[AMD GPU] Issue with dispatch dimension beyon 32 Qubits #993
Description
Describe the issue
Hi all,
We're running QSim on AMD GPUs, there is an issue with the dispatch dimensions that occurs beyond the 31 Qubits, which goes beyond the device limit. Please see more details below:
This is running on MI300x (192 GB Memory), ROCm 7.1.0.
When running more than 31 Qubits we get:
CUDA error: invalid configuration argument
[PATH] ... /statespace_cuda.h:93
When running using AMD_LOG_LEVEL=5, we get the following debug info:
:3:hip_module.cpp :744 : 3639998735920 us: [pid:627351 tid: 0x73b21e4f6740] hipLaunchKernel ( 0x73b10daa38a0, {8388608,1,1}, {512,1,1}, 0x7ffebbdd0d90, 4096, stream:<null> )
:3:hip_module.cpp :745 : 3639998735931 us: [pid:627351 tid: 0x73b21e4f6740] hipLaunchKernel: Returned hipErrorInvalidConfiguration : : duration: 11 us
This breaks because of this condition in the dispatch dimension:
I believe the solution is simply change the dispatch dimensions for a larger number of qubits.
If anyone is interested in working on this, I may be able to provide access to the GPU for debugging. Reach out to me.
Tell us how to reproduce the issue
Run any circuit on AMD GPU with more than 31 qubits
Tell us the version of qsim or qsimcirq (if relevant)
No response
Tell us the computing environment (if relevant)
No response