Forward Socket¶
Introduction¶
The forward socket is a new feature added to StreamPU to improve the
performance and the flexibility in some applications. As mentioned in the
Socket section, the SFWD works as an input and output at the same
time. It receives its dataptr from the input bound socket and this same
pointer is sent to all the output bound sockets, which means that all the
consecutive tasks bound by SFWD share the same memory space.
graph LR;
A(FWD)-->B(FWD); A(FWD)-.->K{MEM};
B(FWD)-->C(FWD); B(FWD)-.->K{MEM};
C(FWD)-->F(FWD); C(FWD)-.->K{MEM};
F(FWD)-.->K{MEM};Technical Improvements¶
The implementation of the forward socket for sequences was mainly straightforward because it behaves the same way as the input and output sockets. We just had to distinguish when it is used as an input and when it is used as an output. However, the most challenging part was to combine forward socket with the pipeline. Especially when forward sockets are bound from one stage to an other.
Forward Sockets and Pipelines¶
As explained in the Adaptor section, a pool of buffers
is used between each stage of the pipeline. The adaptor gets a buffer from
this pool and uses it to update the output socket of its pull task
(dataptr attribute). This output socket is then bound to the input socket of
the next tasks. In other words, all the input sockets connected to the pull
output socket need to be updated with the new dataptr address.
The forward sockets are all pointing to the same dataptr, so getting a new
buffer means that we have to update the dataptr of all the consecutive bound
forward sockets to this new memory space. In the sequence, the
same update needs to be done in the reversed way when the dataptr is exchanged
at the end of the stage. For that, we added two recursive methods as explained
in the sequence section (see
explore_thread_rec() and
explore_thread_rec_reverse()
methods).
Moreover, when multiple forward sockets are crossing pipeline
stages, we need to check if these forward sockets are pointing to a same
dataptr or to different dataptrs. In the case where multiple forward sockets
are crossing pipeline stages and are pointing to the same
dataptr, only one buffer need to be created in the
adaptors. To detect this, a map to record the previous
dataptr is used (see the fwd_source variable in the code).
Tests¶
Some specific tests have been added to the project to validate the robustness of the forward socket implementation.
Specific for Forward Socket¶
test-pipeline-double-chain.
SIO and
SFWD, both on the same stage.
test-complex-pipeline-full-fwd.
SFWD bound to two SFWD in two
different stages, and how the buffer exchange behaves with connections
between distant stages \(S1\) and \(S4\).
test-complex-pipeline-mix-fwd.
SFWD bound to three SFWD in three
different stages (\(S1 \rightarrow S2\), \(S1 \rightarrow S4\) and \(S1
\rightarrow S5\)), and how the buffer exchange behaves with connections
between distant stages. Additionally, a traditional relay task (\(t4\) with
an input and an output socket) has been added in stage \(S2\). The \(t7\)
compare task ensures that all the 3 SFWD have the same contents.
It is expected that the final values in \(t8\) are \(init + 2\) and the final
values in \(t9\) are \(init + 1\). In this test, task to task binding is used
to ensure that \(t4\) is executed before \(t6\) and \(t8\) is executed before
\(t9\) (see the oriented dashed lines).
