1336 lines
		
	
	
		
			45 KiB
		
	
	
	
		
			ArmAsm
		
	
	
	
			
		
		
	
	
			1336 lines
		
	
	
		
			45 KiB
		
	
	
	
		
			ArmAsm
		
	
	
	
/****************************************Implementation**Details**********************************************/
 | 
						|
/*                                                                                                           */
 | 
						|
/*   Lets denote (a,a1i) complex which is mathematically a+a1*i                                              */
 | 
						|
/*   Complex number multiplication: (a,a1i)*(b,b1i)                                                          */
 | 
						|
/*   As i*i=-1 .The multiplication result will be:                                                           */
 | 
						|
/*   (a+a1*i)(b+b1*i)=a*b+a1*i*b1*i+ a1*i*b+a*b1*i=a*b-a1*b1 + (a1*b+a*b1)*i which is (ab-a1b1,a1b+ab1)      */
 | 
						|
/*   so let  c= ab-a1b1 , ci=a1b+ab1 then                                                                    */
 | 
						|
/*   c=c+a*b-a1*b1    => c=a*b-( a1*b1-c)  => c= a1*b1-c then c=a*b-c two mseb                               */
 | 
						|
/*   ci=ci+a1*b+a*b1  => ci= a1*b+ci then ci= a*b1+ci                                                        */
 | 
						|
/*   For simd real and imaginary parts will be grouped together                                              */
 | 
						|
/*   such (realA,realK) and (imageA ,imageK)                                                                 */
 | 
						|
/*   Simd(0,1)=(a*b,k*b)-((ai*bi,ki*bi)-Simd(0,1))                                                           */
 | 
						|
/*   SimdI(0,1)=SimdI(0,1)+(a*bi,k*bi)+(ai*b,ki*b)                                                           */
 | 
						|
/*                                                                                                           */
 | 
						|
/*                                                                                                           */
 | 
						|
/*   for defined(NR) || defined(NC) || defined(TR) || defined(TC)                                            */
 | 
						|
/*   (a+a1*I)(b-b1*I)=ab+a1*b1+I(a1b-ab1)                                                                    */
 | 
						|
/*                                                                                                           */
 | 
						|
/*   c=c+ab+a1b1  => c=a1b1+c;c=ab+c                                                                         */
 | 
						|
/*   ci=ci+a1b-ab1 => ci=a1*b-(ab1-ci) => ci=ab1-ci; ci=a1*b-ci                                              */
 | 
						|
/*                                                                                                           */
 | 
						|
/*                                                                                                           */
 | 
						|
/*   for  defined(RN) || defined(RT) || defined(CN) || defined(CT)                                           */
 | 
						|
/*   (a-a1*I)(b+b1*I)=ab+a1*b1+I(-a1b+ab1)                                                                   */
 | 
						|
/*                                                                                                           */
 | 
						|
/*   c=c+ab+a1b1  => c=a1b1+c;c=ab+c                                                                         */
 | 
						|
/*   ci=ci+a1b-ab1 => ci=a*b1-(a1b-ci) => ci=a1b-ci; ci=a*b1-ci                                              */
 | 
						|
/*                                                                                                           */
 | 
						|
/*                                                                                                           */
 | 
						|
/*   for defined(RR) || defined(RC) || defined(CR) || defined(CC)                                            */
 | 
						|
/*   (a-a1*I)(b-b1*I)=ab-a1*b1+I(-a1b-ab1)                                                                   */
 | 
						|
/*                                                                                                           */
 | 
						|
/*   c= a1*b1-c then c=a*b-c                                                                                 */
 | 
						|
/*   ci = ci-a1*b -a*b1;                                                                                     */
 | 
						|
/*   as ibm z13 only has x*z-m x*z+m  instructions implementation will  be changed a bit                     */
 | 
						|
/*   Assuming  ci=0; and cix=cix+a1b+ab1 ;   ci=ci-cix will work                                             */
 | 
						|
/*   cix= a*b1+cix ; cix= a1*b+cix  (two madb) ci=ci-cix (sign change if ci=0)                               */
 | 
						|
/*   As c=0   then                                                                                           */
 | 
						|
/*   c=a*b-c then c=a1*b1-c => c=(a1*b1-(a*b-c))  which is -1*( a*b -(a1*b1-c))                              */
 | 
						|
/*                                                                                                           */
 | 
						|
/*   Values will be equal to (-c) and (-ci)                                                                  */
 | 
						|
/*   To change sign it'll be multiplied by -1*(alpha+alpha_i)                                                */
 | 
						|
/*   This is done once:                                                                                      */
 | 
						|
/*   lcdbr ALPHA_I,ALPHA_I                                                                                   */
 | 
						|
/*   lcdbr ALPHA ,ALPHA                                                                                      */
 | 
						|
/*************************************************************************************************************/
 | 
						|
 | 
						|
/*************************Zero vectors***************************************/
 | 
						|
/*zero vectors for 4x4 */
 | 
						|
.macro ZERO_ZCVEC_4x4
 | 
						|
    vzero  %v16
 | 
						|
    vzero  %v17
 | 
						|
    vzero  %v18
 | 
						|
    vzero  %v19
 | 
						|
    vzero  %v20
 | 
						|
    vzero  %v21
 | 
						|
    vzero  %v22
 | 
						|
    vzero  %v23
 | 
						|
    vzero  %v24
 | 
						|
    vzero  %v25
 | 
						|
    vzero  %v26
 | 
						|
    vzero  %v27
 | 
						|
    vzero  %v28
 | 
						|
    vzero  %v29
 | 
						|
    vzero  %v30
 | 
						|
    vzero  %v31
 | 
						|
.endm
 | 
						|
 | 
						|
/*zero vectors for */
 | 
						|
.macro ZERO_ZCVEC_2x4
 | 
						|
    vzero  %v16
 | 
						|
    vzero  %v17
 | 
						|
    vzero  %v18
 | 
						|
    vzero  %v19
 | 
						|
    vzero  %v20
 | 
						|
    vzero  %v21
 | 
						|
    vzero  %v22
 | 
						|
    vzero  %v23
 | 
						|
.endm
 | 
						|
 | 
						|
/*zero vectors for */
 | 
						|
.macro ZERO_ZCVEC_1x4
 | 
						|
    vzero  %v16
 | 
						|
    vzero  %v17
 | 
						|
    vzero  %v18
 | 
						|
    vzero  %v19
 | 
						|
.endm
 | 
						|
 | 
						|
/*zero vectors for */
 | 
						|
.macro ZERO_ZCVEC_4x2
 | 
						|
   ZERO_ZCVEC_2x4
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZERO_ZCVEC_4x1
 | 
						|
   ZERO_ZCVEC_1x4
 | 
						|
.endm
 | 
						|
 | 
						|
/*zero vectors for */
 | 
						|
.macro ZERO_ZCVEC_2x2
 | 
						|
    vzero  %v16
 | 
						|
    vzero  %v17
 | 
						|
    vzero  %v20
 | 
						|
    vzero  %v21
 | 
						|
.endm
 | 
						|
 | 
						|
/*zero vectors for */
 | 
						|
.macro ZERO_ZCVEC_1x2
 | 
						|
    vzero  %v16
 | 
						|
    vzero  %v17
 | 
						|
.endm
 | 
						|
 | 
						|
/*zero vectors for */
 | 
						|
.macro ZERO_ZCVEC_2x1
 | 
						|
    vzero  %v16
 | 
						|
    vzero  %v17
 | 
						|
.endm
 | 
						|
 | 
						|
/*zero vectors for 1x1*/
 | 
						|
.macro ZERO_ZCVEC_1x1
 | 
						|
    lzdr %f6
 | 
						|
    lzdr %f7
 | 
						|
.endm
 | 
						|
 | 
						|
 | 
						|
/*
 | 
						|
  Calculate for 4x2 inner
 | 
						|
*/
 | 
						|
.macro CalcComplex_4x2 vResR1, vResI1, vResR2, vResI2, vResR3, vResI3, vResR4, vResI4, vr1, vi1, vr2, vi2, vrB, viB,vrB2, viB2
 | 
						|
 | 
						|
  #if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 | 
						|
    vfmsdb \vResR1, \vi1, \viB, \vResR1
 | 
						|
    vfmadb \vResI1, \vr1, \viB, \vResI1
 | 
						|
    vfmsdb \vResR2, \vi2, \viB, \vResR2
 | 
						|
    vfmadb \vResI2, \vr2, \viB, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR3, \vi1, \viB2, \vResR3
 | 
						|
    vfmadb \vResI3, \vr1, \viB2, \vResI3
 | 
						|
    vfmsdb \vResR4, \vi2, \viB2, \vResR4
 | 
						|
    vfmadb \vResI4, \vr2, \viB2, \vResI4
 | 
						|
 | 
						|
    vfmsdb \vResR1, \vr1, \vrB, \vResR1
 | 
						|
    vfmadb \vResI1, \vi1, \vrB, \vResI1
 | 
						|
    vfmsdb \vResR2, \vr2, \vrB, \vResR2
 | 
						|
    vfmadb \vResI2, \vi2, \vrB, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR3, \vr1, \vrB2, \vResR3
 | 
						|
    vfmadb \vResI3, \vi1, \vrB2, \vResI3
 | 
						|
    vfmsdb \vResR4, \vr2, \vrB2, \vResR4
 | 
						|
    vfmadb \vResI4, \vi2, \vrB2, \vResI4
 | 
						|
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
 | 
						|
    vfmadb \vResR1, \vi1, \viB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vr1, \viB, \vResI1
 | 
						|
    vfmadb \vResR2, \vi2, \viB, \vResR2
 | 
						|
    vfmsdb \vResI2, \vr2, \viB, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR3, \vi1, \viB2, \vResR3
 | 
						|
    vfmsdb \vResI3, \vr1, \viB2, \vResI3
 | 
						|
    vfmadb \vResR4, \vi2, \viB2, \vResR4
 | 
						|
    vfmsdb \vResI4, \vr2, \viB2, \vResI4
 | 
						|
 | 
						|
    vfmadb \vResR1, \vr1, \vrB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vi1, \vrB, \vResI1
 | 
						|
    vfmadb \vResR2, \vr2, \vrB, \vResR2
 | 
						|
    vfmsdb \vResI2, \vi2, \vrB, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR3, \vr1, \vrB2, \vResR3
 | 
						|
    vfmsdb \vResI3, \vi1, \vrB2, \vResI3
 | 
						|
    vfmadb \vResR4, \vr2, \vrB2, \vResR4
 | 
						|
    vfmsdb \vResI4, \vi2, \vrB2, \vResI4
 | 
						|
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
 | 
						|
    vfmadb \vResR1, \vi1, \viB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vi1, \vrB, \vResI1
 | 
						|
    vfmadb \vResR2, \vi2, \viB, \vResR2
 | 
						|
    vfmsdb \vResI2, \vi2, \vrB, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR3, \vi1, \viB2, \vResR3
 | 
						|
    vfmsdb \vResI3, \vi1, \vrB2, \vResI3
 | 
						|
    vfmadb \vResR4, \vi2, \viB2, \vResR4
 | 
						|
    vfmsdb \vResI4, \vi2, \vrB2, \vResI4
 | 
						|
 | 
						|
    vfmadb \vResR1, \vr1, \vrB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vr1, \viB, \vResI1
 | 
						|
    vfmadb \vResR2, \vr2, \vrB, \vResR2
 | 
						|
    vfmsdb \vResI2, \vr2, \viB, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR3, \vr1, \vrB2, \vResR3
 | 
						|
    vfmsdb \vResI3, \vr1, \viB2, \vResI3
 | 
						|
    vfmadb \vResR4, \vr2, \vrB2, \vResR4
 | 
						|
    vfmsdb \vResI4, \vr2, \viB2, \vResI4
 | 
						|
  #endif
 | 
						|
  #if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
 | 
						|
 | 
						|
    vfmsdb \vResR1, \vr1, \vrB, \vResR1
 | 
						|
    vfmadb \vResI1, \vi1, \vrB, \vResI1
 | 
						|
    vfmsdb \vResR2, \vr2, \vrB, \vResR2
 | 
						|
    vfmadb \vResI2, \vi2, \vrB, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR3, \vr1, \vrB2, \vResR3
 | 
						|
    vfmadb \vResI3, \vi1, \vrB2, \vResI3
 | 
						|
    vfmsdb \vResR4, \vr2, \vrB2, \vResR4
 | 
						|
    vfmadb \vResI4, \vi2, \vrB2, \vResI4
 | 
						|
 | 
						|
    vfmsdb \vResR1, \vi1, \viB, \vResR1
 | 
						|
    vfmadb \vResI1, \vr1, \viB, \vResI1
 | 
						|
    vfmsdb \vResR2, \vi2, \viB, \vResR2
 | 
						|
    vfmadb \vResI2, \vr2, \viB, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR3, \vi1, \viB2, \vResR3
 | 
						|
    vfmadb \vResI3, \vr1, \viB2, \vResI3
 | 
						|
    vfmsdb \vResR4, \vi2, \viB2, \vResR4
 | 
						|
    vfmadb \vResI4, \vr2, \viB2, \vResI4
 | 
						|
 | 
						|
 | 
						|
  #endif
 | 
						|
 | 
						|
.endm
 | 
						|
 | 
						|
/*
 | 
						|
  Calculate for 2x4 inner
 | 
						|
*/
 | 
						|
.macro CalcComplex_2x4 vResR1, vResI1, vResR2, vResI2, vResR3, vResI3, vResR4, vResI4, vr1, vi1, vr2, vi2, vrB, viB,vrB2, viB2
 | 
						|
 | 
						|
  #if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 | 
						|
    vfmsdb \vResR1, \vi1, \viB, \vResR1
 | 
						|
    vfmadb \vResI1, \vr1, \viB, \vResI1
 | 
						|
    vfmsdb \vResR2, \vi2, \viB, \vResR2
 | 
						|
    vfmadb \vResI2, \vr2, \viB, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR3, \vi1, \viB2, \vResR3
 | 
						|
    vfmadb \vResI3, \vr1, \viB2, \vResI3
 | 
						|
    vfmsdb \vResR4, \vi2, \viB2, \vResR4
 | 
						|
    vfmadb \vResI4, \vr2, \viB2, \vResI4
 | 
						|
 | 
						|
    vfmsdb \vResR1, \vr1, \vrB, \vResR1
 | 
						|
    vfmadb \vResI1, \vi1, \vrB, \vResI1
 | 
						|
    vfmsdb \vResR2, \vr2, \vrB, \vResR2
 | 
						|
    vfmadb \vResI2, \vi2, \vrB, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR3, \vr1, \vrB2, \vResR3
 | 
						|
    vfmadb \vResI3, \vi1, \vrB2, \vResI3
 | 
						|
    vfmsdb \vResR4, \vr2, \vrB2, \vResR4
 | 
						|
    vfmadb \vResI4, \vi2, \vrB2, \vResI4
 | 
						|
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
 | 
						|
    vfmadb \vResR1, \vi1, \viB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vr1, \viB, \vResI1
 | 
						|
    vfmadb \vResR2, \vi2, \viB, \vResR2
 | 
						|
    vfmsdb \vResI2, \vr2, \viB, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR3, \vi1, \viB2, \vResR3
 | 
						|
    vfmsdb \vResI3, \vr1, \viB2, \vResI3
 | 
						|
    vfmadb \vResR4, \vi2, \viB2, \vResR4
 | 
						|
    vfmsdb \vResI4, \vr2, \viB2, \vResI4
 | 
						|
 | 
						|
    vfmadb \vResR1, \vr1, \vrB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vi1, \vrB, \vResI1
 | 
						|
    vfmadb \vResR2, \vr2, \vrB, \vResR2
 | 
						|
    vfmsdb \vResI2, \vi2, \vrB, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR3, \vr1, \vrB2, \vResR3
 | 
						|
    vfmsdb \vResI3, \vi1, \vrB2, \vResI3
 | 
						|
    vfmadb \vResR4, \vr2, \vrB2, \vResR4
 | 
						|
    vfmsdb \vResI4, \vi2, \vrB2, \vResI4
 | 
						|
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
 | 
						|
    vfmadb \vResR1, \vi1, \viB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vi1, \vrB, \vResI1
 | 
						|
    vfmadb \vResR2, \vi2, \viB, \vResR2
 | 
						|
    vfmsdb \vResI2, \vi2, \vrB, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR3, \vi1, \viB2, \vResR3
 | 
						|
    vfmsdb \vResI3, \vi1, \vrB2, \vResI3
 | 
						|
    vfmadb \vResR4, \vi2, \viB2, \vResR4
 | 
						|
    vfmsdb \vResI4, \vi2, \vrB2, \vResI4
 | 
						|
 | 
						|
    vfmadb \vResR1, \vr1, \vrB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vr1, \viB, \vResI1
 | 
						|
    vfmadb \vResR2, \vr2, \vrB, \vResR2
 | 
						|
    vfmsdb \vResI2, \vr2, \viB, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR3, \vr1, \vrB2, \vResR3
 | 
						|
    vfmsdb \vResI3, \vr1, \viB2, \vResI3
 | 
						|
    vfmadb \vResR4, \vr2, \vrB2, \vResR4
 | 
						|
    vfmsdb \vResI4, \vr2, \viB2, \vResI4
 | 
						|
  #endif
 | 
						|
  #if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
 | 
						|
 | 
						|
    vfmsdb \vResR1, \vr1, \vrB, \vResR1
 | 
						|
    vfmadb \vResI1, \vi1, \vrB, \vResI1
 | 
						|
    vfmsdb \vResR2, \vr2, \vrB, \vResR2
 | 
						|
    vfmadb \vResI2, \vi2, \vrB, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR3, \vr1, \vrB2, \vResR3
 | 
						|
    vfmadb \vResI3, \vi1, \vrB2, \vResI3
 | 
						|
    vfmsdb \vResR4, \vr2, \vrB2, \vResR4
 | 
						|
    vfmadb \vResI4, \vi2, \vrB2, \vResI4
 | 
						|
 | 
						|
    vfmsdb \vResR1, \vi1, \viB, \vResR1
 | 
						|
    vfmadb \vResI1, \vr1, \viB, \vResI1
 | 
						|
    vfmsdb \vResR2, \vi2, \viB, \vResR2
 | 
						|
    vfmadb \vResI2, \vr2, \viB, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR3, \vi1, \viB2, \vResR3
 | 
						|
    vfmadb \vResI3, \vr1, \viB2, \vResI3
 | 
						|
    vfmsdb \vResR4, \vi2, \viB2, \vResR4
 | 
						|
    vfmadb \vResI4, \vr2, \viB2, \vResI4
 | 
						|
 | 
						|
 | 
						|
  #endif
 | 
						|
 | 
						|
.endm
 | 
						|
 | 
						|
/*
 | 
						|
  Calculate for 2x2 inner
 | 
						|
*/
 | 
						|
.macro CalcComplex_2x2 vResR1, vResI1,vResR2, vResI2, vR1, vI1, vRB, vIB,  vRB2, vIB2
 | 
						|
  #if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 | 
						|
    vfmsdb \vResR1, \vI1, \vIB, \vResR1
 | 
						|
    vfmadb \vResI1, \vR1, \vIB, \vResI1
 | 
						|
 | 
						|
    vfmsdb \vResR2, \vI1, \vIB2, \vResR2
 | 
						|
    vfmadb \vResI2, \vR1, \vIB2, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR1, \vR1, \vRB, \vResR1
 | 
						|
    vfmadb \vResI1, \vI1, \vRB, \vResI1
 | 
						|
 | 
						|
    vfmsdb \vResR2, \vR1, \vRB2, \vResR2
 | 
						|
    vfmadb \vResI2, \vI1, \vRB2, \vResI2
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
 | 
						|
    vfmadb \vResR1, \vI1, \vIB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vR1, \vIB, \vResI1
 | 
						|
 | 
						|
    vfmadb \vResR2, \vI1, \vIB2, \vResR2
 | 
						|
    vfmsdb \vResI2, \vR1, \vIB2, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR1, \vR1, \vRB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vI1, \vRB, \vResI1
 | 
						|
 | 
						|
    vfmadb \vResR2, \vR1, \vRB2, \vResR2
 | 
						|
    vfmsdb \vResI2, \vI1, \vRB2, \vResI2
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
 | 
						|
    vfmadb \vResR1, \vI1, \vIB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vI1, \vRB, \vResI1
 | 
						|
 | 
						|
    vfmadb \vResR2, \vI1, \vIB2, \vResR2
 | 
						|
    vfmsdb \vResI2, \vI1, \vRB2, \vResI2
 | 
						|
 | 
						|
    vfmadb \vResR1, \vR1, \vRB, \vResR1
 | 
						|
    vfmsdb \vResI1, \vR1, \vIB, \vResI1
 | 
						|
 | 
						|
    vfmadb \vResR2, \vR1, \vRB2, \vResR2
 | 
						|
    vfmsdb \vResI2, \vR1, \vIB2, \vResI2
 | 
						|
  #endif
 | 
						|
  #if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
 | 
						|
    vfmsdb \vResR1, \vR1, \vRB, \vResR1
 | 
						|
    vfmadb \vResI1, \vI1, \vRB, \vResI1
 | 
						|
 | 
						|
    vfmsdb \vResR2, \vR1, \vRB2, \vResR2
 | 
						|
    vfmadb \vResI2, \vI1, \vRB2, \vResI2
 | 
						|
 | 
						|
    vfmsdb \vResR1, \vI1, \vIB, \vResR1
 | 
						|
    vfmadb \vResI1, \vR1, \vIB, \vResI1
 | 
						|
 | 
						|
    vfmsdb \vResR2, \vI1, \vIB2, \vResR2
 | 
						|
    vfmadb \vResI2, \vR1, \vIB2, \vResI2
 | 
						|
  #endif
 | 
						|
.endm
 | 
						|
 | 
						|
/*
 | 
						|
  Calculate for 2x1 inner
 | 
						|
*/
 | 
						|
.macro CalcComplex_2x1 vRealResult1, vImageResult1, vReal1, vImage1, vecRealB, vecImageB
 | 
						|
  #if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
 | 
						|
    vfmadb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmadb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
 | 
						|
    vfmadb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmadb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
  #endif
 | 
						|
  #if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
  #endif
 | 
						|
.endm
 | 
						|
 | 
						|
/*
 | 
						|
  Calculate for 1x2 inner
 | 
						|
*/
 | 
						|
.macro CalcComplex_1x2 vRealResult1, vImageResult1, vReal1, vImage1, vecRealB, vecImageB
 | 
						|
  #if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(RN) || defined(CN) || defined(RT) || defined(CT)
 | 
						|
    vfmadb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmadb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(NR) || defined(TR) || defined(NC) || defined(TC)
 | 
						|
    vfmadb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmadb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
  #endif
 | 
						|
  #if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
  #endif
 | 
						|
.endm
 | 
						|
 | 
						|
 | 
						|
/*
 | 
						|
  Calculate for 4x1 inner
 | 
						|
*/
 | 
						|
.macro CalcComplex_4x1 vRealResult1, vImageResult1, vRealResult2, vImageResult2, vReal1, vImage1, vReal2, vImage2, vecRealB, vecImageB
 | 
						|
  #if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
 | 
						|
    vfmadb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmadb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmsdb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
    vfmadb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmadb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmsdb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
 | 
						|
    vfmadb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmadb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmsdb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
    vfmadb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmadb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmsdb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
  #endif
 | 
						|
  #if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
 | 
						|
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
  #endif
 | 
						|
 | 
						|
.endm
 | 
						|
 | 
						|
/*
 | 
						|
  Calculate for 1x4 inner
 | 
						|
*/
 | 
						|
.macro CalcComplex_1x4 vRealResult1, vImageResult1, vRealResult2, vImageResult2, vReal1, vImage1, vReal2, vImage2, vecRealB, vecImageB
 | 
						|
  #if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(RN) || defined(CN) || defined(RT) || defined(CT)
 | 
						|
    vfmadb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmadb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmsdb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
    vfmadb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmadb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmsdb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(NR) || defined(TR) || defined(NC) || defined(TC)
 | 
						|
    vfmadb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmadb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmsdb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
    vfmadb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmsdb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmadb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmsdb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
  #endif
 | 
						|
  #if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
 | 
						|
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
  #endif
 | 
						|
 | 
						|
.endm
 | 
						|
 | 
						|
.macro CalcComplex_1x1 RealResult1, ImageResult1, Real1, Image1, RealB, ImageB
 | 
						|
  #if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 | 
						|
    msdbr \RealResult1, \Image1, \ImageB
 | 
						|
    madbr \ImageResult1, \Real1, \ImageB
 | 
						|
    msdbr \RealResult1, \Real1, \RealB
 | 
						|
    madbr \ImageResult1, \Image1, \RealB
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
 | 
						|
    madbr \RealResult1, \Image1, \ImageB
 | 
						|
    msdbr \ImageResult1, \Real1, \ImageB
 | 
						|
    madbr \RealResult1, \Real1, \RealB
 | 
						|
    msdbr \ImageResult1, \Image1, \RealB
 | 
						|
  #endif
 | 
						|
 | 
						|
  #if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
 | 
						|
    madbr \RealResult1, \Image1, \ImageB
 | 
						|
    msdbr \ImageResult1, \Image1, \RealB
 | 
						|
    madbr \RealResult1, \Real1, \RealB
 | 
						|
    msdbr \ImageResult1, \Real1, \ImageB
 | 
						|
  #endif
 | 
						|
  #if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
 | 
						|
    msdbr \RealResult1, \Real1, \RealB
 | 
						|
    madbr \ImageResult1, \Image1, \RealB
 | 
						|
    msdbr \RealResult1, \Image1, \ImageB
 | 
						|
    madbr \ImageResult1, \Real1, \ImageB
 | 
						|
  #endif
 | 
						|
.endm
 | 
						|
 | 
						|
#define DISP(ind,stride,disp) (ind*stride+disp)
 | 
						|
#define DISP64(ind,disp) (ind*64+disp)
 | 
						|
#define DISP32(ind,disp) (ind*32+disp)
 | 
						|
#define DISP16(ind,disp) (ind*16+disp)
 | 
						|
#define USE_VLM 1
 | 
						|
 | 
						|
.macro ZCALC_4x4_I   PTR_A_REG,PTR_B_REG,Index,IsLast
 | 
						|
#if defined(USE_VLM)
 | 
						|
    vlm %v4,%v7, DISP64(\Index ,0) (\PTR_A_REG)
 | 
						|
#else
 | 
						|
    vl %v4 , DISP64(\Index ,0) (\PTR_A_REG)
 | 
						|
    vl %v5 , DISP64(\Index ,16)(\PTR_A_REG)
 | 
						|
    vl %v6 , DISP64(\Index ,32)(\PTR_A_REG)
 | 
						|
    vl %v7 , DISP64(\Index ,48)(\PTR_A_REG)
 | 
						|
#endif
 | 
						|
 | 
						|
    vlrepg %v9,  DISP64(\Index ,0)(\PTR_B_REG)
 | 
						|
    vlrepg %v10 , DISP64(\Index ,8)(\PTR_B_REG)
 | 
						|
    vlrepg %v11,  DISP64(\Index ,16)(\PTR_B_REG)
 | 
						|
    vlrepg %v12 , DISP64(\Index ,24)(\PTR_B_REG)
 | 
						|
 | 
						|
    vpdi %v1,%v4,%v5,0
 | 
						|
    vpdi %v5,%v4,%v5,0b101
 | 
						|
    vpdi %v3,%v6,%v7,0
 | 
						|
    vpdi %v7,%v6,%v7,0b101
 | 
						|
 | 
						|
    CalcComplex_4x2 %v16,%v17,%v18,%v19,%v20,%v21,%v22,%v23,%v1,%v5,%v3,%v7,%v9,%v10,%v11,%v12
 | 
						|
 | 
						|
    vlrepg %v9,  DISP64(\Index ,32)(\PTR_B_REG)
 | 
						|
    vlrepg %v10 , DISP64(\Index ,40)(\PTR_B_REG)
 | 
						|
    vlrepg %v11,  DISP64(\Index ,48)(\PTR_B_REG)
 | 
						|
    vlrepg %v12 , DISP64(\Index ,56)(\PTR_B_REG)
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP64(\Index ,64)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_4x2 %v24,%v25,%v26,%v27,%v28,%v29,%v30,%v31,%v1,%v5,%v3,%v7,%v9,%v10,%v11,%v12
 | 
						|
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP64(\Index ,64)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_4x2_I   PTR_A_REG,PTR_B_REG,Index,IsLast
 | 
						|
#if defined(USE_VLM)
 | 
						|
    vlm %v4,%v7, DISP64(\Index ,0) (\PTR_A_REG)
 | 
						|
#else
 | 
						|
    vl %v4 , DISP64(\Index ,0) (\PTR_A_REG)
 | 
						|
    vl %v5 , DISP64(\Index ,16)(\PTR_A_REG)
 | 
						|
    vl %v6 , DISP64(\Index ,32)(\PTR_A_REG)
 | 
						|
    vl %v7 , DISP64(\Index ,48)(\PTR_A_REG)
 | 
						|
#endif
 | 
						|
    vlrepg %v9,  DISP32(\Index ,0)(\PTR_B_REG)
 | 
						|
    vlrepg %v10 , DISP32(\Index ,8)(\PTR_B_REG)
 | 
						|
    vlrepg %v11,  DISP32(\Index ,16)(\PTR_B_REG)
 | 
						|
    vlrepg %v12 , DISP32(\Index ,24)(\PTR_B_REG)
 | 
						|
 | 
						|
    vpdi %v1,%v4,%v5,0
 | 
						|
    vpdi %v5,%v4,%v5,0b101
 | 
						|
    vpdi %v3,%v6,%v7,0
 | 
						|
    vpdi %v7,%v6,%v7,0b101
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP64(\Index ,64)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_4x2 %v16,%v17,%v18,%v19,%v20,%v21,%v22,%v23,%v1,%v5,%v3,%v7,%v9,%v10,%v11,%v12
 | 
						|
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP32(\Index ,32)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_2x4_I    PTR_A_REG,PTR_B_REG,Index,IsLast
 | 
						|
#if defined(USE_VLM)
 | 
						|
    vlm %v4,%v7, DISP64(\Index ,0) (\PTR_B_REG)
 | 
						|
#else
 | 
						|
    vl %v4 , DISP64(\Index ,0) (\PTR_B_REG)
 | 
						|
    vl %v5 , DISP64(\Index ,16)(\PTR_B_REG)
 | 
						|
    vl %v6 , DISP64(\Index ,32)(\PTR_B_REG)
 | 
						|
    vl %v7 , DISP64(\Index ,48)(\PTR_B_REG)
 | 
						|
#endif
 | 
						|
    vlrepg %v9,  DISP32(\Index ,0)(\PTR_A_REG)
 | 
						|
    vlrepg %v10 , DISP32(\Index ,8)(\PTR_A_REG)
 | 
						|
    vlrepg %v11,  DISP32(\Index ,16)(\PTR_A_REG)
 | 
						|
    vlrepg %v12 , DISP32(\Index ,24)(\PTR_A_REG)
 | 
						|
 | 
						|
    vpdi %v1,%v4,%v5,0
 | 
						|
    vpdi %v5,%v4,%v5,0b101
 | 
						|
    vpdi %v3,%v6,%v7,0
 | 
						|
    vpdi %v7,%v6,%v7,0b101
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP64(\Index ,64)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_2x4 %v16,%v17,%v18,%v19,%v20,%v21,%v22,%v23,%v1,%v5,%v3,%v7,%v9,%v10,%v11,%v12
 | 
						|
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP32(\Index ,32)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_4x1_I   PTR_A_REG,PTR_B_REG,Index,IsLast
 | 
						|
#if defined(USE_VLM)
 | 
						|
    vlm %v4,%v7, DISP64(\Index ,0) (\PTR_A_REG)
 | 
						|
#else
 | 
						|
    vl %v4 , DISP64(\Index ,0) (\PTR_A_REG)
 | 
						|
    vl %v5 , DISP64(\Index ,16)(\PTR_A_REG)
 | 
						|
    vl %v6 , DISP64(\Index ,32)(\PTR_A_REG)
 | 
						|
    vl %v7 , DISP64(\Index ,48)(\PTR_A_REG)
 | 
						|
#endif
 | 
						|
    vlrepg %v9,  DISP16(\Index ,0)(\PTR_B_REG)
 | 
						|
    vlrepg %v10 , DISP16(\Index ,8)(\PTR_B_REG)
 | 
						|
 | 
						|
    vpdi %v1,%v4,%v5,0
 | 
						|
    vpdi %v11,%v4,%v5,0b101
 | 
						|
    vpdi %v3,%v6,%v7,0
 | 
						|
    vpdi %v12,%v6,%v7,0b101
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP64(\Index ,64)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_4x1 %v16,%v17,%v18,%v19,%v1,%v11,%v3,%v12,%v9,%v10
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP16(\Index ,16)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_1x4_I    PTR_A_REG,PTR_B_REG,Index,IsLast
 | 
						|
#if defined(USE_VLM)
 | 
						|
    vlm %v4,%v7, DISP64(\Index ,0) (\PTR_B_REG)
 | 
						|
#else
 | 
						|
    vl %v4 , DISP64(\Index ,0) (\PTR_B_REG)
 | 
						|
    vl %v5 , DISP64(\Index ,16)(\PTR_B_REG)
 | 
						|
    vl %v6 , DISP64(\Index ,32)(\PTR_B_REG)
 | 
						|
    vl %v7 , DISP64(\Index ,48)(\PTR_B_REG)
 | 
						|
#endif
 | 
						|
    vlrepg %v9,  DISP16(\Index ,0)(\PTR_A_REG)
 | 
						|
    vlrepg %v10 , DISP16(\Index ,8)(\PTR_A_REG)
 | 
						|
 | 
						|
    vpdi %v1,%v4,%v5,0
 | 
						|
    vpdi %v11,%v4,%v5,0b101
 | 
						|
    vpdi %v3,%v6,%v7,0
 | 
						|
    vpdi %v12,%v6,%v7,0b101
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP64(\Index ,64)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_1x4 %v16,%v17,%v18,%v19,%v1,%v11,%v3,%v12,%v9,%v10
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP16(\Index ,16)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_2x2_I   PTR_A_REG,PTR_B_REG ,Index,IsLast
 | 
						|
    vl %v1 , DISP32(\Index ,0)(\PTR_A_REG)
 | 
						|
    vl %v3 , DISP32(\Index ,16)(\PTR_A_REG)
 | 
						|
    vlrepg %v9,  DISP32(\Index ,0)(\PTR_B_REG)
 | 
						|
    vlrepg %v10 , DISP32(\Index ,8)(\PTR_B_REG)
 | 
						|
    vlrepg %v11,  DISP32(\Index ,16)(\PTR_B_REG)
 | 
						|
    vlrepg %v12 , DISP32(\Index ,24)(\PTR_B_REG)
 | 
						|
    vpdi %v5,%v1,%v3,0
 | 
						|
    vpdi %v6,%v1,%v3,0b101
 | 
						|
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP32(\Index ,32)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_2x2 %v16,%v17,%v20,%v21,%v5,%v6, %v9,%v10,%v11,%v12
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP32(\Index ,32)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_2x1_I   PTR_A_REG,PTR_B_REG ,Index,IsLast
 | 
						|
    vl %v1 , DISP32(\Index ,0)(\PTR_A_REG)
 | 
						|
    vl %v3 , DISP32(\Index ,16)(\PTR_A_REG)
 | 
						|
    vlrepg %v6,  DISP16(\Index ,0)(\PTR_B_REG)
 | 
						|
    vlrepg %v7 , DISP16(\Index ,8)(\PTR_B_REG)
 | 
						|
    vpdi %v4,%v1,%v3,0
 | 
						|
    vpdi %v5,%v1,%v3,0b101
 | 
						|
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP32(\Index ,32)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_2x1 %v16,%v17,%v4,%v5,%v6,%v7
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP16(\Index ,16)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_1x2_I   PTR_A_REG,PTR_B_REG ,Index,IsLast
 | 
						|
    vl %v1 , DISP32(\Index ,0)(\PTR_B_REG)
 | 
						|
    vl %v3 , DISP32(\Index ,16)(\PTR_B_REG)
 | 
						|
    vlrepg %v6,  DISP16(\Index ,0)(\PTR_A_REG)
 | 
						|
    vlrepg %v7 , DISP16(\Index ,8)(\PTR_A_REG)
 | 
						|
    vpdi %v4,%v1,%v3,0
 | 
						|
    vpdi %v5,%v1,%v3,0b101
 | 
						|
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP32(\Index ,32)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_1x2 %v16,%v17,%v4,%v5,%v6,%v7
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP16(\Index ,16)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_1x1_I   PTR_A_REG,PTR_B_REG ,Index,IsLast
 | 
						|
    ld %f1 , DISP16(\Index ,0)(\PTR_A_REG)
 | 
						|
    ld %f3 , DISP16(\Index ,8)(\PTR_A_REG)
 | 
						|
    ld %f4 , DISP16(\Index ,0)(\PTR_B_REG)
 | 
						|
    ld %f5 , DISP16(\Index ,8)(\PTR_B_REG)
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_A_REG, DISP16(\Index ,16)(\PTR_A_REG)
 | 
						|
  .endif
 | 
						|
    CalcComplex_1x1 %f6,%f7,%f1,%f3,%f4,%f5
 | 
						|
  .if \IsLast==1
 | 
						|
    la \PTR_B_REG, DISP16(\Index ,16)(\PTR_B_REG)
 | 
						|
  .endif
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_4x4   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_4x4_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
.macro ZCALC_4x2   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_4x2_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
.macro ZCALC_4x1   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_4x1_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_4x4_4   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_4x4_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_4x4_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_4x4_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_4x4_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
.macro ZCALC_4x2_4   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_4x2_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_4x2_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_4x2_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_4x2_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
.macro ZCALC_4x1_4   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_4x1_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_4x1_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_4x1_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_4x1_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_2x4_4   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_2x4_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_2x4_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_2x4_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_2x4_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_2x4    PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_2x4_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_1x4_4   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_1x4_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_1x4_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_1x4_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_1x4_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_1x4    PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_1x4_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
.macro ZCALC_2x2   PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_2x2_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_2x2_4    PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_2x2_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_2x2_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_2x2_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_2x2_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_2x1    PTR_A_REG,PTR_B_REG
 | 
						|
   ZCALC_2x1_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_2x1_4    PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_2x1_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_2x1_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_2x1_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_2x1_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
 | 
						|
 | 
						|
.macro ZCALC_1x2_4    PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_1x2_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_1x2_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_1x2_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_1x2_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_1x2    PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_1x2_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_1x1_4    PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_1x1_I \PTR_A_REG,\PTR_B_REG,0,0
 | 
						|
    ZCALC_1x1_I \PTR_A_REG,\PTR_B_REG,1,0
 | 
						|
    ZCALC_1x1_I \PTR_A_REG,\PTR_B_REG,2,0
 | 
						|
    ZCALC_1x1_I \PTR_A_REG,\PTR_B_REG,3,1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZCALC_1x1    PTR_A_REG,PTR_B_REG
 | 
						|
    ZCALC_1x1_I \PTR_A_REG,\PTR_B_REG,0,1
 | 
						|
.endm
 | 
						|
 | 
						|
 | 
						|
 | 
						|
/*****************************STORE RESULTS************************************/
 | 
						|
.macro CalcMultAlpha_4x1 vRealResult1, vImageResult1, vRealResult2, vImageResult2, vReal1, vImage1, vReal2, vImage2, vecRealB, vecImageB
 | 
						|
  #if defined (TRMMKERNEL)
 | 
						|
    vfmdb \vRealResult1, \vImage1, \vecImageB
 | 
						|
    vfmdb \vImageResult1, \vReal1, \vecImageB
 | 
						|
    vfmdb \vRealResult2, \vImage2, \vecImageB
 | 
						|
    vfmdb \vImageResult2, \vReal2, \vecImageB
 | 
						|
  #else
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vImage2, \vecImageB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vReal2, \vecImageB, \vImageResult2
 | 
						|
#endif
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
    vfmsdb \vRealResult2, \vReal2, \vecRealB, \vRealResult2
 | 
						|
    vfmadb \vImageResult2, \vImage2, \vecRealB, \vImageResult2
 | 
						|
 | 
						|
.endm
 | 
						|
 | 
						|
.macro CalcMultAlpha_2x1 vRealResult1, vImageResult1, vReal1, vImage1, vecRealB, vecImageB
 | 
						|
  #if defined (TRMMKERNEL)
 | 
						|
    vfmdb \vRealResult1, \vImage1, \vecImageB
 | 
						|
    vfmdb \vImageResult1, \vReal1, \vecImageB
 | 
						|
#else
 | 
						|
    vfmsdb \vRealResult1, \vImage1, \vecImageB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vReal1, \vecImageB, \vImageResult1
 | 
						|
#endif
 | 
						|
    vfmsdb \vRealResult1, \vReal1, \vecRealB, \vRealResult1
 | 
						|
    vfmadb \vImageResult1, \vImage1, \vecRealB, \vImageResult1
 | 
						|
.endm
 | 
						|
 | 
						|
.macro CalcMultAlpha_1x1    RealResult1, ImageResult1, Real1, Image1, RealB, ImageB
 | 
						|
 | 
						|
    msdbr \RealResult1, \Image1, \ImageB
 | 
						|
    madbr \ImageResult1, \Real1, \ImageB
 | 
						|
    msdbr \RealResult1, \Real1, \RealB
 | 
						|
    madbr \ImageResult1, \Image1, \RealB
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZSTORE_4x4  ALPHA_VECREG,ALPHA_VECI,CIJ_REG , LDC_BYTE_ORIGINAL ,LC1,LC2
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
   vl %v1 , 0(\CIJ_REG)
 | 
						|
    vl %v4 , 16(\CIJ_REG)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vl %v7 , 32(\CIJ_REG)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v6 ,  48 (\CIJ_REG)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
#endif
 | 
						|
    la \LC1,0(\LDC_BYTE_ORIGINAL, \LDC_BYTE_ORIGINAL)
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v16,%v17,%v18,%v19,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    la \LC2,0(\LC1,\LDC_BYTE_ORIGINAL )
 | 
						|
    vpdi %v17, %v3,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG)
 | 
						|
    vpdi %v18, %v1 ,%v6,0
 | 
						|
    vst %v17,16(\CIJ_REG)
 | 
						|
    vpdi %v19, %v1 ,%v6,0b0101
 | 
						|
    vst %v18,32(\CIJ_REG)
 | 
						|
    vst %v19,48(\CIJ_REG)
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vl %v4 , 16(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vl %v7 , 32(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v6 ,  48 (\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v20,%v21,%v22,%v23,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3 ,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v18, %v1 ,%v6,0
 | 
						|
    vst %v17,16(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v19, %v1 ,%v6,0b0101
 | 
						|
    vst %v18,32(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vst %v19,48(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
 | 
						|
#if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG,\LC1)
 | 
						|
    vl %v4 , 16(\CIJ_REG,\LC1)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vl %v7 , 32(\CIJ_REG,\LC1)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v6 ,  48 (\CIJ_REG,\LC1)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v24,%v25,%v26,%v27,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3 ,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG,\LC1)
 | 
						|
    vpdi %v18, %v1 ,%v6,0
 | 
						|
    vst %v17,16(\CIJ_REG,\LC1)
 | 
						|
    vpdi %v19, %v1 ,%v6,0b0101
 | 
						|
    vst %v18,32(\CIJ_REG,\LC1)
 | 
						|
    vst %v19,48(\CIJ_REG,\LC1)
 | 
						|
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG,\LC2)
 | 
						|
    vl %v4 , 16(\CIJ_REG,\LC2)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vl %v7 , 32(\CIJ_REG,\LC2)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v6 ,  48 (\CIJ_REG,\LC2)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v28,%v29,%v30,%v31,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3 ,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG,\LC2)
 | 
						|
    vpdi %v18, %v1 ,%v6,0
 | 
						|
    vst %v17,16(\CIJ_REG,\LC2)
 | 
						|
    vpdi %v19, %v1 ,%v6,0b0101
 | 
						|
    vst %v18,32(\CIJ_REG,\LC2)
 | 
						|
    vst %v19,48(\CIJ_REG,\LC2)
 | 
						|
    la \CIJ_REG,64(\CIJ_REG)
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZSTORE_4x2  ALPHA_VECREG,ALPHA_VECI,CIJ_REG , LDC_BYTE_ORIGINAL
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG)
 | 
						|
    vl %v4 , 16(\CIJ_REG)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vl %v7 , 32(\CIJ_REG)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v6 ,  48 (\CIJ_REG)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v16,%v17,%v18,%v19,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG)
 | 
						|
    vpdi %v18, %v1 ,%v6,0
 | 
						|
    vst %v17,16(\CIJ_REG)
 | 
						|
    vpdi %v19, %v1 ,%v6,0b0101
 | 
						|
    vst %v18,32(\CIJ_REG)
 | 
						|
    vst %v19,48(\CIJ_REG)
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vl %v4 , 16(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vl %v7 , 32(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v6 ,  48 (\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v20,%v21,%v22,%v23,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v20, %v3 ,%v4,0
 | 
						|
    vpdi %v21, %v3 ,%v4,0b0101
 | 
						|
    vst %v20,0(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v22, %v1 ,%v6,0
 | 
						|
    vst %v21,16(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v23, %v1 ,%v6,0b0101
 | 
						|
    vst %v22,32(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vst %v23,48(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    la \CIJ_REG,64(\CIJ_REG)
 | 
						|
.endm
 | 
						|
.macro ZSTORE_4x1  ALPHA_VECREG,ALPHA_VECI,CIJ_REG , LDC_BYTE_ORIGINAL
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG)
 | 
						|
    vl %v4 , 16(\CIJ_REG)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vl %v7 , 32(\CIJ_REG)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v6 ,  48 (\CIJ_REG)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v16,%v17,%v18,%v19,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG)
 | 
						|
    vpdi %v18, %v1 ,%v6,0
 | 
						|
    vst %v17,16(\CIJ_REG)
 | 
						|
    vpdi %v19, %v1 ,%v6,0b0101
 | 
						|
    vst %v18,32(\CIJ_REG)
 | 
						|
    vst %v19,48(\CIJ_REG)
 | 
						|
    la \CIJ_REG,64(\CIJ_REG)
 | 
						|
.endm
 | 
						|
.macro ZSTORE_1x4  ALPHA_VECREG,ALPHA_VECI,CIJ_REG , LDC_BYTE_ORIGINAL,LC1,LC2
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG)
 | 
						|
    la \LC1,0(\LDC_BYTE_ORIGINAL, \LDC_BYTE_ORIGINAL)
 | 
						|
    vl %v4 , 0(\CIJ_REG,  \LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    la \LC2,0(\LC1,\LDC_BYTE_ORIGINAL )
 | 
						|
    vl %v7 , 0(\CIJ_REG,  \LC1)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v6 ,  0 (\CIJ_REG,\LC2)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
#else
 | 
						|
    la \LC1,0(\LDC_BYTE_ORIGINAL, \LDC_BYTE_ORIGINAL)
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v16,%v17,%v18,%v19,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
#if defined(TRMMKERNEL)
 | 
						|
    la \LC2,0(\LC1,\LDC_BYTE_ORIGINAL )
 | 
						|
#endif
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG)
 | 
						|
    vpdi %v18, %v1 ,%v6,0
 | 
						|
    vst %v17,0(\CIJ_REG,  \LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v19, %v1 ,%v6,0b0101
 | 
						|
    vst %v18,0(\CIJ_REG,   \LC1)
 | 
						|
    vst %v19,0(\CIJ_REG,\LC2)
 | 
						|
    la \CIJ_REG,16(\CIJ_REG)
 | 
						|
.endm
 | 
						|
.macro ZSTORE_2x4  ALPHA_VECREG,ALPHA_VECI,CIJ_REG , LDC_BYTE_ORIGINAL,LC1,LC2
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG)
 | 
						|
    vl %v26 , 16(\CIJ_REG)
 | 
						|
    la \LC1,0(\LDC_BYTE_ORIGINAL, \LDC_BYTE_ORIGINAL)
 | 
						|
    vl %v4 , 0(\CIJ_REG,  \LDC_BYTE_ORIGINAL)
 | 
						|
    vl %v25 , 16(\CIJ_REG,  \LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vpdi %v24,%v26,%v25,0
 | 
						|
    la \LC2,0(\LC1,\LDC_BYTE_ORIGINAL )
 | 
						|
    vl %v7 , 0(\CIJ_REG,  \LC1)
 | 
						|
    vl %v28 , 16(\CIJ_REG,  \LC1)
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vpdi %v25,%v26,%v25,0b101
 | 
						|
    vl %v6 ,  0 (\CIJ_REG,\LC2)
 | 
						|
    vl %v27 ,  16 (\CIJ_REG,\LC2)
 | 
						|
    vpdi %v1,%v7,%v6,0
 | 
						|
    vpdi %v6,%v7,%v6,0b101
 | 
						|
    vpdi %v26,%v28,%v27,0
 | 
						|
    vpdi %v27,%v28,%v27,0b101
 | 
						|
#else
 | 
						|
   la \LC1,0(\LDC_BYTE_ORIGINAL, \LDC_BYTE_ORIGINAL)
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_4x1 %v3,%v4,%v1,%v6,%v16,%v17,%v18,%v19,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    CalcMultAlpha_4x1 %v24,%v25,%v26,%v27,%v20,%v21,%v22,%v23,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
#if defined(TRMMKERNEL)
 | 
						|
    la \LC2,0(\LC1,\LDC_BYTE_ORIGINAL )
 | 
						|
#endif
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3,%v4,0b0101
 | 
						|
    vpdi %v20, %v24 ,%v25,0
 | 
						|
    vpdi %v21, %v24,%v25,0b0101
 | 
						|
    vpdi %v22, %v26 ,%v27,0
 | 
						|
    vpdi %v23, %v26 ,%v27,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG)
 | 
						|
    vst %v20,16(\CIJ_REG)
 | 
						|
    vpdi %v18, %v1 ,%v6,0
 | 
						|
    vst %v17,0(\CIJ_REG,  \LDC_BYTE_ORIGINAL)
 | 
						|
    vst %v21,16(\CIJ_REG,  \LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v19, %v1 ,%v6,0b0101
 | 
						|
    vst %v18,0(\CIJ_REG,   \LC1)
 | 
						|
    vst %v22,16(\CIJ_REG,   \LC1)
 | 
						|
    vst %v19,0(\CIJ_REG,\LC2)
 | 
						|
    vst %v23,16(\CIJ_REG,\LC2)
 | 
						|
    la \CIJ_REG,32(\CIJ_REG)
 | 
						|
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZSTORE_2x2  ALPHA_VECREG,ALPHA_VECI,CIJ_REG , LDC_BYTE_ORIGINAL
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG)
 | 
						|
    vl %v4 , 16(\CIJ_REG)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
    vl %v5 , 0(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vl %v7 , 16(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v6,%v5,%v7,0
 | 
						|
    vpdi %v7,%v5,%v7,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_2x1 %v3,%v4, %v16,%v17,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    CalcMultAlpha_2x1 %v6,%v7, %v20,%v21 ,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG)
 | 
						|
    vst %v17,16(\CIJ_REG)
 | 
						|
    vpdi %v20, %v6 ,%v7,0
 | 
						|
    vpdi %v21, %v6 ,%v7,0b0101
 | 
						|
    vst %v20,0(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vst %v21,16(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
 | 
						|
    la \CIJ_REG,32(\CIJ_REG)
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZSTORE_2x1  ALPHA_VECREG,ALPHA_VECI,CIJ_REG , LDC_BYTE_ORIGINAL
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG)
 | 
						|
    vl %v4 , 16(\CIJ_REG)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_2x1 %v3,%v4, %v16,%v17,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG)
 | 
						|
    vst %v17,16(\CIJ_REG)
 | 
						|
    la \CIJ_REG,32(\CIJ_REG)
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZSTORE_1x2  ALPHA_VECREG,ALPHA_VECI,CIJ_REG , LDC_BYTE_ORIGINAL
 | 
						|
  #if !defined(TRMMKERNEL)
 | 
						|
    vl %v1 , 0(\CIJ_REG)
 | 
						|
    vl %v4 , 0(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    vpdi %v3,%v1,%v4,0
 | 
						|
    vpdi %v4,%v1,%v4,0b101
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_2x1 %v3,%v4, %v16,%v17,\ALPHA_VECREG,\ALPHA_VECI
 | 
						|
    vpdi %v16, %v3 ,%v4,0
 | 
						|
    vpdi %v17, %v3,%v4,0b0101
 | 
						|
    vst %v16,0(\CIJ_REG)
 | 
						|
    vst %v17,0(\CIJ_REG,\LDC_BYTE_ORIGINAL)
 | 
						|
    la \CIJ_REG,16(\CIJ_REG)
 | 
						|
.endm
 | 
						|
 | 
						|
.macro ZSTORE_1x1  ALPHA_RR,ALPHA_RI ,CIJ_REG
 | 
						|
  #if defined (TRMMKERNEL)
 | 
						|
    lzdr %f1
 | 
						|
    lzdr %f4
 | 
						|
#else
 | 
						|
    ld %f1 , 0(\CIJ_REG)
 | 
						|
    ld %f4 , 8(\CIJ_REG )
 | 
						|
#endif
 | 
						|
    CalcMultAlpha_1x1 %f1,%f4, %f6,%f7,\ALPHA_RR,\ALPHA_RI
 | 
						|
    std %f1,0(\CIJ_REG)
 | 
						|
    std %f4,8(\CIJ_REG)
 | 
						|
    la \CIJ_REG,16(\CIJ_REG)
 | 
						|
.endm
 | 
						|
 | 
						|
/****************************TRMM POINTER REFRESH MACROSES*************************/
 | 
						|
 | 
						|
.macro RefreshPointers  PTR_A,PTR_B,OFF_VAL,B_VAL,C_A,C_B
 | 
						|
  #if (defined(LEFT) &&  defined(TRANSA)) ||  (!defined(LEFT) && !defined(TRANSA))
 | 
						|
    /*  ptrbb = bb;*/
 | 
						|
    lgr \PTR_B,\B_VAL    /*refresh BPOINT*/
 | 
						|
 | 
						|
  #else
 | 
						|
    /*  ptrba  =ptrba+ off*C_A;
 | 
						|
    ptrbb = bb + off*C_B;*/
 | 
						|
.if \C_B==4
 | 
						|
  .if \C_A==4
 | 
						|
    sllg \PTR_B, \OFF_VAL,6
 | 
						|
    agr \PTR_A,\PTR_B /*ptrba+off*4**/
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
  .elseif \C_A==2
 | 
						|
    sllg \PTR_B, \OFF_VAL,5
 | 
						|
    la \PTR_A,0(\PTR_A,\PTR_B) /*ptrba+off*2**/
 | 
						|
    agr \PTR_B, \PTR_B
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
 | 
						|
  .elseif \C_A==1
 | 
						|
    sllg \PTR_B, \OFF_VAL,4
 | 
						|
    agr \PTR_A,\PTR_B /*ptrba+off*4**/
 | 
						|
    sllg \PTR_B, \OFF_VAL,6
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
  .endif
 | 
						|
 | 
						|
.elseif \C_B==2
 | 
						|
  .if \C_A==4
 | 
						|
    sllg \PTR_B, \OFF_VAL,5
 | 
						|
    la \PTR_A,0(\PTR_A,\PTR_B) /*ptrba+off*2**/
 | 
						|
    agr \PTR_A,\PTR_B /*ptrba+off*2**/
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
  .elseif \C_A==2
 | 
						|
    sllg \PTR_B, \OFF_VAL,5
 | 
						|
    agr \PTR_A,\PTR_B /*ptrba+off*2**/
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
  .elseif \C_A==1
 | 
						|
    sllg \PTR_B, \OFF_VAL,4
 | 
						|
    la \PTR_A,0(\PTR_A,\PTR_B) /*ptrba+off*1**/
 | 
						|
    agr \PTR_B,\PTR_B /* off+off**/
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
  .endif
 | 
						|
 | 
						|
.elseif \C_B==1
 | 
						|
  .if \C_A==4
 | 
						|
    sllg \PTR_B, \OFF_VAL,6
 | 
						|
    agr \PTR_A,\PTR_B /*ptrba+off*4**/
 | 
						|
    sllg \PTR_B, \OFF_VAL,4
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
  .elseif \C_A==2
 | 
						|
    sllg \PTR_B, \OFF_VAL,4
 | 
						|
    la \PTR_A,0(\PTR_A,\PTR_B) /*ptrba+off*1**/
 | 
						|
    agr \PTR_A,\PTR_B /*ptrba+off*1**/
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
 | 
						|
  .elseif \C_A==1
 | 
						|
    sllg \PTR_B, \OFF_VAL,4
 | 
						|
    agr \PTR_A,\PTR_B /*ptrba+off*1**/
 | 
						|
    la \PTR_B,0(\B_VAL,\PTR_B)    /*refresh BPOINT*/
 | 
						|
  .endif
 | 
						|
.endif
 | 
						|
 | 
						|
  #endif
 | 
						|
.endm
 | 
						|
 | 
						|
/**/
 | 
						|
.macro RefreshTempBk TEMP_VAL,BK_VAL,OFF_VAL,INCR_A,INCR_B
 | 
						|
  #if (defined(LEFT) && !defined(TRANSA)) ||  (!defined(LEFT) && defined(TRANSA))
 | 
						|
        /* temp = bk-off;*/
 | 
						|
    sgrk \TEMP_VAL,\BK_VAL,\OFF_VAL
 | 
						|
 | 
						|
  #elif defined(LEFT)
 | 
						|
        /* temp = off+INCR_A; // number of values in A */
 | 
						|
    la \TEMP_VAL,\INCR_A(\OFF_VAL)
 | 
						|
  #else
 | 
						|
        /* temp = off+INCR_B  // number of values in B*/
 | 
						|
    la \TEMP_VAL,\INCR_B(\OFF_VAL)
 | 
						|
  #endif
 | 
						|
 | 
						|
.endm
 | 
						|
 | 
						|
.macro RefreshPointersAndOFF TEMP_VAL,BK_VAL,OFF_VAL,PTR_A,C_A,C_B
 | 
						|
 | 
						|
  #if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
 | 
						|
    /*temp = bk - off;*/
 | 
						|
    sgrk \TEMP_VAL,\BK_VAL,\OFF_VAL
 | 
						|
  #ifdef LEFT
 | 
						|
    /*temp -= 8; // number of values in A*/
 | 
						|
    lay \TEMP_VAL,-\C_A(\TEMP_VAL)
 | 
						|
  #else
 | 
						|
    /*temp -= 4; // number of values in B*/
 | 
						|
    lay \TEMP_VAL,-\C_B(\TEMP_VAL)
 | 
						|
  #endif
 | 
						|
    /*ptrba += temp*C_A;
 | 
						|
    ptrbb += temp*C_B;*/
 | 
						|
 | 
						|
  .if \C_A==4
 | 
						|
    sllg \TEMP_VAL, \TEMP_VAL,6 /*temp*4*/
 | 
						|
  .elseif \C_A==2
 | 
						|
    sllg \TEMP_VAL, \TEMP_VAL,5 /*temp*2*/
 | 
						|
  .elseif \C_A==1
 | 
						|
    sllg \TEMP_VAL, \TEMP_VAL,4 /*temp*1*/
 | 
						|
  .endif
 | 
						|
    la \PTR_A,0(\PTR_A,\TEMP_VAL) /*ptrba+temp*C_A*/
 | 
						|
  #endif
 | 
						|
 | 
						|
  #ifdef LEFT
 | 
						|
    /*off += \c_A; // number of values in A*/
 | 
						|
    aghi \OFF_VAL,\C_A
 | 
						|
  #endif
 | 
						|
.endm
 | 
						|
 |