[kstenerud]

It can't be done, because the next bytes are dependent upon the first byte (which only works in limited circumstances, and where you have constant spacing between the values).

ULEB128 works in SIMD because there's only one dependent bit per byte, so you can speculatively decode and then correct later cheaply. Bijou requires you to check the first byte and then branch based on the value using all 8 bits in the decision matrix (to handle branches 0-247, 248, 249, 250, 251, 252, 253, 254, 255). This absolutely DESTROYS any parallelization opportunities.

Not to mention that non-canonical sized ints (3, 5, 6, 7) have abysmal performance compared to unaligned 2, 4, and 8 byte reads on modern processors.

[jaen]

Right, I think we have a slightly different definition of SIMD: You mean byte-parallel, I mean "doable with SIMD instructions". I also didn't imply the performance would be better than other methods...

Even though decoding the lengths must be serial (since's there's no unambiguous way to differentiate a tag and data byte), it's still doable within the wider SIMD registers, so there's some theoretical efficiency gain to be had (depending on the shape of the data).

On a general note, the continuation bit and prefix byte forms are equivalent, you just broadcast the prefix byte and compare against an increasing vector to convert it to a mask. Yeah, there's probably more fiddly SIMD if there are multiple prefixes in the register, but doable (it's just not byte-parallel, you eg. unroll the serial decode loop 8 times or whatever your maximum output byte width is, and mask out).

Simplified:

  // Just maps a byte to its position in the register
  __m128i idx = _mm_setr_epi8(0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15);
  // Broadcast the prefix
  __m128i nn = _mm_set1_epi8((char)prefix_byte);
  // Get applicable locations: prefix_byte contains the length, if byte_pos < len, the corresponding byte will be set
  __m128i m = _mm_cmpgt_epi8(nn, idx);
  // If you *really* want a high-bit mask:
  m = _mm_and_si128(m, _mm_set1_epi8((char)0x80));

[kstenerud]

Yeah, sorry, I didn't say that very well. Single value decoding of Bijou values is of course trivial in SIMD, but the performance benefits of SIMD come from deterministic boundaries across a window. ULEB128's continuation bit is fixed position, so it's data independent. One pmovmskb gives you every boundary in the window.

Interleaved Bijou has no such signal (tag and payload bytes both span 0x00–0xFF), so finding the boundaries is a dependent per-value walk with no opportunities for parallelism.

[jaen]

There's still speculation though - if eg. most values are of 1 or 2-byte length, you can speculate that any control-valued byte is actually control. You can even do a compensation pass to try to fix some amount of mis-speculations, and then bomb out if that fails.

With that, it's mostly byte-parallel (though data-dependent as I mentioned).

[NooneAtAll3]

would vector instruction be of any help? (variable length simd)