Generating TV and Launch Pattern Files#
Since the cuBB 22-2.2 release, the test vectors are not included in the release package. You must generate the TV files before running cuPHY examples or cuBB end-to-end test.
Note
TV generation is NOT supported on ARM because Matlab Compiler SDK doesn’t support it yet.
Using Aerial Python mcore Module#
No Matlab license is required to generate TV files using the Aerial Python mcore module. The cuBB container already has
aerial_mcore installed.
To generate the test vectors required for end-to-end testing, follow these steps:
Install the MATLAB Runtime and supporting apt packages inside the Aerial container. Note that following these instructions accepts the MATHWORKS license.
sudo apt update && sudo apt install -y unzip wget https://ssd.mathworks.com/supportfiles/downloads/R2023a/Release/1/deployment_files/installer/complete/glnxa64/MATLAB_Runtime_R2023a_Update_1_glnxa64.zip mkdir unzip && cd unzip unzip ../MATLAB_Runtime_R2023a_Update_1_glnxa64.zip sudo ./install -mode silent -agreeToLicense yes cd .. && rm -rf MATLAB_Runtime_R2023a_Update_1_glnxa64.zip unzip sudo apt install -y libxcomposite1 libnss3 libxrandr-dev libatk1.0-0 libatk-bridge2.0-0 libx11-xcb-dev libxcb-dri3-0 libxcursor-dev libxdamage-dev libxi-dev libdrm-dev libgbm-dev libasound-dev libcups2-dev libxtst-dev
These instructions can be run directly inside the Aerial container as shown above, or with a Dockerfile against the Aerial container so that a new container with the MATLAB Runtime and its dependencies are only installed once. Create a Dockerfile with the following contents by running the following command. Note that the double backslash ensures a single backslash in the Dockerfile.
mkdir temp_dockerfile && cd temp_dockerfile cat << EOF > Dockerfile FROM nvcr.io/qhrjhjrvlsbu/aerial-cuda-accelerated-ran:25-1-cubb USER root RUN apt update && apt install -y unzip RUN wget https://ssd.mathworks.com/supportfiles/downloads/R2023a/Release/1/deployment_files/installer/complete/glnxa64/MATLAB_Runtime_R2023a_Update_1_glnxa64.zip && \\ mkdir unzip && \\ cd unzip && \\ unzip ../MATLAB_Runtime_R2023a_Update_1_glnxa64.zip && \\ ./install -mode silent -agreeToLicense yes && \\ cd .. && \\ rm -rf MATLAB_Runtime_R2023a_Update_1_glnxa64.zip unzip RUN apt install -y \\ libxcomposite1 \\ libnss3 \\ libxrandr-dev \\ libatk1.0-0 \\ libatk-bridge2.0-0 \\ libx11-xcb-dev \\ libxcb-dri3-0 \\ libxcursor-dev \\ libxdamage-dev \\ libxi-dev \\ libdrm-dev \\ libgbm-dev \\ libasound-dev \\ libcups2-dev \\ libxtst-dev USER aerial EOF
Next execute the following command to create the new Aerial container with the MATLAB runtime and its dependencies. This needs to be executed from a machine that has docker installed. See instructions here
docker build -t aerial-cuda-accelerated-ran:25-1-cubb-matlab-runtime-enabled .
This will create the aerial-cuda-accelerated-ran image with a 25-1-cubb-matlab-runtime-enabled tag. Use the 25-1-cubb-matlab-runtime-enabled tagged image in place of the Aerial container when generating TVs using aerial_mcore.
Run the following inside the Aerial container. It completes in less than a minute.
cd ${cuBB_SDK}/5GModel/aerial_mcore/examples source ../scripts/setup.sh ../scripts/gen_e2e_ota_tvs.sh ls -lh GPU_test_input/ cp GPU_test_input/* ${cuBB_SDK}/testVectors/
The following is example output from the above commands:
aerial@c_aerial_aerial:/opt/nvidia/cuBB/5GModel/aerial_mcore$ source ./scripts/setup.sh [Aerial Python]aerial@c_aerial_aerial:/opt/nvidia/cuBB/5GModel/aerial_mcore$ ./scripts/gen_e2e_ota_tvs.sh Finished genCuPhyChEstCoeffs Elapsed time: 1.166473150253296 seconds [Aerial Python]aerial@c_aerial_aerial:/opt/nvidia/cuBB/5GModel/aerial_mcore$ ls -lh ../GPU_test_input/ -rw-rw-r-- 1 aerial aerial 90K Oct 17 2023 ../cuPhyChEstCoeffs.h5
Note
The
cuPhyChEstCoeffs.h5file can be found in the/opt/nvidia/cuBB/testVectorsdirectory of both the x86 and ARM containers.
Copy the output to the
testVectorsfolder.
To generate all of the TV files, including files that are not necessary for E2E testing, follow these steps:
Run the following commands inside the Aerial container.
cd ${cuBB_SDK}/5GModel/aerial_mcore/examples source ../scripts/setup.sh export REGRESSION_MODE=1 time python3 ./example_5GModel_regression.py allChannels echo $? ls -alF GPU_test_input/ du -h GPU_test_input/
Note
The TV generation may take a few hours on the devkit with the current isocpus parameter setting in the kernel command line. The host must have at least 64GB of memory and 430GB of available disk space. Hyperthreading must be enabled.
Review the output from the above commands; an example is shown below. The “real” time takes less than one hour on a 24-core x86 host. The
echo $?command shows the exit code of the process, which should be 0, while a non-zero exit code indicates a failure.Channel Compliance_Test Error Test_Vector Error Performance_Test Fail ------------------------------------------------------------------------------ SSB 37 0 42 0 0 0 PDCCH 71 0 80 0 0 0 PDSCH 274 0 286 0 0 0 CSIRS 86 0 87 0 0 0 DLMIX 0 0 1049 0 0 0 PRACH 60 0 60 0 48 0 PUCCH 469 0 469 0 96 0 PUSCH 388 0 398 0 41 0 SRS 125 0 125 0 0 0 ULMIX 0 0 576 0 0 0 BFW 58 0 58 0 0 0 ------------------------------------------------------------------------------ Total 1568 0 3230 0 185 0 Total time for runRegression is 2147 seconds Parallel pool using the 'local' profile is shutting down. real 36m51.931s user 585m1.704s sys 10m28.322s
To generate the launch pattern for each test case using cubb_scripts, follow these steps:
Run the following commands:
cd $cuBB_SDK cd cubb_scripts python3 auto_lp.py -i ../5GModel/aerial_mcore/examples/GPU_test_input -t launch_pattern_nrSim.yaml
Copy the launch pattern and TV files to the
testVectorsrepo:cd $cuBB_SDK cp ./5GModel/aerial_mcore/examples/GPU_test_input/*h5 ./testVectors/. cp ./5GModel/aerial_mcore/examples/GPU_test_input/launch_pattern* ./testVectors/multi-cell/.`
Using Matlab#
To generate TV files using Matlab:
Run the following command in Matlab:
cd('nr_matlab'); startup; [nTC, errCnt] = runRegression({'TestVector'}, {'allChannels'}, 'compact', [0, 1] );
All the cuPHY TVs are generated and stored under
nr_matlab/GPU_test_input.Generate the launch pattern for each test case using cubb_scripts:
cd $cuBB_SDK cd cubb_scripts python3 auto_lp.py -i ../5GModel/nr_matlab/GPU_test_input -t launch_pattern_nrSim.yaml
Copy the launch pattern and TV files to testVectors repo.
cd $cuBB_SDK
cp ./5GModel/nr_matlab/GPU_test_input/TVnr_* ./testVectors/.
cp ./5GModel/nr_matlab/GPU_test_input/launch_pattern* ./testVectors/multi-cell/.