Mark negative indices support for gather as optional (#681)

Signed-off-by: Kevin Chen <[email protected]>

CMakeLists.txt

@@ -74,6 +74,11 @@ if(BUILD_ONNXIFI)
   set(ONNXIFI_SOURCES onnx_trt_backend.cpp)
 endif()
 
+# Build with negative indices support for Gather:
+if (DEFINED SUPPORT_NEGATIVE_GATHER)
+  add_definitions("-DSUPPORT_NEGATIVE_GATHER=1")
+endif()
+
 # Build executables if BUILD_LIBRARY_ONLY flag is not set
 if (NOT DEFINED BUILD_LIBRARY_ONLY)
   set(EXECUTABLE_SOURCES

README.md

@@ -61,7 +61,7 @@ Once you have cloned the repository, you can build the parser libraries and exec
     // Ensure that you update your LD_LIBRARY_PATH to pick up the location of the newly built library:
     export LD_LIBRARY_PATH=$PWD:$LD_LIBRARY_PATH
 
-For building only the libraries, append `-DBUILD_LIBRARY_ONLY=1` to the CMake build command.
+For building only the libraries, append `-DBUILD_LIBRARY_ONLY=1` to the CMake build command. If your model has Gather or GatherElements operations with negative indices, add `-DSUPPORT_NEGATIVE_GATHER` to the build command. Note that enabling negative-indices gather will have a performance impact on gathers with non-negative indices.
 
 ## Executable Usage
 

builtin_op_importers.cpp

@@ -1201,8 +1201,11 @@ DEFINE_BUILTIN_OP_IMPORTER(Gather)
     TRT_CHECK(convertAxis(axis, nbDims));
     LOG_VERBOSE("Using Gather axis: " << axis);
 
-    // Convert any negative indices to positive ones
+    // Support for negative indices can be enabled through adding -DSUPPORT_NEGATIVE_GATHER=1 in the CMake build command.
+    // This will unnecessarily reduce performance of networks that use only non-negative Gather indices.
+#if SUPPORT_NEGATIVE_GATHER
     indices = convertGatherIndices(ctx, data, indices, axis);
+#endif // SUPPORT_NEGATIVE_GATHER
 
     auto* layer = ctx->network()->addGather(*data, *indices, axis);
     ctx->registerLayer(layer, getNodeName(node));
@@ -1251,8 +1254,11 @@ DEFINE_BUILTIN_OP_IMPORTER(GatherElements)
     int32_t axis = attrs.get<int32_t>("axis", 0);
     int32_t dataNbDims = daDims.nbDims;
 
-    // Convert any negative indices to positive ones
+    // Support for negative indices can be enabled through adding -DSUPPORT_NEGATIVE_GATHER=1 in the CMake build command.
+    // This will unnecessarily reduce performance of networks that use only non-negative Gather indices.
+#if SUPPORT_NEGATIVE_GATHER
     index = convertGatherIndices(ctx, data, index, axis);
+#endif // SUPPORT_NEGATIVE_GATHER
 
     TRT_CHECK(convertAxis(axis, dataNbDims));
     LOG_VERBOSE("Using Gather axis: " << axis);

onnx2trt_utils.cpp

@@ -425,15 +425,21 @@ nvinfer1::ITensor* createZeroTensor(IImporterContext* ctx, nvinfer1::ITensor* da
 
 nvinfer1::ITensor* convertGatherIndices(IImporterContext* ctx, nvinfer1::ITensor* data, nvinfer1::ITensor* indices, int32_t axis)
 {
-    // Create a condition tensor that is 1 for the elements in indices that are < 0 or 0 otherwise
-    auto condition = ctx->network()->addElementWise(*indices, *createZeroTensor(ctx, indices), nvinfer1::ElementWiseOperation::kLESS)->getOutput(0);
+    const int32_t n = indices->getDimensions().nbDims;
     auto axisLength = getAxisLength(ctx, data, axis);
-    broadcastTensors(ctx, axisLength, indices);
-    // Create a shifted tensor that is indices + axisLength
-    auto shifted = ctx->network()->addElementWise(*indices, *axisLength, nvinfer1::ElementWiseOperation::kSUM)->getOutput(0);
-    // Select between the shifted and original data based on condition
-    auto select = ctx->network()->addSelect(*condition, *shifted, *indices);
-    return select->getOutput(0);
+    broadcastTensor(ctx, axisLength, n);
+
+    // The formula here implements "indices < 0 ? indices + axisLength : indices"
+    // via the formula "indices - axisLength * max(-1, min(0, indices))".
+    // Think of the "max(-1, min(0, indices))" as extracting the sign bit from the indices.
+    const nvinfer1::Dims d = makeDims(n, 1);
+    auto zero = addConstantScalar(ctx, 0, ::ONNX_NAMESPACE::TensorProto::INT32, d)->getOutput(0);
+    auto minusOne = addConstantScalar(ctx, -1, ::ONNX_NAMESPACE::TensorProto::INT32, d)->getOutput(0);
+    auto min = ctx->network()->addElementWise(*zero, *indices, nvinfer1::ElementWiseOperation::kMIN)->getOutput(0);
+    auto max = ctx->network()->addElementWise(*minusOne, *min, nvinfer1::ElementWiseOperation::kMAX)->getOutput(0);
+    auto prod = ctx->network()->addElementWise(*max, *axisLength, nvinfer1::ElementWiseOperation::kPROD)->getOutput(0);
+    auto sub = ctx->network()->addElementWise(*indices, *prod, nvinfer1::ElementWiseOperation::kSUB)->getOutput(0);
+    return sub;
 }
 
 template <typename DataType>