4 ms·

A lookup table in memory can only be accessed an element at a time, so it gets bottlenecked in the execution units on the load and address generation traffic. T

A lookup table in memory can only be accessed an element at a time, so it gets bottlenecked in the execution units on the load and address generation traffic. This algorithm uses 8-bit integers, so the vectorized version is processing between 16-64 elements per operation depending on the vector width. It's even worse if the 8-bit divide is integrated with other vector operations as then the lookup table method also has insert/extract overhead to exchange the values between the vector and integer units. A hybrid approach using small in-register lookup tables in the vector unit (pshufb/tbl) can be lucrative, but are very limited in table size.

mithametacs 2y ago
I've never thought of a micro-lookup-table! That's cool. What have you used those for?
- ack_complete 2y ago
  Typically just like conventional lookup tables, where you can get the table size down small enough. Indexed palette / quantization coding is a case where this can often work. It's pretty niche given the limitations, but if you can make it work it's often a major speedup since you're able to do 16/32/64 lookups in parallel.
- glangdale 2y ago
  Not OP, but we use these tables all the time in Hyperscan (for string and character class matching) and it's a long-standing technique to use it for things like SIMD population count (obsoleted now if you have the right AVX-512 ISA, ofc).
- anonymoushn 2y ago
  You can test whether a register full of bytes belong to a class of bytes with the high bit unset and distinct lower nibbles in 2 instructions (each with 1 cycle latency). For example the characters that commonly occur in text and must be escaped in json are "\"\\\r\n\t" (5 different bytes). Their lower nibbles are: \": 0010 \t: 1001 \n: 1010 \\: 1100 \r: 1101 Since there are no duplicate lower nibbles, we can test for membership in this set in 2 instructions. If we want to get the value into a general-purpose register and branch on whether it's 0 or not, that's 2 more instructions. For larger sets or sets determined at runtime, see here: http://0x80.pl/notesen/2018-10-18-simd-byte-lookup.html http://0x80.pl/notesen/2018-10-18-simd-byte-lookup.html

mithametacs 2y ago

I've never thought of a micro-lookup-table! That's cool. What have you used those for?

ack_complete 2y ago
Typically just like conventional lookup tables, where you can get the table size down small enough. Indexed palette / quantization coding is a case where this can often work. It's pretty niche given the limitations, but if you can make it work it's often a major speedup since you're able to do 16/32/64 lookups in parallel.
glangdale 2y ago
Not OP, but we use these tables all the time in Hyperscan (for string and character class matching) and it's a long-standing technique to use it for things like SIMD population count (obsoleted now if you have the right AVX-512 ISA, ofc).
anonymoushn 2y ago
You can test whether a register full of bytes belong to a class of bytes with the high bit unset and distinct lower nibbles in 2 instructions (each with 1 cycle latency). For example the characters that commonly occur in text and must be escaped in json are "\"\\\r\n\t" (5 different bytes). Their lower nibbles are: \": 0010 \t: 1001 \n: 1010 \\: 1100 \r: 1101 Since there are no duplicate lower nibbles, we can test for membership in this set in 2 instructions. If we want to get the value into a general-purpose register and branch on whether it's 0 or not, that's 2 more instructions. For larger sets or sets determined at runtime, see here: http://0x80.pl/notesen/2018-10-18-simd-byte-lookup.html http://0x80.pl/notesen/2018-10-18-simd-byte-lookup.html