#1073: Caching hash method pointers globally to improve performances

[haxelion]

This PR fixes the performance issue with TCrypto hash functions we've observed as explained in issue intel/confidential-computing.sgx.sdk#168.

Summary of the changes

• common/inc/internal/se_tcrypto_common.h: add a global structure containing IppsHashMethod pointers when TCrypto is build with IPP
• sdk/tlibcrypto/ipp/init_tcrypto_lib.cpp: populate the global hash methods structure after IPP has been initialized with CPUID values.
• sdk/tlibcrypto/Makefile: avoid SHA1 deprecation warning when building init_tcrypto_lib.cpp.
• other sdk/tlibcrypto/ipp/*.cpp files: switch to use the global hash method structure.

I've also performed the change in RSA and ECC code but haven't benchmarked it. However I don't expect a significant performance uplift in those cases. I simply did it for consistency.

I'm also unsure of what code conventions I should follow so please don't hesitate to ask for such changes.

Performance data

The following benchmark were run on a dual Intel(R) Xeon(R) Silver 4310 system. The support of SHANI amplify the magnitude of the issue so systems without SHANI support will see a smaller uplift.

Benchmarks can be replicated using the enclave code here: https://github.com/haxelion/sgx_tcrypto_bench

The benchmark code perform 1,000,000 iteration of the selected algorithm inside an ECALL. Several of those ECALL are called in parallel from multiple threads. What is expected is that, up to the core count, there shouldn't be any noticeable slowdown.

I've also attached the detailed console output to this PR: Intel TCrypto benchmark data.md.

SHA1

Thread count	1	2	6	12	24
Without fix	87 ms	727 ms	1093 ms	1977 ms	3684 ms
With fix	83 ms	83 ms	83 ms	83 ms	84 ms

24 threads speed-up: 44x

SHA256

Thread count	1	2	6	12	24
Without fix	114 ms	884 ms	1671 ms	2883 ms	4732 ms
With fix	107 ms	108 ms	107 ms	110 ms	109 ms

24 threads speed-up: 43x

HMAC-SHA256

Thread count	1	2	6	12	24
Without fix	640 ms	766 ms	3404 ms	6183 ms	8875 ms
With fix	636 ms	636 ms	634 ms	640 ms	637 ms

24 threads speed-up: 14x

SHA384

Thread count	1	2	6	12	24
Without fix	332 ms	1018 ms	1451 ms	2263 ms	4074 ms
With fix	333 ms	333 ms	333 ms	333 ms	333 ms

24 threads speed-up: 12x

[bot-cc1s-branch-sync]

Important

🔀 The Intel® SGX SDK codebase has moved

Thank you for your contribution!
The code this pull request modifies now resides in a new repository,¹ so, unfortunately, it can no longer be merged here as-is. If your change is still relevant, please re-open it against the SDK repository, where this code is now maintained.

If you have any questions or need assistance migrating, please ask here. As we complete the migration, older pull requests may eventually be closed. No work is lost, though, and you are welcome to re-open your change in the new repository at any time.

Footnotes

1. The affected sources were migrated from confidential-computing.sgx to the dedicated confidential-computing.sgx.sdk repo. The migration covered the SDK sources (sdk/), shared headers (common/), samples (SampleCode/), and the runtime libraries (psw/urts, psw/enclave_common). ↩