creating a C++ API for triu

NVIDIA · Dec 23, 2024 · 5f82c65 · 5f82c65
1 parent e92c575
commit 5f82c65
Show file tree

Hide file tree

Showing 2 changed files with 38 additions and 15 deletions.
diff --git a/csrc/ops/arith.cpp b/csrc/ops/arith.cpp
@@ -504,14 +504,30 @@ TensorView* triu(TensorView* tv, Val* offset) {
       tv->nDims(),
       "D tensor");
 
+  // Let's say we want a triu of a 2D tensor of shape [2, 4]
+  // We broadcast the iota of the outer dim
+  // [0    [0, 0, 0, 0]
+  // 1] -> [1, 1, 1, 1]
+  // We broadcast the iota of the inner dim
+  // [0, 1, 2, 3] -> [0, 1, 2, 3]
+  //                 [0, 1, 2, 3]
+  // Using LE on the bcast tensors we get the mask
+  //[0, 0, 0, 0]  LE [0, 1, 2, 3]
+  //[1, 1, 1, 1]     [0, 1, 2, 3]
+  // Gives:
+  //[1, 0, 0, 0]
+  //[0, 1, 0, 0]
+  // If triu has an offset of k, we shift/subtract the iota of the columns by k
+  // before broadcasting and comparing with the iota of the rows.
+  auto dims = tv->domain()->logical().size();
   auto tv_rows = iota(
-      tv->domain()->logical()[1]->extent(),
+      tv->domain()->logical()[dims - 2]->extent(),
       IrBuilder::create<Val>(0, DataType::Index),
       IrBuilder::create<Val>(1, DataType::Index),
       DataType::Int);
 
   auto tv_columns = iota(
-      tv->domain()->logical()[2]->extent(),
+      tv->domain()->logical()[dims - 1]->extent(),
       SimplifyingIrBuilder::mulExpr(
           offset, IrBuilder::create<Val>(-1, DataType::Index)),
       IrBuilder::create<Val>(1, DataType::Index),

diff --git a/tests/cpp/test_tensor_factories.cpp b/tests/cpp/test_tensor_factories.cpp
@@ -231,25 +231,32 @@ TEST_F(TensorFactoryTest, StandaloneIota) {
 }
 
 TEST_F(TensorFactoryTest, SimpleTriu) {
-  auto fusion = std::make_unique<Fusion>();
-
-  FusionGuard gf(fusion.get());
+  std::vector<std::vector<int64_t>> input_sizes = {
+      {64, 64}, {4, 16}, {16, 4}, {16, 8, 32}};
+  auto offsets = {0, 1, 2, -1, -2, 200, -200};
 
-  auto tv_to_triu_on = makeSymbolicTensor(3, DataType::Half);
-  fusion->addInput(tv_to_triu_on);
+  for (auto input_size : input_sizes) {
+    for (auto offset : offsets) {
+      auto fusion = std::make_unique<Fusion>();
+      FusionGuard fg(fusion.get());
 
-  int64_t k_factor = -2;
-  auto out = triu(tv_to_triu_on, IrBuilder::create<Val>(k_factor, DataType::Int));
-  fusion->addOutput(out);
+      auto tv_to_triu_on =
+          makeSymbolicTensor(input_size.size(), DataType::Half);
+      fusion->addInput(tv_to_triu_on);
 
+      auto out =
+          triu(tv_to_triu_on, IrBuilder::create<Val>(offset, DataType::Int));
+      fusion->addOutput(out);
 
-  auto options = at::TensorOptions().dtype(at::kHalf).device(at::kCUDA);
-  auto in_tensor = at::randn({4, 4, 8}, options);
+      auto options = at::TensorOptions().dtype(at::kHalf).device(at::kCUDA);
+      auto in_tensor = at::randn(input_size, options);
 
-  FusionExecutorCache executor_cache(std::move(fusion));
-  auto cg_outputs = executor_cache.runFusionWithInputs({in_tensor});
+      FusionExecutorCache executor_cache(std::move(fusion));
+      auto cg_outputs = executor_cache.runFusionWithInputs({in_tensor});
 
-  EXPECT_TRUE(cg_outputs[0].allclose(at::triu(in_tensor, k_factor), .001, .001));
+      EXPECT_TRUE(at::equal(cg_outputs[0], at::triu(in_tensor, offset)));
+    }
+  }
 }
 
 TEST_F(TensorFactoryTest, StandaloneARange) {