My Project
Unified Backend

Introduction

The Unified backend was introduced in ArrayFire with version 3.2. While this is not an independent backend, it allows the user to switch between the different ArrayFire backends (CPU, CUDA and OpenCL) at runtime.

Compiling with Unified

The steps to compile with the unified backend are the same as compiling with any of the other backends. The only change being that the executable needs to be linked with the af library (libaf.so (Linux), libaf.dylib (OSX), af.lib (Windows)).

Check the Using with Linux, OSX, Windows for more details.

To use with CMake, use the ArrayFire_Unified_LIBRARIES variable.

Using the Unified Backend

The Unified backend will try to dynamically load the backend libraries. The priority of backends is CUDA -> OpenCL -> CPU

The most important aspect to note here is that all the libraries the ArrayFire libs depend on need to be in the environment paths

  • LD_LIBRARY_PATH -> Linux, Unix, OSX
  • DYLD_LIBRARY_PATH -> OSX
  • PATH -> Windows

If any of the libs are missing, then the library will fail to load and the backend will be marked as unavailable.

Optionally, The ArrayFire libs may be present in AF_PATH or AF_BUILD_PATH environment variables if the path is not in the system paths. These are treated as fallback paths in case the files are not found in the system paths. However, all the other upstream libraries for ArrayFire libs must be present in the system path variables shown above.

Special Mention: CUDA NVVM

For the CUDA backend, ensure that the CUDA NVVM libs/dlls are in the path. These can be easily missed since CUDA installation does not add the paths by default.

On Linux and OSX, add /usr/local/cuda/nvvm/(lib or lib64) to LD_LIBRARY_PATH or DYLD_LIBRARY_PATH.

On Windows, you can set up a post build event that copys the NVVM dlls to the executable directory by using the following commands:

echo copy "$(CUDA_PATH)\nvvm\bin\nvvm64*.dll" "$(OutDir)"
copy "$(CUDA_PATH)\nvvm\bin\nvvm64*.dll" "$(OutDir)"
if errorlevel 1 (
echo "CUDA NVVM DLLs copy failed due to missing files."
exit /B 0
)

This ensures that the NVVM DLLs are copied if present, but does not fail the build if the copy fails. This is how ArrayFire ships it's examples.

The other option is to set %CUDA_PATH%/nvvm/bin in the PATH environment variable.

Switching Backends

The af_backend enum stores the possible backends. To select a backend, call the af::setBackend function as shown below.

af::setBackend(AF_BACKEND_OPENCL); // Sets CUDA as current backend

To get the count of the number of backends available (the number of libaf* backend libraries loaded successfully), call the af::getBackendCount function.

Example

This example is shortened form of basic.cpp.

#include <arrayfire.h>
void testBackend()
{
}
int main()
{
try {
printf("Trying CPU Backend\n");
testBackend();
} catch (af::exception& e) {
printf("Caught exception when trying CPU backend\n");
fprintf(stderr, "%s\n", e.what());
}
try {
printf("Trying CUDA Backend\n");
testBackend();
} catch (af::exception& e) {
printf("Caught exception when trying CUDA backend\n");
fprintf(stderr, "%s\n", e.what());
}
try {
printf("Trying OpenCL Backend\n");
testBackend();
} catch (af::exception& e) {
printf("Caught exception when trying OpenCL backend\n");
fprintf(stderr, "%s\n", e.what());
}
return 0;
}

This output would be:

Trying CPU Backend
ArrayFire v3.2.0 (CPU, 64-bit Linux, build fc7630f)
[0] Intel: Intel(R) Core(TM) i7-4770K CPU @ 3.50GHz Max threads(8)
af::randu(5, 4)
[5 4 1 1]
    0.0000     0.2190     0.3835     0.5297
    0.1315     0.0470     0.5194     0.6711
    0.7556     0.6789     0.8310     0.0077
    0.4587     0.6793     0.0346     0.3834
    0.5328     0.9347     0.0535     0.0668

Trying CUDA Backend
ArrayFire v3.2.0 (CUDA, 64-bit Linux, build fc7630f)
Platform: CUDA Toolkit 7.5, Driver: 355.11
[0] Quadro K5000, 4093 MB, CUDA Compute 3.0
af::randu(5, 4)
[5 4 1 1]
    0.7402     0.4464     0.7762     0.2920
    0.9210     0.6673     0.2948     0.3194
    0.0390     0.1099     0.7140     0.8109
    0.9690     0.4702     0.3585     0.1541
    0.9251     0.5132     0.6814     0.4452

Trying OpenCL Backend
ArrayFire v3.2.0 (OpenCL, 64-bit Linux, build fc7630f)
[0] NVIDIA  : Quadro K5000
-1- INTEL   : Intel(R) Core(TM) i7-4770K CPU @ 3.50GHz
af::randu(5, 4)
[5 4 1 1]
    0.4107     0.0081     0.6600     0.1046
    0.8224     0.3775     0.0764     0.8827
    0.9518     0.3027     0.0901     0.1647
    0.1794     0.6456     0.5933     0.8060
    0.4198     0.5591     0.1098     0.5938

Dos and Don'ts

It is very easy to run into exceptions if you are not careful with the switching of backends.

Don't: Do not use arrays between different backends

ArrayFire checks the input arrays to functions for mismatches with the active backend. If an array created on one backend, but used when another backend is set to active, an exception with code 503 (AF_ERR_ARR_BKND_MISMATCH) is thrown.

#include <arrayfire.h>
int main()
{
try {
af::array A = af::randu(5, 5);
af::array B = af::constant(10, 5, 5);
af::array C = af::matmul(A, B); // This will throw an exception
} catch (af::exception& e) {
fprintf(stderr, "%s\n", e.what());
}
return 0;
}

Do: Use a naming scheme to track arrays and backends

We recommend that you use a technique to track the arrays on the backends. One suggested technique would be to use a suffix of _cpu, _cuda, _opencl with the array names. So an array created on the CUDA backend would be named myarray_cuda.

If you have not used the af::setBackend function anywhere in your code, then you do not have to worry about this as all the arrays will be created on the same default backend.

Don't: Do not use custom kernels (CUDA/OpenCL) with the Unified backend

This is another area that is a no go when using the Unified backend. It not recommended that you use custom kernels with unified backend. This is mainly becuase the Unified backend is meant to be ultra portable and should use only ArrayFire and native CPU code.

af::matmul
AFAPI array matmul(const array &lhs, const array &rhs, const matProp optLhs=AF_MAT_NONE, const matProp optRhs=AF_MAT_NONE)
Matrix multiply of two arrays.
af::info
AFAPI void info()
af::constant
array constant(T val, const dim4 &dims, const dtype ty=(af_dtype) dtype_traits< T >::ctype)
af::array
A multi dimensional data container.
Definition: array.h:32
af_print
#define af_print(...)
Definition: util.h:140
AF_BACKEND_OPENCL
OpenCL Compute Backend.
Definition: defines.h:372
af::randu
AFAPI array randu(const dim4 &dims, const dtype ty=f32)
af::exception
Definition: exception.h:24
arrayfire.h
af::copy
AFAPI void copy(array &dst, const array &src, const index &idx0, const index &idx1=span, const index &idx2=span, const index &idx3=span)
Copy the values of an input array based on index.
AF_BACKEND_CUDA
CUDA Compute Backend.
Definition: defines.h:371
af::exception::what
virtual const char * what() const
Definition: exception.h:45
af::setBackend
AFAPI void setBackend(const Backend bknd)
AF_BACKEND_CPU
CPU a.k.a sequential algorithms.
Definition: defines.h:370