This partially reverts commit 1e9b4c5d84. The make helpers for running the inference code generator were somewhat of a hack due to the desire to integrate a target-specific preprocessor binary as part of the process. Since that is no longer necessary, we can now switch over to just using Bazel for running the code generator.
BUG=b/294230402
The TFLM code_generator takes a TFLite model and generates C/C++ source code that can then be compiled into a binary. This change adds a Bazel macro for invoking the code generator and creating a cc_library with the resulting sources. It also updates the checked-in hello_world example with an appropriate BUILD file and updates the script to use Bazel instead of make.
BUG=b/294230402
The codegen process is a multi-step process that requires compiling,
executing code under simulation, and executing python scripts. To
simplify this workflow, this commit adds Make helper functions for
generating inference source code from a model and creating a binary with
it.
It also updates the hello world example to use these helpers and adds an
update script for keeping the checked in generated source in sync.
BUG=cleanup
The TFLM kernels use the MicroContext and MicroGraph interfaces to fetch
eval tensors and access other parts of the graph. Since codegen does not
have the MicroInterpreter and related objects to serve those, this PR
introduces a new MicroCodegenContext class that serves as both the
MicroContext and MicroGraph.
The MicroCodegenContext is configured with a span of Subgraphs, each of
which includes the inputs, outputs, nodes, tensors and an invocation
function. The code generator will create the data and functions needed
for each Subgraph and initialize the MicroCodegenContext with it. By
having the re-usable MicroCodegenContext, the code generator won't have
to generate nearly as much code.
BUG=b/295174086
The RISC-V toolchain we're using failed to properly discern the TfLiteEvalTensor type for use with an assignment operator. It produced compiler errors for "no match for 'operator=' with operand types TfLiteEvalTensor and a brace-closed initializer list. This PR resolves this issue by explicitly using the TfLiteEvalTensor constructor for the initializer list. We also apply this approach to the TfLiteNode initialization as well, just for consistency.
BUG=#2195
This PR generates all of the TfLiteEvalTensors for the graph. It also generates all of the static buffers used for tensors that were present in the flatbuffer.
BUG=b/295077140
Each operator has its own set of BuiltinOptions, which will translate into an op-specific TfLite Params structure. This change turns Operator into a base class with common functionality, and adds a FullyConnected subclass that knows how to generate the TfLiteFullyConnectedParams struct.
BUG=b/295175961
This PR adds the generation of empty TfLiteContext and TfLiteNode structures, and generates appropriate invocation methods for each subgraph. It also generates the input, output, and intermediate arrays for each TfLiteNode.
BUG=b/295175961
As the next step in the codegen experiment, we want to generate the invoke calls for each layer. This is slightly challenging with the existing sources, as kernels only expose a registration function, not their individual Eval functions. In an effort to keep the code churn to a minimum, this PR introduces an inference only registration structure and function. It includes just two function pointers: invoke and reset. For this CL, we've only introduced it for FullyConnected.
In the code generator, this PR creates a new op_table array in the generated source, with an enum for lookup. It also generates an invoke function for each subgraph, that calls each operator's invoke function.
BUG=295174388
This PR adds a codegen inference example for the hello world model to demonstrate how to invoke the code generator and build the generated source. For now, we're just checking the generated source into the repo to skip over building out the make rules and also ensuring the generated source complies with formatting rules.
This also fixes a minor formatting issue in the source templates, as clang-format now properly complained about it.
BUG=b/295390000
RJ Ascani