Fix GEMM importer (#828)

Signed-off-by: Kevin Chen <[email protected]>
onnx · Apr 11, 2022 · 5f27e45 · 5f27e45
1 parent 0031a42
commit 5f27e45
Show file tree

Hide file tree

Showing 2 changed files with 7 additions and 114 deletions.
diff --git a/builtin_op_importers.cpp b/builtin_op_importers.cpp
@@ -1523,87 +1523,12 @@ DEFINE_BUILTIN_OP_IMPORTER(Gemm)
     bool transA = attrs.get("transA", false);
     bool transB = attrs.get("transB", false);
     nvinfer1::ITensor& inputA = convertToTensor(inputs.at(0), ctx);
+    nvinfer1::ITensor& inputB = convertToTensor(inputs.at(1), ctx);
     // Validate inputs
-    ASSERT(inputs.at(0).shape().nbDims == 2 && inputs.at(1).shape().nbDims == 2 && "GEMM must have 2D inputs!", ErrorCode::kINVALID_NODE);
+    ASSERT(inputA.getDimensions().nbDims == 2 && inputB.getDimensions().nbDims == 2 && "GEMM must have 2D inputs!", ErrorCode::kINVALID_NODE);
     // TRT does not support INT32 input types for this node
-    ASSERT(!inputs.at(0).isInt32() && !inputs.at(1).isInt32()
-        && "TensorRT doesn't support INT32 inputs for GEMM!", ErrorCode::kUNSUPPORTED_NODE);
-    // Use FC if it is likely to be faster - which is usually when no Shuffles are required.
-    bool canUseFC = inputs.at(0).is_tensor() && inputs.at(1).is_weights() && alpha == 1.f
-        && beta == 1.f && inputs.at(0).tensor().getDimensions().nbDims == 2 && inputs.at(1).weights().shape.nbDims == 2;
-    canUseFC &= inputs.size() < 3 || (inputs.at(2).is_weights() && inputs.at(2).weights().shape.nbDims == 1);
-    if (canUseFC)
-    {
-        LOG_VERBOSE("GEMM: using FC layer instead of MM because all criteria were met.");
-        const std::vector<int> axesInput{2, 3};
-        nvinfer1::ITensor* inputAExtendDim = unsqueezeTensor(ctx, node, inputA, axesInput);
-
-        ShapedWeights weights = inputs.at(1).weights();
-        if (!transB)
-        {
-            auto transposedWeights = ctx->createTempWeights(weights.type, weights.shape);
-            ASSERT(transposeWeights(weights, {1, 0}, &transposedWeights, ctx), ErrorCode::kUNSUPPORTED_NODE);
-            weights = transposedWeights;
-        }
-        ShapedWeights biases{};
-        if (inputs.size() > 2)
-        {
-            biases = inputs.at(2).weights();
-        }
-        nvinfer1::IFullyConnectedLayer* fc = ctx->network()->addFullyConnected(*inputAExtendDim, biases.shape.d[0], weights, biases);
-        // Register layer, along with refittable kernel weights and bias weights (if any)
-        ctx->registerLayer(fc, getNodeName(node));
-        ctx->network()->setWeightsName(weights, weights.getName());
-        if (inputs.size() == 3)
-        {
-            ctx->network()->setWeightsName(biases, inputs.at(2).weights().getName());
-        }
-        const std::vector<int> axesOutput{2, 3};
-        return {{squeezeTensor(ctx, node, *fc->getOutput(0), axesOutput)}};
-    }
-
-    nvinfer1::ITensor* inputB {nullptr};
-
-    // If input B is a constant, we transpose at parse time if necessary,
-    // because In some cases, A * Bt is much slower than A * B.
-    if (inputs.at(1).is_weights())
-    {
-        ShapedWeights weights = inputs.at(1).weights();
-        if (transB)
-        {
-            auto transposedWeights = ctx->createTempWeights(weights.type, weights.shape);
-            ASSERT(transposeWeights(weights, {1, 0}, &transposedWeights, ctx) && "Failed to transpose input tensor B.", ErrorCode::kUNSUPPORTED_NODE);
-            weights = transposedWeights;
-            // Since we've already transposed now, we can set transpose to false.
-            transB = false;
-        }
-        nvinfer1::IConstantLayer* weightsLayer
-            = ctx->network()->addConstant(weights.shape, static_cast<nvinfer1::Weights>(weights));
-        // Map the constant layer to the weights name.
-        ctx->registerLayer(weightsLayer, node.input(1));
-        ctx->network()->setWeightsName(weights, weights.getName());
-        inputB = weightsLayer->getOutput(0);
-    }
-    else
-    {
-        inputB = &inputs.at(1).tensor();
-    }
-
-    nvinfer1::ITensor* inputASqueezed = &inputA;
-    nvinfer1::Dims newDims = squeeze_trailing_dims(inputA.getDimensions());
-    // When A has more than 2 dimensions, it needs to be flattened.
-    if (newDims.nbDims > 2)
-    {
-        newDims = nvinfer1::Dims{1, {-1}};
-    }
-    // Due to other TRT layers, inputA may sometimes have trailing 1s that need to be removed.
-    if (newDims.nbDims < inputA.getDimensions().nbDims)
-    {
-        nvinfer1::IShuffleLayer* squeeze = ctx->network()->addShuffle(inputA);
-        squeeze->setReshapeDimensions(newDims);
-        squeeze->setZeroIsPlaceholder(false);
-        inputASqueezed = squeeze->getOutput(0);
-    }
+    ASSERT(!inputs.at(0).isInt32() && !inputs.at(1).isInt32() && "TensorRT doesn't support INT32 inputs for GEMM!",
+        ErrorCode::kUNSUPPORTED_NODE);
 
     const auto getMatrixOp = [](const nvinfer1::ITensor& input, bool transpose) {
         if (input.getDimensions().nbDims == 1)
@@ -1617,13 +1542,12 @@ DEFINE_BUILTIN_OP_IMPORTER(Gemm)
         return nvinfer1::MatrixOperation::kNONE;
     };
 
-    nvinfer1::MatrixOperation opA = getMatrixOp(*inputASqueezed, transA);
-    nvinfer1::MatrixOperation opB = getMatrixOp(*inputB, transB);
+    nvinfer1::MatrixOperation opA = getMatrixOp(inputA, transA);
+    nvinfer1::MatrixOperation opB = getMatrixOp(inputB, transB);
 
     LOG_VERBOSE("Using opA: " << static_cast<int>(opA) << " opB: " << static_cast<int>(opB));
-    LOG_VERBOSE("GEMM: A, after squeezing: " << inputASqueezed->getDimensions());
 
-    nvinfer1::IMatrixMultiplyLayer* matmul = ctx->network()->addMatrixMultiply(*inputASqueezed, opA, *inputB, opB);
+    nvinfer1::IMatrixMultiplyLayer* matmul = ctx->network()->addMatrixMultiply(inputA, opA, inputB, opB);
     ctx->registerLayer(matmul, getNodeName(node));
     nvinfer1::ITensor* matmulTensor = matmul->getOutput(0);
 
@@ -1655,12 +1579,6 @@ DEFINE_BUILTIN_OP_IMPORTER(Gemm)
                 *betaConstantTensor, *biasTensor, nvinfer1::ElementWiseOperation::kPROD);
             biasTensor = scaledBias->getOutput(0);
         }
-        // A*B may be lower rank than C in TRT, so need to squeeze C.
-        if (ctx->getOpsetVersion() < 7 && !attrs.get("broadcast", false))
-        {
-            nvinfer1::Dims squeezeDims = squeeze_leading_dims(biasTensor->getDimensions());
-            biasTensor = reshapeTensor(ctx, *biasTensor, squeezeDims);
-        }
         CHECK(broadcastTensors(ctx, matmulTensor, biasTensor));
         nvinfer1::IElementWiseLayer* biasAdd
             = ctx->network()->addElementWise(*matmulTensor, *biasTensor, nvinfer1::ElementWiseOperation::kSUM);

diff --git a/trt_utils.hpp b/trt_utils.hpp
@@ -102,31 +102,6 @@ inline nvinfer1::Permutation remove_first_dim(nvinfer1::Permutation const& perm)
     return new_perm;
 }
 
-inline nvinfer1::Dims squeeze_trailing_dims(nvinfer1::Dims const& dims)
-{
-    nvinfer1::Dims new_dims = dims;
-    // Note: TRT requires at least one dimension, so we don't squeeze [1]->[]
-    while (new_dims.nbDims > 1 && new_dims.d[new_dims.nbDims - 1] == 1)
-    {
-        --new_dims.nbDims;
-    }
-    return new_dims;
-}
-
-inline nvinfer1::Dims squeeze_leading_dims(const nvinfer1::Dims& dims)
-{
-    nvinfer1::Dims newDims;
-    // Copy dims only if a non-1 has been seen already.
-    bool non1Seen{false};
-    newDims.nbDims = std::copy_if(dims.d, dims.d + dims.nbDims, newDims.d,
-                         [&non1Seen](int x) {
-                             non1Seen = (x != 1) ? true : non1Seen;
-                             return non1Seen;
-                         })
-        - newDims.d;
-    return newDims;
-}
-
 inline nvinfer1::DimsHW operator-(nvinfer1::DimsHW dims)
 {
     return nvinfer1::DimsHW(-dims.h(), -dims.w());