TEE-ACM2

Transformations for Energy Efficient Accelerated Chain Matrix Multiplication

Compile :

$> module load cuda cmake
$> mkdir BUILD && cd BUILD && cmake ..
$> make

Execute:

./mat_test -n -k -run
n : number of matrices to compute
k : kernel to launch
```
  0 : TEE-ACM2
```
```
  1 : cuBLAS using the OP_Count Tree
```
```
  2 : cuBLAS in order
```
run : parameter to generate the sizes

CUDA kernels:

cupy.cuh

Fast implementation (5% close to cuBLAS) for single matrix multiplication
Manually overlap communication and computation (copy data from global memory to registers for the next iteration while computing the current one)

tee_acm.cuh

Energy efficient implementation for single matrix multiplication (up to 10% energy savings)
No overlap

fused_left_mat_mult.cuh

Fused kernel implementation (80% slower than running two cuBLAS kernels)
Each thread in a CUDA thread block compute an entire row of the final R tile
Atomic adds to the global memory

no_atomic_adds_left_mat_mult.cuh

Fused kernel implementation (500% slower than running two cuBLAS kernels)
Each thread in a CUDA thread block compute an entire row of the final R matrix
No atomic operations
Occupancy problems

Copyright

© 2022. Triad National Security, LLC. All rights reserved. This program was produced under U.S. Government contract 89233218CNA000001 for Los Alamos National Laboratory (LANL), which is operated by Triad National Security, LLC for the U.S. Department of Energy/National Nuclear Security Administration. All rights in the program are reserved by Triad National Security, LLC, and the U.S. Department of Energy/National Nuclear Security Administration. The Government is granted for itself and others acting on its behalf a nonexclusive, paid-up, irrevocable worldwide license in this material to reproduce, prepare derivative works, distribute copies to the public, perform publicly and display publicly, and to permit others to do so.

This program is open source under BSD-3 License. Details can be found in LICENSE.

lanl/TEE-ACM2