Embedded Files
Project Overview
Project Overview
Language Used: C
Language Used: C
The objective of this project was to create a C-to-Python interface - a student implementation of numpy. Our version, numc, deals specifically with matrix add, subtraction, absolute value, negation, matrix multiply, and matrix power.
The objective of this project was to create a C-to-Python interface - a student implementation of numpy. Our version, numc, deals specifically with matrix add, subtraction, absolute value, negation, matrix multiply, and matrix power.
Usage
Usage
Each operation is used by overloading the corresponding operators (to add two matrices, use mat1 + mat2, etc.) Some example usage taken from the assignment specifications along with set and get methods and indexing are shown below.
Each operation is used by overloading the corresponding operators (to add two matrices, use mat1 + mat2, etc.) Some example usage taken from the assignment specifications along with set and get methods and indexing are shown below.
Performance Benchmarks
Performance Benchmarks
The minimum specifications for our implementation are as follows:
The minimum specifications for our implementation are as follows:
Simple (Add, Subtract, Absolute Value, Negation): At least 5x faster than the naïve implementation.
Matrix Multiplication: At least 95x times faster than the naïve implementation,
Matrix Power: At least 1700x times faster than the naïve implementation.
Strategies:
Strategies:
In order to meet the following performance benchmarks, we used optimization techniques such as:
In order to meet the following performance benchmarks, we used optimization techniques such as:
Removing abstraction barriers to avoid setting up stack frames.
Manual loop unrolling.
Vectorization with SIMD instructions (assuming the CPU has access to Intel AVX instructions).
Cache blocking and amortized timing for matrix transpose.
Using cache block-friendly indices for matrix multiplication.
Storing data in a 1D array instead of a 2D array for matrices.
Using OpenMP to parallelize loops.
Page updated
Google Sites
Report abuse