[llvm] 5ce4c9a - [mlgo] Use TFLite for 'development' mode.
Mircea Trofin via llvm-commits
llvm-commits at lists.llvm.org
Wed Aug 24 16:07:49 PDT 2022
Author: Mircea Trofin
Date: 2022-08-24T16:07:24-07:00
New Revision: 5ce4c9aa0405bb28dbe374aa741044025d63f05c
URL: https://github.com/llvm/llvm-project/commit/5ce4c9aa0405bb28dbe374aa741044025d63f05c
DIFF: https://github.com/llvm/llvm-project/commit/5ce4c9aa0405bb28dbe374aa741044025d63f05c.diff
LOG: [mlgo] Use TFLite for 'development' mode.
TLite is a lightweight, statically linkable[1], model evaluator, supporting a
subset of what the full tensorflow library does, sufficient for the
types of scenarios we envision having. It is also faster.
We still use saved models as "source of truth" - 'release' mode's AOT
starts from a saved model; and the ML training side operates in terms of
saved models.
Using TFLite solves the following problems compared to using the full TF
C API:
- a compiler-friendly implementation for runtime-loadable (as opposed
to AOT-embedded) models: it's statically linked; it can be built via
cmake;
- solves an issue we had when building the compiler with both AOT and
full TF C API support, whereby, due to a packaging issue on the TF
side, we needed to have the pip package and the TF C API library at
the same version. We have no such constraints now.
The main liability is it supporting a subset of what the full TF
framework does. We do not expect that to cause an issue, but should that
be the case, we can always revert back to using the full framework
(after also figuring out a way to address the problems that motivated
the move to TFLite).
Details:
This change switches the development mode to TFLite. Models are still
expected to be placed in a directory - i.e. the parameters to clang
don't change; what changes is the directory content: we still need
an `output_spec.json` file; but instead of the saved_model protobuf and
the `variables` directory, we now just have one file, `model.tflite`.
The change includes a utility showing how to take a saved model and
convert it to TFLite, which it uses for testing.
The full TF implementation can still be built (not side-by-side). We
intend to remove it shortly, after patching downstream dependencies. The
build behavior, however, prioritizes TFLite - i.e. trying to enable both
full TF C API and TFLite will just pick TFLite.
[1] thanks to @petrhosek's changes to TFLite's cmake support and its deps!
Added:
llvm/lib/Analysis/TFLiteUtils.cpp
llvm/lib/Analysis/models/saved-model-to-tflite.py
llvm/unittests/Analysis/Inputs/ir2native_x86_64_model/model.tflite
Modified:
llvm/CMakeLists.txt
llvm/cmake/modules/TensorFlowCompile.cmake
llvm/include/llvm/Config/llvm-config.h.cmake
llvm/lib/Analysis/CMakeLists.txt
llvm/lib/Analysis/TFUtils.cpp
llvm/test/CodeGen/MLRegalloc/dev-mode-log-2-fcts.ll
llvm/test/CodeGen/MLRegalloc/dev-mode-logging.ll
llvm/test/CodeGen/MLRegalloc/dev-rel-equivalence.ll
llvm/test/Transforms/Inline/ML/bounds-checks-rewards.ll
llvm/test/Transforms/Inline/ML/development-training-log.ll
llvm/test/Transforms/Inline/ML/ml-test-development-mode.ll
llvm/unittests/Analysis/TFUtilsTest.cpp
utils/bazel/llvm_configs/llvm-config.h.cmake
Removed:
################################################################################
diff --git a/llvm/CMakeLists.txt b/llvm/CMakeLists.txt
index b52e2d04b646f..67cb5ce7d0e5f 100644
--- a/llvm/CMakeLists.txt
+++ b/llvm/CMakeLists.txt
@@ -944,15 +944,17 @@ if(LLVM_INCLUDE_TESTS)
get_errc_messages(LLVM_LIT_ERRC_MESSAGES)
endif()
-# For up-to-date instructions for installing the Tensorflow dependency, refer to
+# For up-to-date instructions for installing the TFLite dependency, refer to
# the bot setup script: https://github.com/google/ml-compiler-opt/blob/main/buildbot/buildbot_init.sh
-# In this case, the latest C API library is available for download from
-# https://www.tensorflow.org/install/lang_c.
-# We will expose the conditional compilation variable,
-# LLVM_HAVE_TF_API, through llvm-config.h, so that a user of the LLVM library may
-# also leverage the dependency.
set(TENSORFLOW_C_LIB_PATH "" CACHE PATH "Path to TensorFlow C library install")
-if (TENSORFLOW_C_LIB_PATH)
+set(LLVM_HAVE_TFLITE "" CACHE BOOL "Use tflite")
+if (LLVM_HAVE_TFLITE)
+ find_package(protobuf REQUIRED)
+ find_package(tensorflow-lite REQUIRED)
+ set(LLVM_HAVE_TF_API "ON" CACHE BOOL "Full Tensorflow API available")
+ set(LLVM_PROTOBUF_OUT_DIR ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/protobuf_gen)
+ include_directories(${LLVM_PROTOBUF_OUT_DIR})
+elseif (TENSORFLOW_C_LIB_PATH)
find_library(tensorflow_c_api tensorflow PATHS ${TENSORFLOW_C_LIB_PATH}/lib NO_DEFAULT_PATH REQUIRED)
# Currently, the protobuf headers are distributed with the pip package that corresponds to the version
# of the C API library.
@@ -989,7 +991,12 @@ if (NOT TENSORFLOW_AOT_PATH STREQUAL "")
set(TENSORFLOW_AOT_COMPILER
"${TENSORFLOW_AOT_PATH}/../../../../bin/saved_model_cli"
CACHE PATH "Path to the Tensorflow AOT compiler")
- include_directories(${TENSORFLOW_AOT_PATH}/include)
+ # This needs to happen to avoid clashing protobuf codegen when building both AOT and development mode.
+ # We plan to replace protobuf with a simpler alternative, so this will go away.
+ file(COPY ${TENSORFLOW_AOT_PATH}/include DESTINATION ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/tensorflow
+ PATTERN "google/*" EXCLUDE
+ PATTERN "*.pb.h" EXCLUDE)
+ include_directories(${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/tensorflow/include)
add_subdirectory(${TENSORFLOW_AOT_PATH}/xla_aot_runtime_src
${CMAKE_ARCHIVE_OUTPUT_DIRECTORY}/tf_runtime)
install(TARGETS tf_xla_runtime EXPORT LLVMExports
diff --git a/llvm/cmake/modules/TensorFlowCompile.cmake b/llvm/cmake/modules/TensorFlowCompile.cmake
index d5ff44c9330d1..45ec3fe7c327e 100644
--- a/llvm/cmake/modules/TensorFlowCompile.cmake
+++ b/llvm/cmake/modules/TensorFlowCompile.cmake
@@ -116,3 +116,18 @@ function(tf_find_and_compile model default_url default_path test_model_generator
set(MLLinkDeps ${MLLinkDeps} tf_xla_runtime PARENT_SCOPE)
add_definitions(-DLLVM_HAVE_TF_AOT_${fname_allcaps})
endfunction()
+
+function(build_proto)
+ foreach (P ${ARGV})
+ set(PB_SRCS ${PB_SRCS} ${LLVM_PROTOBUF_OUT_DIR}/${P}.pb.cc)
+ set(PB_HDRS ${PB_HDRS} ${LLVM_PROTOBUF_OUT_DIR}/${P}.pb.h)
+ set(PBS ${PBS} ${TENSORFLOW_SRC_DIR}/${P}.proto)
+ endforeach()
+ add_custom_command(OUTPUT ${PB_SRCS} ${PB_HDRS}
+ COMMAND protobuf::protoc
+ ARGS --proto_path=${TENSORFLOW_SRC_DIR} --cpp_out=${LLVM_PROTOBUF_OUT_DIR} ${PBS})
+ set_source_files_properties(${PB_SRCS} PROPERTIES
+ GENERATED 1)
+ set(GeneratedMLSources ${GeneratedMLSources} ${PB_SRCS} PARENT_SCOPE)
+ set(MLDeps ${MLDeps} ${MLDeps} PARENT_SCOPE)
+endfunction()
diff --git a/llvm/include/llvm/Config/llvm-config.h.cmake b/llvm/include/llvm/Config/llvm-config.h.cmake
index 96fbe6f771d2a..b289db1d46ec4 100644
--- a/llvm/include/llvm/Config/llvm-config.h.cmake
+++ b/llvm/include/llvm/Config/llvm-config.h.cmake
@@ -101,6 +101,9 @@
/* Define if LLVM was built with a dependency to the libtensorflow dynamic library */
#cmakedefine LLVM_HAVE_TF_API
+/* Define if LLVM is using tflite instead of libtensorflow */
+#cmakedefine LLVM_HAVE_TFLITE
+
/* Define to 1 if you have the <sysexits.h> header file. */
#cmakedefine HAVE_SYSEXITS_H ${HAVE_SYSEXITS_H}
diff --git a/llvm/lib/Analysis/CMakeLists.txt b/llvm/lib/Analysis/CMakeLists.txt
index a122851e5495f..4bf14a86225d8 100644
--- a/llvm/lib/Analysis/CMakeLists.txt
+++ b/llvm/lib/Analysis/CMakeLists.txt
@@ -18,7 +18,17 @@ if (DEFINED LLVM_HAVE_TF_AOT OR DEFINED LLVM_HAVE_TF_API)
endif()
if (DEFINED LLVM_HAVE_TF_API)
- list(APPEND MLLinkDeps ${tensorflow_c_api} ${tensorflow_fx})
+ if (DEFINED LLVM_HAVE_TFLITE)
+ build_proto(
+ tensorflow/core/protobuf/error_codes
+ tensorflow/core/example/feature
+ tensorflow/core/example/example)
+ list(APPEND MLLinkDeps
+ tensorflow-lite::tensorflow-lite
+ protobuf::libprotobuf)
+ else()
+ list(APPEND MLLinkDeps ${tensorflow_c_api} ${tensorflow_fx})
+ endif()
endif()
endif()
@@ -130,6 +140,7 @@ add_llvm_component_library(LLVMAnalysis
SyncDependenceAnalysis.cpp
SyntheticCountsUtils.cpp
TFUtils.cpp
+ TFLiteUtils.cpp
TargetLibraryInfo.cpp
TargetTransformInfo.cpp
TensorSpec.cpp
diff --git a/llvm/lib/Analysis/TFLiteUtils.cpp b/llvm/lib/Analysis/TFLiteUtils.cpp
new file mode 100644
index 0000000000000..96eabb2b2c384
--- /dev/null
+++ b/llvm/lib/Analysis/TFLiteUtils.cpp
@@ -0,0 +1,232 @@
+//===- TFUtils.cpp - tensorflow evaluation utilities ----------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements utilities for interfacing with tensorflow C APIs.
+//
+//===----------------------------------------------------------------------===//
+#include "llvm/Config/config.h"
+#if defined(LLVM_HAVE_TFLITE)
+
+#include "llvm/ADT/Twine.h"
+#include "llvm/Analysis/Utils/TFUtils.h"
+#include "llvm/Support/Base64.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/JSON.h"
+#include "llvm/Support/MemoryBuffer.h"
+#include "llvm/Support/Path.h"
+#include "llvm/Support/raw_ostream.h"
+
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/model_builder.h"
+#include "tensorflow/lite/op_resolver.h"
+
+#include <cassert>
+#include <numeric>
+
+using namespace llvm;
+
+namespace llvm {
+class EvaluationResultImpl {
+public:
+ EvaluationResultImpl(const std::vector<const TfLiteTensor *> &Outputs)
+ : Outputs(Outputs){};
+
+ const TfLiteTensor *getOutput(size_t I) { return Outputs[I]; }
+
+ EvaluationResultImpl(const EvaluationResultImpl &) = delete;
+ EvaluationResultImpl(EvaluationResultImpl &&Other) = delete;
+
+private:
+ const std::vector<const TfLiteTensor *> Outputs;
+};
+
+class TFModelEvaluatorImpl {
+public:
+ TFModelEvaluatorImpl(StringRef SavedModelPath,
+ const std::vector<TensorSpec> &InputSpecs,
+ function_ref<TensorSpec(size_t)> GetOutputSpecs,
+ size_t OutputSpecsSize, const char *Tags);
+
+ bool isValid() const { return IsValid; }
+ size_t outputSize() const { return Output.size(); }
+
+ std::unique_ptr<EvaluationResultImpl> evaluate() {
+ Interpreter->Invoke();
+ return std::make_unique<EvaluationResultImpl>(Output);
+ }
+
+ const std::vector<TfLiteTensor *> &getInput() const { return Input; }
+
+ ~TFModelEvaluatorImpl();
+
+private:
+ std::unique_ptr<tflite::FlatBufferModel> Model;
+
+ /// The objects necessary for carrying out an evaluation of the SavedModel.
+ /// They are expensive to set up, and we maintain them accross all the
+ /// evaluations of the model.
+ std::unique_ptr<tflite::Interpreter> Interpreter;
+
+ /// The input tensors. We set up the tensors once and just mutate theirs
+ /// scalars before each evaluation. The input tensors keep their value after
+ /// an evaluation.
+ std::vector<TfLiteTensor *> Input;
+
+ /// The output nodes.
+ std::vector<const TfLiteTensor *> Output;
+
+ void invalidate() { IsValid = false; }
+
+ bool IsValid = true;
+
+ /// Reusable utility for ensuring we can bind the requested Name to a node in
+ /// the SavedModel Graph.
+ bool checkReportAndInvalidate(const TfLiteTensor *Tensor,
+ const TensorSpec &Spec);
+};
+
+} // namespace llvm
+
+TFModelEvaluatorImpl::TFModelEvaluatorImpl(
+ StringRef SavedModelPath, const std::vector<TensorSpec> &InputSpecs,
+ function_ref<TensorSpec(size_t)> GetOutputSpecs, size_t OutputSpecsSize,
+ const char *Tags = "serve")
+ : Input(InputSpecs.size()), Output(OutputSpecsSize) {
+ // FIXME: make ErrorReporter a member (may also need subclassing
+ // StatefulErrorReporter) to easily get the latest error status, for
+ // debugging.
+ tflite::StderrReporter ErrorReporter;
+ SmallVector<char, 128> TFLitePathBuff;
+ llvm::sys::path::append(TFLitePathBuff, SavedModelPath, "model.tflite");
+ StringRef TFLitePath(TFLitePathBuff.data(), TFLitePathBuff.size());
+ Model = tflite::FlatBufferModel::BuildFromFile(TFLitePath.str().c_str(),
+ &ErrorReporter);
+ if (!Model) {
+ invalidate();
+ return;
+ }
+
+ tflite::ops::builtin::BuiltinOpResolver Resolver;
+ tflite::InterpreterBuilder Builder(*Model, Resolver);
+ Builder(&Interpreter);
+
+ if (!Interpreter ||
+ Interpreter->AllocateTensors() != TfLiteStatus::kTfLiteOk) {
+ invalidate();
+ return;
+ }
+ // Known inputs and outputs
+ StringMap<int> InputsMap;
+ StringMap<int> OutputsMap;
+ for (size_t I = 0; I < Interpreter->inputs().size(); ++I)
+ InputsMap[Interpreter->GetInputName(I)] = I;
+ for (size_t I = 0; I < Interpreter->outputs().size(); ++I)
+ OutputsMap[Interpreter->GetOutputName(I)] = I;
+
+ for (size_t I = 0; I < InputSpecs.size(); ++I) {
+ auto &InputSpec = InputSpecs[I];
+ auto MapI = InputsMap.find(InputSpec.name() + ":" +
+ std::to_string(InputSpec.port()));
+ if (MapI == InputsMap.end()) {
+ Input[I] = nullptr;
+ continue;
+ }
+ Input[I] = Interpreter->tensor(MapI->second);
+ if (!checkReportAndInvalidate(Input[I], InputSpec))
+ return;
+ std::memset(Input[I]->data.data, 0,
+ InputSpecs[I].getTotalTensorBufferSize());
+ }
+
+ for (size_t I = 0; I < OutputSpecsSize; ++I) {
+ auto OutputSpec = GetOutputSpecs(I);
+ Output[I] = Interpreter->output_tensor(
+ OutputsMap[OutputSpec.name() + ":" +
+ std::to_string(OutputSpec.port())]);
+ if (!checkReportAndInvalidate(Output[I], OutputSpec))
+ return;
+ }
+}
+
+TFModelEvaluator::TFModelEvaluator(
+ StringRef SavedModelPath, const std::vector<TensorSpec> &InputSpecs,
+ function_ref<TensorSpec(size_t)> GetOutputSpecs, size_t OutputSpecsSize,
+ const char *Tags)
+ : Impl(new TFModelEvaluatorImpl(SavedModelPath, InputSpecs, GetOutputSpecs,
+ OutputSpecsSize, Tags)) {
+ if (!Impl->isValid())
+ Impl.reset();
+}
+
+TFModelEvaluator::TFModelEvaluator(StringRef SavedModelPath,
+ const std::vector<TensorSpec> &InputSpecs,
+ const std::vector<TensorSpec> &OutputSpecs,
+ const char *Tags)
+ : TFModelEvaluator(
+ SavedModelPath, InputSpecs, [&](size_t I) { return OutputSpecs[I]; },
+ OutputSpecs.size(), Tags) {}
+
+TFModelEvaluatorImpl::~TFModelEvaluatorImpl() {}
+
+bool TFModelEvaluatorImpl::checkReportAndInvalidate(const TfLiteTensor *Tensor,
+ const TensorSpec &Spec) {
+ if (!Tensor) {
+ errs() << "Could not find TF_Output named: " + Spec.name();
+ IsValid = false;
+ }
+ if (Spec.getTotalTensorBufferSize() != Tensor->bytes)
+ IsValid = false;
+
+ // If the total sizes match, there could still be a mismatch in the shape.
+ // We ignore that for now.
+
+ return IsValid;
+}
+
+Optional<TFModelEvaluator::EvaluationResult> TFModelEvaluator::evaluate() {
+ if (!isValid())
+ return None;
+ return EvaluationResult(Impl->evaluate());
+}
+
+void *TFModelEvaluator::getUntypedInput(size_t Index) {
+ TfLiteTensor *T = Impl->getInput()[Index];
+ if (!T)
+ return nullptr;
+ return T->data.data;
+}
+
+TFModelEvaluator::EvaluationResult::EvaluationResult(
+ std::unique_ptr<EvaluationResultImpl> Impl)
+ : Impl(std::move(Impl)) {}
+
+TFModelEvaluator::EvaluationResult::EvaluationResult(EvaluationResult &&Other)
+ : Impl(std::move(Other.Impl)) {}
+
+TFModelEvaluator::EvaluationResult &
+TFModelEvaluator::EvaluationResult::operator=(EvaluationResult &&Other) {
+ Impl = std::move(Other.Impl);
+ return *this;
+}
+
+void *TFModelEvaluator::EvaluationResult::getUntypedTensorValue(size_t Index) {
+ return Impl->getOutput(Index)->data.data;
+}
+
+const void *
+TFModelEvaluator::EvaluationResult::getUntypedTensorValue(size_t Index) const {
+ return Impl->getOutput(Index)->data.data;
+}
+
+TFModelEvaluator::EvaluationResult::~EvaluationResult() {}
+TFModelEvaluator::~TFModelEvaluator() {}
+
+#endif // defined(LLVM_HAVE_TF_API)
diff --git a/llvm/lib/Analysis/TFUtils.cpp b/llvm/lib/Analysis/TFUtils.cpp
index de5dde8d2ea4f..980ee5908f98f 100644
--- a/llvm/lib/Analysis/TFUtils.cpp
+++ b/llvm/lib/Analysis/TFUtils.cpp
@@ -10,7 +10,7 @@
//
//===----------------------------------------------------------------------===//
#include "llvm/Config/config.h"
-#if defined(LLVM_HAVE_TF_API)
+#if defined(LLVM_HAVE_TF_API) && !defined(LLVM_HAVE_TFLITE)
#include "llvm/ADT/Twine.h"
#include "llvm/Analysis/Utils/TFUtils.h"
diff --git a/llvm/lib/Analysis/models/saved-model-to-tflite.py b/llvm/lib/Analysis/models/saved-model-to-tflite.py
new file mode 100644
index 0000000000000..e9d45fdf983b7
--- /dev/null
+++ b/llvm/lib/Analysis/models/saved-model-to-tflite.py
@@ -0,0 +1,37 @@
+"""Convert a saved model to tflite model.
+
+Usage: python3 saved-model-to-tflite.py <mlgo saved_model_dir> <tflite dest_dir>
+
+The <tflite dest_dir> will contain:
+ model.tflite: this is the converted saved model
+ output_spec.json: the output spec, copied from the saved_model dir.
+"""
+
+import tensorflow as tf
+import os
+import sys
+from tf_agents.policies import greedy_policy
+
+
+def main(argv):
+ assert len(argv) == 3
+ sm_dir = argv[1]
+ tfl_dir = argv[2]
+ tf.io.gfile.makedirs(tfl_dir)
+ tfl_path = os.path.join(tfl_dir, 'model.tflite')
+ converter = tf.lite.TFLiteConverter.from_saved_model(sm_dir)
+ converter.target_spec.supported_ops = [
+ tf.lite.OpsSet.TFLITE_BUILTINS,
+ ]
+ tfl_model = converter.convert()
+ with tf.io.gfile.GFile(tfl_path, 'wb') as f:
+ f.write(tfl_model)
+
+ json_file = 'output_spec.json'
+ src_json = os.path.join(sm_dir, json_file)
+ if tf.io.gfile.exists(src_json):
+ tf.io.gfile.copy(src_json,
+ os.path.join(tfl_dir, json_file))
+
+if __name__ == '__main__':
+ main(sys.argv)
diff --git a/llvm/test/CodeGen/MLRegalloc/dev-mode-log-2-fcts.ll b/llvm/test/CodeGen/MLRegalloc/dev-mode-log-2-fcts.ll
index c22d95d698a42..14183ca8a9b6a 100644
--- a/llvm/test/CodeGen/MLRegalloc/dev-mode-log-2-fcts.ll
+++ b/llvm/test/CodeGen/MLRegalloc/dev-mode-log-2-fcts.ll
@@ -11,8 +11,9 @@
; RUN: sed -i 's/\\n/ /g' %t1
; RUN: FileCheck --input-file %t1 %s
-; RUN: rm -rf %t && mkdir %t
-; RUN: %python %S/../../../lib/Analysis/models/gen-regalloc-eviction-test-model.py %t
+; RUN: rm -rf %t %t_savedmodel
+; RUN: %python %S/../../../lib/Analysis/models/gen-regalloc-eviction-test-model.py %t_savedmodel
+; RUN: %python %S/../../../lib/Analysis/models/saved-model-to-tflite.py %t_savedmodel %t
; RUN: llc -mtriple=x86_64-linux-unknown -regalloc=greedy -regalloc-enable-advisor=development \
; RUN: -regalloc-training-log=%t2 -tfutils-text-log -regalloc-model=%t < %s
; RUN: sed -i 's/ \+/ /g' %t2
diff --git a/llvm/test/CodeGen/MLRegalloc/dev-mode-logging.ll b/llvm/test/CodeGen/MLRegalloc/dev-mode-logging.ll
index 351d46376cb83..e1a3dada665eb 100644
--- a/llvm/test/CodeGen/MLRegalloc/dev-mode-logging.ll
+++ b/llvm/test/CodeGen/MLRegalloc/dev-mode-logging.ll
@@ -12,8 +12,9 @@
; RUN: FileCheck --input-file %t1 %s --check-prefixes=CHECK,NOML
; RUN:
diff %t1 %S/Inputs/reference-log-noml.txt
-; RUN: rm -rf %t && mkdir %t
-; RUN: %python %S/../../../lib/Analysis/models/gen-regalloc-eviction-test-model.py %t
+; RUN: rm -rf %t_savedmodel %t
+; RUN: %python %S/../../../lib/Analysis/models/gen-regalloc-eviction-test-model.py %t_savedmodel
+; RUN: %python %S/../../../lib/Analysis/models/saved-model-to-tflite.py %t_savedmodel %t
; RUN: llc -mtriple=x86_64-linux-unknown -regalloc=greedy -regalloc-enable-advisor=development \
; RUN: -regalloc-training-log=%t2 -tfutils-text-log -regalloc-model=%t < %S/Inputs/input.ll
; RUN: sed -i 's/ \+/ /g' %t2
diff --git a/llvm/test/CodeGen/MLRegalloc/dev-rel-equivalence.ll b/llvm/test/CodeGen/MLRegalloc/dev-rel-equivalence.ll
index 6f1b265480b42..43c363750f939 100644
--- a/llvm/test/CodeGen/MLRegalloc/dev-rel-equivalence.ll
+++ b/llvm/test/CodeGen/MLRegalloc/dev-rel-equivalence.ll
@@ -11,7 +11,8 @@
; RUN: llc -mtriple=x86_64-linux-unknown -regalloc=greedy -regalloc-enable-advisor=release \
; RUN: %S/Inputs/input.ll -o %t.release
-; RUN: rm -rf %t && mkdir %t
+; RUN: rm -rf %t %t_savedmodel
+; RUN: %python %S/../../../../lib/Analysis/models/saved-model-to-tflite.py %t_savedmodel %t
; RUN: %python %S/../../../lib/Analysis/models/gen-regalloc-eviction-test-model.py %t
; RUN: llc -mtriple=x86_64-linux-unknown -regalloc=greedy -regalloc-enable-advisor=development \
; RUN: -regalloc-model=%t %S/Inputs/input.ll -o %t.development
diff --git a/llvm/test/Transforms/Inline/ML/bounds-checks-rewards.ll b/llvm/test/Transforms/Inline/ML/bounds-checks-rewards.ll
index 48c2e0a301a94..d272bb0f7a838 100644
--- a/llvm/test/Transforms/Inline/ML/bounds-checks-rewards.ll
+++ b/llvm/test/Transforms/Inline/ML/bounds-checks-rewards.ll
@@ -8,7 +8,9 @@
;
; Generate mock model
; RUN: rm -rf %t
-; RUN: %python %S/../../../../lib/Analysis/models/gen-inline-oz-test-model.py %t
+; RUN: rm -rf %t_savedmodel
+; RUN: %python %S/../../../../lib/Analysis/models/gen-inline-oz-test-model.py %t_savedmodel
+; RUN: %python %S/../../../../lib/Analysis/models/saved-model-to-tflite.py %t_savedmodel %t
;
; When the bounds are very wide ("no bounds"), all inlinings happen.
; RUN: opt -passes=scc-oz-module-inliner -ml-inliner-ir2native-model=%S/../../../../unittests/Analysis/Inputs/ir2native_x86_64_model -ml-inliner-model-under-training=%t -training-log=- -tfutils-text-log -enable-ml-inliner=development -ml-advisor-size-increase-threshold=10.0 -S < %s 2>&1 | FileCheck %s --check-prefix=CHECK --check-prefix=NOBOUNDS
diff --git a/llvm/test/Transforms/Inline/ML/development-training-log.ll b/llvm/test/Transforms/Inline/ML/development-training-log.ll
index 7d3b592075828..a0b8d730a19b3 100644
--- a/llvm/test/Transforms/Inline/ML/development-training-log.ll
+++ b/llvm/test/Transforms/Inline/ML/development-training-log.ll
@@ -1,8 +1,9 @@
; Test that we can produce a log if we have or do not have a model, in development mode.
; REQUIRES: have_tf_api
; Generate mock model
-; RUN: rm -rf %t
-; RUN: %python %S/../../../../lib/Analysis/models/gen-inline-oz-test-model.py %t
+; RUN: rm -rf %t_savedmodel %t
+; RUN: %python %S/../../../../lib/Analysis/models/gen-inline-oz-test-model.py %t_savedmodel
+; RUN: %python %S/../../../../lib/Analysis/models/saved-model-to-tflite.py %t_savedmodel %t
;
; RUN: opt -enable-ml-inliner=development -passes=scc-oz-module-inliner -training-log=- -tfutils-text-log -ml-inliner-model-under-training=%t -ml-inliner-ir2native-model=%S/../../../../unittests/Analysis/Inputs/ir2native_x86_64_model -S < %s | FileCheck %s
; RUN: opt -enable-ml-inliner=development -passes=scc-oz-module-inliner -training-log=- -tfutils-text-log -ml-inliner-model-under-training=%t -ml-inliner-ir2native-model=%S/../../../../unittests/Analysis/Inputs/ir2native_x86_64_model -ml-inliner-output-spec-override=%S/Inputs/test_output_spec.json -S < %s | FileCheck %s --check-prefixes=EXTRA-OUTPUTS,CHECK
diff --git a/llvm/test/Transforms/Inline/ML/ml-test-development-mode.ll b/llvm/test/Transforms/Inline/ML/ml-test-development-mode.ll
index d902b9e4c778f..3bfb9b628949f 100644
--- a/llvm/test/Transforms/Inline/ML/ml-test-development-mode.ll
+++ b/llvm/test/Transforms/Inline/ML/ml-test-development-mode.ll
@@ -6,7 +6,9 @@
; for the 'release' mode.
;
; REQUIRES: have_tf_api
-; RUN: rm -rf %t && mkdir %t
-; RUN: %python %S/../../../../lib/Analysis/models/gen-inline-oz-test-model.py %t
+; RUN: rm -rf %t
+; RUN: rm -rf %t_savedmodel
+; RUN: %python %S/../../../../lib/Analysis/models/gen-inline-oz-test-model.py %t_savedmodel
+; RUN: %python %S/../../../../lib/Analysis/models/saved-model-to-tflite.py %t_savedmodel %t
; RUN: opt -passes=scc-oz-module-inliner -enable-ml-inliner=default -S < %S/Inputs/test-module.ll 2>&1 | FileCheck %S/Inputs/test-module.ll --check-prefix=DEFAULT
; RUN: opt -passes=scc-oz-module-inliner -enable-ml-inliner=development -ml-inliner-model-under-training=%t -S < %S/Inputs/test-module.ll 2>&1 | FileCheck %S/Inputs/test-module.ll --check-prefix=CHECK
diff --git a/llvm/unittests/Analysis/Inputs/ir2native_x86_64_model/model.tflite b/llvm/unittests/Analysis/Inputs/ir2native_x86_64_model/model.tflite
new file mode 100644
index 0000000000000..7d4097256911b
Binary files /dev/null and b/llvm/unittests/Analysis/Inputs/ir2native_x86_64_model/model.tflite
diff er
diff --git a/llvm/unittests/Analysis/TFUtilsTest.cpp b/llvm/unittests/Analysis/TFUtilsTest.cpp
index 6d70ab08636ed..fe3b115822bee 100644
--- a/llvm/unittests/Analysis/TFUtilsTest.cpp
+++ b/llvm/unittests/Analysis/TFUtilsTest.cpp
@@ -89,15 +89,6 @@ TEST(TFUtilsTest, EvalError) {
TensorSpec::createSpec<float>("StatefulPartitionedCall", {1})};
TFModelEvaluator Evaluator(getModelPath(), InputSpecs, OutputSpecs);
- EXPECT_TRUE(Evaluator.isValid());
-
- int32_t *V = Evaluator.getInput<int32_t>(0);
- // Fill it up with 1's, we know the output.
- for (auto I = 0; I < KnownSize; ++I) {
- V[I] = 1;
- }
- auto ER = Evaluator.evaluate();
- EXPECT_FALSE(ER.hasValue());
EXPECT_FALSE(Evaluator.isValid());
}
diff --git a/utils/bazel/llvm_configs/llvm-config.h.cmake b/utils/bazel/llvm_configs/llvm-config.h.cmake
index 96fbe6f771d2a..b289db1d46ec4 100644
--- a/utils/bazel/llvm_configs/llvm-config.h.cmake
+++ b/utils/bazel/llvm_configs/llvm-config.h.cmake
@@ -101,6 +101,9 @@
/* Define if LLVM was built with a dependency to the libtensorflow dynamic library */
#cmakedefine LLVM_HAVE_TF_API
+/* Define if LLVM is using tflite instead of libtensorflow */
+#cmakedefine LLVM_HAVE_TFLITE
+
/* Define to 1 if you have the <sysexits.h> header file. */
#cmakedefine HAVE_SYSEXITS_H ${HAVE_SYSEXITS_H}
More information about the llvm-commits
mailing list