Onednn backend

[borg323]

This is the long awaited onednn backend (formerly dnnl, formerly mkl-dnn). To build pass -Donednn=true -Ddnnl_dir=path/to/dnnl/library to meson. This works with both cpu and (intel) gpu. For gpu you will need to build the dnnl library yourself (or ask me for a dll).

For best performance use just one search thread when running on a cpu as a second search thread interferes with the onednn computing threads (onednn on a cpu uses all available cores by default).

There are several backend options, the most important are:

option	values	default	comment
gpu	empty or integer	empty	select gpu to use, empty for cpu
winograd	empty, true or false	empty	Set to true to use Winograd 3x3 convolution on cpu, false for direct convolution, empty to let the library chose. Currently dnnl after v2.0.0 may get this wrong, and is not supported on all processors so may exit with a not very informative error.
fp16	true or false	true on gpu, false on cpu	Use fp16 (or bf16 if you cpu supports it) for computation.
batch	integer	64 for fp16, 32 for fp32	The minimum batch size to use, as the library prepares kernels on startup. Set to 0 for dynamic kernel recompilation on every batch (not recommended).
steps	integer	2	Prepare kernels for this many multiples of batch size.
threads	empty or integer	empty	Number of cpu threads to use, empty to let the library decide.

[borg323]

This is dnnl 1.8.0 compiled with both gpu and cpu support: dnnl.dll.zip

Example command line for benchmark:
./lc0 benchmark -w 703810.pb.gz --threads=1 --backend=onednn --backend-opts=gpu=0

[aochoam]

@borg323 Result of the Sanity checking the dx12 driver.
_
| _ | |
|_ |_ |_| v0.28.2 built Dec 13 2021
Detected 4 core(s) and 8 thread(s) in 1 group(s).
Group 0 has 4 core(s) and 8 thread(s).
Found pb network file: C:\Arena\Engines\lc0_dx12/771479.pb.gz
Creating backend [check]...
Working backend set to dx12.
Reference backend set to eigen.
Creating backend [dx12]...
Creating backend [eigen]...
Using Eigen version 3.3.7
Eigen max batch size is 256.
Check mode: check only with relative tolerance 1.0e-04, absolute tolerance 5.0e-01.
Check rate: 100%.

Position: 1/1 rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1
*** ERROR check failed for a batch of 1 policy incorrect (but value ok).
*** ERROR check failed for a batch of 19 policy incorrect (but value ok).
Benchmark time 6218 ms, 20 nodes, 21 nps, move b1c3
*** ERROR check failed for a batch of 20 policy incorrect (but value ok).
Benchmark time 6260 ms, 21 nodes, 21 nps, move b1a3
Benchmark time 10007 ms, 21 nodes, 4 nps, move b1a3
bestmove b1a3
*** ERROR check failed for a batch of 144 both value and policy incorrect.
*** ERROR check failed for a batch of 256 both value and policy incorrect.

===========================
Total time (ms) : 18655
Nodes searched : 421
Nodes/second : 23
Presione una tecla para continuar . . .

[aochoam]

This version does not work for me.