atl_sjik48x48x48tn48x48x0_a1_b0.asm

来自「基于Blas CLapck的.用过的人知道是干啥的」· 汇编 代码 · 共 1,351 行 · 第 1/2 页

	%ifdef PREC_DST4
		mov [esp+20],ecx
		add ecx,[esp+12]

		prefetchw [ecx-2*64]
		prefetchw [ecx-1*64]

		prefetchw [ecx+0*64]
		prefetchw [ecx+1*64]
		nop

		prefetchw [ecx+2*64-1]
		mov ecx,[esp+20]
	%endif



	%ifdef PREB_DST4
		prefetch [ebx+48*4-2*64]
		fnop
		mov edx,edx

		prefetch [ebx+48*4-1*64]
		nop

		prefetch [ebx+48*4+0*64]
		nop

		prefetch [ebx+48*4+1*64]
		nop

		prefetch [ebx+48*4+2*64]
		nop
	%endif



	fstp dword [ecx+ELM19]
	fxch st3
	fstp dword [ecx+ELM20]

	fxch st1
	fstp dword [ecx+ELM21]
	fstp dword [ecx+ELM22]

	fstp dword [ecx+ELM23]
	fstp dword [ecx+ELM24]
	add eax,edx



	fld dword [ebx+ELM1]			;01+5
	fld dword [eax+DOTP2]
	fmul st0,st1

	fld dword [eax+DOTP3]
	fmul st0,st2
	fld dword [eax+DOTP1]

	rep
	fmul st0,st3
	fxch st0,st3
	fld dword [eax+DOTP5]

	rep
	fmul st0,st1
	rep
	fld dword [eax+DOTP6]
	mov edx,edx

	fmul st0,st2
	fld dword [ebx+ELM2]
	fld dword [eax+DOTP4]

	rep
	fmulp st4,st0
	add eax,byte 30*4
	mov edx,edx



	OPERATION 2,3				;02+5
	OPERATION 3,4				;03+5
	OPERATION 4,5				;04+5
	OPERATION 5,6				;05+5
	OPERATION 6,7				;06+5
	OPERATION 7,8				;07+5
	OPERATION 8,9				;08+5
	OPERATION 9,10				;09+5
	OPERATION 10,11				;10+5
	OPERATION 11,12				;11+5
	OPERATION 12,13				;12+5
	OPERATION 13,14				;13+5
	OPERATION 14,15				;14+5
	OPERATION 15,16				;15+5
	OPERATION 16,17				;16+5
	OPERATION 17,18				;17+5
	OPERATION 18,19				;18+5
	OPERATION 19,20				;19+5
	OPERATION 20,21				;20+5
	OPERATION 21,22				;21+5
	OPERATION 22,23				;22+5
	OPERATION 23,24				;23+5
	OPERATION 24,25				;24+5
	OPERATION 25,26				;25+5
	OPERATION 26,27				;26+5
	OPERATION 27,28				;27+5
	OPERATION 28,29				;28+5
	OPERATION 29,30				;29+5
	OPERATION 30,31				;30+5
	OPERATION 31,32				;31+5
	OPERATION 32,33				;32+5
	OPERATION 33,34				;33+5
	OPERATION 34,35				;34+5
	OPERATION 35,36				;35+5
	OPERATION 36,37				;36+5
	OPERATION 37,38				;37+5
	OPERATION 38,39				;38+5
	OPERATION 39,40				;39+5
	OPERATION 40,41				;40+5
	OPERATION 41,42				;41+5
	OPERATION 42,43				;42+5
	OPERATION 43,44				;43+5
	OPERATION 44,45				;44+5
	OPERATION 45,46				;45+5
	OPERATION 46,47				;45+5
	OPERATION 47,48				;45+5
	fld dword [eax+DOTP1+ELM48]		;48+5
	fmul st0,st1
	faddp st7

	fld dword [eax+DOTP2+ELM48]
	fmul st0,st1
	faddp st6

	fld dword [eax+DOTP3+ELM48]
	fmul st0,st1
	faddp st5

	fld dword [eax+DOTP4+ELM48]
	fmul st0,st1
	faddp st4

	fld dword [eax+DOTP5+ELM48]
	fmul st0,st1
	faddp st3

	rep
	fmul dword [eax+DOTP6+ELM48]
	faddp st1
	fxch st5



	%ifdef PREC_DST3
		mov [esp+20],ecx
		add ecx,[esp+12]

		prefetchw [ecx-2*64]
		prefetchw [ecx-1*64]

		prefetchw [ecx+0*64]
		prefetchw [ecx+1*64]
		nop

		prefetchw [ecx+2*64-1]
		mov ecx,[esp+20]
	%endif



	%ifdef PREB_DST3
		prefetch [ebx+48*4-2*64]
		fnop
		mov edx,edx

		prefetch [ebx+48*4-1*64]
		nop

		prefetch [ebx+48*4+0*64]
		nop

		prefetch [ebx+48*4+1*64]
		nop

		prefetch [ebx+48*4+2*64]
		nop
	%endif



	fstp dword [ecx+ELM25]
	fxch st3
	fstp dword [ecx+ELM26]

	fxch st1
	fstp dword [ecx+ELM27]
	fstp dword [ecx+ELM28]

	fstp dword [ecx+ELM29]
	fstp dword [ecx+ELM30]
	add eax,edx



	fld dword [ebx+ELM1]			;01+6
	fld dword [eax+DOTP2]
	fmul st0,st1

	fld dword [eax+DOTP3]
	fmul st0,st2
	fld dword [eax+DOTP1]

	rep
	fmul st0,st3
	fxch st0,st3
	fld dword [eax+DOTP5]

	rep
	fmul st0,st1
	rep
	fld dword [eax+DOTP6]
	mov edx,edx

	fmul st0,st2
	fld dword [ebx+ELM2]
	fld dword [eax+DOTP4]

	rep
	fmulp st4,st0
	add eax,byte 30*4
	mov edx,edx



	OPERATION 2,3				;02+6
	OPERATION 3,4				;03+6
	OPERATION 4,5				;04+6
	OPERATION 5,6				;05+6
	OPERATION 6,7				;06+6
	OPERATION 7,8				;07+6
	OPERATION 8,9				;08+6
	OPERATION 9,10				;09+6
	OPERATION 10,11				;10+6
	OPERATION 11,12				;11+6
	OPERATION 12,13				;12+6
	OPERATION 13,14				;13+6
	OPERATION 14,15				;14+6
	OPERATION 15,16				;15+6
	OPERATION 16,17				;16+6
	OPERATION 17,18				;17+6
	OPERATION 18,19				;18+6
	OPERATION 19,20				;19+6
	OPERATION 20,21				;20+6
	OPERATION 21,22				;21+6
	OPERATION 22,23				;22+6
	OPERATION 23,24				;23+6
	OPERATION 24,25				;24+6
	OPERATION 25,26				;25+6
	OPERATION 26,27				;26+6
	OPERATION 27,28				;27+6
	OPERATION 28,29				;28+6
	OPERATION 29,30				;29+6
	OPERATION 30,31				;30+6
	OPERATION 31,32				;31+6
	OPERATION 32,33				;32+6
	OPERATION 33,34				;33+6
	OPERATION 34,35				;34+6
	OPERATION 35,36				;35+6
	OPERATION 36,37				;36+6
	OPERATION 37,38				;37+6
	OPERATION 38,39				;38+6
	OPERATION 39,40				;39+6
	OPERATION 40,41				;40+6
	OPERATION 41,42				;41+6
	OPERATION 42,43				;42+6
	OPERATION 43,44				;43+6
	OPERATION 44,45				;44+6
	OPERATION 45,46				;45+6
	OPERATION 46,47				;45+6
	OPERATION 47,48				;45+6
	fld dword [eax+DOTP1+ELM48]		;48+6
	fmul st0,st1
	faddp st7

	fld dword [eax+DOTP2+ELM48]
	fmul st0,st1
	faddp st6

	fld dword [eax+DOTP3+ELM48]
	fmul st0,st1
	faddp st5

	fld dword [eax+DOTP4+ELM48]
	fmul st0,st1
	faddp st4

	fld dword [eax+DOTP5+ELM48]
	fmul st0,st1
	faddp st3

	rep
	fmul dword [eax+DOTP6+ELM48]
	faddp st1
	fxch st5



	%ifdef PREC_DST2
		mov [esp+20],ecx
		add ecx,[esp+12]

		prefetchw [ecx-2*64]
		prefetchw [ecx-1*64]

		prefetchw [ecx+0*64]
		prefetchw [ecx+1*64]
		nop

		prefetchw [ecx+2*64-1]
		mov ecx,[esp+20]
	%endif



	%ifdef PREB_DST2
		prefetch [ebx+48*4-2*64]
		fnop
		mov edx,edx

		prefetch [ebx+48*4-1*64]
		nop

		prefetch [ebx+48*4+0*64]
		nop

		prefetch [ebx+48*4+1*64]
		nop

		prefetch [ebx+48*4+2*64]
		nop
	%endif



	rep
	fstp dword [ecx+ELM31]
	fxch st3
	fstp dword [ecx+ELM32]

	fxch st1
	fstp dword [ecx+ELM33]
	fstp dword [ecx+ELM34]

	fstp dword [ecx+ELM35]
	fstp dword [ecx+ELM36]
	add eax,edx



	fld dword [ebx+ELM1]			;01+7
	fld dword [eax+DOTP2]
	fmul st0,st1

	fld dword [eax+DOTP3]
	fmul st0,st2
	fld dword [eax+DOTP1]

	rep
	fmul st0,st3
	fxch st0,st3
	fld dword [eax+DOTP5]

	rep
	fmul st0,st1
	rep
	fld dword [eax+DOTP6]
	mov edx,edx

	fmul st0,st2
	fld dword [ebx+ELM2]
	fld dword [eax+DOTP4]

	rep
	fmulp st4,st0
	add eax,byte 30*4
	mov edx,edx



	OPERATION 2,3				;02+7
	OPERATION 3,4				;03+7
	OPERATION 4,5				;04+7
	OPERATION 5,6				;05+7
	OPERATION 6,7				;06+7
	OPERATION 7,8				;07+7
	OPERATION 8,9				;08+7
	OPERATION 9,10				;09+7
	OPERATION 10,11				;10+7
	OPERATION 11,12				;11+7
	OPERATION 12,13				;12+7
	OPERATION 13,14				;13+7
	OPERATION 14,15				;14+7
	OPERATION 15,16				;15+7
	OPERATION 16,17				;16+7
	OPERATION 17,18				;17+7
	OPERATION 18,19				;18+7
	OPERATION 19,20				;19+7
	OPERATION 20,21				;20+7
	OPERATION 21,22				;21+7
	OPERATION 22,23				;22+7
	OPERATION 23,24				;23+7
	OPERATION 24,25				;24+7
	OPERATION 25,26				;25+7
	OPERATION 26,27				;26+7
	OPERATION 27,28				;27+7
	OPERATION 28,29				;28+7
	OPERATION 29,30				;29+7
	OPERATION 30,31				;30+7
	OPERATION 31,32				;31+7
	OPERATION 32,33				;32+7
	OPERATION 33,34				;33+7
	OPERATION 34,35				;34+7
	OPERATION 35,36				;35+7
	OPERATION 36,37				;36+7
	OPERATION 37,38				;37+7
	OPERATION 38,39				;38+7
	OPERATION 39,40				;39+7
	OPERATION 40,41				;40+7
	OPERATION 41,42				;41+7
	OPERATION 42,43				;42+7
	OPERATION 43,44				;43+7
	OPERATION 44,45				;44+7
	OPERATION 45,46				;45+7
	OPERATION 46,47				;45+7
	OPERATION 47,48				;45+7
	fld dword [eax+DOTP1+ELM48]		;48+7
	fmul st0,st1
	faddp st7

	fld dword [eax+DOTP2+ELM48]
	fmul st0,st1
	faddp st6

	fld dword [eax+DOTP3+ELM48]
	fmul st0,st1
	faddp st5

	fld dword [eax+DOTP4+ELM48]
	fmul st0,st1
	faddp st4

	fld dword [eax+DOTP5+ELM48]
	fmul st0,st1
	faddp st3

	rep
	fmul dword [eax+DOTP6+ELM48]
	faddp st1
	fxch st5



	%ifdef PREC_DST1
		mov [esp+20],ecx
		add ecx,[esp+12]

		prefetchw [ecx-2*64]
		prefetchw [ecx-1*64]

		prefetchw [ecx+0*64]
		prefetchw [ecx+1*64]
		nop

		prefetchw [ecx+2*64-1]
		mov ecx,[esp+20]
	%endif



	%ifdef PREB_DST1
		prefetch [ebx+48*4-2*64]
		fnop
		mov edx,edx

		prefetch [ebx+48*4-1*64]
		nop

		prefetch [ebx+48*4+0*64]
		nop

		prefetch [ebx+48*4+1*64]
		nop

		prefetch [ebx+48*4+2*64]
		nop
	%endif



	fstp dword [ecx+ELM37]
	fxch st3
	fstp dword [ecx+ELM38]

	fxch st1
	fstp dword [ecx+ELM39]
	fstp dword [ecx+ELM40]

	fstp dword [ecx+ELM41]
	fstp dword [ecx+ELM42]
	add eax,edx



	fld dword [ebx+ELM1]			;01+8
	fld dword [eax+DOTP2]
	fmul st0,st1

	fld dword [eax+DOTP3]
	fmul st0,st2
	fld dword [eax+DOTP1]

	rep
	fmul st0,st3
	fxch st0,st3
	fld dword [eax+DOTP5]

	rep
	fmul st0,st1
	rep
	fld dword [eax+DOTP6]
	mov edx,edx

	fmul st0,st2
	fld dword [ebx+ELM2]
	fld dword [eax+DOTP4]

	rep
	fmulp st4,st0
	add eax,byte 30*4
	mov edx,edx



	OPERATION 2,3				;02+8
	OPERATION 3,4				;03+8
	OPERATION 4,5				;04+8
	OPERATION 5,6				;05+8
	OPERATION 6,7				;06+8
	OPERATION 7,8				;07+8
	OPERATION 8,9				;08+8
	OPERATION 9,10				;09+8
	OPERATION 10,11				;10+8
	OPERATION 11,12				;11+8
	OPERATION 12,13				;12+8
	OPERATION 13,14				;13+8
	OPERATION 14,15				;14+8
	OPERATION 15,16				;15+8
	OPERATION 16,17				;16+8
	OPERATION 17,18				;17+8
	OPERATION 18,19				;18+8
	OPERATION 19,20				;19+8
	OPERATION 20,21				;20+8
	OPERATION 21,22				;21+8
	OPERATION 22,23				;22+8
	OPERATION 23,24				;23+8
	OPERATION 24,25				;24+8
	OPERATION 25,26				;25+8
	OPERATION 26,27				;26+8
	OPERATION 27,28				;27+8
	OPERATION 28,29				;28+8
	OPERATION 29,30				;29+8
	OPERATION 30,31				;30+8
	OPERATION 31,32				;31+8
	OPERATION 32,33				;32+8
	OPERATION 33,34				;33+8
	OPERATION 34,35				;34+8
	OPERATION 35,36				;35+8
	OPERATION 36,37				;36+8
	OPERATION 37,38				;37+8
	OPERATION 38,39				;38+8
	OPERATION 39,40				;39+8
	OPERATION 40,41				;40+8
	OPERATION 41,42				;41+8
	OPERATION 42,43				;42+8
	OPERATION 43,44				;43+8
	OPERATION 44,45				;44+8
	OPERATION 45,46				;45+8
	OPERATION 46,47				;45+8
	OPERATION 47,48				;45+8
	fld dword [eax+DOTP1+ELM48]		;48+8
	fmul st0,st1
	faddp st7

	fld dword [eax+DOTP2+ELM48]
	fmul st0,st1
	faddp st6

	fld dword [eax+DOTP3+ELM48]
	fmul st0,st1
	faddp st5

	fld dword [eax+DOTP4+ELM48]
	fmul st0,st1
	faddp st4

	fld dword [eax+DOTP5+ELM48]
	fmul st0,st1
	faddp st3

	rep
	fmul dword [eax+DOTP6+ELM48]
	faddp st1
	fxch st5



	%ifdef PREA_EN
		mov [esp+20],edx		;save edx in t1
		mov edx,[esp+16]		;&A+1->edx

		lea edx,[edx+ebx]
		prefetch [edx-2*64]
		nop

		prefetch [edx-1*64]
		prefetch [edx+0*64]
		nop

		prefetch [edx+1*64]
		prefetch [edx+2*64-8]

		mov edx,[esp+20]		;restore edx
		mov eax,eax
		fnop
	%endif



	fstp dword [ecx+ELM43]
	fxch st3
	fstp dword [ecx+ELM44]

	fxch st1
	fstp dword [ecx+ELM45]
	fstp dword [ecx+ELM46]

	fstp dword [ecx+ELM47]
	fstp dword [ecx+ELM48]
	sub ebx,edi				;next column of B


	mov eax,[esp+4]				;reset eax
	add ecx,[esp+12]			;next column of C (+ldc*4)

	dec dword [esp+8]			;dec counter
	jnz near loopj_


end_
	femms

	pop ebp
	add esp,byte 5*4			;remove local variables

	pop edi					;restore registers
	pop esi
	pop ebx

	leave                			;mov esp,ebp / pop ebp
	ret