Examples

Steps to run a ge-spmm example

Requirement: GPU Compute-Capability >= SM 70, nvcc >= 11.0

You could build the whole project once by the following code

cd ..
make exp

Or you could follow these steps to run spmm only.

Step 1: build ge-spmm library

cd ../src/ge-spmm
make
cd ../../example

Step 2: build the example

cd ge-spmm
make
# will generate spmm.out

Step 3: run example

./spmm.out ../data/p2p-Gnutella31.mtx


./spmm.out ../data/p2p-Gnutella31.mtx 32 # set arbitrary #columns in rhs dense matrix

Example output (on V100, cuda v11.1)

Finish reading matrix 62586 rows, 62586 columns, 147892 nnz.
Ignore original values and use randomly generated values.
[Cusparse] Report: spmm A(62586 x 62586) * B(62586 x 32) sparsity 0.000038 (nnz=147892)
 Time 0.076032 (ms), Throughput 124.487694 (gflops).
[GE-SpMM][Alg: 0] Report: spmm A(62586 x 62586) * B(62586 x 32) sparsity 0.000038 (nnz=147892)
 Time 0.045675 (ms), Throughput 207.228882 (gflops).
[GE-SpMM][Alg: 1] Report: spmm A(62586 x 62586) * B(62586 x 32) sparsity 0.000038 (nnz=147892)
 Time 0.231848 (ms), Throughput 40.824486 (gflops).
[GE-SpMM][Alg: 2] Report: spmm A(62586 x 62586) * B(62586 x 32) sparsity 0.000038 (nnz=147892)
 Time 0.076493 (ms), Throughput 123.738281 (gflops).
[GE-SpMM][Alg: 3] Report: spmm A(62586 x 62586) * B(62586 x 32) sparsity 0.000038 (nnz=147892)
 Time 0.857259 (ms), Throughput 11.041107 (gflops).
[GE-SpMM][Alg: 8] Report: spmm A(62586 x 62586) * B(62586 x 32) sparsity 0.000038 (nnz=147892)
 Time 0.226008 (ms), Throughput 41.879375 (gflops).
[GE-SpMM][Alg: 9] Report: spmm A(62586 x 62586) * B(62586 x 32) sparsity 0.000038 (nnz=147892)
 Time 0.044950 (ms), Throughput 210.568878 (gflops).

Steps to run a gspmm-fp example

Make sure you have installed the requirements as follows:

• torch >= 1.8.0
• ninja
• scipy

Step Enter the example's folder:

cd gspmm-fp 

Then you could run our example by

python gspmm-exp.py [k]

Here k is the feature length of the input dense matrix.

Example output (on RTX3090, cuda v11.2)

Loading extension module spmm...
running u_sub_e_sum our time is: 0.0004

Note here we use JIT, so it is normal to wait longer when compiling the project the first time.