You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
To split work in a multi-GPU setting, we need to find sets of equal / compatible GPUs on a system.
On a system with 4x Nvidia RTX 3090, sycl::platform::get_platforms() returns four distinct platforms that stringify to sycl::platform(vendor="NVIDIA Corporation", name="NVIDIA CUDA BACKEND"), but compare unequal with operator== and do not produce the same hash. I would expect all devices to share a platform in this case.
Comparing the backend for finding a set of equal GPUs is not enough either, since DPC++ produces the same backend enumerator at least for the Intel(R) OpenCL and Intel(R) FPGA Emulation Platform for OpenCL(TM) platforms which clearly do not originate from a multi-GPU situation.
Describe the bug
To split work in a multi-GPU setting, we need to find sets of equal / compatible GPUs on a system.
On a system with 4x Nvidia RTX 3090,
sycl::platform::get_platforms()
returns four distinct platforms that stringify tosycl::platform(vendor="NVIDIA Corporation", name="NVIDIA CUDA BACKEND")
, but compare unequal withoperator==
and do not produce the same hash. I would expect all devices to share a platform in this case.Comparing the
backend
for finding a set of equal GPUs is not enough either, since DPC++ produces the same backend enumerator at least for theIntel(R) OpenCL
andIntel(R) FPGA Emulation Platform for OpenCL(TM)
platforms which clearly do not originate from a multi-GPU situation.To reproduce
Environment
Additional context
No response
The text was updated successfully, but these errors were encountered: