OpenCL est une forme abrégée pour "Open Computing Language". C'est un langage de programmation qui peut être utilisé sur diverses plates-formes, principalement pour le calcul accéléré. En raison de la diversité de ses applicabilités sur plusieurs plates-formes, il est le plus souvent appelé langage informatique multiplateforme. Vous pouvez écrire des programmes sur OpenCL et les exécuter sur une variété d'appareils, y compris des processeurs, des GPU, des FPGA et bien plus encore.
Dans ce guide, je vais me concentrer uniquement sur les GPU. J'ai travaillé avec des GPU NVIDIA et AMD et je vais vous montrer comment vous pouvez les rendre opérationnels avec OpenCL de la manière la plus simple possible.
Bien que j'aie utilisé Ubuntu pour le système hôte, la partie Docker est applicable à toutes les autres distributions Linux.
Prérequis
- Carte graphique NVIDIA/AMD
- Ubuntu Linux 20.04.2 LTS Desktop/Server 64 bits
- Docker (pour une utilisation spécifique à l'application)
Passons aux détails !
Configuration d'OpenCL pour les GPU NVIDIA
Je vais d'abord vous montrer comment vous assurer qu'OpenCL fonctionne sur votre bureau/serveur Ubuntu principal. Une fois cela fait, je vous montrerai comment exécuter des conteneurs Docker dans le même but avec le GPU NVIDIA.
Exécuter OpenCL sur le système hôte
Sur un nouveau système Ubuntu, vous devez d'abord installer le pilote propriétaire NVIDIA et CUDA. Ce dernier garantit que vous obtenez le framework OpenCL fourni avec lui. Enfin, installez le clinfo programme pour vous assurer que OpenCL est correctement installé, vous montrant en détail les spécifications OpenCL de votre GPU NVIDIA. Voyons comment :
Vérifier le pilote recommandé
Utilisez les appareils ubuntu-drivers devices commande pour récupérer le nom de votre pilote recommandé :
[email protected]:~$ ubuntu-drivers devices
== /sys/devices/pci0000:00/0000:00:01.0/0000:01:00.0 ==
modalias : pci:v000010DEd00001C8Csv00001025sd00001265bc03sc00i00
vendor : NVIDIA Corporation
model : GP107M [GeForce GTX 1050 Ti Mobile]
driver : nvidia-driver-460 - distro non-free recommended
driver : nvidia-driver-418-server - distro non-free
driver : nvidia-driver-390 - distro non-free
driver : nvidia-driver-450-server - distro non-free
driver : nvidia-driver-465 - distro non-free
driver : nvidia-driver-460-server - distro non-free
driver : xserver-xorg-video-nouveau - distro free builtin
Ci-dessus, notez que le pilote recommandé est nvidia-driver-460 .
Installer tous les packages nécessaires
Alors installons le pilote recommandé avec CUDA et le clinfo package mentionné précédemment dans cette section :
sudo apt install nvidia-driver-460 nvidia-cuda-toolkit clinfoUne fois les trois packages ci-dessus installés, redémarrez votre bureau/serveur Ubuntu.
Vérifiez votre configuration OpenCL
[email protected]:~$ clinfo
Number of platforms 1
Platform Name NVIDIA CUDA
Platform Vendor NVIDIA Corporation
Platform Version OpenCL 1.2 CUDA 9.1.84
Platform Profile FULL_PROFILE
Platform Extensions cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_nv_create_buffer
Platform Extensions function suffix NV
Platform Name NVIDIA CUDA
Number of devices 1
Device Name GeForce GTX 1050 Ti
Device Vendor NVIDIA Corporation
Device Vendor ID 0x10de
Device Version OpenCL 1.2 CUDA
Driver Version 390.143
Device OpenCL C Version OpenCL C 1.2
Device Type GPU
Device Topology (NV) PCI-E, 01:00.0
Device Profile FULL_PROFILE
Device Available Yes
Compiler Available Yes
Linker Available Yes
Max compute units 6
Max clock frequency 1620MHz
Compute Capability (NV) 6.1
Device Partition (core)
Max number of sub-devices 1
Supported partition types None
Max work item dimensions 3
Max work item sizes 1024x1024x64
Max work group size 1024
Preferred work group size multiple 32
Warp size (NV) 32
Preferred / native vector sizes
char 1 / 1
short 1 / 1
int 1 / 1
long 1 / 1
half 0 / 0 (n/a)
float 1 / 1
double 1 / 1 (cl_khr_fp64)
Half-precision Floating-point support (n/a)
Single-precision Floating-point support (core)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Correctly-rounded divide and sqrt operations Yes
Double-precision Floating-point support (cl_khr_fp64)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Address bits 64, Little-Endian
Global memory size 4236312576 (3.945GiB)
Error Correction support No
Max memory allocation 1059078144 (1010MiB)
Unified memory for Host and Device No
Integrated memory (NV) No
Minimum alignment for any data type 128 bytes
Alignment of base address 4096 bits (512 bytes)
Global Memory cache type Read/Write
Global Memory cache size 98304 (96KiB)
Global Memory cache line size 128 bytes
Image support Yes
Max number of samplers per kernel 32
Max size for 1D images from buffer 134217728 pixels
Max 1D or 2D image array size 2048 images
Max 2D image size 16384x32768 pixels
Max 3D image size 16384x16384x16384 pixels
Max number of read image args 256
Max number of write image args 16
Local memory type Local
Local memory size 49152 (48KiB)
Registers per block (NV) 65536
Max number of constant args 9
Max constant buffer size 65536 (64KiB)
Max size of kernel argument 4352 (4.25KiB)
Queue properties
Out-of-order execution Yes
Profiling Yes
Prefer user sync for interop No
Profiling timer resolution 1000ns
Execution capabilities
Run OpenCL kernels Yes
Run native kernels No
Kernel execution timeout (NV) Yes
Concurrent copy and kernel execution (NV) Yes
Number of async copy engines 2
printf() buffer size 1048576 (1024KiB)
Built-in kernels
Device Extensions cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_nv_create_buffer
NULL platform behavior
clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...) NVIDIA CUDA
clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...) Success [NV]
clCreateContext(NULL, ...) [default] Success [NV]
clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM) Invalid device type for platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL) No platform
ICD loader properties
ICD loader Name OpenCL ICD Loader
ICD loader Vendor OCL Icd free software
ICD loader Version 2.2.11
ICD loader Profile OpenCL 2.1
Notez qu'ici, seul le nom de la plate-forme est "NVIDIA CUDA". Mais CUDA et OpenCL sont différents l'un de l'autre.
C'est ça! Vous pouvez désormais exécuter des applications OpenCL avec votre GPU NVIDIA sur votre système hôte !
OpenCL sur Docker pour les GPU NVIDIA
Maintenant qu'OpenCL est opérationnel sur votre système bare metal, voyons comment vous pouvez l'installer sur un conteneur Docker !
Installer le runtime de conteneur NVIDIA
Ici, vous devez en plus installer le nvidia-container-runtime emballer.
Pour pouvoir l'installer, vous devez d'abord ajouter les détails du référentiel. Assurez-vous que Curl est installé si vous ne l'avez pas déjà sur votre système.
sudo apt install curl
curl -s -L https://nvidia.github.io/nvidia-container-runtime/gpgkey | \
sudo apt-key add -
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.list | \
sudo tee /etc/apt/sources.list.d/nvidia-container-runtime.list
sudo apt update
sudo apt install nvidia-container-runtimeCréation du Dockerfile
Il est nécessaire que vous répliquiez tout ce que vous avez fait sur le système hôte sur une nouvelle image fraîche afin que vous puissiez l'utiliser pour lancer nos applications OpenCL personnalisées sur un conteneur (nous en parlerons plus tard).
Créez un nouveau répertoire pour votre projet NVIDIA GPU OpenCL et déplacez-vous dedans :
mkdir nvidia-opencl
cd nvidia-opencl
Utilisez votre éditeur de texte préféré (Vim/Nano ou tout autre) pour créer le Dockerfile suivant et enregistrez-le :
FROM ubuntu:20.04
ARG DEBIAN_FRONTEND=noninteractive
RUN apt-get update && apt-get -y upgrade \
&& apt-get install -y \
apt-utils \
unzip \
tar \
curl \
xz-utils \
ocl-icd-libopencl1 \
opencl-headers \
clinfo \
;
RUN mkdir -p /etc/OpenCL/vendors && \
echo "libnvidia-opencl.so.1" > /etc/OpenCL/vendors/nvidia.icd
ENV NVIDIA_VISIBLE_DEVICES all
ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
Construire le Dockerfile
Alors maintenant que vous avez le Dockerfile nécessaire pour commencer, construisons-le. Je nomme l'image comme nvidia-opencl :
docker build -t nvidia-opencl .
Lancer le conteneur OpenCL
Sur la base de la nouvelle image que vous venez de créer, il est temps de lancer le nouveau conteneur OpenCL !
Tout d'abord, autorisez votre nom d'utilisateur Linux sur la machine locale à se connecter à l'affichage X Windows avec la commande suivante :
xhost +local:username
Avec la commande suivante, vous pouvez maintenant entrer directement dans le shell du conteneur local en fonction de la nouvelle image que vous venez de créer :
docker run --rm -it --gpus all -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=$DISPLAY nvidia-opencl
Vérifiez votre configuration OpenCL sur Docker
Maintenant que vous êtes à l'intérieur du shell du conteneur, vous pouvez exécuter le clinfo pour vérifier votre configuration OpenCL comme vous l'avez fait sur le système hôte bare-metal :
[email protected]:/# clinfo
Number of platforms 1
Platform Name NVIDIA CUDA
Platform Vendor NVIDIA Corporation
Platform Version OpenCL 1.2 CUDA 9.1.84
Platform Profile FULL_PROFILE
Platform Extensions cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_nv_create_buffer
Platform Extensions function suffix NV
Platform Name NVIDIA CUDA
Number of devices 1
Device Name GeForce GTX 1050 Ti
Device Vendor NVIDIA Corporation
Device Vendor ID 0x10de
Device Version OpenCL 1.2 CUDA
Driver Version 390.143
Device OpenCL C Version OpenCL C 1.2
Device Type GPU
Device Topology (NV) PCI-E, 01:00.0
Device Profile FULL_PROFILE
Device Available Yes
Compiler Available Yes
Linker Available Yes
Max compute units 6
Max clock frequency 1620MHz
Compute Capability (NV) 6.1
Device Partition (core)
Max number of sub-devices 1
Supported partition types None
Supported affinity domains (n/a)
Max work item dimensions 3
Max work item sizes 1024x1024x64
Max work group size 1024
Preferred work group size multiple 32
Warp size (NV) 32
Preferred / native vector sizes
char 1 / 1
short 1 / 1
int 1 / 1
long 1 / 1
half 0 / 0 (n/a)
float 1 / 1
double 1 / 1 (cl_khr_fp64)
Half-precision Floating-point support (n/a)
Single-precision Floating-point support (core)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Correctly-rounded divide and sqrt operations Yes
Double-precision Floating-point support (cl_khr_fp64)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Address bits 64, Little-Endian
Global memory size 4236312576 (3.945GiB)
Error Correction support No
Max memory allocation 1059078144 (1010MiB)
Unified memory for Host and Device No
Integrated memory (NV) No
Minimum alignment for any data type 128 bytes
Alignment of base address 4096 bits (512 bytes)
Global Memory cache type Read/Write
Global Memory cache size 98304 (96KiB)
Global Memory cache line size 128 bytes
Image support Yes
Max number of samplers per kernel 32
Max size for 1D images from buffer 134217728 pixels
Max 1D or 2D image array size 2048 images
Max 2D image size 16384x32768 pixels
Max 3D image size 16384x16384x16384 pixels
Max number of read image args 256
Max number of write image args 16
Local memory type Local
Local memory size 49152 (48KiB)
Registers per block (NV) 65536
Max number of constant args 9
Max constant buffer size 65536 (64KiB)
Max size of kernel argument 4352 (4.25KiB)
Queue properties
Out-of-order execution Yes
Profiling Yes
Prefer user sync for interop No
Profiling timer resolution 1000ns
Execution capabilities
Run OpenCL kernels Yes
Run native kernels No
Kernel execution timeout (NV) Yes
Concurrent copy and kernel execution (NV) Yes
Number of async copy engines 2
printf() buffer size 1048576 (1024KiB)
Built-in kernels (n/a)
Device Extensions cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_nv_create_buffer
NULL platform behavior
clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...) NVIDIA CUDA
clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...) Success [NV]
clCreateContext(NULL, ...) [default] Success [NV]
clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM) Invalid device type for platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL) No platform
ICD loader properties
ICD loader Name OpenCL ICD Loader
ICD loader Vendor OCL Icd free software
ICD loader Version 2.2.11
ICD loader Profile OpenCL 2.1
[email protected]:/#
Qu'est-ce que ça veut dire? Cela signifie que vous pouvez maintenant exécuter n'importe quelle application OpenCL à partir de ce conteneur ! Vous n'auriez qu'à remodifier le Dockerfile et ce sera tout.
Vous pouvez également travailler avec des applications Python qui nécessitent un backend OpenCL. Vérifiez ma couverture précédente qui peut très bien servir de compagnon pratique à cet article. Vous voudrez peut-être le vérifier et jouer avec les Dockerfiles.
Configuration d'OpenCL pour les GPU AMD
Je vais d'abord vous montrer comment vous assurer qu'OpenCL fonctionne sur votre bureau/serveur Ubuntu principal. Une fois cela fait, je vous montrerai comment exécuter des conteneurs Docker dans le même but avec le GPU AMD.
Exécuter OpenCL sur le système hôte
Sur un nouveau système Ubuntu, vous devez d'abord télécharger les "pilotes AMDGPU" à partir de la page de support AMD. Pour une configuration à l'épreuve du temps, vous n'avez besoin d'installer OpenCL pour les GPU AMD hérités et plus récents qu'après avoir obtenu l'archive d'installation (tar.xz).
Enfin, installez le clinfo programme pour vous assurer que vous avez correctement installé OpenCL, vous montrant en détail les spécifications OpenCL de votre GPU AMD. Mais l'ensemble du processus peut être un peu plus délicat que prévu. Voyons comment.
Télécharger les pilotes AMDGPU avec Curl
Naviguez sur la page de support AMD et téléchargez le pilote approprié avec Curl. Assurez-vous que Curl est installé.
sudo apt install curl
curl -e https://drivers.amd.com/drivers/linux -O https://drivers.amd.com/drivers/linux/amdgpu-pro-21.10-1247438-ubuntu-20.04.tar.xz
Installation, anomalies et leurs solutions
Extrayez l'archive :
tar -Jxvf amdgpu-pro-21.10-1247438-ubuntu-20.04.tar.xz
Déplacez-vous dans le nouveau répertoire :
cd amdgpu-pro-21.10-1247438-ubuntu-20.04
Maintenant, je vais installer OpenCL pour les anciens et les nouveaux GPU :
./amdgpu-install --opencl=legacy,rocr --headless --no-dkms
Pour un aperçu complet de son utilisation, vous pouvez utiliser la commande ./amdgpu-install -h pour en savoir plus sur le fonctionnement fondamental du script. C'est similaire à une entrée man pour la commande. Le --headless l'option spécifie uniquement la prise en charge d'OpenCL et --no-dkms lui dit de NE PAS installer le amdgpu-dkms et le amdgpu-dkms-firmware paquets dans le noyau. Vous n'en avez pas besoin.
Depuis un certain temps, il a été constaté que même si vous spécifiez le --no-dkms option, le script ne se soucie pas de se conformer et procède à l'installation de ces packages inutiles. En ajouter plus, si j'autorisais amdgpu-dkms pour installer et modifier la configuration du noyau, le système refuserait de redémarrer ou de s'arrêter par la suite ! Cela s'est produit après avoir reçu une mise à jour du noyau à partir des référentiels Ubuntu.
Dans un tel cas, voici ce que j'ai fait :
J'ai installé manuellement les packages suivants en utilisant dpkg -i package-name.deb , présent dans le répertoire extrait :
amdgpu-pin_21.10-1247438_all.deb
amdgpu-core_21.10-1247438_all.deb
amdgpu-pro-core_21.10-1247438_all.deb
libdrm-amdgpu-common_1.0.0-1247438_all.deb
libdrm2-amdgpu_2.4.100-1247438_amd64.deb
libdrm-amdgpu-amdgpu1_2.4.100-1247438_amd64.deb
hsakmt-roct-amdgpu_1.0.9-1247438_amd64.deb
hsa-runtime-rocr-amdgpu_1.3.0-1247438_amd64.deb
comgr-amdgpu-pro_2.0.0-1247438_amd64.deb
hip-rocr-amdgpu-pro_21.10-1247438_amd64.deb
ocl-icd-libopencl1-amdgpu-pro_21.10-1247438_amd64.deb
clinfo-amdgpu-pro_21.10-1247438_amd64.deb
opencl-rocr-amdgpu-pro_21.10-1247438_amd64.deb
libllvm11.0-amdgpu_11.0-1247438_amd64.deb
Cela garantissait que amdgpu-dkms et amdgpu-dkms-firmware pourraient être évités et laisser le noyau intact. Notez également que j'ai téléchargé l'ancien pilote 21.10 même si la version 21.30 la plus récente et la plus récente est disponible. La raison étant que ce dernier refuse de reconnaître mon GPU Radeon VII en donnant une "erreur HSA" lorsque j'exécute clinfo plus tard :
HSA Error: Incompatible kernel and userspace, Vega 20 [Radeon VII] disabled. Upgrade amdgpu.
Après m'être occupé de ces anomalies, j'ai pu obtenir clinfo pour signaler mon GPU correctement.
Installer le package clinfo
Installez le clinfo package comme vous l'avez fait précédemment pour les GPU NVIDIA :
sudo apt install clinfo
Vérifiez votre configuration OpenCL
[email protected]:~$ clinfo
Number of platforms 1
Platform Name AMD Accelerated Parallel Processing
Platform Vendor Advanced Micro Devices, Inc.
Platform Version OpenCL 2.0 AMD-APP (3246.0)
Platform Profile FULL_PROFILE
Platform Extensions cl_khr_icd cl_amd_event_callback
Platform Extensions function suffix AMD
Platform Name AMD Accelerated Parallel Processing
Number of devices 1
Device Name gfx906:sramecc-:xnack-
Device Vendor Advanced Micro Devices, Inc.
Device Vendor ID 0x1002
Device Version OpenCL 2.0
Driver Version 3246.0 (HSA1.1,LC)
Device OpenCL C Version OpenCL C 2.0
Device Type GPU
Device Board Name (AMD) Vega 20 [Radeon VII]
Device Topology (AMD) PCI-E, 0a:00.0
Device Profile FULL_PROFILE
Device Available Yes
Compiler Available Yes
Linker Available Yes
Max compute units 60
SIMD per compute unit (AMD) 4
SIMD width (AMD) 16
SIMD instruction width (AMD) 1
Max clock frequency 1801MHz
Graphics IP (AMD) 9.0
Device Partition (core)
Max number of sub-devices 60
Supported partition types None
Supported affinity domains (n/a)
Max work item dimensions 3
Max work item sizes 1024x1024x1024
Max work group size 256
Preferred work group size (AMD) 256
Max work group size (AMD) 1024
Preferred work group size multiple 64
Wavefront width (AMD) 64
Preferred / native vector sizes
char 4 / 4
short 2 / 2
int 1 / 1
long 1 / 1
half 1 / 1 (cl_khr_fp16)
float 1 / 1
double 1 / 1 (cl_khr_fp64)
Half-precision Floating-point support (cl_khr_fp16)
Denormals No
Infinity and NANs No
Round to nearest No
Round to zero No
Round to infinity No
IEEE754-2008 fused multiply-add No
Support is emulated in software No
Single-precision Floating-point support (core)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Correctly-rounded divide and sqrt operations Yes
Double-precision Floating-point support (cl_khr_fp64)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Address bits 64, Little-Endian
Global memory size 17163091968 (15.98GiB)
Global free memory (AMD) 16760832 (15.98GiB)
Global memory channels (AMD) 128
Global memory banks per channel (AMD) 4
Global memory bank width (AMD) 256 bytes
Error Correction support No
Max memory allocation 14588628168 (13.59GiB)
Unified memory for Host and Device No
Shared Virtual Memory (SVM) capabilities (core)
Coarse-grained buffer sharing Yes
Fine-grained buffer sharing Yes
Fine-grained system sharing No
Atomics No
Minimum alignment for any data type 128 bytes
Alignment of base address 1024 bits (128 bytes)
Preferred alignment for atomics
SVM 0 bytes
Global 0 bytes
Local 0 bytes
Max size for global variable 14588628168 (13.59GiB)
Preferred total size of global vars 17163091968 (15.98GiB)
Global Memory cache type Read/Write
Global Memory cache size 16384 (16KiB)
Global Memory cache line size 64 bytes
Image support Yes
Max number of samplers per kernel 26287
Max size for 1D images from buffer 134217728 pixels
Max 1D or 2D image array size 8192 images
Base address alignment for 2D image buffers 256 bytes
Pitch alignment for 2D image buffers 256 pixels
Max 2D image size 16384x16384 pixels
Max 3D image size 16384x16384x8192 pixels
Max number of read image args 128
Max number of write image args 8
Max number of read/write image args 64
Max number of pipe args 16
Max active pipe reservations 16
Max pipe packet size 1703726280 (1.587GiB)
Local memory type Local
Local memory size 65536 (64KiB)
Local memory syze per CU (AMD) 65536 (64KiB)
Local memory banks (AMD) 32
Max number of constant args 8
Max constant buffer size 14588628168 (13.59GiB)
Preferred constant buffer size (AMD) 16384 (16KiB)
Max size of kernel argument 1024
Queue properties (on host)
Out-of-order execution No
Profiling Yes
Queue properties (on device)
Out-of-order execution Yes
Profiling Yes
Preferred size 262144 (256KiB)
Max size 8388608 (8MiB)
Max queues on device 1
Max events on device 1024
Prefer user sync for interop Yes
Number of P2P devices (AMD) 0
P2P devices (AMD) <printDeviceInfo:147: get number of CL_DEVICE_P2P_DEVICES_AMD : error -30>
Profiling timer resolution 1ns
Profiling timer offset since Epoch (AMD) 0ns (Thu Jan 1 05:30:00 1970)
Execution capabilities
Run OpenCL kernels Yes
Run native kernels No
Thread trace supported (AMD) No
Number of async queues (AMD) 8
Max real-time compute queues (AMD) 8
Max real-time compute units (AMD) 60
printf() buffer size 4194304 (4MiB)
Built-in kernels (n/a)
Device Extensions cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_amd_device_attribute_query cl_amd_media_ops cl_amd_media_ops2 cl_khr_image2d_from_buffer cl_khr_subgroups cl_khr_depth_images cl_amd_copy_buffer_p2p cl_amd_assembly_program
NULL platform behavior
clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...) No platform
clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...) No platform
clCreateContext(NULL, ...) [default] No platform
clCreateContext(NULL, ...) [other] Success [AMD]
clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT) Success (1)
Platform Name AMD Accelerated Parallel Processing
Device Name gfx906:sramecc-:xnack-
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU) Success (1)
Platform Name AMD Accelerated Parallel Processing
Device Name gfx906:sramecc-:xnack-
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL) Success (1)
Platform Name AMD Accelerated Parallel Processing
Device Name gfx906:sramecc-:xnack-
So, now you can run OpenCL applications with your AMD GPU on your host system!
OpenCL on Docker for AMD GPUs
How about doing the same through Docker containers? Let's see how much it contrasts with NVIDIA GPUs.
Creating the Dockerfile
Create a new directory for your AMD GPU OpenCL project and move into it:
mkdir amd-opencl
cd amd-opencl
Use your favorite text editor (Vim/Nano or any other) to create the following Dockerfile and save it:
FROM ubuntu:20.04
ARG DEBIAN_FRONTEND=noninteractive
RUN apt-get update && apt-get -y upgrade \
&& apt-get install -y \
initramfs-tools \
apt-utils \
unzip \
tar \
curl \
xz-utils \
ocl-icd-libopencl1 \
opencl-headers \
clinfo \
;
ARG AMD_DRIVER=amdgpu-pro-21.10-1247438-ubuntu-20.04.tar.xz
ARG AMD_DRIVER_URL=https://drivers.amd.com/drivers/linux
RUN mkdir -p /tmp/opencl-driver-amd
WORKDIR /tmp/opencl-driver-amd
RUN curl --referer $AMD_DRIVER_URL -O $AMD_DRIVER_URL/$AMD_DRIVER; \
tar -Jxvf $AMD_DRIVER; \
cd amdgpu-pro-*; \
./amdgpu-install --opencl=legacy,rocr --headless --no-dkms -y; \
rm -rf /tmp/opencl-driver-amd;
RUN mkdir -p /etc/OpenCL/vendors && \
echo "libamdocl64.so" > /etc/OpenCL/vendors/amdocl64.icd
RUN ln -s /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 /usr/lib/libOpenCL.so
WORKDIR /
I had to add the initramfs-tools package since the amdgpu-dkms and amdgpu-dkms-firmware would still be installed. I kept it this way since in this case, the reboot and shutdown issues I mentioned earlier are irrelevant for containers.
Alternatively, you could still use the dpkg -i method in the Dockerfile.
Building the Dockerfile
So now that you have the necessary Dockerfile to get started, let's build it. I'm naming the image as amd-opencl :
docker build -t amd-opencl .
Add your username to the video &render groups
For the AMD GPU Docker container to work flawlessly, it is better you also add your username to the video and render groups:
sudo usermod -a -G video $LOGNAME
sudo usermod -a -G render $LOGNAMELaunch the OpenCL Container
Based on the new image that you just built, it's time to launch the new OpenCL container!
Permit your Linux username on the local machine to connect to the X windows display with the following command:
xhost +local:username
With the following command, you can now directly enter the local container's shell based on the new image just created:
docker run --rm -it --device=/dev/kfd --device=/dev/dri --group-add video --group-add render -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=$DISPLAY amd-openclVerify your OpenCL configuration on Docker
Now that you are inside the container shell, you can run the clinfo command to verify your OpenCL configuration just like you did on the bare-metal host system:
[email protected]:/# clinfo
Platform Name AMD Accelerated Parallel Processing
Number of devices 1
Device Name gfx906:sramecc-:xnack-
Device Vendor Advanced Micro Devices, Inc.
Device Vendor ID 0x1002
Device Version OpenCL 2.0
Driver Version 3246.0 (HSA1.1,LC)
Device OpenCL C Version OpenCL C 2.0
Device Type GPU
Device Board Name (AMD) Device 66af
Device Topology (AMD) PCI-E, 0a:00.0
Device Profile FULL_PROFILE
Device Available Yes
Compiler Available Yes
Linker Available Yes
Max compute units 60
SIMD per compute unit (AMD) 4
SIMD width (AMD) 16
SIMD instruction width (AMD) 1
Max clock frequency 1801MHz
Graphics IP (AMD) 9.0
Device Partition (core)
Max number of sub-devices 60
Supported partition types None
Supported affinity domains (n/a)
Max work item dimensions 3
Max work item sizes 1024x1024x1024
Max work group size 256
Preferred work group size (AMD) 256
Max work group size (AMD) 1024
Preferred work group size multiple 64
Wavefront width (AMD) 64
Preferred / native vector sizes
char 4 / 4
short 2 / 2
int 1 / 1
long 1 / 1
half 1 / 1 (cl_khr_fp16)
float 1 / 1
double 1 / 1 (cl_khr_fp64)
Half-precision Floating-point support (cl_khr_fp16)
Denormals No
Infinity and NANs No
Round to nearest No
Round to zero No
Round to infinity No
IEEE754-2008 fused multiply-add No
Support is emulated in software No
Single-precision Floating-point support (core)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Correctly-rounded divide and sqrt operations Yes
Double-precision Floating-point support (cl_khr_fp64)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Address bits 64, Little-Endian
Global memory size 17163091968 (15.98GiB)
Global free memory (AMD) 16760832 (15.98GiB)
Global memory channels (AMD) 128
Global memory banks per channel (AMD) 4
Global memory bank width (AMD) 256 bytes
Error Correction support No
Max memory allocation 14588628168 (13.59GiB)
Unified memory for Host and Device No
Shared Virtual Memory (SVM) capabilities (core)
Coarse-grained buffer sharing Yes
Fine-grained buffer sharing Yes
Fine-grained system sharing No
Atomics No
Minimum alignment for any data type 128 bytes
Alignment of base address 1024 bits (128 bytes)
Preferred alignment for atomics
SVM 0 bytes
Global 0 bytes
Local 0 bytes
Max size for global variable 14588628168 (13.59GiB)
Preferred total size of global vars 17163091968 (15.98GiB)
Global Memory cache type Read/Write
Global Memory cache size 16384 (16KiB)
Global Memory cache line size 64 bytes
Image support Yes
Max number of samplers per kernel 26287
Max size for 1D images from buffer 134217728 pixels
Max 1D or 2D image array size 8192 images
Base address alignment for 2D image buffers 256 bytes
Pitch alignment for 2D image buffers 256 pixels
Max 2D image size 16384x16384 pixels
Max 3D image size 16384x16384x8192 pixels
Max number of read image args 128
Max number of write image args 8
Max number of read/write image args 64
Max number of pipe args 16
Max active pipe reservations 16
Max pipe packet size 1703726280 (1.587GiB)
Local memory type Local
Local memory size 65536 (64KiB)
Local memory syze per CU (AMD) 65536 (64KiB)
Local memory banks (AMD) 32
Max number of constant args 8
Max constant buffer size 14588628168 (13.59GiB)
Preferred constant buffer size (AMD) 16384 (16KiB)
Max size of kernel argument 1024
Queue properties (on host)
Out-of-order execution No
Profiling Yes
Queue properties (on device)
Out-of-order execution Yes
Profiling Yes
Preferred size 262144 (256KiB)
Max size 8388608 (8MiB)
Max queues on device 1
Max events on device 1024
Prefer user sync for interop Yes
Number of P2P devices (AMD) 0
P2P devices (AMD) <printDeviceInfo:147: get number of CL_DEVICE_P2P_DEVICES_AMD : error -30>
Profiling timer resolution 1ns
Profiling timer offset since Epoch (AMD) 0ns (Thu Jan 1 00:00:00 1970)
Execution capabilities
Run OpenCL kernels Yes
Run native kernels No
Thread trace supported (AMD) No
Number of async queues (AMD) 8
Max real-time compute queues (AMD) 8
Max real-time compute units (AMD) 60
printf() buffer size 4194304 (4MiB)
Built-in kernels (n/a)
Device Extensions cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_amd_device_attribute_query cl_amd_media_ops cl_amd_media_ops2 cl_khr_image2d_from_buffer cl_khr_subgroups cl_khr_depth_images cl_amd_copy_buffer_p2p cl_amd_assembly_program
NULL platform behavior
clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...) No platform
clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...) No platform
clCreateContext(NULL, ...) [default] No platform
clCreateContext(NULL, ...) [other] Success [AMD]
clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT) Success (1)
Platform Name AMD Accelerated Parallel Processing
Device Name gfx906:sramecc-:xnack-
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU) Success (1)
Platform Name AMD Accelerated Parallel Processing
Device Name gfx906:sramecc-:xnack-
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL) Success (1)
Platform Name AMD Accelerated Parallel Processing
Device Name gfx906:sramecc-:xnack-
[email protected]:/#
And that's how you can run OpenCL applications inside an AMD GPU container!
Note that the xhost command being used for both the NVIDIA and AMD GPU containers is necessary every time you want to run them from a new terminal.
Conseils bonus
If you happen to own multiple GPUs on a single system and want to be specific about running the containers, you can do that as well. Read on.
NVIDIA GPUs
Based on how clinfo reports NVIDIA GPU information, they are classified on Docker as 0 , 1 , 2 and so on. So, say you have three NVIDIA GPUs and want the container to see only GPU 0(the first one), the corresponding command would have to be revised as:
docker run --rm -it --gpus 0 -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=$DISPLAY nvidia-openclAMD GPUs
Similarly, based on how clinfo reports AMD GPU information, they are classified on Docker as /dev/dri/card0 , /dev/dri/card1 , /dev/dri/card2 and so on. So, say you have three AMD GPUs and want the container to see only the first, use the following command instead:
docker run --rm -it --device=/dev/kfd --device=/dev/dri/card0 --device=/dev/dri/renderD128 --group-add video --group-add render -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=$DISPLAY amd-opencl
As per the above command, note that renderD128 corresponds to card0 , both of which relate to the first AMD GPU. On the same lines, renderD129 would correspond to card1 for the second AMD GPU and so on. The "renderD" value is incremental and therefore for the third GPU, it would be renderD130 corresponding to card2 . You can know these metrics in detail by running the ls -l /dev/dri/by-path commande.
Personal notes
Since the last 7 years, I've been actively involved with research that focuses on harnessing the computational power of Graphics Processing Units (GPUs) to understand biological phenomena.
For more than a year now, I've been working on Dockerizing CellModeller, which is my primary research software that I've been working with, to understand multicellularity and at the same time also contributing on its development as a software.
Even though the AMD GPU containerization process can be a bit tedious and tricky, I still liked the way it works without the need of an additional runtime package necessary for NVIDIA GPU containers.
For the entire endeavour, the following references were extremely helpful:
Congleton, N., 2020. Install OpenCL For The AMDGPU Open Source Drivers On Debian and Ubuntu . [online] LinuxConfig.org. Available at: https://linuxconfig.org/install-opencl-for-the-amdgpu-open-source-drivers-on-debian-and-ubuntu [Accessed June 23 2021].
My heartfelt thanks to all three authors!
There are so many applications out there on the accelerated computing domain that need OpenCL running on the backend for both GPU vendors. One good example is [email protected] and its specific GPU requirements.
Do let me know your thoughts about this intriguing adventure with OpenCL, GPUs, Linux and finally, Docker! If you have any feedback or suggestions, please let me know in the comment section below.