DoS performs optimal SPMM by estimating the SPMM latency when the right matrix is transformed to dense format or remains in sparse format based on matrix multiplication information.
You can use this system in the following order.
Step 1. Write matrix multiplication information
(row of left matrix, column of left matrix, column of right matrix, density of left matrix, density of right matrix).
Step 2. Click Submit.
Step 3. Check Sparse X Sparse Latency and Sparse X Dense Latency according to matrix multiplication information.
Click either “Test Data A” or “Test Data B” to fill in the input matrix dimensions to query.
Two sample test data represent cases where different SPMM implementations show performance variation.
| Matrix info | Input |
|---|---|
| NR L The number of rows for a left matrix |
|
| NC L The number of columns for a left matrix |
|
| NC R The number of columns for a right matrix |
|
| D L The density of the left matrix |
|
| D R The density of the right matrix |
|
| SPMM Method | Latency (ms) |
|---|---|
| Sparse X Sparse | |
| Sparse X Dense |
| Optimal Method | X times improvement |
|---|---|
| Sparse X Dense |
Latency Prediction Result
according to SPMM Method
SPMM Method
Latency Prediction Result according to SPMM Method
SPMM Method