pfor: move ISA checks into turbopfor package (compile-time)

Also build with GOAMD64=v4, as 3 of our 6 SIMD kernels are AVX512 anyway.

This only saves a few instructions, but is straight-forward enough.
The effect is more pronounced on the go-stream implementation,
where the per-value overhead is higher in general.

pfordec % make bench-relative
[…]
benchstat -filter '/impl:go .unit:(Mval/s)' baseline.txt bench.txt
goos: linux
goarch: amd64
pkg: github.com/Debian/dcs/internal/turbopfor/pfordec
cpu: AMD Ryzen 9 9950X3D 16-Core Processor
                                              │ baseline.txt │              bench.txt              │
                                              │    Mval/s    │    Mval/s     vs base               │
Decode/n=2048/vals=all-zero/impl=go              21.42k ± 2%   22.51k ±  2%   +5.11% (p=0.002 n=6)
Decode/n=2048/vals=all-constant/impl=go          21.65k ± 1%   21.87k ±  2%   +1.02% (p=0.009 n=6)
Decode/n=2048/vals=bitpacking-bw1/impl=go        30.20k ± 1%   19.26k ±  0%  -36.24% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-bw2/impl=go        33.84k ± 1%   34.80k ±  1%   +2.84% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-bw7/impl=go        32.34k ± 0%   33.20k ±  0%   +2.66% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-bw1-exc/impl=go    8.017k ± 0%   9.052k ±  1%  +12.91% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-bw2-exc/impl=go    9.060k ± 1%   9.168k ±  0%   +1.20% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-bw7-exc/impl=go    8.983k ± 1%   9.189k ±  1%   +2.29% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-vb-exc/impl=go     19.71k ± 1%   19.38k ±  0%   -1.66% (p=0.002 n=6)
Decode/n=2048/vals=sparse-exc/impl=go            7.135k ± 0%   7.159k ±  1%        ~ (p=0.394 n=6)
Decode/n=2048/vals=sparse-vb-exc/impl=go         12.80k ± 3%   12.71k ±  1%        ~ (p=0.394 n=6)
Decode/n=2048/vals=debian-mix/impl=go            11.87k ± 2%   12.16k ±  1%   +2.52% (p=0.002 n=6)
Decode/n=2039/vals=all-zero/impl=go              21.54k ± 0%   22.09k ±  1%   +2.57% (p=0.002 n=6)
Decode/n=2039/vals=all-constant/impl=go          21.58k ± 1%   21.57k ±  9%        ~ (p=0.818 n=6)
Decode/n=2039/vals=bitpacking-bw1/impl=go        23.43k ± 1%   16.68k ±  0%  -28.80% (p=0.002 n=6)
Decode/n=2039/vals=bitpacking-bw2/impl=go        25.81k ± 1%   26.03k ±  1%   +0.84% (p=0.004 n=6)
Decode/n=2039/vals=bitpacking-bw7/impl=go        25.37k ± 4%   25.29k ±  2%        ~ (p=0.699 n=6)
Decode/n=2039/vals=bitpacking-bw1-exc/impl=go    5.958k ± 0%   6.388k ±  0%   +7.23% (p=0.002 n=6)
Decode/n=2039/vals=bitpacking-bw2-exc/impl=go    6.558k ± 0%   6.584k ±  1%   +0.40% (p=0.022 n=6)
Decode/n=2039/vals=bitpacking-bw7-exc/impl=go    6.467k ± 0%   6.535k ±  0%   +1.06% (p=0.002 n=6)
Decode/n=2039/vals=bitpacking-vb-exc/impl=go     16.38k ± 1%   16.41k ±  1%        ~ (p=0.732 n=6)
Decode/n=2039/vals=sparse-exc/impl=go            5.944k ± 0%   5.921k ±  1%        ~ (p=0.307 n=6)
Decode/n=2039/vals=sparse-vb-exc/impl=go         12.27k ± 5%   12.88k ±  0%   +4.99% (p=0.002 n=6)
Decode/n=2039/vals=debian-mix/impl=go            7.632k ± 2%   7.840k ±  1%   +2.73% (p=0.002 n=6)
Decode/n=160/vals=all-zero/impl=go               19.66k ± 1%   18.16k ± 16%   -7.63% (p=0.002 n=6)
Decode/n=160/vals=all-constant/impl=go           14.89k ± 3%   14.51k ±  2%   -2.54% (p=0.041 n=6)
Decode/n=160/vals=bitpacking-bw1/impl=go         6.840k ± 1%   6.902k ±  1%   +0.91% (p=0.002 n=6)
Decode/n=160/vals=bitpacking-bw2/impl=go         6.922k ± 1%   6.995k ±  0%        ~ (p=0.132 n=6)
Decode/n=160/vals=bitpacking-bw7/impl=go         6.961k ± 3%   7.168k ±  1%   +2.97% (p=0.026 n=6)
Decode/n=160/vals=bitpacking-bw1-exc/impl=go     2.168k ± 3%   2.218k ±  2%        ~ (p=0.065 n=6)
Decode/n=160/vals=bitpacking-bw2-exc/impl=go     2.228k ± 1%   2.230k ±  1%        ~ (p=0.738 n=6)
Decode/n=160/vals=bitpacking-bw7-exc/impl=go     2.167k ± 1%   2.182k ±  0%   +0.69% (p=0.041 n=6)
Decode/n=160/vals=bitpacking-vb-exc/impl=go      6.605k ± 4%   6.612k ±  0%        ~ (p=1.000 n=6)
Decode/n=160/vals=sparse-exc/impl=go             2.554k ± 1%   2.576k ±  1%   +0.88% (p=0.002 n=6)
Decode/n=160/vals=sparse-vb-exc/impl=go          11.05k ± 8%   12.18k ±  6%  +10.24% (p=0.015 n=6)
Decode/n=160/vals=debian-mix/impl=go             2.231k ± 2%   2.228k ±  1%        ~ (p=0.786 n=6)
geomean                                          10.11k        10.04k         -0.68%
[…]
benchstat -filter '/impl:go-stream .unit:(Mval/s)' baseline.txt bench.txt
goos: linux
goarch: amd64
pkg: github.com/Debian/dcs/internal/turbopfor/pfordec
cpu: AMD Ryzen 9 9950X3D 16-Core Processor
                                                     │ baseline.txt │             bench.txt             │
                                                     │    Mval/s    │   Mval/s     vs base              │
Decode/n=2048/vals=all-zero/impl=go-stream              20.21k ± 1%   20.41k ± 1%       ~ (p=0.093 n=6)
Decode/n=2048/vals=all-constant/impl=go-stream          20.35k ± 2%   20.68k ± 2%  +1.62% (p=0.041 n=6)
Decode/n=2048/vals=bitpacking-bw1/impl=go-stream        24.71k ± 0%   24.84k ± 1%  +0.55% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-bw2/impl=go-stream        30.56k ± 0%   30.72k ± 0%  +0.52% (p=0.015 n=6)
Decode/n=2048/vals=bitpacking-bw7/impl=go-stream        30.48k ± 1%   30.73k ± 0%  +0.80% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-bw1-exc/impl=go-stream    8.728k ± 0%   8.891k ± 0%  +1.87% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-bw2-exc/impl=go-stream    8.861k ± 0%   8.919k ± 1%  +0.65% (p=0.019 n=6)
Decode/n=2048/vals=bitpacking-bw7-exc/impl=go-stream    8.816k ± 1%   9.023k ± 0%  +2.35% (p=0.002 n=6)
Decode/n=2048/vals=bitpacking-vb-exc/impl=go-stream     18.69k ± 0%   18.85k ± 0%  +0.87% (p=0.002 n=6)
Decode/n=2048/vals=sparse-exc/impl=go-stream            7.006k ± 1%   7.061k ± 1%  +0.79% (p=0.043 n=6)
Decode/n=2048/vals=sparse-vb-exc/impl=go-stream         12.13k ± 3%   12.31k ± 0%       ~ (p=0.065 n=6)
Decode/n=2048/vals=debian-mix/impl=go-stream            11.53k ± 2%   11.68k ± 0%  +1.29% (p=0.015 n=6)
Decode/n=2039/vals=all-zero/impl=go-stream              20.22k ± 0%   20.19k ± 2%       ~ (p=0.485 n=6)
Decode/n=2039/vals=all-constant/impl=go-stream          20.23k ± 1%   20.40k ± 4%       ~ (p=0.240 n=6)
Decode/n=2039/vals=bitpacking-bw1/impl=go-stream        20.11k ± 0%   20.13k ± 0%       ~ (p=0.132 n=6)
Decode/n=2039/vals=bitpacking-bw2/impl=go-stream        24.16k ± 0%   24.18k ± 0%  +0.10% (p=0.006 n=6)
Decode/n=2039/vals=bitpacking-bw7/impl=go-stream        23.84k ± 0%   23.94k ± 0%  +0.43% (p=0.002 n=6)
Decode/n=2039/vals=bitpacking-bw1-exc/impl=go-stream    6.285k ± 1%   6.332k ± 1%  +0.74% (p=0.002 n=6)
Decode/n=2039/vals=bitpacking-bw2-exc/impl=go-stream    6.391k ± 0%   6.431k ± 1%  +0.63% (p=0.002 n=6)
Decode/n=2039/vals=bitpacking-bw7-exc/impl=go-stream    6.317k ± 0%   6.425k ± 0%  +1.70% (p=0.002 n=6)
Decode/n=2039/vals=bitpacking-vb-exc/impl=go-stream     15.60k ± 1%   15.66k ± 0%       ~ (p=0.331 n=6)
Decode/n=2039/vals=sparse-exc/impl=go-stream            5.832k ± 0%   5.820k ± 2%       ~ (p=1.000 n=6)
Decode/n=2039/vals=sparse-vb-exc/impl=go-stream         11.98k ± 2%   12.39k ± 1%  +3.40% (p=0.002 n=6)
Decode/n=2039/vals=debian-mix/impl=go-stream            7.436k ± 1%   7.615k ± 1%  +2.41% (p=0.002 n=6)
Decode/n=160/vals=all-zero/impl=go-stream               17.25k ± 5%   17.53k ± 3%  +1.62% (p=0.039 n=6)
Decode/n=160/vals=all-constant/impl=go-stream           12.61k ± 1%   12.21k ± 1%  -3.16% (p=0.002 n=6)
Decode/n=160/vals=bitpacking-bw1/impl=go-stream         6.781k ± 1%   6.798k ± 0%  +0.24% (p=0.002 n=6)
Decode/n=160/vals=bitpacking-bw2/impl=go-stream         7.146k ± 3%   7.184k ± 1%       ~ (p=0.240 n=6)
Decode/n=160/vals=bitpacking-bw7/impl=go-stream         7.510k ± 3%   7.609k ± 1%  +1.33% (p=0.011 n=6)
Decode/n=160/vals=bitpacking-bw1-exc/impl=go-stream     2.197k ± 1%   2.171k ± 1%       ~ (p=0.061 n=6)
Decode/n=160/vals=bitpacking-bw2-exc/impl=go-stream     2.164k ± 2%   2.156k ± 1%       ~ (p=0.310 n=6)
Decode/n=160/vals=bitpacking-bw7-exc/impl=go-stream     2.150k ± 1%   2.168k ± 1%  +0.88% (p=0.019 n=6)
Decode/n=160/vals=bitpacking-vb-exc/impl=go-stream      6.646k ± 1%   6.644k ± 1%       ~ (p=0.909 n=6)
Decode/n=160/vals=sparse-exc/impl=go-stream             2.530k ± 1%   2.549k ± 1%  +0.75% (p=0.015 n=6)
Decode/n=160/vals=sparse-vb-exc/impl=go-stream          10.39k ± 6%   11.19k ± 4%  +7.69% (p=0.015 n=6)
Decode/n=160/vals=debian-mix/impl=go-stream             2.181k ± 1%   2.163k ± 1%  -0.85% (p=0.006 n=6)
geomean                                                 9.716k        9.800k       +0.87%

pforenc % make bench-relative
[…]
benchstat -filter '/impl:go .unit:(Mval/s)' baseline.txt bench.txt
goos: linux
goarch: amd64
pkg: github.com/Debian/dcs/internal/turbopfor/pforenc
cpu: AMD Ryzen 9 9950X3D 16-Core Processor
                                              │ baseline.txt │             bench.txt             │
                                              │    Mval/s    │   Mval/s     vs base              │
Encode/n=2048/vals=all-zero/impl=go             19.21k ±  1%   19.17k ± 6%       ~ (p=0.310 n=6)
Encode/n=2048/vals=all-constant/impl=go         19.25k ±  1%   19.06k ± 6%       ~ (p=0.310 n=6)
Encode/n=2048/vals=bitpacking-bw1/impl=go       6.649k ±  0%   7.283k ± 3%  +9.53% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-bw2/impl=go       6.424k ±  0%   6.975k ± 2%  +8.58% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-bw7/impl=go       5.253k ±  0%   5.559k ± 0%  +5.84% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-bw1-exc/impl=go   2.055k ± 10%   2.151k ± 8%  +4.67% (p=0.039 n=6)
Encode/n=2048/vals=bitpacking-bw2-exc/impl=go   2.083k ±  1%   2.086k ± 3%       ~ (p=0.149 n=6)
Encode/n=2048/vals=bitpacking-bw7-exc/impl=go   2.232k ±  1%   2.301k ± 1%  +3.09% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-vb-exc/impl=go    1.696k ±  3%   1.728k ± 2%  +1.86% (p=0.002 n=6)
Encode/n=2048/vals=sparse-exc/impl=go           2.997k ±  1%   3.138k ± 1%  +4.70% (p=0.002 n=6)
Encode/n=2048/vals=sparse-vb-exc/impl=go        1.825k ±  1%   1.868k ± 1%  +2.33% (p=0.002 n=6)
Encode/n=2048/vals=debian-mix/impl=go           2.150k ±  2%   2.245k ± 1%  +4.42% (p=0.002 n=6)
Encode/n=2039/vals=all-zero/impl=go             18.80k ±  2%   18.70k ± 1%       ~ (p=0.699 n=6)
Encode/n=2039/vals=all-constant/impl=go         18.48k ±  4%   18.72k ± 1%       ~ (p=0.290 n=6)
Encode/n=2039/vals=bitpacking-bw1/impl=go       6.014k ±  0%   6.474k ± 2%  +7.64% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw2/impl=go       5.739k ± 10%   6.125k ± 0%  +6.73% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw7/impl=go       4.668k ±  1%   4.931k ± 2%  +5.63% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw1-exc/impl=go   1.907k ±  1%   1.986k ± 0%  +4.14% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw2-exc/impl=go   1.867k ±  3%   1.974k ± 3%  +5.70% (p=0.004 n=6)
Encode/n=2039/vals=bitpacking-bw7-exc/impl=go   1.932k ±  7%   1.995k ± 0%       ~ (p=0.084 n=6)
Encode/n=2039/vals=bitpacking-vb-exc/impl=go    1.563k ±  3%   1.579k ± 3%       ~ (p=0.143 n=6)
Encode/n=2039/vals=sparse-exc/impl=go           2.974k ±  1%   3.135k ± 0%  +5.41% (p=0.002 n=6)
Encode/n=2039/vals=sparse-vb-exc/impl=go        1.843k ±  4%   1.944k ± 0%  +5.48% (p=0.002 n=6)
Encode/n=2039/vals=debian-mix/impl=go           1.952k ±  1%   2.050k ± 1%  +4.97% (p=0.002 n=6)
Encode/n=160/vals=all-zero/impl=go              15.93k ±  6%   15.65k ± 5%       ~ (p=0.818 n=6)
Encode/n=160/vals=all-constant/impl=go          15.56k ±  5%   15.54k ± 4%       ~ (p=1.000 n=6)
Encode/n=160/vals=bitpacking-bw1/impl=go        3.454k ±  0%   3.464k ± 1%       ~ (p=0.247 n=6)
Encode/n=160/vals=bitpacking-bw2/impl=go        3.295k ±  0%   3.296k ± 2%       ~ (p=0.558 n=6)
Encode/n=160/vals=bitpacking-bw7/impl=go        2.736k ±  0%   2.742k ± 0%       ~ (p=0.078 n=6)
Encode/n=160/vals=bitpacking-bw1-exc/impl=go    1.207k ±  0%   1.213k ± 1%       ~ (p=0.199 n=6)
Encode/n=160/vals=bitpacking-bw2-exc/impl=go    1.208k ±  3%   1.177k ± 0%  -2.57% (p=0.019 n=6)
Encode/n=160/vals=bitpacking-bw7-exc/impl=go    1.260k ±  1%   1.263k ± 1%       ~ (p=0.102 n=6)
Encode/n=160/vals=bitpacking-vb-exc/impl=go     1.067k ±  1%   1.034k ± 0%  -3.14% (p=0.002 n=6)
Encode/n=160/vals=sparse-exc/impl=go            2.627k ±  6%   2.635k ± 0%       ~ (p=0.056 n=6)
Encode/n=160/vals=sparse-vb-exc/impl=go         1.874k ±  5%   1.849k ± 0%       ~ (p=0.368 n=6)
Encode/n=160/vals=debian-mix/impl=go            1.202k ±  0%   1.161k ± 0%  -3.45% (p=0.002 n=6)
geomean                                         3.363k         3.439k       +2.28%
[…]
benchstat -filter '/impl:go-stream .unit:(Mval/s)' baseline.txt bench.txt
goos: linux
goarch: amd64
pkg: github.com/Debian/dcs/internal/turbopfor/pforenc
cpu: AMD Ryzen 9 9950X3D 16-Core Processor
                                                     │ baseline.txt │              bench.txt              │
                                                     │    Mval/s    │    Mval/s     vs base               │
Encode/n=2048/vals=all-zero/impl=go-stream             2.208k ±  6%   2.193k ±  1%        ~ (p=0.394 n=6)
Encode/n=2048/vals=all-constant/impl=go-stream         2.122k ±  5%   2.056k ±  7%        ~ (p=0.180 n=6)
Encode/n=2048/vals=bitpacking-bw1/impl=go-stream       1.776k ±  0%   1.827k ±  1%   +2.84% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-bw2/impl=go-stream       1.758k ±  0%   1.821k ±  2%   +3.53% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-bw7/impl=go-stream       1.651k ±  0%   1.723k ±  1%   +4.36% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-bw1-exc/impl=go-stream   1.069k ±  9%   1.149k ±  2%   +7.39% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-bw2-exc/impl=go-stream   1.120k ±  1%   1.165k ±  2%   +4.06% (p=0.002 n=6)
Encode/n=2048/vals=bitpacking-bw7-exc/impl=go-stream   1.163k ±  4%   1.131k ± 19%        ~ (p=0.167 n=6)
Encode/n=2048/vals=bitpacking-vb-exc/impl=go-stream     979.5 ±  4%    979.3 ±  1%        ~ (p=1.000 n=6)
Encode/n=2048/vals=sparse-exc/impl=go-stream           1.347k ±  0%   1.385k ±  0%   +2.78% (p=0.002 n=6)
Encode/n=2048/vals=sparse-vb-exc/impl=go-stream        1.040k ±  5%   1.060k ±  4%   +1.88% (p=0.037 n=6)
Encode/n=2048/vals=debian-mix/impl=go-stream           1.138k ±  1%   1.179k ±  2%   +3.60% (p=0.002 n=6)
Encode/n=2039/vals=all-zero/impl=go-stream             2.223k ± 12%   2.289k ±  0%   +3.01% (p=0.026 n=6)
Encode/n=2039/vals=all-constant/impl=go-stream         2.068k ±  8%   2.273k ±  1%   +9.91% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw1/impl=go-stream       1.736k ±  0%   1.843k ±  1%   +6.19% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw2/impl=go-stream       1.714k ±  4%   1.823k ±  0%   +6.36% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw7/impl=go-stream       1.597k ±  1%   1.703k ±  4%   +6.60% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw1-exc/impl=go-stream   1.071k ±  1%   1.105k ±  2%   +3.22% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw2-exc/impl=go-stream   1.079k ±  4%   1.121k ±  1%   +3.89% (p=0.002 n=6)
Encode/n=2039/vals=bitpacking-bw7-exc/impl=go-stream   1.115k ±  0%   1.147k ±  5%        ~ (p=0.058 n=6)
Encode/n=2039/vals=bitpacking-vb-exc/impl=go-stream     970.1 ±  3%    993.3 ±  5%        ~ (p=0.394 n=6)
Encode/n=2039/vals=sparse-exc/impl=go-stream           1.353k ±  4%   1.383k ±  3%   +2.18% (p=0.024 n=6)
Encode/n=2039/vals=sparse-vb-exc/impl=go-stream        1.077k ±  4%   1.124k ±  3%   +4.32% (p=0.002 n=6)
Encode/n=2039/vals=debian-mix/impl=go-stream           1.067k ±  3%   1.065k ±  2%        ~ (p=0.732 n=6)
Encode/n=160/vals=all-zero/impl=go-stream              2.046k ±  1%   1.825k ±  1%  -10.76% (p=0.002 n=6)
Encode/n=160/vals=all-constant/impl=go-stream          2.038k ±  2%   1.817k ±  1%  -10.84% (p=0.002 n=6)
Encode/n=160/vals=bitpacking-bw1/impl=go-stream        1.419k ±  0%   1.330k ±  3%   -6.27% (p=0.002 n=6)
Encode/n=160/vals=bitpacking-bw2/impl=go-stream        1.384k ±  0%   1.269k ±  3%   -8.31% (p=0.002 n=6)
Encode/n=160/vals=bitpacking-bw7/impl=go-stream        1.266k ±  0%   1.155k ±  1%   -8.73% (p=0.002 n=6)
Encode/n=160/vals=bitpacking-bw1-exc/impl=go-stream     798.7 ±  1%    801.8 ±  0%        ~ (p=0.065 n=6)
Encode/n=160/vals=bitpacking-bw2-exc/impl=go-stream     797.0 ±  2%    787.2 ±  2%        ~ (p=0.554 n=6)
Encode/n=160/vals=bitpacking-bw7-exc/impl=go-stream     821.3 ±  2%    825.1 ±  0%        ~ (p=0.065 n=6)
Encode/n=160/vals=bitpacking-vb-exc/impl=go-stream      751.3 ±  2%    721.0 ±  3%   -4.03% (p=0.002 n=6)
Encode/n=160/vals=sparse-exc/impl=go-stream            1.296k ±  3%   1.235k ±  3%   -4.71% (p=0.002 n=6)
Encode/n=160/vals=sparse-vb-exc/impl=go-stream         1.080k ±  3%   1.000k ±  4%   -7.41% (p=0.002 n=6)
Encode/n=160/vals=debian-mix/impl=go-stream             797.3 ±  1%    772.1 ±  4%        ~ (p=0.394 n=6)
geomean                                                1.292k         1.294k         +0.15%

Author: stapelberg

internal/turbopfor/cpu_amd64.go

@@ -0,0 +1,12 @@
+//go:build goexperiment.simd && amd64 && !amd64.v3
+
+package turbopfor
+
+import "simd/archsimd"
+
+// Compiled with GOAMD64=v1 or =v2,
+// need to use runtime detection to find supported ISAs
+
+var HasAVX2 = archsimd.X86.AVX2()
+
+var HasAVX512 = archsimd.X86.AVX512()

internal/turbopfor/cpu_amd64v3.go

@@ -0,0 +1,9 @@
+//go:build goexperiment.simd && amd64 && amd64.v3 && !amd64.v4
+
+package turbopfor
+
+import "simd/archsimd"
+
+const HasAVX2 = true
+
+var HasAVX512 = archsimd.X86.AVX512()

internal/turbopfor/cpu_amd64v4.go

@@ -0,0 +1,7 @@
+//go:build goexperiment.simd && amd64 && amd64.v4
+
+package turbopfor
+
+const HasAVX2 = true
+
+const HasAVX512 = true

internal/turbopfor/cpu_nosimd.go

@@ -0,0 +1,7 @@
+//go:build !goexperiment.simd || !amd64
+
+package turbopfor
+
+const HasAVX2 = false
+
+const HasAVX512 = false

internal/turbopfor/pfordec/Makefile

@@ -1,5 +1,5 @@
 # Required for the Go compiler to intrinsify LZCNT, for example.
-export GOAMD64=v3
+export GOAMD64=v4
 # Required for SIMD support with Go 1.27.
 # Always set (even before adding SIMD code) so the baseline is stable.
 export GOEXPERIMENT=simd

internal/turbopfor/pfordec/bitunpack256v_amd64.go

@@ -2,10 +2,14 @@
 
 package pfordec
 
-import "simd/archsimd"
+import (
+	"simd/archsimd"
+
+	"github.com/Debian/dcs/internal/turbopfor"
+)
 
 func bitunpack256v32(fullinput []byte, fulloutput []uint32, nbits int) (read int) {
-	if !hasAVX2 {
+	if !turbopfor.HasAVX2 {
 		return bitunpack256v32Scalar(fullinput, fulloutput, nbits)
 	}
 	output := (*[256]uint32)(fulloutput[:256])

internal/turbopfor/pfordec/bitunpack256vex_amd64.go

@@ -5,12 +5,14 @@ package pfordec
 import (
 	"math/bits"
 	"simd/archsimd"
+
+	"github.com/Debian/dcs/internal/turbopfor"
 )
 
 // bitunpack256v32Ex is like bitunpack256v32, but with exception decoding fused
 // into the same loop (instead of a separate pass).
 func bitunpack256v32Ex(input []byte, fulloutput []uint32, nbits int, exmap *[32]byte, exceptions *[256]uint32) (read int) {
-	if !hasAVX512 {
+	if !turbopfor.HasAVX512 {
 		return bitunpack256v32ExScalar(input, fulloutput, nbits, exmap, exceptions)
 	}
 	output := (*[256]uint32)(fulloutput[:256])

internal/turbopfor/pfordec/bitunpack_amd64.go

@@ -4,9 +4,9 @@ package pfordec
 
 import (
 	"simd/archsimd"
-)
 
-var hasAVX512 = archsimd.X86.AVX512()
+	"github.com/Debian/dcs/internal/turbopfor"
+)
 
 // seqUnpackConsts are the per-lane constants of bitunpackSIMD for one
 // bit width. This allows keeping the dispatch unit busy with vector
@@ -32,7 +32,7 @@ var precomputed = func() (t [33]seqUnpackConsts) {
 }()
 
 func bitunpack(input []byte, output []uint32, bitWidth int) (read int) {
-	if !hasAVX2 {
+	if !turbopfor.HasAVX2 {
 		return bitunpackScalar(input, output, bitWidth)
 	}
 	if bitWidth == 0 {

internal/turbopfor/pfordec/constant_amd64.go

@@ -2,12 +2,14 @@
 
 package pfordec
 
-import "simd/archsimd"
+import (
+	"simd/archsimd"
 
-var hasAVX2 = archsimd.X86.AVX2()
+	"github.com/Debian/dcs/internal/turbopfor"
+)
 
 func fillConstant(output []uint32, val uint32) {
-	if !hasAVX2 {
+	if !turbopfor.HasAVX2 {
 		fillConstantScalar(output, val)
 		return
 	}