tulip_rs 0.1.15

//use crate::common::validate_inputs;
use crate::common_simd::options::{validate_inputs, validate_options};
use crate::indicators::simd_indicators::road_train::{Asset, Driver, PrimeMover};
use crate::indicators::simd_indicators::wma_simd::SimdState;
use crate::indicators::wma::{
    min_data, multiplier, output_length, IndicatorState, State, INPUTS_WIDTH, OPTIONS_WIDTH,
};
use crate::types::IndicatorError;
use std::simd::Simd;
struct Params {
    multipliers: (f64, f64, f64),
    period: usize,
}
/// SIMD driver for the Weighted Moving Average (WMA) indicator, processing `N` option-set lanes per scheduling epoch.
struct WmaDriver {
    want_optional_outputs: bool,
}

impl Driver<State, Params> for WmaDriver {
    /// Processes one epoch of output bars for `N` option-set lanes simultaneously using SIMD.
    fn next_run<const N: usize>(
        &mut self,
        inputs: Vec<Vec<&[f64]>>,
        mut outputs: Vec<Vec<&mut [f64]>>,
        mut states: Vec<&mut State>,
        options: Vec<Option<&Params>>,
    ) {
        let len = outputs[0][0].len();

        // Optimization 1: Direct array construction instead of collect+try_into
        let mut state = SimdState::new(&states);
        let mut i = [0usize; N];
        let multiplier_simd = {
            let mut multipliers = ([0.0; N], [0.0; N], [0.0; N]);

            for (lane, &option) in options.iter().enumerate() {
                if let Some(param) = option {
                    i[lane] = param.period;
                    multipliers.0[lane] = param.multipliers.0;
                    multipliers.1[lane] = param.multipliers.1;
                    multipliers.2[lane] = param.multipliers.2;
                }
            }
            (
                Simd::from_array(multipliers.0),
                Simd::from_array(multipliers.1),
                Simd::from_array(multipliers.2),
            )
        };

        // Optimization 2: Pre-compute all input and output pointers
        let input_ptrs = crate::extract_input_ptrs!(inputs, N, real_ptrs);
        let (wma_line_ptr, sma_line_ptr) =
            crate::extract_output_ptrs!(outputs, N, wma_line_ptr, sma_line_ptr);

        // Optimization 3: Simplified main loop with pre-computed offsets
        for j in 0..len {
            let old_vals = crate::extract_simd_inputs_at_index!(j, N, old @ input_ptrs);
            let new_vals = crate::extract_simd_inputs_at_index_array!(i, N, new @ input_ptrs);
            // Get new and old values using pre-computed pointers
            let (wma, sma) = state.calc_simd(old_vals, new_vals, multiplier_simd);

            // Store results using pre-computed pointers
            crate::write_simd_at_indices!(N, j,
                wma_line_ptr => wma
            );
            crate::store_simd_optional_outputs!(j, N,
                self.want_optional_outputs, sma_line_ptr => sma
            );

            for i in i.iter_mut() {
                *i += 1;
            }
        }

        state.write_states(&mut states);
    }
}

/// Calculates the Weighted Moving Average (WMA) for one shared asset across `N` different
/// option sets simultaneously using SIMD parallelism.
///
/// Uses the [`PrimeMover`] scheduler to batch option sets into SIMD-width groups.
///
/// # Arguments
/// * `inputs` - Shared input data: `inputs[0]` is `&[f64]` containing `real` (price series).
/// * `options` - An array of `N` option sets; `options[i]` is `&[f64; OPTIONS_WIDTH]` containing
///   `[period]` for option set `i`.
/// * `optional_outputs` - Optional slice controlling extra output series;
///   index 0 enables `sma`.
///
/// # Returns
/// `Ok((outputs, states))` where `outputs[i][0]` is `wma` and `outputs[i][1]` is `sma`
/// (empty unless requested) for option set `i`,
/// and `states[i]` is the final [`IndicatorState`] for option set `i`.
/// Returns `Err(IndicatorError)` if any input slice is too short or any option set is invalid.
pub fn indicator_by_options<const N: usize>(
    inputs: &[&[f64]; INPUTS_WIDTH],
    options: &[&[f64; OPTIONS_WIDTH]; N],
    optional_outputs: Option<&[bool]>,
) -> Result<(Vec<Vec<Vec<f64>>>, Vec<IndicatorState>), IndicatorError> {
    validate_inputs::<OPTIONS_WIDTH>(inputs, options, min_data)?;
    validate_options(options, None)?;
    let params: [Params; N] = std::array::from_fn(|i| Params {
        multipliers: multiplier(options[i][0] as usize),
        period: options[i][0] as usize,
    });

    let mut road_train = PrimeMover::<N, State, Params>::new();
    let mut output_buffers = Vec::with_capacity(N);
    let mut want_optional_outputs = false;

    for (i, param) in params.iter().enumerate() {
        let asset_inputs = vec![inputs[0]];
        let (wma_line, sma_line) = {
            let capacity = output_length(inputs[0].len(), options[i]);
            (
                crate::uninit_vec!(f64, capacity),
                crate::init_optional_outputs_eff!(
                    optional_outputs, &[false],
                    sma_line: capacity
                ),
            )
        };
        if i == 0 {
            (_, want_optional_outputs) = crate::calc_want_flags!(sma_line);
        }
        let mut output_buffer = vec![wma_line, sma_line];
        let state = State::init_state(&inputs[0][0..param.period]);

        //let adosc_len = output_buffer[0].len();
        let mut asset_outputs = Vec::with_capacity(output_buffer.len());

        for j in 0..output_buffer.len() {
            unsafe {
                //let slice_len = output_buffer.len() - starts[j];
                // Get a mutable reference to the output buffer for this asset
                let output_buffer = &mut output_buffer[j];
                asset_outputs.push(std::slice::from_raw_parts_mut(
                    output_buffer.as_mut_ptr(), //slice from
                    output_buffer.len(),        // slice to
                ));
            }
        }
        road_train.add_asset(Asset::new(
            asset_inputs,
            asset_outputs,
            i,
            param.period,
            param.period,
            state,
            Some(param),
        ));
        output_buffers.push(output_buffer);
    }
    let mut driver = WmaDriver {
        want_optional_outputs,
    };
    let states_vec = road_train.drive(&mut driver);

    let mut indicator_states = Vec::with_capacity(N);
    for (state, param) in states_vec.into_iter().zip(params.into_iter()) {
        indicator_states.push(IndicatorState::new(
            unsafe { inputs.get_unchecked(0) },
            param.multipliers,
            state,
            param.period,
        ));
    }
    Ok((output_buffers, indicator_states))
}