unrollcuda 0.0.5

Loop unrolling and batching for CUDA

unrollcuda

Loop unrolling and batching for CUDA
The core idea of this solution is to give a way to solve the following tasks:

1. Use Loop unrolling to compute in CUDA any size and any count of dimensions array
2. Use Batching to compute any size array, even if it s big that can't be fitted in GPU memory

Requirements:

CUDA
Python

Getting Started

Installation

pip install unrollcuda

Usage

Invert values in a multi-dimensional boolean array

invert.cu

__global__ void unroll(
    bool *arr,
    unsigned int *shape,
    unsigned long long gpu_arr_size,
    unsigned long long shape_total,
    unsigned long long dimensions_count,
    unsigned long long step,
    unsigned char order,
    unsigned long long batch_start
)
{
    unsigned long long idx = threadIdx.x + blockIdx.x * blockDim.x;
    unsigned long long idx_full;
    unsigned int i = 0;
    unsigned int *indices = new unsigned int[dimensions_count]; // array to hold the computed indices
    unsigned long long tmp;
    
    idx_full = i * step + idx;

    while (idx_full < shape_total && idx_full < gpu_arr_size)
    {
        tmp = idx_full + batch_start; // add batch_start to account for the offset
        // Compute the indices
        for (unsigned int j = 0; j < dimensions_count; ++j)
        {
            unsigned int dimension = (order == 0) ? dimensions_count - j - 1 : j;
            // Modulo by the dimension size
            indices[dimension] = tmp % shape[dimension];
            // Divide by the dimension size
            tmp /= shape[dimension];
        }
        //printf("idx_full: %llu, idx: %llu, batch_start: %llu\n", idx_full, idx, batch_start);
        
        for (unsigned int j = 0; j < dimensions_count; ++j)
        {
            // j is the dimension
            
            // Your code ++
            // Invert the value in arr
            arr[idx_full] = !arr[idx_full];
            // Your code --
            
            break;
        }
        i += 1;
        idx_full = i * step + idx;
    }
    // Free the memory
    delete[] indices;
}

invert.py

import numpy as np
from unrollcuda import unrollcuda


def main():

    dimensions = [3, 4]
    shape = [int(size) for size in dimensions]
    # random boolean values
    arr = np.random.choice(
        a=[False, True],
        size=shape,
        p=[0.5, 0.5],
        )

    with open('invert.cu', 'r') as f:
        kernel_code = f.read()

    uc = unrollcuda(kernel_code)
    arr_new = uc.inference(arr)

    # Prepare the test array
    arr_test = arr.copy()
    print('\nOriginal array:\n', arr_test)
    # Convert all False values to True and vice versa
    arr_test = np.logical_not(arr_test)
    print('\nTest array:\n', arr_test)

    # Check the result
    result_check = np.array_equal(arr_new, arr_test)
    print('\nResult check: ', result_check)


if __name__ == '__main__':
    main()