benchdnn: prohibit --mode=CP

[kwieloch-intel]

This PR prohibits --mode=CP in benchdnn and auto-enables no_ref_memory for --mode=P on Intel GPUs to align performance data filling with --mode=F.

JIRA: MFDNN-14789

---

Problem description

Two issues with benchdnn performance mode data filling on Intel GPUs:

1. --mode=CP is broken: correctness + performance combined mode is unused and currently broken on both CPU and GPU. More importantly, correctness validation requires specific low-range data filling (less randomness in mantissa to minimize round-off errors), which makes the data compressible by the GPU driver and produces unreliable performance numbers.

2. --mode=P vs --mode=F data mismatch: in --mode=P, fill_random_real_dense() overwrites GPU buffers with deterministic CPU-generated data via reorder(), replacing the incompressible Philox PRNG data written by gpu_fill_random() during dnn_mem_t::initialize(). This leads to different (and potentially compressible) data patterns compared to --mode=F.

---

Proposed Solution

Two changes in engine_t constructor (dnnl_common.cpp):

1. Prohibit --mode=CP: if both corr and perf mode bits are set, emit an error message and fail. Users should run --mode=C and --mode=P (or --mode=F) separately.

2. Auto-enable no_ref_memory for --mode=P on Intel GPUs: when running perf-only mode (perf && !corr) with an Intel GPU engine, automatically set the no_ref_memory modifier. This skips CPU reference memory allocation and fill_random_real_dense() calls, so GPU buffers retain incompressible Philox PRNG data, aligning --mode=P data filling with --mode=F.

   For CPU, --mode=F uses 0x3F memset (different from --mode=P which uses fill_random_real_dense), so the existing behavior is intentionally preserved, no change for CPU.

---

Example

Before: --mode=CP silently runs but produces NaN in correctness output (broken) and unreliable perf numbers:

> benchdnn --mode=CP --matmul --engine=gpu --dt=f16 128x128:128x128

[   0][DST][0:0] exp_f32:     2438.02 exp:        2438 got:        -nan diff:     nan rdiff:     nan
[   1][DST][0:1] exp_f32:     115.184 exp:     115.188 got:        -nan diff:     nan rdiff:     nan
[   2][DST][0:2] exp_f32:     258.576 exp:       258.5 got:        -nan diff:     nan rdiff:     nan
...
perf,gpu,jit:gemm:any,,--mode=CP --matmul --engine=gpu --dt=f16:f16:f16 128x128:128x128,0.0041943,...
tests:1 passed:1 skipped:0 ... failed:0

After: --mode=CP is explicitly prohibited:

> benchdnn --mode=CP --matmul --engine=gpu --dt=f16 128x128:128x128

Error: --mode=CP is not supported. Use --mode=C and --mode=P (or --mode=F) separately.

---

[dzarukin]

All those parts to be handled in utils/parser.cpp, not here.

[kwieloch-intel]

fbfda9c I’ve moved it to utils/parser.cpp.

[dzarukin]

I'm sorry I was not quite clear in my intentions. There's already logic handling this: https://github.com/uxlfoundation/oneDNN/blob/main/tests/benchdnn/utils/parser.cpp#L1568

In this function you'll find places for both updates you want to do.