xe: jit: gemm: Add k-parallelism parameter to Xe2 gemm kernels

[Simonsays095]

Addresses MFDNN-13067. The Xe2 dynamic quantization kernels have some parameters that need to be tweaked:

• ar -> sr br: This results in faster execution and should be the default in all cases
• 2 k-parallel kernels need to have ikr added to the strategy to be valid
• di cc can be removed from 2 kernels

The result should be a few correctness passes that used to fail (below), and slightly more optimized execution.

$ benchdnn --mode=C --matmul --engine=gpu --dt=s8:s4:f16 --stag=ab --wtag=ba --dtag=ab --attr-scales=src0:per_ocic:f16:1x128+wei:per_ocic:f16:128x1 --attr-fpmath=f16:true 1x4096:4096x4096
[   0][DST][0:0] exp_f32:     -213.25 exp:     -213.25 got:    -53.3125 diff: 159.938 rdiff:    0.75
[   1][DST][0:1] exp_f32:    -137.375 exp:    -137.375 got:    -68.6875 diff: 68.6875 rdiff:     0.5
[   2][DST][0:2] exp_f32:    -109.625 exp:    -109.625 got:     -219.25 diff: 109.625 rdiff:       1
[   4][DST][0:4] exp_f32:     798.812 exp:         799 got:      199.75 diff:  599.25 rdiff:    0.75
[   6][DST][0:6] exp_f32:      36.375 exp:      36.375 got:       145.5 diff: 109.125 rdiff:       3
[   7][DST][0:7] exp_f32:      415.75 exp:      415.75 got:     207.875 diff: 207.875 rdiff:     0.5
[   9][DST][0:9] exp_f32:    -464.938 exp:        -465 got:       -1860 diff:    1395 rdiff:       3
[  10][DST][0:10] exp_f32:      89.875 exp:      89.875 got:      179.75 diff:  89.875 rdiff:       1
[  11][DST][0:11] exp_f32:    -466.938 exp:        -467 got:        -934 diff:     467 rdiff:       1
[  12][DST][0:12] exp_f32:     -46.375 exp:     -46.375 got:    -11.5938 diff: 34.7812 rdiff:    0.75
0:FAILED (errors:4092 total:4096) __REPRO: --matmul --engine=gpu --dt=s8:s4:f16 --stag=ab --wtag=ba --dtag=ab --attr-scales=src:per_ocic:f16:1x128+wei:per_ocic:f16:128x1 --attr-fpmath=f16:true 1x4096:4096x4096
tests:1 passed:0 skipped:0 mistrusted:0 unimplemented:0 invalid_arguments:0 failed:1 listed:0
total: 0.43s; fill: 0.09s (21%); compute_ref: 0.15s (36%); compare: 0.00s (0%);

[Simonsays095]

make test
disable test_device_cpu
enable test_device_gpu

[petercad]

@Simonsays095 can you run Xe2 perf testing?

[petercad]

After modifying strategies in the catalog "by hand," you'll want to update the embedded DriverInfo structs with ktool --reinfo. In this case only the strategies you added ikr to really need it.