Radeon Performance Primitives Library

Radeon Performance Primitives (RPP) library is a comprehensive high-performance computer vision library for AMD (CPU and GPU) with HIP and OpenCL back-ends.

Latest Release

Top level design

RPP is developed for Linux operating system.

Prerequisites

• OS

  • Ubuntu 18.04/20.04
  • CentOS/RHEL 8.0

• ROCm supported hardware

• ROCm

• Clang Version 6.0

  sudo apt-get install clang-6.0
  sudo ln -s /usr/bin/clang-6.0 /usr/bin/clang
  sudo ln -s /usr/bin/clang++-6.0 /usr/bin/clang++
  

• CMake Version 3.5 or higher

  sudo apt-get install cmake
  

• Boost Version 1.72 (install from source: Easy Build and Install - https://www.boost.org/doc/libs/1_72_0/more/getting_started/unix-variants.html#easy-build-and-install)

  wget https://boostorg.jfrog.io/artifactory/main/release/1.72.0/source/boost_1_72_0.tar.gz
  tar -xzvf boost_1_72_0.tar.gz
  cd boost_1_72_0
  ./bootstrap.sh
  ./b2 install
  

• IEEE 754-based half-precision floating-point library - half.hpp

  wget https://sourceforge.net/projects/half/files/half/1.12.0/half-1.12.0.zip
  unzip half-1.12.0.zip -d half-files
  sudo cp half-files/include/half.hpp /usr/local/include/
  

Prerequisites for Test Suite

• OpenCV 3.4.0

  wget https://github.com/opencv/opencv/archive/3.4.0.zip
  unzip 3.4.0.zip
  cd opencv-3.4.0/
  mkdir build
  cd build
  cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=/usr/local ..
  make -j<number of threads>
  sudo make install
  

Supported Functionalities and Variants

Supported Functionalities List

Supported Functionalities Samples

Supported Variants

• Packed 3 Channel (RGB) - PKD3 images
• Planar 3 Channel (RGB) - PLN3 images
• Planar 1 Channel (Greyscale) - PLN1 images
• Affinity - Host and GPU
• Batch Processing - 26 variants
• ROI variants
• Padded variants

Instructions to build the library

$ git clone https://github.com/GPUOpen-ProfessionalCompute-Libraries/rpp.git
$ cd rpp
$ mkdir build
$ cd build
$ cmake -DBACKEND=OCL .. #for OCL and HOST
or
$ cmake -DBACKEND=HIP .. #for HIP
$ make -j16
$ sudo make install

Instructions to test functionalities after HOST installation

$ cd rpp/utilities/rpp-unittests/HOST_NEW
$ ./testAllScript.sh

Instructions to test functionalities after OCL installation

$ cd rpp/utilities/rpp-unittests/OCL_NEW
$ ./testAllScript.sh

Instructions to test functionalities after HIP installation

$ cd rpp/utilities/rpp-unittests/HIP_NEW
$ ./testAllScript.sh

MIVisionX(OpenVX) Support

Extended RPP support as a functionality through OpenVX MIVisionX (Find build instructions and build the amd_rpp library)

Miscellaneous examples

RPP stand-alone batch processing code snippet (HOST)

// Initializations
int noOfImages = 32;
int channels = 3;
RppiSize maxSize;
maxSize.width = 224;
maxSize.height = 224;

// Allocate host memory and/or obtain input data
unsigned long long ioBufferSize = noOfImages * maxSize.width * maxSize.height * channels;
Rpp8u *input = (Rpp8u *)calloc(ioBufferSize, sizeof(Rpp8u));
Rpp8u *output = (Rpp8u *)calloc(ioBufferSize, sizeof(Rpp8u));
// Get the data for a batch of 224x224 images into 'input' here

// Initialize values for any necessary parameters to the RPP function being called
Rpp32f alpha[noOfImages];
Rpp32f beta[noOfImages];
for (int i = 0; i < noOfImages; i++)
{
    alpha[i] = 1.75;
    beta[i] = 50;
    srcSize[i].width = 224;
    srcSize[i].height = 224;
}

// Create handle
rppHandle_t handle;
rppCreateWithBatchSize(&handle, noOfImages);

// Call the RPP API for the specific variant required (pkd3/pln3/pln1)
rppi_brightness_u8_pkd3_batchPD_host(input, srcSize, maxSize, output, alpha, beta, noOfImages, handle);

RPP stand-alone batch processing code snippet (OCL)

// Initializations
int noOfImages = 32;
int channels = 3;
RppiSize maxSize;
maxSize.width = 224;
maxSize.height = 224;

// Allocate host memory and/or obtain input data
unsigned long long ioBufferSize = noOfImages * maxSize.width * maxSize.height * channels;
Rpp8u *input = (Rpp8u *)calloc(ioBufferSize, sizeof(Rpp8u));
Rpp8u *output = (Rpp8u *)calloc(ioBufferSize, sizeof(Rpp8u));
// Get the data for a batch of 224x224 images into 'input' here

// OCL initializations, allocate device memory and copy input data to device
cl_mem d_input, d_output;
cl_platform_id platform_id;
cl_device_id device_id;
cl_context theContext;
cl_command_queue theQueue;
cl_int err;
err = clGetPlatformIDs(1, &platform_id, NULL);
err |= clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_GPU, 1, &device_id, NULL);
theContext = clCreateContext(0, 1, &device_id, NULL, NULL, &err);
theQueue = clCreateCommandQueueWithProperties(theContext, device_id, 0, &err);
d_input = clCreateBuffer(theContext, CL_MEM_READ_ONLY, ioBufferSize * sizeof(Rpp8u), NULL, NULL);
d_output = clCreateBuffer(theContext, CL_MEM_READ_ONLY, ioBufferSize * sizeof(Rpp8u), NULL, NULL);
err |= clEnqueueWriteBuffer(theQueue, d_input, CL_TRUE, 0, ioBufferSize * sizeof(Rpp8u), input, 0, NULL, NULL);
err |= clEnqueueWriteBuffer(theQueue, d_output, CL_TRUE, 0, ioBufferSize * sizeof(Rpp8u), output, 0, NULL, NULL);

// Initialize values for any necessary parameters to the RPP function being called
Rpp32f alpha[noOfImages];
Rpp32f beta[noOfImages];
for (int i = 0; i < noOfImages; i++)
{
    alpha[i] = 1.75;
    beta[i] = 50;
    srcSize[i].width = 224;
    srcSize[i].height = 224;
}

// Create handle
rppHandle_t handle;
rppCreateWithStreamAndBatchSize(&handle, theQueue, noOfImages);

// Call the RPP API for the specific variant required (pkd3/pln3/pln1)
rppi_brightness_u8_pkd3_batchPD_gpu(d_input, srcSize, maxSize, d_output, alpha, beta, noOfImages, handle);

// Copy output data back to host
clEnqueueReadBuffer(theQueue, d_output, CL_TRUE, 0, ioBufferSize * sizeof(Rpp8u), output, 0, NULL, NULL);

RPP stand-alone batch processing code snippet (HIP)

// Initializations
int noOfImages = 32;
int channels = 3;
RppiSize maxSize;
maxSize.width = 224;
maxSize.height = 224;

// Allocate host memory and/or obtain input data
unsigned long long ioBufferSize = noOfImages * maxSize.width * maxSize.height * channels;
Rpp8u *input = (Rpp8u *)calloc(ioBufferSize, sizeof(Rpp8u));
Rpp8u *output = (Rpp8u *)calloc(ioBufferSize, sizeof(Rpp8u));
// Get the data for a batch of 224x224 images into 'input' here

// HIP initializations, allocate device memory and copy input data to device
int *d_input, d_output;
hipMalloc(&d_input, ioBufferSize * sizeof(Rpp8u));
hipMalloc(&d_output, ioBufferSize * sizeof(Rpp8u));
hipMemcpy(d_input, input, ioBufferSize * sizeof(Rpp8u), hipMemcpyHostToDevice);
hipMemcpy(d_output, output, ioBufferSize * sizeof(Rpp8u), hipMemcpyHostToDevice);

// Initialize values for any necessary parameters to the RPP function being called
Rpp32f alpha[noOfImages];
Rpp32f beta[noOfImages];
for (int i = 0; i < noOfImages; i++)
{
    alpha[i] = 1.75;
    beta[i] = 50;
    srcSize[i].width = 224;
    srcSize[i].height = 224;
}

// Create handle
rppHandle_t handle;
hipStream_t stream;
hipStreamCreate(&stream);
rppCreateWithStreamAndBatchSize(&handle, stream, noOfImages);

// Call the RPP API for the specific variant required (pkd3/pln3/pln1)
rppi_brightness_u8_pkd3_batchPD_gpu(d_input, srcSize, maxSize, d_output, alpha, beta, noOfImages, handle);

// Copy output data back to host
hipMemcpy(output, d_output, ioBufferSize * sizeof(Rpp8u), hipMemcpyDeviceToHost);