Fix bug introduced by #259

PR #259 switched from TC function names to using whole TCs
(canonicalized) as part of compilation cache keys. However, the final
step of the autotuner, that is retrieving the best found options, was
not updated and it used just the function name. As a result, an empty
optional<CudaMappingOptions> was always returned.

Author: Theodoros Theodoridis

include/tc/autotuner/utils/utils.h

@@ -21,6 +21,7 @@
 #include "tc/core/cuda/cuda.h"
 #include "tc/core/cuda/cuda_mapping_options.h"
 #include "tc/core/utils/dlpack.h"
+#include "tc/lang/tree.h"
 
 #include <llvm/ADT/Optional.h>
 
@@ -59,6 +60,8 @@ llvm::Optional<CudaMappingOptions> getBestOptions(
     const std::vector<const DLTensor*>& inputs,
     const std::vector<const DLTensor*>& outputs);
 
+std::string canonicalTC(const lang::TreeRef& tc);
+
 } // namespace autotune
 } // namespace tc
 

src/autotuner/genetic_autotuner.cc

@@ -186,7 +186,7 @@ llvm::Optional<CudaMappingOptions> GeneticAutotuner::tune(
 
   CHECK_GT(inputs.size(), 0);
   return tc::autotune::getBestOptions(
-      tcName, inputs.begin()->second, outputPtrs);
+      canonicalTC(tcNameMap_.at(tcName)), inputs.begin()->second, outputPtrs);
 }
 
 } // namespace detail

src/autotuner/utils/utils.cc

@@ -23,7 +23,6 @@
 #include "tc/lang/canonicalize.h"
 #include "tc/lang/parser.h"
 #include "tc/lang/sema.h"
-#include "tc/lang/tree.h"
 
 namespace tc {
 namespace autotune {
@@ -74,20 +73,18 @@ std::vector<OptionsWithMedianTime> getOptionsAndMedianRuntimes(
   return c;
 }
 
-namespace {
 std::string canonicalTC(const lang::TreeRef& tc) {
   std::stringstream ss;
-  ss << lang::canonicalize(tc);
+  ss << lang::canonicalize(lang::Sema().checkFunction(tc));
   return ss.str();
 }
-} // namespace
 
 std::vector<CudaMappingOptions> restoreCandidates(
     const lang::TreeRef& tc,
     const std::vector<const DLTensor*>& inputs,
     const std::vector<const DLTensor*>& outputs) {
-  auto candidates = getOptionsAndMedianRuntimes(
-      canonicalTC(lang::Sema().checkFunction(tc)), inputs, outputs);
+  auto candidates =
+      getOptionsAndMedianRuntimes(canonicalTC(tc), inputs, outputs);
   LOG_IF(INFO, candidates.size() < FLAGS_tuner_gen_restore_number)
       << "Requested " << FLAGS_tuner_gen_restore_number
       << " candidates but there are only " << candidates.size() << " in cache.";