Method build

src/framework/modules/conv_modules.cpp:233–291

    // (mirrors the Metal supports gate); otherwise keep the mul_mat + add chain
    const bool use_acc = in_channels >= 64 && output_frames > 8;
    ggml_tensor * acc = nullptr;
    for (int64_t tap = 0; tap < kernel_size; ++tap) {
        // columns[c, j] = input[tap * dilation + j, c]: contiguous column view of input_cf.
        auto * columns = ggml_view_2d(
            ctx.ggml,
            input_cf,
            in_channels,
            output_frames,
            input_cf->nb[1],
            static_cast<size_t>(tap * dilation) * in_channels * sizeof(float));
        auto * tap_weights = ggml_view_2d(
            ctx.ggml,
            weight_taps,
            in_channels,
            out_channels,
            weight_taps->nb[1],
            static_cast<size_t>(tap) * in_channels * sizeof(float));
        if (acc == nullptr) {
            acc = ggml_mul_mat(ctx.ggml, tap_weights, columns);
        } else if (use_acc) {
            acc = ggml_mul_mat_acc(ctx.ggml, tap_weights, columns, acc);
        } else {
            acc = ggml_add(ctx.ggml, acc, ggml_mul_mat(ctx.ggml, tap_weights, columns));
        }
    }
    return acc;
}

// Metal fast path for stride-1 conv1d on the time-fast [frames, channels] layout used by
// the audio codecs. ggml_conv_1d materializes an im2col matrix whose kernel taps are
// strided gathers in this layout (~200 ms per conv at [569k, 96] on M4); instead, transpose
// the input to channel-fast once, run one contiguous GEMM per kernel tap, and transpose
// the accumulator back (~3-6x faster).
core::TensorValue build_conv1d_pertap_fast_path(
    core::ModuleBuildContext & ctx,
    const Conv1dConfig & config,
    const core::TensorValue & input,
    const core::TensorValue & weight_f32,
    const core::TensorShape & output_shape) {
    // channel-fast copy of the input: [IC, frames]; kernel taps become contiguous columns.
    auto * input_cf = ggml_cont(ctx.ggml, ggml_transpose(ctx.ggml, input.tensor));
    auto * acc = conv1d_pertap_gemm_channel_fast(
        ctx,
        input_cf,
        weight_f32.tensor,
        config.in_channels,
        config.out_channels,
        config.kernel_size,
        config.dilation,
        output_shape.dims[2]);
    // mul_mat yields [OC, frames]; restore the canonical [frames, OC] orientation.
    return core::wrap_tensor(
        ggml_cont(ctx.ggml, ggml_transpose(ctx.ggml, acc)),
        output_shape,
        GGML_TYPE_F32);
}

int64_t conv2d_output_dim(int64_t input, int kernel, int stride, int padding, int dilation) {