xe: grp_gemm: Add microkernel based implementation of grouped GEMM #4505
Conversation
Only supports power of 2 shapes and f16xf16:f32 types
- Adjust global work size calculation for accurate group partitioning - Correct m_all calculation to divide by num_groups for per-group M
github-actions
bot
added
platform:gpu-intel
| @@ -0,0 +1,168 @@ | |||
|
|
|||
| #include "dnnl_test_common.hpp" | |||
There was a problem hiding this comment.
what is missing from benchdnn, so that this test is required..?
There was a problem hiding this comment.
Nothing is missing. I want more control over how the data is initialized so that its easier to test/debug new configurations. I will probably remove this in the future.
There was a problem hiding this comment.
I want more control over how the data is initialized
Could you clarify, so that I could possibly make it better/easier?
I will probably remove this in the future
I agree, we would need this removed prior merging if it is covered by benchdnn..
There was a problem hiding this comment.
It's difficult to initialize the input data with known values to identify incorrect results. For example I want to find out why I am getting an incorrect value at a certain location in the destination buffer. With my own tests I can make my inputs to be known values to debug. I don't know of a way to do that with benchdnn.
mzhukova
force-pushed
the
mzhukova/main/poc-grouped-mem
branch
2 times, most recently
from
be4fed9 to
0ea335d
Compare
3 participants
This pull request introduces support for a new "grouped micro GEMM" (General Matrix Multiply) implementation, along with corresponding kernel code, and enhancements to the testing utilities to support grouped memory.