|
|
|
In *Overview of Algorithm and Data structures* we presented the different code paths in the function.
|
|
|
|
It is important to consider carefully the implications of this when vectorizing.
|
|
|
|
Vectorization is based on SIMD processing (single instruction, multiple data), but different code paths require different instructions.
|
|
|
|
With the RISC-V vector extension, this can be overcame with the help of masked instructions, which allows restricting writing the result of a vector instructions to only certain elements using a bitmask.
|
|
|
|
|
|
|
|
For instance, which proportion of the atom pair interactions (or inner loop iterations) belong to the *do nothing* group?
|
|
|
|
Even when using masked instructions, we can avoid updating data for *do nothing* interactions, but the execution time required for processing this data cannot be avoided.
|
|
|
|
So, as opposed to the serial version, a *do nothing* interaction has the same cost in time as any other atom in the vectorized version with masked instructions.
|
|
|
|
|
|
|
|
Before starting working on the vectorization, the code was modified to count the number of interactions that belong to each category.
|
|
|
|
The flowchart shows the average number number of interactions (for a single `i` atom in a timestep) that belong to each category, and the arrows show the same information in percentage form.
|
|
|
|
Black values show data for the default protein input, while red values correspond to the modified input described in section *Loop size*.
|
|
|
|
|
|
|
|
We can see how the proportion of "do nothing" elements in the regular input is about 42%.
|
|
|
|
We deemed to extract the not "do-nothing" elements would be too costly, since the proportion is too high, and the accelerator lacks the `vcompress` [instruction](https://github.com/riscv/riscv-v-spec/blob/0.7.1/v-spec.adoc#176-vector-compress-instruction) that implements this (see [ISA support](https://repo.hca.bsc.es/gitlab/EPI/RTL/Vector_Accelerator/-/wikis/VPU/ISA-support)).
|
|
|
|
For this reason, we decided to use the masking approach, even if it makes "do nothing" elements as slow as the rest.
|
|
|
|
|
|
|
|
This type of "masking" approach is not suitable for the elements labeled as "slow" (the ones involving `sqrt` and `exp`), since all elements would need a computation time of "slow" and "fast" combined.
|
|
|
|
The fact that there are so few "slow" elements (around 0.3%) makes it feasible to try to use the "vextract" method.
|
|
|
|
Since the instruction is unavailable, we used a loop of `vmfirst` in order to mask the "slow" elements in the vector register and update them separately using the serial function `compute_iterj_special`.
|
|
|
|
|
|
|
|
The modified input manages to reduce the proportion of interactions that belong to the *do nothing* and *slow* categories.
|
|
|
|
It may be interesting to test how the modified input affects performance in both serial and vectorized versions. |
