Gfx12.5 (DG2) will use DPAS instructions to accelerate the
implementation. Earlier platforms will use equivalent discrete
instructions (basically subgroup operations). Gfx12 (Tigerlake) will use
DP4A for 8-bit integer matrix multiplication. Older platforms, which
lack DP4A, will use a suboptimal instruction sequence. There is plenty
of room for improvement here.
On DG2 (Gfx12.5) gets the following results from the CTS:
Test run totals:
Passed: 1642/13982 (11.7%)
Failed: 0/13982 (0.0%)
Not supported: 12340/13982 (88.3%)
Warnings: 0/13982 (0.0%)
Waived: 0/13982 (0.0%)
On DG2 (Gfx12.5) with forced lowering, Raptor Lake (Gfx12) and Ice Lake
(Gfx11):
Test run totals:
Passed: 1662/13982 (11.9%)
Failed: 0/13982 (0.0%)
Not supported: 12320/13982 (88.1%)
Warnings: 0/13982 (0.0%)
Waived: 0/13982 (0.0%)
The difference in the number of tests run is due to
saturatingAccumulation not being set on DG2 when DPAS is used. There is
a comment in "intel/dev: Advertise integer configs with
saturatingAccumulation too" that explains how this could be added should
the need arise.
v2: Prefix type names with INTEL_CMAT_. Suggested by Lionel.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25994>
c6d44284aa