Cupy, CUDA Bilinear interpolation

Ultra fast Bilinear interpolation in image resize with CUDA.

lerp.py : Concept and code base (*single thread, may take a while to run).
resize_ker.cu : CUDA test case in C.
resize.py : Cupy example

(*PyCUDA(deprecated) is no longer support , use cupy instead )

Requirements:

• GPU (compute capability: 3.0 or above, testing platform: 7.5)
• CUDA driver
• Docker and nvidia docker

---

Pros:

• support Batch image.
• no shared object .so and .dll binary file
• Install cupy and use
• Compatible to Numpy library
• pass the GPU array to TensorRT directly.

Cons:

• still need the concept of CUDA programming
• SourceModule have to write in C CUDA, including all CUDA kernel and device code

---

Quick Start

# Pull docker image
docker run -it --runtime=nvidia royinx/cuda_resize bash

# For Cupy implementation
python3 resize.py

# For concept
python3 lerp.py

# For CUDA kernel testing
nvcc resize_free.cu -o resize_free.o && ./resize_free.o

# For benmarking
wget http://images.cocodataset.org/zips/val2017.zip
unzip val2017.zip
python3 benchmark.py

Build

git clone https://github.com/royinx/CUDA_Resize.git
cd CUDA_Resize
docker build -t lerp_cuda .
docker run -it --runtime=nvidia -v ${PWD}:/py -w /py lerp_cuda bash

Advance Metrics

docker run -it --privileged --runtime=nvidia -p 20072:22 -v ${PWD}:/py -w /py lerp_cuda bash
sh -c 'echo 1 >/proc/sys/kernel/perf_event_paranoid'
nvcc resize_free.cu -o resize_free.o
nsys profile ./resize_free.o

ncu -o metrics /bin/python3 resize.py  > profile_log
ncu -o metrics /bin/python3 resize.py

Remark: Development platform is in dockerfile.opencv with OpenCV in C for debugging

Function Working well in pycuda container, you dont need to build OpenCV.

---

Benchmark

2080ti

ratio = 2080ti (ms) / Ryzen 2700x (ms)

time (us/img)

shared memory

(Deprecated) [w/o smem] AWS g4dn.xlarge (Tesla T4)

ratio = T4 (ms) per img / Xeon Platinum 8259CL (ms) per img

(ms) per img on T4