vector-multiplication-cuda - Comparing various launch configs for CUDA based vector multiply

Part of the report Parallelizing PageRank for a Volta GPU.

--

Comparing various launch configs for CUDA based vector multiply.

Two floating-point vectors x and y, with number of elements from 1E+6 to 1E+9 were multiplied using CUDA. Each element count was attempted with various CUDA launch configs, running each config 5 times to get a good time measure. Multiplication here represents any memory-aligned independent operation, or a map() operation. Results indicate that a grid_limit of 16384/32768, and a block_size of 128/256 to be suitable for both float and double. Using a grid_limit of MAX and a block_size of 256 could be a decent choice.

All outputs are saved in out and a small part of the output is listed here. Nsight Compute profile results are saved in prof. Some charts are also included below, generated from sheets. This experiment was done with guidance from Prof. Dip Sankar Banerjee and Prof. Kishore Kothapalli.

```bash $ nvcc -std=c++17 -Xcompiler -O3 main.cu $ ./a.out

...

# Elements 1e+07

[00002.138 ms] [1.644725] multiplySeq

[00000.243 ms] [1.644725] multiplyCuda<<<1024, 32>>>

[00000.179 ms] [1.644725] multiplyCuda<<<1024, 64>>>

[00000.170 ms] [1.644725] multiplyCuda<<<1024, 128>>>

[00000.169 ms] [1.644725] multiplyCuda<<<1024, 256>>>

[00000.172 ms] [1.644725] multiplyCuda<<<1024, 512>>>

[00000.168 ms] [1.644725] multiplyCuda<<<1024, 1024>>>

[00000.181 ms] [1.644725] multiplyCuda<<<2048, 32>>>

[00000.171 ms] [1.644725] multiplyCuda<<<2048, 64>>>

[00000.170 ms] [1.644725] multiplyCuda<<<2048, 128>>>

[00000.170 ms] [1.644725] multiplyCuda<<<2048, 256>>>

[00000.168 ms] [1.644725] multiplyCuda<<<2048, 512>>>

[00000.162 ms] [1.644725] multiplyCuda<<<2048, 1024>>>

[00000.186 ms] [1.644725] multiplyCuda<<<4096, 32>>>

[00000.169 ms] [1.644725] multiplyCuda<<<4096, 64>>>

[00000.172 ms] [1.644725] multiplyCuda<<<4096, 128>>>

[00000.167 ms] [1.644725] multiplyCuda<<<4096, 256>>>

[00000.168 ms] [1.644725] multiplyCuda<<<4096, 512>>>

[00000.167 ms] [1.644725] multiplyCuda<<<4096, 1024>>>

[00000.189 ms] [1.644725] multiplyCuda<<<8192, 32>>>

[00000.171 ms] [1.644725] multiplyCuda<<<8192, 64>>>

[00000.169 ms] [1.644725] multiplyCuda<<<8192, 128>>>

[00000.166 ms] [1.644725] multiplyCuda<<<8192, 256>>>

[00000.167 ms] [1.644725] multiplyCuda<<<8192, 512>>>

[00000.168 ms] [1.644725] multiplyCuda<<<8192, 1024>>>

[00000.181 ms] [1.644725] multiplyCuda<<<16384, 32>>>

[00000.168 ms] [1.644725] multiplyCuda<<<16384, 64>>>

[00000.167 ms] [1.644725] multiplyCuda<<<16384, 128>>>

[00000.166 ms] [1.644725] multiplyCuda<<<16384, 256>>>

[00000.167 ms] [1.644725] multiplyCuda<<<16384, 512>>>

[00000.168 ms] [1.644725] multiplyCuda<<<16384, 1024>>>

[00000.184 ms] [1.644725] multiplyCuda<<<32768, 32>>>

[00000.168 ms] [1.644725] multiplyCuda<<<32768, 64>>>

[00000.167 ms] [1.644725] multiplyCuda<<<32768, 128>>>

[00000.166 ms] [1.644725] multiplyCuda<<<32768, 256>>>

[00000.166 ms] [1.644725] multiplyCuda<<<32768, 512>>>

[00000.169 ms] [1.644725] multiplyCuda<<<32768, 1024>>>

...

```

References

• CUDA by Example :: Jason Sanders, Edward Kandrot

- C++
Published by wolfram77 over 4 years ago