src: cpu: aarch64: injectors: eltwise: Improve performance of tanh for block size 16

[vishwascm]

Description

Performance Improvement: Eltwise Tanh performance improvement for block size 16

Major Code changes:
• Added a new function tanh_polynomial_approx_compute_vector_fwd(const TRegS &vmm_src) for
computing tanh.
• Added new tanh constants and polynomial constants table.

Checklist

General

All the tests are carried on A64FX machine which has block size 16:

• [✓] Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?

1. make test

95% tests passed, 11 tests failed out of 200
Total Test time (real) = 3790.05 sec
The following tests FAILED:
	 55 - test_convolution_backward_data_f32 (Subprocess aborted)
	123 - test_graph_c_api_compile_parametrized_usm_cpu (Failed)
	153 - test_graph_unit_dnnl_conv_usm_cpu (Failed)
	157 - test_graph_unit_dnnl_group_norm_usm_cpu (Failed)
	159 - test_graph_unit_dnnl_large_partition_usm_cpu (Failed)
	160 - test_graph_unit_dnnl_layer_norm_usm_cpu (Failed)
	161 - test_graph_unit_dnnl_matmul_usm_cpu (Failed)
	162 - test_graph_unit_dnnl_mqa_decomp_usm_cpu (Failed)
	163 - test_graph_unit_dnnl_pool_usm_cpu (Failed)
	168 - test_graph_unit_dnnl_sdp_decomp_usm_cpu (Failed)
	169 - test_graph_unit_dnnl_softmax_usm_cpu (Failed)
Errors while running CTest
Output from these tests are in: /home/vishwas/oneDNN/build/Testing/Temporary/LastTest.log
Use "--rerun-failed --output-on-failure" to re-run the failed cases verbosely.
make: *** [Makefile:71: test] Error 8

2. ./benchdnn --eltwise --batch=inputs/eltwise/test_eltwise_all

tests:75713 passed:6771 skipped:68907 mistrusted:35 unimplemented:0 invalid_arguments:0 failed:0 listed:0
total: 125.07s; fill: 29.31s (23%); compute_ref: 15.50s (12%); compare: 27.88s (22%);

3. make test_benchdnn_*

make: *** No rule to make target 'test_benchdnn_*'.  Stop.

gtests

4. ./test_eltwise

----------] Global test environment tear-down
[==========] 44 tests from 5 test suites ran. (2752 ms total)
[  PASSED  ] 24 tests.
[  SKIPPED ] 20 tests, listed below:
[  SKIPPED ] Test_Eltwise_EF/eltwise_test_t.TestsEltwise/4
[  SKIPPED ] Test_Eltwise_EF/eltwise_test_t.TestsEltwise/5
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/0
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/1
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/2
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/3
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/4
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/5
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/6
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/7
[  SKIPPED ] EltwiseSimpleBF16/eltwise_test_t.TestsEltwise/8
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/0
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/1
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/2
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/3
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/4
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/5
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/6
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/7
[  SKIPPED ] EltwiseSimpleF16/eltwise_test_t.TestsEltwise/8

Note: All above results are same with and without the code changes.

• [✓] Have you formatted the code using clang-format? - Yes

Performance improvements

• [✓ ] Have you submitted performance data that demonstrates performance improvements?
  
  

[theComputeKid]

Thanks for this. Can you let me know what cmake command line options you use when building to get these perf results?

[vishwascm]

Thanks for this. Can you let me know what cmake command line options you use when building to get these perf results?

@theComputeKid following steps were used to get perf results (benchdnn was used):

cd build
cmake ..
make -j
cd tests/benchdnn
taskset -c 1,2,3,4 ./benchdnn --eltwise --mode=p  --alg=gelu_tanh 3x17x2x5x3
taskset -c 1,2,3,4 ./benchdnn --conv --mode=p --attr-post-ops=eltwise_tanh ic3oc64_ih224oh112kh7sh2dh0ph3_iw224ow112kw7sw2dw0pw3

Above test is for 4 Core.

[theComputeKid]

Sorry, I should have clarified, I was interested in the CMake options during the configure phase. Particularly, I wanted to know whether you compile with -DONEDNN_WERROR=ON, as I have found that the JIT codebase for aarch64 produces a lot of warnings that prevent us from turning on the flag. Could you just confirm that no new warnings are added by your changes? Thanks.

[vishwascm]

Sorry, I should have clarified, I was interested in the CMake options during the configure phase. Particularly, I wanted to know whether you compile with -DONEDNN_WERROR=ON, as I have found that the JIT codebase for aarch64 produces a lot of warnings that prevent us from turning on the flag. Could you just confirm that no new warnings are added by your changes? Thanks.

@theComputeKid I did not compile with -DONEDNN_WERROR=ON.

[theComputeKid]

Can you still please confirm that no new warnings are added due to your changes? I.e. the number of warnings emitted before and after your changes (if any) are the same or less.

[jondea]

When you say block size 16, do you mean this should only affect machines with an SVE vector length of 512? Block size is a somewhat overloaded term, especially with different data types.

[vishwascm]

Can you still please confirm that no new warnings are added due to your changes? I.e. the number of warnings emitted before and after your changes (if any) are the same or less.

@theComputeKid Yeah, there are no new warnings due to the changes.

[vishwascm]

When you say block size 16, do you mean this should only affect machines with an SVE vector length of 512? Block size is a somewhat overloaded term, especially with different data types.

@jondea Yes, it is for SVE_512, for fp32 datatype.

[abhijain1204fujitsu]

@jondea Kindly support to check the changes and let us know if some information or changes are required.

[theComputeKid]

Can you please rebase and push to force pipelines to run again. Thanks.