rt_utils.rs get_sola_offset 中的卷积似乎引入了额外计算量 #1

New issue

Closed

opened 2026-10-20 09:52:55 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-20 09:52:55 +00:00

Owner

Originally created by @TYPEmber on GitHub (May 23, 2024).
Original GitHub issue: https://github.com/RVC-Project/obs-rvc/issues/3

这两个地方似乎都只是想要做一维卷积，但是却都先转换到了三维再进行卷积，这样会引入额外的计算。

    let conv_input_size = buffer_frame_size + search_frame_size;
    let conv_input = input_buffer
        .slice(s![..conv_input_size])
        .into_shape((1, 1, conv_input_size))?;
    let sola_buffer_view = sola_buffer
        .into_shape((1, 1, buffer_frame_size))?;

    let cor_nom = conv_input.conv_fft(
        &sola_buffer_view, ndarray_conv::ConvMode::Valid, 
        ndarray_conv::PaddingMode::Zeros
    )?;

    let cor_den_filler = ndarray::Array3::<f32>::ones((1, 1, buffer_frame_size));
    let mut cor_den = conv_input.mapv(|x| x.powi(2)).conv_fft(
        &cor_den_filler, ndarray_conv::ConvMode::Valid, 
        ndarray_conv::PaddingMode::Zeros
    )?;

Originally created by @TYPEmber on GitHub (May 23, 2024). Original GitHub issue: https://github.com/RVC-Project/obs-rvc/issues/3 这两个地方似乎都只是想要做一维卷积，但是却都先转换到了三维再进行卷积，这样会引入额外的计算。 ```rust let conv_input_size = buffer_frame_size + search_frame_size; let conv_input = input_buffer .slice(s![..conv_input_size]) .into_shape((1, 1, conv_input_size))?; let sola_buffer_view = sola_buffer .into_shape((1, 1, buffer_frame_size))?; let cor_nom = conv_input.conv_fft( &sola_buffer_view, ndarray_conv::ConvMode::Valid, ndarray_conv::PaddingMode::Zeros )?; let cor_den_filler = ndarray::Array3::<f32>::ones((1, 1, buffer_frame_size)); let mut cor_den = conv_input.mapv(|x| x.powi(2)).conv_fft( &cor_den_filler, ndarray_conv::ConvMode::Valid, ndarray_conv::PaddingMode::Zeros )?; ```

bsakaguchi referenced this issue

2026-10-20 09:52:55 +00:00

[GH-ISSUE #1] Retrieval-based Voice Conversion VST plugin For DAW (Cubase, Nuendo) #3

bsakaguchi closed this issue

2026-10-20 09:52:55 +00:00

bsakaguchi commented

2026-10-20 09:52:56 +00:00

Author

Owner

@wfjsw commented on GitHub (May 24, 2024):

因为 torch 做 conv1d 的时候是这样的，我就这么写了。看起来好像确实没有必要。

@wfjsw commented on GitHub (May 24, 2024): 因为 torch 做 conv1d 的时候是这样的，我就这么写了。看起来好像确实没有必要。