Skip to content

D25 cooking #20

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
alion02 merged 3 commits into from
Dec 26, 2024
Merged

D25 cooking #20

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
837b67d
wip
alion02 Dec 26, 2024
fc3d0c9
wip2
alion02 Dec 26, 2024
b5269dd
cook up a jump table
alion02 Dec 26, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
269 changes: 153 additions & 116 deletions src/day25.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,43 +5,43 @@ use super::*;
unsafe fn inner1(s: &[u8]) -> u32 {
static mut LOCKS: [u32; 250] = [0; 250];
static mut KEYS: [u32x8; 32] = [Simd::from_array([!0; 8]); 32];
static mut BUCKETS: [[u32; 250]; 6] = [[0; 250]; 6];

let locks = LOCKS.as_mut_ptr();
let keys = KEYS.as_mut_ptr();
let buckets = &mut BUCKETS;
let indices = &mut [0usize; 6];

asm!(
"jmp 20f",
"21:",
"mov [{locks}], {mask:e}",
"add {locks}, 4",
"vpcmpeqb {chunk}, {vec_ascii_hash}, [{ptr} + {i} - 43]",
"vpmovmskb {mask:e}, {chunk}",
"test {mask:l}, 1",
"jz 23f",
"22:",
"mov [{locks}], {mask:e}",
"add {locks}, 4",
"add {i:e}, -86",
"add {i:e}, -43",
"jl 30f",
"20:",
"vpcmpeqb {chunk}, {vec_ascii_hash}, [{ptr} + {i}]",
"vpmovmskb {mask:e}, {chunk}",
"test {mask:l}, 1",
"jnz 21b",
"mov [{keys}], {mask:e}",
"add {keys}, 4",
"vpcmpeqb {chunk}, {vec_ascii_hash}, [{ptr} + {i} - 43]",
"vpmovmskb {mask:e}, {chunk}",
"test {mask:l}, 1",
"jnz 22b",
"23:",
"mov [{keys}], {mask:e}",
"add {keys}, 4",
"add {i:e}, -86",
"andn {height:e}, {mask:e}, {col_mask:e}",
"popcnt {height:e}, {height:e}",
"imul {bucket:e}, {height:e}, 1000",
"lea {bucket}, [{buckets} + {bucket}]",
"mov {idx}, [{indices} + {height} * 8]",
"mov [{bucket} + {idx}], {mask:e}",
"add {idx}, 4",
"mov [{indices} + {height} * 8], {idx}",
"add {i:e}, -43",
"jge 20b",
"30:",
locks = inout(reg) locks => _,
keys = inout(reg) keys => _,
buckets = in(reg) buckets,
indices = in(reg) indices,
height = out(reg) _,
bucket = out(reg) _,
idx = out(reg) _,
col_mask = in(reg) 0b1000001000001000001000001000,
mask = out(reg) _,
i = inout(reg) 43usize * 499 + 3 => _,
ptr = in(reg) s.as_ptr(),
Expand All @@ -50,6 +50,32 @@ unsafe fn inner1(s: &[u8]) -> u32 {
options(nostack),
);

asm!(
"23:",
"mov {idx}, [{index}]",
"add {idx}, -4",
"jl 22f",
"20:",
"mov {tmp:e}, [{bucket} + {idx}]",
"mov [{keys}], {tmp:e}",
"add {keys}, 4",
"21:",
"add {idx}, -4",
"jge 20b",
"22:",
"add {bucket}, -1000",
"add {index}, -8",
"cmp {bucket}, {buckets}",
"jge 23b",
keys = inout(reg) keys => _,
index = inout(reg) indices.as_mut_ptr().add(5) => _,
buckets = in(reg) buckets.as_mut_ptr(),
bucket = inout(reg) buckets.as_mut_ptr().add(5) => _,
tmp = out(reg) _,
idx = out(reg) _,
options(nostack),
);

let mut sums = i32x8::splat(0);

asm!(
Expand All @@ -65,109 +91,120 @@ unsafe fn inner1(s: &[u8]) -> u32 {
"vmovdqa {cache10}, [rip + {keys}+288]",
"vmovdqa {cache11}, [rip + {keys}+320]",
"vmovdqa {cache12}, [rip + {keys}+352]",
"lea {reladdrs}, [rip + 200f]",
"mov {r2:e}, [{reladdrs}]",
"add {r2}, {reladdrs}",
"mov {r1}, [{table}]",
"mov [{table}], {r2}",
"mov {r2}, {r1}",
"shr {r2}, 5",
"mov {r2:e}, [{reladdrs} + {r2} * 4]",
"add {r2}, {reladdrs}",
"add {r1}, [{table} + 8]",
"mov [{table} + 8], {r2}",
"mov {r2}, {r1}",
"shr {r2}, 5",
"mov {r2:e}, [{reladdrs} + {r2} * 4]",
"add {r2}, {reladdrs}",
"add {r1}, [{table} + 16]",
"mov [{table} + 16], {r2}",
"mov {r2}, {r1}",
"shr {r2}, 5",
"mov {r2:e}, [{reladdrs} + {r2} * 4]",
"add {r2}, {reladdrs}",
"add {r1}, [{table} + 24]",
"mov [{table} + 24], {r2}",
"mov {r2}, {r1}",
"shr {r2}, 5",
"mov {r2:e}, [{reladdrs} + {r2} * 4]",
"add {r2}, {reladdrs}",
"add {r1}, [{table} + 32]",
"mov [{table} + 32], {r2}",
"shr {r1}, 5",
"mov {r1:e}, [{reladdrs} + {r1} * 4]",
"add {r1}, {reladdrs}",
"mov [{table} + 40], {r1}",
"20:",
"vpbroadcastd {lock}, [{locks} + {i}]",
"vpand {tmp}, {lock}, {cache1}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache2}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache3}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache4}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache5}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache6}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache7}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache8}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache9}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache10}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache11}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, {cache12}",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+384]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+416]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+448]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+480]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+512]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+544]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+576]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+608]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+640]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+672]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+704]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+736]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+768]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+800]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+832]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+864]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+896]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+928]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+960]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"vpand {tmp}, {lock}, [rip + {keys}+992]",
"vpcmpeqd {tmp}, {tmp}, {vzero}",
"vpsubd {sums}, {sums}, {tmp}",
"andn {height:e}, {col_mask:e}, [{locks} + {i}]",
"popcnt {height:e}, {height:e}",
"jmp [{table} + {height} * 8]",
"200:",
".long 201f-200b",
".long 202f-200b",
".long 203f-200b",
".long 204f-200b",
".long 205f-200b",
".long 206f-200b",
".long 207f-200b",
".long 208f-200b",
".long 209f-200b",
".long 210f-200b",
".long 211f-200b",
".long 212f-200b",
".long 213f-200b",
".long 214f-200b",
".long 215f-200b",
".long 216f-200b",
".long 217f-200b",
".long 218f-200b",
".long 219f-200b",
".long 220f-200b",
".long 221f-200b",
".long 222f-200b",
".long 223f-200b",
".long 224f-200b",
".long 225f-200b",
".long 226f-200b",
".long 227f-200b",
".long 228f-200b",
".long 229f-200b",
".long 230f-200b",
".long 231f-200b",
".long 232f-200b",
"201:", "vpand {tmp}, {lock}, [rip + {keys}+992]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"202:", "vpand {tmp}, {lock}, [rip + {keys}+960]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"203:", "vpand {tmp}, {lock}, [rip + {keys}+928]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"204:", "vpand {tmp}, {lock}, [rip + {keys}+896]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"205:", "vpand {tmp}, {lock}, [rip + {keys}+864]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"206:", "vpand {tmp}, {lock}, [rip + {keys}+832]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"207:", "vpand {tmp}, {lock}, [rip + {keys}+800]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"208:", "vpand {tmp}, {lock}, [rip + {keys}+768]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"209:", "vpand {tmp}, {lock}, [rip + {keys}+736]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"210:", "vpand {tmp}, {lock}, [rip + {keys}+704]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"211:", "vpand {tmp}, {lock}, [rip + {keys}+672]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"212:", "vpand {tmp}, {lock}, [rip + {keys}+640]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"213:", "vpand {tmp}, {lock}, [rip + {keys}+608]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"214:", "vpand {tmp}, {lock}, [rip + {keys}+576]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"215:", "vpand {tmp}, {lock}, [rip + {keys}+544]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"216:", "vpand {tmp}, {lock}, [rip + {keys}+512]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"217:", "vpand {tmp}, {lock}, [rip + {keys}+480]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"218:", "vpand {tmp}, {lock}, [rip + {keys}+448]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"219:", "vpand {tmp}, {lock}, [rip + {keys}+416]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"220:", "vpand {tmp}, {lock}, [rip + {keys}+384]", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"221:", "vpand {tmp}, {lock}, {cache12}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"222:", "vpand {tmp}, {lock}, {cache11}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"223:", "vpand {tmp}, {lock}, {cache10}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"224:", "vpand {tmp}, {lock}, {cache9}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"225:", "vpand {tmp}, {lock}, {cache8}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"226:", "vpand {tmp}, {lock}, {cache7}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"227:", "vpand {tmp}, {lock}, {cache6}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"228:", "vpand {tmp}, {lock}, {cache5}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"229:", "vpand {tmp}, {lock}, {cache4}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"230:", "vpand {tmp}, {lock}, {cache3}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"231:", "vpand {tmp}, {lock}, {cache2}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"232:", "vpand {tmp}, {lock}, {cache1}", "vpcmpeqd {tmp}, {tmp}, {vzero}", "vpsubd {sums}, {sums}, {tmp}",
"add {i:e}, -4",
"jge 20b",
keys = sym KEYS,
locks = in(reg) locks,
i = inout(reg) 996usize => _,
height = out(reg) _,
col_mask = in(reg) !0b1000001000001000001000001000u32,
reladdrs = out(reg) _,
r1 = out(reg) _,
r2 = out(reg) _,
table = in(reg) indices,
vzero = in(ymm_reg) u32x8::splat(0),
tmp = out(ymm_reg) _,
lock = out(ymm_reg) _,
Expand All @@ -184,7 +221,7 @@ unsafe fn inner1(s: &[u8]) -> u32 {
cache10 = out(ymm_reg) _,
cache11 = out(ymm_reg) _,
cache12 = out(ymm_reg) _,
options(nostack, readonly),
options(nostack),
);

sums.reduce_sum() as u32
Expand Down
Loading