Function activation1d

src/framework/audio/flashsr.cpp:236–285

}

core::TensorValue activation1d(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & input,
    const ActivationParams & params,
    const FlashSrWeights & weights) {
    constexpr int64_t pad = kFlashSrActivationKernel / kFlashSrActivationRatio - 1;
    constexpr int64_t crop_left = pad * kFlashSrActivationRatio +
                                  (kFlashSrActivationKernel - kFlashSrActivationRatio) / 2;
    constexpr int64_t crop_right = pad * kFlashSrActivationRatio +
                                   (kFlashSrActivationKernel - kFlashSrActivationRatio + 1) / 2;
    auto padded = replicate_pad1d(ctx, input, pad, pad);
    auto up = modules::ConvTranspose1dModule({
        kFlashSrChannels,
        kFlashSrChannels,
        kFlashSrActivationKernel,
        kFlashSrActivationRatio,
        0,
        1,
        false,
    }).build(ctx, padded, weights.upsample_filter);
    up = modules::SliceModule({2, crop_left, up.shape.dims[2] - crop_left - crop_right}).build(ctx, up);

    auto alpha = core::reshape_tensor(ctx, params.alpha, core::TensorShape::from_dims({1, kFlashSrChannels, 1}));
    alpha = core::wrap_tensor(ggml_repeat(ctx.ggml, alpha.tensor, up.tensor), up.shape, GGML_TYPE_F32);
    auto inv_beta = core::reshape_tensor(ctx, params.inv_beta, core::TensorShape::from_dims({1, kFlashSrChannels, 1}));
    inv_beta = core::wrap_tensor(ggml_repeat(ctx.ggml, inv_beta.tensor, up.tensor), up.shape, GGML_TYPE_F32);
    auto periodic = core::wrap_tensor(
        ggml_sqr(ctx.ggml, ggml_sin(ctx.ggml, ggml_mul(ctx.ggml, up.tensor, alpha.tensor))),
        up.shape,
        GGML_TYPE_F32);
    auto snake = core::wrap_tensor(
        ggml_add(ctx.ggml, up.tensor, ggml_mul(ctx.ggml, periodic.tensor, inv_beta.tensor)),
        up.shape,
        GGML_TYPE_F32);

    auto down_padded = replicate_pad1d(
        ctx,
        snake,
        kFlashSrActivationKernel / 2 - 1,
        kFlashSrActivationKernel / 2);
    return modules::Conv1dModule({
        kFlashSrChannels,
        kFlashSrChannels,
        kFlashSrActivationKernel,
        kFlashSrActivationRatio,
        0,
        1,
        false,
    }).build(ctx, down_padded, weights.downsample_filter);
}

core::TensorValue resblock(
    core::ModuleBuildContext & ctx,