Improve setting launching bounds on preset kernels #64
Conversation
tests/gpu/kernel_preset_bounds.cpp
Outdated
| #include "gpu_common.h" | ||
|
|
||
| __global__ __attribute__((annotate("jit"))) | ||
| __launch_bounds__(128, 4) void kernel() { |
There was a problem hiding this comment.
I don't recall the semantics. Is it always mandatory to pass the second value? If it is NOT you should add a test for such a case. Cause I am guessing this:
assert(MetadataNode->getNumOperands() == 3);
will fail
There was a problem hiding this comment.
Further, I would consider having the bounds as <1, 1>. To match the call site and indirectly test case in which proteus does not call the function correctly (for whatever reason).
There was a problem hiding this comment.
No, it's not mandatory. The assert will not fail because this is the format of the metadata.
There was a problem hiding this comment.
I'm using different launch bounds than the actual launch dimensions because I can visually check the IR if the optimization happened. Checking will be automated in the future.
- Replace metadata if already set (CUDA) - Add test kernel_preset_bounds
ggeorgakoudis
force-pushed
the
test-kernel-with-preset-bounds
branch
from
75ec319 to
9858438
Compare
|
Rebased ☝️ |
|
@koparasy @davidbeckingsale Given the discussion in slack on the necessity to override launch bounds for HIP, can we approve this PR? |
Closes #27