config.
-     */
-    int Init(const Config& config = Config());
-
-    /**
-     * Extract log-mel features from float32 audio samples [num_samples].
-     * Returns features of shape [num_frames, n_mels].
-     */
-    std::vector<float> Extract(const std::vector<float>& samples, int sample_rate);
-
-    /**
-     * Get the number of frames for given number of samples.
-     */
-    static int ComputeNumFrames(int num_samples, int hop_length);
-
-    const Config& GetConfig() const { return m_config; }
-
-private:
-    Config m_config;
-
-    // Mel filterbank matrix [n_freqs, n_mels]
-    std::vector<float> m_mel_basis;
-    int m_n_freqs;
-
-    // Hann window
-    std::vector<float> m_window;
-
-    void CreateMelFilterbank();
-    void CreateWindow();
-
-    // Simple STFT
-    void ComputeSTFT(const std::vector<float>& samples,
-                     std::vector<float>& spec_real,
-                     std::vector<float>& spec_imag,
-                     int& num_frames);
-};

cpp/src/middleware/ax_npu_runner.cc

@@ -1,131 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#include "middleware/ax_npu_runner_impl.hpp"
-
-#if defined(ENV_AX_LOCAL_API_ENABLE)
-middleware::npu_runner::npu_runner() {
-    this->impl_ = new impl();
-}
-
-middleware::npu_runner::~npu_runner() {
-    delete this->impl_;
-}
-
-bool middleware::npu_runner::init(const std::string& config_file, const uint32_t& index, const uint32_t& kind) {
-    std::ignore = config_file;
-    std::ignore = index;
-    return this->impl_->init(kind);
-}
-
-bool middleware::npu_runner::final() {
-    return this->impl_->final();
-}
-
-bool middleware::npu_runner::load(const std::string& model_path) {
-    return this->impl_->load(model_path);
-}
-
-bool middleware::npu_runner::prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) {
-    return this->impl_->prepare(input_cached, output_cached, group, batch);
-}
-
-bool middleware::npu_runner::run(const bool& parallel) {
-    return this->impl_->run(parallel);
-}
-
-uint32_t middleware::npu_runner::get_input_count() const {
-    return this->impl_->get_input_count();
-}
-
-uint32_t middleware::npu_runner::get_output_count() const {
-    return this->impl_->get_output_count();
-}
-
-std::string middleware::npu_runner::get_input_name(const uint32_t& index) const {
-    return this->impl_->get_input_name(index);
-}
-
-std::string middleware::npu_runner::get_output_name(const uint32_t& index) const {
-    return this->impl_->get_output_name(index);
-}
-
-void *middleware::npu_runner::get_input_pointer(const uint32_t& index) const {
-    return this->impl_->get_input_pointer(index);
-}
-
-void *middleware::npu_runner::get_output_pointer(const uint32_t& index) const {
-    return this->impl_->get_output_pointer(index);
-}
-
-uintmax_t middleware::npu_runner::get_input_size(const uint32_t& index) const {
-    return this->impl_->get_input_size(index);
-}
-
-uintmax_t middleware::npu_runner::get_output_size(const uint32_t& index) const {
-    return this->impl_->get_output_size(index);
-}
-
-uintmax_t middleware::npu_runner::get_shape_group_count() const {
-    return this->impl_->get_shape_group_count();
-}
-
-bool middleware::npu_runner::flush_input() const {
-    return this->impl_->flush_input();
-}
-
-bool middleware::npu_runner::invalidate_output() const {
-    return this->impl_->invalidate_output();
-}
-
-bool middleware::npu_runner::feed(const std::string& input_folder, const std::string& stimulus_name) const {
-    return this->impl_->feed(input_folder, stimulus_name);
-}
-
-bool middleware::npu_runner::verify(const std::string& output_folder, const std::string& stimulus_name) const {
-    return this->impl_->verify(output_folder, stimulus_name);
-}
-
-bool middleware::npu_runner::save(const std::string& output_folder, const std::string& stimulus_name) const {
-    return this->impl_->save(output_folder, stimulus_name);
-}
-
-std::string middleware::npu_runner::get_library_version() const {
-    return this->impl_->get_library_version();
-}
-
-std::string middleware::npu_runner::get_model_version() const {
-    return this->impl_->get_model_version();
-}
-
-int32_t middleware::npu_runner::get_model_type() const {
-    return this->impl_->get_model_type();
-}
-
-int32_t middleware::npu_runner::get_npu_type() const {
-    return this->impl_->get_npu_type();
-}
-
-int32_t middleware::npu_runner::get_batch_size() const {
-    return this->impl_->get_batch_size();
-}
-
-intmax_t middleware::npu_runner::get_sys_usage() const {
-    return this->impl_->get_sys_usage();
-}
-
-intmax_t middleware::npu_runner::get_cmm_usage() const {
-    return this->impl_->get_cmm_usage();
-}
-
-void middleware::npu_runner::sleep_for(const uint32_t sleep_duration) const {
-    impl::sleep_for(sleep_duration);
-}
-#endif

cpp/src/middleware/ax_npu_runner.hpp

@@ -1,65 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#if defined(ENV_AX_LOCAL_API_ENABLE)
-#include "middleware/runner.hpp"
-
-namespace middleware {
-
-class npu_runner final : public runner {
-public:
-    npu_runner();
-    ~npu_runner() override;
-
-    [[nodiscard]] bool init(const std::string& config_file, const uint32_t& index, const uint32_t& kind) override;
-    [[nodiscard]] bool final() override;
-
-    [[nodiscard]] bool load(const std::string& model_path) override;
-    [[nodiscard]] bool prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) override;
-
-    [[nodiscard]] bool run(const bool& parallel) override;
-
-    [[nodiscard]] uint32_t get_input_count() const override;
-    [[nodiscard]] uint32_t get_output_count() const override;
-    [[nodiscard]] std::string get_input_name(const uint32_t& index) const override;
-    [[nodiscard]] std::string get_output_name(const uint32_t& index) const override;
-    [[nodiscard]] void *get_input_pointer(const uint32_t& index) const override;
-    [[nodiscard]] void *get_output_pointer(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_input_size(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_output_size(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_shape_group_count() const override;
-
-    [[nodiscard]] bool flush_input() const override;
-    [[nodiscard]] bool invalidate_output() const override;
-
-    [[nodiscard]] bool feed(const std::string& input_folder, const std::string& stimulus_name) const override;
-    [[nodiscard]] bool verify(const std::string& output_folder, const std::string& stimulus_name) const override;
-    [[nodiscard]] bool save(const std::string& output_folder, const std::string& stimulus_name) const override;
-
-    [[nodiscard]] std::string get_library_version() const override;
-    [[nodiscard]] std::string get_model_version() const override;
-
-    [[nodiscard]] int32_t get_model_type() const override;
-    [[nodiscard]] int32_t get_npu_type() const override;
-    [[nodiscard]] int32_t get_batch_size() const override;
-    [[nodiscard]] intmax_t get_sys_usage() const override;
-    [[nodiscard]] intmax_t get_cmm_usage() const override;
-
-    void sleep_for(uint32_t sleep_duration) const override;
-
-private:
-    struct impl;
-    impl *impl_;
-};
-#endif
-
-}

cpp/src/middleware/ax_npu_runner_impl.hpp

@@ -1,640 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#if defined(ENV_AX_LOCAL_API_ENABLE)
-#include "middleware/ax_npu_runner.hpp"
-
-#include "utilities/scalar_guard.hpp"
-#include "utilities/file.hpp"
-#include "utilities/file_mapper.hpp"
-#include "utilities/log.hpp"
-
-#include <ax_hrtimer.h>
-#include <ax_sys_api.h>
-#include <ax_engine_api.h>
-
-#include <mutex>
-
-constexpr ::AX_U32 align_size = 128;
-constexpr ::AX_S8 token_name[] = "toolkit";
-
-struct middleware::npu_runner::impl {
-    impl() = default;
-
-    ~impl() {
-        std::ignore = this->final();
-    }
-
-    [[nodiscard]] bool init(const uint32_t& kind) {
-        std::lock_guard guard(this->mutex_);
-
-        // 0. check the NPU kind
-        if (kind >= ::AX_ENGINE_VIRTUAL_NPU_BUTT) {
-            utilities::glog.print(utilities::log::type::error, "Specified NPU kind{%d} is out of range{total %d}.\n", kind, ::AX_ENGINE_VIRTUAL_NPU_BUTT);
-            return false;
-        }
-
-        // 1. init ax system, using scalar_guard to ensure the finalization
-        auto sys_guard = utilities::scalar_guard<::AX_S32>(
-            ::AX_SYS_Init(),
-            [](const ::AX_S32& state) {
-                if (0 == state) {
-                    std::ignore = ::AX_SYS_Deinit();
-                }
-            });
-
-        // 2. check the initialization result
-        if (0 != sys_guard.get()) {
-            utilities::glog.print(utilities::log::type::error, "AX SYS init failed{0x%08X}.\n", sys_guard.get());
-            return false;
-        }
-        utilities::glog.print(utilities::log::type::info, "AX SYS inited.\n");
-
-        // 3. init ax engine
-        ::AX_ENGINE_NPU_ATTR_T attr{};
-        attr.eHardMode = static_cast<::AX_ENGINE_NPU_MODE_T>(kind);
-        if (const int ret = ::AX_ENGINE_Init(&attr); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Init AX ENGINE as kind{%d} failed{0x%08X}.\n", kind, ret);
-            return false;
-        }
-        utilities::glog.print(utilities::log::type::info, "AX ENGINE inited.\n");
-
-        // 4. disable guard of env
-        sys_guard.get() = -1;
-
-        // 5. set the flag
-        this->is_initialized_ = true;
-        return true;
-    }
-
-    [[nodiscard]] bool final() {
-        auto flag = true;
-
-        std::lock_guard guard(this->mutex_);
-
-        if (this->is_initialized_) {
-            if (nullptr != this->handle_) {
-                for (auto& input : inputs_) {
-                    if (0 != input.phyAddr) {
-                        flag &= 0 == ::AX_SYS_MemFree(input.phyAddr, input.pVirAddr);
-                        input.phyAddr = 0;
-                    }
-                }
-                for (auto& output : outputs_) {
-                    if (0 != output.phyAddr) {
-                        flag &= 0 == ::AX_SYS_MemFree(output.phyAddr, output.pVirAddr);
-                        output.phyAddr = 0;
-                    }
-                }
-
-                flag &= 0 == ::AX_ENGINE_DestroyHandle(this->handle_);
-            }
-
-            flag &= 0 == ::AX_ENGINE_Deinit();
-            flag &= 0 == ::AX_SYS_Deinit();
-            this->is_initialized_ = false;
-        }
-
-        return flag;
-    }
-
-    [[nodiscard]] bool load(const std::string& model_path) {
-        std::lock_guard guard(this->mutex_);
-
-        if (!this->is_initialized_) {
-            utilities::glog.print(utilities::log::type::error, "axcl is not initialized.\n");
-            return false;
-        }
-
-        if (nullptr != this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model is already loaded.\n");
-            return false;
-        }
-
-        if (!utilities::exists(model_path)
-            || !utilities::is_regular_file(model_path)
-            || 0 == utilities::file_size(model_path)
-            || utilities::error_size == utilities::file_size(model_path)) {
-            utilities::glog.print(utilities::log::type::error, "Model file{%s} error, please check it.\n", model_path.c_str());
-            return false;
-        }
-
-        const auto model_map = utilities::file_mapper(model_path);
-        if (nullptr == model_map.get()) {
-            utilities::glog.print(utilities::log::type::error, "Model file{%s} mapping failed.\n", model_path.c_str());
-            return false;
-        }
-
-        const auto model_name = utilities::get_file_name(model_path);
-        ::AX_ENGINE_HANDLE_EXTRA_T extra_param{};
-        extra_param.pName = reinterpret_cast<::AX_S8*>(const_cast<char*>(model_name.c_str()));
-        if (const auto ret = ::AX_ENGINE_CreateHandleV2(&this->handle_, model_map.get(), model_map.size(), &extra_param); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Create model{%s} handle failed.\n", model_path.c_str());
-            return false;
-        }
-
-        if (const auto ret = ::AX_ENGINE_CreateContextV2(this->handle_, &this->context_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Create model{%s} context failed.\n", model_path.c_str());
-            return false;
-        }
-
-        return true;
-    }
-
-    [[nodiscard]] bool prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) {
-        std::lock_guard guard(this->mutex_);
-
-        // 0. check the handle
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is null.\n");
-            return false;
-        }
-
-        // 1. get the count of shape group
-        if (const auto ret = ::AX_ENGINE_GetGroupIOInfoCount(this->handle_, &this->group_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model shape group count failed{0x%08X}.\n", ret);
-            return false;
-        }
-
-        // 2. check the group index
-        if (group >= this->group_) {
-            utilities::glog.print(utilities::log::type::error, "Model group{%d} is out of range{total %d}.\n", group, this->group_);
-            return false;
-        }
-
-        // 3. get the IO info
-        if (1 == this->group_) {
-            if (const auto ret = ::AX_ENGINE_GetIOInfo(this->handle_, &info_); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error, "Get model IO info failed{0x%08X}.\n", ret);
-                return false;
-            }
-        } else {
-            if (const auto ret = ::AX_ENGINE_GetGroupIOInfo(this->handle_, group, &info_); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error, "Get model group{index: %d} IO info failed{0x%08X}.\n", group, ret);
-                return false;
-            }
-        }
-
-        // 4. check the batch size
-        this->batch = (0 == batch ? 1 : batch);
-        this->io_.nBatchSize = this->batch;
-
-        // 5. prepare the input and output
-        this->inputs_.resize(this->info_->nInputSize, ::AX_ENGINE_IO_BUFFER_T{});
-        this->io_.pInputs = this->inputs_.data();
-        this->io_.nInputSize = this->info_->nInputSize;
-
-        this->outputs_.resize(this->info_->nOutputSize, ::AX_ENGINE_IO_BUFFER_T{});
-        this->io_.pOutputs = this->outputs_.data();
-        this->io_.nOutputSize = this->info_->nOutputSize;
-
-        this->input_cached_.resize(inputs_.size(), input_cached);
-        this->output_cached_.resize(outputs_.size(), output_cached);
-
-        // 6. prepare the memory, inputs
-        for (::AX_U32 i = 0; i < this->info_->nInputSize; i++) {
-            const auto& meta = this->info_->pInputs[i];
-            auto& io = this->io_.pInputs[i];
-            io.nSize = meta.nSize * this->batch;
-
-            if (this->input_cached_[i]) {
-                if (const auto ret = ::AX_SYS_MemAllocCached(&io.phyAddr, &io.pVirAddr, io.nSize, align_size, token_name); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Memory allocation for tensor {%s} failed{0x%08X}.\n", meta.pName, ret);
-                    return false;
-                }
-                utilities::glog.print(utilities::log::type::info, "Memory for tensor {%s} is allocated.\n", meta.pName);
-            } else {
-                if (const auto ret = ::AX_SYS_MemAlloc(&io.phyAddr, &io.pVirAddr, io.nSize, align_size, token_name); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Memory allocation for tensor {%s} failed{0x%08X}.\n", meta.pName, ret);
-                    return false;
-                }
-                utilities::glog.print(utilities::log::type::info, "Memory for tensor {%s} is allocated.\n", meta.pName);
-            }
-
-            // clean memory, some cases model may need to clean memory
-            memset(io.pVirAddr, 0, io.nSize);
-
-            if (this->output_cached_[i]) {
-                if (const auto ret = ::AX_SYS_MflushCache(io.phyAddr, io.pVirAddr, io.nSize); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Memory flush cache for tensor {%s} failed{0x%08X}.\n", meta.pName, ret);
-                    return false;
-                }
-            }
-
-            utilities::glog.print(utilities::log::type::info, "Set tensor { name: %s, phy: %p, vir: %p, size: %u Bytes }.\n",
-                meta.pName, reinterpret_cast<void*>(io.phyAddr), io.pVirAddr, io.nSize);
-        }
-
-        // 7. prepare the memory, outputs
-        for (::AX_U32 i = 0; i < this->info_->nOutputSize; i++) {
-            const auto& meta = this->info_->pOutputs[i];
-            auto& io = this->io_.pOutputs[i];
-            io.nSize = meta.nSize * this->batch;
-
-            if (this->output_cached_[i]) {
-                if (const auto ret = ::AX_SYS_MemAllocCached(&io.phyAddr, &io.pVirAddr, io.nSize, align_size, token_name); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Memory allocation for tensor {%s} failed{0x%08X}.\n", meta.pName, ret);
-                    return false;
-                }
-                utilities::glog.print(utilities::log::type::info, "Memory for tensor {%s} is allocated.\n", meta.pName);
-            } else {
-                if (const auto ret = ::AX_SYS_MemAlloc(&io.phyAddr, &io.pVirAddr, io.nSize, align_size, token_name); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Memory allocation for tensor {%s} failed{0x%08X}.\n", meta.pName, ret);
-                    return false;
-                }
-                utilities::glog.print(utilities::log::type::info, "Memory for tensor {%s} is allocated.\n", meta.pName);
-            }
-
-            // clean memory, some cases model may need to clean memory
-            memset(io.pVirAddr, 0, io.nSize);
-
-            if (this->output_cached_[i]) {
-                if (const auto ret = ::AX_SYS_MflushCache(io.phyAddr, io.pVirAddr, io.nSize); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Memory flush cache for tensor {%s} failed{0x%08X}.\n", meta.pName, ret);
-                    return false;
-                }
-            }
-
-            utilities::glog.print(utilities::log::type::info, "Set tensor { name: %s, phy: %p, vir: %p, size: %u Bytes }.\n",
-                meta.pName, reinterpret_cast<void*>(io.phyAddr), io.pVirAddr, io.nSize);
-        }
-
-        return true;
-    }
-
-    [[nodiscard]] bool run(const bool& parallel) {
-        std::lock_guard guard(this->mutex_);
-
-        this->io_.nParallelRun = (parallel ? 1 : 0);
-
-        if (nullptr == this->handle_ || nullptr == this->context_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return false;
-        }
-
-        if (0 == this->io_.nInputSize || 0 == this->io_.nOutputSize) {
-            utilities::glog.print(utilities::log::type::error, "Model io is not set, prepare first.\n");
-            return false;
-        }
-
-        ::AX_S32 ret = 0;
-        if (1 == this->group_) {
-            ret = ::AX_ENGINE_RunSyncV2(this->handle_, this->context_, &this->io_);
-        } else {
-            ret = ::AX_ENGINE_RunGroupIOSync(this->handle_, this->context_, this->group_, &this->io_);
-        }
-        if (0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Run model failed{0x%08X}.\n", ret);
-            return false;
-        }
-        utilities::glog.print(utilities::log::type::info, "Running done.\n");
-
-        return true;
-    }
-
-    [[nodiscard]] uint32_t get_input_count() {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        return this->info_->nInputSize;
-    }
-
-    [[nodiscard]] uint32_t get_output_count() {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        return this->info_->nOutputSize;
-    }
-
-    [[nodiscard]] std::string get_input_name(const uint32_t& index) {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return {};
-        }
-
-        if (index >= this->info_->nInputSize) {
-            return {};
-        }
-        return std::string{this->info_->pInputs[index].pName};
-    }
-
-    [[nodiscard]] std::string get_output_name(const uint32_t& index) {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return {};
-        }
-
-        if (index >= this->info_->nOutputSize) {
-            return {};
-        }
-        return std::string{this->info_->pOutputs[index].pName};
-    }
-
-    [[nodiscard]] void *get_input_pointer(const uint32_t& index) {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return nullptr;
-        }
-
-        if (index >= this->info_->nInputSize) {
-            return nullptr;
-        }
-        return reinterpret_cast<void *>(this->io_.pInputs[index].phyAddr);
-    }
-
-    [[nodiscard]] void *get_output_pointer(const uint32_t& index) {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return nullptr;
-        }
-
-        if (index >= this->info_->nOutputSize) {
-            return nullptr;
-        }
-        return reinterpret_cast<void *>(this->io_.pOutputs[index].phyAddr);
-    }
-
-    [[nodiscard]] uintmax_t get_input_size(const uint32_t& index) {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        if (index >= this->info_->nInputSize) {
-            return 0;
-        }
-        return this->io_.pInputs[index].nSize;
-    }
-
-    [[nodiscard]] uintmax_t get_output_size(const uint32_t& index) {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        if (index >= this->info_->nOutputSize) {
-            return 0;
-        }
-        return this->io_.pOutputs[index].nSize;
-    }
-
-    [[nodiscard]] uintmax_t get_shape_group_count() {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        return this->group_;
-    }
-
-    [[nodiscard]] bool flush_input() {
-        std::lock_guard guard(this->mutex_);
-
-        for (::AX_U32 i = 0; i < this->io_.nInputSize; i++) {
-            const auto buffer = &(this->io_.pInputs[i]);
-            if (const auto ret = ::AX_SYS_MflushCache(buffer->phyAddr, buffer->pVirAddr, buffer->nSize); 0 != ret) {
-                utilities::glog.print(utilities::log::type::warn, "Flush input{index: %d} failed{0x%08X}.\n", i, ret);
-                return false;
-            }
-        }
-        return true;
-    }
-
-    [[nodiscard]] bool invalidate_output() {
-        std::lock_guard guard(this->mutex_);
-
-        for (::AX_U32 i = 0; i < this->io_.nOutputSize; i++) {
-            const auto buffer = &(this->io_.pOutputs[i]);
-            if (const auto ret = ::AX_SYS_MinvalidateCache(buffer->phyAddr, buffer->pVirAddr, buffer->nSize); 0 != ret) {
-                utilities::glog.print(utilities::log::type::warn, "Invalidate output{index: %d} failed{0x%08X}.\n", i, ret);
-                return false;
-            }
-        }
-        return true;
-    }
-
-    [[nodiscard]] bool feed(const std::string& input_folder, const std::string& stimulus_name) {
-        for (uint32_t i = 0; i < this->get_input_count(); i++) {
-            const auto file_name = utilities::get_legal_name(this->get_input_name(i));
-            auto file_path = input_folder;
-            file_path.append("/").append(stimulus_name).append("/").append(file_name).append(".bin");
-
-            if (!read(file_path, this->inputs_[i].pVirAddr, this->inputs_[i].nSize)) {
-                utilities::glog.print(utilities::log::type::error, "Read tensor {idx: %d, name: %s} file {%s} failed.\n", i, file_name.c_str(), file_path.c_str());
-                return false;
-            }
-
-            if (this->input_cached_[i]) {
-                if (const auto ret = ::AX_SYS_MflushCache(this->inputs_[i].phyAddr, this->inputs_[i].pVirAddr, this->inputs_[i].nSize); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Flush tensor {idx: %d, name: %s} cache failed{%d}.\n", i, file_name.c_str(), ret);
-                    return false;
-                }
-                utilities::glog.print(utilities::log::type::info, "Tensor {idx: %d, name: %s} cache flushed.\n", i, file_name.c_str());
-            }
-        }
-        return true;
-    }
-
-    [[nodiscard]] bool verify(const std::string& output_folder, const std::string& stimulus_name) {
-        for (uint32_t i = 0; i < this->get_output_count(); i++) {
-            const auto file_name = utilities::get_legal_name(this->get_output_name(i));
-            auto file_path = output_folder;
-            file_path.append("/").append(stimulus_name).append("/").append(file_name).append(".bin");
-
-            std::vector<uint8_t> file_buffer(this->get_output_size(i));
-            if (!read(file_path, file_buffer.data(), file_buffer.size())) {
-                utilities::glog.print(utilities::log::type::error, "Read tensor {idx: %d, name: %s} file {%s} failed.\n", i, file_name.c_str(), file_path.c_str());
-                return false;
-            }
-
-            if (this->output_cached_[i]) {
-                if (const auto ret = ::AX_SYS_MinvalidateCache(this->outputs_[i].phyAddr, this->outputs_[i].pVirAddr, this->outputs_[i].nSize); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Invalidate tensor {idx: %d, name: %s} cache failed{%d}.\n", i, file_name.c_str(), ret);
-                    return false;
-                }
-                utilities::glog.print(utilities::log::type::info, "Tensor {idx: %d, name: %s} cache invalidated.\n", i, file_name.c_str());
-            }
-
-            if (!runner::verify(file_buffer.data(), this->outputs_[i].pVirAddr, this->outputs_[i].nSize)) {
-                utilities::glog.print(utilities::log::type::error, "Verify tensor {idx: %d, name: %s} failed.\n", i, file_name.c_str());
-                return false;
-            }
-        }
-        return true;
-    }
-
-    [[nodiscard]] bool save(const std::string& output_folder, const std::string& stimulus_name) {
-        auto file_folder = output_folder;
-        file_folder.append("/").append(stimulus_name);
-
-        if (!utilities::exists(file_folder) && !utilities::create_directory(file_folder)) {
-            utilities::glog.print(utilities::log::type::error, "Create folder {%s} failed.\n", file_folder.c_str());
-            return false;
-        }
-
-        for (uint32_t i = 0; i < this->get_output_count(); i++) {
-            const auto file_name = utilities::get_legal_name(this->get_output_name(i));
-            auto file_path = file_folder;
-            file_path.append("/").append(file_name).append(".bin");
-
-            if (this->output_cached_[i]) {
-                if (const auto ret = ::AX_SYS_MinvalidateCache(this->outputs_[i].phyAddr, this->outputs_[i].pVirAddr, this->outputs_[i].nSize); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Invalidate tensor {idx: %d, name: %s} cache failed{%d}.\n", i, file_name.c_str(), ret);
-                    return false;
-                }
-                utilities::glog.print(utilities::log::type::info, "Tensor {idx: %d, name: %s} cache invalidated.\n", i, file_name.c_str());
-            }
-
-            if (!write(file_path, this->outputs_[i].pVirAddr, this->outputs_[i].nSize)) {
-                utilities::glog.print(utilities::log::type::error, "Write tensor {idx: %d, name: %s} file {%s} failed.\n", i, file_name.c_str(), file_path.c_str());
-                return false;
-            }
-        }
-        return true;
-    }
-
-    [[nodiscard]] std::string get_library_version() {
-        std::lock_guard guard(this->mutex_);
-
-        const auto ver = ::AX_ENGINE_GetVersion();
-        if (nullptr == ver) {
-            utilities::glog.print(utilities::log::type::error, "Get library version failed.\n");
-            return "";
-        }
-        return {ver};
-    }
-
-    [[nodiscard]] std::string get_model_version() {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return "";
-        }
-        return {::AX_ENGINE_GetModelToolsVersion(this->handle_)};
-    }
-
-    [[nodiscard]] int32_t get_model_type() {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return -1;
-        }
-
-        ::AX_ENGINE_MODEL_TYPE_T type;
-        if (const auto ret = ::AX_ENGINE_GetHandleModelType(this->handle_, &type); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model type failed{0x%08X}.\n", ret);
-            return -1;
-        }
-        return type;
-    }
-
-    [[nodiscard]] int32_t get_npu_type() {
-        std::lock_guard guard(this->mutex_);
-
-        if (!this->is_initialized_) {
-            utilities::glog.print(utilities::log::type::error, "axcl is not initialized.\n");
-            return false;
-        }
-
-        ::AX_ENGINE_NPU_ATTR_T attr{};
-        if (const auto ret = ::AX_ENGINE_GetVNPUAttr(&attr); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get visual NPU type failed{0x%08X}.\n", ret);
-            return -1;
-        }
-        return attr.eHardMode;
-    }
-
-    [[nodiscard]] int32_t get_batch_size() {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return -1;
-        }
-
-        return static_cast<int32_t>(this->batch);
-    }
-
-    [[nodiscard]] intmax_t get_sys_usage() {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return -1;
-        }
-
-        return 0;
-    }
-
-    [[nodiscard]] intmax_t get_cmm_usage() {
-        std::lock_guard guard(this->mutex_);
-
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return -1;
-        }
-
-        ::AX_ENGINE_CMM_INFO usage{};
-        if (const auto ret = ::AX_ENGINE_GetCMMUsage(this->handle_, &usage); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model usage failed{0x%08X}.\n", ret);
-            return -1;
-        }
-
-        return usage.nCMMSize;
-    }
-
-    static void sleep_for(const uint32_t sleep_duration) {
-        ::AX_SYS_Msleep(sleep_duration);
-    }
-
-private:
-    uint32_t group_ = 0;
-    uint32_t batch = 0;
-
-    ::AX_ENGINE_HANDLE handle_{};
-    ::AX_ENGINE_CONTEXT_T context_{};
-    ::AX_ENGINE_IO_INFO_T* info_{};
-    ::AX_ENGINE_IO_T io_{};
-    std::vector<::AX_ENGINE_IO_BUFFER_T> inputs_;
-    std::vector<::AX_ENGINE_IO_BUFFER_T> outputs_;
-
-    std::vector<bool> input_cached_;
-    std::vector<bool> output_cached_;
-
-    bool is_initialized_ = false;
-    std::mutex mutex_;
-};
-#endif

cpp/src/middleware/axcl_base.cc

@@ -1,241 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#include "middleware/axcl_base.hpp"
-
-
-#if defined(ENV_AXCL_RUNTIME_API_ENABLE) || defined(ENV_AXCL_NATIVE_API_ENABLE)
-
-#include "utilities/scalar_guard.hpp"
-#include "utilities/file.hpp"
-#include "utilities/log.hpp"
-
-#include <axcl.h>
-
-#include <mutex>
-
-static int static_module_count{0};
-static std::mutex static_module_mutex;
-
-static bool module_init(const std::string& config, const uint32_t& index, const middleware::axcl_base::npu_func& func) {
-    // 1. init axcl, using scalar_guard to ensure the finalization
-    const auto cfg_flag = utilities::exists(config) && utilities::is_regular_file(config);
-    utilities::glog.print(utilities::log::type::info, "axcl initializing...\n");
-    auto env_guard = utilities::scalar_guard<int32_t>(
-        axclInit(cfg_flag ? config.c_str() : nullptr),
-        [](const int32_t& code) {
-            if (0 == code) {
-                std::ignore = axclFinalize();
-            }
-        }
-    );
-
-    // 2. check the initialization result
-    if (const int ret = env_guard.get(); 0 != ret) {
-        utilities::glog.print(utilities::log::type::error, "Init axcl failed{0x%08X}.\n", ret);
-        return false;
-    }
-    utilities::glog.print(utilities::log::type::info, "axcl inited.\n");
-
-    // 3. get device list
-    axclrtDeviceList lst;
-    if (const auto ret = axclrtGetDeviceList(&lst); 0 != ret || 0 == lst.num) {
-        utilities::glog.print(utilities::log::type::error,
-            "Get axcl device failed{0x%08X}, find total %d device.\n", ret, lst.num);
-        return false;
-    }
-
-    // 4. check the device index
-    if (index >= lst.num) {
-        utilities::glog.print(utilities::log::type::error,
-            "Specified device index{%d} is out of range{total %d}.\n", index, lst.num);
-        return false;
-    }
-
-    // 5. set device
-    if (const auto ret = axclrtSetDevice(lst.devices[index]); 0 != ret) {
-        utilities::glog.print(utilities::log::type::error,
-            "Set axcl device as index{%d} failed{0x%08X}.\n", index, ret);
-        return false;
-    }
-    utilities::glog.print(utilities::log::type::info,
-        "Select axcl device{index: %d} as {%d}.\n", index, lst.devices[index]);
-
-    // 6. init NPU
-    if (!func()) {
-        utilities::glog.print(utilities::log::type::error, "Init NPU failed.\n");
-        return false;
-    }
-
-    // 7. disengage guard of env
-    env_guard.get() = -1;
-    return true;
-}
-
-static bool module_final(const middleware::axcl_base::npu_func& func) {
-    const auto flag = func();
-    const auto ret = axclFinalize();
-    return flag && (0 == ret);
-}
-
-bool middleware::axcl_base::init(const std::string &config, const uint32_t &index, const npu_func &func) {
-    bool flag = false;
-    {
-        std::lock_guard lock(static_module_mutex);
-        if (static_module_count == 0) {
-            flag = module_init(config, index, func);
-            if (flag) {
-                static_module_count++;
-            }
-        } else {
-            static_module_count++;
-            flag = true;
-        }
-    }
-
-    return flag;
-}
-
-bool middleware::axcl_base::final(const npu_func& func) {
-    bool flag = false;
-    {
-        std::lock_guard lock(static_module_mutex);
-        if (static_module_count == 1) {
-            flag = module_final(func);
-            if (flag) {
-                static_module_count--;
-            }
-        } else if (static_module_count > 1) {
-            static_module_count--;
-            flag = true;
-        }
-    }
-
-    return flag;
-}
-
-bool middleware::axcl_base::flush_input() const {
-    return true;
-}
-
-bool middleware::axcl_base::invalidate_output() const {
-    return true;
-}
-
-[[nodiscard]] bool middleware::axcl_base::feed(const std::string& input_folder, const std::string& stimulus_name) const {
-    for (uint32_t i = 0; i < this->get_input_count(); i++) {
-        if (const auto ret = feed(input_folder, stimulus_name, this->get_input_name(i), this->get_input_pointer(i), this->get_input_size(i) * this->get_batch_size()); !ret) {
-            utilities::glog.print(utilities::log::type::error, "Feed tensor {idx: %d, name: %s} failed.\n", i, this->get_input_name(i).c_str());
-            return false;
-        }
-    }
-    return true;
-}
-
-[[nodiscard]] bool middleware::axcl_base::verify(const std::string& output_folder, const std::string& stimulus_name) const {
-    for (uint32_t i = 0; i < this->get_output_count(); i++) {
-        if (const auto ret = verify(output_folder, stimulus_name, this->get_output_name(i),this->get_output_pointer(i), this->get_output_size(i) * this->get_batch_size()); !ret) {
-            utilities::glog.print(utilities::log::type::error, "Verify tensor {idx: %d, name: %s} failed.\n", i, this->get_output_name(i).c_str());
-            return false;
-        }
-    }
-    return true;
-}
-
-[[nodiscard]] bool middleware::axcl_base::save(const std::string& output_folder, const std::string& stimulus_name) const {
-    auto file_folder = output_folder;
-    file_folder.append("/").append(stimulus_name);
-
-    if (!utilities::exists(file_folder) && !utilities::create_directory(file_folder)) {
-        utilities::glog.print(utilities::log::type::error, "Create folder {%s} failed.\n", file_folder.c_str());
-        return false;
-    }
-
-    for (uint32_t i = 0; i < this->get_output_count(); i++) {
-        if (const auto ret = save(output_folder, stimulus_name, this->get_output_name(i), this->get_output_pointer(i), this->get_output_size(i) * this->get_batch_size()); !ret) {
-            utilities::glog.print(utilities::log::type::error, "Save tensor {idx: %d, name: %s} failed.\n", i, this->get_output_name(i).c_str());
-            return false;
-        }
-    }
-    return true;
-}
-
-bool middleware::axcl_base::feed(const std::string& folder, const std::string& stimulus_name, const std::string& tensor_name, void* address, const uintmax_t size) {
-    const auto file_name = utilities::get_legal_name(tensor_name);
-    auto file_path(folder);
-    file_path.append("/").append(stimulus_name).append("/").append(file_name).append(".bin");
-
-    std::vector<uint8_t> tensor_buffer(size);
-    if (!read(file_path, tensor_buffer.data(), tensor_buffer.size())) {
-        utilities::glog.print(utilities::log::type::error,
-            "Read tensor {name: %s} stimulus file {%s} failed.\n", tensor_name.c_str(), file_path.c_str());
-        return false;
-    }
-
-    if (const auto ret = axclrtMemcpy(address, tensor_buffer.data(), size, AXCL_MEMCPY_HOST_TO_DEVICE); 0 != ret) {
-        utilities::glog.print(utilities::log::type::error,
-            "Receive tensor {name: %s, addr: 0x%08X, size: %ld} from host {addr: 0x%08X} failed{0x%08X}.\n",
-            tensor_name.c_str(), address, size, tensor_buffer.data(), ret);
-        return false;
-    }
-
-    return true;
-}
-
-bool middleware::axcl_base::verify(const std::string& folder, const std::string& stimulus_name, const std::string& tensor_name, void* address, const uintmax_t size) {
-    const auto file_name = utilities::get_legal_name(tensor_name);
-    auto file_path(folder);
-    file_path.append("/").append(stimulus_name).append("/").append(file_name).append(".bin");
-
-    std::vector<uint8_t> file_buffer(size);
-    if (!read(file_path, file_buffer.data(), file_buffer.size())) {
-        utilities::glog.print(utilities::log::type::error,
-            "Read tensor {name: %s} file {%s} failed.\n", file_name.c_str(), file_path.c_str());
-        return false;
-    }
-
-    std::vector<uint8_t> tensor_buffer(size);
-
-    if (const auto ret = axclrtMemcpy(tensor_buffer.data(), address, size, AXCL_MEMCPY_DEVICE_TO_HOST); 0 != ret) {
-        utilities::glog.print(utilities::log::type::error,
-            "Send tensor {name: %s, addr: 0x%08X, size: %ld} to host {addr: 0x%08X} failed{0x%08X}.\n",
-            file_name.c_str(), address, size, tensor_buffer.data(), ret);
-        return false;
-    }
-
-    if (!runner::verify(file_buffer.data(), tensor_buffer.data(), file_buffer.size())) {
-        utilities::glog.print(utilities::log::type::error, "Verify tensor {name: %s} failed.\n", file_name.c_str());
-        return false;
-    }
-
-    return true;
-}
-
-bool middleware::axcl_base::save(const std::string& output_folder, const std::string& stimulus_name, const std::string& tensor_name, void* address, const uintmax_t size) {
-    const auto file_name = utilities::get_legal_name(tensor_name);
-    auto file_path(output_folder);
-    file_path.append("/").append(stimulus_name).append("/").append(file_name).append(".bin");
-
-    std::vector<uint8_t> tensor_buffer(size);
-    if (const auto ret = axclrtMemcpy(tensor_buffer.data(), address, size, AXCL_MEMCPY_DEVICE_TO_HOST); 0 != ret) {
-        utilities::glog.print(utilities::log::type::error,
-            "Send tensor {name: %s, addr: 0x%08X, size: %ld} to host {addr: 0x%08X} failed{0x%08X}.\n",
-            file_name.c_str(), address, size, tensor_buffer.data(), ret);
-        return false;
-    }
-
-    if (!write(file_path, tensor_buffer.data(), tensor_buffer.size())) {
-        utilities::glog.print(utilities::log::type::error, "Write tensor {name: %s} file {%s} failed.\n", file_name.c_str(), file_path.c_str());
-        return false;
-    }
-
-    return true;
-}
-#endif

cpp/src/middleware/axcl_base.hpp

@@ -1,42 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include "middleware/runner.hpp"
-
-#if defined(ENV_AXCL_RUNTIME_API_ENABLE) || defined(ENV_AXCL_NATIVE_API_ENABLE)
-
-#include <functional>
-
-namespace middleware {
-
-class axcl_base : public runner {
-public:
-    using npu_func = std::function<bool()>;
-
-    [[nodiscard]] static bool init(const std::string& config, const uint32_t& index, const npu_func& func);
-    [[nodiscard]] static bool final(const npu_func& func);
-
-    [[nodiscard]] bool flush_input() const override;
-    [[nodiscard]] bool invalidate_output() const override;
-
-    [[nodiscard]] bool feed(const std::string& input_folder, const std::string& stimulus_name) const override;
-    [[nodiscard]] bool verify(const std::string& output_folder, const std::string& stimulus_name) const override;
-    [[nodiscard]] bool save(const std::string& output_folder, const std::string& stimulus_name) const override;
-
-private:
-    [[nodiscard]] static bool feed(const std::string& folder, const std::string& stimulus_name, const std::string& tensor_name, void* address, uintmax_t size) ;
-    [[nodiscard]] static bool verify(const std::string &folder, const std::string &stimulus_name, const std::string &tensor_name, void *address, uintmax_t size) ;
-    [[nodiscard]] static bool save(const std::string& output_folder, const std::string& stimulus_name, const std::string& tensor_name, void* address, uintmax_t size);
-};
-
-}
-#endif

cpp/src/middleware/axcl_native_runner.cc

@@ -1,105 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#include "middleware/axcl_native_runner_impl.hpp"
-
-#if defined(ENV_AXCL_NATIVE_API_ENABLE)
-middleware::native_runner::native_runner() {
-    this->impl_ = new impl();
-}
-
-middleware::native_runner::~native_runner() {
-    delete this->impl_;
-}
-
-bool middleware::native_runner::init(const std::string& config_file, const uint32_t& device_index, const uint32_t& kind) {
-    return this->impl_->init(config_file, device_index, kind);
-}
-
-bool middleware::native_runner::final() {
-    return this->impl_->final();
-}
-
-bool middleware::native_runner::load(const std::string& model_path) {
-    return this->impl_->load(model_path);
-}
-
-bool middleware::native_runner::prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) {
-    return this->impl_->prepare(input_cached, output_cached, group, batch);
-}
-
-bool middleware::native_runner::run(const bool& parallel) {
-    return this->impl_->run(parallel);
-}
-
-uint32_t middleware::native_runner::get_input_count() const {
-    return this->impl_->get_input_count();
-}
-
-uint32_t middleware::native_runner::get_output_count() const {
-    return this->impl_->get_output_count();
-}
-
-std::string middleware::native_runner::get_input_name(const uint32_t& index) const {
-    return this->impl_->get_input_name(index);
-}
-
-std::string middleware::native_runner::get_output_name(const uint32_t& index) const {
-    return this->impl_->get_output_name(index);
-}
-
-void *middleware::native_runner::get_input_pointer(const uint32_t& index) const {
-    return this->impl_->get_input_pointer(index);
-}
-
-void *middleware::native_runner::get_output_pointer(const uint32_t& index) const {
-    return this->impl_->get_output_pointer(index);
-}
-
-uintmax_t middleware::native_runner::get_input_size(const uint32_t& index) const {
-    return this->impl_->get_input_size(index);
-}
-
-uintmax_t middleware::native_runner::get_output_size(const uint32_t& index) const {
-    return this->impl_->get_output_size(index);
-}
-
-uintmax_t middleware::native_runner::get_shape_group_count() const {
-    return this->impl_->get_shape_group_count();
-}
-
-std::string middleware::native_runner::get_library_version() const {
-    return impl::get_library_version();
-}
-
-std::string middleware::native_runner::get_model_version() const {
-    return this->impl_->get_model_version();
-}
-
-int32_t middleware::native_runner::get_model_type() const {
-    return this->impl_->get_model_type();
-}
-
-int32_t middleware::native_runner::get_npu_type() const {
-    return this->impl_->get_npu_type();
-}
-
-int32_t middleware::native_runner::get_batch_size() const {
-    return this->impl_->get_batch_size();
-}
-
-intmax_t middleware::native_runner::get_sys_usage() const {
-    return this->impl_->get_sys_usage();
-}
-
-intmax_t middleware::native_runner::get_cmm_usage() const {
-    return this->impl_->get_cmm_usage();
-}
-#endif

cpp/src/middleware/axcl_native_runner.hpp

@@ -1,57 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#if defined(ENV_AXCL_NATIVE_API_ENABLE)
-
-#include "middleware/axcl_base.hpp"
-
-namespace middleware {
-
-class native_runner final : public axcl_base {
-public:
-    native_runner();
-    ~native_runner() override;
-
-    [[nodiscard]] bool init(const std::string& config_file, const uint32_t& device_index, const uint32_t& kind) override;
-    [[nodiscard]] bool final() override;
-
-    [[nodiscard]] bool load(const std::string& model_path) override;
-    [[nodiscard]] bool prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) override;
-
-    [[nodiscard]] bool run(const bool& parallel) override;
-
-    [[nodiscard]] uint32_t get_input_count() const override;
-    [[nodiscard]] uint32_t get_output_count() const override;
-    [[nodiscard]] std::string get_input_name(const uint32_t& index) const override;
-    [[nodiscard]] std::string get_output_name(const uint32_t& index) const override;
-    [[nodiscard]] void *get_input_pointer(const uint32_t& index) const override;
-    [[nodiscard]] void *get_output_pointer(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_input_size(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_output_size(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_shape_group_count() const override;
-
-    [[nodiscard]] std::string get_library_version() const override;
-    [[nodiscard]] std::string get_model_version() const override;
-
-    [[nodiscard]] int32_t get_model_type() const override;
-    [[nodiscard]] int32_t get_npu_type() const override;
-    [[nodiscard]] int32_t get_batch_size() const override;
-    [[nodiscard]] intmax_t get_sys_usage() const override;
-    [[nodiscard]] intmax_t get_cmm_usage() const override;
-
-private:
-    struct impl;
-    impl *impl_;
-};
-
-}
-#endif

cpp/src/middleware/axcl_native_runner_impl.hpp

@@ -1,504 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include "middleware/axcl_native_runner.hpp"
-
-#include "utilities/scalar_guard.hpp"
-#include "utilities/file.hpp"
-#include "utilities/file_mapper.hpp"
-#include "utilities/log.hpp"
-
-#include <axcl.h>
-
-#if defined(ENV_AXCL_NATIVE_API_ENABLE)
-#define MODEL_LEADING_NAME "axcl npu "
-
-struct middleware::native_runner::impl {
-    impl() = default;
-
-    ~impl() {
-        std::ignore = this->final();
-    }
-
-    [[nodiscard]] bool init(const std::string& config_file, const uint32_t& index, const uint32_t& kind) {
-        // 0. check the NPU kind
-        if (kind >= ::AX_ENGINE_VIRTUAL_NPU_BUTT) {
-            utilities::glog.print(utilities::log::type::error,
-                "Specified NPU kind{%d} is out of range{total %d}.\n", kind, ::AX_ENGINE_VIRTUAL_NPU_BUTT);
-            return false;
-        }
-
-        // 1. build the npu init function
-        auto npu_init_func = [&kind]() {
-            ::AX_ENGINE_NPU_ATTR_T attr{};
-            attr.eHardMode = static_cast<::AX_ENGINE_NPU_MODE_T>(kind);
-            if (const int ret = ::AXCL_ENGINE_Init(&attr); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error,
-                    "Init axcl NPU as kind{%d} failed{0x%08X}.\n", kind, ret);
-                return false;
-            }
-            utilities::glog.print(utilities::log::type::info, "axcl NPU inited.\n");
-            return true;
-        };
-
-        // 2. execute initialization
-        if (const auto ret = axcl_base::init(config_file, index, npu_init_func); !ret) {
-            return false;
-        }
-
-        // 3. set the flag
-        this->is_initialized_ = true;
-        return true;
-    }
-
-    [[nodiscard]] bool final() {
-        auto flag = true;
-
-        if (this->is_initialized_) {
-            if (nullptr != this->handle_) {
-                for (auto& input : inputs_) {
-                    if (0 != input.phyAddr) {
-                        flag &= 0 == ::axclrtFree(reinterpret_cast<void *>(input.phyAddr));
-                        input.phyAddr = 0;
-                    }
-                }
-                for (auto& output : outputs_) {
-                    if (0 != output.phyAddr) {
-                        flag &= 0 == ::axclrtFree(reinterpret_cast<void *>(output.phyAddr));
-                        output.phyAddr = 0;
-                    }
-                }
-
-                flag &= 0 == ::AXCL_ENGINE_DestroyHandle(this->handle_);
-            }
-
-            flag &= 0 == axcl_base::final([&]() {
-               if (const auto ret = ::AXCL_ENGINE_Deinit(); 0 != ret) {
-                   utilities::glog.print(utilities::log::type::error, "Final axcl NPU failed{0x%08X}.\n", ret);
-                   return false;
-               }
-               utilities::glog.print(utilities::log::type::info, "axcl NPU finalized.\n");
-               return true;
-           });
-
-            this->is_initialized_ = false;
-        }
-
-        return flag;
-    }
-
-    [[nodiscard]] bool load(const std::string& model_path) {
-        if (!this->is_initialized_) {
-            utilities::glog.print(utilities::log::type::error, "axcl is not initialized.\n");
-            return false;
-        }
-
-        if (nullptr != this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model is already loaded.\n");
-            return false;
-        }
-
-        if (!utilities::exists(model_path)
-            || !utilities::is_regular_file(model_path)
-            || 0 == utilities::file_size(model_path)
-            || utilities::error_size == utilities::file_size(model_path)) {
-            utilities::glog.print(utilities::log::type::error,
-                "Model file{%s} error, please check it.\n", model_path.c_str());
-            return false;
-        }
-
-        const auto file = utilities::file_mapper(model_path);
-        if (nullptr == file.get()) {
-            utilities::glog.print(utilities::log::type::error,
-                "Model file{%s} mapping failed.\n", model_path.c_str());
-            return false;
-        }
-
-        const auto file_size = utilities::file_size(model_path);
-        auto device = utilities::scalar_guard<void*>(
-            [&file_size]() -> void* {
-                void *mem = nullptr;
-                if (const auto ret = ::axclrtMalloc(&mem, file_size, ::axclrtMemMallocPolicy{}); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error,
-                        "Malloc on device for model file failed{0x%08X}.\n", ret);
-                }
-                return mem;
-            },
-            [](void*& dev) { if (dev != nullptr) std::ignore = ::axclrtFree(dev); }
-        );
-
-        if (nullptr == device.get()) {
-            utilities::glog.print(utilities::log::type::error,
-                "Malloc on device for file{%s} failed.\n", model_path.c_str());
-            return false;
-        }
-
-        if (const auto ret = ::axclrtMemcpy(device.get(), file.get(), file_size, ::AXCL_MEMCPY_HOST_TO_DEVICE); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error,
-                "Send model file{%s} to device failed{0x%08X}.\n", model_path.c_str(), ret);
-            return false;
-        }
-
-        const auto model_name = std::string(MODEL_LEADING_NAME) + utilities::get_file_name(model_path);
-        ::AX_ENGINE_HANDLE_EXTRA_T extra_param{};
-        extra_param.pName = reinterpret_cast<::AX_S8 *>(const_cast<char *>(model_name.c_str()));
-        if (const auto ret = ::AXCL_ENGINE_CreateHandleV2(&this->handle_, device.get(), file_size, &extra_param); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error,
-                "Create model{%s} handle failed.\n", model_path.c_str());
-            return false;
-        }
-
-        if (const auto ret = ::AXCL_ENGINE_CreateContextV2(this->handle_, &this->context_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error,
-                "Create model{%s} context failed.\n", model_path.c_str());
-            return false;
-        }
-
-        return true;
-    }
-
-    [[nodiscard]] bool prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) {
-        std::ignore = input_cached;
-        std::ignore = output_cached;
-
-        // 0. check the handle
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is null.\n");
-            return false;
-        }
-
-        // 1. get the count of shape group
-        if (const auto ret = ::AXCL_ENGINE_GetGroupIOInfoCount(this->handle_, &this->group_count_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error,
-                "Get model shape group count failed{0x%08X}.\n", ret);
-            return false;
-        }
-
-        // 2. check the group index
-        if (group >= this->group_count_) {
-            utilities::glog.print(utilities::log::type::error,
-                "Model group{%d} is out of range{total %d}.\n", group, this->group_);
-            return false;
-        }
-        this->group_ = group;
-
-        // 3. get the IO info
-        if (1 == this->group_count_) {
-            if (const auto ret = ::AXCL_ENGINE_GetIOInfo(this->handle_, &info_); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error,
-                    "Get model IO info failed{0x%08X}.\n", ret);
-                return false;
-            }
-        } else {
-            if (const auto ret = ::AXCL_ENGINE_GetGroupIOInfo(this->handle_, group, &info_); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error,
-                    "Get model group{index: %d} IO info failed{0x%08X}.\n", group, ret);
-                return false;
-            }
-        }
-
-        // 4. check the batch size
-        this->batch_ = (0 == batch ? 1 : batch);
-        this->io_.nBatchSize = this->batch_;
-
-        // 5. prepare the input and output
-        this->inputs_.resize(this->info_->nInputSize, ::AX_ENGINE_IO_BUFFER_T{});
-        this->io_.pInputs = this->inputs_.data();
-        this->io_.nInputSize = this->info_->nInputSize;
-
-        this->outputs_.resize(this->info_->nOutputSize, ::AX_ENGINE_IO_BUFFER_T{});
-        this->io_.pOutputs = this->outputs_.data();
-        this->io_.nOutputSize = this->info_->nOutputSize;
-
-        // 6. prepare the memory, inputs
-        for (::AX_U32 i = 0; i < this->info_->nInputSize; i++) {
-            const auto meta = &(this->info_->pInputs[i]);
-            this->io_.pInputs[i].nSize = meta->nSize * this->batch_;
-
-            if (const auto ret = ::axclrtMalloc(
-                reinterpret_cast<void **>(&this->io_.pInputs[i].phyAddr),
-                this->io_.pInputs[i].nSize,
-                ::axclrtMemMallocPolicy{}); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error,
-                    "Memory allocation for tensor {%s} failed{0x%08X}.\n", meta->pName, ret);
-                return false;
-            }
-            utilities::glog.print(utilities::log::type::info, "Memory for tensor {%s} is allocated.\n", meta->pName);
-
-            // clean memory, some cases model may need to clean memory
-            //::axclrtMemset(reinterpret_cast<void *>(this->io_.pInputs[i].phyAddr), 0, this->io_.pInputs[i].nSize);
-
-            utilities::glog.print(utilities::log::type::info,
-                "Set tensor { name: %s, phy: %p, vir: %p, size: %u Bytes }.\n",
-                meta->pName,
-                reinterpret_cast<void*>(this->io_.pInputs[i].phyAddr),
-                this->io_.pInputs[i].pVirAddr,
-                this->io_.pInputs[i].nSize);
-        }
-
-        // 7. prepare the memory, outputs
-        for (::AX_U32 i = 0; i < this->info_->nOutputSize; i++) {
-            const auto meta = &(this->info_->pOutputs[i]);
-            this->io_.pOutputs[i].nSize = meta->nSize * this->batch_;
-
-            if (const auto ret = ::axclrtMalloc(
-                reinterpret_cast<void **>(&this->io_.pOutputs[i].phyAddr),
-                this->io_.pOutputs[i].nSize,
-                ::axclrtMemMallocPolicy{}); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error,
-                    "Memory allocation for tensor {%s} failed{0x%08X}.\n", meta->pName, ret);
-                return false;
-            }
-            utilities::glog.print(utilities::log::type::info, "Memory for tensor {%s} is allocated.\n", meta->pName);
-
-            // clean memory, some cases model may need to clean memory
-            //::axclrtMemset(reinterpret_cast<void *>(this->io_.pOutputs[i].phyAddr), 0, this->io_.pOutputs[i].nSize);
-
-            utilities::glog.print(utilities::log::type::info,
-                "Set tensor { name: %s, phy: %p, vir: %p, size: %u Bytes }.\n",
-                meta->pName,
-                reinterpret_cast<void*>(this->io_.pOutputs[i].phyAddr),
-                this->io_.pOutputs[i].pVirAddr,
-                this->io_.pOutputs[i].nSize);
-        }
-
-        return true;
-    }
-
-    [[nodiscard]] bool run(const bool& parallel) {
-        this->io_.nParallelRun = (parallel ? 1 : 0);
-
-        if (nullptr == this->handle_ || nullptr == this->context_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return false;
-        }
-
-        if (0 == this->io_.nInputSize || 0 == this->io_.nOutputSize) {
-            utilities::glog.print(utilities::log::type::error, "Model io is not set, prepare first.\n");
-            return false;
-        }
-
-        ::AX_S32 ret = 0;
-        // Note: this is a workaround for the case that the model has only ONE group,
-        //       currently AXCL_ENGINE_GetGroupIOInfo will return error.
-        //       the bug of libax_engine.so will be fixed in the future.
-        if (1 == this->group_count_) {
-            ret = ::AXCL_ENGINE_RunSyncV2(this->handle_, this->context_, &this->io_);
-        } else {
-            ret = ::AXCL_ENGINE_RunGroupIOSync(this->handle_, this->context_, this->group_, &this->io_);
-        }
-        if (0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Run model failed{0x%08X}.\n", ret);
-            return false;
-        }
-        utilities::glog.print(utilities::log::type::info, "Running done.\n");
-
-        return true;
-    }
-
-    [[nodiscard]] uint32_t get_input_count() const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        return this->info_->nInputSize;
-    }
-
-    [[nodiscard]] uint32_t get_output_count() const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        return this->info_->nOutputSize;
-    }
-
-    [[nodiscard]] std::string get_input_name(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return {};
-        }
-
-        if (index >= this->info_->nInputSize) {
-            return {};
-        }
-        return std::string{this->info_->pInputs[index].pName};
-    }
-
-    [[nodiscard]] std::string get_output_name(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return {};
-        }
-
-        if (index >= this->info_->nOutputSize) {
-            return {};
-        }
-        return std::string{this->info_->pOutputs[index].pName};
-    }
-
-    [[nodiscard]] void *get_input_pointer(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return nullptr;
-        }
-
-        if (index >= this->info_->nInputSize) {
-            return nullptr;
-        }
-        return reinterpret_cast<void *>(this->io_.pInputs[index].phyAddr);
-    }
-
-    [[nodiscard]] void *get_output_pointer(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return nullptr;
-        }
-
-        if (index >= this->info_->nOutputSize) {
-            return nullptr;
-        }
-        return reinterpret_cast<void *>(this->io_.pOutputs[index].phyAddr);
-    }
-
-    [[nodiscard]] uintmax_t get_input_size(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        if (index >= this->info_->nInputSize) {
-            return 0;
-        }
-        return this->io_.pInputs[index].nSize;
-    }
-
-    [[nodiscard]] uintmax_t get_output_size(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        if (index >= this->info_->nOutputSize) {
-            return 0;
-        }
-        return this->io_.pOutputs[index].nSize;
-    }
-
-    [[nodiscard]] uintmax_t get_shape_group_count() const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        ::AX_U32 group = 0;
-        if (const auto ret = ::AXCL_ENGINE_GetGroupIOInfoCount(this->handle_, &group); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model shape group count failed{0x%08X}.\n", ret);
-            return 0;
-        }
-        return group;
-    }
-
-    [[nodiscard]] static std::string get_library_version() {
-        int32_t major = 0, minor = 0, patch = 0;
-        if (const auto ret = ::axclrtGetVersion(&major, &minor, &patch); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get library version failed{0x%08X}.\n", ret);
-            return "";
-        }
-        return {std::to_string(major) + "." + std::to_string(minor) + "." + std::to_string(patch) + " " + ::AXCL_ENGINE_GetVersion()};
-    }
-
-    [[nodiscard]] std::string get_model_version() const {
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return "";
-        }
-        return {::AXCL_ENGINE_GetModelToolsVersion(this->handle_)};
-    }
-
-    [[nodiscard]] int32_t get_model_type() const {
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return -1;
-        }
-
-        ::AX_ENGINE_MODEL_TYPE_T type;
-        if (const auto ret = ::AXCL_ENGINE_GetHandleModelType(this->handle_, &type); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model type failed{0x%08X}.\n", ret);
-            return -1;
-        }
-        return type;
-    }
-
-    [[nodiscard]] int32_t get_npu_type() const {
-        if (!this->is_initialized_) {
-            utilities::glog.print(utilities::log::type::error, "axcl is not initialized.\n");
-            return false;
-        }
-
-        ::AX_ENGINE_NPU_ATTR_T attr{};
-        if (const auto ret = ::AXCL_ENGINE_GetVNPUAttr(&attr); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get visual NPU type failed{0x%08X}.\n", ret);
-            return -1;
-        }
-        return attr.eHardMode;
-    }
-
-    [[nodiscard]] int32_t get_batch_size() const {
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return -1;
-        }
-
-        return static_cast<int32_t>(this->batch_);
-    }
-
-    [[nodiscard]] intmax_t get_sys_usage() const {
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return -1;
-        }
-
-        return 0;
-    }
-
-    [[nodiscard]] intmax_t get_cmm_usage() const {
-        if (nullptr == this->handle_) {
-            utilities::glog.print(utilities::log::type::error, "Model handle is not set, load model first.\n");
-            return -1;
-        }
-
-        ::AX_ENGINE_CMM_INFO usage{};
-        if (const auto ret = ::AXCL_ENGINE_GetCMMUsage(this->handle_, &usage); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model usage failed{0x%08X}.\n", ret);
-            return -1;
-        }
-
-        return usage.nCMMSize;
-    }
-
-private:
-    uint32_t group_count_ = 0;
-    uint32_t group_ = 0;
-    uint32_t batch_ = 0;
-
-    ::AX_ENGINE_HANDLE handle_{};
-    ::AX_ENGINE_CONTEXT_T context_{};
-    ::AX_ENGINE_IO_INFO_T* info_{};
-    ::AX_ENGINE_IO_T io_{};
-    std::vector<::AX_ENGINE_IO_BUFFER_T> inputs_;
-    std::vector<::AX_ENGINE_IO_BUFFER_T> outputs_;
-
-    bool is_initialized_ = false;
-};
-#endif

cpp/src/middleware/axcl_runtime_runner.cc

@@ -1,105 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#include "middleware/axcl_runtime_runner_impl.hpp"
-
-#if defined(ENV_AXCL_RUNTIME_API_ENABLE)
-middleware::runtime_runner::runtime_runner() {
-    this->impl_ = new impl();
-}
-
-middleware::runtime_runner::~runtime_runner() {
-    delete this->impl_;
-}
-
-bool middleware::runtime_runner::init(const std::string& config_file, const uint32_t& device_index, const uint32_t& kind) {
-    return this->impl_->init(config_file, device_index, kind);
-}
-
-bool middleware::runtime_runner::final() {
-    return this->impl_->final();
-}
-
-bool middleware::runtime_runner::load(const std::string& model_path) {
-    return this->impl_->load(model_path);
-}
-
-bool middleware::runtime_runner::prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) {
-    return this->impl_->prepare(input_cached, output_cached, group, batch);
-}
-
-bool middleware::runtime_runner::run(const bool& parallel) {
-    return this->impl_->run(parallel);
-}
-
-uint32_t middleware::runtime_runner::get_input_count() const {
-    return this->impl_->get_input_count();
-}
-
-uint32_t middleware::runtime_runner::get_output_count() const {
-    return this->impl_->get_output_count();
-}
-
-std::string middleware::runtime_runner::get_input_name(const uint32_t& index) const {
-    return this->impl_->get_input_name(index);
-}
-
-std::string middleware::runtime_runner::get_output_name(const uint32_t& index) const {
-    return this->impl_->get_output_name(index);
-}
-
-void *middleware::runtime_runner::get_input_pointer(const uint32_t& index) const {
-    return this->impl_->get_input_pointer(index);
-}
-
-void *middleware::runtime_runner::get_output_pointer(const uint32_t& index) const {
-    return this->impl_->get_output_pointer(index);
-}
-
-uintmax_t middleware::runtime_runner::get_input_size(const uint32_t& index) const {
-    return this->impl_->get_input_size(index);
-}
-
-uintmax_t middleware::runtime_runner::get_output_size(const uint32_t& index) const {
-    return this->impl_->get_output_size(index);
-}
-
-uintmax_t middleware::runtime_runner::get_shape_group_count() const {
-    return this->impl_->get_shape_group_count();
-}
-
-std::string middleware::runtime_runner::get_library_version() const {
-    return impl::get_library_version();
-}
-
-std::string middleware::runtime_runner::get_model_version() const {
-    return this->impl_->get_model_version();
-}
-
-int32_t middleware::runtime_runner::get_model_type() const {
-    return this->impl_->get_model_type();
-}
-
-int32_t middleware::runtime_runner::get_npu_type() const {
-    return this->impl_->get_npu_type();
-}
-
-int32_t middleware::runtime_runner::get_batch_size() const {
-    return this->impl_->get_batch_size();
-}
-
-intmax_t middleware::runtime_runner::get_sys_usage() const {
-    return this->impl_->get_sys_usage();
-}
-
-intmax_t middleware::runtime_runner::get_cmm_usage() const {
-    return this->impl_->get_cmm_usage();
-}
-#endif

cpp/src/middleware/axcl_runtime_runner.hpp

@@ -1,56 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include "middleware/axcl_base.hpp"
-
-namespace middleware {
-
-#if defined(ENV_AXCL_RUNTIME_API_ENABLE)
-class runtime_runner final : public axcl_base {
-public:
-    runtime_runner();
-    ~runtime_runner() override;
-
-    [[nodiscard]] bool init(const std::string& config_file, const uint32_t& device_index, const uint32_t& kind) override;
-    [[nodiscard]] bool final() override;
-
-    [[nodiscard]] bool load(const std::string& model_path) override;
-    [[nodiscard]] bool prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) override;
-
-    [[nodiscard]] bool run(const bool& parallel) override;
-
-    [[nodiscard]] uint32_t get_input_count() const override;
-    [[nodiscard]] uint32_t get_output_count() const override;
-    [[nodiscard]] std::string get_input_name(const uint32_t& index) const override;
-    [[nodiscard]] std::string get_output_name(const uint32_t& index) const override;
-    [[nodiscard]] void *get_input_pointer(const uint32_t& index) const override;
-    [[nodiscard]] void *get_output_pointer(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_input_size(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_output_size(const uint32_t& index) const override;
-    [[nodiscard]] uintmax_t get_shape_group_count() const override;
-
-    [[nodiscard]] std::string get_library_version() const override;
-    [[nodiscard]] std::string get_model_version() const override;
-
-    [[nodiscard]] int32_t get_model_type() const override;
-    [[nodiscard]] int32_t get_npu_type() const override;
-    [[nodiscard]] int32_t get_batch_size() const override;
-    [[nodiscard]] intmax_t get_sys_usage() const override;
-    [[nodiscard]] intmax_t get_cmm_usage() const override;
-
-private:
-    struct impl;
-    impl *impl_;
-};
-#endif
-
-}

cpp/src/middleware/axcl_runtime_runner_impl.hpp

@@ -1,482 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include "middleware/axcl_runtime_runner.hpp"
-
-#include "middleware/axcl_base.hpp"
-#include "utilities/scalar_guard.hpp"
-#include "utilities/file.hpp"
-#include "utilities/log.hpp"
-
-#include <axcl.h>
-
-#if defined(ENV_AXCL_RUNTIME_API_ENABLE)
-struct middleware::runtime_runner::impl {
-    impl() = default;
-
-    ~impl() {
-        std::ignore = this->final();
-    }
-
-    [[nodiscard]] bool init(const std::string& config_file, const uint32_t& index, const uint32_t& kind) {
-        // 0. check the NPU kind
-        if (kind > AXCL_VNPU_LITTLE_BIG) {
-            utilities::glog.print(utilities::log::type::error,
-                "Specified NPU kind{%d} is out of range{total %d}.\n", kind, static_cast<int>(AXCL_VNPU_LITTLE_BIG) + 1);
-            return false;
-        }
-
-        // 1. build the npu init function
-        auto npu_init_func = [&kind]() {
-            if (const int ret = axclrtEngineInit(static_cast<axclrtEngineVNpuKind>(kind)); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error, "Init AXCLRT Engine as kind{%d} failed{0x%08X}.\n", kind, ret);
-                return false;
-            }
-            utilities::glog.print(utilities::log::type::info, "AXCLRT Engine inited.\n");
-            return true;
-        };
-
-        // 2. execute initialization
-        if (const auto ret = axcl_base::init(config_file, index, npu_init_func); !ret) {
-            return false;
-        }
-
-        // 3. set the flag
-        this->is_initialized_ = true;
-        return true;
-    }
-
-    [[nodiscard]] bool final() {
-        auto flag = true;
-
-        if (this->is_initialized_) {
-            if (0 != this->model_id_) {
-                for (auto& input : inputs_) {
-                    if (nullptr != input) {
-                        std::ignore = axclrtFree(input);
-                        input = nullptr;
-                    }
-                }
-                for (auto& output : outputs_) {
-                    if (nullptr != output) {
-                        std::ignore = axclrtFree(output);
-                        output = nullptr;
-                    }
-                }
-
-                flag &= 0 == axclrtEngineDestroyIOInfo(this->info_);
-                flag &= 0 == axclrtEngineDestroyIO(this->io_);
-                flag &= 0 == axclrtEngineUnload(this->model_id_);
-            }
-
-             flag &= 0 == axcl_base::final([&]() {
-                if (const auto ret = axclrtEngineFinalize(); 0 != ret) {
-                    utilities::glog.print(utilities::log::type::error, "Final AXCLRT Engine failed{0x%08X}.\n", ret);
-                    return false;
-                }
-                utilities::glog.print(utilities::log::type::info, "AXCLRT Engine finalized.\n");
-                return true;
-            });
-
-            this->is_initialized_ = false;
-        }
-
-        return flag;
-    }
-
-    [[nodiscard]] bool load(const std::string& model_path) {
-        if (!this->is_initialized_) {
-            utilities::glog.print(utilities::log::type::error, "axcl is not initialized.\n");
-            return false;
-        }
-
-        if (0 != this->model_id_) {
-            utilities::glog.print(utilities::log::type::error, "Model is already loaded.\n");
-            return false;
-        }
-
-        if (!utilities::exists(model_path)
-            || !utilities::is_regular_file(model_path)
-            || 0 == utilities::file_size(model_path)
-            || utilities::error_size == utilities::file_size(model_path)) {
-            utilities::glog.print(utilities::log::type::error, "Model file{%s} error, please check it.\n", model_path.c_str());
-            return false;
-        }
-
-        if (const auto ret = axclrtEngineLoadFromFile(model_path.c_str(), &this->model_id_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Create model{%s} handle failed.\n", model_path.c_str());
-            return false;
-        }
-
-        if (const auto ret = axclrtEngineCreateContext(this->model_id_, &this->context_id_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Create model{%s} context failed.\n", model_path.c_str());
-            return false;
-        }
-
-        return true;
-    }
-
-    [[nodiscard]] bool prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) {
-        std::ignore = input_cached;
-        std::ignore = output_cached;
-
-        // 0. check the handle
-        if (0 == this->model_id_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return false;
-        }
-
-        // 1. get the IO info
-        if (const auto ret = axclrtEngineGetIOInfo(this->model_id_, &this->info_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model io info failed{0x%08X}.\n", ret);
-            return false;
-        }
-
-        // 2. get the count of shape group
-        int32_t total_group = 0;
-        if (const auto ret = axclrtEngineGetShapeGroupsCount(this->info_, &total_group); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model shape group count failed{0x%08X}.\n", ret);
-            return false;
-        }
-
-        // 3. check the group index
-        if (group >= static_cast<decltype(group)>(total_group)) {
-            utilities::glog.print(utilities::log::type::error, "Model group{%d} is out of range{total %d}.\n", group, total_group);
-            return false;
-        }
-        this->group_ = static_cast<int32_t>(group);
-
-        // 4. check the batch size
-        this->batch_ = (0 == batch ? 1 : batch);
-
-        // 5. get the count of inputs
-        uint32_t input_count = 0;
-        if (input_count = axclrtEngineGetNumInputs(this->info_); 0 == input_count) {
-            utilities::glog.print(utilities::log::type::error, "Get model input count failed.\n");
-            return false;
-        }
-
-        // 6. get the count of outputs
-        uint32_t output_count = 0;
-        if (output_count = axclrtEngineGetNumOutputs(this->info_); 0 == output_count) {
-            utilities::glog.print(utilities::log::type::error, "Get model output count failed.\n");
-            return false;
-        }
-
-        // 7. prepare the input and output
-        this->inputs_.resize(input_count, nullptr);
-        this->inputs_size_.resize(input_count, 0);
-        this->outputs_.resize(output_count, nullptr);
-        this->outputs_size_.resize(output_count, 0);
-
-        // 8. prepare the memory, inputs
-        for (uint32_t i = 0; i < input_count; i++) {
-            uint32_t original_size = 0;
-            if (original_size = axclrtEngineGetInputSizeByIndex(this->info_, group, i); 0 == original_size) {
-                utilities::glog.print(utilities::log::type::error, "Get model input{index: %d} size failed.\n", i);
-                return false;
-            }
-
-            this->inputs_size_[i] = original_size * this->batch_;
-
-            if (const auto ret = axclrtMalloc(&this->inputs_[i], this->inputs_size_[i], axclrtMemMallocPolicy{}); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error, "Memory allocation for tensor{index: %d} failed{0x%08X}.\n", i, ret);
-                return false;
-            }
-            utilities::glog.print(utilities::log::type::info, "Memory for tensor{index: %d} is allocated.\n", i);
-
-            // clean memory, some cases model may need to clean memory
-            // axclrtMemset(this->inputs_[i], 0, size);
-
-            utilities::glog.print(utilities::log::type::info,
-                "Set tensor { index: %d, address: %p, size: %u Bytes }.\n", i, this->inputs_[i], this->inputs_size_[i]);
-        }
-
-        // 9. prepare the memory, outputs
-        for (uint32_t i = 0; i < output_count; i++) {
-            uint32_t original_size = 0;
-            if (original_size = axclrtEngineGetOutputSizeByIndex(this->info_, group, i); 0 == original_size) {
-                utilities::glog.print(utilities::log::type::error, "Get model output{index: %d} size failed.\n", i);
-                return false;
-            }
-
-            this->outputs_size_[i] = original_size * this->batch_;
-
-            if (const auto ret = axclrtMalloc(&this->outputs_[i], this->outputs_size_[i], axclrtMemMallocPolicy{}); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error, "Memory allocation for tensor{index: %d} failed{0x%08X}.\n", i, ret);
-                return false;
-            }
-            utilities::glog.print(utilities::log::type::info, "Memory for tensor{index: %d} is allocated.\n", i);
-
-            // clean memory, some cases model may need to clean memory
-            // axclrtMemset(this->outputs_[i], 0, size);
-
-            utilities::glog.print(utilities::log::type::info,
-                "Set tensor { index: %d, address: %p, size: %u Bytes }.\n", i, this->outputs_[i], this->outputs_size_[i]);
-        }
-
-        // 10. create the IO
-        if (const auto ret = axclrtEngineCreateIO(this->info_, &this->io_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Create model io failed{0x%08X}.\n", ret);
-            return false;
-        }
-        utilities::glog.print(utilities::log::type::info, "AXCLRT Engine inited.\n");
-
-        // 11. set the input and output buffer
-        for (uint32_t i = 0; i < input_count; i++) {
-            if (const auto ret = axclrtEngineSetInputBufferByIndex(this->io_, i, this->inputs_[i], this->inputs_size_[i]); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error, "Set input buffer{index: %d} failed{0x%08X}.\n", i, ret);
-                return false;
-            }
-        }
-        for (uint32_t i = 0; i < output_count; i++) {
-            if (const auto ret = axclrtEngineSetOutputBufferByIndex(this->io_, i, this->outputs_[i], this->outputs_size_[i]); 0 != ret) {
-                utilities::glog.print(utilities::log::type::error, "Set output buffer{index: %d} failed{0x%08X}.\n", i, ret);
-                return false;
-            }
-        }
-
-        // 12. set the batch size
-        if (const auto ret = axclrtEngineSetDynamicBatchSize(this->io_, this->batch_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Set batch size{%d} failed{0x%08X}.\n", this->batch_, ret);
-            return false;
-        }
-
-        return true;
-    }
-
-    [[nodiscard]] bool run(const bool& parallel) const {
-        std::ignore = parallel;
-
-        if (0 == this->model_id_ || 0 == this->context_id_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return false;
-        }
-
-        if (nullptr == this->io_) {
-            utilities::glog.print(utilities::log::type::error, "Model io is not set, prepare first.\n");
-            return false;
-        }
-
-        if (const auto ret = axclrtEngineExecute(this->model_id_, this->context_id_, this->group_, this->io_); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Run model failed{0x%08X}.\n", ret);
-            return false;
-        }
-        utilities::glog.print(utilities::log::type::info, "Running done.\n");
-
-        return true;
-    }
-
-    [[nodiscard]] uint32_t get_input_count() const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        return axclrtEngineGetNumInputs(this->info_);
-    }
-
-    [[nodiscard]] uint32_t get_output_count() const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        return axclrtEngineGetNumOutputs(this->info_);
-    }
-
-    [[nodiscard]] std::string get_input_name(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return {};
-        }
-
-        if (index >= get_input_count()) {
-            return {};
-        }
-        return std::string{axclrtEngineGetInputNameByIndex(this->info_, index)};
-    }
-
-    [[nodiscard]] std::string get_output_name(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return {};
-        }
-
-        if (index >= get_output_count()) {
-            return {};
-        }
-        return std::string{axclrtEngineGetOutputNameByIndex(this->info_, index)};
-    }
-
-    [[nodiscard]] void *get_input_pointer(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return nullptr;
-        }
-
-        if (index >= get_input_count()) {
-            return nullptr;
-        }
-        return this->inputs_[index];
-    }
-
-    [[nodiscard]] void *get_output_pointer(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return nullptr;
-        }
-
-        if (index >= get_output_count()) {
-            return nullptr;
-        }
-        return this->outputs_[index];
-    }
-
-    [[nodiscard]] uintmax_t get_input_size(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        if (index >= get_input_count()) {
-            return 0;
-        }
-        return axclrtEngineGetInputSizeByIndex(this->info_, this->group_, index);
-    }
-
-    [[nodiscard]] uintmax_t get_output_size(const uint32_t& index) const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        if (index >= get_output_count()) {
-            return 0;
-        }
-        return axclrtEngineGetOutputSizeByIndex(this->info_, this->group_, index);
-    }
-
-    [[nodiscard]] uintmax_t get_shape_group_count() const {
-        if (nullptr == this->info_) {
-            utilities::glog.print(utilities::log::type::error, "Model io info is not set, prepare first.\n");
-            return 0;
-        }
-
-        int group = 0;
-        if (const auto ret = axclrtEngineGetShapeGroupsCount(this->info_, &group); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model shape group count failed{0x%08X}.\n", ret);
-            return 0;
-        }
-        return group;
-    }
-
-    [[nodiscard]] static std::string get_library_version() {
-        int32_t major = 0, minor = 0, patch = 0;
-        if (const auto ret = axclrtGetVersion(&major, &minor, &patch); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get library version failed{0x%08X}.\n", ret);
-            return "";
-        }
-        return {std::to_string(major) + "." + std::to_string(minor) + "." + std::to_string(patch)};
-    }
-
-    [[nodiscard]] std::string get_model_version() const {
-        if (0 == this->model_id_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return "";
-        }
-        return {axclrtEngineGetModelCompilerVersion(this->model_id_)};
-    }
-
-    [[nodiscard]] int32_t get_model_type() const {
-        if (0 == this->model_id_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return -1;
-        }
-
-        axclrtEngineModelKind type;
-        if (const auto ret = axclrtEngineGetModelTypeFromModelId(this->model_id_, &type); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model type failed{0x%08X}.\n", ret);
-            return -1;
-        }
-        return type;
-    }
-
-    [[nodiscard]] int32_t get_npu_type() const {
-        if (!this->is_initialized_) {
-            utilities::glog.print(utilities::log::type::error, "axcl is not initialized.\n");
-            return false;
-        }
-
-        axclrtEngineVNpuKind kind;
-        if (const auto ret = axclrtEngineGetVNpuKind(&kind); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get visual NPU type failed{0x%08X}.\n", ret);
-            return -1;
-        }
-        return kind;
-    }
-
-    [[nodiscard]] int32_t get_batch_size() const {
-        if (0 == this->model_id_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return -1;
-        }
-
-        return static_cast<int32_t>(this->batch_);
-    }
-
-    [[nodiscard]] intmax_t get_sys_usage() const {
-        if (0 == this->model_id_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return -1;
-        }
-
-        int64_t sys_size = 0, cmm_size = 0;
-        if (const auto ret = axclrtEngineGetUsageFromModelId(this->model_id_, &sys_size, &cmm_size); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model usage failed{0x%08X}.\n", ret);
-            return -1;
-        }
-
-        return sys_size;
-    }
-
-    [[nodiscard]] intmax_t get_cmm_usage() const {
-        if (0 == this->model_id_) {
-            utilities::glog.print(utilities::log::type::error, "Model id is not set, load model first.\n");
-            return -1;
-        }
-
-        int64_t sys_size = 0, cmm_size = 0;
-        if (const auto ret = axclrtEngineGetUsageFromModelId(this->model_id_, &sys_size, &cmm_size); 0 != ret) {
-            utilities::glog.print(utilities::log::type::error, "Get model usage failed{0x%08X}.\n", ret);
-            return -1;
-        }
-
-        return cmm_size;
-    }
-
-private:
-    int32_t group_ = 0;
-    uint32_t batch_ = 0;
-
-    uint64_t model_id_{};
-    uint64_t context_id_{};
-    axclrtEngineIOInfo info_{};
-    axclrtEngineIO io_{};
-    std::vector<void*> inputs_;
-    std::vector<void*> outputs_;
-    std::vector<uintmax_t> inputs_size_;
-    std::vector<uintmax_t> outputs_size_;
-
-    bool is_initialized_ = false;
-};
-#endif

cpp/src/middleware/runner.cc

@@ -1,88 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#include "middleware/runner.hpp"
-
-#include "utilities/file.hpp"
-#include "utilities/log.hpp"
-
-#include <chrono>
-#include <thread>
-
-void middleware::runner::sleep_for(const uint32_t sleep_duration) const {
-    std::this_thread::sleep_for(std::chrono::milliseconds(sleep_duration));
-}
-
-std::vector<std::string> middleware::runner::read(const std::string& list_file) {
-    if (!utilities::exists(list_file) || !utilities::is_regular_file(list_file)) {
-        utilities::glog.print(utilities::log::type::error,
-            "Stimulus list file{%s} is not exist.\n", list_file.c_str());
-        return {};
-    }
-
-    std::vector<std::string> list;
-
-    std::ifstream list_stream;
-    list_stream.open(list_file, std::ios::in);
-    if (!list_stream.is_open()) {
-        utilities::glog.print(utilities::log::type::error,
-            "Read stimulus list file{%s} failed.\n", list_file.c_str());
-        return {};
-    }
-
-    std::string file_path;
-    while (std::getline(list_stream, file_path)) {
-        list.push_back(file_path);
-    }
-    list_stream.close();
-
-    return list;
-}
-
-bool middleware::runner::read(const std::string& binary_file, void* address, const uintmax_t& size) {
-    if (!utilities::exists(binary_file) || !utilities::is_regular_file(binary_file)) {
-        utilities::glog.print(utilities::log::type::error,
-            "Stimulus file{%s} is not exist.\n", binary_file.c_str());
-        return false;
-    }
-
-    const auto file_size = utilities::file_size(binary_file);
-    if (utilities::error_size == file_size) {
-        utilities::glog.print(utilities::log::type::error, "Get stimulus file size failed.\n");
-        return false;
-    }
-
-    if (file_size != size) {
-        utilities::glog.print(utilities::log::type::error,
-            "Size of stimulus file{%s} is not equal the specified{file: %d vs tensor: %d}.\n", binary_file.c_str(), file_size, size);
-        return false;
-    }
-
-    return utilities::read(binary_file, address, size);
-}
-
-bool middleware::runner::write(const std::string& binary_file, const void* address, const uintmax_t& size) {
-    return utilities::write(binary_file, address, size);
-}
-
-bool middleware::runner::verify(const void* dst, const void* src, const uintmax_t& size) {
-    const auto* src_ptr = static_cast<const uint8_t*>(src);
-    const auto* dst_ptr = static_cast<const uint8_t*>(dst);
-
-    for (uintmax_t i = 0; i < size; i++) {
-        if (src_ptr[i] != dst_ptr[i]) {
-            utilities::glog.print(utilities::log::type::error,
-                "Verify stimulus failed{offset: 0x%x; 0x%02X vs. 0x%02X}.\n", i, src_ptr[i], dst_ptr[i]);
-            return false;
-        }
-    }
-
-    return true;
-}

cpp/src/middleware/runner.hpp

@@ -1,67 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include <string>
-#include <vector>
-#include <cstdint>
-
-namespace middleware {
-
-class runner {
-public:
-    runner() = default;
-    virtual ~runner() = default;
-
-    [[nodiscard]] virtual bool init(const std::string& config_file, const uint32_t& device_index, const uint32_t& kind) = 0;
-    [[nodiscard]] virtual bool final() = 0;
-
-    [[nodiscard]] virtual bool load(const std::string& model_path) = 0;
-    [[nodiscard]] virtual bool prepare(const bool& input_cached, const bool& output_cached, const uint32_t& group, const uint32_t& batch) = 0;
-
-    [[nodiscard]] virtual bool run(const bool& parallel) = 0;
-
-    [[nodiscard]] virtual uint32_t get_input_count() const = 0;
-    [[nodiscard]] virtual uint32_t get_output_count() const = 0;
-    [[nodiscard]] virtual std::string get_input_name(const uint32_t& index) const = 0;
-    [[nodiscard]] virtual std::string get_output_name(const uint32_t& index) const = 0;
-    [[nodiscard]] virtual void *get_input_pointer(const uint32_t& index) const = 0;
-    [[nodiscard]] virtual void *get_output_pointer(const uint32_t& index) const = 0;
-    [[nodiscard]] virtual uintmax_t get_input_size(const uint32_t& index) const = 0;
-    [[nodiscard]] virtual uintmax_t get_output_size(const uint32_t& index) const = 0;
-
-    [[nodiscard]] virtual uintmax_t get_shape_group_count() const = 0;
-
-    [[nodiscard]] virtual bool flush_input() const = 0;
-    [[nodiscard]] virtual bool invalidate_output() const = 0;
-
-    [[nodiscard]] virtual bool feed(const std::string& input_folder, const std::string& stimulus_name) const = 0;
-    [[nodiscard]] virtual bool verify(const std::string& output_folder, const std::string& stimulus_name) const = 0;
-    [[nodiscard]] virtual bool save(const std::string& output_folder, const std::string& stimulus_name) const = 0;
-
-    [[nodiscard]] virtual std::string get_library_version() const = 0;
-    [[nodiscard]] virtual std::string get_model_version() const = 0;
-
-    [[nodiscard]] virtual int32_t get_model_type() const = 0;
-    [[nodiscard]] virtual int32_t get_npu_type() const = 0;
-    [[nodiscard]] virtual int32_t get_batch_size() const = 0;
-    [[nodiscard]] virtual intmax_t get_sys_usage() const = 0;
-    [[nodiscard]] virtual intmax_t get_cmm_usage() const = 0;
-
-    virtual void sleep_for(uint32_t sleep_duration) const;
-
-    [[nodiscard]] static std::vector<std::string> read(const std::string& list_file);
-    [[nodiscard]] static bool read(const std::string& binary_file, void* address, const uintmax_t& size);
-    [[nodiscard]] static bool write(const std::string& binary_file, const void* address, const uintmax_t& size);
-    [[nodiscard]] static bool verify(const void* dst, const void* src, const uintmax_t& size);
-};
-
-}

cpp/src/tokenizer.cpp

@@ -1,306 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * Tokenizer implementation. Uses pinyin_table.hpp (auto-generated from pypinyin.pinyin_dict)
- * for Chinese character → pinyin token mapping, matching Python LocalEmiliaTokenizer output.
- **************************************************************************************************/
-
-#include "tokenizer.hpp"
-#include "src/pinyin_table.hpp"
-
-#include <fstream>
-#include <sstream>
-#include <regex>
-#include <cctype>
-#include <algorithm>
-
-Tokenizer::Tokenizer() : m_pad_id(0), m_loaded(false) {}
-
-int Tokenizer::Load(const std::string& token_file) {
-    std::ifstream file(token_file);
-    if (!file.is_open()) {
-        printf("Failed to open token file: %s\n", token_file.c_str());
-        return -1;
-    }
-
-    m_token2id.clear();
-    std::string line;
-    while (std::getline(file, line)) {
-        if (line.empty()) continue;
-        // Format: token\tid
-        size_t tab_pos = line.rfind('\t');
-        if (tab_pos == std::string::npos) continue;
-
-        std::string token = line.substr(0, tab_pos);
-        int id = std::stoi(line.substr(tab_pos + 1));
-        m_token2id[token] = id;
-    }
-    file.close();
-
-    // Build reverse mapping
-    int max_id = 0;
-    for (auto& kv : m_token2id) {
-        if (kv.second > max_id) max_id = kv.second;
-    }
-    m_id2token.resize(max_id + 1);
-    for (auto& kv : m_token2id) {
-        m_id2token[kv.second] = kv.first;
-    }
-
-    auto it = m_token2id.find("_");
-    if (it != m_token2id.end()) {
-        m_pad_id = it->second;
-    } else {
-        m_pad_id = 0;
-        printf("WARNING: pad token '_' not found in token file, using 0 as pad_id\n");
-    }
-
-    m_loaded = true;
-    printf("Loaded %zu tokens from %s, pad_id=%d\n", m_token2id.size(), token_file.c_str(), m_pad_id);
-    return 0;
-}
-
-int Tokenizer::TokenToId(const std::string& token) const {
-    auto it = m_token2id.find(token);
-    return (it != m_token2id.end()) ? it->second : m_pad_id;
-}
-
-bool Tokenizer::IsChinese(char c) {
-    return (c >= 0x4E00 && c <= 0x9FFF);
-}
-
-bool Tokenizer::IsAlphabet(char c) {
-    return (c >= 'A' && c <= 'Z') || (c >= 'a' && c <= 'z');
-}
-
-std::string Tokenizer::MapPunctuation(const std::string& text) {
-    // Map full-width Chinese punctuation to ASCII equivalents.
-    // UTF-8 code points:
-    //   \xEF\xBC\x8C = U+FF0C (，)  \xEF\xBC\x9A = U+FF1A (：)
-    //   \xE3\x80\x82 = U+3002 (。)  \xE3\x80\x81 = U+3001 (、)
-    //   \xEF\xBC\x81 = U+FF01 (！)  \xE2\x80\x9C = U+201C (")
-    //   \xEF\xBC\x9F = U+FF1F (？)  \xE2\x80\x9D = U+201D (")
-    //   \xEF\xBC\x9B = U+FF1B (；)  \xE2\x80\x98 = U+2018 (')
-    //   \xE2\x80\xA6 = U+2026 (…)   \xE2\x80\x99 = U+2019 (')
-    //   \xE2\x8B\xAF = U+22EF (⋯)
-    static const std::pair<std::string, std::string> replacements[] = {
-        {"\xEF\xBC\x8C", ","},   // ，
-        {"\xE3\x80\x82", "."},   // 。
-        {"\xEF\xBC\x81", "!"},   // ！
-        {"\xEF\xBC\x9F", "?"},   // ？
-        {"\xEF\xBC\x9B", ";"},   // ；
-        {"\xEF\xBC\x9A", ":"},   // ：
-        {"\xE3\x80\x81", ","},   // 、
-        {"\xE2\x80\x9C", "\""},  // "
-        {"\xE2\x80\x9D", "\""},  // "
-        {"\xE2\x80\x98", "'"},   // '
-        {"\xE2\x80\x99", "'"},   // '
-        {"\xE2\x80\xA6", "..."}, // …
-        {"\xE2\x8B\xAF", "..."}, // ⋯
-    };
-    std::string result = text;
-    for (auto& rep : replacements) {
-        size_t pos = 0;
-        while ((pos = result.find(rep.first, pos)) != std::string::npos) {
-            result.replace(pos, rep.first.length(), rep.second);
-            pos += rep.second.length();
-        }
-    }
-    return result;
-}
-
-std::vector<std::pair<std::string, std::string>>
-Tokenizer::GetSegments(const std::string& text) const {
-    std::vector<std::pair<std::string, std::string>> segments;
-    if (text.empty()) return segments;
-
-    std::string current;
-    std::string current_lang;
-
-    auto char_lang = [](char c) -> std::string {
-        if (IsChinese(c)) return "zh";
-        if (IsAlphabet(c)) return "en";
-        return "other";
-    };
-
-    for (size_t i = 0; i < text.size(); ) {
-        // Handle UTF-8 multi-byte characters
-        size_t char_len = 1;
-        unsigned char c = static_cast<unsigned char>(text[i]);
-        if ((c & 0x80) == 0) char_len = 1;
-        else if ((c & 0xE0) == 0xC0) char_len = 2;
-        else if ((c & 0xF0) == 0xE0) char_len = 3;
-        else if ((c & 0xF8) == 0xF0) char_len = 4;
-
-        std::string ch = text.substr(i, char_len);
-        std::string lang;
-
-        if (char_len > 1) {
-            lang = "zh"; // Assume multi-byte UTF-8 is Chinese
-        } else if (IsAlphabet(ch[0])) {
-            lang = "en";
-        } else {
-            lang = "other";
-        }
-
-        if (current.empty()) {
-            current = ch;
-            current_lang = lang;
-        } else if (current_lang == "other" || lang == "other" || current_lang == lang) {
-            current += ch;
-            if (current_lang == "other" && lang != "other") current_lang = lang;
-        } else {
-            segments.push_back({current, current_lang});
-            current = ch;
-            current_lang = lang;
-        }
-        i += char_len;
-    }
-    if (!current.empty()) {
-        segments.push_back({current, current_lang});
-    }
-    return segments;
-}
-
-std::vector<std::string> Tokenizer::TokenizeZh(const std::string& text) const {
-    // Character-by-character pinyin tokenization using pinyin_table.hpp.
-    // Matches Python LocalEmiliaTokenizer (without jieba word segmentation).
-    std::vector<std::string> tokens;
-    std::string normalized = MapPunctuation(text);
-
-    for (size_t i = 0; i < normalized.size(); ) {
-        unsigned char c = static_cast<unsigned char>(normalized[i]);
-        size_t char_len = 1;
-        if ((c & 0x80) == 0) char_len = 1;
-        else if ((c & 0xE0) == 0xC0) char_len = 2;
-        else if ((c & 0xF0) == 0xE0) char_len = 3;
-        else if ((c & 0xF8) == 0xF0) char_len = 4;
-
-        std::string ch = normalized.substr(i, char_len);
-        i += char_len;
-
-        // Punctuation or ASCII
-        if (char_len == 1) {
-            std::string mapped = MapPunctuation(ch);
-            if (mapped != ch && m_token2id.find(mapped) != m_token2id.end()) {
-                tokens.push_back(mapped);
-                continue;
-            }
-            if (m_token2id.find(ch) != m_token2id.end()) {
-                tokens.push_back(ch);
-                continue;
-            }
-        }
-
-        // Chinese character: lookup pinyin token IDs via generated table
-        if (char_len == 3) {
-            uint32_t cp = ((unsigned char)ch[0] & 0x0F) << 12
-                        | ((unsigned char)ch[1] & 0x3F) << 6
-                        | ((unsigned char)ch[2] & 0x3F);
-            auto ids = pinyin_lookup(cp);
-            for (int tid : ids) {
-                if (tid >= 0 && tid < (int)m_id2token.size() && !m_id2token[tid].empty()) {
-                    tokens.push_back(m_id2token[tid]);
-                }
-            }
-        }
-    }
-    return tokens;
-}
-
-std::vector<std::string> Tokenizer::TokenizeEn(const std::string& text) const {
-    // Simplified English tokenization: character-by-character phoneme mapping
-    // In production, use the Python tokenizer with piper_phonemize
-    std::vector<std::string> tokens;
-    std::string lower = text;
-    std::transform(lower.begin(), lower.end(), lower.begin(), ::tolower);
-    lower = MapPunctuation(lower);
-
-    for (size_t i = 0; i < lower.size(); ++i) {
-        char c = lower[i];
-        if (c == ' ') continue;
-        if (ispunct(c)) {
-            // Map common punctuation
-            switch (c) {
-                case ',': tokens.push_back(","); break;
-                case '.': tokens.push_back("."); break;
-                case '!': tokens.push_back("!"); break;
-                case '?': tokens.push_back("?"); break;
-                case ';': tokens.push_back(";"); break;
-                case ':': tokens.push_back(":"); break;
-                case '\'': tokens.push_back("'"); break;
-                case '"': tokens.push_back("\""); break;
-                default: break;
-            }
-            continue;
-        }
-        // Output character as-is (phoneme-based tokens)
-        std::string token(1, c);
-        if (m_token2id.find(token) != m_token2id.end()) {
-            tokens.push_back(token);
-        }
-    }
-    return tokens;
-}
-
-std::vector<int> Tokenizer::TextToTokenIds(const std::string& text) const {
-    if (!m_loaded) return {};
-
-    std::string normalized = MapPunctuation(text);
-    auto segments = GetSegments(normalized);
-
-    std::vector<int> token_ids;
-    for (auto& seg : segments) {
-        std::vector<std::string> tokens;
-        if (seg.second == "zh") {
-            tokens = TokenizeZh(seg.first);
-        } else if (seg.second == "en") {
-            tokens = TokenizeEn(seg.first);
-        }
-        // "other" segments: try individual chars as tokens
-        else {
-            for (char c : seg.first) {
-                std::string token(1, c);
-                if (m_token2id.find(token) != m_token2id.end()) {
-                    token_ids.push_back(m_token2id.at(token));
-                }
-            }
-            continue;
-        }
-
-        for (auto& token : tokens) {
-            auto it = m_token2id.find(token);
-            if (it != m_token2id.end()) {
-                token_ids.push_back(it->second);
-            }
-        }
-    }
-    return token_ids;
-}
-
-std::vector<std::vector<int>> Tokenizer::TextsToTokenIds(const std::vector<std::string>& texts) const {
-    std::vector<std::vector<int>> result;
-    result.reserve(texts.size());
-    for (auto& text : texts) {
-        result.push_back(TextToTokenIds(text));
-    }
-    return result;
-}
-
-void Tokenizer::BuildCatTokens(const std::vector<int>& prompt_tokens,
-                                const std::vector<int>& text_tokens,
-                                int max_tokens,
-                                std::vector<int32_t>& out_cat_tokens) const {
-    out_cat_tokens.assign(max_tokens, static_cast<int32_t>(m_pad_id));
-
-    size_t pos = 0;
-    for (size_t i = 0; i < prompt_tokens.size() && pos < (size_t)max_tokens; ++i) {
-        out_cat_tokens[pos++] = static_cast<int32_t>(prompt_tokens[i]);
-    }
-    for (size_t i = 0; i < text_tokens.size() && pos < (size_t)max_tokens; ++i) {
-        out_cat_tokens[pos++] = static_cast<int32_t>(text_tokens[i]);
-    }
-    if (pos < (size_t)max_tokens) {
-        out_cat_tokens[pos] = static_cast<int32_t>(m_pad_id);
-    }
-}

cpp/src/tokenizer.hpp

@@ -1,79 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * Tokenizer: loads tokens.txt and maps token strings to IDs.
- *
- * For board deployment, tokenization should ideally be done on a host machine
- * and token IDs passed directly. This class provides basic tokenization
- * capability for convenience.
- **************************************************************************************************/
-
-#pragma once
-
-#include <string>
-#include <vector>
-#include <unordered_map>
-#include <cstdint>
-
-class Tokenizer {
-public:
-    Tokenizer();
-
-    /**
-     * Load tokens.txt file. Format: token\tid per line.
-     */
-    int Load(const std::string& token_file);
-
-    /**
-     * Convert a single text to token IDs using simplified rules.
-     * For production use, pre-compute token IDs with the Python tokenizer.
-     */
-    std::vector<int> TextToTokenIds(const std::string& text) const;
-
-    /**
-     * Convert multiple texts to token IDs.
-     */
-    std::vector<std::vector<int>> TextsToTokenIds(const std::vector<std::string>& texts) const;
-
-    /**
-     * Get the pad token ID ("_" token).
-     */
-    int GetPadId() const { return m_pad_id; }
-
-    /**
-     * Get vocabulary size.
-     */
-    int GetVocabSize() const { return static_cast<int>(m_token2id.size()); }
-
-    /**
-     * Lookup a single token. Returns pad_id if not found.
-     */
-    int TokenToId(const std::string& token) const;
-
-    /**
-     * Build concatenated token array [prompt_tokens + text_tokens + pad] padded to max_tokens.
-     */
-    void BuildCatTokens(const std::vector<int>& prompt_tokens,
-                        const std::vector<int>& text_tokens,
-                        int max_tokens,
-                        std::vector<int32_t>& out_cat_tokens) const;
-
-    bool IsLoaded() const { return m_loaded; }
-
-private:
-    std::unordered_map<std::string, int> m_token2id;
-    std::vector<std::string> m_id2token;
-    int m_pad_id;
-    bool m_loaded;
-
-    // Simplified text→tokens for Chinese (char-by-char pinyin-like tokens)
-    std::vector<std::string> TokenizeZh(const std::string& text) const;
-    // Simplified text→tokens for English (basic phoneme rules)
-    std::vector<std::string> TokenizeEn(const std::string& text) const;
-    // Split text into language segments
-    std::vector<std::pair<std::string, std::string>> GetSegments(const std::string& text) const;
-
-    static bool IsChinese(char c);
-    static bool IsAlphabet(char c);
-    static std::string MapPunctuation(const std::string& text);
-};

cpp/src/utilities/file.hpp

@@ -1,251 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include <fstream>
-#include <cctype>
-
-#if defined(ENV_HAS_STD_FILESYSTEM)
-#include <filesystem>
-#endif // ENV_HAS_STD_FILESYSTEM
-#if defined(ENV_HAS_POSIX_FILE_STAT)
-#include <sys/types.h>
-#include <sys/stat.h>
-#include <dirent.h>
-#include <cstring>
-#include "utilities/scalar_guard.hpp"
-#endif // ENV_HAS_POSIX_FILE_STAT
-#if defined(ENV_HAS_WIN_API)
-#include <Windows.h>
-#endif // ENV_HAS_WIN_API
-
-#include <string>
-
-namespace utilities {
-constexpr auto error_size = static_cast<uintmax_t>(-1);
-
-enum class file_type {
-    none,
-    not_found,
-    regular,
-    directory,
-    symlink,
-    block,
-    character,
-    fifo,
-    socket,
-    unknown,
-};
-
-inline bool exists(const std::string& path) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    return std::filesystem::exists(path);
-#elif defined(ENV_HAS_POSIX_FILE_STAT)
-    struct ::stat path_stat{};
-    return 0 == ::stat(path.c_str(), &path_stat);
-#else
-#error "Unsupported platform, native file system API is required."
-#endif
-}
-
-inline uintmax_t file_size(const std::string& path) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    return std::filesystem::file_size(path);
-#elif defined(ENV_HAS_POSIX_FILE_STAT)
-    struct ::stat path_stat{};
-    if (0 == ::stat(path.c_str(), &path_stat)) {
-        return path_stat.st_size;
-    }
-    return error_size;
-#else
-#error "Unsupported platform, native file system API is required."
-#endif
-}
-
-inline file_type status(const std::string& path) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    if (!exists(path)) {
-        return file_type::not_found;
-    }
-
-    const auto status = std::filesystem::status(path);
-    if (std::filesystem::is_regular_file(status)) {
-        return file_type::regular;
-    }
-    if (std::filesystem::is_directory(status)) {
-        return file_type::directory;
-    }
-    if (std::filesystem::is_symlink(status)) {
-        return file_type::symlink;
-    }
-    if (std::filesystem::is_block_file(status)) {
-        return file_type::block;
-    }
-    if (std::filesystem::is_character_file(status)) {
-        return file_type::character;
-    }
-    if (std::filesystem::is_fifo(status)) {
-        return file_type::fifo;
-    }
-    if (std::filesystem::is_socket(status)) {
-        return file_type::socket;
-    }
-    return file_type::unknown;
-#elif defined(ENV_HAS_POSIX_FILE_STAT)
-    struct ::stat path_stat{};
-    if (0 != ::stat(path.c_str(), &path_stat)) {
-        return file_type::not_found;
-    }
-
-    if (S_ISREG(path_stat.st_mode)) {
-        return file_type::regular;
-    }
-    if (S_ISDIR(path_stat.st_mode)) {
-        return file_type::directory;
-    }
-    if (S_ISLNK(path_stat.st_mode)) {
-        return file_type::symlink;
-    }
-    if (S_ISBLK(path_stat.st_mode)) {
-        return file_type::block;
-    }
-    if (S_ISCHR(path_stat.st_mode)) {
-        return file_type::character;
-    }
-    if (S_ISFIFO(path_stat.st_mode)) {
-        return file_type::fifo;
-    }
-    if (S_ISSOCK(path_stat.st_mode)) {
-        return file_type::socket;
-    }
-    return file_type::unknown;
-#else
-#error "Unsupported platform, native file system API is required."
-#endif
-}
-
-inline bool is_regular_file(const std::string& path) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    return std::filesystem::is_regular_file(path);
-#elif defined(ENV_HAS_POSIX_FILE_STAT)
-    struct ::stat path_stat{};
-    return 0 == ::stat(path.c_str(), &path_stat) && S_ISREG(path_stat.st_mode);
-#else
-#error "Unsupported platform, native file system API is required."
-#endif
-}
-
-inline bool is_directory(const std::string& path) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    return std::filesystem::is_directory(path);
-#elif defined(ENV_HAS_POSIX_FILE_STAT)
-    struct ::stat path_stat{};
-    return 0 == ::stat(path.c_str(), &path_stat) && S_ISDIR(path_stat.st_mode);
-#else
-#error "Unsupported platform, native file system API is required."
-#endif
-}
-
-inline bool is_empty(const std::string& path) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    return std::filesystem::is_empty(path);
-#elif defined(ENV_HAS_POSIX_FILE_STAT)
-    if (!exists(path))
-        return false;
-
-    if (is_regular_file(path))
-        return 0 == file_size(path);
-    if (is_directory(path)) {
-        auto dir = scalar_guard<DIR*>(::opendir(path.c_str()), ::closedir);
-        if (nullptr == dir.get()) {
-            return false;
-        }
-
-        struct ::dirent* entry;
-        while ((entry = ::readdir(dir.get())) != nullptr) {
-            if (0 != ::strcmp(entry->d_name, ".") && 0 != ::strcmp(entry->d_name, "..")) {
-                return false;
-            }
-        }
-        return true;
-    }
-    return false;
-#else
-#error "Unsupported platform, native file system API is required."
-#endif
-}
-
-inline bool read(const std::string& file, void* data, const uintmax_t& size) {
-    std::ifstream stream;
-    stream.open(file, std::ios_base::binary | std::ios_base::in);
-    if (stream.is_open()) {
-        stream.read(static_cast<char*>(data), static_cast<std::streamsize>(size));
-        stream.close();
-        return true;
-    }
-    return false;
-}
-
-inline bool write(const std::string& file, const void* data, const uintmax_t& size) {
-    std::ofstream stream;
-    stream.open(file, std::ios_base::binary | std::ios_base::out);
-    if (stream.is_open()) {
-        stream.write(static_cast<const char*>(data), static_cast<std::streamsize>(size));
-        stream.close();
-        return true;
-    }
-    return false;
-}
-
-inline bool create_directory(const std::string& path) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    return std::filesystem::create_directories(path);
-#elif defined(ENV_HAS_POSIX_FILE_STAT)
-    return 0 == ::mkdir(path.c_str(), 0755);
-#else
-#error "Unsupported platform, native file system API is required."
-#endif
-}
-
-inline std::string get_file_name(const std::string& file) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    return std::filesystem::path(file).filename().string();
-#else
-    if (const auto pos = file.find_last_of("/\\"); std::string::npos != pos) {
-        return file.substr(pos + 1);
-    }
-    return file;
-#endif
-}
-
-inline std::string get_file_extension(const std::string& file) {
-#if defined(ENV_HAS_STD_FILESYSTEM)
-    return std::filesystem::path(file).extension().string();
-#else
-    if (const auto pos = file.find_last_of('.'); std::string::npos != pos) {
-        return file.substr(pos + 1);
-    }
-    return "";
-#endif
-}
-
-inline std::string get_legal_name(const std::string& name) {
-    std::string temp = name;
-
-    size_t pos = temp.find_last_of("/\\");
-    if (pos != std::string::npos) pos = 0;
-
-    for (size_t i = pos; i < temp.size(); ++i)
-        if (!(::isalnum(temp[i]) || '-' == temp[i] || '_' == temp[i])) temp[i] = '_';
-    return temp;
-}
-
-}

cpp/src/utilities/file_mapper.hpp

@@ -1,86 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#if defined(ENV_HAS_POSIX_FILE_STAT)
-#include "utilities/scalar_guard.hpp"
-#include <sys/mman.h>
-#include <fcntl.h>
-#include <unistd.h>
-#endif // ENV_HAS_POSIX_FILE_STAT
-
-#if defined(ENV_HAS_WIN_API)
-#include <Windows.h>
-#endif // ENV_HAS_WIN_API
-
-#include <utilities/file.hpp>
-
-namespace utilities {
-
-struct file_mapper {
-    file_mapper() = delete;
-
-    explicit file_mapper(const std::string& path) {
-#if defined(ENV_HAS_POSIX_FILE_STAT)
-
-        auto fd_guard = scalar_guard<int>(
-            ::open(path.c_str(), O_RDONLY),
-            [](const int& fd) { if (fd != -1) ::close(fd); }
-        );
-
-        if (-1 == fd_guard.get()) {
-            return;
-        }
-
-        auto size = file_size(path);
-        auto map_guard = scalar_guard<void*>(
-            ::mmap(nullptr, size, PROT_READ, MAP_SHARED, fd_guard.get(), 0),
-            [&size](void*& addr) { if (MAP_FAILED != addr && nullptr != addr) ::munmap(addr, size); }
-        );
-
-        if (MAP_FAILED == map_guard.get()) {
-            return;
-        }
-
-        this->size_ = size;
-        std::swap(this->fd_, fd_guard.get());
-        std::swap(this->buffer_, map_guard.get());
-#endif // ENV_HAS_POSIX_FILE_STAT
-    }
-
-    ~file_mapper() {
-#if defined(ENV_HAS_POSIX_FILE_STAT)
-        if (nullptr != this->buffer_) {
-            ::munmap(this->buffer_, this->size_);
-            this->buffer_ = nullptr;
-        }
-        if (-1 != this->fd_) {
-            ::close(this->fd_);
-            this->fd_ = -1;
-        }
-#endif // ENV_HAS_POSIX_FILE_STAT
-    }
-
-    [[nodiscard]] void* get() const {
-        return this->buffer_;
-    }
-
-    [[nodiscard]] uint64_t size() const {
-        return this->size_;
-    }
-
-private:
-    int fd_ = -1;
-    void* buffer_ = nullptr;
-    uint64_t size_ = 0;
-};
-
-}

cpp/src/utilities/log.hpp

@@ -1,64 +0,0 @@
-/*
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- * Author: wanglusheng@axera-tech.com
- */
-
-#pragma once
-
-#include <cstdarg>
-#include <cstdio>
-
-namespace utilities {
-    class log
-    {
-    public:
-        enum class type {
-            debug = 0,
-            info  = 1,
-            warn = 2,
-            error = 3,
-        };
-        
-        log() = default;
-        explicit log(const type& mode) : type_(mode) {}
-        ~log() = default;
-
-        void print(const type& mode, const char *fmt, ...) const {
-            if (mode >= this->type_) {
-                switch (mode) {
-                    case type::debug:
-                        ::fprintf(stdout, "[DEBUG] ");
-                    break;
-                    case type::info:
-                        ::fprintf(stdout, "[INFO] ");
-                    break;
-                    case type::warn:
-                        ::fprintf(stdout, "[WARNING] ");
-                    break;
-                    case type::error:
-                        ::fprintf(stdout, "[ERROR] ");
-                    break;
-                }
-
-                ::va_list args;
-                va_start(args, fmt);
-                ::vprintf( fmt, args);
-                va_end(args);
-            }
-            ::fflush(stdout);
-        }
-        void set_level(const type& mode) {
-            this->type_ = mode;
-        }
-
-    private:
-        type type_ = type::warn;
-    };
-
-    static log glog;
-}

cpp/src/utilities/scalar_guard.hpp

@@ -1,60 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include <functional>
-#include <vector>
-#include <utility>
-#include <stdexcept>
-
-namespace utilities {
-
-    template<typename T>
-    class scalar_guard {
-    public:
-        scalar_guard() = delete;
-
-        scalar_guard(T resource, std::function<void(T&)> destructor)
-            : instance_(resource), destructor_(std::move(destructor)) {}
-
-        scalar_guard(std::function<T()> resource_creator, std::function<void(T&)> destructor)
-            : instance_(resource_creator()), destructor_(std::move(destructor)) {}
-
-        scalar_guard(scalar_guard&& other) noexcept
-            : instance_(std::exchange(other.instance_, T{})), destructor_(std::move(other.destructor_)) {}
-
-        scalar_guard& operator=(scalar_guard&& other) noexcept {
-            if (this != &other) {
-                instance_ = std::exchange(other.instance_, T{});
-                destructor_ = std::move(other.destructor_);
-            }
-            return *this;
-        }
-
-        scalar_guard(const scalar_guard&) = delete;
-        scalar_guard& operator=(const scalar_guard&) = delete;
-
-        ~scalar_guard() noexcept {
-            if (destructor_) {
-                destructor_(instance_);
-            }
-        }
-
-        T& get() {
-            return instance_;
-        }
-
-    private:
-        T instance_;
-        std::function<void(T&)> destructor_;
-    };
-
-}

cpp/src/utilities/timer.hpp

@@ -1,84 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include <chrono>
-
-namespace utilities {
-
-class timer {
-public:
-    using nanoseconds = std::chrono::nanoseconds;
-    using microseconds = std::chrono::microseconds;
-    using milliseconds = std::chrono::milliseconds;
-    using seconds = std::chrono::seconds;
-    using minutes = std::chrono::minutes;
-    using hours = std::chrono::hours;
-
-    timer() {
-        start();
-    }
-
-    void start() {
-        stop();
-        this->start_ = this->end_;
-    }
-
-    void stop() {
-#ifdef _MSC_VER
-        this->end_ = std::chrono::system_clock::now();
-#else
-        this->end_ = std::chrono::high_resolution_clock::now();
-#endif
-    }
-
-    template<typename T>
-    float elapsed() {
-        if (this->end_ <= this->start_) {
-            this->stop();
-        }
-
-        if (std::is_same_v<T, std::chrono::nanoseconds>) {
-            const auto t = std::chrono::duration_cast<std::chrono::nanoseconds>(this->end_ - this->start_).count();
-            return static_cast<float>(t);
-        }
-        if (std::is_same_v<T, std::chrono::microseconds>) {
-            const auto t = std::chrono::duration_cast<std::chrono::nanoseconds>(this->end_ - this->start_).count();
-            return static_cast<float>(t) / 1000.f;
-        }
-        if (std::is_same_v<T, std::chrono::milliseconds>) {
-            const auto t = std::chrono::duration_cast<std::chrono::microseconds>(this->end_ - this->start_).count();
-            return static_cast<float>(t) / 1000.f;
-        }
-        if (std::is_same_v<T, std::chrono::seconds>) {
-            const auto t = std::chrono::duration_cast<std::chrono::milliseconds>(this->end_ - this->start_).count();
-            return static_cast<float>(t) / 1000.f;
-        }
-        if (std::is_same_v<T, std::chrono::minutes>) {
-            const auto t = std::chrono::duration_cast<std::chrono::milliseconds>(this->end_ - this->start_).count();
-            return static_cast<float>(t) / (60.f * 1000.f);
-        }
-        if (std::is_same_v<T, std::chrono::hours>) {
-            const auto t = std::chrono::duration_cast<std::chrono::seconds>(this->end_ - this->start_).count();
-            return static_cast<float>(t) / (60.f * 60.f);
-        }
-        return 0.f;
-    }
-
-    float elapsed() {
-        return elapsed<std::chrono::milliseconds>();
-    }
-
-private:
-    std::chrono::system_clock::time_point start_, end_;
-};
-
-}

cpp/src/utilities/vector_guard.hpp

@@ -1,64 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2024 Axera Semiconductor Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor Co., Ltd.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include <functional>
-#include <vector>
-
-namespace utilities {
-
-    template<typename T>
-    class vector_guard {
-    public:
-        vector_guard() = delete;
-
-        vector_guard(std::vector<T> resources, std::function<void(T&)> destructor)
-            : instance_(std::move(resources)), destructor_(std::move(destructor)) {}
-
-        vector_guard(std::function<std::vector<T>()> resource_creator, std::function<void(T&)> destructor)
-            : instance_(resource_creator()), destructor_(std::move(destructor)) {}
-
-        vector_guard(vector_guard&& other) noexcept
-            : instance_(std::exchange(other.instance_, std::vector<T>{})), destructor_(std::move(other.destructor_)) {}
-
-        vector_guard& operator=(vector_guard&& other) noexcept {
-            if (this != &other) {
-                instance_ = std::exchange(other.instance_, std::vector<T>{});
-                destructor_ = std::move(other.destructor_);
-            }
-            return *this;
-        }
-
-        vector_guard(const vector_guard&) = delete;
-        vector_guard& operator=(const vector_guard&) = delete;
-
-        ~vector_guard() noexcept {
-            if (destructor_) {
-                for (auto& resource : instance_) {
-                    destructor_(resource);
-                }
-            }
-        }
-
-        std::vector<T>& get() {
-            return instance_;
-        }
-
-        T* data() {
-            return instance_.data();
-        }
-
-    private:
-        std::vector<T> instance_;
-        std::function<void(T&)> destructor_;
-    };
-
-}

cpp/src/vocoder.cpp

@@ -1,190 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * Vocoder implementation using kissfft for fast IRFFT.
- **************************************************************************************************/
-
-#include "src/vocoder.hpp"
-
-#include <cmath>
-#include <cstring>
-#include <cstdio>
-#include <algorithm>
-
-#ifndef M_PI
-#define M_PI 3.14159265358979323846
-#endif
-
-Vocoder::Vocoder() : m_has_init(false), m_n_freqs(0) {}
-Vocoder::~Vocoder() {
-    m_session.reset();
-    if (m_fft_cfg)  { rnn_fft_free(m_fft_cfg, 0);  m_fft_cfg = nullptr; }
-    m_ifft_cfg = nullptr;  // shared with m_fft_cfg
-}
-
-int Vocoder::Init(const Config& cfg) {
-    m_cfg = cfg;
-    m_n_freqs = cfg.n_fft / 2 + 1;  // 513
-
-    BuildWindow();
-
-    // Init kissfft once
-    m_fft_cfg = rnn_fft_alloc(cfg.n_fft, NULL, NULL, 0);
-    if (!m_fft_cfg) {
-        printf("Failed to init kissfft for vocoder\n");
-        return -1;
-    }
-    m_ifft_cfg = m_fft_cfg;
-
-    // Load backbone (or full) model
-    m_session = std::make_unique<EngineWrapper>();
-    if (m_session->Init(cfg.model_path.c_str(), 0, cfg.axclConfig) != 0) {
-        printf("Failed to load vocoder model: %s\n", cfg.model_path.c_str());
-        return -1;
-    }
-
-    // Load head_linear model (optional, for split mode)
-    if (!cfg.head_model_path.empty()) {
-        m_session_head = std::make_unique<EngineWrapper>();
-        if (m_session_head->Init(cfg.head_model_path.c_str(), 0, cfg.axclConfig) != 0) {
-            printf("Failed to load vocoder head model: %s\n", cfg.head_model_path.c_str());
-            return -1;
-        }
-    }
-
-    m_has_init = true;
-    printf("Vocoder initialized: n_fft=%d, hop=%d, n_mels=%d%s\n",
-           cfg.n_fft, cfg.hop_length, cfg.n_mels,
-           m_session_head ? " (split mode)" : "");
-    return 0;
-}
-
-void Vocoder::BuildWindow() {
-    int n = m_cfg.n_fft;
-    m_window.resize(n);
-    m_window_sq.resize(n);
-    for (int i = 0; i < n; ++i) {
-        m_window[i] = 0.5f * (1.0f - std::cos(2.0f * M_PI * i / (n - 1)));
-        m_window_sq[i] = m_window[i] * m_window[i];
-    }
-}
-
-int Vocoder::Decode(const std::vector<float>& mel, int T, float feat_scale, std::vector<float>& audio) {
-    if (!m_has_init || mel.empty()) return -1;
-
-    int feat_dim = m_cfg.n_mels;      // 100
-    int n_fft = m_cfg.n_fft;          // 1024
-    int n_freqs = m_n_freqs;          // 513
-    int hop = m_cfg.hop_length;       // 256
-
-    // 1. Undo feat_scale and transpose to [1, n_mels, T_model]
-    float inv_scale = (feat_scale != 0.0f) ? (1.0f / feat_scale) : 1.0f;
-    int T_model = 620;
-    std::vector<float> mel_padded(T_model * feat_dim, 0.0f);
-    for (int t = 0; t < T; ++t) {
-        for (int c = 0; c < feat_dim; ++c) {
-            mel_padded[c * T_model + t] = mel[t * feat_dim + c] * inv_scale;
-        }
-    }
-
-    // 2. Run NPU inference
-    std::vector<float> real_spec, imag_spec;
-    int spec_size;
-
-    if (m_session_head) {
-        // Split mode: backbone → features → head_linear → (real, imag)
-        int in_idx = m_session->GetInputIndex("mel");
-        if (in_idx < 0) in_idx = 0;
-        m_session->SetInput(mel_padded.data(), in_idx);
-        if (m_session->RunSync() != 0) { printf("Vocoder backbone failed\n"); return -1; }
-
-        int feat_idx = m_session->GetOutputIndex("features");
-        if (feat_idx < 0) feat_idx = 0;
-        std::vector<float> features(m_session->GetOutputSize(feat_idx) / sizeof(float));
-        m_session->GetOutput(features.data(), feat_idx);
-
-        int h_in = m_session_head->GetInputIndex("features");
-        if (h_in < 0) h_in = 0;
-        m_session_head->SetInput(features.data(), h_in);
-        if (m_session_head->RunSync() != 0) { printf("Vocoder head failed\n"); return -1; }
-
-        int real_idx = m_session_head->GetOutputIndex("real");
-        int imag_idx = m_session_head->GetOutputIndex("imag");
-        if (real_idx < 0) real_idx = 0;
-        if (imag_idx < 0) imag_idx = 1;
-        spec_size = m_session_head->GetOutputSize(real_idx);
-        real_spec.resize(spec_size / sizeof(float));
-        imag_spec.resize(spec_size / sizeof(float));
-        m_session_head->GetOutput(real_spec.data(), real_idx);
-        m_session_head->GetOutput(imag_spec.data(), imag_idx);
-    } else {
-        // Full model: mel → (real, imag)
-        int in_idx = m_session->GetInputIndex("mel");
-        if (in_idx < 0) in_idx = 0;
-        m_session->SetInput(mel_padded.data(), in_idx);
-        if (m_session->RunSync() != 0) { printf("Vocoder NPU inference failed\n"); return -1; }
-
-        int real_idx = m_session->GetOutputIndex("real");
-        int imag_idx = m_session->GetOutputIndex("imag");
-        if (real_idx < 0) real_idx = 0;
-        if (imag_idx < 0) imag_idx = 1;
-        spec_size = m_session->GetOutputSize(real_idx);
-        real_spec.resize(spec_size / sizeof(float));
-        imag_spec.resize(spec_size / sizeof(float));
-        m_session->GetOutput(real_spec.data(), real_idx);
-        m_session->GetOutput(imag_spec.data(), imag_idx);
-    }
-
-    // 3. kissfft IRFFT + window + overlap-add
-    int out_len = (T - 1) * hop + n_fft;
-    audio.assign(out_len, 0.0f);
-    std::vector<float> envelope(out_len, 0.0f);
-
-    std::vector<kiss_fft_cpx> cx_in(n_fft), cx_out(n_fft);
-
-    for (int t = 0; t < T; ++t) {
-        int spec_offset = t * n_freqs;
-
-        // Build full Hermitian spectrum from [0..N/2]
-        cx_in[0].r = real_spec[spec_offset + 0];
-        cx_in[0].i = 0.0f;
-        for (int k = 1; k < n_freqs - 1; ++k) {
-            cx_in[k].r = real_spec[spec_offset + k];
-            cx_in[k].i = imag_spec[spec_offset + k];
-            cx_in[n_fft - k].r =  real_spec[spec_offset + k];
-            cx_in[n_fft - k].i = -imag_spec[spec_offset + k];
-        }
-        cx_in[n_freqs - 1].r = real_spec[spec_offset + (n_freqs - 1)];
-        cx_in[n_freqs - 1].i = 0.0f;  // Nyquist imag must be 0
-
-        rnn_ifft(m_ifft_cfg, cx_in.data(), cx_out.data(), 0);
-
-        // Window + overlap-add
-        int pos = t * hop;
-        for (int n = 0; n < n_fft; ++n) {
-            float sample = (cx_out[n].r / n_fft) * m_window[n];
-            int p = pos + n;
-            if (p < out_len) {
-                audio[p] += sample;
-                envelope[p] += m_window_sq[n];
-            }
-        }
-    }
-
-    // 4. Normalize by window envelope
-    for (int i = 0; i < out_len; ++i) {
-        if (envelope[i] > 1e-10f) {
-            audio[i] /= envelope[i];
-        }
-    }
-
-    // 5. Center trim (n_fft/2 from each side, matching torch.istft center=True)
-    int pad = n_fft / 2;
-    int trim_len = out_len - 2 * pad;
-    if (trim_len > 0) {
-        std::vector<float> trimmed(audio.begin() + pad, audio.begin() + pad + trim_len);
-        audio = std::move(trimmed);
-    }
-
-    return 0;
-}

cpp/src/vocoder.hpp

@@ -1,68 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * Vocoder: mel features → audio via vocos_full.axmodel + IRFFT + overlap-add.
- **************************************************************************************************/
-
-#pragma once
-
-#include <vector>
-#include <string>
-#include <memory>
-
-#include "src/EngineWrapper.hpp"
-#include "kiss_fft.h"
-
-class Vocoder {
-public:
-    struct Config {
-        std::string model_path;         // full model (or backbone in split mode)
-        std::string head_model_path;    // head_linear model (optional, for split mode)
-        const char* axclConfig = nullptr;
-        int n_fft = 1024;
-        int hop_length = 256;
-        int n_mels = 100;
-        int sample_rate = 24000;
-    };
-
-    Vocoder();
-    ~Vocoder();
-
-    /**
-     * Initialize: load vocos_full.axmodel and pre-compute IRFFT basis.
-     */
-    int Init(const Config& cfg);
-
-    /**
-     * Decode mel features to audio waveform.
-     * @param mel        Mel features [T, n_mels] (flat, row-major)
-     * @param T          Number of mel frames
-     * @param feat_scale Feature scaling factor to undo (Python: features/feat_scale before vocoder)
-     * @param audio      Output audio samples (24kHz float32)
-     */
-    int Decode(const std::vector<float>& mel, int T, float feat_scale, std::vector<float>& audio);
-
-    bool HasInit() const { return m_has_init; }
-
-private:
-    Config m_cfg;
-    bool m_has_init;
-    std::unique_ptr<EngineWrapper> m_session;
-    std::unique_ptr<EngineWrapper> m_session_head;  // optional, for split mode
-
-    // kissfft configs (shared forward/inverse twiddles)
-    kiss_fft_state* m_fft_cfg = nullptr;
-    kiss_fft_state* m_ifft_cfg = nullptr;
-
-    // Hann window
-    std::vector<float> m_window;
-
-    // Window squared for envelope normalization
-    std::vector<float> m_window_sq;
-
-    int m_n_freqs;
-
-    void BuildWindow();
-
-    static void Fft(std::vector<float>& real, std::vector<float>& imag, bool inverse);
-};

cpp/src/wav_writer.hpp

@@ -1,84 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * Simple WAV file writer for float32 mono audio.
- **************************************************************************************************/
-
-#pragma once
-
-#include <vector>
-#include <string>
-#include <fstream>
-#include <cstdint>
-#include <cstring>
-
-class WavWriter {
-public:
-    /**
-     * Write float32 mono audio samples to a WAV file.
-     *
-     * @param path        Output file path
-     * @param samples     Float32 audio samples in range [-1.0, 1.0]
-     * @param sample_rate Sample rate in Hz (e.g., 24000)
-     * @return true on success
-     */
-    static bool Write(const std::string& path,
-                      const std::vector<float>& samples,
-                      int sample_rate) {
-        std::ofstream file(path, std::ios::binary);
-        if (!file.is_open()) return false;
-
-        int num_samples = static_cast<int>(samples.size());
-        int byte_rate = sample_rate * 2;  // 16-bit mono
-        int data_size = num_samples * 2;
-        int chunk_size = 36 + data_size;
-
-        // RIFF header
-        Write4CC(file, "RIFF");
-        WriteU32LE(file, chunk_size);
-        Write4CC(file, "WAVE");
-
-        // fmt chunk
-        Write4CC(file, "fmt ");
-        WriteU32LE(file, 16);           // chunk size
-        WriteU16LE(file, 1);            // PCM format
-        WriteU16LE(file, 1);            // mono
-        WriteU32LE(file, sample_rate);
-        WriteU32LE(file, byte_rate);
-        WriteU16LE(file, 2);            // block align
-        WriteU16LE(file, 16);           // bits per sample
-
-        // data chunk
-        Write4CC(file, "data");
-        WriteU32LE(file, data_size);
-
-        // Convert float32 to int16 (with clipping)
-        for (float sample : samples) {
-            // Clamp to [-1, 1]
-            if (sample > 1.0f) sample = 1.0f;
-            if (sample < -1.0f) sample = -1.0f;
-            int16_t val = static_cast<int16_t>(sample * 32767.0f);
-            WriteU16LE(file, static_cast<uint16_t>(val));
-        }
-
-        file.close();
-        return true;
-    }
-
-private:
-    static void WriteU16LE(std::ofstream& f, uint16_t v) {
-        f.put(v & 0xFF);
-        f.put((v >> 8) & 0xFF);
-    }
-
-    static void WriteU32LE(std::ofstream& f, uint32_t v) {
-        f.put(v & 0xFF);
-        f.put((v >> 8) & 0xFF);
-        f.put((v >> 16) & 0xFF);
-        f.put((v >> 24) & 0xFF);
-    }
-
-    static void Write4CC(std::ofstream& f, const char* s) {
-        f.write(s, 4);
-    }
-};

cpp/src/zipvoice_engine.cpp

@@ -1,616 +0,0 @@
-/**************************************************************************************************
- * ZipVoice AXERA C++ Port
- *
- * ZipVoiceEngine implementation.
- **************************************************************************************************/
-
-#include "zipvoice_engine.hpp"
-
-#include <fstream>
-#include <sstream>
-#include <cmath>
-#include <cstring>
-#include <algorithm>
-#include <sys/time.h>
-
-#ifndef M_PI
-#define M_PI 3.14159265358979323846
-#endif
-
-// Simple JSON value parser (no external dependency)
-// Only parses the simple structures needed by manifest.json and runtime_config.json
-namespace {
-
-std::string json_get_string(const std::string& json, const std::string& key) {
-    std::string search = "\"" + key + "\"";
-    size_t pos = json.find(search);
-    if (pos == std::string::npos) return "";
-
-    pos = json.find(':', pos + search.length());
-    if (pos == std::string::npos) return "";
-
-    pos = json.find('"', pos + 1);
-    if (pos == std::string::npos) return "";
-
-    size_t end = json.find('"', pos + 1);
-    if (end == std::string::npos) return "";
-
-    return json.substr(pos + 1, end - pos - 1);
-}
-
-int json_get_int(const std::string& json, const std::string& key, int default_val = 0) {
-    std::string search = "\"" + key + "\"";
-    size_t pos = json.find(search);
-    if (pos == std::string::npos) return default_val;
-
-    pos = json.find(':', pos + search.length());
-    if (pos == std::string::npos) return default_val;
-
-    // Skip whitespace
-    pos++;
-    while (pos < json.length() && (json[pos] == ' ' || json[pos] == '\t' || json[pos] == '\n')) pos++;
-    if (pos >= json.length()) return default_val;
-
-    // Parse number
-    std::string num_str;
-    while (pos < json.length() && (std::isdigit(json[pos]) || json[pos] == '-' || json[pos] == '.')) {
-        num_str += json[pos];
-        pos++;
-    }
-    if (num_str.empty()) return default_val;
-
-    // Check if float
-    if (num_str.find('.') != std::string::npos) {
-        return static_cast<int>(std::stof(num_str));
-    }
-    return std::stoi(num_str);
-}
-
-float json_get_float(const std::string& json, const std::string& key, float default_val = 0.0f) {
-    std::string search = "\"" + key + "\"";
-    size_t pos = json.find(search);
-    if (pos == std::string::npos) return default_val;
-
-    pos = json.find(':', pos + search.length());
-    if (pos == std::string::npos) return default_val;
-
-    pos++;
-    while (pos < json.length() && (json[pos] == ' ' || json[pos] == '\t' || json[pos] == '\n')) pos++;
-    if (pos >= json.length()) return default_val;
-
-    std::string num_str;
-    while (pos < json.length() && (std::isdigit(json[pos]) || json[pos] == '-' || json[pos] == '.')) {
-        num_str += json[pos];
-        pos++;
-    }
-    if (num_str.empty()) return default_val;
-    return std::stof(num_str);
-}
-
-std::vector<std::string> json_get_string_array(const std::string& json, const std::string& key) {
-    std::vector<std::string> result;
-    std::string search = "\"" + key + "\"";
-    size_t pos = json.find(search);
-    if (pos == std::string::npos) return result;
-
-    pos = json.find('[', pos + search.length());
-    if (pos == std::string::npos) return result;
-
-    size_t end = json.find(']', pos);
-    if (end == std::string::npos) return result;
-
-    std::string array_str = json.substr(pos + 1, end - pos - 1);
-
-    size_t start = 0;
-    while (start < array_str.length()) {
-        size_t q1 = array_str.find('"', start);
-        if (q1 == std::string::npos) break;
-        size_t q2 = array_str.find('"', q1 + 1);
-        if (q2 == std::string::npos) break;
-        result.push_back(array_str.substr(q1 + 1, q2 - q1 - 1));
-        start = q2 + 1;
-    }
-    return result;
-}
-
-std::string read_file_content(const std::string& path) {
-    std::ifstream file(path);
-    if (!file.is_open()) return "";
-    std::stringstream ss;
-    ss << file.rdbuf();
-    return ss.str();
-}
-
-} // anonymous namespace
-
-ZipVoiceEngine::ZipVoiceEngine()
-    : m_has_init(false), m_decoder_seq_len(0), m_decoder_has_padding_mask(false) {}
-
-ZipVoiceEngine::~ZipVoiceEngine() {
-    m_sessions.clear();
-}
-
-double ZipVoiceEngine::GetCurrentTimeMs() {
-    struct timeval tv;
-    gettimeofday(&tv, nullptr);
-    return tv.tv_sec * 1000.0 + tv.tv_usec / 1000.0;
-}
-
-int ZipVoiceEngine::Init(const std::string& model_dir, const char* axclConfig) {
-    if (LoadConfig(model_dir) != 0) return -1;
-    if (LoadManifest(model_dir) != 0) return -1;
-    if (LoadModels(model_dir, axclConfig) != 0) return -1;
-    if (LoadDecoderMetadata() != 0) return -1;
-
-    m_has_init = true;
-    printf("ZipVoiceEngine initialized: max_tokens=%d, max_feat_len=%d, feat_dim=%d, num_step=%d\n",
-           m_config.max_tokens, m_config.max_feat_len, m_config.feat_dim, m_config.num_step);
-    return 0;
-}
-
-int ZipVoiceEngine::LoadConfig(const std::string& model_dir) {
-    std::string config_path = model_dir + "/runtime_config.json";
-    std::string json = read_file_content(config_path);
-    if (!json.empty()) {
-        m_config.max_tokens = json_get_int(json, "max_tokens", 384);
-        m_config.max_feat_len = json_get_int(json, "max_feat_len", 1024);
-        m_config.feat_dim = json_get_int(json, "feat_dim", 100);
-        m_config.sampling_rate = json_get_int(json, "sampling_rate", 24000);
-        m_config.hop_length = json_get_int(json, "hop_length", 256);
-        m_config.num_step = json_get_int(json, "num_step", 10);
-        m_config.t_shift = json_get_float(json, "t_shift", 0.5f);
-        m_config.guidance_scale = json_get_float(json, "guidance_scale", 1.0f);
-    }
-    m_config.model_dir = model_dir;
-    return 0;
-}
-
-int ZipVoiceEngine::LoadManifest(const std::string& model_dir) {
-    std::string manifest_path = model_dir + "/decoder4_split_manifest.json";
-    std::string json = read_file_content(manifest_path);
-    if (json.empty()) {
-        printf("Failed to read manifest: %s\n", manifest_path.c_str());
-        return -1;
-    }
-
-    // Parse encoder info
-    m_encoder_info.name = json_get_string(json, "name");
-    // Need to parse the nested "encoder" object
-    size_t enc_pos = json.find("\"encoder\"");
-    if (enc_pos != std::string::npos) {
-        size_t obj_start = json.find('{', enc_pos);
-        size_t obj_end = json.find('}', obj_start);
-        if (obj_start != std::string::npos && obj_end != std::string::npos) {
-            std::string enc_json = json.substr(obj_start, obj_end - obj_start + 1);
-            m_encoder_info.name = json_get_string(enc_json, "name");
-            m_encoder_info.file = json_get_string(enc_json, "file");
-            m_encoder_info.inputs = json_get_string_array(enc_json, "inputs");
-            m_encoder_info.outputs = json_get_string_array(enc_json, "outputs");
-        }
-    }
-
-    // Parse decoder_parts array
-    size_t dec_pos = json.find("\"decoder_parts\"");
-    if (dec_pos != std::string::npos) {
-        size_t arr_start = json.find('[', dec_pos);
-        if (arr_start != std::string::npos) {
-            m_decoder_parts.clear();
-            size_t pos = arr_start + 1;
-            int depth = 0;
-            std::string part_json;
-            for (; pos < json.length(); ++pos) {
-                if (json[pos] == '{') depth++;
-                if (depth > 0) part_json += json[pos];
-                if (json[pos] == '}') {
-                    depth--;
-                    if (depth == 0) {
-                        // Parse this part
-                        ModelInfo info;
-                        info.name = json_get_string(part_json, "name");
-                        info.file = json_get_string(part_json, "file");
-                        info.inputs = json_get_string_array(part_json, "inputs");
-                        info.outputs = json_get_string_array(part_json, "outputs");
-                        m_decoder_parts.push_back(info);
-                        part_json.clear();
-                    }
-                }
-            }
-        }
-    }
-
-    printf("Manifest: encoder=%s, decoder_parts=%zu\n",
-           m_encoder_info.name.c_str(), m_decoder_parts.size());
-    for (size_t i = 0; i < m_decoder_parts.size(); ++i) {
-        printf("  part%zu: %s (%s) in=%zu out=%zu\n", i,
-               m_decoder_parts[i].name.c_str(), m_decoder_parts[i].file.c_str(),
-               m_decoder_parts[i].inputs.size(), m_decoder_parts[i].outputs.size());
-    }
-    return 0;
-}
-
-int ZipVoiceEngine::LoadModels(const std::string& model_dir, const char* axclConfig) {
-    // Load encoder
-    std::string enc_path = model_dir + "/" + m_encoder_info.file;
-    auto enc = std::make_unique<EngineWrapper>();
-    if (enc->Init(enc_path.c_str(), 0, axclConfig) != 0) {
-        printf("Failed to load encoder: %s\n", enc_path.c_str());
-        return -1;
-    }
-    m_sessions[m_encoder_info.name] = std::move(enc);
-
-    // Load decoder parts
-    for (auto& part : m_decoder_parts) {
-        std::string path = model_dir + "/" + part.file;
-        auto sess = std::make_unique<EngineWrapper>();
-        if (sess->Init(path.c_str(), 0, axclConfig) != 0) {
-            printf("Failed to load decoder part: %s\n", path.c_str());
-            return -1;
-        }
-        m_sessions[part.name] = std::move(sess);
-    }
-
-    printf("Loaded %zu models\n", m_sessions.size());
-    return 0;
-}
-
-int ZipVoiceEngine::LoadDecoderMetadata() {
-    if (m_decoder_parts.empty()) return -1;
-
-    auto& part0 = m_decoder_parts[0];
-    auto it = m_sessions.find(part0.name);
-    if (it == m_sessions.end()) return -1;
-
-    auto& sess = it->second;
-
-    // Check if padding_mask is an input
-    int pad_idx = sess->GetInputIndex("padding_mask");
-    m_decoder_has_padding_mask = (pad_idx >= 0);
-
-    // Get sequence length from x input shape
-    m_decoder_seq_len = m_config.max_feat_len;
-
-    // Try to get seq_len from model metadata
-    // The shape info is in the AX engine io_info; for simplicity we use config value
-
-    printf("Decoder metadata: seq_len=%d, has_padding_mask=%d\n",
-           m_decoder_seq_len, m_decoder_has_padding_mask ? 1 : 0);
-    return 0;
-}
-
-std::vector<float> ZipVoiceEngine::GetTimesteps(int num_step, float t_shift) const {
-    std::vector<float> ts(num_step + 1);
-    for (int i = 0; i <= num_step; ++i) {
-        float t = static_cast<float>(i) / num_step;
-        ts[i] = t_shift * t / (1.0f + (t_shift - 1.0f) * t);
-    }
-    return ts;
-}
-
-int ZipVoiceEngine::RunEncoder(const std::vector<int32_t>& cat_tokens,
-                                std::vector<float>& out_encoded) {
-    auto it = m_sessions.find(m_encoder_info.name);
-    if (it == m_sessions.end()) return -1;
-
-    auto& sess = it->second;
-
-    // Set input
-    int input_idx = sess->GetInputIndex(m_encoder_info.inputs[0].c_str());
-    if (input_idx < 0) input_idx = 0;
-
-    // Need to copy to non-const buffer for the engine API
-    std::vector<int32_t> input_copy = cat_tokens;
-    sess->SetInput(input_copy.data(), input_idx);
-
-    // Run
-    if (sess->RunSync() != 0) return -1;
-
-    // Get output
-    int output_idx = sess->GetOutputIndex(m_encoder_info.outputs[0].c_str());
-    if (output_idx < 0) output_idx = 0;
-
-    int output_size = sess->GetOutputSize(output_idx);
-    out_encoded.resize(output_size / sizeof(float));
-    sess->GetOutput(out_encoded.data(), output_idx);
-
-    return 0;
-}
-
-int ZipVoiceEngine::DurationExpand(const std::vector<float>& encoded,
-                                    int prompt_tokens_len,
-                                    int text_tokens_len,
-                                    int prompt_features_len,
-                                    float speed,
-                                    std::vector<float>& out_text_condition,
-                                    int& out_features_len) {
-    int total_tokens_len = prompt_tokens_len + text_tokens_len;
-    if (total_tokens_len <= 0) return -1;
-
-    // Compute target features length
-    int features_len = static_cast<int>(
-        std::ceil(static_cast<float>(prompt_features_len) / prompt_tokens_len
-                  * total_tokens_len / speed)
-    );
-    if (features_len > m_config.max_feat_len) {
-        features_len = m_config.max_feat_len;
-    }
-
-    int feat_dim = m_config.feat_dim;
-
-    // encoded shape: [1, max_tokens, feat_dim] (flat array)
-    // Extract the no-pad portion
-    int token_dur = features_len / total_tokens_len;
-    int max_tokens = m_config.max_tokens;
-
-    // out_text_condition: [1, features_len, feat_dim]
-    out_text_condition.assign(features_len * feat_dim, 0.0f);
-
-    // Repeat each token embedding token_dur times
-    for (int t = 0; t < total_tokens_len; ++t) {
-        int base_idx = t * feat_dim;
-        for (int d = 0; d < token_dur; ++d) {
-            int frame_idx = t * token_dur + d;
-            if (frame_idx >= features_len) break;
-            int dst_idx = frame_idx * feat_dim;
-            std::copy(encoded.begin() + base_idx,
-                      encoded.begin() + base_idx + feat_dim,
-                      out_text_condition.begin() + dst_idx);
-        }
-    }
-
-    // Fill residual frames with last token embedding
-    int filled = total_tokens_len * token_dur;
-    int residual = features_len - filled;
-    if (residual > 0) {
-        int last_base = total_tokens_len * feat_dim;
-        for (int d = 0; d < residual; ++d) {
-            int dst_idx = (filled + d) * feat_dim;
-            std::copy(encoded.begin() + last_base,
-                      encoded.begin() + last_base + feat_dim,
-                      out_text_condition.begin() + dst_idx);
-        }
-    }
-
-    out_features_len = features_len;
-    return 0;
-}
-
-int ZipVoiceEngine::RunDecoderPart(const ModelInfo& part,
-                                    std::map<std::string, std::vector<float>>& values,
-                                    const std::vector<uint8_t>* padding_mask_data,
-                                    std::vector<uint8_t>* padding_mask2_out) {
-    auto it = m_sessions.find(part.name);
-    if (it == m_sessions.end()) return -1;
-    auto& sess = it->second;
-
-    // Set inputs in order
-    for (size_t i = 0; i < part.inputs.size(); ++i) {
-        const std::string& expected_name = part.inputs[i];
-
-        int input_idx = sess->GetInputIndex(expected_name.c_str());
-        if (input_idx < 0) input_idx = static_cast<int>(i);
-
-        // padding_mask: use uint8 data from caller (part0 only)
-        if (expected_name == "padding_mask") {
-            if (padding_mask_data) {
-                sess->SetInput((void*)padding_mask_data->data(), input_idx);
-                continue;
-            }
-        }
-        // padding_mask2: use raw uint8 data from part0's output
-        if (expected_name == "padding_mask2") {
-            if (padding_mask2_out && !padding_mask2_out->empty()) {
-                sess->SetInput((void*)padding_mask2_out->data(), input_idx);
-                continue;
-            }
-            printf("ERROR: padding_mask2 not available for '%s'\n", part.name.c_str());
-            return -1;
-        }
-
-        // Find the value in float map
-        auto val_it = values.find(expected_name);
-        if (val_it == values.end()) {
-            printf("Missing input '%s' for model '%s'\n",
-                   expected_name.c_str(), part.name.c_str());
-            return -1;
-        }
-
-        sess->SetInput((void*)val_it->second.data(), input_idx);
-    }
-
-    // Run
-    if (sess->RunSync() != 0) return -1;
-
-    // Get outputs
-    for (size_t i = 0; i < part.outputs.size(); ++i) {
-        const std::string& name = part.outputs[i];
-
-        int output_idx = sess->GetOutputIndex(name.c_str());
-        if (output_idx < 0) output_idx = static_cast<int>(i);
-
-        int size = sess->GetOutputSize(output_idx);
-        if (size <= 0) {
-            printf("Invalid output size for '%s' in '%s'\n",
-                   name.c_str(), part.name.c_str());
-            return -1;
-        }
-
-        // padding_mask2 is uint8 — read raw bytes, don't convert to float
-        if (name == "padding_mask2" && padding_mask2_out) {
-            padding_mask2_out->resize(size);
-            sess->GetOutput(padding_mask2_out->data(), output_idx);
-            continue;
-        }
-
-        std::vector<float> output_data(size / sizeof(float));
-        sess->GetOutput(output_data.data(), output_idx);
-        values[name] = std::move(output_data);
-    }
-
-    return 0;
-}
-
-int ZipVoiceEngine::Sample(const std::vector<int32_t>& cat_tokens,
-                            int prompt_tokens_len,
-                            int text_tokens_len,
-                            const std::vector<float>& prompt_features,
-                            int prompt_features_len,
-                            float speed,
-                            float guidance_scale,
-                            int seed,
-                            std::vector<float>& out_features,
-                            Timing& out_timing) {
-    double t_total_start = GetCurrentTimeMs();
-
-    // 1. Run encoder
-    double t_start = GetCurrentTimeMs();
-    std::vector<float> encoded;
-    if (RunEncoder(cat_tokens, encoded) != 0) return -1;
-    out_timing.encoder_time_sec = static_cast<float>(GetCurrentTimeMs() - t_start) / 1000.0f;
-
-    // encoded shape: [max_tokens * feat_dim] (flattened, batch=1)
-    int feat_dim = m_config.feat_dim;
-
-    // 2. Duration expand
-    t_start = GetCurrentTimeMs();
-    std::vector<float> text_condition;
-    int features_len;
-    if (DurationExpand(encoded, prompt_tokens_len, text_tokens_len,
-                       prompt_features_len, speed,
-                       text_condition, features_len) != 0) return -1;
-    out_timing.duration_expand_time_sec = static_cast<float>(GetCurrentTimeMs() - t_start) / 1000.0f;
-
-    // 3. Prepare decoder inputs
-    int seq_len = m_decoder_seq_len > 0 ? m_decoder_seq_len : m_config.max_feat_len;
-
-    if (features_len > seq_len) {
-        printf("features_len=%d exceeds decoder seq_len=%d\n", features_len, seq_len);
-        return -1;
-    }
-    if (prompt_features_len > seq_len) {
-        printf("prompt_features_len=%d exceeds decoder seq_len=%d\n", prompt_features_len, seq_len);
-        return -1;
-    }
-
-    // text_cond_padded: [1, seq_len, feat_dim] (zeros, filled up to features_len)
-    std::vector<float> text_cond_padded(seq_len * feat_dim, 0.0f);
-    std::copy(text_condition.begin(),
-              text_condition.begin() + features_len * feat_dim,
-              text_cond_padded.begin());
-
-    // speech_cond_padded: [1, seq_len, feat_dim]
-    std::vector<float> speech_cond_padded(seq_len * feat_dim, 0.0f);
-    std::copy(prompt_features.begin(),
-              prompt_features.begin() + prompt_features_len * feat_dim,
-              speech_cond_padded.begin());
-    // DEBUG
-
-    // padding_mask: [1, seq_len] as uint8_t (bool), NOT float32!
-    // Python: np.zeros((1, seq_len), dtype=np.bool_)
-    std::vector<uint8_t> padding_mask(seq_len, 0);
-    for (int i = features_len; i < seq_len; ++i) {
-        padding_mask[i] = 1;
-    }
-
-    // x: random init [1, seq_len, feat_dim]
-    std::vector<float> x(seq_len * feat_dim, 0.0f);
-    // Simple LCG random
-    uint32_t rng_state = static_cast<uint32_t>(seed > 0 ? seed : 42);
-    for (int i = 0; i < features_len * feat_dim; ++i) {
-        rng_state = rng_state * 1103515245 + 12345;
-        // Box-Muller for normal distribution
-        float u1 = static_cast<float>(rng_state & 0x7FFFFFFF) / 0x7FFFFFFF;
-        rng_state = rng_state * 1103515245 + 12345;
-        float u2 = static_cast<float>(rng_state & 0x7FFFFFFF) / 0x7FFFFFFF;
-        x[i] = std::sqrt(-2.0f * std::log(std::max(u1, 1e-10f)))
-               * std::cos(2.0f * M_PI * u2);
-    }
-
-    // Time steps
-    std::vector<float> timesteps = GetTimesteps(m_config.num_step, m_config.t_shift);
-
-    // Guidance scale as single float array
-    std::vector<float> gs = { guidance_scale };
-
-    // 4. Flow-matching decoder
-    double t_dec_total = 0.0;
-    for (int step = 0; step < m_config.num_step; ++step) {
-        double t_step_start = GetCurrentTimeMs();
-
-        // Prepare inputs for this step
-        std::map<std::string, std::vector<float>> values;
-
-        // t: scalar
-        std::vector<float> t_val = { timesteps[step] };
-        values["t"] = t_val;
-
-        // x: [1, seq_len, feat_dim]
-        values["x"] = x;
-
-        // text_condition
-        values["text_condition"] = text_cond_padded;
-
-        // speech_condition
-        values["speech_condition"] = speech_cond_padded;
-
-        // guidance_scale: scalar
-        values["guidance_scale"] = gs;
-
-        // padding_mask is passed separately as uint8_t (bool), not in values map
-        std::vector<uint8_t> padding_mask2_data;  // populated by part0, used by parts 1-3
-
-        // Run all decoder parts in sequence
-        for (auto& part : m_decoder_parts) {
-            if (RunDecoderPart(part, values, &padding_mask, &padding_mask2_data) != 0) return -1;
-        }
-
-        t_dec_total += GetCurrentTimeMs() - t_step_start;
-
-        // Get v (velocity) from last part output
-        const std::string& final_output_name = m_decoder_parts.back().outputs[0];
-        auto vit = values.find(final_output_name);
-        if (vit == values.end()) {
-            printf("Missing final output '%s'\n", final_output_name.c_str());
-            return -1;
-        }
-
-        const std::vector<float>& v = vit->second;
-
-        // Euler step: x = x + v * dt
-        float dt = timesteps[step + 1] - timesteps[step];
-        for (size_t i = 0; i < x.size(); ++i) {
-            x[i] += v[i] * dt;
-        }
-
-        // Zero out padding region
-        for (int i = features_len * feat_dim; i < seq_len * feat_dim; ++i) {
-            x[i] = 0.0f;
-        }
-    }
-
-    out_timing.decoder_time_sec = static_cast<float>(t_dec_total) / 1000.0f;
-
-    // 5. Extract generated features (excluding prompt region)
-    int generated_frames = features_len - prompt_features_len;
-    if (generated_frames <= 0) {
-        generated_frames = features_len;
-        out_features.assign(x.begin(), x.begin() + features_len * feat_dim);
-    } else {
-        int offset = prompt_features_len * feat_dim;
-        out_features.assign(x.begin() + offset,
-                            x.begin() + features_len * feat_dim);
-    }
-
-    out_timing.generated_frames = generated_frames;
-    out_timing.features_len = features_len;
-    out_timing.total_time_sec = static_cast<float>(GetCurrentTimeMs() - t_total_start) / 1000.0f;
-
-    printf("  encoder: %.3f s  dur_expand: %.3f s  decoder(%d steps): %.3f s (avg %.3f ms/step)  total: %.3f s\n",
-           out_timing.encoder_time_sec,
-           out_timing.duration_expand_time_sec,
-           m_config.num_step,
-           out_timing.decoder_time_sec,
-           out_timing.decoder_time_sec / m_config.num_step * 1000.0f,
-           out_timing.total_time_sec);
-
-    return 0;
-}