Add Kronecker benchmark

This commit adds a Kronecker decomposition benchmark and saves the best
options found by the tuner for future reproducibility.
Since Kronecker Recurrent Units are meant to replace large FC layers we
compare against that baseline.

Author: nicolasvasilache

.jenkins/build.sh

@@ -69,5 +69,8 @@ WITH_CAFFE2=ON CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda CLANG_PREFIX=$(${CONDA_PREF
 python setup.py install
 ./test_python/run_test.sh
 
-FILTER_OUT=MLP_model ./test.sh
+FILTER_OUT="MLP_model kronecker" ./test.sh
+# 2LUT can OOM on smaller Maxwells on our CI machines
 ./build/tc/benchmarks/MLP_model --gtest_filter=-*2LUT*
+# Kronecker xxxAsMatMul can OOM
+./build/tc/benchmarks/kronecker --gtest_filter=-*AsMatMul*

tc/benchmarks/CMakeLists.txt

@@ -18,6 +18,7 @@ find_library(CUDA_CUDNN_LIBRARIES cudnn
 set(EXAMPLES_FILES
   batchmatmul
   group_convolution
+  kronecker
   tmm
   MLP_model
 )

tc/benchmarks/benchmark_fixture.h

@@ -96,8 +96,10 @@ struct Benchmark : public ::testing::Test {
     }
   }
 
-  template <typename CheckFunction>
-  void Check(
+  using CheckFunction = std::function<bool(
+      const std::vector<at::Tensor>& inputs,
+      const std::vector<at::Tensor>& outputs)>;
+  std::vector<at::Tensor> Check(
       const std::string& tc,
       const std::string& name,
       const tc::CudaMappingOptions& mappingOptions,
@@ -184,6 +186,8 @@ struct Benchmark : public ::testing::Test {
     std::cout << "\n\n";
 
 #undef GET_US
+
+    return outputs;
   }
 
   template <typename InitFunction, typename InplaceFunction>
@@ -230,7 +234,6 @@ struct Benchmark : public ::testing::Test {
 #undef GET_US
   }
 
-  template <typename CheckFunction>
   void validateProto(
       std::string cacheFilename,
       const std::string& tc,
@@ -342,8 +345,7 @@ struct Benchmark : public ::testing::Test {
 #undef GET_US
   }
 
-  template <typename CheckFunction>
-  void autotune(
+  std::vector<tc::CudaMappingOptions> autotune(
       std::string cacheFilename,
       std::string resultsFilename,
       std::string tc,
@@ -442,6 +444,10 @@ struct Benchmark : public ::testing::Test {
           << "\n---------------------------------------------------------";
       std::cout << "\n\n";
 #undef GET_US
+
+      return {bestOptions};
     }
+
+    return {};
   }
 };