Skip to content

Commit

Permalink
Implement MicroContext/Graph for codegen (#2218)
Browse files Browse the repository at this point in the history
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
  • Loading branch information
rascani authored Sep 12, 2023
1 parent f800eb8 commit 3323a41
Show file tree
Hide file tree
Showing 9 changed files with 396 additions and 42 deletions.
8 changes: 6 additions & 2 deletions codegen/examples/hello_world/Makefile.inc
Original file line number Diff line number Diff line change
@@ -1,9 +1,13 @@
# TODO(rjascani): The codegen runtime files (ie, in runtime subdir) should be a
# separate library.
CODEGEN_HELLO_WORLD_SRCS := \
$(TENSORFLOW_ROOT)codegen/examples/hello_world/hello_world.cc \
$(TENSORFLOW_ROOT)codegen/examples/hello_world/hello_world_model.cc
$(TENSORFLOW_ROOT)codegen/examples/hello_world/hello_world_model.cc \
$(TENSORFLOW_ROOT)codegen/runtime/micro_codegen_context.cc

CODEGEN_HELLO_WORLD_HDRS := \
$(TENSORFLOW_ROOT)codegen/examples/hello_world/hello_world_model.h
$(TENSORFLOW_ROOT)codegen/examples/hello_world/hello_world_model.h \
$(TENSORFLOW_ROOT)codegen/runtime/micro_codegen_context.h

# Builds a standalone binary.
$(eval $(call microlite_test,codegen_hello_world,\
Expand Down
47 changes: 32 additions & 15 deletions codegen/examples/hello_world/hello_world_model.cc
Original file line number Diff line number Diff line change
Expand Up @@ -17,11 +17,14 @@ limitations under the License.

#include "hello_world_model.h"

#include "codegen/runtime/micro_codegen_context.h"
#include "tensorflow/lite/c/builtin_op_data.h"
#include "tensorflow/lite/c/c_api_types.h"
#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/kernels/internal/compatibility.h"
#include "tensorflow/lite/micro/kernels/micro_ops.h"
#include "tensorflow/lite/micro/micro_common.h"
#include "tensorflow/lite/micro/micro_context.h"

namespace hello_world_model {
namespace {
Expand Down Expand Up @@ -101,6 +104,10 @@ alignas(16) uint8_t buffer_7[16] = {

// buffer_10 is located in the arena

constexpr size_t kSubgraph0Inputs[1] = {0};

constexpr size_t kSubgraph0Outputs[1] = {9};

struct Node0_0 {
struct Inputs {
int size = 3;
Expand Down Expand Up @@ -202,13 +209,34 @@ struct Tensor0_9Dims {
int data[2] = {1, 1};
} tensor0_9_dims;

TfLiteStatus InvokeSubgraph0(TfLiteContext* context,
tflite::Span<TfLiteNode> nodes) {
TFLITE_DCHECK(nodes.size() == 3);
TF_LITE_ENSURE_OK(
context, op_table[OpCode::kFullyConnected].invoke(context, &nodes[0]));
TF_LITE_ENSURE_OK(
context, op_table[OpCode::kFullyConnected].invoke(context, &nodes[1]));
TF_LITE_ENSURE_OK(
context, op_table[OpCode::kFullyConnected].invoke(context, &nodes[2]));

return kTfLiteOk;
}

} // namespace

Model::Model() {
context_.impl_ = nullptr;
Model::Model()
: subgraphs_{
{.inputs = {&kSubgraph0Inputs[0], 1},
.outputs = {&kSubgraph0Outputs[0], 1},
.nodes = {&subgraph0_nodes_[0], 3},
.tensors = {&subgraph0_tensors_[0], 10},
.invoke = &InvokeSubgraph0},
},
micro_context_{&context_, {&subgraphs_[0], 1}} {
context_.impl_ = static_cast<void*>(&micro_context_);
context_.ReportError = nullptr;
context_.GetTensor = nullptr;
context_.GetEvalTensor = nullptr;
context_.GetEvalTensor = tflite::MicroContextGetEvalTensor;
context_.profiler = nullptr;
context_.GetExternalContext = nullptr;
context_.GetScratchBuffer = nullptr;
Expand Down Expand Up @@ -280,17 +308,6 @@ Model::Model() {
.type = kTfLiteInt8};
}

TfLiteStatus Model::Invoke() { return InvokeSubgraph0(); }

TfLiteStatus Model::InvokeSubgraph0() {
TF_LITE_ENSURE_OK(context_, op_table[OpCode::kFullyConnected].invoke(
&context_, &subgraph0_nodes_[0]));
TF_LITE_ENSURE_OK(context_, op_table[OpCode::kFullyConnected].invoke(
&context_, &subgraph0_nodes_[1]));
TF_LITE_ENSURE_OK(context_, op_table[OpCode::kFullyConnected].invoke(
&context_, &subgraph0_nodes_[2]));

return kTfLiteOk;
}
TfLiteStatus Model::Invoke() { return micro_context_.InvokeSubgraph(0); }

} // namespace hello_world_model
5 changes: 3 additions & 2 deletions codegen/examples/hello_world/hello_world_model.h
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@ limitations under the License.

#pragma once

#include "codegen/runtime/micro_codegen_context.h"
#include "tensorflow/lite/c/c_api_types.h"
#include "tensorflow/lite/c/common.h"

Expand All @@ -29,9 +30,9 @@ class Model {
TfLiteStatus Invoke();

private:
TfLiteStatus InvokeSubgraph0();

TfLiteContext context_ = {};
tflite::Subgraph subgraphs_[1];
tflite::MicroCodegenContext micro_context_;
TfLiteNode subgraph0_nodes_[3] = {};
TfLiteEvalTensor subgraph0_tensors_[10] = {};
};
Expand Down
70 changes: 64 additions & 6 deletions codegen/graph.py
Original file line number Diff line number Diff line change
Expand Up @@ -73,6 +73,14 @@ def __init__(self, model: schema_fb.ModelT, buffers: Sequence[tensor.Buffer],
def index(self) -> int:
return self._subgraph_idx

@property
def inputs(self) -> Sequence[int]:
return self._subgraph.inputs

@property
def outputs(self) -> Sequence[int]:
return self._subgraph.outputs

@property
def operators(self) -> Sequence[operator.Operator]:
return self._operators
Expand All @@ -85,6 +93,18 @@ def tensors(self) -> Sequence[tensor.Tensor]:
def needs_zero_length_int_array(self) -> bool:
return any(t.needs_zero_length_int_array for t in self.tensors)

@property
def invoke_fn_name(self) -> str:
return f"InvokeSubgraph{self.index}"

@property
def inputs_array_name(self) -> str:
return f"kSubgraph{self.index}Inputs"

@property
def outputs_array_name(self) -> str:
return f"kSubgraph{self.index}Outputs"

@property
def nodes_array(self) -> str:
return f"subgraph{self.index}_nodes_"
Expand Down Expand Up @@ -116,16 +136,54 @@ def generate_c_node_init(self, indent: str) -> str:
return textwrap.indent("\n".join(node_init_strs), indent)

def generate_c_invoke(self, indent: str) -> str:
invoke_template = string.Template(
"TF_LITE_ENSURE_OK(context_, op_table[${op_code}].invoke(\n"
" &context_, &${node}));\n")
function_template = string.Template(
"TfLiteStatus ${function_name}(TfLiteContext* context,\n"
" tflite::Span<TfLiteNode> nodes) {\n"
" TFLITE_DCHECK(nodes.size() == ${num_nodes});\n"
"${body}\n"
" return kTfLiteOk;\n"
"}")

body_template = string.Template(
" TF_LITE_ENSURE_OK(\n"
" context, op_table[${op_code}].invoke(context, &${node}));\n")
invoke_strs: List[str] = []
for op_idx, op in enumerate(self.operators):
invoke_strs.append(
invoke_template.substitute(
body_template.substitute(
op_code=self._op_codes[op.op_code_index].full_enum_name,
node=self.nodes_element(op_idx)))
return textwrap.indent("".join(invoke_strs), indent)
node=f"nodes[{op_idx}]"))

invoke = function_template.substitute(function_name=self.invoke_fn_name,
num_nodes=len(self.operators),
body="".join(invoke_strs))
return textwrap.indent(invoke, indent)

def generate_c_input_array(self, indent: str) -> str:
return utils.generate_c_int_array(indent, "size_t", self.inputs_array_name,
self.inputs)

def generate_c_output_array(self, indent: str) -> str:
return utils.generate_c_int_array(indent, "size_t",
self.outputs_array_name, self.outputs)

def generate_c_subgraph_init(self, indent: str) -> str:
init_template = string.Template(
"{.inputs = {&${input_array}[0], ${input_size}},\n"
" .outputs = {&${output_array}[0], ${output_size}},\n"
" .nodes = {&${node_array}[0], ${node_size}},\n"
" .tensors = {&${tensor_array}[0], ${tensor_size}},\n"
" .invoke = &${invoke}},")
return textwrap.indent(
init_template.substitute(input_array=self.inputs_array_name,
input_size=len(self.inputs),
output_array=self.outputs_array_name,
output_size=len(self.outputs),
node_array=self.nodes_array,
node_size=len(self.operators),
tensor_array=self.tensors_array,
tensor_size=len(self.tensors),
invoke=self.invoke_fn_name), indent)

@property
def tensors_array(self) -> str:
Expand Down
139 changes: 139 additions & 0 deletions codegen/runtime/micro_codegen_context.cc
Original file line number Diff line number Diff line change
@@ -0,0 +1,139 @@
/* Copyright 2023 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "codegen/runtime/micro_codegen_context.h"

#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/kernels/internal/compatibility.h"
#include "tensorflow/lite/kernels/op_macros.h"
#include "tensorflow/lite/micro/micro_log.h"

namespace tflite {

MicroCodegenContext::MicroCodegenContext(TfLiteContext* context,
Span<Subgraph> subgraphs)
: context_(context), subgraphs_(subgraphs) {}

void* MicroCodegenContext::GetScratchBuffer(int buffer_idx) {
// TODO(rjascani): Implement scratch buffers
return nullptr;
}

TfLiteEvalTensor* MicroCodegenContext::GetEvalTensor(int tensor_idx) {
TFLITE_DCHECK(static_cast<size_t>(tensor_idx) <
subgraphs_[current_subgraph_idx_].tensors.size());
return &subgraphs_[current_subgraph_idx_].tensors[tensor_idx];
}

TfLiteStatus MicroCodegenContext::set_external_context(
void* external_context_payload) {
if (external_context_payload == nullptr ||
external_context_payload_ != nullptr) {
MicroPrintf(
"Attempting to set external context to %x but it was %x already",
external_context_payload, external_context_payload_);
return kTfLiteError;
}

external_context_payload_ = external_context_payload;
return kTfLiteOk;
}

void* MicroCodegenContext::external_context() {
return external_context_payload_;
}

MicroGraph& MicroCodegenContext::graph() { return *this; }

void* MicroCodegenContext::AllocatePersistentBuffer(size_t) {
// Not allowed at Eval
TFLITE_ABORT;
return nullptr;
}

TfLiteStatus MicroCodegenContext::RequestScratchBufferInArena(size_t, int*) {
// Not allowed at Eval
TFLITE_ABORT;
return kTfLiteError;
}

TfLiteTensor* MicroCodegenContext::AllocateTempTfLiteTensor(int) {
// Not allowed at Eval
TFLITE_ABORT;
return nullptr;
}

void MicroCodegenContext::DeallocateTempTfLiteTensor(TfLiteTensor*) {
// Not allowed at Eval
TFLITE_ABORT;
}

uint8_t* MicroCodegenContext::AllocateTempBuffer(size_t, size_t) {
// Not allowed at Eval
TFLITE_ABORT;
return nullptr;
}

void MicroCodegenContext::DeallocateTempBuffer(uint8_t*) {
// Not allowed at Eval
TFLITE_ABORT;
}

TfLiteStatus MicroCodegenContext::InvokeSubgraph(int subgraph_idx) {
TF_LITE_ENSURE(context_,
static_cast<size_t>(subgraph_idx) < subgraphs_.size());
size_t previous_subgraph_idx = current_subgraph_idx_;
current_subgraph_idx_ = subgraph_idx;
TfLiteStatus status =
subgraphs_[subgraph_idx].invoke(context_, subgraphs_[subgraph_idx].nodes);
current_subgraph_idx_ = previous_subgraph_idx;
return status;
}

size_t MicroCodegenContext::NumSubgraphInputs(int subgraph_idx) {
TFLITE_DCHECK(static_cast<size_t>(subgraph_idx) < subgraphs_.size());
return subgraphs_[subgraph_idx].inputs.size();
}

TfLiteEvalTensor* MicroCodegenContext::GetSubgraphInput(int subgraph_idx,
int input_idx) {
TFLITE_DCHECK(static_cast<size_t>(subgraph_idx) < subgraphs_.size());
TFLITE_DCHECK(static_cast<size_t>(input_idx) <
subgraphs_[subgraph_idx].inputs.size());
const size_t tensor_idx = subgraphs_[subgraph_idx].inputs[input_idx];
return &subgraphs_[subgraph_idx].tensors[tensor_idx];
}

size_t MicroCodegenContext::NumSubgraphOutputs(int subgraph_idx) {
TFLITE_DCHECK(static_cast<size_t>(subgraph_idx) < subgraphs_.size());
return subgraphs_[subgraph_idx].outputs.size();
}

TfLiteEvalTensor* MicroCodegenContext::GetSubgraphOutput(int subgraph_idx,
int output_idx) {
TFLITE_DCHECK(static_cast<size_t>(subgraph_idx) < subgraphs_.size());
TFLITE_DCHECK(static_cast<size_t>(output_idx) <
subgraphs_[subgraph_idx].outputs.size());
const size_t tensor_idx = subgraphs_[subgraph_idx].outputs[output_idx];
return &subgraphs_[subgraph_idx].tensors[tensor_idx];
}

int MicroCodegenContext::NumSubgraphs() { return subgraphs_.size(); }

MicroResourceVariables* MicroCodegenContext::GetResourceVariables() {
return nullptr;
}

} // namespace tflite
Loading

0 comments on commit 3323a41

Please sign in to comment.