Pcurl improvements

[semenov-vladyslav]

Ptrit/pcurl implementation using SIMD (see README).

When building with bazel ptrit and pcurl configuration flags must be passed via --define:

• PTRIT_PLATFORM:

  bazel ... --define ptrit_platform=PTRIT_64
  bazel ... --define ptrit_platform=PTRIT_SSE2 --copt=-msse2
  bazel ... --define ptrit_platform=PTRIT_AVX2 --copt=-mavx2
  bazel ... --define ptrit_platform=PTRIT_AVX512 --copt=-mavx512f --copt=-fno-asynchronous-unwind-tables
  bazel ... --define ptrit_platform=PTRIT_NEON --copt=-march=armv7-a --copt=-mfpu=neon

• PTRIT_CVT:

  bazel ... --define ptrit_cvt=PTRIT_CVT_ANDN
  bazel ... --define ptrit_cvt=PTRIT_CVT_ORN

• PCURL_S2_CIRCUIT:

  bazel ... --define pcurl_s2_circuit=PCURL_S2_CIRCUIT4
  bazel ... --define pcurl_s2_circuit=PCURL_S2_CIRCUIT5

• PCURL_S2_ARGS:

  bazel ... --define pcurl_s2_args=PCURL_S2_ARGS_PTR
  bazel ... --define pcurl_s2_args=PCURL_S2_ARGS_LUT

• PCURL_STATE:

  bazel ... --define pcurl_state=PCURL_STATE_SHORT
  bazel ... --define pcurl_state=PCURL_STATE_DOUBLE

• PCURL_SBOX_UNWIND:

  bazel ... --define pcurl_sbox_unwind=PCURL_SBOX_UNWIND_1
  bazel ... --define pcurl_sbox_unwind=PCURL_SBOX_UNWIND_2
  bazel ... --define pcurl_sbox_unwind=PCURL_SBOX_UNWIND_4
  bazel ... --define pcurl_sbox_unwind=PCURL_SBOX_UNWIND_8

TODO:

• [x] Detect BMI instruction in cpuid, needed in order to use _tzcnt_u64 intrinsic under MSVC.
• [x] In ptrits_find_zero_slice return SIZE_MAX instead of PTRIT_SIZE?
• [x] Check for carry after ptrit_increment in pearl diver, carry can indicate exhaustion of the search.
• [x] The original end of ptrit_increment in run_search_thread was HASH_LENGTH_TRIT. Was it an error? It should be end.
• [x] There is no need for statusLock in pearl diver. It locks access to status variable, which is only overwritten with SEARCH_FINISHED. As a compromise atomics can be used instead of lock.
• [x] Search state is incremented in pearl diver as follows: ptrit_increment(state, begin + 1, end). With multiple threads (4+) this search space can overlap?!
• [x] Test on ARM with NEON and on AVX512 under sde.
• [x] ssize_t is not standard type and causes problems on Windows with msvc.
• [x] The bazel command line is too long, add PTRIT_PLATFORM autodetect.
• [x] Drop support for PCURL_S2_ARGS_LUT as it won't be used in production code?
• [ ] Drop support for PCURL_SBOX_UNWIND?
• [x] Ensure the ciri uses PTRIT_SIZE transactions per call to pcurl.

[semenov-vladyslav]

Also, ssize_t causes problems on Windows, as it's not a standard type.

[semenov-vladyslav]

Changed pearl diver search logic a bit: search range is split into fixed and variable ranges. The fixed range is initialized with different n * PTRIT_SIZE integer values, where n is the number of threads. The variable range is initialized with "all -1s", so when ptrit_hincr overflows that means that the variable range is completely exhausted and the search has failed.

[thibault-martinez]

I allowed myself to tick two items of your list because I created the following related issues:

• https://github.com/iotaledger/entangled/issues/1422
• https://github.com/iotaledger/entangled/issues/1414