Add more tests for reduction detection

Both cases test reduction detection for a non-parametric 2-dimensionnal mean.
They differ only by the number of mapped threads, and the expected result.
When mapping to more threads than tiling size, some threads are
considered to do nothing, which prevents the reduction detection from
succeeding.  Fixing this would require to perform launch bounds
tightening before reduction detection.

Author: math-fehr

test/test_cuda_mapper.cc

@@ -926,6 +926,78 @@ def fun(float(N, K) I, float(N) O0) -> (O) {
 )TC");
 }
 
+/*
+ * Check that a 2D mean with these parameters does not produce a library call.
+ * The call is not produced because the band is tiled by 32 and 512 threads are
+ * mapped to the band.
+ * In practice, check that the library call does not appear in the code.
+ */
+TEST_F(PolyhedralMapperTest, Mean2DNonParametric_512threads) {
+  string tc = R"TC(
+def fun(float(36864, 1024) I) -> (O) {
+    O(n) +=! I(n, r_n)
+    O(n) = O(n) / (1024)
+}
+)TC";
+  auto mappingOptions =
+      DefaultOptions()
+          .outerScheduleFusionStrategy(tc::FusionStrategy::Preserve3Coincident)
+          .outerScheduleAllowSkewing(false)
+          .outerSchedulePositiveOrthant(true)
+          .intraTileScheduleFusionStrategy(tc::FusionStrategy::Min)
+          .intraTileScheduleAllowSkewing(false)
+          .intraTileSchedulePositiveOrthant(true)
+          .fixParametersBeforeScheduling(false)
+          .tile(18, 32)
+          .unroll(16)
+          .tileImperfectlyNested(false)
+          .matchLibraryCalls(true)
+          .mapToThreads({512})
+          .mapToBlocks({16384})
+          .useSharedMemory(true)
+          .usePrivateMemory(false)
+          .unrollCopyShared(true);
+
+  auto code = codegenMapped(tc, mappingOptions);
+  using tc::code::cuda::kCUBReductionName;
+  EXPECT_TRUE(code.find(kCUBReductionName) == std::string::npos);
+}
+
+/*
+ * Check that a 2D mean with these parameters produce a reduction library call.
+ * In practice, check that the library call appears in the code.
+ */
+TEST_F(PolyhedralMapperTest, Mean2DNonParametric_32threads) {
+  string tc = R"TC(
+def fun(float(36864, 1024) I) -> (O) {
+    O(n) +=! I(n, r_n)
+    O(n) = O(n) / (1024)
+}
+)TC";
+  auto mappingOptions =
+      DefaultOptions()
+          .outerScheduleFusionStrategy(tc::FusionStrategy::Preserve3Coincident)
+          .outerScheduleAllowSkewing(false)
+          .outerSchedulePositiveOrthant(true)
+          .intraTileScheduleFusionStrategy(tc::FusionStrategy::Min)
+          .intraTileScheduleAllowSkewing(false)
+          .intraTileSchedulePositiveOrthant(true)
+          .fixParametersBeforeScheduling(false)
+          .tile(18, 32)
+          .unroll(16)
+          .tileImperfectlyNested(false)
+          .matchLibraryCalls(true)
+          .mapToThreads({32})
+          .mapToBlocks({16384})
+          .useSharedMemory(true)
+          .usePrivateMemory(false)
+          .unrollCopyShared(true);
+
+  auto code = codegenMapped(tc, mappingOptions);
+  using tc::code::cuda::kCUBReductionName;
+  EXPECT_TRUE(code.find(kCUBReductionName) != std::string::npos);
+}
+
 static const string kTcMM = R"TC(
 def fun(float(M, K) A, float(K, N) B) -> (C) {
     C(m, n) +=! A(m, r_k) * B(r_k, n)