-
Notifications
You must be signed in to change notification settings - Fork 130
[ESIMD] Add tests for new esimd::xmx:dpas API #1281
Conversation
/verify with intel/llvm#6834 |
FYI I am seeing the same windows failures in intel/llvm#6532. |
Signed-off-by: Vyacheslav N Klochkov <[email protected]>
Signed-off-by: Vyacheslav N Klochkov <[email protected]>
618363b
to
b9bbed9
Compare
Signed-off-by: Vyacheslav N Klochkov <[email protected]>
/verify with intel/llvm#6834 |
CI reported these fails, unrelated to 2 new tests: [2022-09-22T00:34:31.113Z] Failed Tests (3): |
Passed &= tests<8, 8, s2, s4>(Q, Print); | ||
Passed &= tests<8, 1, s2, s8>(Q, Print); | ||
Passed &= tests<8, 4, s8, s4>(Q, Print); | ||
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For follow-up in later PR: please test other supported types
// Test unsigned 2-bit integers./ | ||
Passed &= tests<8, 8, u2, u2>(Q, Print); | ||
Passed &= tests<8, 4, u2, u2>(Q, Print); | ||
// TODO: enable this case when the problem with simd constructor |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is there internal tracker on this problem?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, I created one.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||
// | ||
//===----------------------------------------------------------------------===// | ||
// REQUIRES: gpu-intel-pvc || esimd_emulator |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For follow-up in later PR: dg2 should also work, right?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
There will be more tests with ExecutionSize=8 (testing ACM/dg2)
bool Print = argc > 1 && std::string(argv[1]) == "-debug"; | ||
bool Passed = true; | ||
|
||
// Test unsigned 2-bit integers./ |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
typo - 2-bit integer -> half float
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Will fix in next PR adding more tests. Thank you.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Fixed here: #1291
|
||
// This test verifies DPAS support for 2,4,8-bit integers. | ||
|
||
#include "dpas_common.hpp" |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Nit: just curious - why not simply template by element types in a single test source instead of including common header?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
just wanted to have different tests for different groups of types: int separate, fp16 - separate, etc.
Otherwise, it would be too big test running too long.
dpas_int.cpp already has too many test cases in it. And it needs some more extension to test the cases when precision is not specified.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM, but a follow-up PR is desirable.
* [ESIMD] Add tests for new esimd::xmx:dpas API Signed-off-by: Vyacheslav N Klochkov <[email protected]>
[ESIMD] Remove invalid rules in some of test .xmls
…1281) * [ESIMD] Add tests for new esimd::xmx:dpas API Signed-off-by: Vyacheslav N Klochkov <[email protected]>
Signed-off-by: Vyacheslav N Klochkov [email protected]