simd-lang

A language that makes it easy to write succinct SIMD code.

Vectorized iteration with cross-iteration state. The stream, over, and carry keywords make SIMD code read like a normal loop. Hits ~73% of simdjson stage-1 throughput on M1 (~3.5 GB/s) and ~76% for full DOM parse (~1.15 GB/s).

simdexamples/json_stage1.simd

 1fn json_stage1(input: ptr[u8], positions: ptr[i32]) -> i32[1] {
 2    stream chunk: u8[64] over input carry (prev_in_string: u64[1] = 0,
 3                                           pos: i32[1] = 0) {
 4        quote_vec      = chunk == '"'
 5        bs_vec         = chunk == '\\'
 6        real_quote_vec = quote_vec & ~lane_shr(bs_vec, 1)
 7 
 8        structural_vec = (chunk == '{') | (chunk == '}') | (chunk == '[')
 9                       | (chunk == ']') | (chunk == ':') | (chunk == ',')
10 
11        quote_bits           = to_bitmask(real_quote_vec)
12        in_string_raw        = clmul(quote_bits, ~0)
13        in_string_with_carry = in_string_raw ^ prev_in_string
14 
15        real_structural_bits = to_bitmask(structural_vec) & ~in_string_with_carry
16        real_structural_vec  = from_bitmask(real_structural_bits, 64)
17 
18        indices = iota(64) + chunk_offset
19        compressstore(positions, pos, indices, real_structural_vec)
20 
21        carry prev_in_string = in_string_with_carry >> 63 * ~0
22        carry pos            = pos + popcount(real_structural_vec)
23    }
24    return pos
25}