[Math-atlas-commits] CVS: AtlasBase/kernel/ClintWhaley ATL_smm14x1x84_sse.c,NONE,1.1 ATL_smm14x1x84_sseCU.c,NONE,1.1

[R. C. W. <rw...@us...>]

Update of /cvsroot/math-atlas/AtlasBase/kernel/ClintWhaley
In directory sc8-pr-cvs1:/tmp/cvs-serv24925

Added Files:
	ATL_smm14x1x84_sse.c ATL_smm14x1x84_sseCU.c 
Log Message:
hammer sgemm kernels

--- NEW FILE: ATL_smm14x1x84_sse.c ---
#if !defined(NB) || (NB == 0)
   #error "NB must be a compile-time constant!"
#endif

#if (NB/14)*14 != NB
   #error "NB must be multiple of 14!"
#endif

#if !defined(ATL_GAS_x8664)
   #error "This kernel requires a gas x86-64 assembler!"
#endif

#
#  Prefetch defines
#
#define pref2(mem) prefetcht1   mem
#define prefB(mem) prefetcht0   mem
#define prefC(mem) prefetchw    mem

[...2320 lines suppressed...]
#       End unrolled final M-loop iteration
#       stM
#
#
#	pC += incCn;  pA -= NBNB;  pB += NB;
#
	addq	incCn, pC
	subq	$NBNBso-NB14so, pA
	addq	$NBso, pB
#
#	while (pB != stN);
#
	cmp	pB, stN
	jne	UNLOOP

#
#	Restore callee-saved iregs
#
	movq	%r11, %rbp
	ret

--- NEW FILE: ATL_smm14x1x84_sseCU.c ---
#if !defined(KB) || (KB == 0)
   #error "KB must be a compile-time constant!"
#endif

#ifndef MB
   #define MB 0
#endif
#if (MB/14)*14 != MB
   #error "MB must be multiple of 14!"
#endif

#if !defined(ATL_GAS_x8664)
   #error "This kernel requires a gas x86-64 assembler!"
#endif

#
#  Prefetch defines
#
#define pref2(mem) prefetcht1   mem
[...2414 lines suppressed...]
#       End unrolled final M-loop iteration
#       stM
#
#
#	pC += incCn;  pA -= NBNB;  pB += NB;
#
	addq	incCn, pC
	movq	pA0, pA
	addq	$NBso, pB
#
#	while (pB != stN);
#
	cmp	pB, stN
	jne	UNLOOP

#
#	Restore callee-saved iregs
#
	movq	%r11, %rbp
	ret