Speedup are mostly fake data?

[jiajun-c]

. The best speedups are both in the matrix ‘bibd_20_10’,

Test bibd_20_10 in A100, performance are really bad.... DASP performance are slower than cusparse in many matrix.

[ZeneleLu]

Hi,

Thanks for your feedback. Please note that the code currently available on GitHub is not exactly the same as the one used in the paper. After the paper was published, some researchers pointed out that certain long-row cases produced incorrect results. I have since revised and updated the implementation. As a result, for matrices dominated by long rows, the current version may show performance differences compared to what was reported in the paper.

I reran tests on the same code and suspect that your evaluation might not include sufficient warm-up iterations.
When I perform 100 warm-up runs followed by 1000 timing runs, DASP consistently outperforms cuSPARSE.

If I instead do only 10 warm-ups and 10 timing runs, both DASP and cuSPARSE show significantly lower performance.

In addition, I conducted large-scale testing of the latest GitHub version on an A100 GPU, using both CUDA 12.0 (the version used in the paper) and CUDA 12.4 (the version you used):

CUDA 12.0: DASP outperforms cuSPARSE on 2098 out of 2872 matrices, with a maximum speedup of 7.05×, observed on wiki-Talk.mtx.

CUDA 12.4: DASP outperforms cuSPARSE on 1830 out of 2872 matrices, with a maximum speedup of 6.99×, also on wiki-Talk.mtx.

If you need, I can also help rerun the test for your specific case.

[jiajun-c]

Since dasp only use 1/8 of mma result and tensor core flops is just twice of cuda core in fp64 in A100. How can it be faster than cuda core?

[jiajun-c]

wiki-Talk are mainly compute by cuda core and it can be faster if you only use cuda core, even faster than 11x. So what's the meaning use tensor core?

[ZeneleLu]

The performance gain of DASP does not come solely from the Tensor Core hardware itself. It also benefits from the data layout and scheduling strategies we designed to better utilize Tensor Cores.

To enable Tensor Core computation, we store the matrix in a block-based format, which improves data locality during computation. In addition, DASP classifies rows into three categories and applies different processing strategies to each, which helps mitigate load imbalance among warps.

These optimizations contribute to the performance improvement observed in DASP together.

[jiajun-c]

Cuda core can also compute the layout you use for tensor core, tensor core just make it slower in this job.

[jiajun-c]

The paper AmgT of your lab also have the same problem in spmv part. I think tensor core is not suitable for all job.

[ZeneleLu]

You are right, but you are not the first to find this.

DASP was never intended to be faster than all SpMV methods. It leveraged new hardware and introduced new data structures and algorithms, achieving performance improvements on some matrices, which is the true purpose of DASP.

Overall, you are welcome to improve it and make it even faster.

[jiajun-c]

The cusparseSPMV compute the , while DASP only compute the y = Ax + y
The cusparse use another kernel called vector_scalar_multiply_kernel to compte beta * Y while dasp ignore it.

This kernel take 10% ~ 20% time in cuSPARSE spmv, and it is also timed when dasp are compared with cusparse. It is fair for cusparse?