make sharedDepth mapping option tunable

Introduce a tunable parameter for sharedDepth added by a preceding
commit.  Select the initial range as [0,7] because the range of
privateDepth goes until 10, and 3 innermost loops may be mapped to
threads making it impossible to promote below them.

Author: ftynse

tc/autotuner/autotuner-inl.h

@@ -374,6 +374,8 @@ void setupTuningParameters(
       RangeParameter(powers2(FLAGS_tuner_max_unroll_size), "unroll");
   configuration.privateDepth =
       RangeParameter({0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10}, "pdepth");
+  configuration.sharedDepth =
+      RangeParameter({0, 1, 2, 3, 4, 5, 6, 7}, "sdepth");
 }
 } // namespace
 

tc/autotuner/parameters.cc

@@ -240,6 +240,7 @@ void TuningConfiguration::applyToParameters(
   useReadOnlyCache.apply(f);
   matchLibraryCalls.apply(f);
   privateDepth.apply(f);
+  sharedDepth.apply(f);
 }
 
 bool TuningConfiguration::isValid() const {
@@ -275,6 +276,7 @@ std::vector<ParameterView> TuningConfiguration::collectParameters() {
   params.emplace_back(useReadOnlyCache);
   params.emplace_back(matchLibraryCalls);
   params.emplace_back(privateDepth);
+  params.emplace_back(sharedDepth);
 
   return params;
 }
@@ -306,6 +308,7 @@ void TuningConfiguration::fromCudaMappingOptions(
   unrollCopyShared.selectValue(options.proto().unroll_copy_shared());
   useReadOnlyCache.selectValue(options.proto().use_readonly_cache());
   privateDepth.selectFromValue(options.proto().private_depth());
+  sharedDepth.selectFromValue(options.proto().shared_depth());
 }
 
 void TuningConfiguration::fromCpuMappingOptions(
@@ -335,6 +338,7 @@ void TuningConfiguration::applyToCudaMappingOptions(
   options.unrollCopyShared(unrollCopyShared.value());
   options.useReadOnlyCache(useReadOnlyCache.value());
   options.privateDepth(privateDepth.value());
+  options.sharedDepth(sharedDepth.value());
 }
 
 void TuningConfiguration::applyToCpuMappingOptions(

tc/autotuner/parameters.h

@@ -191,6 +191,7 @@ class TuningConfiguration {
   BoolParameter useReadOnlyCache;
   BoolParameter matchLibraryCalls;
   RangeParameter privateDepth;
+  RangeParameter sharedDepth;
 
  private:
   std::vector<std::function<bool(const TuningConfiguration&)>> validators_;
@@ -229,6 +230,7 @@ class TuningParameterFixer {
   llvm::Optional<bool> useReadOnlyCache;
   llvm::Optional<bool> matchLibraryCalls;
   llvm::Optional<uint32_t> privateDepth;
+  llvm::Optional<uint32_t> sharedDepth;
 
   friend class TuningConfiguration;
 };