[SYCL][E2E] Fix invalid discard_write in atomic_memory_order_seq_cst #10900
Conversation
The atomic_memory_order_seq_cst incorrectly uses discard_write access after initializing its memory. This causes failures on some hardware. This commit fixes this invalid access mode. Signed-off-by: Larsen, Steffen <[email protected]>
| @@ -40,7 +37,7 @@ void check(queue &q, buffer<int, 2> &res_buf, size_t N_iters) { | |||
| q.submit([&](handler &cgh) { | |||
| auto res = res_buf.template get_access<access::mode::read>(cgh); | |||
| auto checked = | |||
| checked_buf.template get_access<access::mode::discard_write>(cgh); | |||
There was a problem hiding this comment.
Is it a temporary problem with discard_write caused by, e.g., some bug in GPU driver, or we should simply always use write instead of discard_write in such scenarios? If this is temporary problem in GPU driver, should we wait until GPU driver is updated on CI machines?
There was a problem hiding this comment.
I believe it is more of a misunderstanding of what discard_write does. It is now deprecated, but is the same as write and no_init. The problem here is that since we tell the device allocation that we don't care about the initial value of the memory, it is free to write back whatever value we don't write, while for this test we actually initialize the memory ourselves earlier.
Signed-off-by: Larsen, Steffen <[email protected]>
|
HIP failures reported in #10460. |
The atomic_memory_order_seq_cst incorrectly uses discard_write access after initializing its memory. This causes failures on some hardware. This commit fixes this invalid access mode.