[GPU] enable simd16 version for convolution_gpu_mmad_b_fs_yx_fsv32 #32501
Conversation
michal-miotk
changed the title
michal-miotk
changed the title
Lyamin-Roman
left a comment
Lyamin-Roman
left a comment
There was a problem hiding this comment.
LGTM, but please add tests
src/plugins/intel_gpu/src/kernel_selector/cl_kernels/convolution_gpu_mmad_b_fs_yx_fsv32.cl
Outdated
Show resolved
Hide resolved
| os_is_yx_isa8_osv8_isv4, ///< format for weights for MMAD convolution | ||
| os_is_zyx_isa8_osv8_isv4, ///< format for weights for MMAD convolution | ||
| os_is_yx_isa8_osv16_isv4, ///< format for weights for fully connected MMAD | ||
| os_is_zyx_isa8_osv16_isv4, ///< format for weights for fully connected MMAD |
There was a problem hiding this comment.
random spot) could you check why onednn convolution kernel is not chosen? For GPU with XMX, our expectation is to use OneDNN convolutions, instead of cldnn convolutions.
There was a problem hiding this comment.
because weights are u8
There was a problem hiding this comment.
oh I see. I think we need to file a ticket to onednn team for supporting it. Performance won't be very good with cldnn kernel..
There was a problem hiding this comment.
ticket created: https://jira.devtools.intel.com/browse/MFDNN-14355
|
|
||
| class convolution_random_smoke_test : public testing::TestWithParam<convolution_random_test_all_params> {}; | ||
| class convolution_random_fsv32_test : public testing::TestWithParam<convolution_random_test_all_params> {}; | ||
|
|
There was a problem hiding this comment.
random spot)
- How does it perform on old platforms with SIMD16? If your new implementation performs good, I'd suggest to remove simd8 implementation. That will simplify the code.
- Does this test code validate simd16 version on old platforms? We don't have precommit test for the new platforms.
There was a problem hiding this comment.
on model from ticket on intelUHD770 it is 34fps(simd8) vs 39fps(simd16)
There was a problem hiding this comment.
on model from ticket on intel iris (DUT2069-ADLP) it is 36fps(simd8) vs 42fps(simd16)
There was a problem hiding this comment.
on model from ticket on A770 it is (simd8) 438fps(simd8) vs 418fps(simd16)
…penvinotoolkit#32501) ### Details: - new platforms not support simd8 (LNL, BMG) ### Tickets: - 174772
3 participants
Details:
Tickets: