Merge pull request #302 from NVIDIA/topk

Adding support for aten::topk

Author: narendasan

core/conversion/converters/BUILD

@@ -53,6 +53,7 @@ cc_library(
         "impl/stack.cpp",
         "impl/lstm_cell.cpp",
         "impl/unsqueeze.cpp",
+        "impl/topk.cpp",
     ],
     deps = [
         "@tensorrt//:nvinfer",

core/conversion/converters/impl/topk.cpp

@@ -0,0 +1,62 @@
+#include "NvInfer.h"
+#include "core/conversion/converters/converters.h"
+#include "core/conversion/tensorcontainer/TensorContainer.h"
+#include "core/util/prelude.h"
+#include "torch/torch.h"
+
+#include <ATen/ATen.h>
+#include <vector>
+
+namespace trtorch {
+namespace core {
+namespace conversion {
+namespace converters {
+namespace impl {
+namespace {
+
+auto topk_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns().pattern(
+    {"aten::topk(Tensor self, int k, int dim=-1, bool largest=True, bool sorted=True) -> (Tensor values, Tensor indices)",
+     [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+       auto self = args[0].ITensorOrFreeze(ctx);
+       auto k = args[1].unwrapToInt();
+       auto dim = args[2].unwrapToInt();
+       auto largest = args[3].unwrapToBool();
+       LOG_DEBUG(
+           "Note: sorted argument is not used in TensorRT for aten::topk, results will depend on the value of largest");
+       // auto sorted = args[4].unwrapToBool(); # Currently unused
+
+       auto selfDim = util::toVec(self->getDimensions());
+
+       // reduceAxes	The reduction dimensions. The bit in position i of bitmask reduceAxes corresponds to explicit
+       // dimension i of the result. E.g., the least significant bit corresponds to the first explicit dimension and the
+       // next to least significant bit corresponds to the second explicit dimension.
+
+       if (dim < 0) {
+         dim = selfDim.size() + dim;
+       }
+
+       uint32_t shiftDim = 1 << dim;
+
+       LOG_DEBUG("Output topk reduce dim: " << dim);
+
+       auto TopKOperation = largest ? (nvinfer1::TopKOperation::kMAX) : (nvinfer1::TopKOperation::kMIN);
+
+       auto new_layer = ctx->net->addTopK(*self, TopKOperation, k, shiftDim);
+
+       TRTORCH_CHECK(new_layer, "Unable to create topk layer from node: " << *n);
+
+       auto out0 = ctx->AssociateValueAndTensor(n->outputs()[0], new_layer->getOutput(0));
+       auto out1 = ctx->AssociateValueAndTensor(n->outputs()[1], new_layer->getOutput(1));
+
+       LOG_DEBUG("Output tensor(0) shape: " << out0->getDimensions());
+       LOG_DEBUG("Output tensor(1) shape: " << out1->getDimensions());
+
+       return true;
+     }});
+
+} // namespace
+} // namespace impl
+} // namespace converters
+} // namespace conversion
+} // namespace core
+} // namespace trtorch

tests/core/conversion/converters/BUILD

@@ -71,6 +71,10 @@ converter_test(
   name = "test_stack"
 )
 
+converter_test(
+  name = "test_topk"
+)
+
 converter_test(
   name = "test_lstm_cell"
 )
@@ -103,6 +107,7 @@ test_suite(
     ":test_stack",
     ":test_lstm_cell",
     ":test_unsqueeze",
-    ":test_squeeze"
+    ":test_squeeze",
+    ":test_topk",
   ]
 )

tests/core/conversion/converters/test_topk.cpp

@@ -0,0 +1,30 @@
+#include <string>
+#include "core/compiler.h"
+#include "gtest/gtest.h"
+#include "tests/util/util.h"
+#include "torch/csrc/jit/ir/irparser.h"
+
+TEST(Converters, ATenTopKConvertsCorrectly) {
+  const auto graph = R"IR(
+      graph(%0 : Tensor):
+        %1 : int = prim::Constant[value=20]()
+        %2 : int = prim::Constant[value=-1]()
+        %3 : bool = prim::Constant[value=1]()
+        %4 : bool = prim::Constant[value=1]()
+        %5 : Tensor, %6 : Tensor = aten::topk(%0, %1, %2, %3, %4)
+        return (%5, %6))IR";
+
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(graph, &*g);
+
+  auto in = at::rand({10, 10, 100}, {at::kCUDA});
+
+  auto params = trtorch::core::conversion::get_named_params(g->inputs(), {});
+  auto jit_results = trtorch::tests::util::RunGraph(g, params, {in});
+
+  params = trtorch::core::conversion::get_named_params(g->inputs(), {});
+  auto trt_results = trtorch::tests::util::RunGraphEngine(g, params, {in});
+
+  ASSERT_TRUE(trtorch::tests::util::almostEqual(jit_results[0], trt_results[0].reshape_as(jit_results[0]), 2e-6));
+  ASSERT_TRUE(trtorch::tests::util::almostEqual(jit_results[1], trt_results[1].reshape_as(jit_results[1]), 2e-6));
+}