End-to-End ResNet (C++/CUDA/cuBLAS) Build Instructions

Requirements

• CUDA Toolkit (with cuBLAS)
• NVIDIA GPU
• g++ (for host code)

Build

nvcc -o resnet_main resnet_main.cu -lcublas

Run

./resnet_main

Notes

• This is a minimal demo: no data/weight loading, only random values.
• Only a single residual block is implemented for demonstration.
• Uses im2col for convolution and cuBLAS for matrix multiplication.
• Extend as needed for full ResNet.

ResNet-34 (C++/CUDA/cuBLAS)

Structure

• core/: Tensor, im2col, cuBLAS wrappers
• layers/: Conv2D, BatchNorm, ReLU
• blocks/: BasicBlock (residual)
• models/: ResNet34
• utils/: Utilities
• main.cpp: Entry point

Build

nvcc -o resnet34 main.cpp \
    core/tensor.cpp core/im2col.cu core/cublas_utils.cpp \
    layers/conv2d.cpp layers/batchnorm.cpp layers/relu.cpp \
    blocks/basic_block.cpp models/resnet34.cpp \
    -lcublas

Run

./resnet34

Notes

• All weights and input are random.
• Only forward pass is implemented.
• Output is the first value of the final feature map.