Describe the issue:
When compiling Grid and running with a single GPU, running e.g. Benchmark_ITT gives the error:
accelerator_barrier(): Cuda error invalid configuration argument
Digging into this, this is due to line 137 of Grid/threads/Accelerator.h
dim3 cu_blocks ((num1+nt-1)/nt,num2,1); \
For reasons I haven't dug deep enough to understand, when running with 1 GPU, then (num1+nt-1)/nt (or in the specific case that fails—called from WilsonKernelsImplementation.h—(sz+nt-1)/nt) gets set to zero, which isn't a valid block count.
As a workaround, changing line 137 to
dim3 cu_blocks ((num1+nt-1)/nt == 0 ? 1 : (num1+nt-1)/nt,num2,1); \
allows the code to run correctly.
Code example:
N/A
Target platform:
Tested on Grace Hopper Arm+H100, Leicester Arm+A100, and AMD Rome + A100 in Swansea.
Configure options:
../configure --enable-comms=none --enable-simd=GPU --enable-accelerator=cuda CXX=nvcc --disable-zmobius --disable-gparity 'CXXFLAGS=-g -gencode arch=compute_90,code=sm_90 -std=c++17 -DEIGEN_DONT_VECTORIZE'
Describe the issue:
When compiling Grid and running with a single GPU, running e.g. Benchmark_ITT gives the error:
Digging into this, this is due to line 137 of Grid/threads/Accelerator.h
For reasons I haven't dug deep enough to understand, when running with 1 GPU, then (num1+nt-1)/nt (or in the specific case that fails—called from WilsonKernelsImplementation.h—(sz+nt-1)/nt) gets set to zero, which isn't a valid block count.
As a workaround, changing line 137 to
allows the code to run correctly.
Code example:
Target platform:
Tested on Grace Hopper Arm+H100, Leicester Arm+A100, and AMD Rome + A100 in Swansea.
Configure options: