pascalpolygon Posts - Answer Overflow

pascalpolygon

•Created by pascalpolygon on 4/25/2024 in #questions

Efficient Mojo Tensor To Numpy Array Conversion

Hello, I am attempting to convert a Mojo Tensor to a Numpy array without looping or at least with sub-polynomial time. My approach is to get the pointer to the Tensor and use Python's ctypes.from_address() But it doesn't work - the output np_array is not the same as the tensor...maybe Tensor has some header bytes? Or the pointer does not point to a contiguous memory block? I would like to understand the issue with this code and get some inputs on how to achieve fast conversions for potentially large Mojo Tensors. Here is the code:

def main():
    var ctypes = Python.import_module("ctypes")
    var np = Python.import_module("numpy")

    var mojo_tensor =  Tensor[DType.int32](
        TensorShape(4, 4),
        1, 0, 0, 0,
        0, 1, 0, 0,
        0, 0, 1, 0,
        0, 0, 0, 1,
    )

    var mojo_tensor_ptr = mojo_tensor.data().__int__()
  
    var num_elements = mojo_tensor.num_elements().__int__()
    var tensor_type = mojo_tensor.type

    
    #Load the hopefully contiguous tensor into a ctypes array
    var int32_array = (ctypes.c_int32 * num_elements).from_address(mojo_tensor_ptr.__int__())
    var np_array = np.ctypeslib.as_array(int32_array).astype(np.int32)

    print("Numpy array: ", np_array)

def main():
    var ctypes = Python.import_module("ctypes")
    var np = Python.import_module("numpy")

    var mojo_tensor =  Tensor[DType.int32](
        TensorShape(4, 4),
        1, 0, 0, 0,
        0, 1, 0, 0,
        0, 0, 1, 0,
        0, 0, 0, 1,
    )

    var mojo_tensor_ptr = mojo_tensor.data().__int__()
  
    var num_elements = mojo_tensor.num_elements().__int__()
    var tensor_type = mojo_tensor.type

    
    #Load the hopefully contiguous tensor into a ctypes array
    var int32_array = (ctypes.c_int32 * num_elements).from_address(mojo_tensor_ptr.__int__())
    var np_array = np.ctypeslib.as_array(int32_array).astype(np.int32)

    print("Numpy array: ", np_array)

Thank you.

2 replies

MModular

•Created by pascalpolygon on 4/3/2024 in #questions

Is there a simple way to implement Numpy-style ndarray slicing on Mojo Tensors ?

I have a rank 4 tensor where the first dimension is an index and the last 3 represent a Voxel Grid.

  # Initialize the observation space as a 4D tensor
        self.observation_space = Tensor[DType.int32](num_builds, build_volume[0], build_volume[1], build_volume[2])

  # Initialize the observation space as a 4D tensor
        self.observation_space = Tensor[DType.int32](num_builds, build_volume[0], build_volume[1], build_volume[2])

I want to get the voxel grid given a build_id. I wrote a function to do this but I am wondering if I there's a Mojo way to do this.

2 replies

MModular

•Created by pascalpolygon on 3/13/2024 in #questions

Numpy: Import from source dir error

I am trying to import Numpy in my mojo project using

from python import Python

fn main() raises:
    var np = Python.import_module("numpy")

from python import Python

fn main() raises:
    var np = Python.import_module("numpy")

But I get the following Error:

Unhandled exception caught during execution: Error importing numpy: you should not try to import numpy from
        its source directory; please exit the numpy source tree, and relaunch
        your python interpreter from there.
mojo: error: execution exited with a non-zero result: 1

Unhandled exception caught during execution: Error importing numpy: you should not try to import numpy from
        its source directory; please exit the numpy source tree, and relaunch
        your python interpreter from there.
mojo: error: execution exited with a non-zero result: 1

5 replies

MModular

•Created by pascalpolygon on 2/29/2024 in #questions

Invalid Parameter to Vectorize from example

Hello! I just got started with Mojo, been playing with in for the past 2 days. I run into an a syntax error from just copying the matmul example.

@mojo
from algorithm.functional import vectorize

fn matmul_tiled(inout C: Matrix, A: Matrix, B: Matrix):
    @parameter
    fn calc_row(m: Int):
        @parameter
        fn calc_tile[tile_x: Int, tile_y: Int](x: Int, y: Int):
            for k in range(y, y + tile_y):

                @parameter
                fn dot[nelts: Int](n: Int):
                    C.store(
                        m,
                        n + x,
                        C.load[nelts](m, n + x)
                        + A[m, k] * B.load[nelts](k, n + x),
                    )

                vectorize[dot, nelts, size=tile_x]()

        tile[calc_tile, tile_n, tile_k](C.cols, B.rows)

    parallelize[calc_row](C.rows, C.rows)

@mojo
from algorithm.functional import vectorize

fn matmul_tiled(inout C: Matrix, A: Matrix, B: Matrix):
    @parameter
    fn calc_row(m: Int):
        @parameter
        fn calc_tile[tile_x: Int, tile_y: Int](x: Int, y: Int):
            for k in range(y, y + tile_y):

                @parameter
                fn dot[nelts: Int](n: Int):
                    C.store(
                        m,
                        n + x,
                        C.load[nelts](m, n + x)
                        + A[m, k] * B.load[nelts](k, n + x),
                    )

                vectorize[dot, nelts, size=tile_x]()

        tile[calc_tile, tile_n, tile_k](C.cols, B.rows)

    parallelize[calc_row](C.rows, C.rows)

On this line vectorize[dot, nelts, size=tile_x]() I get the error : Cannot pass 'fn[Int](n = Int) capturing -> None' value, parameter expected 'Int'mojo

4 replies

Gaming

Programming