From ac494c0d04e12c83b38cab845578b5c147696232 Mon Sep 17 00:00:00 2001
From: traz <wangqian10@iscas.ac.cn>
Date: Wed, 6 Apr 2011 10:36:44 +0000
Subject: [PATCH] New kernel in LOONGSON3A.

---
 kernel/mips64/gemm_kernel_loongson3a.S | 1631 ++++++++++++++++++++++++
 1 file changed, 1631 insertions(+)
 create mode 100644 kernel/mips64/gemm_kernel_loongson3a.S

diff --git a/kernel/mips64/gemm_kernel_loongson3a.S b/kernel/mips64/gemm_kernel_loongson3a.S
new file mode 100644
index 000000000..d19d65469
--- /dev/null
+++ b/kernel/mips64/gemm_kernel_loongson3a.S
@@ -0,0 +1,1631 @@
+#define gsLQC1(base,fq,ft,offset) .word(0x32<<26|base<<21|ft<<16|0x1<<15|offset<<6|0x1<<5|fq)
+#define gsSQC1(base,fq,ft,offset) .word(0x3A<<26|base<<21|ft<<16|0x1<<15|offset<<6|0x1<<5|fq)
+#define REALNAME ASMNAME 
+#define PROLOGUE \
+	.text ;\
+	.set	mips64 ;\
+	.align 5 ;\
+	.globl	REALNAME ;\
+	.ent	REALNAME ;\
+	.type	REALNAME, @function ;\
+REALNAME: ;\
+	.set	noreorder ;\
+	.set	nomacro
+
+#define EPILOGUE \
+	.set	macro ;\
+	.set	reorder ;\
+	.end	REALNAME
+#define BASE_SHIFT 3
+#define FETCH	ld
+
+#define M	$4
+#define	N	$5
+#define	K	$6
+#define A	$8
+#define B	$9
+#define C	$10
+#define LDC	$11
+
+#define AO	$12
+#define BO	$13
+
+#define I	$2
+#define J	$3
+#define L	$7
+
+#define CO1	$14
+#define CO2	$15
+#define CO3	$16
+#define CO4	$17
+
+#define KCO	$18
+#define MCO	$19
+#define NCO	$20
+
+#define SPANB	$21
+#define SPANC	$22
+#define PREB	$23
+#define PREA	$24
+#define SPANA	$25
+
+#define ALPHA	$f15
+
+#define R8	8
+#define	R9	9
+#define R14	14
+#define R15	15
+#define R16	16
+#define R17 17
+
+#define	t11	$f30
+#define	t21	$f31
+#define	t31	$f28
+#define	t41	$f29
+
+#define	t12	$f26
+#define	t22	$f27
+#define	t32	$f24
+#define	t42	$f25
+
+#define	t13	$f22
+#define	t23	$f23
+#define	t33	$f20
+#define	t43	$f21
+
+#define	t14	$f18
+#define	t24	$f19
+#define	t34	$f16
+#define	t44	$f17
+
+#define	c11	$f0
+#define	c21	$f1
+#define	c31	$f2
+#define	c41	$f3
+
+#define	c12	$f4
+#define	c22	$f5
+#define	c32	$f6
+#define	c42	$f7
+
+#define	c13	$f8
+#define	c23	$f9
+#define	c33	$f10
+#define c43	$f11
+
+#define	c14	$f12
+#define	c24	$f13
+#define	c34	$f14
+#define	c44	$f0
+
+#define	a0	$f0
+#define	a1	$f1
+#define	a2	$f2
+#define	a3	$f3
+#define	a4	$f4
+#define	a5	$f5
+#define	a6	$f6
+#define	a7	$f7
+#define	b0	$f8
+#define	b1	$f9
+#define	b2	$f10
+#define b3	$f11
+#define	b4	$f12
+#define	b5	$f13
+#define	b6	$f14
+#define	b7	$f15
+
+#define F31 31
+#define F30 30
+#define F29 29
+#define F28 28
+#define F27 27
+#define F26 26
+#define F25 25
+#define F24 24 
+#define F23 23
+#define F22 22
+#define F21 21
+#define F20 20
+#define F19 19
+#define F18 18
+#define F17 17
+#define F16 16 
+#define F15 15
+#define F14 14
+#define F13 13
+#define F12 12
+#define F11 11
+#define F10 10
+#define F9 9
+#define F8 8
+#define F7 7
+#define F6 6
+#define F5 5
+#define F4 4 
+#define F3 3 
+#define F2 2 
+#define F1 1 
+#define F0 0
+
+	PROLOGUE
+	
+	daddiu	$sp, $sp, -160
+	sd	$16,   0($sp)
+	sd	$17,   8($sp)
+	sd	$18,  16($sp)
+	sd	$19,  24($sp)
+	sd	$20,  32($sp)
+	sd	$21,  40($sp)
+	sd	$22,  48($sp)
+	sdc1	$f24, 56($sp)
+	sdc1	$f25, 64($sp)
+	sdc1	$f26, 72($sp)
+	sdc1	$f27, 80($sp)
+	sdc1	$f28, 88($sp)
+	sd	$23,  96($sp)
+	sd	$24, 104($sp)
+	sd	$25, 112($sp)
+	sdc1	$f20,120($sp)
+	sdc1	$f21,128($sp)
+	sdc1	$f22,136($sp)
+	sdc1	$f23,144($sp)
+
+
+	.align	5					#  BACKUP
+.L0_N4:							#  Loop N
+	sdc1	ALPHA,152($sp)		#  Backup	ALPHA
+	move	MCO,M				#  Backup	M
+
+	move	NCO,N				#  Backup	N
+	move	KCO,K				#  Backup	K
+
+	move	AO,A				#  Backup	A_addr
+	move	BO,B				#  Backup	B_addr
+	
+	dsll	LDC,LDC,3			#  LDC*8Byte
+	dsll	SPANB,KCO,5			#  SPANB=KC*NR(4)*8Byte=KC*2^5
+	
+	dsll	SPANA,KCO,5			#  SPANA = KCO*4mr*8Byte
+	dsra	N,NCO,2				#  N=NCO/2
+	beq		N,$0,.L0_N2			#  N=0,NCO<4
+	dsll	SPANC,LDC,2			#  SPANC=LDC*4
+
+.L0_N4_Lb:
+	move	CO1,C				#  Set C	
+	dsra	M,MCO,2				#  M=MCO/2
+	
+	move	A,AO				#  Reset A
+	daddu	CO2,CO1,LDC
+
+	daddu	CO3,CO2,LDC
+	daddu	PREB,BO,SPANB		#  PreB point next panelB
+
+	daddu	CO4,CO3,LDC
+	beqz	M,.L14_M2
+	daddu	PREA,AO,SPANA
+
+.L10:
+	dmtc1	$0,t11
+	mov.d	t21,t11
+	gsLQC1(R8,F1,F0,0)			#a0,a1
+
+	mov.d	t31,t11
+	mov.d	t41,t11
+	gsLQC1(R9,F9,F8,0)			#b0,b1
+
+	mov.d	t12,t11
+	mov.d	t22,t11
+	gsLQC1(R8,F3,F2,1)			#a2,a3
+	
+	mov.d	t32,t11
+	mov.d	t42,t11
+	gsLQC1(R9,F11,F10,1)		#b2,b3
+
+	dsra	K,KCO,2				#  K=KCO/2
+	mov.d	t13,t11
+	
+	mov.d	t23,t11
+	mov.d	t33,t11
+	
+	mov.d	t43,t11
+	mov.d	t14,t11
+	
+	mov.d	t24,t11
+	mov.d	t34,t11
+	
+	mov.d	t44,t11
+	beqz	K,.L15
+	nop
+
+.L11:							#  N=M=K=4
+	gsLQC1(R8,F5,F4,2)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	gsLQC1(R9,F13,F12,2)		#  R9=B
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+	gsLQC1(R8,F7,F6,3)
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+	
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t32,t32,a2,b1
+	madd.d	t42,t42,a3,b1
+
+	FETCH		$0,(PREB)
+	madd.d	t13,t13,a0,b2
+	madd.d	t23,t23,a1,b2
+	
+	madd.d	t14,t14,a0,b3
+	madd.d	t24,t24,a1,b3
+	
+	FETCH		$0,(PREA)
+	madd.d	t33,t33,a2,b2
+	madd.d	t43,t43,a3,b2
+
+	madd.d	t34,t34,a2,b3
+	madd.d	t44,t44,a3,b3
+								#load2 comp1
+.L12:
+	gsLQC1(R8,F1,F0,4)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F9,F8,4)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+
+	gsLQC1(R8,F3,F2,5)
+	madd.d	t31,t31,a6,b4
+	madd.d	t41,t41,a7,b4
+
+	gsLQC1(R9,F11,F10,5)
+	madd.d	t32,t32,a6,b5
+	madd.d	t42,t42,a7,b5
+
+	FETCH		$0,32(PREB)
+	madd.d	t13,t13,a4,b6
+	madd.d	t23,t23,a5,b6
+
+	madd.d	t14,t14,a4,b7
+	madd.d	t24,t24,a5,b7
+
+	FETCH		$0,32(PREA)
+	madd.d	t33,t33,a6,b6
+	madd.d	t43,t43,a7,b6
+	
+	madd.d	t34,t34,a6,b7
+	madd.d	t44,t44,a7,b7
+
+.L13:
+	gsLQC1(R8,F5,F4,6)	
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	gsLQC1(R9,F13,F12,6)
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+	gsLQC1(R8,F7,F6,7)
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+
+	gsLQC1(R9,F15,F14,7)
+	madd.d	t32,t32,a2,b1
+	madd.d	t42,t42,a3,b1
+	daddu	A,A,128				#  A+=4(mr)*4(kr)*8Byte=128
+
+	FETCH		$0,64(PREB)
+	madd.d	t13,t13,a0,b2
+	madd.d	t23,t23,a1,b2
+	daddu	B,B,128
+
+	madd.d	t14,t14,a0,b3
+	madd.d	t24,t24,a1,b3
+
+	FETCH		$0,64(PREA)
+	madd.d	t33,t33,a2,b2
+	madd.d	t43,t43,a3,b2
+
+	madd.d	t34,t34,a2,b3
+	madd.d	t44,t44,a3,b3
+	
+.L14:
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+
+	gsLQC1(R8,F3,F2,1)
+	madd.d	t31,t31,a6,b4
+	madd.d	t41,t41,a7,b4
+	daddiu	K,K,-1
+
+	gsLQC1(R9,F11,F10,1)
+	madd.d	t32,t32,a6,b5
+	madd.d	t42,t42,a7,b5
+
+	FETCH		$0,96(PREB)
+	madd.d	t13,t13,a4,b6
+	madd.d	t23,t23,a5,b6
+
+	madd.d	t14,t14,a4,b7
+	madd.d	t24,t24,a5,b7
+
+	FETCH		$0,96(PREA)
+	madd.d	t33,t33,a6,b6
+	madd.d	t43,t43,a7,b6
+	daddu	PREB,PREB,128
+	
+	madd.d	t34,t34,a6,b7
+	daddu	PREA,PREA,128
+	bnez 	K,.L11
+	madd.d	t44,t44,a7,b7
+
+.L15:							#  N=4 M=4 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L18
+	nop
+
+.L16:			
+	gsLQC1(R8,F5,F4,2)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	gsLQC1(R9,F13,F12,2)		#  R9=B
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+	gsLQC1(R8,F7,F6,3)
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t32,t32,a2,b1
+	madd.d	t42,t42,a3,b1
+	daddu	A,A,64				#  A+=4(mr)*2(kr)*8Byte=64
+
+	FETCH		$0,0(PREB)
+	madd.d	t13,t13,a0,b2
+	madd.d	t23,t23,a1,b2
+	daddu	B,B,64
+
+	madd.d	t14,t14,a0,b3
+	madd.d	t24,t24,a1,b3
+
+	FETCH		$0,0(PREA)
+	madd.d	t33,t33,a2,b2
+	madd.d	t43,t43,a3,b2
+
+	madd.d	t34,t34,a2,b3
+	madd.d	t44,t44,a3,b3
+								
+.L17:
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+
+	gsLQC1(R8,F3,F2,1)
+	madd.d	t31,t31,a6,b4
+	madd.d	t41,t41,a7,b4
+
+	gsLQC1(R9,F11,F10,1)
+	madd.d	t32,t32,a6,b5
+	madd.d	t42,t42,a7,b5
+
+	FETCH		$0,32(PREB)
+	madd.d	t13,t13,a4,b6
+	madd.d	t23,t23,a5,b6
+
+	madd.d	t14,t14,a4,b7
+	madd.d	t24,t24,a5,b7
+
+	FETCH		$0,32(PREA)
+	madd.d	t33,t33,a6,b6
+	madd.d	t43,t43,a7,b6
+	daddu	PREB,PREB,64
+	
+	madd.d	t34,t34,a6,b7
+	madd.d	t44,t44,a7,b7
+	daddu	PREA,PREA,64
+	
+.L18:							#  N=4, M=4, K=1
+	and		K,KCO,1
+	beqz	K,.L19				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	
+	FETCH		$0,0(PREB)
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	A,A,32				#  A+=4(mr)*1(kr)*8Byte=32
+
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+	daddu	B,B,32
+
+	FETCH		$0,0(PREA)
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+	daddu	PREB,PREB,32
+
+	madd.d	t32,t32,a2,b1
+	madd.d	t42,t42,a3,b1
+	daddu	PREA,PREA,32
+
+	madd.d	t13,t13,a0,b2
+	madd.d	t23,t23,a1,b2
+
+	madd.d	t14,t14,a0,b3
+	madd.d	t24,t24,a1,b3
+
+	madd.d	t33,t33,a2,b2
+	madd.d	t43,t43,a3,b2
+
+	madd.d	t34,t34,a2,b3
+	madd.d	t44,t44,a3,b3
+
+.L19:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	ldc1	c21,8(CO1)			
+	ldc1	c31,16(CO1)
+	ldc1	c41,24(CO1)
+
+	ldc1	c12,0(CO2)
+	madd.d	t11,c11,t11,ALPHA
+	ldc1	c22,8(CO2)
+	madd.d	t21,c21,t21,ALPHA
+	ldc1	c32,16(CO2)
+	madd.d	t31,c31,t31,ALPHA
+	ldc1	c42,24(CO2)
+	madd.d	t41,c41,t41,ALPHA
+
+	ldc1	c13,0(CO3)
+	madd.d	t12,c12,t12,ALPHA
+	ldc1	c23,8(CO3)
+	madd.d	t22,c22,t22,ALPHA
+	ldc1	c33,16(CO3)
+	madd.d	t32,c32,t32,ALPHA
+	ldc1	c43,24(CO3)
+	madd.d	t42,c42,t42,ALPHA
+
+	ldc1	c14,0(CO4)
+	madd.d	t13,c13,t13,ALPHA
+	ldc1	c24,8(CO4)
+	madd.d	t23,c23,t23,ALPHA
+	ldc1	c34,16(CO4)
+	madd.d	t33,c33,t33,ALPHA
+	ldc1	c44,24(CO4)
+	madd.d	t43,c43,t43,ALPHA
+
+	sdc1	t11,0(CO1)
+	madd.d	t14,c14,t14,ALPHA
+	sdc1	t21,8(CO1)
+	madd.d	t24,c24,t24,ALPHA
+	sdc1	t31,16(CO1)
+	madd.d	t34,c34,t34,ALPHA
+	sdc1	t41,24(CO1)
+	madd.d	t44,c44,t44,ALPHA
+	daddiu	M,M,-1				#  M--
+
+	sdc1	t12,0(CO2)
+	sdc1	t22,8(CO2)
+	sdc1	t32,16(CO2)
+	sdc1	t42,24(CO2)
+
+	sdc1	t13,0(CO3)
+	sdc1	t23,8(CO3)
+	sdc1	t33,16(CO3)
+	sdc1	t43,24(CO3)
+
+	FETCH	$0,32(CO1)
+	FETCH	$0,32(CO2)
+	FETCH	$0,32(CO3)
+	FETCH	$0,32(CO4)
+
+	sdc1	t14,0(CO4)
+	daddu	CO1,CO1,32			#  COx += 4*8Byte
+	sdc1	t24,8(CO4)
+	daddu	CO2,CO2,32
+	sdc1	t34,16(CO4)
+	daddu	CO3,CO3,32
+	sdc1	t44,24(CO4)
+	move	B,BO				#  Reset B
+	daddu	PREB,BO,SPANB
+	bnez	M,.L10				#  M!=0
+	daddu	CO4,CO4,32
+
+
+
+.L14_M2:
+	and		M,MCO,2				#  Remainder M = 2
+	beqz	M,.L14_M1			
+	nop
+
+.L20:
+	dmtc1	$0,t11
+	mov.d	t21,t11
+	gsLQC1(R8,F1,F0,0)			#a0,a1
+
+	mov.d	t12,t11
+	mov.d	t22,t11
+	gsLQC1(R9,F9,F8,0)			#b0,b1
+	
+	dsra	K,KCO,2				#  K=KCO/2
+	mov.d	t13,t11
+	gsLQC1(R9,F11,F10,1)		#b2,b3
+	
+	mov.d	t23,t11
+	mov.d	t14,t11
+	
+	mov.d	t24,t11
+	beqz	K,.L25
+	nop
+
+.L21:							#  N=4 m=2,=K=4
+	gsLQC1(R8,F5,F4,1)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	gsLQC1(R9,F13,F12,2)		#  R9=B
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t13,t13,a0,b2
+	madd.d	t23,t23,a1,b2
+	
+	gsLQC1(R8,F3,F2,2)
+	madd.d	t14,t14,a0,b3
+	madd.d	t24,t24,a1,b3
+	
+	gsLQC1(R9,F9,F8,4)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F11,F10,5)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+
+	gsLQC1(R8,F7,F6,3)	
+	madd.d	t13,t13,a4,b6
+	madd.d	t23,t23,a5,b6
+
+	madd.d	t14,t14,a4,b7
+	madd.d	t24,t24,a5,b7
+
+	gsLQC1(R9,F13,F12,6)
+	madd.d	t11,t11,a2,b0
+	madd.d	t21,t21,a3,b0
+	daddu	A,A,64				#  A+=2(mr)*4(kr)*8Byte=64
+
+	gsLQC1(R9,F15,F14,7)
+	madd.d	t12,t12,a2,b1
+	madd.d	t22,t22,a3,b1
+	daddiu	K,K,-1
+
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t13,t13,a2,b2
+	madd.d	t23,t23,a3,b2
+	daddu	B,B,128				#  B+=4(nr)*4(kr)*8Byte=128
+
+	madd.d	t14,t14,a2,b3
+	madd.d	t24,t24,a3,b3
+
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t11,t11,a6,b4
+	madd.d	t21,t21,a7,b4
+
+	gsLQC1(R9,F11,F10,1)
+	madd.d	t12,t12,a6,b5
+	madd.d	t22,t22,a7,b5
+
+	madd.d	t13,t13,a6,b6
+	madd.d	t23,t23,a7,b6
+
+	madd.d	t14,t14,a6,b7
+	bnez 	K,.L21
+	madd.d	t24,t24,a7,b7
+
+.L25:							#  N=4 M=2 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L28
+	nop
+
+.L26:			
+	gsLQC1(R8,F5,F4,1)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	gsLQC1(R9,F13,F12,2)		#  R9=B
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+	daddu	A,A,32				#  A+=2(mr)*2(kr)*8Byte=32
+
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t13,t13,a0,b2
+	madd.d	t23,t23,a1,b2
+	daddu	B,B,64
+
+	madd.d	t14,t14,a0,b3
+	madd.d	t24,t24,a1,b3
+
+.L27:
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+
+	gsLQC1(R9,F11,F10,1)
+	madd.d	t13,t13,a4,b6
+	madd.d	t23,t23,a5,b6
+
+	madd.d	t14,t14,a4,b7
+	madd.d	t24,t24,a5,b7
+	
+.L28:							#  N=4, M=2, K=1
+	and		K,KCO,1
+	beqz	K,.L29				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	A,A,16				#  A+=2(mr)*1(kr)*8Byte=16
+	daddu	B,B,32
+
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+	madd.d	t13,t13,a0,b2
+	madd.d	t23,t23,a1,b2
+
+	madd.d	t14,t14,a0,b3
+	madd.d	t24,t24,a1,b3
+
+.L29:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	ldc1	c21,8(CO1)			
+
+	ldc1	c12,0(CO2)
+	ldc1	c22,8(CO2)
+	
+	ldc1	c13,0(CO3)
+	madd.d	t11,c11,t11,ALPHA
+	ldc1	c23,8(CO3)
+	madd.d	t21,c21,t21,ALPHA
+
+	ldc1	c14,0(CO4)
+	madd.d	t12,c12,t12,ALPHA
+	ldc1	c24,8(CO4)
+	madd.d	t22,c22,t22,ALPHA
+
+	sdc1	t11,0(CO1)
+	madd.d	t13,c13,t13,ALPHA
+	sdc1	t21,8(CO1)
+	madd.d	t23,c23,t23,ALPHA
+
+	sdc1	t12,0(CO2)
+	madd.d	t14,c14,t14,ALPHA
+	sdc1	t22,8(CO2)
+	madd.d	t24,c24,t24,ALPHA
+
+	sdc1	t13,0(CO3)
+	move	B,BO				#  Reset B
+	sdc1	t23,8(CO3)
+	daddu	CO1,CO1,16			#  COx += 2*8Byte
+
+	FETCH	$0,0(CO1)
+	FETCH	$0,16(CO2)
+	FETCH	$0,16(CO3)
+	FETCH	$0,16(CO4)
+
+	sdc1	t14,0(CO4)
+	daddu	CO2,CO2,16
+	sdc1	t24,8(CO4)
+	daddu	CO3,CO3,16
+	daddu	CO4,CO4,16
+
+
+
+.L14_M1:
+	and		M,MCO,1				#  Remainder M = 1
+	beqz	M,.L0_N4_Loop		#  M = 0, finishing one panel B
+	nop
+
+.L30:
+	ldc1	a0,0(A)
+	dsra	K,KCO,2				#  K=KCO/2
+	gsLQC1(R9,F9,F8,0)			#b0,b1
+	dmtc1	$0,t11
+	gsLQC1(R9,F11,F10,1)		#b2,b3
+	mov.d	t12,t11
+	mov.d	t13,t11
+	beqz	K,.L25
+	mov.d	t14,t11
+
+.L31:							#  N=4 m=1,=K=4
+	ldc1	a1,8(A)
+	
+	gsLQC1(R9,F13,F12,2)		#  R9=B
+	madd.d	t11,t11,a0,b0
+	madd.d	t12,t12,a0,b1
+	
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t13,t13,a0,b2
+	madd.d	t14,t14,a0,b3
+	
+	ldc1	a2,16(A)
+	
+	gsLQC1(R9,F9,F8,4)
+	madd.d	t11,t11,a1,b4
+	madd.d	t12,t12,a1,b5
+	
+	gsLQC1(R9,F11,F10,5)
+	madd.d	t13,t13,a1,b6
+	madd.d	t14,t14,a1,b7
+
+	ldc1	a3,24(A)
+	daddiu	K,K,-1
+	
+	gsLQC1(R9,F13,F12,6)
+	madd.d	t11,t11,a2,b0
+	madd.d	t12,t12,a2,b1
+	daddu	A,A,32				#  A+=1(mr)*4(kr)*8Byte=64
+	
+	gsLQC1(R9,F15,F14,7)
+	madd.d	t13,t13,a2,b2
+	madd.d	t14,t14,a2,b3
+	daddu	B,B,128				#  B+=4(nr)*4(kr)*8Byte=128
+
+	ldc1	a0,0(A)
+	
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t11,t11,a3,b4
+	madd.d	t12,t12,a3,b5
+	
+	gsLQC1(R9,F11,F10,1)
+	madd.d	t13,t13,a3,b6
+	bnez 	K,.L31
+	madd.d	t14,t14,a3,b7
+
+.L35:							#  N=4 M=1 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L38
+	nop
+
+.L36:			
+	ldc1	a1,8(A)	
+	
+	gsLQC1(R9,F13,F12,2)		#  R9=B
+	madd.d	t11,t11,a0,b0
+	madd.d	t12,t12,a0,b1
+	daddu	A,A,16				#  A+=1(mr)*2(kr)*8Byte=32
+	
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t13,t13,a0,b2
+	madd.d	t14,t14,a0,b3
+	daddu	B,B,64
+
+
+.L37:
+	ldc1	a0,0(A)
+	
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t11,t11,a1,b4
+	madd.d	t12,t12,a1,b5
+	
+	gsLQC1(R9,F11,F10,1)
+	madd.d	t13,t13,a1,b6
+	madd.d	t14,t14,a1,b7
+	
+.L38:							#  N=4, M=1, K=1
+	and		K,KCO,1
+	beqz	K,.L39				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	
+	madd.d	t11,t11,a0,b0
+	madd.d	t12,t12,a0,b1
+	daddu	A,A,8				#  A+=1(mr)*1(kr)*8Byte=16
+	daddu	B,B,32
+	
+	madd.d	t13,t13,a0,b2
+	madd.d	t14,t14,a0,b3
+
+.L39:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	ldc1	c12,0(CO2)
+	ldc1	c13,0(CO3)
+	ldc1	c14,0(CO4)
+	
+	madd.d	t11,c11,t11,ALPHA
+	madd.d	t12,c12,t12,ALPHA
+	madd.d	t13,c13,t13,ALPHA
+	madd.d	t14,c14,t14,ALPHA
+
+	sdc1	t11,0(CO1)
+	sdc1	t12,0(CO2)
+	sdc1	t13,0(CO3)
+	sdc1	t14,0(CO4)
+
+
+.L0_N4_Loop:
+	daddu	BO,BO,SPANB			#  BO point to next panel B
+	daddiu	N,N,-1				#  N--
+	daddu	C,C,SPANC			#  C pointe to next panel C
+	bnez	N,.L0_N4_Lb			#  N!=0
+	move	B,BO				#  Set B
+
+
+
+	.align	5					
+.L0_N2:
+	and		N,NCO,2				#  Remainder N = 2
+	beqz	N,.L0_N1			#  N=0,NCO<2
+	dsll	SPANC,LDC,1			#  SPANC=LDC*2
+
+.L0_N2_Lb:
+	move	CO1,C				#  Set C	
+	dsra	M,MCO,2				#  M=MCO/2
+	
+	dsll	SPANB,KCO,4			#  SPANB=KC*NR(2)*8Byte=KC*16=KC*2^4
+	move	A,AO				#  Reset A
+
+	daddu	CO2,CO1,LDC
+	beqz	M,.L12_M2
+	daddu	PREA,AO,SPANA
+
+.L40:						
+	dmtc1	$0,t11
+	mov.d	t21,t11
+	gsLQC1(R8,F1,F0,0)			#a0,a1
+
+	mov.d	t31,t11
+	mov.d	t41,t11
+	gsLQC1(R9,F9,F8,0)			#b0,b1
+
+	dsra	K,KCO,2				#  K=KCO/2
+	mov.d	t12,t11
+	gsLQC1(R8,F3,F2,1)			#a2,a3
+	
+	mov.d	t22,t11
+	mov.d	t32,t11
+	
+	mov.d	t42,t11
+	beqz	K,.L45
+	nop
+
+.L41:							#  N=2,M=K=4
+	gsLQC1(R8,F5,F4,2)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	gsLQC1(R9,F13,F12,1)		#  R9=B
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+	gsLQC1(R8,F7,F6,3)
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+	
+	FETCH		$0,(PREA)
+	madd.d	t32,t32,a2,b1
+	madd.d	t42,t42,a3,b1
+
+.L42:
+	gsLQC1(R8,F1,F0,4)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F11,F10,2)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+
+	gsLQC1(R8,F3,F2,5)
+	madd.d	t31,t31,a6,b4
+	madd.d	t41,t41,a7,b4
+
+	FETCH		$0,32(PREA)
+	madd.d	t32,t32,a6,b5
+	madd.d	t42,t42,a7,b5
+
+.L43:
+	gsLQC1(R8,F5,F4,6)	
+	madd.d	t11,t11,a0,b2
+	madd.d	t21,t21,a1,b2
+
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t12,t12,a0,b3
+	madd.d	t22,t22,a1,b3
+
+	gsLQC1(R8,F7,F6,7)
+	madd.d	t31,t31,a2,b2
+	madd.d	t41,t41,a3,b2
+	daddu	B,B,64				#  B+=2(nr)*4(kr)*8Byte=64
+
+	FETCH		$0,64(PREA)
+	madd.d	t32,t32,a2,b3
+	madd.d	t42,t42,a3,b3
+	daddu	A,A,128				#  A+=4(mr)*4(kr)*8Byte=128
+
+.L44:
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a4,b6
+	madd.d	t21,t21,a5,b6
+	daddiu	K,K,-1
+
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t12,t12,a4,b7
+	madd.d	t22,t22,a5,b7
+	daddu	PREA,PREA,128
+
+	gsLQC1(R8,F3,F2,1)
+	madd.d	t31,t31,a6,b6
+	madd.d	t41,t41,a7,b6
+
+	FETCH		$0,-32(PREA)
+	madd.d	t32,t32,a6,b7
+	bnez 	K,.L41
+	madd.d	t42,t42,a7,b7
+
+
+.L45:							#  N=2 M=4 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L48
+	nop
+
+.L46:			
+	gsLQC1(R8,F5,F4,2)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	gsLQC1(R9,F13,F12,1)		#  R9=B
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+	gsLQC1(R8,F7,F6,3)
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+	daddu	B,B,32				#  B+=2(nr)*2(kr)*8Byte=32
+
+	FETCH		$0,0(PREA)
+	madd.d	t32,t32,a2,b1
+	madd.d	t42,t42,a3,b1
+	daddu	A,A,64				#  A+=4(mr)*2(kr)*8Byte=64
+
+.L47:
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+
+	gsLQC1(R8,F3,F2,1)
+	madd.d	t31,t31,a6,b4
+	madd.d	t41,t41,a7,b4
+
+	FETCH		$0,32(PREA)
+	madd.d	t32,t32,a6,b5
+	madd.d	t42,t42,a7,b5
+	daddu	PREA,PREA,64
+
+	
+.L48:							#  N=2, M=4, K=1
+	and		K,KCO,1
+	beqz	K,.L49				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	
+	FETCH		$0,0(PREA)
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	A,A,32				#  A+=4(mr)*1(kr)*8Byte=32
+
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+	daddu	B,B,32
+	daddu	PREA,PREA,32
+
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+
+	madd.d	t32,t32,a2,b1
+	madd.d	t42,t42,a3,b1
+
+.L49:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	ldc1	c21,8(CO1)			
+	ldc1	c31,16(CO1)
+	ldc1	c41,24(CO1)
+
+	ldc1	c12,0(CO2)
+	madd.d	t11,c11,t11,ALPHA
+	ldc1	c22,8(CO2)
+	madd.d	t21,c21,t21,ALPHA
+	ldc1	c32,16(CO2)
+	madd.d	t31,c31,t31,ALPHA
+	ldc1	c42,24(CO2)
+	madd.d	t41,c41,t41,ALPHA
+
+	sdc1	t11,0(CO1)
+	madd.d	t12,c12,t12,ALPHA
+	sdc1	t21,8(CO1)
+	madd.d	t22,c22,t22,ALPHA
+	sdc1	t31,16(CO1)
+	madd.d	t32,c32,t32,ALPHA
+	sdc1	t41,24(CO1)
+	madd.d	t42,c42,t42,ALPHA
+	daddiu	M,M,-1				#  M--
+
+	sdc1	t12,0(CO2)
+	sdc1	t22,8(CO2)
+	sdc1	t32,16(CO2)
+	sdc1	t42,24(CO2)
+
+	FETCH	$0,32(CO1)
+	FETCH	$0,32(CO2)
+
+	daddu	CO1,CO1,32			#  COx += 4*8Byte
+	daddu	CO2,CO2,32
+	bnez	M,.L40				#  M!=0
+	move	B,BO				#  Reset B
+
+
+.L12_M2:
+	and		M,MCO,2				#  Remainder M = 2
+	beqz	M,.L12_M1			
+	nop
+
+.L50:
+	dsra	K,KCO,2				#  K=KCO/2
+	dmtc1	$0,t11
+	gsLQC1(R8,F1,F0,0)			#a0,a1
+
+	mov.d	t21,t11
+	mov.d	t12,t11
+	gsLQC1(R9,F9,F8,0)			#b0,b1
+	
+	mov.d	t22,t11
+	beqz	K,.L55
+	nop
+
+.L51:							#  N=2 m=2,=K=4
+	gsLQC1(R8,F5,F4,1)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	gsLQC1(R9,F13,F12,1)		#  R9=B
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+	gsLQC1(R8,F3,F2,2)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F11,F10,2)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+	daddiu	K,K,-1
+
+	gsLQC1(R8,F7,F6,3)	
+	madd.d	t11,t11,a2,b2
+	madd.d	t21,t21,a3,b2
+	daddu	A,A,64				#  A+=2(mr)*4(kr)*8Byte=64
+
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t12,t12,a2,b3
+	madd.d	t22,t22,a3,b3
+	daddu	B,B,64				#  B+=2(nr)*4(kr)*8Byte=128
+
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a6,b6
+	madd.d	t21,t21,a7,b6
+
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t12,t12,a6,b7
+	bnez 	K,.L51
+	madd.d	t22,t22,a7,b7
+
+.L55:							#  N=2 M=2 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L58
+	nop
+
+.L56:			
+	gsLQC1(R8,F5,F4,1)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	A,A,32				#  A+=2(mr)*2(kr)*8Byte=32
+
+	gsLQC1(R9,F13,F12,1)		#  R9=B
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+	daddu	B,B,32
+
+.L57:
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t12,t12,a4,b5
+	madd.d	t22,t22,a5,b5
+
+	
+.L58:							#  N=2, M=2, K=1
+	and		K,KCO,1
+	beqz	K,.L59				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	A,A,16				#  A+=2(mr)*1(kr)*8Byte=16
+	daddu	B,B,16
+
+	madd.d	t12,t12,a0,b1
+	madd.d	t22,t22,a1,b1
+
+
+.L59:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	ldc1	c21,8(CO1)			
+	ldc1	c12,0(CO2)
+	ldc1	c22,8(CO2)
+	
+	madd.d	t11,c11,t11,ALPHA
+	madd.d	t21,c21,t21,ALPHA
+	madd.d	t12,c12,t12,ALPHA
+	madd.d	t22,c22,t22,ALPHA
+
+	sdc1	t11,0(CO1)
+	sdc1	t21,8(CO1)
+	sdc1	t12,0(CO2)
+	move	B,BO				#  Reset B
+	sdc1	t22,8(CO2)
+	daddu	CO1,CO1,16			#  COx += 2*8Byte
+	daddu	CO2,CO2,16
+
+	FETCH	$0,0(CO1)
+	FETCH	$0,0(CO2)
+
+
+.L12_M1:
+	and		M,MCO,1				#  Remainder M = 1
+	beqz	M,.L0_N2_Loop		#  M = 0, finishing one panel B
+	nop
+
+.L60:
+	dsra	K,KCO,2				#  K=KCO/2
+	dmtc1	$0,t11
+	ldc1	a0,0(A)
+
+	mov.d	t21,t11
+	mov.d	t12,t11
+	gsLQC1(R9,F9,F8,0)			#b0,b1
+	
+	mov.d	t22,t11
+	beqz	K,.L65
+	nop
+
+.L61:							#  N=2 m=1,=K=4
+	ldc1	a4,8(A)
+	gsLQC1(R9,F13,F12,1)		#  R9=B
+	ldc1	a2,16(A)
+	madd.d	t11,t11,a0,b0
+	madd.d	t12,t12,a0,b1
+
+	gsLQC1(R9,F11,F10,2)
+	madd.d	t11,t11,a4,b4
+	madd.d	t12,t12,a4,b5
+	daddiu	K,K,-1
+
+	ldc1	a6,24(A)	
+	madd.d	t11,t11,a2,b2
+	daddu	A,A,32				#  A+=1(mr)*4(kr)*8Byte=64
+	
+	gsLQC1(R9,F15,F14,3)
+	madd.d	t12,t12,a2,b3
+	daddu	B,B,64				#  B+=2(nr)*4(kr)*8Byte=128
+
+	ldc1	a0,0(A)
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t11,t11,a6,b6
+	bnez 	K,.L61
+	madd.d	t12,t12,a6,b7
+
+.L65:							#  N=2 M=1 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L68
+	nop
+
+.L66:			
+	ldc1	a4,8(A)
+	gsLQC1(R9,F13,F12,1)		#  R9=B
+	madd.d	t11,t11,a0,b0
+	madd.d	t12,t12,a0,b1
+	daddu	A,A,16				#  A+=1(mr)*2(kr)*8Byte=32
+	daddu	B,B,32
+
+.L67:
+	ldc1	a0,0(A)
+	gsLQC1(R9,F9,F8,0)
+	madd.d	t11,t11,a4,b4
+	madd.d	t12,t12,a4,b5
+
+	
+.L68:							#  N=2, M=1, K=1
+	and		K,KCO,1
+	beqz	K,.L69				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	
+	madd.d	t11,t11,a0,b0
+	madd.d	t12,t12,a0,b1
+	daddu	A,A,8				#  A+=1(mr)*1(kr)*8Byte=16
+	daddu	B,B,16
+
+
+.L69:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	ldc1	c12,0(CO2)
+	
+	madd.d	t11,c11,t11,ALPHA
+	madd.d	t12,c12,t12,ALPHA
+
+	sdc1	t11,0(CO1)
+	move	B,BO				#  Reset B
+	sdc1	t12,0(CO2)
+	daddu	CO1,CO1,8			#  COx += 2*8Byte
+	daddu	CO2,CO2,8
+
+	FETCH	$0,0(CO1)
+	FETCH	$0,0(CO2)
+
+
+.L0_N2_Loop:
+	daddu	BO,BO,SPANB			#  BO+=KC*2N
+	move	B,BO				#  Set B
+	daddu	C,C,SPANC			#  C+=LDC*2
+
+
+
+	.align	5					
+.L0_N1:
+	and		N,NCO,1				#  Remainder N = 1
+	beqz	N,.L999				#  N=0,NCO<1
+	nop
+
+	move	CO1,C				#  Set C	
+	dsra	M,MCO,2				#  M=MCO/2
+	
+	move	A,AO				#  Reset A
+	beqz	M,.L11_M2
+	daddu	PREA,AO,SPANA
+
+
+.L70:						
+	dsra	K,KCO,2				#  K=KCO/2
+	ldc1	b0,0(B)
+	dmtc1	$0,t11
+	gsLQC1(R8,F1,F0,0)			#a0,a1
+	mov.d	t21,t11
+	gsLQC1(R8,F3,F2,1)			#a2,a3
+	mov.d	t31,t11
+	beqz	K,.L75
+	mov.d	t41,t11
+
+.L71:							#  N=1,M=K=4
+	ldc1	b4,8(B)
+	gsLQC1(R8,F5,F4,2)			#  R8=A
+	gsLQC1(R8,F7,F6,3)
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	FETCH		$0,(PREA)
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+	
+
+.L72:
+	ldc1	b2,16(B)
+	gsLQC1(R8,F1,F0,4)
+	gsLQC1(R8,F3,F2,5)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	FETCH		$0,32(PREA)
+	madd.d	t31,t31,a6,b4
+	madd.d	t41,t41,a7,b4
+
+
+.L73:
+	ldc1	b6,24(B)
+	gsLQC1(R8,F5,F4,6)	
+	gsLQC1(R8,F7,F6,7)
+	madd.d	t11,t11,a0,b2
+	madd.d	t21,t21,a1,b2
+	daddu	B,B,32				#  B+=1(nr)*4(kr)*8Byte=64
+
+	FETCH		$0,64(PREA)
+	madd.d	t31,t31,a2,b2
+	madd.d	t41,t41,a3,b2
+	daddu	A,A,128				#  A+=4(mr)*4(kr)*8Byte=128
+
+.L74:
+	ldc1	b0,0(B)
+	gsLQC1(R8,F1,F0,0)
+	daddu	PREA,PREA,128
+	gsLQC1(R8,F3,F2,1)
+	madd.d	t11,t11,a4,b6
+	madd.d	t21,t21,a5,b6
+	daddiu	K,K,-1
+
+	FETCH		$0,-32(PREA)
+	madd.d	t31,t31,a6,b6
+	bnez 	K,.L71
+	madd.d	t41,t41,a7,b6
+
+
+
+.L75:							#  N=2 M=4 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L78
+	nop
+
+.L76:			
+	ldc1	b4,8(B)
+	gsLQC1(R8,F5,F4,2)			#  R8=A
+	gsLQC1(R8,F7,F6,3)
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	B,B,16				#  B+=1(nr)*2(kr)*8Byte=32
+
+	FETCH		$0,0(PREA)
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+	daddu	A,A,64				#  A+=4(mr)*2(kr)*8Byte=64
+
+.L77:
+	ldc1	b0,0(B)
+	gsLQC1(R8,F1,F0,0)
+	gsLQC1(R8,F3,F2,1)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	FETCH		$0,32(PREA)
+	madd.d	t31,t31,a6,b4
+	madd.d	t41,t41,a7,b4
+	daddu	PREA,PREA,64
+
+
+	
+.L78:							#  N=2, M=4, K=1
+	and		K,KCO,1
+	beqz	K,.L79				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	
+	FETCH		$0,0(PREA)
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	A,A,32				#  A+=4(mr)*1(kr)*8Byte=32
+
+	madd.d	t31,t31,a2,b0
+	madd.d	t41,t41,a3,b0
+	daddu	B,B,8
+	daddu	PREA,PREA,32
+
+
+.L79:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	ldc1	c21,8(CO1)			
+	ldc1	c31,16(CO1)
+	ldc1	c41,24(CO1)
+
+	madd.d	t11,c11,t11,ALPHA
+	madd.d	t21,c21,t21,ALPHA
+	madd.d	t31,c31,t31,ALPHA
+	madd.d	t41,c41,t41,ALPHA
+
+	sdc1	t11,0(CO1)
+	sdc1	t21,8(CO1)
+	sdc1	t31,16(CO1)
+	sdc1	t41,24(CO1)
+	daddiu	M,M,-1				#  M--
+
+	FETCH	$0,32(CO1)
+	daddu	CO1,CO1,32			#  COx += 4*8Byte
+	bnez	M,.L70				#  M!=0
+	move	B,BO				#  Reset B
+
+
+
+.L11_M2:
+	and		M,MCO,2				#  Remainder M = 2
+	beqz	M,.L11_M1			
+	nop
+
+.L80:						
+	dsra	K,KCO,2				#  K=KCO/2
+	ldc1	b0,0(B)
+	dmtc1	$0,t11
+	gsLQC1(R8,F1,F0,0)			#a0,a1
+	mov.d	t21,t11
+	beqz	K,.L85
+	nop
+
+.L81:							#  N=1,M=2,K=4
+	ldc1	b4,8(B)
+	gsLQC1(R8,F5,F4,1)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+
+	ldc1	b2,16(B)
+	gsLQC1(R8,F3,F2,2)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	ldc1	b6,24(B)
+	daddu	B,B,32				#  B+=1(nr)*4(kr)*8Byte=32
+	
+	gsLQC1(R8,F7,F6,3)
+	madd.d	t11,t11,a2,b2
+	madd.d	t21,t21,a3,b2
+	daddu	A,A,64				#  A+=2(mr)*4(kr)*8Byte=64
+
+	ldc1	b0,0(B)
+	daddiu	K,K,-1
+	
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a6,b6
+	bnez 	K,.L81
+	madd.d	t21,t21,a7,b6
+
+
+.L85:							#  N=2 M=4 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L88
+	nop
+
+.L86:			
+	ldc1	b4,8(B)
+	daddu	B,B,16				#  B+=1(nr)*2(kr)*8Byte=16
+	
+	gsLQC1(R8,F5,F4,1)			#  R8=A
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	A,A,32				#  A+=2(mr)*2(kr)*8Byte=32
+
+	ldc1	b0,0(B)
+	gsLQC1(R8,F1,F0,0)
+	madd.d	t11,t11,a4,b4
+	madd.d	t21,t21,a5,b4
+
+	
+.L88:							#  N=2, M=4, K=1
+	and		K,KCO,1
+	beqz	K,.L89				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	
+	madd.d	t11,t11,a0,b0
+	madd.d	t21,t21,a1,b0
+	daddu	A,A,16				#  A+=2(mr)*1(kr)*8Byte=16
+	daddu	B,B,8
+
+
+.L89:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	ldc1	c21,8(CO1)			
+
+	madd.d	t11,c11,t11,ALPHA
+	madd.d	t21,c21,t21,ALPHA
+
+	sdc1	t11,0(CO1)
+	sdc1	t21,8(CO1)
+
+	FETCH	$0,16(CO1)
+	daddu	CO1,CO1,16			#  COx += 2*8Byte
+	move	B,BO				#  Reset B
+
+
+.L11_M1:
+	and		M,MCO,1				#  Remainder M = 1
+	beqz	M,.L999				#  M = 0, End
+	nop
+
+.L90:						
+	dsra	K,KCO,2				#  K=KCO/2
+	ldc1	b0,0(B)
+	ldc1	a0,0(A)
+	beqz	K,.L95
+	dmtc1	$0,t11
+
+.L91:							#  N=1,M=1,K=4
+	ldc1	b4,8(B)
+	ldc1	a4,8(A)
+	ldc1	b2,16(B)
+	ldc1	a2,16(A)
+	ldc1	b6,24(B)
+	ldc1	a6,24(A)
+	
+	madd.d	t11,t11,a0,b0
+	madd.d	t11,t11,a4,b4
+	daddu	B,B,32				#  B+=1(nr)*4(kr)*8Byte=32
+	daddu	A,A,32				#  A+=1(mr)*4(kr)*8Byte=32
+	madd.d	t11,t11,a2,b2
+	madd.d	t11,t11,a6,b6
+	daddiu	K,K,-1
+	
+	ldc1	b0,0(B)
+	bnez 	K,.L91
+	ldc1	a0,0(A)
+
+
+.L95:							#  N=2 M=4 K=2
+	and		K,KCO,2				#  k = KCO&2
+	beqz	K,.L98
+	nop
+
+.L96:			
+	ldc1	b4,8(B)
+	ldc1	a4,8(A)
+	
+	madd.d	t11,t11,a0,b0
+	madd.d	t11,t11,a4,b4
+	daddu	B,B,16				#  B+=1(nr)*2(kr)*8Byte=16
+	daddu	A,A,16				#  A+=1(mr)*2(kr)*8Byte=32
+
+	ldc1	b0,0(B)
+	ldc1	a0,0(A)
+
+	
+.L98:							#  N=2, M=4, K=1
+	and		K,KCO,1
+	beqz	K,.L99				#  
+	ldc1	ALPHA,152($sp)		#  Get ALPHA
+	madd.d	t11,t11,a0,b0
+
+
+.L99:							#  Write Back
+	ldc1	c11,0(CO1)			#  Fetch 16 C
+	madd.d	t11,c11,t11,ALPHA
+	sdc1	t11,0(CO1)
+
+
+
+
+.L999:							#  End
+	ld	$16,   0($sp)
+	ld	$17,   8($sp)
+	ld	$18,  16($sp)
+	ld	$19,  24($sp)
+	ld	$20,  32($sp)
+	ld	$21,  40($sp)
+	ld	$22,  48($sp)
+	ldc1	$f24, 56($sp)
+	ldc1	$f25, 64($sp)
+	ldc1	$f26, 72($sp)
+	ldc1	$f27, 80($sp)
+	ldc1	$f28, 88($sp)
+	ld	$23,  96($sp)
+	ld	$24, 104($sp)
+	ld	$25, 112($sp)
+	ldc1	$f20,120($sp)
+	ldc1	$f21,128($sp)
+	ldc1	$f22,136($sp)
+	ldc1	$f23,144($sp)
+
+	j	$31
+	daddiu	$sp, $sp, 160
+
+	EPILOGUE