LiteRT コンパイラ プラグイン

// Name associated with the manufacturer this plugin relates to.
LITERT_CAPI_EXPORT const char* LiteRtGetCompilerPluginSocManufacturer();

// Create and initialize the plugin instance.
LITERT_CAPI_EXPORT LiteRtStatus
LiteRtCreateCompilerPlugin(LiteRtCompilerPlugin* compiler_plugin,
                           LiteRtEnvironmentOptions env, LiteRtOptions options);

// Choose ops for compilation.
// This is the PARTITION step.
LITERT_CAPI_EXPORT LiteRtStatus LiteRtCompilerPluginPartition(
    LiteRtCompilerPlugin compiler_plugin, const char* soc_model,
    LiteRtSubgraph subgraph, LiteRtOpList selected_ops);

// Prepare result to pass to the runtime for given model containing partitioned
// subgraphs. This is the COMPILE step.
LITERT_CAPI_EXPORT LiteRtStatus LiteRtCompilerPluginCompile(
    LiteRtCompilerPlugin compiler_plugin, const char* soc_model,
    LiteRtModel partitions, LiteRtCompiledResult* compiled_result);

LITERT_CAPI_EXPORT LiteRtStatus LiteRtCompilerPluginPartition(
    LiteRtCompilerPlugin compiler_plugin, const char* soc_model,
    LiteRtSubgraph subgraph, LiteRtOpList selected_ops);

LITERT_CAPI_EXPORT LiteRtStatus LiteRtCompilerPluginCompile(
    LiteRtCompilerPlugin compiler_plugin, const char* soc_model,
    LiteRtModel partitions, LiteRtCompiledResult* compiled_result);

// Define the plugin's internal state structure
struct LiteRtCompilerPluginT {};

// Identify the manufacturer
const char* LiteRtGetCompilerPluginSocManufacturer() {
  return "AcmeCorp"; // Example manufacturer name
}

// Specify the supported hardware (in this example, it supports kLiteRtHwAcceleratorNpu)
LiteRtStatus LiteRtGetCompilerPluginSupportedHardware(
    LiteRtCompilerPlugin compiler_plugin,
    LiteRtHwAccelerators* supported_hardware) {
  // ... argument checking ...
  *supported_hardware = kLiteRtHwAcceleratorNpu;
  return kLiteRtStatusOk;
}

LiteRtStatus LiteRtCompilerPluginPartition(LiteRtCompilerPlugin compiler_plugin,
                                          const char* soc_model,
                                          LiteRtSubgraph subgraph,
                                          LiteRtOpList selected_ops) {

  // Iterate over ops and check criteria for selection
  // (using a C++ wrapper namespace '::litert' for convenience).
  // `subgraph` is a single subgraph from the original model, as such
  // this function will be called for each subgraph in the original model.

  ::litert::Subgraph main_subgraph(subgraph);
  for (const auto& op : main_subgraph.Ops()) {
    // 1. Check a constraint: require all tensors to be Float32
    bool only_f32 = true;
    // ... logic to check input/output types ...
    if (!only_f32) {
      continue;
    }

    // 2. Check op codes and push to selected_ops list
    if (op.Code() == kLiteRtOpCodeTflMul) {
      LITERT_RETURN_IF_ERROR(LiteRtPushOp(selected_ops, op.Get(), 0));
    } else if (op.Code() == kLiteRtOpCodeTflSub) {
      LITERT_RETURN_IF_ERROR(LiteRtPushOp(selected_ops, op.Get(), 0));
    } else if (op.Code() == kLiteRtOpCodeShloComposite) {
      // Example of checking composite op options
      // ... logic to check for "odml.rms_norm" name ...
      LITERT_RETURN_IF_ERROR(LiteRtPushOp(selected_ops, op.Get(), 0));
    }
  }
  return kLiteRtStatusOk;
}

// Internal structure defining the compiled output
struct LiteRtCompiledResultT {
  std::vector<std::string> byte_code;   // The hardware bytecode buffers
  std::vector<std::string> per_op_data; // Per-call metadata (CallInfo)
};

LiteRtStatus LiteRtCompilerPluginCompile(
    LiteRtCompilerPlugin compiler_plugin, const char* soc_model,
    LiteRtModel partitions, LiteRtCompiledResult* compiled_result) {

  // 1. Create the internal result structure
  auto model = litert::Model::CreateFromNonOwnedHandle(partitions);
  const auto num_partitions = model.NumSubgraphs();
  auto result = std::make_unique<LiteRtCompiledResultT>();
  result->byte_code.resize(num_partitions);
  result->per_op_data.resize(num_partitions);

  // 2. Iterate and compile each partition
  for (auto i = 0; i < num_partitions; ++i) {
    // CompileSinglePartition is an internal helper that converts the subgraph
    // into the target hardware's format and stores it in result->byte_code.
    // In the case of the example this is just a stringification of the graph.

    // ... internal call to CompileSinglePartition ...
    // Example: result.byte_code[i] = generated_hw_code;
    // Example: result.per_op_data[i] = absl::StrFormat("Partition_%d", i);

    // The "per_op_data" is a unique identifier associated to the `ith` partition.
    // This is analogous to the name of a function in a library.
    // This is only meaningful when the plugin is preparing single modules with multiple entry points.
  }

  // 3. Pass ownership of the result back to the framework
  *compiled_result = result.release();

  return kLiteRtStatusOk;
}

// Functions to expose the compiled result data to the framework
LiteRtStatus LiteRtGetCompiledResultByteCode(
    LiteRtCompiledResult compiled_result, LiteRtParamIndex byte_code_idx,
    const void** byte_code, size_t* byte_code_size) {
  // ... implementation reads from compiled_result->byte_code ...
}
// ... other LiteRtGetCompiledResult* functions ...