[SYCL] Add new tests for int8 and bfloat16 automatic transpose and VNNI transform #1415
Conversation
yubingex007-a11y
commented
Nov 24, 2022
yubingex007-a11y
commented
- Add new tests for automatic transpose and VNNI transform
- fix a bug in bfloat16's testcase
- add testcases for subB(int8, colmajor/rowmajor) and subA(int8, colmajor/rowmajor)
yubingex007-a11y
requested review from
v-klochkov
and removed request for
a team
yubingex007-a11y
changed the title
yubingex007-a11y
changed the title
|
@dkhaldi i fixed the bug in testcases itself and provide more cases for int8. |
|
Sorry, I am on vacation. Adding intel/llvm-reviewers-runtime again. |
| using namespace sycl::ext::oneapi::experimental::matrix; | ||
| using bfloat16 = sycl::ext::oneapi::experimental::bfloat16; | ||
|
|
||
| #define SG_SZ 8 |
There was a problem hiding this comment.
Please transform the tests to follow the existing https://github.com/intel/llvm-test-suite/blob/intel/SYCL/Matrix/joint_matrix_bfloat16.cpp
SG size needs to be 16 for PVC and SPR.
sg size = 8 test needs to be in XMX8 folder
| #define SG_SZ 8 | ||
|
|
||
| #define TM 8 | ||
| #define TN 8 |
| bool res = true; | ||
| for (int i = 0; i < MATRIX_M; i++) { | ||
| for (int j = 0; j < MATRIX_N; j++) { | ||
| if (C[i][j] != D[i][j]) { |
There was a problem hiding this comment.
if ((fabs(C[i][j]) - fabs(D[i][j])) > BF16_EPSILON)
see https://github.com/intel/llvm-test-suite/blob/intel/SYCL/Matrix/joint_matrix_bfloat16_impl.hpp
| using namespace sycl::ext::oneapi::experimental::matrix; | ||
| using bfloat16 = sycl::ext::oneapi::experimental::bfloat16; | ||
|
|
||
| #define SG_SZ 8 |
| @@ -0,0 +1,170 @@ | |||
| //==-------joint_matrix_bfloat16_row_major.cpp - DPC++ joint_matrix--------==// | |||
There was a problem hiding this comment.
i really has question about naming the testcase here.
SYCL/Matrix/joint_matrix_bfloat16_col_majorA.cpp(A is colmajor and B is colmajor, we can test colmajor=>packed and colmajor=>rowmajor)
i suggest change it into joint_matrix_bfloat16_colmajorA_colmajorB.cpp
joint_matrix_bfloat16_row_major.cpp => joint_matrix_bfloat16_rowmajorA_rowmajorB.cpp(we can test colmajor=>packed)
with joint_matrix_bfloat16_colmajorA_colmajorB.cpp and joint_matrix_bfloat16_rowmajorA_rowmajorB.cpp, we can test all the auto-vnnitransform cases.
| @@ -0,0 +1,170 @@ | |||
| //==----- joint_matrix_bfloat16_col_major.cpp - DPC++ joint_matrix---------==// | |||
| bool res = true; | ||
| for (int i = 0; i < MATRIX_M; i++) { | ||
| for (int j = 0; j < MATRIX_N; j++) { | ||
| if (C[i][j] != D[i][j]) { |
yubingex007-a11y
force-pushed
the
row-col-major-int8-bfloat16
branch
from
6599813 to
98f24ea
Compare
yubingex007-a11y
requested review from
dkhaldi
and removed request for
a team
|
Sorry, but I just realized after you made the change that we will never get XMX8 to support transpose and VNNI transform. This will be only PVC feature (XMX16). So we should put back the three tests you have into one file. There is no need to split it since we won't have them in XMX8 folder. |
maybe let's keep the code style with other cases? we may have other arch in the future. |
Fine by me |
|
ping? @intel/llvm-reviewers-runtime |
| // RUN: %clangxx -fsycl %s -o %t.out | ||
| // RUN: %CPU_RUN_PLACEHOLDER %t.out | ||
| // RUN: %GPU_RUN_PLACEHOLDER %t.out | ||
| // CHECK: passed |
There was a problem hiding this comment.
This only has an effect if you run the output through FileCheck. It would be easier to just have the main functions return a non-0 value if they fail.
There was a problem hiding this comment.
maybe later we change all of it. other tesctcases have such issue as well.
There was a problem hiding this comment.
If this is indeed a problem for the other tests, I am okay with merging it as-is, but it should be addressed ASAP. Note that anything that can cause "failed" to be printed will not currently be considered a failure by the test system for these tests.
steffenlarsen
changed the title
…NI transform (intel#1415) - Add new tests for automatic transpose and VNNI transform - fix a bug in bfloat16's testcase - add testcases for subB(int8, colmajor/rowmajor) and subA(int8, colmajor/rowmajor)
…NI transform (intel/llvm-test-suite#1415) - Add new tests for automatic transpose and VNNI transform - fix a bug in bfloat16's testcase - add testcases for subB(int8, colmajor/rowmajor) and subA(int8, colmajor/rowmajor)
