topic I improve sparsity,the first in Intel® oneAPI Math Kernel Library

mkl_sparse_s_spmmd is slower than tensorflow tf.sparse_matmul

zhou__jianqian — Tue, 17 Sep 2019 07:08:02 GMT

I have two sparse matrix,both of them sparsity is about 55%, I use mkl_sparse_s_spmmd function and pack the funtion .so dynamic library.in python program import .so.I compare .so and th.sparse_matmul use the same dataset.but I found the tf.sparse_matmul performance is good than the .so.

I compile MKL tensorflow,and test in this tensorflow.

55% of sparsity - this is

Gennady_F_Intel — Fri, 20 Sep 2019 06:41:15 GMT

55% of sparsity - this is dense but not sparse matrixes. what is the problem size? Probably, it makes sense to try the "classical" sgemm if the RAM size will allow doing that.

I improve sparsity,the first

zhou__jianqian — Fri, 20 Sep 2019 06:54:49 GMT

I improve sparsity,the first matrix is(1507, 256),which sparsity is 70% more,and The other one is(256,31140),which sparsity is 70% more,here is my code. and I test in Intel(R) Xeon(R) Silver.

Meanwhile,when use mkl sparse

zhou__jianqian — Fri, 20 Sep 2019 09:04:57 GMT

Meanwhile,when use mkl sparse product,the CPU Utilization ratio is lower than tf.matmul() and tf.sparse_matmul()

nevertheless, why don't you

Gennady_F_Intel — Fri, 20 Sep 2019 13:38:30 GMT

nevertheless, why don't you want to call sgemm? Your typical problem sizes are not too big. You may convert from csr to dense representation and make the [s,d]gemm call. I do believe the performance and scalability will be pretty fine.

Thank you for your answer，at

zhou__jianqian — Mon, 23 Sep 2019 07:08:56 GMT

Thank you for your answer，at frist I think my matrix is sparse，so I replace dense matrix product with sparse matrix product，meanwhile,I enlarge my matrix size and sparsity,the size is13000*256,256*31140,the sparsity is 80%,I originally think this will improve my performance，but the running result shows my idea is wrong.