Reduce normal register loads #263
Conversation
|
@larq/compute-engine any idea why the CI isn't running? |
Oh, I think we've configured CI to only run on push, since this is a PR from a fork, it doesn't work. |
AdamHillier
force-pushed
the
reduce-reads
branch
from
4866629
to
1ec2183
Compare
|
The |
|
Now there's different CI failures that are unrelated to the code. The Git checkout is failing for some bizarre reason. |
Looks like this might be a problem with the git submodule checkout, though I don't see any failure on master. |
I just rebased and it seems the checkout fine now. After actions/checkout#173 is merged we should upgrade to the new checkout actions which hopefully will be less flaky. |
* Move two CortexM tests from GTest to TFLite testing GTest is replaced with the TFLiteMicro testing framework, making it possible to run these tests also on the CortexM. The two tests for which this is done don't used advanced features (parameter sweep) and therefore work equally well without GTest. * Add two existing tests for the CorexM target Added two existing tests for BGEMV and the quantized multiplier to the build_lcem script to run on CortexM. The tests are slightly modified since std::cerr and the C++ random number generator are not available. * Fix Bazel linter issues * Fix compilation error with std::round for RISC-V * Fix issues for RISC-V target * Improve QEMU check in build script * Fix error reporting and micro compiler flags * Fix issue with tflite error reporting Co-authored-by: Cedric Nugteren <cedric@plumerai.com>
What do these changes do?
This change swaps some of the (normal, non-SIMD) registers used for intermediate results in the ARM64 binary conv kernels, in such a way that reloading several parameters from memory is no longer necessary.
Two additional registers are used by the kernel (which were previously unused):
x14andx15.Three
ldrmemory read instructions have been eliminated.How Has This Been Tested?
The ARM64 tests pass locally.
Benchmark Results
Benchmark results (μs) on a Raspberry Pi 4 (1 thread) on QuickNet Large: 78828.5 -> 78580.3.
Results are the average of 500 runs but the results are so close that there probably isn't a meaningful difference here.