conv based speed stft

src/layer/spectrogram.cpp

@@ -207,4 +207,4 @@ int Spectrogram::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
     return 0;
 }
 
-} // namespace ncnn
+} // namespace ncnn

src/layer/spectrogram.h

@@ -33,4 +33,4 @@ class Spectrogram : public Layer
 
 } // namespace ncnn
 
-#endif // LAYER_SPECTROGRAM_H
+#endif // LAYER_SPECTROGRAM_H

src/layer/x86/spectrogram_x86.cpp

@@ -0,0 +1,277 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "spectrogram_x86.h"
+
+namespace ncnn {
+
+Spectrogram_x86::Spectrogram_x86()
+    : conv_transpose(0)
+{
+    one_blob_only = true;
+    support_inplace = false;
+}
+
+Spectrogram_x86::~Spectrogram_x86()
+{
+    delete conv_transpose;
+}
+
+int Spectrogram_x86::load_param(const ParamDict& pd)
+{
+    n_fft = pd.get(0, 0);
+    power = pd.get(1, 0);
+    hoplen = pd.get(2, n_fft / 4);
+    winlen = pd.get(3, n_fft);
+    window_type = pd.get(4, 0);
+    center = pd.get(5, 1);
+    pad_type = pd.get(6, 2);
+    normalized = pd.get(7, 0);
+    onesided = pd.get(8, 1);
+
+    // assert winlen <= n_fft
+    // generate window
+    window_data.create(n_fft);
+    {
+        float* p = window_data;
+        for (int i = 0; i < (n_fft - winlen) / 2; i++)
+        {
+            *p++ = 0.f;
+        }
+        if (window_type == 0)
+        {
+            // all ones
+            for (int i = 0; i < winlen; i++)
+            {
+                *p++ = 1.f;
+            }
+        }
+        if (window_type == 1)
+        {
+            // hann window
+            for (int i = 0; i < winlen; i++)
+            {
+                *p++ = 0.5f * (1 - cosf(2 * 3.14159265358979323846 * i / winlen));
+            }
+        }
+        if (window_type == 2)
+        {
+            // hamming window
+            for (int i = 0; i < winlen; i++)
+            {
+                *p++ = 0.54f - 0.46f * cosf(2 * 3.14159265358979323846 * i / winlen);
+            }
+        }
+        for (int i = 0; i < n_fft - winlen - (n_fft - winlen) / 2; i++)
+        {
+            *p++ = 0.f;
+        }
+
+        // pre-calculated window norm factor
+        if (normalized == 2)
+        {
+            float sqsum = 0.f;
+            for (int i = 0; i < n_fft; i++)
+            {
+                sqsum += window_data[i] * window_data[i];
+            }
+            float scale = 1.f / sqrt(sqsum);
+
+            for (int i = 0; i < n_fft; i++)
+            {
+                window_data[i] *= scale;
+            }
+        }
+    }
+
+    Mat theta;
+    if (onesided)
+    {
+        n_freq = n_fft / 2 + 1;
+    }
+    else
+    {
+        n_freq = n_fft;
+    }
+    theta.create(n_fft, n_freq, size_t(8));
+
+    for (int i = 0; i < n_freq; i++)
+    {
+        for (int j = 0; j < n_fft; j++)
+        {
+            theta.row<double>(i)[j] = 2 * 3.14159265358979323846 * i * j / n_fft;
+        }
+    }
+
+    Mat real_basis, imag_basis;
+    real_basis.create(n_fft, n_freq, size_t(8));
+    imag_basis.create(n_fft, n_freq, size_t(8));
+
+    for (int i = 0; i < n_freq; i++)
+    {
+        for (int j = 0; j < n_fft; j++)
+        {
+            real_basis.row<double>(i)[j] = cos(theta.row<double>(i)[j]);
+            imag_basis.row<double>(i)[j] = -sin(theta.row<double>(i)[j]);
+        }
+    }
+
+    // multiply window
+    for (int i = 0; i < n_freq; i++)
+    {
+        for (int j = 0; j < n_fft; j++)
+        {
+            real_basis.row<double>(i)[j] *= window_data[j];
+            imag_basis.row<double>(i)[j] *= window_data[j];
+        }
+    }
+
+    if (normalized == 1)
+    {
+        double scale = 1.f / sqrt(n_fft);
+        for (int i = 0; i < n_freq; i++)
+        {
+            for (int j = 0; j < n_fft; j++)
+            {
+                real_basis.row<double>(i)[j] *= scale;
+                imag_basis.row<double>(i)[j] *= scale;
+            }
+        }
+    }
+
+    conv_data.create(n_fft, 1, n_freq * 2);
+
+    for (int i = 0; i < n_freq; i++)
+    {
+        for (int j = 0; j < n_fft; j++)
+        {
+            conv_data.channel(i).row<float>(0)[j] = (float)real_basis.row<double>(i)[j];
+            conv_data.channel(i + n_freq).row<float>(0)[j] = (float)imag_basis.row<double>(i)[j];
+        }
+    }
+
+    conv_transpose = ncnn::create_layer("Convolution1D");
+    ncnn::ParamDict conv_transpose_pd;
+
+    conv_transpose_pd.set(0, 2 * n_freq); // num_output
+    conv_transpose_pd.set(1, n_fft);      // kernel_w
+    conv_transpose_pd.set(3, hoplen);     // stride_w
+    conv_transpose_pd.set(19, 1);         // dynamic_weight
+
+    conv_transpose->load_param(conv_transpose_pd);
+
+    return 0;
+}
+
+int Spectrogram_x86::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
+{
+    // https://pytorch.org/audio/stable/generated/torchaudio.functional.spectrogram.html
+
+    // TODO custom window
+
+    Mat bottom_blob_bordered = bottom_blob;
+    if (center == 1)
+    {
+        Option opt_b = opt;
+        opt_b.blob_allocator = opt.workspace_allocator;
+        if (pad_type == 0)
+            copy_make_border(bottom_blob, bottom_blob_bordered, 0, 0, n_fft / 2, n_fft / 2, BORDER_CONSTANT, 0.f, opt_b);
+        if (pad_type == 1)
+            copy_make_border(bottom_blob, bottom_blob_bordered, 0, 0, n_fft / 2, n_fft / 2, BORDER_REPLICATE, 0.f, opt_b);
+        if (pad_type == 2)
+            copy_make_border(bottom_blob, bottom_blob_bordered, 0, 0, n_fft / 2, n_fft / 2, BORDER_REFLECT, 0.f, opt_b);
+    }
+
+    const int size = bottom_blob_bordered.w;
+
+    // const int frames = size / hoplen + 1;
+    const int frames = (size - n_fft) / hoplen + 1;
+
+    const size_t elemsize = bottom_blob_bordered.elemsize;
+
+    if (elemsize != sizeof(float))
+    {
+        return -100;
+    }
+
+    if (power == 0)
+    {
+        top_blob.create(2, frames, n_freq, elemsize, opt.blob_allocator);
+    }
+    else
+    {
+        top_blob.create(frames, n_freq, elemsize, opt.blob_allocator);
+    }
+    if (top_blob.empty())
+        return -100;
+
+    std::vector<Mat> inputs;
+    inputs.push_back(bottom_blob_bordered);
+    inputs.push_back(conv_data);
+
+    std::vector<Mat> outputs;
+    outputs.push_back(Mat());
+
+    Option opt_conv = opt;
+    opt_conv.use_packing_layout = false;
+
+    conv_transpose->create_pipeline(opt_conv);
+    conv_transpose->forward(inputs, outputs, opt_conv);
+    conv_transpose->destroy_pipeline(opt_conv);
+
+    Mat conv_top_blob = outputs[0]; // (2 * n_freq, frames)
+    float* conv_top_data = conv_top_blob;
+
+    if (power == 0) // as complex
+    {
+        // copy
+        for (int i = 0; i < frames; i++)
+        {
+            for (int j = 0; j < n_freq; j++)
+            {
+                top_blob.channel(j).row<float>(i)[0] = conv_top_data[j * frames + i];
+                top_blob.channel(j).row<float>(i)[1] = conv_top_data[(j + n_freq) * frames + i];
+            }
+        }
+    }
+    else
+    {
+        if (power == 1) // magnitude sqrt(re * re + im * im);
+        {
+            // copy
+            for (int i = 0; i < frames; i++)
+            {
+                for (int j = 0; j < n_freq; j++)
+                {
+                    top_blob.row<float>(j)[i] = sqrtf(conv_top_data[j * frames + i] * conv_top_data[j * frames + i] + conv_top_data[(j + n_freq) * frames + i] * conv_top_data[(j + n_freq) * frames + i]);
+                }
+            }
+        }
+        else if (power == 2) // power re * re + im * im;
+        {
+            // copy
+            for (int i = 0; i < frames; i++)
+            {
+                for (int j = 0; j < n_freq; j++)
+                {
+                    top_blob.row<float>(j)[i] = conv_top_data[j * frames + i] * conv_top_data[j * frames + i] + conv_top_data[(j + n_freq) * frames + i] * conv_top_data[(j + n_freq) * frames + i];
+                }
+            }
+        }
+    }
+
+    return 0;
+}
+
+} // namespace ncnn

src/layer/x86/spectrogram_x86.h

@@ -0,0 +1,53 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#ifndef LAYER_SPECTROGRAM_X86_H
+#define LAYER_SPECTROGRAM_X86_H
+
+#include "spectrogram.h"
+
+namespace ncnn {
+
+class Spectrogram_x86 : public Spectrogram
+{
+public:
+    Spectrogram_x86();
+    ~Spectrogram_x86();
+
+    virtual int load_param(const ParamDict& pd);
+
+    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
+
+public:
+    int n_fft;
+    int power;
+    int hoplen;
+    int winlen;
+    int window_type; // 0=ones 1=hann 2=hamming
+    int center;
+    int pad_type;   // 0=CONSTANT 1=REPLICATE 2=REFLECT
+    int normalized; // 0=disabled 1=sqrt(n_fft) 2=window-l2-energy
+    int onesided;
+
+    int n_freq;
+
+    Mat window_data;
+    Mat conv_data;
+
+    Layer* conv_transpose;
+};
+
+} // namespace ncnn
+
+#endif // LAYER_SPECTROGRAM_X86_H