import jax
import jax.numpy as jnp
import kernex as kex
import matplotlib.pyplot as plt

# see https://nbviewer.org/github/barbagroup/CFDPython/blob/master/lessons/01_Step_1.ipynb

tmax,xmax = 0.5,2.0
nt,nx = 151,51
dt,dx = tmax/(nt-1) , xmax/(nx-1)
u = jnp.ones([nt,nx])
c = 0.5

# kscan moves sequentially in row-major order and updates in-place using lax.scan.

F = kernex.kscan(
        kernel_size = (3,3),
        padding = ((1,1),(1,1)),
        named_axis={0:'n',1:'i'},  # n for time axis , i for spatial axis (optional naming)
        relative=True
        )


# boundary condtion as a function
def bc(u):
    return 1

# initial condtion as a function
def ic1(u):
    return 1

def ic2(u):
    return 2

def linear_convection(u):
    return ( u['i','n-1'] - (c*dt/dx) * (u['i','n-1'] - u['i-1','n-1']) )


F[:,0]  = F[:,-1] = bc # assign 1 for left and right boundary for all t

# square wave initial condition
F[:,:int((nx-1)/4)+1] = F[:,int((nx-1)/2):] = ic1
F[0:1, int((nx-1)/4)+1 : int((nx-1)/2)] = ic2

# assign linear convection function for
# interior spatial location [1:-1]
# and start from t>0  [1:]
F[1:,1:-1] = linear_convection

kx_solution = F(jnp.array(u))

plt.figure(figsize=(20,7))
for line in kx_solution[::20]:
    plt.plot(jnp.linspace(0,xmax,nx),line)

import jax
import jax.numpy as jnp
import kernex as kex

def gaussian_blur(image, sigma, kernel_size):
    x = jnp.linspace(-(kernel_size - 1) / 2.0, (kernel_size- 1) / 2.0, kernel_size)
    w = jnp.exp(-0.5 * jnp.square(x) * jax.lax.rsqrt(sigma))
    w = jnp.outer(w, w)
    w = w / w.sum()

    @kex.kmap(kernel_size=(kernel_size, kernel_size), padding="same")
    def conv(x):
        return jnp.sum(x * w)

    return conv(image)

import jax
import jax.numpy as jnp
import kernex as kex

@jax.jit
@jax.vmap
@kex.kmap(
kernel_size= (3,3),
padding = ('same','same'))
def kernex_depthwise_conv2d(x,w): # Channel-first depthwise convolution # jax.debug.print("x=\n{a}\nw=\n{b} \n\n",a=x, b=w)
return jnp.sum(x\*w)

h,w,c = 5,5,2
k=3

x = jnp.arange(1,h*w*c+1).reshape(c,h,w)
w = jnp.arange(1,k*k*c+1).reshape(c,k,k)
print(kernex_depthwise_conv2d(x,w))</summary>

@jax.vmap # vectorize over the channel dimension
@kex.kmap(kernel_size=(3,3), strides=(2,2))
def maxpool_2d(x):
    # define the kernel for the Maxpool operation over the spatial dimensions
    return jnp.max(x)

@jax.vmap # vectorize over the channel dimension
@kex.kmap(kernel_size=(3,3), strides=(2,2))
def avgpool_2d(x):
    # define the kernel for the Average pool operation over the spatial dimensions
    return jnp.mean(x)

# lets solve dydt = y, where y0 = 1 and y(t)=e^t
# using Runge-Kutta 4th order method
# f(t,y) = y
import jax.numpy as jnp
import matplotlib.pyplot as plt
import kernex as kex


t = jnp.linspace(0, 1, 5)
y = jnp.zeros(5)
x = jnp.stack([y, t], axis=0)
dt = t[1] - t[0]  # 0.1
f = lambda tn, yn: yn


def ic(x):
    """ initial condition y0 = 1 """
    return 1.


def rk4(x):
    """ runge kutta 4th order integration step """
    # ┌────┬────┬────┐      ┌──────┬──────┬──────┐
    # │ y0 │*y1*│ y2 │      │[0,-1]│[0, 0]│[0, 1]│
    # ├────┼────┼────┤ ==>  ├──────┼──────┼──────┤
    # │ t0 │ t1 │ t2 │      │[1,-1]│[1, 0]│[1, 1]│
    # └────┴────┴────┘      └──────┴──────┴──────┘
    t0 = x[1, -1]
    y0 = x[0, -1]
    k1 = dt * f(t0, y0)
    k2 = dt * f(t0 + dt / 2, y0 + 1 / 2 * k1)
    k3 = dt * f(t0 + dt / 2, y0 + 1 / 2 * k2)
    k4 = dt * f(t0 + dt, y0 + k3)
    yn_1 = y0 + 1 / 6 * (k1 + 2 * k2 + 2 * k3 + k4)
    return yn_1


F = kex.kscan(kernel_size=(2, 3), relative=True, padding=((0, 1)))  # kernel size = 3

F[0:1, 1:] = rk4
F[0, 0] = ic
# compile the solver
solver = jax.jit(F.__call__)
y = solver(x)[0, :]

plt.plot(t, y, '-o', label='rk4')
plt.plot(t, jnp.exp(t), '-o', label='analytical')
plt.legend()