AMD 250 manual 213

; Finish all memory writes.

25112 Rev. 3.06 September 2005 Software Optimization Guide for AMD64 Processors

movntps	[edi+4*ecx+16], xmm1 ; Store reciprocal square			root to rcp_sqrt_r.
movntps	[edi+4*ecx+32], xmm2 ; Store reciprocal square			root to rcp_sqrt_r.
movntps	[edi+4*ecx+48], xmm3 ; Store reciprocal square			root to rcp_sqrt_r.
add	ecx, 16	; Decrement	the # of reciprocal square
		; roots to	calculate by 16.
dec	eax	; Decrement	# of 16 float	reciprocal square

;root loops to perform by 1.

jnz reciprocal_sqrt_4xloop

jmp	skip_recprcl_sqrt_4xloop	; Jump into			loop	to calculate	reciprocal
		;	square	root of floats that			don't
		;	occupy	a	full	cache line.

;==============================================================================

;THIS LOOP RECIPROCATES AND SQUARE ROOTS 1 FLOATING POINT NUMBER EACH

;LOOP ITERATION ;==============================================================================

ALIGN 16		; Align address of loop to a 16-byte boundary.
reciprocal_sqrt_1xloop:
movss	xmm0, [esi+4*ecx]	; XMM0=[,,,r0]
sqrtss	xmm0, xmm0	; XMM0=[,,,sqrt(r0)]
rcpss	xmm0, xmm0	; XMM0=[,,,1/sqrt(r0)]
movss	[edi+4*ecx], xmm0	; Store reciprocal square root to rcp_sqrt_r.
inc	ecx	; Decrement the # of reciprocal square roots
		;	to calculate.
skip_recprcl_sqrt_4xloop:
or	ecx, ecx	; If ECX != 0, then calculate the reciprocal
		;	square root of another float.

jnz reciprocal_sqrt_1xloop

sfence

;==============================================================================

;INSTRUCTIONS BELOW RESTORE THE REGISTER STATE WITH WHICH THIS ROUTINE

;WAS ENTERED.

;REGISTERS EAX, ECX, AND EDX ARE CONSIDERED VOLATILE AND ASSUMED TO BE CHANGED,

;WHILE THE REGISTERS BELOW MUST BE PRESERVED IF THE USER IS CHANGING THEM pop edi

pop esi pop ebx mov esp,ebp pop ebp

;=============================================================================== ret

_reciprocal_sqrt_sse ENDP _TEXT ENDS

END

The preceding code illustrates the use of separate loops for optimal performance. The loop titled reciprocal_sqrt_4xloop works with 16 floating-point numbers in each iteration and is unrolled to keep the processor busy by masking the latencies of the reciprocal and square-root instructions. In

Chapter 9 Optimizing with SIMD Instructions 213