AVX-512 horizontal multiply for k1om (Intel Xeon Phi Knights Corner) using Intel compiler

reduce_mul.c

// AVX-512 horizontal multiply for k1om (Intel Xeon Phi Knights Corner)
//
// (c) 2016 Dmitry Mikushin [email protected]
//
// $ icc -mmic -std=c99 -O3 reduce_mul.c -o reduce_mul
// $ micnativeloadex ./reduce_mul
// -0.004276 vs -0.004276 

#include <immintrin.h>
#include <stdio.h>
#include <stdlib.h>

static __m512i reverse_mask;

__attribute__((constructor)) void init_consts()
{
	reverse_mask = _mm512_set_epi32(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
}

static inline __m512d _mm512_reverse_pd(__m512d val)
{
	return (__m512d)_mm512_permutevar_epi32(reverse_mask, (__m512i)val);
}

static inline double _mm512_reduce_mul_pd(__m512d val)
{
	__m512d hgfe_dcba = val;
	__m512d ghef_cdab = _mm512_swizzle_pd(val, _MM_SWIZ_REG_CDAB);
	__m512d fehg_badc = _mm512_swizzle_pd(val, _MM_SWIZ_REG_BADC);
	__m512d efgh_abcd = _mm512_swizzle_pd(_mm512_swizzle_pd(val, _MM_SWIZ_REG_BADC), _MM_SWIZ_REG_CDAB);

	hgfe_dcba = _mm512_mul_pd(_mm512_mul_pd(hgfe_dcba, ghef_cdab), _mm512_mul_pd(fehg_badc, efgh_abcd));
	__m512d abcd_efgh = _mm512_reverse_pd(hgfe_dcba);
	
	double values[8] __attribute__((aligned(64)));
	_mm512_store_pd(values, _mm512_mul_pd(hgfe_dcba, abcd_efgh));
	return values[0];
}

int main(int argc, char* argv[])
{
	// Generate random input vector of [-1, 1] values.
	double values[8] __attribute__((aligned(64)));
	for (int i = 0; i < 8; i++)
		values[i] = 2 * (0.5 - rand() / (double)RAND_MAX);
	__m512d val = _mm512_load_pd(values);

	double vector = _mm512_reduce_mul_pd(val);
	double scalar = values[0];
	for (int i = 1; i < 8; i++)
		scalar *= values[i];

	printf("%f vs %f\n", vector, scalar);

	fflush(stdout);

	return 0;
}