benchdnn: respect accumulation mode in safe digits calculation

[echeresh]

Jira: MFDNN-14670

PR addresses two issues:

1. benchdnn doesn't respect acc-mode during data filling (e.g. see the matmul failure below)
2. GPU convolution/deconvolution is not aligned with documentation and can use f16 accumulator even when acc-mode doesn't allow it: link
   • It's a FIXME but at this point it's easier to wait until XeLP/XeLPG are officially deprecated

Matmul failures on XeLP/XeLPG - note --attr-acc-mode=any:

$ ./build/tests/benchdnn/benchdnn -v5 --matmul --engine=gpu --dt=f16:f16:f32 --attr-acc-mode=any -v5 128x10240:10240x128
create: --matmul --engine=gpu --dt=f16:f16:f32 --attr-acc-mode=any 128x10240:10240x128
oneDNN implementation: jit:gemm:any
CPU reference oneDNN implementation: brg_matmul:avx2
run: --matmul --engine=gpu --dt=f16:f16:f32 --attr-acc-mode=any 128x10240:10240x128
run ref: --matmul --engine=cpu --attr-acc-mode=any 128x10240:10240x128
[   0][DST][0:0] exp_f32:       -2613 exp:       -2613 got:       -2612 diff:       1 rdiff:0.000382702
[   4][DST][0:4] exp_f32:        2365 exp:        2365 got:        2364 diff:       1 rdiff:0.000422833
[  10][DST][0:10] exp_f32:       -2169 exp:       -2169 got:       -2168 diff:       1 rdiff:0.000461042
[  16][DST][0:16] exp_f32:       -3379 exp:       -3379 got:       -3380 diff:       1 rdiff:0.000295946
[  25][DST][0:25] exp_f32:        2821 exp:        2821 got:        2822 diff:       1 rdiff:0.000354484
[  29][DST][0:29] exp_f32:        2707 exp:        2707 got:        2708 diff:       1 rdiff:0.000369413
[  33][DST][0:33] exp_f32:        2263 exp:        2263 got:        2264 diff:       1 rdiff:0.000441891
[  38][DST][0:38] exp_f32:       -5186 exp:       -5186 got:       -5184 diff:       2 rdiff:0.000385654
[  44][DST][0:44] exp_f32:       -2287 exp:       -2287 got:       -2288 diff:       1 rdiff:0.000437254
[  54][DST][0:54] exp_f32:       -2293 exp:       -2293 got:       -2292 diff:       1 rdiff:0.00043611
[COMPARE_STATS][DST]: trh=0 err_max_diff:       6 err_max_rdiff:0.00101868 all_max_diff:       6 all_max_rdiff:0.00101868
[PRIM_REF][INFO]: L2_size:524288 bytes; per_core_L3_size:2359296 bytes; nthr:24; impl_name:brg_matmul:avx2
0:FAILED (errors:3689 total:16384) (627 ms) __REPRO: --matmul --engine=gpu --dt=f16:f16:f32 --attr-acc-mode=any 128x10240:10240x128

[rjoursler]

GPU convolution/deconvolution is not aligned with documentation and can use f16 accumulator even when acc-mode doesn't allow it: link
It's a FIXME but at this point it's easier to wait until XeLP/XeLPG are officially deprecated

I think we should just fix this (unless it causes functionality regressions). It looks like is was just missed when implementing #1540.

[echeresh]

GPU convolution/deconvolution is not aligned with documentation and can use f16 accumulator even when acc-mode doesn't allow it: link
It's a FIXME but at this point it's easier to wait until XeLP/XeLPG are officially deprecated

I think we should just fix this (unless it causes functionality regressions). It looks like is was just missed when implementing #1540.

I agree, it's tracked in MFDNN-13271
I currently lack the bandwidth for it so that's why I came up with this more targeted fix in benchdnn to fix the existing Nightly failures.