atl_smm4x4x16_us.c

来自「基于Blas CLapck的.用过的人知道是干啥的」· C语言 代码 · 共 1,551 行 · 第 1/3 页

C
1,551
字号
							ldd	[pA2+16], rA2	fadds	rC32, m3, rC32				fmuls	ra3, rb1, m3							ldd	[pA3+16], rA3	fadds	rC03, m0, rC03				fmuls	rA0, rB0, m0							ldd	[pB1+16], rB1	fadds	rC13, m1, rC13				fmuls	rA1, rB0, m1	fadds	rC23, m2, rC23				fmuls	rA2, rB0, m2	fadds	rC33, m3, rC33				fmuls	rA3, rB0, m3!!	K=4 iteration!	fadds	rC00, m0, rC00				fmuls	rA0, rB1, m0	fadds	rC10, m1, rC10				fmuls	rA1, rB1, m1	fadds	rC20, m2, rC20				fmuls	rA2, rB1, m2	fadds	rC30, m3, rC30				fmuls	rA3, rB1, m3	fadds	rC01, m0, rC01				fmuls	ra0, rb0, m0	fadds	rC11, m1, rC11				fmuls	ra1, rb0, m1	fadds	rC21, m2, rC21				fmuls	ra2, rb0, m2	fadds	rC31, m3, rC31				fmuls	ra3, rb0, m3							ldd	[pB2+16], rB0	fadds	rC00, m0, rC00				fmuls	ra0, rb1, m0	fadds	rC10, m1, rC10				fmuls	ra1, rb1, m1	fadds	rC20, m2, rC20				fmuls	ra2, rb1, m2	fadds	rC30, m3, rC30				fmuls	ra3, rb1, m3							ldd	[pB3+16], rB1	fadds	rC01, m0, rC01				fmuls	rA0, rB0, m0	fadds	rC11, m1, rC11				fmuls	rA1, rB0, m1	fadds	rC21, m2, rC21				fmuls	rA2, rB0, m2	fadds	rC31, m3, rC31				fmuls	rA3, rB0, m3	fadds	rC02, m0, rC02				fmuls	rA0, rB1, m0	fadds	rC12, m1, rC12				fmuls	rA1, rB1, m1	fadds	rC22, m2, rC22				fmuls	rA2, rB1, m2	fadds	rC32, m3, rC32				fmuls	rA3, rB1, m3	fadds	rC03, m0, rC03				fmuls	ra0, rb0, m0	fadds	rC13, m1, rC13				fmuls	ra1, rb0, m1	fadds	rC23, m2, rC23				fmuls	ra2, rb0, m2	fadds	rC33, m3, rC33				fmuls	ra3, rb0, m3							ldd	[pB0+24], rB0	fadds	rC02, m0, rC02				fmuls	ra0, rb1, m0							ldd	[pA0+24], rA0	fadds	rC12, m1, rC12				fmuls	ra1, rb1, m1							ldd	[pA1+24], rA1	fadds	rC22, m2, rC22				fmuls	ra2, rb1, m2							ldd	[pA2+24], rA2	fadds	rC32, m3, rC32				fmuls	ra3, rb1, m3							ldd	[pA3+24], rA3	fadds	rC03, m0, rC03				fmuls	rA0, rB0, m0							ldd	[pB1+24], rB1	fadds	rC13, m1, rC13				fmuls	rA1, rB0, m1	fadds	rC23, m2, rC23				fmuls	rA2, rB0, m2	fadds	rC33, m3, rC33				fmuls	rA3, rB0, m3!!	K=6 iteration!	fadds	rC00, m0, rC00				fmuls	rA0, rB1, m0	fadds	rC10, m1, rC10				fmuls	rA1, rB1, m1	fadds	rC20, m2, rC20				fmuls	rA2, rB1, m2	fadds	rC30, m3, rC30				fmuls	rA3, rB1, m3	fadds	rC01, m0, rC01				fmuls	ra0, rb0, m0	fadds	rC11, m1, rC11				fmuls	ra1, rb0, m1	fadds	rC21, m2, rC21				fmuls	ra2, rb0, m2	fadds	rC31, m3, rC31				fmuls	ra3, rb0, m3							ldd	[pB2+24], rB0	fadds	rC00, m0, rC00				fmuls	ra0, rb1, m0	fadds	rC10, m1, rC10				fmuls	ra1, rb1, m1	fadds	rC20, m2, rC20				fmuls	ra2, rb1, m2	fadds	rC30, m3, rC30				fmuls	ra3, rb1, m3							ldd	[pB3+24], rB1	fadds	rC01, m0, rC01				fmuls	rA0, rB0, m0	fadds	rC11, m1, rC11				fmuls	rA1, rB0, m1	fadds	rC21, m2, rC21				fmuls	rA2, rB0, m2	fadds	rC31, m3, rC31				fmuls	rA3, rB0, m3	fadds	rC02, m0, rC02				fmuls	rA0, rB1, m0	fadds	rC12, m1, rC12				fmuls	rA1, rB1, m1	fadds	rC22, m2, rC22				fmuls	rA2, rB1, m2	fadds	rC32, m3, rC32				fmuls	rA3, rB1, m3	fadds	rC03, m0, rC03				fmuls	ra0, rb0, m0	fadds	rC13, m1, rC13				fmuls	ra1, rb0, m1	fadds	rC23, m2, rC23				fmuls	ra2, rb0, m2	fadds	rC33, m3, rC33				fmuls	ra3, rb0, m3							ldd	[pB0+32], rB0	fadds	rC02, m0, rC02				fmuls	ra0, rb1, m0							ldd	[pA0+32], rA0	fadds	rC12, m1, rC12				fmuls	ra1, rb1, m1							ldd	[pA1+32], rA1	fadds	rC22, m2, rC22				fmuls	ra2, rb1, m2							ldd	[pA2+32], rA2	fadds	rC32, m3, rC32				fmuls	ra3, rb1, m3							ldd	[pA3+32], rA3	fadds	rC03, m0, rC03				fmuls	rA0, rB0, m0							ldd	[pB1+32], rB1	fadds	rC13, m1, rC13				fmuls	rA1, rB0, m1	fadds	rC23, m2, rC23				fmuls	rA2, rB0, m2	fadds	rC33, m3, rC33				fmuls	rA3, rB0, m3!!	K=8 iteration!	fadds	rC00, m0, rC00				fmuls	rA0, rB1, m0	fadds	rC10, m1, rC10				fmuls	rA1, rB1, m1	fadds	rC20, m2, rC20				fmuls	rA2, rB1, m2	fadds	rC30, m3, rC30				fmuls	rA3, rB1, m3	fadds	rC01, m0, rC01				fmuls	ra0, rb0, m0	fadds	rC11, m1, rC11				fmuls	ra1, rb0, m1	fadds	rC21, m2, rC21				fmuls	ra2, rb0, m2	fadds	rC31, m3, rC31				fmuls	ra3, rb0, m3							ldd	[pB2+32], rB0	fadds	rC00, m0, rC00				fmuls	ra0, rb1, m0	fadds	rC10, m1, rC10				fmuls	ra1, rb1, m1	fadds	rC20, m2, rC20				fmuls	ra2, rb1, m2	fadds	rC30, m3, rC30				fmuls	ra3, rb1, m3							ldd	[pB3+32], rB1	fadds	rC01, m0, rC01				fmuls	rA0, rB0, m0	fadds	rC11, m1, rC11				fmuls	rA1, rB0, m1	fadds	rC21, m2, rC21				fmuls	rA2, rB0, m2	fadds	rC31, m3, rC31				fmuls	rA3, rB0, m3							prefW2([pC0+incCn])							prefW2([pC1+incCn])							prefW2([pC2+incCn])							prefW2([pC3+incCn])	fadds	rC02, m0, rC02				fmuls	rA0, rB1, m0	fadds	rC12, m1, rC12				fmuls	rA1, rB1, m1	fadds	rC22, m2, rC22				fmuls	rA2, rB1, m2	fadds	rC32, m3, rC32				fmuls	rA3, rB1, m3	fadds	rC03, m0, rC03				fmuls	ra0, rb0, m0	fadds	rC13, m1, rC13				fmuls	ra1, rb0, m1	fadds	rC23, m2, rC23				fmuls	ra2, rb0, m2	fadds	rC33, m3, rC33				fmuls	ra3, rb0, m3							ldd	[pB0+40], rB0	fadds	rC02, m0, rC02				fmuls	ra0, rb1, m0							ldd	[pA0+40], rA0	fadds	rC12, m1, rC12				fmuls	ra1, rb1, m1							ldd	[pA1+40], rA1	fadds	rC22, m2, rC22				fmuls	ra2, rb1, m2							ldd	[pA2+40], rA2	fadds	rC32, m3, rC32				fmuls	ra3, rb1, m3							ldd	[pA3+40], rA3	fadds	rC03, m0, rC03				fmuls	rA0, rB0, m0							ldd	[pB1+40], rB1	fadds	rC13, m1, rC13				fmuls	rA1, rB0, m1	fadds	rC23, m2, rC23				fmuls	rA2, rB0, m2	fadds	rC33, m3, rC33				fmuls	rA3, rB0, m3!!	K=10 iteration!	fadds	rC00, m0, rC00				fmuls	rA0, rB1, m0	fadds	rC10, m1, rC10				fmuls	rA1, rB1, m1	fadds	rC20, m2, rC20				fmuls	rA2, rB1, m2	fadds	rC30, m3, rC30				fmuls	rA3, rB1, m3	fadds	rC01, m0, rC01				fmuls	ra0, rb0, m0	fadds	rC11, m1, rC11				fmuls	ra1, rb0, m1	fadds	rC21, m2, rC21				fmuls	ra2, rb0, m2	fadds	rC31, m3, rC31				fmuls	ra3, rb0, m3							ldd	[pB2+40], rB0	fadds	rC00, m0, rC00				fmuls	ra0, rb1, m0	fadds	rC10, m1, rC10				fmuls	ra1, rb1, m1	fadds	rC20, m2, rC20				fmuls	ra2, rb1, m2	fadds	rC30, m3, rC30				fmuls	ra3, rb1, m3							ldd	[pB3+40], rB1	fadds	rC01, m0, rC01				fmuls	rA0, rB0, m0	fadds	rC11, m1, rC11				fmuls	rA1, rB0, m1	fadds	rC21, m2, rC21				fmuls	rA2, rB0, m2	fadds	rC31, m3, rC31				fmuls	rA3, rB0, m3	fadds	rC02, m0, rC02				fmuls	rA0, rB1, m0	fadds	rC12, m1, rC12				fmuls	rA1, rB1, m1	fadds	rC22, m2, rC22				fmuls	rA2, rB1, m2	fadds	rC32, m3, rC32				fmuls	rA3, rB1, m3	fadds	rC03, m0, rC03				fmuls	ra0, rb0, m0	fadds	rC13, m1, rC13				fmuls	ra1, rb0, m1	fadds	rC23, m2, rC23				fmuls	ra2, rb0, m2	fadds	rC33, m3, rC33				fmuls	ra3, rb0, m3							ldd	[pB0+48], rB0	fadds	rC02, m0, rC02				fmuls	ra0, rb1, m0							ldd	[pA0+48], rA0	fadds	rC12, m1, rC12				fmuls	ra1, rb1, m1							ldd	[pA1+48], rA1	fadds	rC22, m2, rC22				fmuls	ra2, rb1, m2							ldd	[pA2+48], rA2	fadds	rC32, m3, rC32				fmuls	ra3, rb1, m3							ldd	[pA3+48], rA3	fadds	rC03, m0, rC03				fmuls	rA0, rB0, m0							ldd	[pB1+48], rB1	fadds	rC13, m1, rC13				fmuls	rA1, rB0, m1	fadds	rC23, m2, rC23				fmuls	rA2, rB0, m2	fadds	rC33, m3, rC33				fmuls	rA3, rB0, m3!!	K=12 iteration!	fadds	rC00, m0, rC00				fmuls	rA0, rB1, m0	fadds	rC10, m1, rC10				fmuls	rA1, rB1, m1	fadds	rC20, m2, rC20				fmuls	rA2, rB1, m2	fadds	rC30, m3, rC30				fmuls	rA3, rB1, m3	fadds	rC01, m0, rC01				fmuls	ra0, rb0, m0	fadds	rC11, m1, rC11				fmuls	ra1, rb0, m1	fadds	rC21, m2, rC21				fmuls	ra2, rb0, m2	fadds	rC31, m3, rC31				fmuls	ra3, rb0, m3							ldd	[pB2+48], rB0	fadds	rC00, m0, rC00				fmuls	ra0, rb1, m0	fadds	rC10, m1, rC10				fmuls	ra1, rb1, m1	fadds	rC20, m2, rC20				fmuls	ra2, rb1, m2	fadds	rC30, m3, rC30				fmuls	ra3, rb1, m3							ldd	[pB3+48], rB1	fadds	rC01, m0, rC01				fmuls	rA0, rB0, m0	fadds	rC11, m1, rC11				fmuls	rA1, rB0, m1	fadds	rC21, m2, rC21				fmuls	rA2, rB0, m2	fadds	rC31, m3, rC31				fmuls	rA3, rB0, m3	fadds	rC02, m0, rC02				fmuls	rA0, rB1, m0	fadds	rC12, m1, rC12				fmuls	rA1, rB1, m1	fadds	rC22, m2, rC22				fmuls	rA2, rB1, m2	fadds	rC32, m3, rC32				fmuls	rA3, rB1, m3	fadds	rC03, m0, rC03				fmuls	ra0, rb0, m0	fadds	rC13, m1, rC13				fmuls	ra1, rb0, m1	fadds	rC23, m2, rC23				fmuls	ra2, rb0, m2	fadds	rC33, m3, rC33				fmuls	ra3, rb0, m3							ldd	[pB0+56], rB0	fadds	rC02, m0, rC02				fmuls	ra0, rb1, m0							ldd	[pA0+56], rA0	fadds	rC12, m1, rC12				fmuls	ra1, rb1, m1							ldd	[pA1+56], rA1	fadds	rC22, m2, rC22				fmuls	ra2, rb1, m2							ldd	[pA2+56], rA2	fadds	rC32, m3, rC32				fmuls	ra3, rb1, m3							ldd	[pA3+56], rA3	fadds	rC03, m0, rC03				fmuls	rA0, rB0, m0							ldd	[pB1+56], rB1	fadds	rC13, m1, rC13				fmuls	rA1, rB0, m1	fadds	rC23, m2, rC23				fmuls	rA2, rB0, m2	fadds	rC33, m3, rC33				fmuls	rA3, rB0, m3!!	K=14 iteration!	fadds	rC00, m0, rC00				fmuls	rA0, rB1, m0	fadds	rC10, m1, rC10				fmuls	rA1, rB1, m1	fadds	rC20, m2, rC20				fmuls	rA2, rB1, m2	fadds	rC30, m3, rC30				fmuls	rA3, rB1, m3	fadds	rC01, m0, rC01				fmuls	ra0, rb0, m0	fadds	rC11, m1, rC11				fmuls	ra1, rb0, m1	fadds	rC21, m2, rC21				fmuls	ra2, rb0, m2	fadds	rC31, m3, rC31				fmuls	ra3, rb0, m3							ldd	[pB2+56], rB0	fadds	rC00, m0, rC00				fmuls	ra0, rb1, m0	fadds	rC10, m1, rC10				fmuls	ra1, rb1, m1	fadds	rC20, m2, rC20				fmuls	ra2, rb1, m2	fadds	rC30, m3, rC30				fmuls	ra3, rb1, m3							ldd	[pB3+56], rB1	fadds	rC01, m0, rC01				fmuls	rA0, rB0, m0	fadds	rC11, m1, rC11				fmuls	rA1, rB0, m1	fadds	rC21, m2, rC21				fmuls	rA2, rB0, m2	fadds	rC31, m3, rC31				fmuls	rA3, rB0, m3	fadds	rC02, m0, rC02				fmuls	rA0, rB1, m0	fadds	rC12, m1, rC12				fmuls	rA1, rB1, m1	fadds	rC22, m2, rC22				fmuls	rA2, rB1, m2	fadds	rC32, m3, rC32				fmuls	rA3, rB1, m3	fadds	rC03, m0, rC03				fmuls	ra0, rb0, m0							prefR2([pfA])							prefR2([pfB])	fadds	rC13, m1, rC13				fmuls	ra1, rb0, m1							add	pfA, 64, pfA	fadds	rC23, m2, rC23				fmuls	ra2, rb0, m2							add	pfB, 64, pfB	fadds	rC33, m3, rC33				fmuls	ra3, rb0, m3	fadds	rC02, m0, rC02				fmuls	ra0, rb1, m0				add	pA0, incAm, pA0	fadds	rC12, m1, rC12				fmuls	ra1, rb1, m1				add	pA1, incAm, pA1	fadds	rC22, m2, rC22				fmuls	ra2, rb1, m2				add	pA2, incAm, pA2	fadds	rC32, m3, rC32				fmuls	ra3, rb1, m3				add	pA3, incAm, pA3	fadds	rC03, m0, rC03				add	pB0, incBm, pB0	fadds	rC13, m1, rC13				add	pB1, incBm, pB1	fadds	rC23, m2, rC23				add	pB2, incBm, pB2	fadds	rC33, m3, rC33				add	pB3, incBm, pB3!!       Write answer back to C!	st	rC00, [pC0]	st	rC10, [pC0+CMUL(4)]	st	rC20, [pC0+CMUL(8)]	st	rC30, [pC0+CMUL(12)]					add	pC0, CMUL(16), pC0	st	rC01, [pC1]	st	rC11, [pC1+CMUL(4)]	st	rC21, [pC1+CMUL(8)]	st	rC31, [pC1+CMUL(12)]					add	pC1, CMUL(16), pC1	st	rC02, [pC2]	st	rC12, [pC2+CMUL(4)]	st	rC22, [pC2+CMUL(8)]	st	rC32, [pC2+CMUL(12)]					add	pC2, CMUL(16), pC2	st	rC03, [pC3]	st	rC13, [pC3+CMUL(4)]	st	rC23, [pC3+CMUL(8)]	st	rC33, [pC3+CMUL(12)]	subcc	II, 4, II	bnz	MLOOP					add	pC3, CMUL(16), pC3	add	pC0, incCn, pC0	add	pC1, incCn, pC1	add	pC2, incCn, pC2	add	pC3, incCn, pC3	add	pA0, incAn, pA0	add	pA1, incAn, pA1	add	pA2, incAn, pA2	add	pA3, incAn, pA3	add	pB0, incBn, pB0	add	pB1, incBn, pB1	add	pB2, incBn, pB2!!       while(N);!	subcc	N, 4, N	bnz	NLOOP	add	pB3, incBn, pB3!!	Epilogue!	ret	restore

⌨️ 快捷键说明

复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?