This is a collection of e2e tests that save a TensorFlow model, compile it with IREE, run it on multiple backends and crosscheck the results.
You will need a TensorFlow 2.0+ nightly installed in your python environment: the python binary in $PYTHON_BIN should be able to import tensorflow and that TensorFlow should be version 2.0+. This can be checked with tensorflow.version.
See Install TensorFlow with pip for instructions.
If you do not have your environment setup to use IREE with Vulkan (see the doc), then you can run the manual test targets with --target_backends=tf,iree_vmla,iree_llvmjit (that is, by omitting iree_vulkan from the list of backends to run the tests on).
The test suites can be run excluding Vulkan by specifying --test_tag_filters="-driver=vulkan" in the bazel test invocation, or by adding test --test_tag_filters="-driver=vulkan" to your user.bazelrc.
tf.ModulesCompatible TensorFlow modules can be compiled to specific IREE backends using IreeCompiledModule. This also optionally saves compilation artifacts to a specified directory. These artifacts include: MLIR across various lowerings, a TensorFlow SavedModel, and the compiled VM FlatBuffer. A basic example of creating and calling an IreeCompiledModule can be found in tf_utils_test.py
When using Keras models or tf.Modules with functions that IREE can't compile, exported_names should be specified. For example:
from pyiree.tf.support import tf_utils vmla_module = tf_utils.IreeCompiledModule( module_class=KerasTFModuleClass, backend_info=tf_utils.BackendInfo('iree_vmla'), exported_names=['predict']) vmla_module.predict(...)
By default the TensorFlow SavedModels will not be kept. This can be overridden via the --keep_saved_model flag.
For locally running tests and iterating on backend development, bazel run is preferred.
# Run math_test on all backends. bazel run :math_test_manual # Run math_test comparing TensorFlow to itself (e.g. to debug randomization). bazel run :math_test_manual -- target_backends=tf # Run math_test comparing the VMLA backend and TensorFlow. bazel run :math_test_manual -- --target_backends=iree_vmla # Run math_test comparing the VMLA backend to itself multiple times. bazel run :math_test_manual -- \ --reference_backend=iree_vmla --target_backends=iree_vmla,iree_vmla # Run math_test and output on failure. bazel test :math_test_manual --test_output=errors # Run an individual test interactively. bazel run :math_test_manual -- --test_output=streamed
For reproducibility of the unit tests CompiledModule() sets the random seeds of tf, numpy and python by calling tf_utils.set_random_seed() before model creation.
Our tests use a class TracedModule to capture and store all of the inputs and outputs of a CompiledModule in a Trace. Each unittest on a TestCase uses the compare_backends method. This method runs the function it is passed with a TracedModule once for each reference and target backend. The inputs and outputs to these modules are then checked for correctness, using the reference backend as a source of truth. For example:
# Compile a `tf.Module` named `SimpleArithmeticModule` into a `CompiledModule`. @tf_test_utils.compile_module(SimpleArithmeticModule) # Inherit from `TracedModuleTestCase`. class SimpleArithmeticTest(tf_test_utils.TracedModuleTestCase): # Unit test. def test_simple_mul(self): # Trace function. def simple_mul(module): # A random seed is automatically set before each call to `simple_mul`. a = tf_utils.uniform([4]) b = np.array([400., 5., 6., 7.], dtype=np.float32) # The inputs `a` and `b` are recorded along with the output `c` c = module.simple_mul(a, b) # The inputs `a` and `b` are recorded along with the (unnamed) output # module.simple_mul returns. module.simple_mul(a, c) # Calls `simple_mul` once for each backend, recording the inputs and outputs # to `module` and then comparing them. self.compare_backends(simple_mul)
Test targets are automatically generated for each test file and for each backend to check numerical correctness against TensorFlow. Tests targets that pass are placed into the e2e_tests test suite. Tests that fail on particular backends are recorded in lists in the BUILD files. For example, if experimental_new_test.py fails on the iree_llvmjit and iree_vulkan backends then the following lines should be added to the BUILD file:
LLVM_FAILING = [ ... "experimental_new_test.py", ... ] VULKAN_FAILING = [ ... "experimental_new_test.py", ... ]
Test targets for these backends are placed into the e2e_tests_failing test suite. Test targets in these test suites can be run as follows:
# Run all e2e tests that are expected to pass. bazel test :e2e_tests # Run all e2e tests that are expected to fail. bazel test :e2e_tests_failing # Run a specific failing e2e test target. # Note that generated test targets are prefixed with their test suite name. bazel test :e2e_tests_failing_broadcasting_test__tf__iree_vulkan
By default, running an E2E test generates a number of compilation, debugging and benchmarking artifacts in /tmp/iree/modules/. The location of these artifacts can be changed via the --artifacts_dir flag. The generated directory structure for each module is as follows:
/tmp/iree/modules/ModuleName
├── tf_input.mlir # MLIR for ModuleName in TF's input dialect
├── iree_input.mlir # tf_input.mlir translated to IREE MLIR
├── backend_name_1 # e.g. iree_vmla, tf or tf_ref
│ ├── compiled.vmfb # flatbuffer of ModuleName compiled to this backend
│ ├── saved_model # Only created if --keep_saved_model is specified.
│ └── traces
│ ├── trace_1 # Directory storing logs and serialization for each trace.
│ │ └── log.txt # A more detailed version of the test logs
│ └── trace_2
│ └── log.txt
└── backend_name_2
└── ...
Traces for a particular test can be loaded via the Trace.load(trace_dir) method. For example:
ref_trace = Trace.load("/tmp/iree/modules/ModuleName/tf_ref/traces/predict/") tar_trace = Trace.load("/tmp/iree/modules/ModuleName/iree_vmla/traces/predict/") abs_diff = np.abs(ref_trace.calls[0].outputs[0] - tar_trace.calls[0].outputs[0]) print(np.mean(abs_diff))
Traces are named after the trace functions defined in their unittests. So in the SimpleArithmeticModule example above, the trace_dir would be /tmp/iree/modules/SimpleArithmeticModule/iree_vmla/traces/simple_mul/.
Abseil flagfiles containing all of the data that iree-benchmark-module needs to run are generated for each Trace in our E2E tests. This allows for any module we test to be easily benchmarked on valid inputs. The process for benchmarking a vision model can thus be reduced to the following:
# Generate benchmarking artifacts for all vision models: bazel test integrations/tensorflow/e2e/keras:vision_external_tests # Benchmark ResNet50 with cifar10 weights on vmla: bazel run iree/tools:iree-benchmark-module -- \ --flagfile=/tmp/iree/modules/ResNet50/cifar10/iree_vmla/traces/predict/flagfile # Benchmark ResNet50 with cifar10 weights on llvmjit: bazel run iree/tools:iree-benchmark-module -- \ --flagfile=/tmp/iree/modules/ResNet50/cifar10/iree_llvmjit/traces/predict/flagfile
Duplicate flags provided after the flagfile will take precedence. For example:
bazel run iree/tools:iree-benchmark-module -- \ --flagfile=/tmp/iree/modules/ResNet50/cifar10/iree_llvmjit/traces/predict/flagfile \ --input_file=/path/to/custom/compiled.vmfb
Currently, this only supports benchmarking the first module call in a trace. We plan to extend this to support benchmarking all of the calls in the trace, and also plan to support verifying outputs during the warm-up phase of the benchmark.
If the compiler fails to compile the program, then it will create a crash reproducer (see MLIR documentation), which then allows reproducing the bug with an appropriate “opt” tool. Further debugging iteration can happen in opt.
TODO(silvasean): debugging miscompiles
See simple_arithmetic_test.py for some basic examples.