FM / checks

/opt/imh-python/lib/python3.9/site-packages/numpy/distutils/checks/

/opt/imh-python/lib/python3.9/site-packages/numpy/distutils/checks

Name	Size	Mode	Actions
cpu_asimd.c	818	0644	edit dl rm
cpu_asimddp.c	432	0644	edit dl rm
cpu_asimdfhm.c	529	0644	edit dl rm
cpu_asimdhp.c	379	0644	edit dl rm
cpu_avx.c	779	0644	edit dl rm
cpu_avx2.c	749	0644	edit dl rm
cpu_avx512cd.c	759	0644	edit dl rm
cpu_avx512f.c	755	0644	edit dl rm
cpu_avx512_clx.c	842	0644	edit dl rm
cpu_avx512_cnl.c	948	0644	edit dl rm
cpu_avx512_icl.c	1004	0644	edit dl rm
cpu_avx512_knl.c	956	0644	edit dl rm
cpu_avx512_knm.c	1132	0644	edit dl rm
cpu_avx512_skx.c	1010	0644	edit dl rm
cpu_avx512_spr.c	904	0644	edit dl rm
cpu_f16c.c	868	0644	edit dl rm
cpu_fma3.c	817	0644	edit dl rm
cpu_fma4.c	301	0644	edit dl rm
cpu_neon.c	600	0644	edit dl rm
cpu_neon_fp16.c	251	0644	edit dl rm
cpu_neon_vfpv4.c	609	0644	edit dl rm
cpu_popcnt.c	1049	0644	edit dl rm
cpu_sse.c	686	0644	edit dl rm
cpu_sse2.c	697	0644	edit dl rm
cpu_sse3.c	689	0644	edit dl rm
cpu_sse41.c	675	0644	edit dl rm
cpu_sse42.c	692	0644	edit dl rm
cpu_ssse3.c	705	0644	edit dl rm
cpu_vsx.c	478	0644	edit dl rm
cpu_vsx2.c	263	0644	edit dl rm
cpu_vsx3.c	250	0644	edit dl rm
cpu_vsx4.c	305	0644	edit dl rm
cpu_vx.c	461	0644	edit dl rm
cpu_vxe.c	788	0644	edit dl rm
cpu_vxe2.c	624	0644	edit dl rm
cpu_xop.c	234	0644	edit dl rm
extra_avx512bw_mask.c	636	0644	edit dl rm
extra_avx512dq_mask.c	504	0644	edit dl rm
extra_avx512f_reduce.c	1595	0644	edit dl rm
extra_vsx4_mma.c	499	0644	edit dl rm
extra_vsx_asm.c	945	0644	edit dl rm
test_flags.c	16	0644	edit dl rm

Edit: /opt/imh-python/lib/python3.9/site-packages/numpy/distutils/checks/extra_avx512f_reduce.c (1595B)

#include <immintrin.h>
/**
 * The following intrinsics don't have direct native support but compilers
 * tend to emulate them.
 * They're usually supported by gcc >= 7.1, clang >= 4 and icc >= 19
 */
int main(void)
{
    __m512  one_ps = _mm512_set1_ps(1.0f);
    __m512d one_pd = _mm512_set1_pd(1.0);
    __m512i one_i64 = _mm512_set1_epi64(1);
    // add
    float sum_ps  = _mm512_reduce_add_ps(one_ps);
    double sum_pd = _mm512_reduce_add_pd(one_pd);
    int sum_int   = (int)_mm512_reduce_add_epi64(one_i64);
        sum_int  += (int)_mm512_reduce_add_epi32(one_i64);
    // mul
    sum_ps  += _mm512_reduce_mul_ps(one_ps);
    sum_pd  += _mm512_reduce_mul_pd(one_pd);
    sum_int += (int)_mm512_reduce_mul_epi64(one_i64);
    sum_int += (int)_mm512_reduce_mul_epi32(one_i64);
    // min
    sum_ps  += _mm512_reduce_min_ps(one_ps);
    sum_pd  += _mm512_reduce_min_pd(one_pd);
    sum_int += (int)_mm512_reduce_min_epi32(one_i64);
    sum_int += (int)_mm512_reduce_min_epu32(one_i64);
    sum_int += (int)_mm512_reduce_min_epi64(one_i64);
    // max
    sum_ps  += _mm512_reduce_max_ps(one_ps);
    sum_pd  += _mm512_reduce_max_pd(one_pd);
    sum_int += (int)_mm512_reduce_max_epi32(one_i64);
    sum_int += (int)_mm512_reduce_max_epu32(one_i64);
    sum_int += (int)_mm512_reduce_max_epi64(one_i64);
    // and
    sum_int += (int)_mm512_reduce_and_epi32(one_i64);
    sum_int += (int)_mm512_reduce_and_epi64(one_i64);
    // or
    sum_int += (int)_mm512_reduce_or_epi32(one_i64);
    sum_int += (int)_mm512_reduce_or_epi64(one_i64);
    return (int)sum_ps + (int)sum_pd + sum_int;
}

#include /** * The following intrinsics don't have direct native support but compilers * tend to emulate them. * They're usually supported by gcc >= 7.1, clang >= 4 and icc >= 19 */ int main(void) { __m512 one_ps = _mm512_set1_ps(1.0f); __m512d one_pd = _mm512_set1_pd(1.0); __m512i one_i64 = _mm512_set1_epi64(1); // add float sum_ps = _mm512_reduce_add_ps(one_ps); double sum_pd = _mm512_reduce_add_pd(one_pd); int sum_int = (int)_mm512_reduce_add_epi64(one_i64); sum_int += (int)_mm512_reduce_add_epi32(one_i64); // mul sum_ps += _mm512_reduce_mul_ps(one_ps); sum_pd += _mm512_reduce_mul_pd(one_pd); sum_int += (int)_mm512_reduce_mul_epi64(one_i64); sum_int += (int)_mm512_reduce_mul_epi32(one_i64); // min sum_ps += _mm512_reduce_min_ps(one_ps); sum_pd += _mm512_reduce_min_pd(one_pd); sum_int += (int)_mm512_reduce_min_epi32(one_i64); sum_int += (int)_mm512_reduce_min_epu32(one_i64); sum_int += (int)_mm512_reduce_min_epi64(one_i64); // max sum_ps += _mm512_reduce_max_ps(one_ps); sum_pd += _mm512_reduce_max_pd(one_pd); sum_int += (int)_mm512_reduce_max_epi32(one_i64); sum_int += (int)_mm512_reduce_max_epu32(one_i64); sum_int += (int)_mm512_reduce_max_epi64(one_i64); // and sum_int += (int)_mm512_reduce_and_epi32(one_i64); sum_int += (int)_mm512_reduce_and_epi64(one_i64); // or sum_int += (int)_mm512_reduce_or_epi32(one_i64); sum_int += (int)_mm512_reduce_or_epi64(one_i64); return (int)sum_ps + (int)sum_pd + sum_int; }