W1034
W1034 — Code on the built-inTopology::GPUuses a tensor that lives in host memory (CPU_DRAM). The built-in GPU lists host memory as visible, so this compiles. A matmul sent to cuBLAS has its operands copied to the device, but a kernel the compiler generates is given the host pointer as it is, and what it does with it depends on the driver: it may reach host memory over the link, one access at a time, or fault. Transfer the tensor toMemory::GPU_HBMbefore the region. Reported once per tensor per region. A later version may make this an error.
No example yet
No fixture in the test suite asserts this code, so there is no program here that is known to trigger it. That is a gap in our coverage rather than a statement about the code.
Contributing one is a good first change — add a fixture under tests/ with a CHECK line naming W1034, and this page picks it up.
Related
A tour of VxThe chapter covering the rule this code enforces.
All diagnosticsEvery code the compiler can emit, grouped by stage.
W1033The loop of a
spawn region runs on one thread although each pass writes only memory of its own, so its passes could run at the same time. The warning points at what stops it: a call the compiler cannot show writes nothing, inline MLIR, a read of a row another pass writes, bounds not known when compiling, and the like.