Add support for IR2Vec

[ChrisCummins]

Moving @anilavakundu's PR (ChrisCummins#3) to the main repo. This supersedes #560.

Adds IR2Vec (https://github.com/IITH-Compilers/IR2Vec) an observation space for compiler gym.

Fixes #449.

[ChrisCummins]

Hi @anilavakundu, thanks for pushing on this! I've left some comments inline.

The one thing that is missing is this code needs tests. Take a look at the tests/llvm/observation_spaces_test.py file. Each observation space needs a test function. At a minimum, you will want to test these properties for each of the four new observation spaces:

• env.observation.spaces["Ir2VecFlowSensitive"].space.dtype
• env.observation.spaces["Ir2VecFlowSensitive"].space.shape
• env.observation.spaces["Ir2VecFlowSensitive"].space.contains()
• env.observation.spaces["Ir2VecFlowSensitive"].deterministic
• env.observation.spaces["Ir2VecFlowSensitive"].platform_dependent

You can use the existing tests (like test_inst2vec_observation_space) as a starting point. To run the tests locally, see INSTALL.md.

Once you've done that, let me know. There's still a couple more things that need doing before merging (adding CMake support, rebasing on top of latest development branch) but I'm happy to do those for you.

Thanks again!

Cheers,
Chris

[ChrisCummins]

I'm not sure about this. This code says that values are in the range [0,∞], but when I run your code I see plenty of negative values:

>>> env.observation["Ir2vecFa"]
array([ 16.67847493,  -8.76068458, -49.90505585,  12.22742195,
        15.365804  , -10.31495722, -42.74864244, -10.26059285,
       -44.02249729,  23.81014121,  29.06372466,  24.08734658,
       -14.95525244,  19.8942756 , -29.91043964,  10.30582115,
       -24.02354845,   0.10980253,  -1.66427926,  14.17916835,
       -34.78192827,  37.14407874,  -8.33318256,  -3.45480279,
       -16.80741089, -32.45384884,  45.50566991,  37.82753753,
       -49.07060102,  -8.93597257, -52.5364784 ,   1.33546551,
       -12.41253508,  29.89899298,  10.97634208,  10.21049925,
        31.45356546,  16.61958681,  13.0980088 ,  -8.284721  ,

What are the bounds for embedding values?

[anilavakundu]

It seems that ScalarRange wasn't a good fit for the shape of the embeddings as the range is not really bounded. I switched this to be of a Sequence type and fixed the length of the sequence type to be of 300 for both max & min. Can you please check the new code?

[ChrisCummins]

What is the shape of the two non-function-level spaces? Is it a single 300 dimension vector? Or a list of 300 dimension vectors?

[anilavakundu]

It's a single 300 dimension vector

[ChrisCummins]

OKay, it should be an int64_list then. You can copy the Autophase space and adjust the dimensionality and limits:

CompilerGym/compiler_gym/envs/llvm/service/ObservationSpaces.cc

Lines 81 to 94 in 1553e1f

	ScalarRange featureSize;
	featureSize.mutable_min()->set_value(0);
	std::vector<ScalarRange> featureSizes;
	featureSizes.reserve(kAutophaseFeatureDim);
	for (size_t i = 0; i < kAutophaseFeatureDim; ++i) {
	featureSizes.push_back(featureSize);
	}
	*space.mutable_int64_range_list()->mutable_range() = {featureSizes.begin(),
	featureSizes.end()};
	space.set_deterministic(true);
	space.set_platform_dependent(false);
	std::vector<int64_t> defaultValue(kAutophaseFeatureDim, 0);
	*space.mutable_default_value()->mutable_int64_list()->mutable_value() = {
	defaultValue.begin(), defaultValue.end()};

If there is no lower bound, remove this line:

        featureSize.mutable_min()->set_value(0);

[anilavakundu]

OKay, it should be an int64_list then. You can copy the Autophase space and adjust the dimensionality and limits:

CompilerGym/compiler_gym/envs/llvm/service/ObservationSpaces.cc

Lines 81 to 94 in 1553e1f

	ScalarRange featureSize;
	featureSize.mutable_min()->set_value(0);
	std::vector<ScalarRange> featureSizes;
	featureSizes.reserve(kAutophaseFeatureDim);
	for (size_t i = 0; i < kAutophaseFeatureDim; ++i) {
	featureSizes.push_back(featureSize);
	}
	*space.mutable_int64_range_list()->mutable_range() = {featureSizes.begin(),
	featureSizes.end()};
	space.set_deterministic(true);
	space.set_platform_dependent(false);
	std::vector<int64_t> defaultValue(kAutophaseFeatureDim, 0);
	*space.mutable_default_value()->mutable_int64_list()->mutable_value() = {
	defaultValue.begin(), defaultValue.end()};

If there is no lower bound, remove this line:

        featureSize.mutable_min()->set_value(0);

Did you mean a double_list ? The values for the embeddings are floating-point numbers

[ChrisCummins]

Ah yes, of course, sorry :)

[ChrisCummins]

Bouncing this back over to #560 now that I've updated onto the new proto schema!