SIMD

Returns the name of the vector-instruction path Open Graph's kernels run on.

const char *og_simd_path(void);

One of "avx512" (AVX-512 F/CD/BW/DQ/VL with the operating system saving the wide registers; bit counts use VPOPCNTDQ where present), "avx2" (AVX2 with FMA, BMI1/2, LZCNT, POPCNT, F16C and MOVBE), "neon" (every ARM64 CPU) or "scalar" (plain C: the reference, and the path of an x86-64 CPU without AVX2). The path is chosen once, at the first use, from what the CPU and operating system support; the environment variable OG_SIMD=scalar|avx2|avx512|neon chooses a lower path, to compare paths on one machine (a path the CPU cannot run is ignored).

Results never depend on the path: element-wise kernels do the plain loop's IEEE operations in its order, never fused, and every floating-point reduction uses one fixed order on every path (element i to lane i mod 16, each lane summed in increasing i, the 16 lanes combined by a fixed tree), so all paths agree to the last bit. The kernels serve bit sets, connected-component reachability counts, Jaccard and Dice similarity, the Fruchterman-Reingold, Kamada-Kawai and DrL layouts, the BLAS-style vector operations and the eigenvalue solvers (eigenvector centrality and others).

Returns

A static string: "avx512", "avx2", "neon" or "scalar".

Complexity

O(1) after the first call (which runs the CPU detection once, thread safe).

Checked

Open Graph extension, no igraph counterpart. tests/simd checks every kernel on every path the CPU runs, bit for bit, and the library functions that use them for identical results on every path.

SIMD

Part of the SIMD extension: it reports the path the kernels run on and runs no kernel itself.

C name

og_simd_path

In notebooks

Og.Native.og_simd_path

igraph 1.0

None: Open Graph's own.

See also

simd_features

#include "og.h"
#include <stdio.h>

int main(void) {
    char cpu[256];
    const char *path, *best, *lowered;
    og_simd_features(cpu, sizeof cpu, &path, &best, &lowered);
    printf("running on %s (best %s)%s%s\n", og_simd_path(), best,
           lowered ? ", lowered by OG_SIMD=" : "", lowered ? lowered : "");
    printf("CPU: %s\n", cpu);
    return 0;
}

Reports what the CPU detection saw and which path it chose.

void og_simd_features(
    char *buf,
    size_t cap,
    const char **path,
    const char **best,
    const char **lowered
);

buf receives the relevant instruction sets the CPU has, separated by spaces (on x86-64 for example "SSE4.2 POPCNT AVX AVX2 FMA ... AVX-512F ... AVX-512VPOPCNTDQ", on ARM64 "NEON"), cut to fit cap bytes. *path receives the path in use (as simd_path), *best the best path this CPU can run, and *lowered the value of OG_SIMD when it chose a lower path, otherwise NULL. Every output may be NULL. On x86-64 the AVX paths need the operating system to save the wide registers (XGETBV), not only the CPUID bits.

Parameters

• buf Output, may be NULL: the instruction sets found.
• cap Size of buf in bytes.
• path Output, may be NULL: the path in use.
• best Output, may be NULL: the best path this CPU runs.
• lowered Output, may be NULL: OG_SIMD's value when it lowered the path, else NULL.

Returns

Nothing.

Complexity

O(cap).

Checked

Open Graph extension, no igraph counterpart.

SIMD

Part of the SIMD extension: it reports what the CPU offers and runs no kernel itself.

C name

og_simd_features

In notebooks

Og.Native.og_simd_features

igraph 1.0

None: Open Graph's own.

See also

simd_path