sse_blas_local_vcdot_double.cc File Reference

Generic Scalar VAXPY routine. More...

#include "qdp_diagnostics.h"
#include <xmmintrin.h>
#include "qdp_config.h"
#include "scalarsite_sse/sse_blas_local_vcdot_double.h"

Go to the source code of this file.

Namespaces
namespace	QDP
	Yet another random number generator.

Macros
#define	CONJMUL(z, x, y)
#define	CONJMADD(z, x, y)

Functions
void	QDP::local_vcdot4 (REAL64 *sum, REAL64 *y, REAL64 *x, int n_4spin)

Detailed Description

Generic Scalar VAXPY routine.

Definition in file sse_blas_local_vcdot_double.cc.

Macro Definition Documentation

CONJMADD

#define CONJMADD	(		z,
			x,
			y )

Value:

{ \
  __m128d t1,t2,t3,t4; \
  t1 = _mm_mul_pd(x,y); \
  t2 = _mm_shuffle_pd(t1,t1,0x1); \
  t3 = _mm_shuffle_pd(y,y,0x1);\
  t4 = _mm_add_pd(t1,t2); \
  t2 = _mm_mul_pd(x,t3); \
  t3 = _mm_shuffle_pd(t2,t2,0x1); \
  t3 = _mm_sub_pd(t2,t3); \
  t4= _mm_shuffle_pd(t4,t3,0x2); \
  z = _mm_add_pd(z,t4); \
}

Definition at line 32 of file sse_blas_local_vcdot_double.cc.

Referenced by QDP::local_vcdot4(), QDP::ssed_m_eq_hm(), QDP::ssed_m_eq_hm_u(), QDP::ssed_m_eq_mh(), QDP::ssed_m_eq_mh_u(), QDP::ssed_m_peq_ahm(), QDP::ssed_m_peq_ahm_u(), QDP::ssed_m_peq_amh(), and QDP::ssed_m_peq_amh_u().

CONJMUL

#define CONJMUL	(		z,
			x,
			y )

Value:

{ \
  __m128d t1,t2,t3,t4; \
  t1 = _mm_mul_pd(x,y); \
  t2 = _mm_shuffle_pd(t1,t1,0x1); \
  t3 = _mm_shuffle_pd(y,y,0x1);\
  z = _mm_add_pd(t1,t2); \
  t2 = _mm_mul_pd(x,t3); \
  t3 = _mm_shuffle_pd(t2,t2,0x1); \
  t3 = _mm_sub_pd(t2,t3);         \
  z= _mm_shuffle_pd(z,t3,0x2); \
}

Definition at line 19 of file sse_blas_local_vcdot_double.cc.

Referenced by QDP::ssed_m_eq_hm(), QDP::ssed_m_eq_hm_u(), QDP::ssed_m_eq_mh(), QDP::ssed_m_eq_mh_u(), QDP::ssed_m_peq_ahm(), QDP::ssed_m_peq_ahm_u(), QDP::ssed_m_peq_amh(), and QDP::ssed_m_peq_amh_u().