FrancescAlted

import numpy as np
import iarray as ia
from iarray import udf
import math


# Params for array construction
shape = (40_000, 40_000)
ia.set_config_defaults(dtype=np.float32, fp_mantissa_bits=15)

FrancescAlted

# Benchmark comparing npy, npz, jdb and blosc2 storage formats

import sys

import numpy as np
import jdata as jd
import blosc2
from time import time

N = 10_000

FrancescAlted

# Examples on getting orthogonal slices

import iarray as ia
import numpy as np


dtype = np.float32

ia.set_config_defaults(dtype=dtype)

FrancescAlted

# Calling scalar UDFs from expressions.
# This is for 1-dim arrays.

from time import time

import numpy as np

import iarray as ia
from iarray import udf

FrancescAlted

FrancescAlted

FrancescAlted

c = ia.matmul(am, bm)

FrancescAlted

from iarray.udf import jit, Array, float64, int64

# Create second array via an UDF
@jit()
def tri(out: Array(float64, 2), x: Array(float64, 2), k: int64) -> int64:
    n = out.window_shape[0]
    m = out.window_shape[1]
    row_start = out.window_start[0]
    col_start = out.window_start[1]
    for i in range(n):

FrancescAlted

# Random values with lossy compression to get some decent compression ratio
am = ia.random.normal(amshape, 3, 2, chunks=amchunks, blocks=amblocks, urlpath=filename, fp_mantissa_bits=10)

FrancescAlted

nrows = 100_000 # number of rows in matrix am
ncols = 25000   # number of columns in first matrix
ncols2 = 1000   # number of columns in second matrix

shape = (nrows, ncols, ncols2)
amshape = (shape[0], shape[1])
bmshape = (shape[1], shape[2])

# Obtain optimal chunk and block shapes
mparams = ia.matmul_params(amshape, bmshape, dtype=np.float64)