chore: import upstream snapshot with attribution

2026-07-13 13:33:03 +08:00
commit 5b57521aa1
8226 changed files with 3425766 additions and 0 deletions
@@ -0,0 +1,51 @@
+//
+//  TRTBatchMatMul.cpp
+//  MNN
+//
+//  Created by MNN on 2021/02/28.
+//  Copyright © 2018, Alibaba Group Holding Limited
+//
+
+#include "TRTBatchMatMul.hpp"
+#include <core/TensorUtils.hpp>
+#include "TRTBackend.hpp"
+
+using namespace std;
+
+namespace MNN {
+
+nvinfer1::MatrixOperation transposeFormat(nvinfer1::ITensor *x, bool transpose) {
+    return transpose ? nvinfer1::MatrixOperation::kTRANSPOSE : nvinfer1::MatrixOperation::kNONE;
+}
+
+TRTBatchMatMul::TRTBatchMatMul(Backend *b, const Op *op, const std::vector<Tensor *> &inputs,
+                     const std::vector<Tensor *> &outputs)
+    : MNN::TRTCommonExecution(b, op) {
+#ifdef TRT_LOG
+    printf("TRTBatchMatMul in\n");
+#endif
+}
+
+std::vector<ITensor *> TRTBatchMatMul::onEncode(const std::vector<ITensor *> &xOp) {
+#ifdef TRT_LOG
+    printf("TRTBatchMatMul in\n");
+#endif
+    auto param       = mOp->main_as_BatchMatMulParam();
+    MNN_ASSERT(mInputs.size() == 2);
+    bool isConst0 = TensorUtils::getDescribe(mInputs[0])->usage == Tensor::InsideDescribe::Usage::CONSTANT;
+    bool isConst1 = TensorUtils::getDescribe(mInputs[1])->usage == Tensor::InsideDescribe::Usage::CONSTANT;
+
+    auto dimSize0 = mInputs[0]->dimensions();
+    auto dimSize1 = mInputs[1]->dimensions();
+
+    auto transpose_a = transposeFormat(xOp[0], param->adjX());
+    auto transpose_b = transposeFormat(xOp[1], param->adjY());
+
+    auto matmul_layer = mTrtBackend->getNetwork()->addMatrixMultiply(*xOp[0], transpose_a, *xOp[1], transpose_b);
+    return {matmul_layer->getOutput(0)};
+
+}
+
+TRTCreatorRegister<TypedCreator<TRTBatchMatMul>> __batch_matmul_op(OpType_BatchMatMul);
+
+} // namespace MNN