fix: Update "reduceAxes" variable in GlobalPoolingConverter function …

…and add corresponding uTests

Signed-off-by: Ruoqian Guo <ruoqiang@nvidia.com>

core/conversion/converters/impl/pooling.cpp

@@ -16,8 +16,9 @@ bool GlobalPoolingConverter(
     nvinfer1::PoolingType pool_type) {
   auto in = args[0].ITensorOrFreeze(ctx);
   nvinfer1::Dims dims = in->getDimensions();
+  auto out_size = util::toDims(args[1].unwrapToIntList());
   // Generate a bitmask of all 1s except the last 2 bits (N and C axes)
-  uint32_t reduceAxes = ((1 << dims.nbDims) - 1) & ~0b11;
+  uint32_t reduceAxes = ((1 << dims.nbDims) - 1) ^ ((1 << (dims.nbDims - out_size.nbDims)) - 1);
   auto* new_layer = ctx->net->addReduce(
       *in,
       pool_type == nvinfer1::PoolingType::kMAX ? nvinfer1::ReduceOperation::kMAX : nvinfer1::ReduceOperation::kAVG,

tests/core/conversion/converters/test_pooling.cpp

@@ -436,6 +436,32 @@ TEST(Converters, ATenAdaptiveAvgPool2DConvertsCorrectly) {
   ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
 }
 
+TEST(Converters, ATenAdaptiveAvgPool2DGlobalPoolingConvertsCorrectly) {
+  const auto graph = R"IR(
+      graph(%0 : Tensor):
+        %2 : int = prim::Constant[value=1]()
+        %3 : int = prim::Constant[value=1]()
+        %6 : int[] = prim::ListConstruct(%2, %3)
+        %10 : Tensor = aten::adaptive_avg_pool2d(%0, %6)
+        return (%10))IR";
+
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(graph, g.get());
+
+  // PyTorch PyTorch adaptive_avg_pool2d needs a 4D input or a 3D input
+  auto in = at::randint(-5, 5, {64, 16, 32, 32}, at::kCUDA);
+
+  auto jit_in = at::clone(in);
+  auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_in});
+
+  auto trt_in = at::clone(in);
+  params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {trt_in});
+
+  ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
+}
+
 TEST(Converters, ATenAdaptiveAvgPool2DConvertsCorrectlyWithDynamicInput) {
   const auto graph = R"IR(
       graph(%0 : Tensor):
@@ -488,6 +514,32 @@ TEST(Converters, ATenAdaptiveAvgPool1DConvertsCorrectly) {
   ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 1.0));
 }
 
+TEST(Converters, ATenAdaptiveAvgPool1DGlobalPoolingConvertsCorrectly) {
+  const auto graph =
+      R"IR(
+      graph(%0 : Tensor):
+        %2 : int = prim::Constant[value=1]()
+        %6 : int[] = prim::ListConstruct(%2)
+        %10 : Tensor = aten::adaptive_avg_pool1d(%0, %6)
+        return (%10))IR";
+
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(graph, g.get());
+
+  // PyTorch adaptive_avg_pool1d needs a 3D input or a 2D input
+  auto in = at::randint(-5, 5, {3, 16}, at::kCUDA);
+
+  auto jit_in = at::clone(in);
+  auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_in});
+
+  auto trt_in = at::clone(in);
+  params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {trt_in});
+
+  ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
+}
+
 TEST(Converters, ATenAdaptiveMaxPool2DConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor):