Create Make helpers for running codegen (#2222)

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

codegen/examples/hello_world/Makefile.inc

@@ -1,14 +1,10 @@
-# 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 \
-$(TENSORFLOW_ROOT)codegen/runtime/micro_codegen_context.cc
+CODEGEN_HELLO_WORLD_MODEL := \
+$(TENSORFLOW_ROOT)tensorflow/lite/micro/examples/hello_world/models/hello_world_int8.tflite
 
-CODEGEN_HELLO_WORLD_HDRS := \
-$(TENSORFLOW_ROOT)codegen/examples/hello_world/hello_world_model.h \
-$(TENSORFLOW_ROOT)codegen/runtime/micro_codegen_context.h
+CODEGEN_HELLO_WORLD_SRCS := \
+$(TENSORFLOW_ROOT)codegen/examples/hello_world/hello_world.cc
 
 # Builds a standalone binary.
-$(eval $(call microlite_test,codegen_hello_world,\
-$(CODEGEN_HELLO_WORLD_SRCS),,))
+$(eval $(call codegen_model_binary,codegen_hello_world,hello_world_model,\
+$(CODEGEN_HELLO_WORLD_MODEL),$(CODEGEN_HELLO_WORLD_SRCS),,))
+

codegen/examples/hello_world/README.md

@@ -1,51 +1,28 @@
 # Codegen Hello World Example
 
-This is a code-generated example of the hello world model. The process is
-currently somewhat involved:
+This is a code-generated example of the hello world model. The generated source
+is checked in for now so that it can be reviewed during the prototyping stage.
 
-## Build the preprocessor for your target
+## Building the example executable
+Please note that this will execute Bazel from make as part of the process.
 
-This creates a target-specific preprocessor binary capable of performing the
-init and prepare stages of the Interpreter and serializing the output. This
-binary can be re-used for multiple models.
-
-### x86
 ```
-make -f tensorflow/lite/micro/tools/make/Makefile codegen_preprocessor
+make -f tensorflow/lite/micro/tools/make/Makefile codegen_hello_world
 ```
 
-## Run the preprocessor
+## Running the example
 
-The preprocessor will take the provided model, create a TFLM Interpreter, and
-allocate tensors. It will then capture and serialize the resulting data
-structures needed for inference. For embedded targets, this should be run under
-simulation.
+TODO(rjascani): The command works, but it'll just crash as we don't have all of
+the data structures fully populated yet.
 
-### x86
 ```
-./gen/linux_x86_64_default/bin/codegen_preprocessor \
-  $(pwd)/tensorflow/lite/micro/examples/hello_world/models/hello_world_int8.tflite \
-  $(pwd)/gen/linux_86_64_default/genfiles/hello_world_int8.ppd
+make -f tensorflow/lite/micro/tools/make/Makefile run_codegen_hello_world
 ```
 
-## Generate the inference code
+## Updating the generated sources
+To update the generated source, you can execute this make target:
 
-To generate the inference code at `codegen/example/hello_world_model.h/.cc`:
-
-### x86
 ```
-bazel run codegen:code_generator -- \
-  --model $(pwd)/tensorflow/lite/micro/examples/hello_world/models/hello_world_int8.tflite \
-  --preprocessed_data $(pwd)/gen/linux_86_64_default/genfiles/hello_world_int8.ppd \
-  --output_dir $(pwd)/codegen/examples/hello_world \
-  --output_name hello_world_model
+./codegen/examples/hello_world/update_example_source.sh
 ```
 
-## Compile the generated inference code
-
- To compile the generated source, you can use the Makefile:
-
-### x86
-```
-make -f tensorflow/lite/micro/tools/make/Makefile codegen_hello_world
-```

codegen/examples/hello_world/update_example_source.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+# 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.
+# ==============================================================================
+
+#
+# Syncs the generated example source code in the repository.
+#
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ROOT_DIR=${SCRIPT_DIR}/../../..
+cd "${ROOT_DIR}"
+
+make -j8 -f tensorflow/lite/micro/tools/make/Makefile codegen_hello_world
+cp ./gen/linux_x86_64_default/genfiles/hello_world_model.h ${SCRIPT_DIR}
+cp ./gen/linux_x86_64_default/genfiles/hello_world_model.cc ${SCRIPT_DIR}

tensorflow/lite/micro/tools/make/Makefile

@@ -21,6 +21,10 @@ endif
 TENSORFLOW_ROOT :=
 RELATIVE_MAKEFILE_DIR := tensorflow/lite/micro/tools/make
 MAKEFILE_DIR := $(TENSORFLOW_ROOT)$(RELATIVE_MAKEFILE_DIR)
+BAZEL_ROOT := $(TENSORFLOW_ROOT)
+ifeq ($(BAZEL_ROOT),)
+ BAZEL_ROOT = .
+endif
 
 # Pull in some convenience functions.
 include $(MAKEFILE_DIR)/helper_functions.inc
@@ -294,6 +298,8 @@ MICRO_LITE_BENCHMARKS := $(wildcard $(TENSORFLOW_ROOT)tensorflow/lite/micro/tool
 MICROLITE_BENCHMARK_SRCS := \
 $(wildcard $(TENSORFLOW_ROOT)tensorflow/lite/micro/tools/benchmarking/*benchmark.cc)
 
+CODEGEN_PREPROCESSOR_PATH := $(BINDIR)codegen_preprocessor
+
 MICRO_LITE_CODEGEN_PREPROCESSOR := $(TENSORFLOW_ROOT)codegen/preprocessor/Makefile.inc
 
 MICRO_LITE_CODEGEN_EXAMPLES := $(shell find $(TENSORFLOW_ROOT)codegen/examples/ -name Makefile.inc)
@@ -552,7 +558,11 @@ include $(MAKEFILE_DIR)/additional_kernels.inc
 MICROLITE_CC_SRCS := $(filter-out $(MICROLITE_TEST_SRCS), $(MICROLITE_CC_BASE_SRCS))
 MICROLITE_CC_SRCS := $(filter-out $(MICROLITE_BENCHMARK_SRCS), $(MICROLITE_CC_SRCS))
 
+CODEGEN_RUNTIME_CC_SRCS := \
+$(TENSORFLOW_ROOT)codegen/runtime/micro_codegen_context.cc
 
+CODEGEN_RUNTIME_CC_HDRS := \
+$(TENSORFLOW_ROOT)codegen/runtime/micro_codegen_context.h
 
 # The download scripts require that the downloads directory already exist for
 # improved error checking. To accomodate that, we first create a downloads
@@ -868,8 +878,6 @@ integration_tests: $(MICROLITE_INTEGRATION_TEST_TARGETS)
 generated_micro_mutable_op_resolver: $(MICROLITE_GEN_OP_RESOLVER_TEST_TARGETS)
 endif
 
-CODEGEN_PREPROCESSOR_PATH := $(BINDIR)codegen_preprocessor
-
 codegen_preprocessor: $(CODEGEN_PREPROCESSOR_PATH)
 
 $(CODEGEN_PREPROCESSOR_PATH): $(CODEGEN_PREPROCESSOR_SRCS) $(MICROLITE_LIB_PATH)

tensorflow/lite/micro/tools/make/helper_functions.inc

@@ -117,3 +117,94 @@ endef
 # 2 - File pattern, e.g: *.h
 recursive_find = $(wildcard $(1)$(2)) $(foreach dir,$(wildcard $(1)*),$(call recursive_find,$(dir)/,$(2)))
 
+# Generates code capable of performing inference without an interpreter. It run
+# the codegen preprocessor and the code generator.
+#
+# Arguments are:
+# 1 - Name of target
+# 2 - Generated source basename
+# 3 - Model
+# Calling eval on the output will create the targets that you need.
+define codegen_model
+# Filter out targets that currently don't support codegen:
+# Bluepill: Is compiled with nostdlib, but preprocessor uses standard library.
+# RISC-V: TODO(b/300484340): qemu-riscv32 currently does not support semi-
+#         hosting, which prevents the preprocessor from opening a file on the
+#         host filesystem.
+# Hexagon: TODO(b/300322637): The hexagon 3.5.1 SDK doesn't provide a working
+#          C++11 stdlib, so preprocessor fails to link.
+ifneq ($(TARGET), $(filter $(TARGET), bluepill riscv32_generic hexagon))
+
+$(1)_MODEL := $(3)
+$(1)_PREPROCESSOR_OUTPUT := $(GENERATED_SRCS_DIR)$(2).ppd
+
+$(1)_GENERATED_SRCS := $(GENERATED_SRCS_DIR)$(2).cc
+$(1)_GENERATED_HDRS := $(GENERATED_SRCS_DIR)$(2).h
+
+$$($(1)_PREPROCESSOR_OUTPUT): $(CODEGEN_PREPROCESSOR_PATH) $$($(1)_MODEL)
+	@mkdir -p $$(dir $$@)
+	$$(RUN_COMMAND) $(CODEGEN_PREPROCESSOR_PATH) \
+	$(abspath $$($(1)_MODEL)) $(abspath $$($(1)_PREPROCESSOR_OUTPUT))
+
+$$($(1)_GENERATED_SRCS) $$($(1)_GENERATED_HDRS): $$($(1)_MODEL) $$($(1)_PREPROCESSOR_OUTPUT)
+	cd $(BAZEL_ROOT) && bazel run //codegen:code_generator -- \
+	--model $(abspath $$($(1)_MODEL)) \
+	--preprocessed_data $(abspath $$($(1)_PREPROCESSOR_OUTPUT)) \
+	--output_dir $(abspath $(GENERATED_SRCS_DIR)) --output_name $(2)
+
+$(1): $$($(1)_GENERATED_SRCS) $$($(1)_GENERATED_HDRS)
+
+endif
+endef # codegen_model
+
+# Generates and compiles code capable of performing inference without an
+# interpreter.
+#
+# Users can use `make run_<target>` to execute the binary in the appropriate
+# simulator.
+#
+# Arguments are:
+# 1 - Name of target
+# 2 - Generated source basename
+# 3 - Model
+# 4 - C/C++ source files
+# 5 - C/C++ header files
+# Calling eval on the output will create the targets that you need.
+define codegen_model_binary
+# Filter out targets that currently don't support codegen:
+# Bluepill: Is compiled with nostdlib, but preprocessor uses standard library.
+# RISC-V: TODO(b/300484340): qemu-riscv32 currently does not support semi-
+#         hosting, which prevents the preprocessor from opening a file on the
+#         host filesystem.
+# Hexagon: TODO(b/300322637): The hexagon 3.5.1 SDK doesn't provide a working
+#          C++11 stdlib, so preprocessor fails to link.
+ifneq ($(TARGET), $(filter $(TARGET), bluepill riscv32_generic hexagon))
+
+$(1)_CODEGEN_SRCS := $(4) $$(CODEGEN_RUNTIME_CC_SRCS)
+$(1)_CODEGEN_HDRS := $(5) $$(CODEGEN_RUNTIME_CC_HDRS)
+
+$(call codegen_model,$(1)_codegen,$(2),$(3))
+
+$(1)_CODEGEN_SRCS += $$($(1)_codegen_GENERATED_SRCS)
+$(1)_CODEGEN_HDRS += $$($(1)_codegen_GENERATED_HDRS)
+
+$(1)_CODEGEN_OBJS := $$(addprefix $$(CORE_OBJDIR), \
+$$(patsubst %.S,%.o,$$(patsubst %.cc,%.o,$$(patsubst %.c,%.o,$$($(1)_CODEGEN_SRCS)))))
+
+$(1)_BINARY := $$(BINDIR)$(1)
+$$($(1)_BINARY): $$($(1)_CODEGEN_OBJS) $$(MICROLITE_LIB_PATH)
+	@mkdir -p $$(dir $$@)
+	$$(CXX) $$(CXXFLAGS) $$(INCLUDES) \
+	-o $$($(1)_BINARY) $$($(1)_CODEGEN_OBJS) \
+	$$(MICROLITE_LIB_PATH) $$(LDFLAGS) $$(MICROLITE_LIBS)
+
+$(1): $$($(1)_BINARY)
+$(1)_bin: $$($(1)_BINARY).bin
+
+MICROLITE_BUILD_TARGETS += $$($(1)_BINARY)
+
+run_$(1): $$($(1)_BINARY)
+	$$(RUN_COMMAND) $$($(1)_BINARY)
+
+endif
+endef # codegen_model_binary

tensorflow/lite/micro/tools/make/targets/cortex_m_qemu_makefile.inc

@@ -44,3 +44,4 @@ EXCLUDED_TESTS := \
 MICROLITE_TEST_SRCS := $(filter-out $(EXCLUDED_TESTS), $(MICROLITE_TEST_SRCS))
 
 TEST_SCRIPT := $(TENSORFLOW_ROOT)tensorflow/lite/micro/testing/test_with_qemu.sh arm $(TARGET_ARCH)
+RUN_COMMAND := qemu-arm -cpu $(TARGET_ARCH)

tensorflow/lite/micro/tools/make/targets/hexagon_makefile.inc

@@ -114,3 +114,4 @@ MICROLITE_TEST_SRCS := $(filter-out $(EXCLUDED_TESTS), $(MICROLITE_TEST_SRCS))
 
 TEST_SCRIPT := $(TENSORFLOW_ROOT)tensorflow/lite/micro/testing/test_hexagon_binary.sh
 SIZE_SCRIPT := $(TENSORFLOW_ROOT)tensorflow/lite/micro/testing/size_hexagon_binary.sh
+RUN_COMMAND := hexagon-sim

tensorflow/lite/micro/tools/make/targets/riscv32_generic_makefile.inc

@@ -45,4 +45,5 @@ MICROLITE_TEST_SRCS := $(filter-out $(EXCLUDED_TESTS), $(MICROLITE_TEST_SRCS))
 LDFLAGS += -mno-relax
 TEST_SCRIPT := $(TENSORFLOW_ROOT)tensorflow/lite/micro/testing/test_with_qemu.sh riscv32 rv32
 SIZE_SCRIPT := ${TENSORFLOW_ROOT}tensorflow/lite/micro/testing/size_riscv32_binary.sh
+RUN_COMMAND := qemu-riscv32 -cpu rv32
 

tensorflow/lite/micro/tools/make/targets/xtensa_makefile.inc

@@ -72,6 +72,7 @@ CXXFLAGS += $(XTENSA_EXTRA_CFLAGS)
 
 TEST_SCRIPT := $(TENSORFLOW_ROOT)tensorflow/lite/micro/testing/test_xtensa_binary.sh
 SIZE_SCRIPT := $(TENSORFLOW_ROOT)tensorflow/lite/micro/testing/size_xtensa_binary.sh
+RUN_COMMAND := xt-run
 
 # TODO(b/158651472): Fix the memory_arena_threshold_test
 # TODO(b/174707181): Fix the micro_interpreter_test