PyPIM: Integrating Digital Processing-in-Memory from Microarchitectural Design to Python Tensors

Overview

This repository includes the framework proposed in the following paper,

Orian Leitersdorf, Ronny Ronen, Shahar Kvatinsky, “PyPIM: Integrating Digital Processing-in-Memory from Microarchitectural Design to Python Tensors,” Accepted to IEEE/ACM MICRO 2024.

The framework enables high-level programming of PIM applications with significant ease. The framework benefits from the high flexibility of tensor-based Python (e.g., NumPy, PyTorch, TensorFlow) to provide the user with simple operations that can already be executed today (through the simulator backend):

>>> import pypim as pim
>>> x = pim.Tensor(8, dtype=pim.float32)
>>> x
Tensor(shape=(8,), dtype=float32): [0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0]
>>> 
>>> x[2] = 2.5
>>> x[3] = 1.25
>>> x[4] = 2.25
>>> x
Tensor(shape=(8,), dtype=float32): [0.0,0.0,2.5,1.25,2.25,0.0,0.0,0.0]
>>> 
>>> x[::2]
TensorView(shape=(4,), dtype=float32, slicing=slice(0, 7, 2)): [0.0,2.5,2.25,0.0]
>>> x[::2].sum()
4.75
>>> x[::2].sort()
TensorView(shape=(4,), dtype=float32, slicing=slice(0, 7, 2)): [0.0,0.0,2.25,2.5]

This also enables the user to assemble new PIM routines from existing arithmetic functions, such as:

import pypim as pim
def myFunc(a: pim.Tensor, b: pim.Tensor):    
    # Parallel multiplication and addition 
    return a * (1 + b)

The repository is split into four parts: (1) the underlying GPU-accelerated simulator, (2) the microarchitectural driver, (3) the development library, and (4) a series of test scripts that serve as the benchmarks of PyPIM.

User Information

Dependencies

The simulation environment is implemented via CUDA to enable fast execution of many samples in parallel. Therefore, the project requires the following dependencies:

1. CUDA 12.0 (with a capable GPU of at least 8GB DRAM)
2. Python installation (tested with 3.10)
3. Compiler for C++ 17 (or higher)

Installation

The development library may be installed using pip from the project directory as follows:

pip install -e .

The installation can then be verified by running python main.py, with the expected result being 32.0.

Organization

The repository is organized into the following directories:

• csrc: this directory contains the C++ source code for the simulator and the driver.
• pypim: this directory contains the Python source code for the development library.
• tests: this directory contains the source code for the tests.
• results: this directory contains the raw results of the tests.
• main.py: this file contains the example script from the paper.

Instruction-Set-Architecture (ISA)

The full instruction-set-architecture (ISA) supported in CUDA-PIM is as follows:

Operation	Int Support	Float Support
Arithmetic
Addition	✓	✓
Subtraction	✓	✓
Multiplication	✓	✓
Division	✓	✓
Modulo	✓
Negation	✓	✓
Comparison
Less than (or equal to)	✓	✓
Greater than (or equal to)	✓	✓
Equal	✓	✓
Bitwise
Bitwise Not	✓	✓
Bitwise And	✓	✓
Bitwise Or	✓	✓
Bitwise Xor	✓	✓
Miscellaneous
Sign	✓	✓
Zero	✓	✓
Abs	✓	✓
Mux	✓	✓