Conversation
github-actions
bot
added
platform:gpu-intel
dyoussif
changed the title
dyoussif
force-pushed
the
dyoussif/gemm_ref_scales
branch
from
a3665fa to
b1d324b
Compare
|
make test |
| c_stride_d3); | ||
| #else | ||
| #if NDIMS == 5 | ||
| scale_off = DST_SCALE_OFF( |
There was a problem hiding this comment.
It looks like DST_SCALE_OFF is only used here so either we should fix that macro or drop it from src/gpu/intel/include/types_specific.h.
| scale_off = scale_off_dst(d0 % DST_D0, m, n, groupSize); | ||
| #else | ||
| scale_off = DST_SCALE_OFF(m, n, 0, 0, 0, groupSize, 1); | ||
| scale_off = scale_off_dst(m, n, groupSize); |
There was a problem hiding this comment.
nit: we could normalize these calls to scale_off = scale_off_dst(n, m, d0 % DST_D0, d1 %DST_D1, ...) and consolidate all the #ifdef logic at the function definition.
| !utils::one_of(DNNL_RUNTIME_DIM_VAL, desc()->m(), | ||
| desc()->n(), desc()->k(), desc()->lda(), | ||
| desc()->ldb(), desc()->ldc(), desc()->batch()), | ||
| VERBOSE_RUNTIMEDIM_UNSUPPORTED); |
There was a problem hiding this comment.
Comment: The OpenCL kernel looks like it supports runtime dimensions, so we are likely just missing some logic in the execute function, it might be good to explain why were are disabling it here (in general, I think ref_gemm should just be removed, but RNN relies on it as I recall).
3 participants
closes MFDNN-14852
MFDNN-14853