Repository navigation
[BUG] Binary release with CUDA failing in Google Colab environment #3142
Description
Activity
Thanks for the link to your notebook. It really helps with the debugging of this issue.
Looks like the nvidia runtime is looking for a different nvrtc-buildtin library than provided by the installer. This is strange because we have tested this sceanario in our CI and it has been working correctly. I will need to investigate why this is happening. in the meantime you can work around this bug by renaming the
/opt/arrayfire/lib64/libnvrtc-builtins.so to /opt/arrayfire/lib64/libnvrtc-builtins.so.11.1@umar456 thanks for checking this out! It looks like CUDA 11.* is not supported in the binary installer https://arrayfire.s3.amazonaws.com/3.8.0/ArrayFire-v3.8.0_Linux_x86_64.sh
You can see the error in my colab notebook:[platform][1623062663][000247] [ ../src/backend/cuda/device_manager.cpp:466 ] CUDA Driver supports up to CUDA 11.3 ArrayFire CUDA Runtime 11.0 [platform][1623062663][000247] [ ../src/backend/cuda/device_manager.cpp:451 ] CUDA driver version(11.3) not part of the CudaToDriverVersion array. Please create an issue or a pull request on the ArrayFire repository to update the CudaToDriverVersion variable with this version of the CUDA runtime.
Today I tried building from source in another colab and it worked, though it took hours to build everything so this is not practical to use in colab.
As I see you have different CUDA 10 and 11 distributions, maybe you can just build one with CUDA 11 support for Linux?
http://arrayfire.s3.amazonaws.com/index.html#!/3.8.0%2FHaving nightly builds would also be awesome: http://arrayfire.s3.amazonaws.com/index.html#!/nightly%2F
That message is not an error. It is targeted towards ArrayFire developers to update the minimum driver requirements for the latest CUDA toolkits. CUDA maintains a forward compatibility and therefore older versions of the toolkit would work and do not need intervention. We use the toolkit versions to display useful error message if the CUDA driver doesn't support a certain version of the CUDA toolkit. Since we do not know the minimum driver versions of the future versions of CUDA we allow the program to execute without those checks if the Toolkit version is not found in the database. In those cases we add the message to only the AF_TRACE logs. If you remove the AF_TRACE environment variable, you will not see that message. The toolkit version has also been updated in the latest master and should be part of the next release(see #3125).
That message is not an error. It is targeted towards ArrayFire developers to update the minimum driver requirements for the latest CUDA toolkits. CUDA maintains a forward compatibility and therefore older versions of the toolkit would work and do not need intervention. We use the toolkit versions to display useful error message if the CUDA driver doesn't support a certain version of the CUDA toolkit. Since we do not know the minimum driver versions of the future versions of CUDA we allow the program to execute without those checks if the Toolkit version is not found in the database. In those cases we add the message to only the AF_TRACE logs. If you remove the AF_TRACE environment variable, you will not see that message. The toolkit version has also been updated in the latest master and should be part of the next release(see #3125).
It makes sense, thanks! Adding this line fixed the issue in my colab:
cp /opt/arrayfire/lib64/libnvrtc-builtins.so /opt/arrayfire/lib64/libnvrtc-builtins.so.11.1
I am setting up google colab playground to run some experiments.
Here is the notebook with helloworld example which fails when running with GPU kernel, while works fine on CPU:
https://colab.research.google.com/drive/1xUVsyKJgD7pvPDRnhliYzzqjcctpK7E7?usp=sharing
Description
Did you build ArrayFire yourself or did you use the official installers
Using binary installer: https://arrayfire.s3.amazonaws.com/3.8.0/ArrayFire-v3.8.0_Linux_x86_64.sh
Which backend is experiencing this issue? (CPU, CUDA, OpenCL)
CUDA throws exception
CPU works as expected
Do you have a workaround?
No
Can the bug be reproduced reliably on your system?
Yes
Run your executable with AF_TRACE=all and AF_PRINT_ERRORS=1 environment
variables set.
The relevant error is likely
Unable to open /root/.arrayfire/KER11230367656483596278_CU_60_AF_38.bin for Tesla P100-PCIE-16GB
You can see full trace in the notebook.
Screenshot or terminal output of the results
N/A, check colab link.
Reproducible Code and/or Steps
Easily reproducible, try running my shared colab notebook.
To ensure you are running on GPU, select Change Runtime -> Change runtime type -> GPU.
System Information
Run one of the following commands based on your OS
No LSB modules are available. Distributor ID: Ubuntu Description: Ubuntu 18.04.5 LTS Release: 18.04 Codename: bionic Failed to initialize NVML: Driver/library version mismatch rocm-smi not found. Number of platforms 1 Platform Name NVIDIA CUDA Platform Vendor NVIDIA Corporation Platform Version OpenCL 3.0 CUDA 11.3.101 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_3d_image_writes cl_khr_byte_addressable_store cl_khr_icd cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_nv_create_buffer cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_nv_kernel_attribute cl_khr_device_uuid Platform Host timer resolution 0ns Platform Extensions function suffix NV Platform Name NVIDIA CUDA Number of devices 1 Device Name Tesla P100-PCIE-16GB Device Vendor NVIDIA Corporation Device Vendor ID 0x10de Device Version OpenCL 3.0 CUDA Driver Version 465.27 Device OpenCL C Version OpenCL C 1.2 Device Type GPU Device Topology (NV) PCI-E, 00:00.4 Device Profile FULL_PROFILE Device Available Yes Compiler Available Yes Linker Available Yes Max compute units 56 Max clock frequency 1328MHz Compute Capability (NV) 6.0 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 Max sub-groups per work group 0 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 17071734784 (15.9GiB) Error Correction support Yes Max memory allocation 4267933696 (3.975GiB) Unified memory for Host and Device No Integrated memory (NV) No Shared Virtual Memory (SVM) capabilities (core) Coarse-grained buffer sharing Yes Fine-grained buffer sharing No Fine-grained system sharing No Atomics No Minimum alignment for any data type 128 bytes Alignment of base address 4096 bits (512 bytes) Preferred alignment for atomics SVM 0 bytes Global 0 bytes Local 0 bytes Max size for global variable 0 Preferred total size of global vars 0 Global Memory cache type Read/Write Global Memory cache size 1376256 (1.312MiB) Global Memory cache line size 128 bytes Image support Yes Max number of samplers per kernel 32 Max size for 1D images from buffer 268435456 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 Max number of read/write image args 0 Max number of pipe args 0 Max active pipe reservations 0 Max pipe packet size 0 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 (on host) Out-of-order execution Yes Profiling Yes Queue properties (on device) Out-of-order execution No Profiling No Preferred size 0 Max size 0 Max queues on device 0 Max events on device 0 Prefer user sync for interop No Profiling timer resolution 1000ns Execution capabilities Run OpenCL kernels Yes Run native kernels No Sub-group independent forward progress No Kernel execution timeout (NV) No Concurrent copy and kernel execution (NV) Yes Number of async copy engines 2 IL version printf() buffer size 1048576 (1024KiB)