tensor-rs 0.5.9

#![allow(clippy::comparison_chain)]
use std::fmt;
use std::cmp;

pub mod compare_tensor;
pub mod convolution;
pub mod elemwise;
pub mod index_slicing;
pub mod linalg;
pub mod reduction;
pub mod rand;

#[cfg(feature = "use-serde")]
use serde::{Serialize, Deserialize};

/// Naive tensor implementation, single thread
#[cfg_attr(feature = "use-serde", derive(Serialize, Deserialize))]
pub struct GenTensor<T> {
    d: Vec<T>, // GenTensor is stored in row major, 
    dim: Vec<usize>, // higher/outer dimension is on the lower index.
}
impl<T> GenTensor<T> where T: num_traits::Float {

    /// Creation Ops
    
    pub fn new() -> GenTensor<T> {
        GenTensor { d: Vec::<T>::new(), dim: Vec::new() }
    }

    /// Create a tensor with given Vec.
    pub fn new_raw(data: &[T], shape: &[usize]) -> GenTensor<T> {
        let new_data = data.to_vec();
        let new_dim = shape.to_vec();
        GenTensor {
            d: new_data,
            dim: new_dim,
        }
    }

    pub fn new_move(data: Vec::<T>, shape: Vec::<usize>) -> GenTensor<T>{
        GenTensor {
            d: data,
            dim: shape,
        }
    }

    pub fn data_copy(&mut self, other: &GenTensor<T>) {
        self.d = other.d.clone();
        self.dim = other.dim.clone();
    }

    /// Convert 1 dim index to multi-dim index.
    pub fn index2dimpos(&self, index: usize) -> Vec::<usize> {
        if index >= self.d.len() {
            panic!("index out of range, {:?}, {:?}", index, self.d.len());
        }
        let mut ret = Vec::new();
        let mut reminder = index;
        for i in &self.stride() {
            //println!("{}", reminder);
            ret.push(reminder/i);
            reminder %= i;
        }
        ret
    }

    /// Convert multi-dim index to 1 dim index.
    pub fn dimpos2index(&self, dimpos: &[usize]) -> usize {
        if dimpos.len() != self.dim.len() {
            panic!("get expects the same dim self.dim: {:?}, o: {:?}", self.dim, dimpos);
        }
        for (i, j) in self.dim.iter().zip(dimpos.iter()) {
            if j >= i {
                panic!("get expects the dim within range self.dim: {:?}, o: {:?}", self.dim, dimpos);
            }
        }
        let mut ret = 0;
        for (st, i) in self.stride().iter().zip(dimpos.iter()) {
            //println!("{}", reminder);
            ret += st*i;
        }
        ret
    }

    // 
    // as_tensor
    // as_strided
    // from_ndarray
    // zeros
    pub fn zeros(size: &[usize]) -> GenTensor<T> {
        let cap = size.iter().product();
        GenTensor {
            d: vec![T::zero(); cap],
            dim: size.to_vec(),
        }
    }
    // zeros_like
    pub fn zeros_like(&self) -> GenTensor<T> {
        let new_data = vec![T::zero(); self.d.len()];
        let new_dim = self.dim.to_vec();
        GenTensor {
            d: new_data,
            dim: new_dim,
        }
    }

    // ones
    pub fn ones(size: &[usize]) -> GenTensor<T> {
        let cap = size.iter().product();
        GenTensor {
            d: vec![T::one(); cap],
            dim: size.to_vec(),
        }
    }
    // ones_like
    pub fn ones_like(&self) -> GenTensor<T> {
        let new_data = vec![T::one(); self.d.len()];
        let new_dim = self.dim.to_vec();
        GenTensor {
            d: new_data,
            dim: new_dim,
        }
    }
    // arange
    pub fn arange(end: usize) -> GenTensor<T> {
        let mut ret = GenTensor::<T>::zeros(&[end]);
        for i in 0..end {
            ret.d[i] = T::from(i).expect("");
        }
        ret
    }
    // range
    // linspace
    // logspace
    pub fn eye(n: usize, m: usize) -> GenTensor<T> {
        let cap = n*m;
        let d = vec![T::zero(); cap];
        let mut ret = GenTensor {
            d,
            dim: [n, m].to_vec(),
        };
        for i in 0..cmp::min(n, m) {
            ret.set(&[i, i], T::one());
        }
        ret
    }
    ////pub fn empty(shape: &[usize]) -> GenTensor<T> {
    ////    let mut elem = 1;
    ////    for i in shape {
    ////        elem *= i;
    ////    }
    ////    
    ////    let mut new_data = Vec::with_capacity(elem);
    ////    unsafe{ new_data.set_len(elem); }
    ////    let new_dim = shape.to_vec();
    ////    GenTensor {
    ////        d: new_data,
    ////        dim: new_dim,
    ////    }
    ////}
    ////// empty_like
    ////pub fn empty_like(&self) -> GenTensor<T> {
    ////    let elem = self.dim.iter().product();
    ////    
    ////    let mut new_data = Vec::with_capacity(elem);
    ////    unsafe{ new_data.set_len(elem); }
    ////    GenTensor {
    ////        d: new_data,
    ////        dim: self.dim.clone(),
    ////    }
    ////}
    // empty_stided
    // full
    // full_like
    // quantize_per_tensor
    // quantize_per_channel
    // 

    /// Create a tensor filled with the same value d
    ///
    /// ```
    /// # use tensor_rs::tensor_impl::gen_tensor::*;
    /// let m1 = GenTensor::<f64>::fill(1., &vec![3,5,2]);
    /// ```
    pub fn fill(d: T, shape: &[usize]) -> GenTensor<T> {
        let mut dsize = 1;
        for i in shape {
            dsize *= *i;
        }
        GenTensor {
            d: vec![d; dsize],
            dim: shape.to_vec(),
        }
    }
    /// assign a row.
    pub fn from_record_f32(&mut self, row: usize, record: &[f32]) -> Result<(), &'static str> {
        for (i, index) in record.iter().zip(0..self.dim[self.dim.len()-1]) {
            self.d[row*self.dim[self.dim.len()-1] + index] = T::from(*i).expect("");
        }
        Ok(())
    }
    pub fn from_record_f64(&mut self, row: usize, record: &[f64]) -> Result<(), &'static str> {
        for (i, index) in record.iter().zip(0..self.dim[self.dim.len()-1]) {
            self.d[row*self.dim[self.dim.len()-1] + index] = T::from(*i).expect("");
        }
        Ok(())
    }

    /// stride() returns a vector contains steps
    /// for each dimension on linear storage.
    /// Right dimension changes fastest.
    /// Right dimension has the stride 1.
    ///
    /// ```
    /// # use tensor_rs::tensor_impl::gen_tensor::*;
    /// let m1 = GenTensor::<f64>::new_raw(&vec![0.; 3*5*2], &vec![3,5,2]);
    /// assert_eq!(m1.stride(), vec![10,2,1]);
    /// ```
    pub fn stride(&self) -> Vec<usize> {
        let mut ret = vec![0; self.dim.len()];
        let dsize = ret.len();
        for i in 0..dsize {
            if i == 0 {
                ret[dsize-1] = 1;
            } else {
                ret[dsize-i-1] = ret[dsize-i]*self.dim[dsize-i];
            }
        }
        ret
    }
    
    /// Return value at the index of the tensor.
    ///
    /// ```
    /// # use tensor_rs::tensor_impl::gen_tensor::*;
    /// let m1 = GenTensor::<f64>::new_raw(&vec![1.,2.,3.,4.,5.,6.], &vec![2,3]);
    /// assert_eq!(m1.get(&vec![1,1]), 5.);
    /// ```
    pub fn get(&self, o: &[usize]) -> T {
        if o.len() != self.dim.len() {
            panic!("get expects the same dim self.dim: {:?}, o: {:?}", self.dim, o);
        }
        for (i, j) in self.dim.iter().zip(o.iter()) {
            if j >= i {
                panic!("get expects the dim within range self.dim: {:?}, o: {:?}", self.dim, o);
            }
        }
        let stride = self.stride();
        let dsize = o.len();
        let mut index = 0;
        //println!("{:?}", stride);
        for i in 0..dsize {
            index += stride[i]*o[i];
        }
        //println!("index: {:?}", index);
        self.d[index]
    }
    pub fn set(&mut self, o: &[usize], v: T) {
        if o.len() != self.dim.len() {
            panic!("get expects the same dim self.dim: {:?}, o: {:?}", self.dim, o);
        }
        for (i, j) in self.dim.iter().zip(o.iter()) {
            if j >= i {
                panic!("get expects the dim within range self.dim: {:?}, o: {:?}", self.dim, o);
            }
        }
        let stride = self.stride();
        let dsize = o.len();
        let mut index = 0;
        for i in 0..dsize {
            index += stride[i]*o[i];
        }
        self.d[index] = v;
    }
    pub fn set_1d(&mut self, o: usize, v: T) {
        if o < self.d.len() {
            self.d[o] = v;
        } else {
            panic!("o {} is beyond limit {}", o, self.d.len());
        }
    }
    pub fn get_mut(&mut self, o: &[usize]) -> &mut T {
        if o.len() != self.dim.len() {
            panic!("get expects the same dim self.dim: {:?}, o: {:?}", self.dim, o);
        }
        for (i, j) in self.dim.iter().zip(o.iter()) {
            if j >= i {
                panic!("get expects the dim within range self.dim: {:?}, o: {:?}", self.dim, o);
            }
        }
        let stride = self.stride();
        let dsize = o.len();
        let mut index = 0;
        for i in 0..dsize {
            index += stride[i]*o[i];
        }
        &mut self.d[index]
    }

    /// dump the underlying vec
    pub fn get_raw(&self) -> Vec<T> {
        self.d.to_vec()
    }
    pub fn get_u8(&self) -> Option<Vec<u8>> {
        let mut ret = Vec::<u8>::with_capacity(self.d.len());
        for i in &self.d {
            let val = i.to_u8()?;
            ret.push(val);
        }
        Some(ret)
    }
    
    /// dump the single value in the tensor
    /// if it is the single value in the tensor.
    pub fn get_scale(&self) -> T {
        if self.d.len() == 1 {
            self.d[0]
        } else {
            panic!("Only one element tensor can get_scale()");
        }
    }
    
    // get NCHW elements
    /// get NCHW elements, always return the size of left most dimension.
    pub fn get_n(&self) -> GenTensor<T> {
        GenTensor {
            d: vec![T::from(self.dim[0]).expect("N")],
            dim: vec![1],
        }
    }
    /// get NCHW elements, always return the size of second left most dimension.
    pub fn get_c(&self) -> GenTensor<T> {
        GenTensor {
            d: vec![T::from(self.dim[1]).expect("N")],
            dim: vec![1],
        }
    }
    /// get NCDHW elements, will require the self.dim has 5 dimensions.
    pub fn get_d(&self) -> GenTensor<T> {
        if self.dim.len() == 5 {
            GenTensor {
                d: vec![T::from(self.dim[2]).expect("N")],
                dim: vec![1],
            }            
        } else {
            panic!("Bad shape for get_D");
        }

    }
    /// get NCDHW elements, will require the self.dim has 5 dimensions or 4 dimensions.
    pub fn get_h(&self) -> GenTensor<T> {
        if self.dim.len() == 5 {
            GenTensor {
                d: vec![T::from(self.dim[3]).expect("N")],
                dim: vec![1],
            }
        } else if self.dim.len() == 4 {
            GenTensor {
                d: vec![T::from(self.dim[2]).expect("N")],
                dim: vec![1],
            }
        } else {
            panic!("Bad shape for get_D");
        }
    }
    /// get NCDHW elements, will require the self.dim has 5 dimensions or 4 dimensions.
    pub fn get_w(&self) -> GenTensor<T> {
        if self.dim.len() == 5 {
            GenTensor {
                d: vec![T::from(self.dim[4]).expect("N")],
                dim: vec![1],
            }
        } else if self.dim.len() == 4 {
            GenTensor {
                d: vec![T::from(self.dim[3]).expect("N")],
                dim: vec![1],
            }
        } else {
            panic!("Bad shape for get_D");
        }
    }

    /// Returns the size of the self tensor.
    pub fn size(&self) -> &Vec<usize> {
        &self.dim
    }
    pub fn get_size(&self) -> &Vec<usize> {
        self.size()
    }
    pub fn get_data(&self) -> &Vec<T> {
        &self.d
    }
    pub fn get_data_mut(&mut self) -> &mut Vec<T> {
        &mut self.d
    }

    /// Returns the total number of elements in the input tensor
    pub fn numel(&self) -> usize {
        self.d.len()
    }

    /// Returns the total number of elements in the input tensor
    pub fn numel_tensor(&self) -> GenTensor<T> {
        GenTensor {
            d: vec![T::from(self.d.len()).expect(""),],
            dim: vec![1],
        }
    }

    /// Return portion of the image.
    /// Every range of each dim with inclusive start and exclusive end.
    /// The pixel can be skipped by setting step larger than 1.
    pub fn get_patch(&self, range: &[(usize, usize)], step: Option<&[usize]>) -> GenTensor<T> {
        if range.len() != self.dim.len() {
            panic!("Expect range covers all dimension range: {:?}, dim: {:?}", range, self.dim);
        }
        let mut step_dim = vec![1; self.dim.len()];
        if let Some(step_val) = step {
            step_dim = step_val.to_vec();
        }

        // index store the the index needs visit at each dim.
        let mut index = Vec::<Vec::<usize>>::new();
        let mut ret_dim = Vec::new();
        for (i, dim_index) in range.iter().zip(0..self.dim.len()) {
            let mut pos = i.0;
            let mut all_index = Vec::new();
            while pos < i.1 {
                all_index.push(pos);
                pos += step_dim[dim_index];
            }
            //println!("{:?}", &all_index);
            ret_dim.push(all_index.len());
            index.push(all_index);
        }
        let mut ret = Self::zeros(&ret_dim);

        let d = self.dim.len();
        let mut pos_index = vec![0; d];
        let mut self_index = vec![0; d];
        loop {
            //println!("pos_index: {:?}", pos_index);
            for i in 0..d {
                self_index[i] = index[i][pos_index[i]];
            }
            let value = self.get(&self_index);
            ret.set(&pos_index, value);

            for dim_index in 0..d {
                pos_index[d-1-dim_index] += 1;
                if pos_index[d-1-dim_index] >= ret.dim[d-1-dim_index] {
                    pos_index[d-1-dim_index] = 0;
                } else {
                    break;
                }
            }

            if pos_index == vec![0; d] {
                break;
            }
        }
        
        ret
    }

    pub fn set_patch(&mut self, val: &GenTensor<T>, range: &[(usize, usize)], step: Option<&[usize]>) {
        if range.len() != self.dim.len() {
            panic!("Expect range covers all dimension range: {:?}, dim: {:?}", range, self.dim);
        }

        let mut step_dim = vec![1; self.dim.len()];
        if let Some(step_val) = step {
            step_dim = step_val.to_vec();
        }

        // index store the the index needs visit at each dim.
        let mut index = Vec::<Vec::<usize>>::new();
        for (i, dim_index) in range.iter().zip(0..self.dim.len()) {
            let mut pos = i.0;
            let mut all_index = Vec::new();
            while pos < i.1 {
                all_index.push(pos);
                pos += step_dim[dim_index];
            }
            //println!("{:?}", &all_index);
            index.push(all_index);
        }

        let d = self.dim.len();
        let mut pos_index = vec![0; d];
        let mut self_index = vec![0; d];
        loop {
            //println!("pos_index: {:?}", pos_index);
            for i in 0..d {
                self_index[i] = index[i][pos_index[i]];
            }
            //let value = self.get(&self_index);
            //ret.set(&pos_index, value);
            self.set(&self_index, val.get(&pos_index));

            for dim_index in 0..d {
                pos_index[d-1-dim_index] += 1;
                if pos_index[d-1-dim_index] >= val.size()[d-1-dim_index] {
                    pos_index[d-1-dim_index] = 0;
                } else {
                    break;
                }
            }
            
            if pos_index == vec![0; d] {
                break;
            }
        }
    }

    pub fn _iter_patch<F>(&self, dim: Option<&[usize]>, keep_dim: bool, closure: F) -> GenTensor<T>
    where F: Fn(&[T]) -> T {
        // take the whole tensor as the patch.
        if dim.is_none() {
            let ret_dim = if keep_dim {
                vec![1; self.size().len()]
            } else {
                vec![1]
            };
            return GenTensor::new_raw(&[closure(self.get_data())], &ret_dim)
        }
        let dim = dim.unwrap();

        // build return tensor dimension.
        let mut ret_dim = Vec::new();
        for i in 0..self.size().len() {
            if dim.contains(&i) {
                if keep_dim {
                    ret_dim.push(1);
                }
            } else {
                ret_dim.push(self.size()[i]);
            }
        }
        let mut ret = Self::zeros(&ret_dim);

         
        let kept_dim: Vec<usize> = (0..self.size().len()).filter(|x| !dim.contains(x)).collect();
        let mut index = vec![0; kept_dim.len()];
        loop {
            let mut patch_index: Vec::<(usize, usize)> = Vec::new();
            let mut output_index: Vec<usize> = Vec::new();
            let mut kept_dim_step = 0;
            for i in 0..self.size().len() {
                if dim.contains(&i) {
                    patch_index.push((0, self.size()[i]));
                    if keep_dim {
                        output_index.push(0);
                    }
                } else {
                    patch_index.push((index[kept_dim_step], index[kept_dim_step]+1));
                    output_index.push(index[kept_dim_step]);
                    kept_dim_step += 1;
                }
            }
            //println!("index: {:?}, patch_index: {:?}, output_index: {:?}", index, patch_index, output_index);

            let value = closure(self.get_patch(&patch_index, None).get_data());
            ret.set(&output_index, value);
            
            for i in 0..index.len() {
                index[kept_dim.len() -i -1] += 1;
                if index[kept_dim.len() -i -1] >= self.size()[kept_dim[kept_dim.len() -i -1]] {
                    index[kept_dim.len() -i -1] = 0;
                } else {
                    break
                }
            }

            if index == vec![0; kept_dim.len()] {
                break
            }
        }
        
        ret
    }

    pub fn _dim_statistic<F>(&self, dim: usize, keepdim: bool, closure: F) -> GenTensor<T>
    where F: Fn(usize, usize, usize, usize, usize) -> T {
        if self.dim.len() <= dim {
            panic!("Tensor has dimension {:?}, mean() get dim of {}", self.dim, dim);
        }
        
        let mut ret_dim;
        if keepdim {
            ret_dim = self.dim.to_vec();
            ret_dim[dim] = 1;
        } else {
            ret_dim = Vec::new();
            for (i, index) in self.dim.iter().zip(0..self.dim.len()) {
                if index != dim {
                    ret_dim.push(*i);
                }
            }
        }
        
        let mut cap = 1;
        for i in &ret_dim {
            cap *= i;
        }

        let mut outer_size = 1;
        let mut inner_size = 1;
        for i in 0..self.dim.len() {
            if i < dim {
                outer_size *= self.dim[i];
            }
            if i > dim {
                inner_size *= self.dim[i];
            }
        }
        
        let mut data = Vec::with_capacity(cap);
        let over = self.dim[dim];
        let stride = self.stride();
        let step = stride[dim];

        for k in 0..outer_size {
            for j in 0..inner_size {
                let val = closure(over, k, j, inner_size, step);
                data.push(val);
            }
        }
        
        GenTensor {
            d: data,
            dim: ret_dim,
        }
    }

    pub fn get_diag(&self) -> GenTensor<T> {
        // TODO: handle batched matrix.
        let n = *self.size().last().unwrap();
        let mut ret = GenTensor::<T>::zeros(&[n]);
        for i in 0..n {
            ret.set(&[i], self.get(&[i, i]))
        }
        ret
    }
    pub fn set_diag(&mut self, o: &GenTensor<T>)  {
        // TODO: handle batched matrix.
        let n = *self.size().last().unwrap();
        for i in 0..n {
            self.set(&[i,i], o.get(&[i]));
        }
    }

    // TODO: handle batched matrix.
    pub fn get_column(&self, i: usize) -> GenTensor<T> {
        let nr = self.size()[self.size().len()-2];
        let mut ret = GenTensor::zeros(&[nr, 1]);
        for r in 0..nr {
            ret.set(&[r, 0], self.get(&[r, i]));
        }
        ret
    }
    // TODO: handle batched matrix.
    pub fn set_column(&mut self, o: &GenTensor<T>, i: usize) {
        let nr = self.size()[self.size().len()-2];
        for r in 0..nr {
            self.set(&[r, i], o.get(&[r, 0]));
        }
    }
    // TODO: handle batched matrix.
    pub fn get_row(&self, i: usize) -> GenTensor<T> {
        let nc = self.size()[self.size().len()-1];
        let mut ret = GenTensor::zeros(&[nc]);
        for c in 0..nc {
            ret.set(&[c], self.get(&[i, c]));
        }
        ret
    }
    // TODO: handle batched matrix.
    pub fn set_row(&mut self, o: &GenTensor<T>, i: usize) {
        let nc = self.size()[self.size().len()-1];
        for c in 0..nc {
            self.set(&[i, c], o.get(&[c]));
        }
    }

    // reduction ops
    // Implementation see the reduction.rs file.

    // Pointwise Ops
    pub fn _pointwise<F>(&self, closure: F) -> GenTensor<T>
    where F: Fn(&T) -> T {
        let mut ret = GenTensor {
            d: Vec::with_capacity(self.d.len()),
            dim: self.dim.clone(),
        };

        for i in &self.d {
            ret.d.push(closure(i));
        }
        ret
    }
    
    pub fn _right_broadcast<F>(&self, o: &GenTensor<T>, closure: F) -> GenTensor<T>
    where F: Fn(&T, &T) -> T {
        let mut ret = GenTensor {
            d: Vec::with_capacity(self.d.len()),
            dim: self.dim.clone(),
        };
        // with same shape.
        if self.d.len() == o.d.len() {
            for (v1, v2) in self.d.iter().zip(o.d.iter()) {
                ret.d.push(closure(v1, v2));
            }
        // right single scale
        } else if o.dim.len() == 1 && o.dim[0] == 1{
            for i in 0..self.d.len() {
                ret.d.push(closure(&self.d[i], &o.d[0]));
            }
        } else {
            if self.d.len() < o.d.len() {
                panic!("right-hand broadcast only.");
            }
            if self.dim.len() <= o.dim.len() {
                panic!("unmatched dimension. {}, {}", self.dim.len(), o.dim.len());
            }
            for i in 0..o.dim.len() {
                if o.dim[o.dim.len()-i-1] != self.dim[self.dim.len()-i-1] {
                    panic!("unmatched size.");
                }
            }

            // do repeat add
            let mut index = 0;
            for i in 0..self.d.len() {
                ret.d.push(closure(&self.d[i], &o.d[index]));
                index += 1;
                if index >= o.d.len() {
                    index = 0;
                }
            }
        }
        ret
    }

    pub fn log10_like(&self) -> GenTensor<T> {
	let new_data = vec![T::from(std::f64::consts::LN_10).unwrap(); self.d.len()];
        let new_dim = self.dim.to_vec();
        GenTensor {
            d: new_data,
            dim: new_dim,
        }
    }
    pub fn log2_like(&self) -> GenTensor<T> {
	let new_data = vec![T::from(std::f64::consts::LN_2).unwrap(); self.d.len()];
        let new_dim = self.dim.to_vec();
        GenTensor {
            d: new_data,
            dim: new_dim,
        }
    }
    
    /// element-wise add with right-hand broadcast.
    ///
    /// ```
    /// # use tensor_rs::tensor_impl::gen_tensor::*;
    /// let m1 = GenTensor::<f64>::new_raw(&vec![1.,2.,3.,4.,], &vec![2,2]);
    /// let m2 = GenTensor::<f64>::new_raw(&vec![1.,2.,3.,4.,], &vec![2,2]);
    /// let m3 = m1.add(&m2);
    /// assert_eq!(m3.get(&vec![0,0]), 2.);
    /// assert_eq!(m3.get(&vec![1,1]), 8.);
    /// ```
    pub fn add(&self, o: &GenTensor<T>) -> GenTensor<T> {
        self._right_broadcast(o, |x, y| *x + *y)
    }
    pub fn sub(&self, o: &GenTensor<T>) -> GenTensor<T> {
        self._right_broadcast(o, |x, y| *x - *y)
    }
    pub fn mul(&self, o: &GenTensor<T>) -> GenTensor<T> {
        self._right_broadcast(o, |x, y| *x * *y)
    }
    pub fn div(&self, o: &GenTensor<T>) -> GenTensor<T> {
        self._right_broadcast(o, |x, y| *x / *y)
    }

    /// matrix multiplication
    ///
    /// ```
    /// # use tensor_rs::tensor_impl::gen_tensor::*;
    /// let m1 = GenTensor::<f64>::new_raw(&vec![1.,2.,3.,4.,5.,6.], &vec![3,2]);
    /// let m2 = GenTensor::<f64>::new_raw(&vec![2.,3.,4.,5.,6.,7.], &vec![2,3]);
    /// let mut result = m1.mm(&m2);
    /// assert!(result == GenTensor::<f64>::new_raw(&vec![12.,15.,18.,26.,33.,40.,40.,51.,62.,], &vec![3,3]), "");
    /// ```
    pub fn mm(&self, o: &GenTensor<T>) -> GenTensor<T>{
        if self.dim.len() != 2 || o.dim.len() != 2 {
            panic!("Not a matrix input.");
        }
        let ls = self.dim[0];
        let rs = o.dim[1];
        let mut ret = GenTensor {
            d: Vec::with_capacity(ls*rs),
            dim: vec![ls, rs],
        };
        let lstride = self.stride();
        let rstride = o.stride();
        for i in 0..ls {
            for j in 0..rs {
                let mut tsum = T::zero();
                for k in 0..self.dim[1] {
                    tsum = tsum
                        + self.d[i*lstride[0] + k] * o.d[k*rstride[0] + j];
                }
                ret.d.push(tsum);
            }
        }
        ret
    }

    pub fn dot(&self, b: &GenTensor<T>) -> T {
        let mut sum = T::zero();
        for (l, m) in self.d.iter().zip(b.d.iter()) {
            sum = (*l)*(*m) + sum;
        }
        sum
    }

    /// Project a vector onto another one.
    // TODO, support batched vector.
    pub fn proj(&self, b: &GenTensor<T>) -> GenTensor<T> {
        let mut sum = T::zero();
        for (l, m) in self.d.iter().zip(b.d.iter()) {
            sum = (*l)*(*m) + sum;
        }
        let mut ret = b.clone();
        for i in ret.d.iter_mut() {
            *i = (*i) * sum;
        }
        ret
    }

    /// matrix multiplication of two tensor
    /// This is also for dot/inner product.
    pub fn matmul(&self, o: &GenTensor<T>) -> GenTensor<T> {
        if self.size()[self.size().len()-1] != o.size()[0] {
            panic!("matmul expect matched size {:?}, {:?}", self.dim, o.dim);
        }
        if self.size().len() == 1 && o.size().len() == 1 {
            panic!("Two vector have not matched size for matmul! {:?}, {:?}", self.numel(), o.numel());
        }
        let inner = o.dim[0];
        let mut cap = 1;
        let mut odim = Vec::new();
        let mut lloop = 1;
        let mut rloop = 1;
        for i in 0..self.dim.len()-1 {
            cap *= self.dim[i];
            odim.push(self.dim[i]);
            lloop *= self.dim[i];
        }
        for i in 1..o.dim.len() {
            cap *= o.dim[i];
            odim.push(o.dim[i]);
            rloop *= o.dim[i];
        }

        let mut ret = GenTensor {
            d: Vec::with_capacity(cap),
            dim: odim,
        };
        
        let lstride = self.stride();
        let rstride = o.stride();
        for i in 0..lloop {
            for j in 0..rloop {
                let mut tsum = T::zero();
                for k in 0..inner {
                    tsum = tsum
                        + self.d[i*lstride[0] + k] * o.d[k*rstride[0] + j];
                }
                ret.d.push(tsum);
            }
        }
        ret
    }

    /// outer product of right-most dimensions.
    pub fn outer(&self, o: &GenTensor<T>, avg: Option<bool>) -> GenTensor<T> {
        let mut dim = Vec::new();
        let mut data;
        let mut cap = 1;
        let mut outer_size = 1;
        let left_dim;
        let right_dim;
        if self.dim.len() == o.dim.len()
            && self.dim[0..self.dim.len()-1] == o.dim[0..self.dim.len()-1] {
                left_dim = self.dim[self.dim.len()-1];
                right_dim = o.dim[self.dim.len()-1];
                for i in 0..self.dim.len()-1 {
                    dim.push(self.dim[i]);
                    cap *= self.dim[i];
                    outer_size *= self.dim[i];
                }
                dim.push(left_dim);
                cap *= left_dim;
                dim.push(right_dim);
                cap *= right_dim;
                if avg.is_some() && avg.unwrap() {
                    data = vec![T::zero(); left_dim*right_dim];
                    dim = vec![left_dim, right_dim];
                } else {
                    data = Vec::with_capacity(cap);
                }
            } else {
                panic!("bad size for outer: {:?}, {:?}", self.dim, o.dim);
            }

        
        if avg.is_some() && avg.unwrap() {
            for k in 0..outer_size {
                let mut new_data = Vec::with_capacity(left_dim*right_dim);
                for i in 0..left_dim {
                    for j in 0..right_dim {
                        new_data.push(self.d[i + k*left_dim] * o.d[j + k*right_dim]);
                    }
                }
                for i in 0..new_data.len() {
                    data[i] = data[i] + new_data[i];
                }
            }
            for i in &mut data {
                *i = *i / T::from(outer_size).expect("");
            }
            GenTensor {
                d: data,
                dim,
            }
        } else {
            for k in 0..outer_size {
                for i in 0..left_dim {
                    for j in 0..right_dim {
                        data.push(self.d[i + k*left_dim] * o.d[j + k*right_dim]);
                    }
                }
            }
            GenTensor {
                d: data,
                dim,
            }
        }
    }

    #[cfg(not(feature = "use-blas-lapack"))]
    pub fn squared_error(t1: &Self, t2: &Self) -> GenTensor<T> {
        let mut ret = GenTensor {
            d: Vec::with_capacity(t1.d.len()),
            dim: t1.dim.to_vec(),
        };
        for (v1, v2) in t1.d.iter().zip(t2.d.iter()) {
            ret.d.push((*v1 - *v2)*(*v1 - *v2));
        }
        ret
    }


    // Comparison Ops


    pub fn all_close(&self, o: &GenTensor<T>) -> GenTensor<T> {
        self.eq_t(o)
    }

    pub fn arg_sort(&self, dim: usize, descending: bool) -> GenTensor<T> {
        let mut d = self.d.to_vec();

        let mut outer_size = 1;
        let mut inner_size = 1;

        for (i, index) in self.dim.iter().zip(0..self.dim.len()) {
            if index < dim {
                outer_size *= i;
            } else if index > dim {
                inner_size *= i;
            }
        }

        let stride = self.stride()[dim];
        let size = self.dim[dim];

        for i in 0..outer_size {
            for j in 0..inner_size {
                let mut collected = Vec::<(T, usize)>::with_capacity(size);
                for k in 0..size {
                    collected.push((self.d[k*stride + j + i*inner_size*size], k));
                }
                collected.sort_unstable_by(|a, b| {
                    let porder = a.0.partial_cmp(&b.0).unwrap();
                    if descending {
                        porder
                    } else {
                        porder.reverse()
                    }
                });
                let (_left, right): (Vec<_>, Vec<_>) = collected.iter().cloned().unzip();
                for k in 0..size {
                    d[k*stride + j + i*inner_size*size] = T::from(right[k]).expect("");
                }
            }
        }

        GenTensor {
            d,
            dim: self.dim.to_vec()
        }
    }

    /// Computes element-wise equality
    /// use eq_t instead, as eq is reserved for == overloading.
    ///
    /// ```
    /// # use tensor_rs::tensor_impl::gen_tensor::*;
    /// let m1 = GenTensor::<f64>::new_raw(&vec![1.,2.,3.,4.,5.,6.], &vec![3,2]);
    /// let m2 = GenTensor::<f64>::new_raw(&vec![1.,2.,3.,4.,5.,6.], &vec![3,2]);
    /// assert_eq!(m1.eq_t(&m2).get(&vec![0,0]), 1.);
    /// assert_eq!(m1.eq_t(&m2).get(&vec![2,1]), 1.);
    /// ```
    pub fn eq_t(&self, o: &GenTensor<T>) -> GenTensor<T> {
        let mut cmp = Vec::<T>::with_capacity(self.d.len());
        for (v1, v2) in self.d.iter().zip(o.d.iter()) {
            if (*v1-*v2).abs() < T::min_positive_value().sqrt() {
                cmp.push(T::one(),);
            } else {
                cmp.push(T::zero());
            }
        }
        GenTensor {
            d: cmp,
            dim: self.dim.to_vec(),
        }
    }

    /// true if two tensors have the same size and elements, false otherwise.
    ///
    /// ```
    /// # use tensor_rs::tensor_impl::gen_tensor::*;
    /// let m1 = GenTensor::<f64>::fill(1., &vec![3,5,2]);
    /// let m2 = GenTensor::<f64>::fill(1., &vec![3,5,2]);
    /// assert_eq!(m1.equal(&m2), true)
    /// ```
    pub fn equal(&self, o: &GenTensor<T>) -> bool {
        if self.dim.len() != o.dim.len() || self.dim != o.dim {
            return false;
        }

        if self.d.len() != o.d.len() {
            return false;
        }
        
        let mut same = true;
        for (v1, v2) in self.d.iter().zip(o.d.iter()) {
            if (*v1-*v2).abs() > T::min_positive_value().sqrt() {
                same = false;
                break;
            }
        }
        same
    }

    pub fn ge(&self, o: &GenTensor<T>) -> GenTensor<T> {
        if self.size() != o.size() {
            panic!("max needs two tensor have the same size, {:?}, {:?}", self.dim, o.dim);
        }
        let mut ret = GenTensor::zeros(&self.dim);

        for ((a, b), c) in self.d.iter().zip(o.d.iter()).zip(ret.d.iter_mut()) {
            if a >= b {
                *c = T::one();
            } else {
                *c = T::zero();
            }
        }
        ret
    }

    pub fn gt(&self, o: &GenTensor<T>) -> GenTensor<T> {
        if self.size() != o.size() {
            panic!("max needs two tensor have the same size, {:?}, {:?}", self.dim, o.dim);
        }
        let mut ret = GenTensor::zeros(&self.dim);

        for ((a, b), c) in self.d.iter().zip(o.d.iter()).zip(ret.d.iter_mut()) {
            if a > b {
                *c = T::one();
            } else {
                *c = T::zero();
            }
        }
        ret
    }

    //pub fn isfinite(&self, o: &GenTensor<T>) -> GenTensor<T> {
    //    GenTensor::new()
    //}
    //
    //pub fn isinf(&self, o: &GenTensor<T>) -> GenTensor<T> {
    //    GenTensor::new()
    //}
    //
    //pub fn isnan(&self, o: &GenTensor<T>) -> GenTensor<T> {
    //    GenTensor::new()
    //}
    //
    //pub fn kthvalue(&self, o: &GenTensor<T>) -> GenTensor<T> {
    //    GenTensor::new()
    //}
    // le
    pub fn le(&self, o: &GenTensor<T>) -> GenTensor<T> {
        if self.size() != o.size() {
            panic!("max needs two tensor have the same size, {:?}, {:?}", self.dim, o.dim);
        }
        let mut ret = GenTensor::zeros(&self.dim);

        for ((a, b), c) in self.d.iter().zip(o.d.iter()).zip(ret.d.iter_mut()) {
            if a <= b {
                *c = T::one();
            } else {
                *c = T::zero();
            }
        }
        ret
    }
    // lt
    pub fn lt(&self, o: &GenTensor<T>) -> GenTensor<T> {
        if self.size() != o.size() {
            panic!("max needs two tensor have the same size, {:?}, {:?}", self.dim, o.dim);
        }
        let mut ret = GenTensor::zeros(&self.dim);

        for ((a, b), c) in self.d.iter().zip(o.d.iter()).zip(ret.d.iter_mut()) {
            if a < b {
                *c = T::one();
            } else {
                *c = T::zero();
            }
        }
        ret
    }
    // max, 
    //pub fn max(&self, o: Option<&GenTensor<T>>, dim: Option<usize>, keep_dim: Option<bool>) -> GenTensor<T> {
    //    if o.is_none() && dim.is_none() && keep_dim.is_none() {
    //        max_all()
    //    } else if o.is_some() && dim.is_none() && keep_dim.is_none() {
    //        max_pair()
    //    }
    //}

    // ne
    pub fn ne(&self, o: &GenTensor<T>) -> GenTensor<T> {
        if self.size() != o.size() {
            panic!("max needs two tensor have the same size, {:?}, {:?}", self.dim, o.dim);
        }
        

        let data = self.d.iter().zip(
            o.d.iter())
            .map(|(x, y)|
                 if *x != *y {
                     T::one()
                 } else {
                     T::zero()
                 }
        ).collect();
        GenTensor {
            d: data,
            dim: self.dim.to_vec(),
        }
    }
    // sort
    // topk
    
}

impl<T> Default for GenTensor<T> where T: num_traits::Float {
    fn default() -> GenTensor<T> {
        GenTensor { d: Vec::<T>::new(), dim: Vec::new() }
    }
}


/// ```
/// # use tensor_rs::tensor_impl::gen_tensor::*;
/// let m1 = GenTensor::<f64>::fill(1., &vec![3,5,2]);
/// let m2 = GenTensor::<f64>::fill(1., &vec![3,5,2]);
/// assert_eq!(m1==m2, true)
/// ```
impl<T> PartialEq for GenTensor<T> where T: num_traits::Float {
    fn eq(&self, other: &Self) -> bool {
        self.equal(other)
    }
}
impl<T> Eq for GenTensor<T> where T: num_traits::Float {}

impl fmt::Display for GenTensor<f32> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        if self.dim.len() == 2 {
            write!(f, "[")?;
            for i in 0..self.dim[0] {
                write!(f, "[")?;
                for j in 0..self.dim[1] {
                    write!(f, "{}, ", self.get(&[i, j]))?;
                }
                writeln!(f, "]")?;
            }
            writeln!(f, "]")
        } else {
            writeln!(f, "{:?}", self.dim)?;
            write!(f, "{:?}", self.d)            
        }
    }
}
impl fmt::Display for GenTensor<f64> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(f, "{:?}", self.dim)?;
        writeln!(f, "{:?}", self.d)
    }
}

impl fmt::Debug for GenTensor<f32> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        if self.numel() > 30 {
            writeln!(f, "size: {:?}", self.dim)?;
            write!(f, "data: {:?}", &self.d[..30])
        } else if self.dim.len() == 2 {
            write!(f, "[")?;
            for i in 0..self.dim[0] {
                write!(f, "[")?;
                for j in 0..self.dim[1] {
                    write!(f, "{}, ", self.get(&[i, j]))?;
                }
                writeln!(f, "]")?;
            }
            writeln!(f, "]")
        } else {
            writeln!(f, "size: {:?}", self.dim)?;
            write!(f, "data: {:?}", self.d)            
        }
    }
}
impl fmt::Debug for GenTensor<f64> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        if self.numel() > 30 {
            writeln!(f, "size: {:?}", self.dim)?;
            write!(f, "data: {:?}", &self.d[..30])
        } else if self.dim.len() == 2 {
            write!(f, "[")?;
            for i in 0..self.dim[0] {
                write!(f, "[")?;
                for j in 0..self.dim[1] {
                    write!(f, "{}, ", self.get(&[i, j]))?;
                }
                writeln!(f, "]")?;
            }
            writeln!(f, "]")
        } else {
            writeln!(f, "{:?}", self.dim)?;
            write!(f, "{:?}", self.d)            
        }
    }
}

impl<T> Clone for GenTensor<T> where T: num_traits::Float {
    fn clone(&self) -> Self {
        GenTensor {
            d: self.d.to_vec(),
            dim: self.dim.to_vec(),
        }
    }
}

#[cfg(feature = "use-blas-lapack")]
use crate::tensor_impl::lapack_tensor::blas_api::BlasAPI;


#[cfg(feature = "use-blas-lapack")]
impl GenTensor<f32> {
    
    pub fn squared_error(t1: &Self, t2: &Self) -> GenTensor<f32> {
        let mut v2 = t2.d.to_vec();
        BlasAPI::<f32>::axpy(t1.d.len(), -1., &t1.d, 1, &mut v2, 1);
        
        let mut ret = GenTensor {
            d: v2,
            dim: t1.dim.to_vec(),
        };
        for i in &mut ret.d {
            *i = (*i)*(*i);
        }
        ret
    }
}
#[cfg(feature = "use-blas-lapack")]
impl GenTensor<f64> {
    
    pub fn squared_error(t1: &Self, t2: &Self) -> GenTensor<f64> {
        let mut v2 = t2.d.to_vec();
        BlasAPI::<f64>::axpy(t1.d.len(), -1., &t1.d, 1, &mut v2, 1);
        
        let mut ret = GenTensor {
            d: v2,
            dim: t1.dim.to_vec(),
        };
        for i in &mut ret.d {
            *i = (*i)*(*i);
        }
        ret
    }
}


#[cfg(test)]
mod tests {
    use super::*;
    

    #[test]
    fn test_index2dimpos() {
        let a = GenTensor::<f32>::zeros(&vec![10, 5, 3, 4]);

        let b = a.index2dimpos(10);
        assert_eq!(b, vec![0, 0, 2, 2]);

    }

    #[test]
    fn test_gentensor() {
        {
            let mut m2 = GenTensor::<f64>::new_raw(&vec![1., 2., 3., 4.,], &vec![2, 2]);
            *m2.get_mut(&vec![0,0]) = 5.;
            assert_eq!(m2.get_raw(), vec![5., 2., 3., 4.,])
        }
    }

    #[test]
    fn test_gen_tensor_get() {
        {
            let m1 = GenTensor::<f64>::fill(1., &vec![10, 3, 28, 30]);
            assert_eq!(m1.get_n().get_raw(), vec![10.]);
            assert_eq!(m1.get_c().get_raw(), vec![3.]);
            assert_eq!(m1.get_h().get_raw(), vec![28.]);
            assert_eq!(m1.get_w().get_raw(), vec![30.]);

            let result = std::panic::catch_unwind(
                ||
                    m1.get_d().get_raw()
            );
            assert!(result.is_err());
        }
    }

    #[test]
    fn outer() {
        let a = GenTensor::<f32>::fill(1., &vec![10, 2]);
        let b = GenTensor::<f32>::fill(1., &vec![10, 3]);
        let c = a.outer(&b, None);
        assert_eq!(*c.size(), vec![10, 2, 3]);
        //println!("{}", c);
        let d = b.outer(&a, None);
        assert_eq!(*d.size(), vec![10, 3, 2]);

        let e = a.outer(&b, Some(true));
        assert_eq!(e, GenTensor::ones(&[2, 3]));
    }

    #[test]
    fn proj() {
        let a = GenTensor::<f32>::fill(1., &vec![2]);
        let b = GenTensor::<f32>::new_raw(&[3., -2.], &[2, 1]);
        assert_eq!(a.proj(&b), b);
    }

    #[test]
    fn get_patch() {
        let a = GenTensor::new_raw(&GenTensor::<f32>::arange(30).get_data(), &[2, 3, 5]);
        let b = a.get_patch(&vec![(0, 2), (0, 2), (2, 3)][..], Option::None);
        assert_eq!(b, GenTensor::<f32>::new_raw(&vec![2.0, 7.0, 17.0, 22.0][..], &vec![2, 2, 1][..]));
    }

    #[test]
    fn set_patch() {
        let mut a = GenTensor::new_raw(&GenTensor::<f32>::arange(30).get_data(), &[2, 3, 5]);
        let b = GenTensor::<f32>::ones(&[1, 3, 5]);
        a.set_patch(&b, &[(1,2), (0,3), (0,5)], None);
        println!("{:?}", a);
        assert_eq!(a, GenTensor::new_raw(&[0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0], &[2, 3, 5]));
    }

    // Pointwise Ops
    use crate::tensor_trait::elemwise::ElemwiseTensorOp;
    #[test]
    fn ceil() {
        let a = GenTensor::<f32>::new_raw(&vec![0.9213,  1.0887, -0.8858, -1.7683],
                                              &vec![4]);
        
        let ret = a.ceil();

        let expect = GenTensor::<f32>::new_raw(&vec![1., 2., 0., -1.], 
                                               &vec![4]);
        assert_eq!(ret, expect);
    }
    
    #[test]
    fn log1pexp() {
        let a = GenTensor::<f32>::new_raw(&vec![0.9213,  1.0887, -0.8858, -1.7683],
                                              &vec![4]);
        
        let ret = a.log1pexp();

        let expect = GenTensor::<f32>::new_raw(&vec![1.2563436, 1.3788694, 0.34527916, 0.15753591], 
                                               &vec![4]);
        assert_eq!(ret, expect);
    }
    
    #[test]
    fn sigmoid() {
        let a = GenTensor::<f32>::new_raw(&vec![0.9213,  1.0887, -0.8858, -1.7683],
                                              &vec![4]);
        
        let ret = a.sigmoid();

        let expect = GenTensor::<f32>::new_raw(&vec![0.71530694, 0.7481369, 0.29197732, 0.14575386], 
                                               &vec![4]);
        assert_eq!(ret, expect);
    }

    #[test]
    fn sign() {
        let a = GenTensor::<f32>::new_raw(&vec![0.9213,  0.0, -0.0, -1.7683],
                                              &vec![4]);
        
        let ret = a.sign();

        let expect = GenTensor::<f32>::new_raw(&vec![1.0, 0.0, 0.0, -1.0],
                                               &vec![4]);
        assert_eq!(ret, expect);
    }
    

    // Comparison Ops
    #[test]
    fn arg_sort() {
        let a = GenTensor::<f32>::new_raw(&vec![0.0785,  1.5267, -0.8521,  0.4065,
                                                    0.1598,  0.0788, -0.0745, -1.2700,
                                                    1.2208,  1.0722, -0.7064,  1.2564,
                                                    0.0669, -0.2318, -0.8229, -0.9280,],
                                              &vec![4, 4]);
        
        let index = a.arg_sort(1, true);

        let expect = GenTensor::<f32>::new_raw(&vec![2., 0., 3., 1., 
                                                     3., 2., 1., 0., 
                                                     2., 1., 0., 3., 
                                                     3., 2., 1., 0.], 
                                               &vec![4, 4]);
        assert_eq!(index, expect);
    }

    #[test]
    fn ne() {
        let a = GenTensor::<f32>::new_raw(&vec![1., 3., 10., 11.], &vec![2,2]);
        let b = GenTensor::<f32>::new_raw(&vec![2., 3., 10., 6.], &vec![2,2]);
        let c = a.ne(&b);
        assert_eq!(c, GenTensor::<f32>::new_raw(&vec![1., 0., 0., 1.], &vec![2,2]));
    }
}