have activelaneid_u32 call ockl_activelane_u32 #1018
Conversation
|
Is there a test that checks that all/most of the __foo functions compile? If so it needs updates too. |
sorry, I should have already had a test ready. Let me cook one up, and update this PR. |
david-salinas
force-pushed
the
activelaneid
branch
from
f74a870
to
716cb85
Compare
david-salinas
requested review from
b-sumner,
scchan,
aaronenyeshi and
AlexVlx
tests/Unit/AMDGPU/activelaneid.cpp
Outdated
| int k1 = int_dist(rd); | ||
| int k2 = int_dist(rd); | ||
| if (k1 != k2) { | ||
| test[i * WAVEFRONT_SIZE + k1] ^= test[i * WAVEFRONT_SIZE + k2] ^= test[i * WAVEFRONT_SIZE + k1] ^= test[i * WAVEFRONT_SIZE + k2]; |
There was a problem hiding this comment.
Looking at this test (I based it on the existing "activelanecount.cpp" test) most of this set up of "test" can be scrapped, I think. We can just create a PFE, with a suitable extent, and an output grid/vector and call activelaneid to verify that we get correct/valid laneid values.
tests/Unit/AMDGPU/activelaneid.cpp
Outdated
|
|
||
| for (int i = 0; i < WAVEFRONT_SIZE; ++i) { | ||
| for (int j = 0; j < WAVEFRONT_SIZE; ++j) { | ||
| #if TEST_DEBUG |
There was a problem hiding this comment.
I would move this outside the entire loop nest
tests/Unit/AMDGPU/activelaneid.cpp
Outdated
| array<uint32_t, 1> output_GPU(GRID_SIZE); | ||
| extent<1> ex(GRID_SIZE); | ||
| parallel_for_each(ex, [&](index<1>& idx) [[hc]] { | ||
| output_GPU(idx) = __activelaneid_u32(); |
There was a problem hiding this comment.
Every lane is sure to be active. Control flow is needed to deactivate some lanes.
|
@david-salinas ping... |
|
I'm not sure I'm seeing the latest version of the test code. I agree it can be small, but the kernel should have some control flow so that we can see that activelaneid is reporting the correct ID on lanes that are truly active. |
|
I don't see any update to the code either |
|
Sorry, I was about to push an update test, when I realized I need to add some control flow to deactivate some lanes. I should have an update shortly. |
david-salinas
force-pushed
the
activelaneid
branch
from
716cb85
to
c8d80ea
Compare
for #1011