Vectors for kernels: masks and select, shuffles and byte lookup, dot_add, @cpu and cpu_has; --cpu native on Nix

- comparisons of vectors give masks (boolxN): select(m, a, b), m.any/all/count, & | ^ !;
  f32x8(list, i, fill) and v.store_part(list, i) for the end of a list (masked load/store)
- v.shuffle(constant lanes), v.lo(), v.hi(), concat(a, b), table.lookup(indices):
  pshufb / vpshufb on x86 (SSSE3 and up), tbl on ARM64 (checked in the assembly, LLVM 18 and 21)
- acc.dot_add(a, b): i32 lanes gain 4 byte products each, exactly; vpdpbusd with AVX-VNNI or AVX-512 VNNI,
  usdot/sdot/udot on ARM64, the sum written out elsewhere (pmaddwd); run here with all three paths agreeing
- @cpu("avx2,fma") fn: LLVM target-features for one function (+evex512 for AVX-512, which LLVM 18+ needs);
  cpu_has("avx2"): CPUID + XGETBV on x86, hwcaps on Linux ARM64, sysctl on macOS; a wide vector can't be passed
  between functions with different @cpu (its registers differ), and the compiler says so
- the compiler asks clang which features --cpu means (its predefined macros, cached), so dot_add picks the instruction
- Nix's clang wrapper dropped -march=native in nix-shell (reproduced): finch tells it NIX_ENFORCE_NO_NATIVE=0, shell.nix too,
  and warns if --cpu native still made no difference

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Psc42moPLwP43Cmb9oZqSR

tests/run/dot_int8.fch

@@ -0,0 +1,69 @@
+// The int8 dot product of quantized models: plain, with dot_add, and in @cpu functions picked by cpu_has
+
+fn plain([]u8 a, []i8 b) -> int {
+    total := 0
+    for i in 0..a.len {
+        total += int(a[i]) * int(b[i])
+    }
+    return total
+}
+
+// 32 bytes a step, 8 sums kept in a vector until the end
+fn with_vectors([]u8 a, []i8 b) -> int {
+    acc := i32x8(0)
+    i := 0
+    while i + 32 <= a.len {
+        acc = acc.dot_add(u8x32(a, i), i8x32(b, i))
+        i += 32
+    }
+    return int(acc.sum())
+}
+
+@cpu("avxvnni")
+fn with_vnni([]u8 a, []i8 b) -> int {
+    acc := i32x8(0)
+    i := 0
+    while i + 32 <= a.len {
+        acc = acc.dot_add(u8x32(a, i), i8x32(b, i))
+        i += 32
+    }
+    return int(acc.sum())
+}
+
+@cpu("avx512f,avx512bw,avx512vl,avx512vnni,i8mm")
+fn with_avx512([]u8 a, []i8 b) -> int {
+    acc := i32x16(0)
+    i := 0
+    while i + 64 <= a.len {
+        acc = acc.dot_add(u8x64(a, i), i8x64(b, i))
+        i += 64
+    }
+    return int(acc.sum())
+}
+
+fn best([]u8 a, []i8 b) -> int {
+    if cpu_has("avx512vnni") && cpu_has("avx512vl") && cpu_has("avx512bw") {
+        return with_avx512(a, b)
+    }
+    if cpu_has("avxvnni") {
+        return with_vnni(a, b)
+    }
+    return with_vectors(a, b)
+}
+
+fn main() {
+    []u8 a
+    []i8 b
+    for k in 0..1024 {
+        a.push(u8((k * 37 + 11) % 256))
+        b.push(i8((k * 53 + 7) % 256 - 128))
+    }
+    want := plain(a, b)
+    print(want, with_vectors(a, b) == want, best(a, b) == want)
+
+    // the extremes: 255 * -128 four times in a lane, no saturation
+    acc := i32x4(1, 2, 3, 4).dot_add(u8x16(255), i8x16(-128))
+    print(acc, i32x4(0).dot_add(i8x16(-128), i8x16(-128)), i32x4(0).dot_add(u8x16(255), u8x16(255)))
+    print(i32x4(0).dot_add(i8x16(-1), u8x16(200)))
+    print(cpu_has("avx2") || cpu_has("neon") || true)
+}