From 29dce62b8f5299201dbe539ccff1b1caf99ad715 Mon Sep 17 00:00:00 2001
From: traz <wangqian10@iscas.ac.cn>
Date: Wed, 11 May 2011 10:44:23 +0000
Subject: [PATCH] Finish dtrsm_kernel_Rx.S on Loongson3A.

---
 kernel/mips64/trsm_kernel_RN_loongson3a.S | 1852 +++++++++++++++++++
 kernel/mips64/trsm_kernel_RT_loongson3a.S | 1958 +++++++++++++++++++++
 2 files changed, 3810 insertions(+)
 create mode 100644 kernel/mips64/trsm_kernel_RN_loongson3a.S
 create mode 100644 kernel/mips64/trsm_kernel_RT_loongson3a.S

diff --git a/kernel/mips64/trsm_kernel_RN_loongson3a.S b/kernel/mips64/trsm_kernel_RN_loongson3a.S
new file mode 100644
index 000000000..790d7c981
--- /dev/null
+++ b/kernel/mips64/trsm_kernel_RN_loongson3a.S
@@ -0,0 +1,1852 @@
+#define REALNAME ASMNAME
+
+#define ASSEMBLER
+#include "common.h"
+
+
+#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 OFFSET	$22
+#define KK	$23
+#define TEMP	$24
+#define AORIG	$25
+
+#define a1	$f0
+#define a2	$f1
+#define a3	$f26
+#define a4	$f27
+
+#define a5	$f28
+#define a6	$f29
+#define	a7	$f30
+#define	a8	$f31
+
+#define b1	$f2
+#define b2	$f3
+#define b3	$f4
+#define b4	$f5
+
+#define b5	$f6
+#define b6	$f7
+#define b7	$f8
+#define b8	$f9
+
+#define t11	$f10
+#define t21	$f11
+#define t31	$f12
+#define	t41	$f13
+
+#define t12	$f14
+#define	t22	$f15
+#define t32	$f16
+#define	t42	$f17
+
+#define	t13	$f18
+#define	t23	$f19
+#define	t33	$f20
+#define	t43	$f21
+
+#define	t14	$f22
+#define	t24	$f23
+#define	t34	$f24
+#define t44	$f25
+
+	PROLOGUE
+	
+	daddiu	$sp, $sp, -144
+
+	SDARG	$16,   0($sp)
+	SDARG	$17,   8($sp)
+	SDARG	$18,  16($sp)
+	SDARG	$19,  24($sp)
+	SDARG	$20,  32($sp)
+	SDARG	$21,  40($sp)
+	sdc1	$f24, 48($sp)
+	sdc1	$f25, 56($sp)
+	sdc1	$f26, 64($sp)
+	sdc1	$f27, 72($sp)
+	sdc1	$f28, 80($sp)
+
+	SDARG	$22,  88($sp)
+	SDARG	$23,  96($sp)
+	SDARG	$24, 104($sp)
+	SDARG	$25, 112($sp)
+
+#ifndef __64BIT__
+	sdc1	$f20,112($sp)
+	sdc1	$f21,120($sp)
+	sdc1	$f22,128($sp)
+	sdc1	$f23,136($sp)
+#endif
+
+											#	RN compute from top to bottom left to right
+	.align	3
+	LDARG	OFFSET, 144($sp)				#	get the last parameter
+	dsll	LDC, LDC, BASE_SHIFT			#	LDC * data_Byte
+
+	neg	KK, OFFSET							#	for RN OFFSET always 0
+
+	dsra	J,  N, 2						#	J = NC/4 
+	blez	J, .L30
+	NOP
+
+.L10:
+	daddiu	J, J, -1
+	
+	move	CO1, C
+	daddu	CO2, C,   LDC
+	daddu	CO3, CO2, LDC
+	daddu	CO4, CO3, LDC
+	
+	move	AO, A							#	A is the retangular matrix and B is the trigular matrix	
+	daddu	C,  CO4, LDC					#	Fixed pointer C
+
+	dsra	I,  M, 2						#	I=MC/4
+	blez	I, .L20
+	NOP
+
+	.align 3
+.L11:
+	MTC	$0,  t11							#	clear results registers
+	MOV	t21, t11
+	MOV	t31, t11
+	MOV	t41, t11
+
+	MOV	t12, t11
+	MOV	t22, t11
+	MOV	t32, t11
+	MOV	t42, t11
+
+	MOV	t13, t11
+	MOV	t23, t11
+	MOV	t33, t11
+	MOV	t43, t11
+	
+	MOV	t14, t11
+	MOV	t24, t11
+	MOV	t34, t11
+	MOV	t44, t11
+	
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+	LD	a2,  1 * SIZE(AO)					#	get 4 a	
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+	LD	b2,  1 * SIZE(B)					#	get 4 b
+	LD	b3,  2 * SIZE(B)
+	LD	b4,  3 * SIZE(B)
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L15
+	move	BO,  B							#	reset B
+
+.L12:
+	LD	a5,  4 * SIZE(AO)
+	LD	a6,  5 * SIZE(AO)
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b5,  4 * SIZE(BO)
+	LD	b6,  5 * SIZE(BO)
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+	MADD	t33, t33, a3, b3
+	MADD	t43, t43, a4, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+	MADD	t34, t34, a3, b4
+	MADD	t44, t44, a4, b4			#	fisrt 	
+
+	LD	a1,  8 * SIZE(AO)
+	LD	a2,  9 * SIZE(AO)
+	LD	a3,  10 * SIZE(AO)
+	LD	a4,  11 * SIZE(AO)
+
+	LD	b1,  8 * SIZE(BO)
+	LD	b2,  9 * SIZE(BO)
+	LD	b3,  10 * SIZE(BO)
+	LD	b4,  11 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t31, t31, a7, b5
+	MADD	t41, t41, a8, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+	MADD	t32, t32, a7, b6
+	MADD	t42, t42, a8, b6
+
+	MADD	t13, t13, a5, b7
+	MADD	t23, t23, a6, b7
+	MADD	t33, t33, a7, b7
+	MADD	t43, t43, a8, b7
+
+	MADD	t14, t14, a5, b8
+	MADD	t24, t24, a6, b8
+	MADD	t34, t34, a7, b8
+	MADD	t44, t44, a8, b8			#	second
+
+	LD	a5,  12 * SIZE(AO)
+	LD	a6,  13 * SIZE(AO)
+	LD	a7,  14 * SIZE(AO)
+	LD	a8,  15 * SIZE(AO)
+
+	LD	b5,  12 * SIZE(BO)
+	LD	b6,  13 * SIZE(BO)
+	LD	b7,  14 * SIZE(BO)
+	LD	b8,  15 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+	MADD	t33, t33, a3, b3
+	MADD	t43, t43, a4, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+	MADD	t34, t34, a3, b4
+	MADD	t44, t44, a4, b4			#	third
+
+	daddiu	AO, AO, 16 * SIZE			#	AO += 4mr*4kr	
+	daddiu	BO, BO, 16 * SIZE			#	BP += 4nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t31, t31, a7, b5
+	MADD	t41, t41, a8, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+	MADD	t32, t32, a7, b6
+	MADD	t42, t42, a8, b6
+
+	MADD	t13, t13, a5, b7
+	MADD	t23, t23, a6, b7
+	MADD	t33, t33, a7, b7
+	MADD	t43, t43, a8, b7
+
+	MADD	t14, t14, a5, b8
+	MADD	t24, t24, a6, b8
+	MADD	t34, t34, a7, b8
+	MADD	t44, t44, a8, b8			#	fouth
+
+	daddiu	L, L, -1
+	bgtz	L, .L12
+	NOP
+
+
+.L15:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L18
+	NOP
+
+	.align	3
+.L16:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+	MADD	t33, t33, a3, b3
+	MADD	t43, t43, a4, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+	MADD	t34, t34, a3, b4
+	MADD	t44, t44, a4, b4		
+
+	daddiu	AO, AO, 4 * SIZE			#	AO += 4mr	
+	daddiu	BO, BO, 4 * SIZE			#	BP += 4nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L16
+	NOP
+
+
+	.align 3
+.L18:									#	.L18 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+	LD	b2,  1 * SIZE(AO)				#	Fixed results
+	LD	b3,  2 * SIZE(AO)				
+	LD	b4,  3 * SIZE(AO)				#	sa stored as col major
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+	SUB	t31, b3, t31
+	SUB	t41, b4, t41
+
+ 	LD	b5,  4 * SIZE(AO)
+	LD	b6,  5 * SIZE(AO)
+	LD	b7,  6 * SIZE(AO)
+	LD	b8,  7 * SIZE(AO)
+
+	SUB	t12, b5, t12
+	SUB	t22, b6, t22
+	SUB	t32, b7, t32
+	SUB	t42, b8, t42
+
+ 	LD	b1,  8 * SIZE(AO)
+	LD	b2,  9 * SIZE(AO)
+	LD	b3, 10 * SIZE(AO)
+	LD	b4, 11 * SIZE(AO)
+
+	SUB	t13, b1, t13
+	SUB	t23, b2, t23
+	SUB	t33, b3, t33
+	SUB	t43, b4, t43
+
+ 	LD	b5, 12 * SIZE(AO)
+	LD	b6, 13 * SIZE(AO)
+	LD	b7, 14 * SIZE(AO)
+	LD	b8, 15 * SIZE(AO)
+
+	SUB	t14, b5, t14
+	SUB	t24, b6, t24
+	SUB	t34, b7, t34
+	SUB	t44, b8, t44
+
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+	MUL	t11, b1, t11
+	MUL	t21, b1, t21
+	MUL	t31, b1, t31
+	MUL	t41, b1, t41
+	NMSUB	t12, t12, b2, t11
+	NMSUB	t22, t22, b2, t21
+	NMSUB	t32, t32, b2, t31
+	NMSUB	t42, t42, b2, t41
+	NMSUB	t13, t13, b3, t11
+	NMSUB	t23, t23, b3, t21
+	NMSUB	t33, t33, b3, t31
+	NMSUB	t43, t43, b3, t41
+	NMSUB	t14, t14, b4, t11
+	NMSUB	t24, t24, b4, t21
+	NMSUB	t34, t34, b4, t31
+	NMSUB	t44, t44, b4, t41
+
+	
+	LD	b5,  5 * SIZE(BO)
+	LD	b6,  6 * SIZE(BO)
+	LD	b7,  7 * SIZE(BO)
+	MUL	t12, b5, t12
+	MUL	t22, b5, t22
+	MUL	t32, b5, t32
+	MUL	t42, b5, t42
+	NMSUB	t13, t13, b6, t12
+	NMSUB	t23, t23, b6, t22
+	NMSUB	t33, t33, b6, t32
+	NMSUB	t43, t43, b6, t42
+	NMSUB	t14, t14, b7, t12
+	NMSUB	t24, t24, b7, t22
+	NMSUB	t34, t34, b7, t32
+	NMSUB	t44, t44, b7, t42
+
+
+
+	LD	b8,  10 * SIZE(BO)
+	LD	b1,  11 * SIZE(BO)
+	MUL	t13, b8, t13
+	MUL	t23, b8, t23
+	MUL	t33, b8, t33
+	MUL	t43, b8, t43
+	NMSUB	t14, t14, b1, t13
+	NMSUB	t24, t24, b1, t23
+	NMSUB	t34, t34, b1, t33
+	NMSUB	t44, t44, b1, t43
+
+
+
+	LD	b2,  15 * SIZE(BO)
+	MUL	t14, b2, t14
+	MUL	t24, b2, t24
+	MUL	t34, b2, t34
+	MUL	t44, b2, t44
+
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+	ST	t21,  1 * SIZE(AO)
+	ST	t31,  2 * SIZE(AO)
+	ST	t41,  3 * SIZE(AO)
+
+	ST	t12,  4 * SIZE(AO)
+	ST	t22,  5 * SIZE(AO)
+	ST	t32,  6 * SIZE(AO)
+	ST	t42,  7 * SIZE(AO)
+
+	ST	t13,  8 * SIZE(AO)
+	ST	t23,  9 * SIZE(AO)
+	ST	t33, 10 * SIZE(AO)
+	ST	t43, 11 * SIZE(AO)
+
+	ST	t14, 12 * SIZE(AO)
+	ST	t24, 13 * SIZE(AO)
+	ST	t34, 14 * SIZE(AO)
+	ST	t44, 15 * SIZE(AO)
+
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	ST	t21,  1 * SIZE(CO1)
+	ST	t31,  2 * SIZE(CO1)
+	ST	t41,  3 * SIZE(CO1)
+	
+	ST	t12,  0 * SIZE(CO2)
+	ST	t22,  1 * SIZE(CO2)
+	ST	t32,  2 * SIZE(CO2)
+	ST	t42,  3 * SIZE(CO2)
+	
+	ST	t13,  0 * SIZE(CO3)
+	ST	t23,  1 * SIZE(CO3)
+	ST	t33,  2 * SIZE(CO3)
+	ST	t43,  3 * SIZE(CO3)
+	
+	ST	t14,  0 * SIZE(CO4)
+	ST	t24,  1 * SIZE(CO4)
+	ST	t34,  2 * SIZE(CO4)
+	ST	t44,  3 * SIZE(CO4)
+
+	daddiu	CO1, CO1, 4 * SIZE			#	fixed address
+	daddiu	CO2, CO2, 4 * SIZE
+	daddiu	CO3, CO3, 4 * SIZE
+	daddiu	CO4, CO4, 4 * SIZE
+
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, 2 + BASE_SHIFT
+	dsll	TEMP, TEMP, 2 + BASE_SHIFT
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+	daddiu	I, I, -1
+	bgtz	I, .L11
+	NOP
+
+	.align 3
+.L20:									
+	andi	I,  M, 2					#	mr=2
+	blez	I, .L50
+	nop
+
+	MTC	$0,  t11							#	clear results registers
+	MOV	t21, t11
+	MOV	t31, t11
+	MOV	t41, t11
+
+	MOV	t12, t11
+	MOV	t22, t11
+	MOV	t32, t11
+	MOV	t42, t11
+
+	MOV	t13, t11
+	MOV	t23, t11
+	MOV	t33, t11
+	MOV	t43, t11
+	
+	MOV	t14, t11
+	MOV	t24, t11
+	MOV	t34, t11
+	MOV	t44, t11
+	
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+	LD	a2,  1 * SIZE(AO)					#	get 4 a	
+
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+	LD	b2,  1 * SIZE(B)					#	get 4 b
+	LD	b3,  2 * SIZE(B)
+	LD	b4,  3 * SIZE(B)
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L25
+	move	BO,  B							#	reset B
+
+.L22:
+	LD	a5,  2 * SIZE(AO)
+	LD	a6,  3 * SIZE(AO)
+
+	LD	b5,  4 * SIZE(BO)
+	LD	b6,  5 * SIZE(BO)
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+
+	LD	a3,  4 * SIZE(AO)
+	LD	a4,  5 * SIZE(AO)
+
+	LD	b1,  8 * SIZE(BO)
+	LD	b2,  9 * SIZE(BO)
+	LD	b3,  10 * SIZE(BO)
+	LD	b4,  11 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+
+	MADD	t13, t13, a5, b7
+	MADD	t23, t23, a6, b7
+
+	MADD	t14, t14, a5, b8
+	MADD	t24, t24, a6, b8
+
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b5,  12 * SIZE(BO)
+	LD	b6,  13 * SIZE(BO)
+	LD	b7,  14 * SIZE(BO)
+	LD	b8,  15 * SIZE(BO)
+
+	MADD	t11, t11, a3, b1
+	MADD	t21, t21, a4, b1
+
+	MADD	t12, t12, a3, b2
+	MADD	t22, t22, a4, b2
+
+	MADD	t13, t13, a3, b3
+	MADD	t23, t23, a4, b3
+
+	MADD	t14, t14, a3, b4
+	MADD	t24, t24, a4, b4
+
+	daddiu	AO, AO, 8 * SIZE			#	AO += 2mr*4kr	
+	daddiu	BO, BO, 16 * SIZE			#	BP += 4nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	MADD	t11, t11, a7, b5
+	MADD	t21, t21, a8, b5
+
+	MADD	t12, t12, a7, b6
+	MADD	t22, t22, a8, b6
+
+	MADD	t13, t13, a7, b7
+	MADD	t23, t23, a8, b7
+
+	MADD	t14, t14, a7, b8
+	MADD	t24, t24, a8, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L22
+	NOP
+
+
+.L25:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L28
+	NOP
+
+	.align	3
+.L26:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+
+	daddiu	AO, AO, 2 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 4 * SIZE			#	BP += 4nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L26
+	NOP
+
+
+	.align 3
+.L28:									#	.L18 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+	LD	b2,  1 * SIZE(AO)				#	Fixed results
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+
+ 	LD	b5,  2 * SIZE(AO)
+	LD	b6,  3 * SIZE(AO)
+
+	SUB	t12, b5, t12
+	SUB	t22, b6, t22
+
+	LD	b3,  4 * SIZE(AO)
+	LD	b4,  5 * SIZE(AO)
+
+	SUB	t13, b3, t13
+	SUB	t23, b4, t23
+
+	LD	b7,  6 * SIZE(AO)
+	LD	b8,  7 * SIZE(AO)
+
+	SUB	t14, b7, t14
+	SUB	t24, b8, t24
+
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+	MUL	t11, b1, t11
+	MUL	t21, b1, t21
+	NMSUB	t12, t12, b2, t11
+	NMSUB	t22, t22, b2, t21
+	NMSUB	t13, t13, b3, t11
+	NMSUB	t23, t23, b3, t21
+	NMSUB	t14, t14, b4, t11
+	NMSUB	t24, t24, b4, t21
+
+	
+	LD	b5,  5 * SIZE(BO)
+	LD	b6,  6 * SIZE(BO)
+	LD	b7,  7 * SIZE(BO)
+	MUL	t12, b5, t12
+	MUL	t22, b5, t22
+	NMSUB	t13, t13, b6, t12
+	NMSUB	t23, t23, b6, t22
+	NMSUB	t14, t14, b7, t12
+	NMSUB	t24, t24, b7, t22
+
+
+
+	LD	b8,  10 * SIZE(BO)
+	LD	b1,  11 * SIZE(BO)
+	MUL	t13, b8, t13
+	MUL	t23, b8, t23
+	NMSUB	t14, t14, b1, t13
+	NMSUB	t24, t24, b1, t23
+
+
+
+	LD	b2,  15 * SIZE(BO)
+	MUL	t14, b2, t14
+	MUL	t24, b2, t24
+
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+	ST	t21,  1 * SIZE(AO)
+
+	ST	t12,  2 * SIZE(AO)
+	ST	t22,  3 * SIZE(AO)
+
+	ST	t13,  4 * SIZE(AO)
+	ST	t23,  5 * SIZE(AO)
+
+	ST	t14,  6 * SIZE(AO)
+	ST	t24,  7 * SIZE(AO)
+
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	ST	t21,  1 * SIZE(CO1)
+	
+	ST	t12,  0 * SIZE(CO2)
+	ST	t22,  1 * SIZE(CO2)
+	
+	ST	t13,  0 * SIZE(CO3)
+	ST	t23,  1 * SIZE(CO3)
+	
+	ST	t14,  0 * SIZE(CO4)
+	ST	t24,  1 * SIZE(CO4)
+
+	daddiu	CO1, CO1, 2 * SIZE			#	fixed address
+	daddiu	CO2, CO2, 2 * SIZE			#	mr=2
+	daddiu	CO3, CO3, 2 * SIZE
+	daddiu	CO4, CO4, 2 * SIZE
+
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, 1 + BASE_SHIFT	#	mr=2
+	dsll	TEMP, TEMP, 2 + BASE_SHIFT
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+	.align 3
+.L50:									
+	andi	I,  M, 1					#	mr=1
+	blez	I, .L29
+	nop
+
+	MTC	$0,  t11							#	clear results registers
+	MOV	t21, t11
+	MOV	t31, t11
+	MOV	t41, t11
+
+	MOV	t12, t11
+	MOV	t22, t11
+	MOV	t32, t11
+	MOV	t42, t11
+
+	MOV	t13, t11
+	MOV	t23, t11
+	MOV	t33, t11
+	MOV	t43, t11
+	
+	MOV	t14, t11
+	MOV	t24, t11
+	MOV	t34, t11
+	MOV	t44, t11
+	
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+	LD	b2,  1 * SIZE(B)					#	get 4 b
+	LD	b3,  2 * SIZE(B)
+	LD	b4,  3 * SIZE(B)
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L55
+	move	BO,  B							#	reset B
+
+.L52:
+	LD	a5,  1 * SIZE(AO)
+
+	LD	b5,  4 * SIZE(BO)
+	LD	b6,  5 * SIZE(BO)
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t12, t12, a1, b2
+	MADD	t13, t13, a1, b3
+	MADD	t14, t14, a1, b4
+
+	LD	a3,  2 * SIZE(AO)
+
+	LD	b1,  8 * SIZE(BO)
+	LD	b2,  9 * SIZE(BO)
+	LD	b3,  10 * SIZE(BO)
+	LD	b4,  11 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t12, t12, a5, b6
+	MADD	t13, t13, a5, b7
+	MADD	t14, t14, a5, b8
+
+	LD	a7,  3 * SIZE(AO)
+
+	LD	b5,  12 * SIZE(BO)
+	LD	b6,  13 * SIZE(BO)
+	LD	b7,  14 * SIZE(BO)
+	LD	b8,  15 * SIZE(BO)
+
+	MADD	t11, t11, a3, b1
+	MADD	t12, t12, a3, b2
+	MADD	t13, t13, a3, b3
+	MADD	t14, t14, a3, b4
+
+	daddiu	AO, AO, 4 * SIZE			#	AO += 1mr*4kr	
+	daddiu	BO, BO, 16 * SIZE			#	BP += 4nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	MADD	t11, t11, a7, b5
+	MADD	t12, t12, a7, b6
+	MADD	t13, t13, a7, b7
+	MADD	t14, t14, a7, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L52
+	NOP
+
+
+.L55:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L58
+	NOP
+
+	.align	3
+.L56:
+	MADD	t11, t11, a1, b1
+	MADD	t12, t12, a1, b2
+	MADD	t13, t13, a1, b3
+	MADD	t14, t14, a1, b4
+
+	daddiu	AO, AO, 1 * SIZE			#	AO += 1mr	
+	daddiu	BO, BO, 4 * SIZE			#	BP += 4nr
+	
+	LD	a1,  0 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L56
+	NOP
+
+
+	.align 3
+.L58:									#	.L18 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+ 	LD	b5,  1 * SIZE(AO)
+	LD	b3,  2 * SIZE(AO)
+	LD	b7,  3 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t12, b5, t12
+	SUB	t13, b3, t13
+	SUB	t14, b7, t14
+
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+	MUL	t11, b1, t11
+	NMSUB	t12, t12, b2, t11
+	NMSUB	t13, t13, b3, t11
+	NMSUB	t14, t14, b4, t11
+
+	
+	LD	b5,  5 * SIZE(BO)
+	LD	b6,  6 * SIZE(BO)
+	LD	b7,  7 * SIZE(BO)
+	MUL	t12, b5, t12
+	NMSUB	t13, t13, b6, t12
+	NMSUB	t14, t14, b7, t12
+
+
+	LD	b8,  10 * SIZE(BO)
+	LD	b1,  11 * SIZE(BO)
+	MUL	t13, b8, t13
+	NMSUB	t14, t14, b1, t13
+
+
+	LD	b2,  15 * SIZE(BO)
+	MUL	t14, b2, t14
+
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+	ST	t12,  1 * SIZE(AO)
+	ST	t13,  2 * SIZE(AO)
+	ST	t14,  3 * SIZE(AO)
+
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	ST	t12,  0 * SIZE(CO2)
+	ST	t13,  0 * SIZE(CO3)
+	ST	t14,  0 * SIZE(CO4)
+
+	daddiu	CO1, CO1, 1 * SIZE			#	fixed address
+	daddiu	CO2, CO2, 1 * SIZE			#	mr=2
+	daddiu	CO3, CO3, 1 * SIZE
+	daddiu	CO4, CO4, 1 * SIZE
+
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, BASE_SHIFT	#	mr=2
+	dsll	TEMP, TEMP, 2 + BASE_SHIFT
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+
+	.align 3
+.L29:
+	move	B,  BO						#	change to next panel of Bj
+	daddiu	KK, KK,  4					#	rectangular data length increase by 4	
+	bgtz	J, .L10
+	NOP
+
+
+	.align 3
+	
+.L30:
+	andi	J,  N, 2
+	blez	J, .L70
+	nop
+
+	move	CO1, C
+	daddu	CO2, C,   LDC
+	
+	move	AO, A							#	A is the retangular matrix and B is the trigular matrix	
+	daddu	C,  CO2, LDC					#	Fixed pointer C
+
+	dsra	I,  M, 2						#	I=MC/4
+	blez	I, .L40
+	NOP
+
+	.align 3
+.L31:
+	MTC	$0,  t11							#	clear results registers
+	MOV	t21, t11
+	MOV	t31, t11
+	MOV	t41, t11
+
+	MOV	t12, t11
+	MOV	t22, t11
+	MOV	t32, t11
+	MOV	t42, t11
+
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+	LD	a2,  1 * SIZE(AO)					#	get 4 a	
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+	LD	b2,  1 * SIZE(B)					#	get 4 b
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L35
+	move	BO,  B							#	reset B
+
+.L32:
+	LD	a5,  4 * SIZE(AO)
+	LD	a6,  5 * SIZE(AO)
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b5,  2 * SIZE(BO)
+	LD	b6,  3 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	LD	a1,  8 * SIZE(AO)
+	LD	a2,  9 * SIZE(AO)
+	LD	a3,  10 * SIZE(AO)
+	LD	a4,  11 * SIZE(AO)
+
+	LD	b3,  4 * SIZE(BO)
+	LD	b4,  5 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t31, t31, a7, b5
+	MADD	t41, t41, a8, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+	MADD	t32, t32, a7, b6
+	MADD	t42, t42, a8, b6
+
+	LD	a5,  12 * SIZE(AO)
+	LD	a6,  13 * SIZE(AO)
+	LD	a7,  14 * SIZE(AO)
+	LD	a8,  15 * SIZE(AO)
+
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a1, b3
+	MADD	t21, t21, a2, b3
+	MADD	t31, t31, a3, b3
+	MADD	t41, t41, a4, b3
+
+	MADD	t12, t12, a1, b4
+	MADD	t22, t22, a2, b4
+	MADD	t32, t32, a3, b4
+	MADD	t42, t42, a4, b4
+
+	daddiu	AO, AO, 16 * SIZE			#	AO += 4mr*4kr	
+	daddiu	BO, BO,  8 * SIZE			#	BP += 2nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	MADD	t11, t11, a5, b7
+	MADD	t21, t21, a6, b7
+	MADD	t31, t31, a7, b7
+	MADD	t41, t41, a8, b7
+
+	MADD	t12, t12, a5, b8
+	MADD	t22, t22, a6, b8
+	MADD	t32, t32, a7, b8
+	MADD	t42, t42, a8, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L32
+	NOP
+
+
+.L35:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L38
+	NOP
+
+	.align	3
+.L36:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	daddiu	AO, AO, 4 * SIZE			#	AO += 4mr	
+	daddiu	BO, BO, 2 * SIZE			#	BP += 2nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L36
+	NOP
+
+
+	.align 3
+.L38:									#	.L38 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+	LD	b2,  1 * SIZE(AO)				#	Fixed results
+	LD	b3,  2 * SIZE(AO)				
+	LD	b4,  3 * SIZE(AO)				#	sa stored as col major
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+	SUB	t31, b3, t31
+	SUB	t41, b4, t41
+
+ 	LD	b5,  4 * SIZE(AO)
+	LD	b6,  5 * SIZE(AO)
+	LD	b7,  6 * SIZE(AO)
+	LD	b8,  7 * SIZE(AO)
+
+	SUB	t12, b5, t12
+	SUB	t22, b6, t22
+	SUB	t32, b7, t32
+	SUB	t42, b8, t42
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	LD	b2,  1 * SIZE(BO)
+	MUL	t11, b1, t11
+	MUL	t21, b1, t21
+	MUL	t31, b1, t31
+	MUL	t41, b1, t41
+	NMSUB	t12, t12, b2, t11
+	NMSUB	t22, t22, b2, t21
+	NMSUB	t32, t32, b2, t31
+	NMSUB	t42, t42, b2, t41
+	
+	LD	b5,  3 * SIZE(BO)
+	MUL	t12, b5, t12
+	MUL	t22, b5, t22
+	MUL	t32, b5, t32
+	MUL	t42, b5, t42
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+	ST	t21,  1 * SIZE(AO)
+	ST	t31,  2 * SIZE(AO)
+	ST	t41,  3 * SIZE(AO)
+
+	ST	t12,  4 * SIZE(AO)
+	ST	t22,  5 * SIZE(AO)
+	ST	t32,  6 * SIZE(AO)
+	ST	t42,  7 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	ST	t21,  1 * SIZE(CO1)
+	ST	t31,  2 * SIZE(CO1)
+	ST	t41,  3 * SIZE(CO1)
+	
+	ST	t12,  0 * SIZE(CO2)
+	ST	t22,  1 * SIZE(CO2)
+	ST	t32,  2 * SIZE(CO2)
+	ST	t42,  3 * SIZE(CO2)
+
+	daddiu	CO1, CO1, 4 * SIZE			#	fixed address
+	daddiu	CO2, CO2, 4 * SIZE
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, 2 + BASE_SHIFT
+	dsll	TEMP, TEMP, 1 + BASE_SHIFT	#	nr=2
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+	daddiu	I, I, -1
+	bgtz	I, .L31
+	NOP
+
+	.align 3
+.L40:
+	andi	I, M,2
+	blez	I,.L60
+	nop
+
+	MTC	$0,  t11							#	clear results registers
+	MOV	t21, t11
+
+	MOV	t12, t11
+	MOV	t22, t11
+
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+	LD	a2,  1 * SIZE(AO)					#	get 4 a	
+
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+	LD	b2,  1 * SIZE(B)					#	get 4 b
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L45
+	move	BO,  B							#	reset B
+
+.L42:
+	LD	a5,  2 * SIZE(AO)
+	LD	a6,  3 * SIZE(AO)
+	LD	b5,  2 * SIZE(BO)
+	LD	b6,  3 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+
+	LD	a3,  4 * SIZE(AO)
+	LD	a4,  5 * SIZE(AO)
+	LD	b3,  4 * SIZE(BO)
+	LD	b4,  5 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a3, b3
+	MADD	t21, t21, a4, b3
+	MADD	t12, t12, a3, b4
+	MADD	t22, t22, a4, b4
+
+	daddiu	AO, AO,  8 * SIZE			#	AO += 2mr*4kr	
+	daddiu	BO, BO,  8 * SIZE			#	BP += 2nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	MADD	t11, t11, a7, b7
+	MADD	t21, t21, a8, b7
+	MADD	t12, t12, a7, b8
+	MADD	t22, t22, a8, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L42
+	NOP
+
+
+.L45:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L48
+	NOP
+
+	.align	3
+.L46:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+
+	daddiu	AO, AO, 2 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 2 * SIZE			#	BP += 2nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L46
+	NOP
+
+
+	.align 3
+.L48:									#	.L48 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+	LD	b2,  1 * SIZE(AO)				#	Fixed results
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+
+ 	LD	b5,  2 * SIZE(AO)
+	LD	b6,  3 * SIZE(AO)
+
+	SUB	t12, b5, t12
+	SUB	t22, b6, t22
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	LD	b2,  1 * SIZE(BO)
+	MUL	t11, b1, t11
+	MUL	t21, b1, t21
+	NMSUB	t12, t12, b2, t11
+	NMSUB	t22, t22, b2, t21
+	
+	LD	b5,  3 * SIZE(BO)
+	MUL	t12, b5, t12
+	MUL	t22, b5, t22
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+	ST	t21,  1 * SIZE(AO)
+	ST	t12,  2 * SIZE(AO)
+	ST	t22,  3 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	ST	t21,  1 * SIZE(CO1)
+	ST	t12,  0 * SIZE(CO2)
+	ST	t22,  1 * SIZE(CO2)
+
+	daddiu	CO1, CO1, 2 * SIZE			#	fixed address
+	daddiu	CO2, CO2, 2 * SIZE
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, 1 + BASE_SHIFT
+	dsll	TEMP, TEMP, 1 + BASE_SHIFT	#	nr=2
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+
+	.align 3
+.L60:
+	andi	I,M,1						#	nr=2 mr=1
+	blez	I,.L39
+	nop
+
+	MTC	$0,  t11							#	clear results registers
+	MOV	t12, t11
+
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+	LD	b2,  1 * SIZE(B)					#	get 4 b
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L65
+	move	BO,  B							#	reset B
+
+.L62:
+	LD	a5,  1 * SIZE(AO)
+	LD	b5,  2 * SIZE(BO)
+	LD	b6,  3 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t12, t12, a1, b2
+
+	LD	a3,  2 * SIZE(AO)
+	LD	b3,  4 * SIZE(BO)
+	LD	b4,  5 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t12, t12, a5, b6
+
+	LD	a7,  3 * SIZE(AO)
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a3, b3
+	MADD	t12, t12, a3, b4
+
+	daddiu	AO, AO,  4 * SIZE			#	AO += 1mr*4kr	
+	daddiu	BO, BO,  8 * SIZE			#	BP += 2nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	MADD	t11, t11, a7, b7
+	MADD	t12, t12, a7, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L62
+	NOP
+
+
+.L65:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L68
+	NOP
+
+	.align	3
+.L66:
+	MADD	t11, t11, a1, b1
+	MADD	t12, t12, a1, b2
+
+	daddiu	AO, AO, 1 * SIZE			#	AO += mr	
+	daddiu	BO, BO, 2 * SIZE			#	BP += 2nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L66
+	NOP
+
+
+	.align 3
+.L68:									#	.L48 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+	LD	b5,  1 * SIZE(AO)				#	Fixed results
+
+	SUB	t11, b1, t11
+	SUB	t12, b5, t12
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	LD	b2,  1 * SIZE(BO)
+	MUL	t11, b1, t11
+	NMSUB	t12, t12, b2, t11
+	
+	LD	b5,  3 * SIZE(BO)
+	MUL	t12, b5, t12
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+	ST	t12,  1 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	ST	t12,  0 * SIZE(CO2)
+
+	daddiu	CO1, CO1, 1 * SIZE			#	fixed address
+	daddiu	CO2, CO2, 1 * SIZE
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, BASE_SHIFT		#	mr=1
+	dsll	TEMP, TEMP, 1 + BASE_SHIFT	#	nr=2
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+
+	.align 3
+.L39:
+	move	B,  BO						#	change to next panel of Bj
+	daddiu	KK, KK,  2					#	rectangular data length increase by 4	
+
+
+
+	.align 3
+
+.L70:
+	andi	J,  N, 1					#	nr=1
+	blez	J, .L999
+	NOP
+
+	move	CO1, C
+	move	AO, A
+
+	daddu	C,  CO1, LDC
+
+	dsra	I,  M, 2					#	I=MC/4
+	blez	I, .L80
+	NOP
+
+	.align 3
+.L71:
+	MTC	$0,  t11							#	clear results registers
+	MOV	t21, t11
+	MOV	t31, t11
+	MOV	t41, t11
+
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+	LD	a2,  1 * SIZE(AO)					#	get 4 a	
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L75
+	move	BO,  B							#	reset B
+
+.L72:
+	LD	a5,  4 * SIZE(AO)
+	LD	a6,  5 * SIZE(AO)
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b5,  1 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	LD	a1,  8 * SIZE(AO)
+	LD	a2,  9 * SIZE(AO)
+	LD	a3,  10 * SIZE(AO)
+	LD	a4,  11 * SIZE(AO)
+
+	LD	b3,  2 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t31, t31, a7, b5
+	MADD	t41, t41, a8, b5
+
+	LD	a5,  12 * SIZE(AO)
+	LD	a6,  13 * SIZE(AO)
+	LD	a7,  14 * SIZE(AO)
+	LD	a8,  15 * SIZE(AO)
+
+	LD	b7,  3 * SIZE(BO)
+
+	MADD	t11, t11, a1, b3
+	MADD	t21, t21, a2, b3
+	MADD	t31, t31, a3, b3
+	MADD	t41, t41, a4, b3
+
+	daddiu	AO, AO, 16 * SIZE			#	AO += 4mr*4kr	
+	daddiu	BO, BO,  4 * SIZE			#	BP += 1nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	MADD	t11, t11, a5, b7
+	MADD	t21, t21, a6, b7
+	MADD	t31, t31, a7, b7
+	MADD	t41, t41, a8, b7
+
+	daddiu	L, L, -1
+	bgtz	L, .L72
+	NOP
+
+
+.L75:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L78
+	NOP
+
+	.align	3
+.L76:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	daddiu	AO, AO, 4 * SIZE			#	AO += 4mr	
+	daddiu	BO, BO, 1 * SIZE			#	BP += 1nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L76
+	NOP
+
+
+	.align 3
+.L78:									#	.L78 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+	LD	b2,  1 * SIZE(AO)				#	Fixed results
+	LD	b3,  2 * SIZE(AO)				
+	LD	b4,  3 * SIZE(AO)				#	sa stored as col major
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+	SUB	t31, b3, t31
+	SUB	t41, b4, t41
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	MUL	t11, b1, t11
+	MUL	t21, b1, t21
+	MUL	t31, b1, t31
+	MUL	t41, b1, t41
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+	ST	t21,  1 * SIZE(AO)
+	ST	t31,  2 * SIZE(AO)
+	ST	t41,  3 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	ST	t21,  1 * SIZE(CO1)
+	ST	t31,  2 * SIZE(CO1)
+	ST	t41,  3 * SIZE(CO1)
+	
+
+	daddiu	CO1, CO1, 4 * SIZE			#	fixed address
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, 2 + BASE_SHIFT
+	dsll	TEMP, TEMP, BASE_SHIFT		#	nr=1
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+	daddiu	I, I, -1
+	bgtz	I, .L71
+	NOP
+
+
+	.align 3
+.L80:
+	andi	I, M, 2						#	mr=2
+	blez	I, .L90
+	nop
+
+	MTC	$0,  t11							#	clear results registers
+	MOV	t21, t11
+
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+	LD	a2,  1 * SIZE(AO)					#	get 4 a	
+
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L85
+	move	BO,  B							#	reset B
+
+.L82:
+	LD	a5,  2 * SIZE(AO)
+	LD	a6,  3 * SIZE(AO)
+
+	LD	b5,  1 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	LD	a3,  4 * SIZE(AO)
+	LD	a4,  5 * SIZE(AO)
+
+	LD	b3,  2 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b7,  3 * SIZE(BO)
+
+	MADD	t11, t11, a3, b3
+	MADD	t21, t21, a4, b3
+
+	daddiu	AO, AO,  8 * SIZE			#	AO += 2mr*4kr	
+	daddiu	BO, BO,  4 * SIZE			#	BP += 1nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	MADD	t11, t11, a7, b7
+	MADD	t21, t21, a8, b7
+
+	daddiu	L, L, -1
+	bgtz	L, .L82
+	NOP
+
+
+.L85:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L88
+	NOP
+
+	.align	3
+.L86:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	daddiu	AO, AO, 2 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 1 * SIZE			#	BP += 1nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L86
+	NOP
+
+
+	.align 3
+.L88:									#	.L88 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+	LD	b2,  1 * SIZE(AO)				#	Fixed results
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	MUL	t11, b1, t11
+	MUL	t21, b1, t21
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+	ST	t21,  1 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	ST	t21,  1 * SIZE(CO1)
+	
+
+	daddiu	CO1, CO1, 2 * SIZE			#	fixed address
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, 1 + BASE_SHIFT
+	dsll	TEMP, TEMP, BASE_SHIFT		#	nr=1
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+
+	.align 3
+.L90:
+	andi	I, M, 1						#	mr=1
+	blez	I, .L79
+	nop
+
+	MTC	$0,  t11							#	clear results registers
+
+	LD	a1,  0 * SIZE(AO)					#	AO point to the beginning address of sa
+	LD	b1,  0 * SIZE(B)					#	B point to the beginning address of every panel Bj
+
+	dsra	L,  KK, 2						#	L=KK/4, KK is the length of the retangular data part of Bj
+	blez	L, .L95
+	move	BO,  B							#	reset B
+
+.L92:
+	LD	a5,  1 * SIZE(AO)
+	LD	b5,  1 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+
+	LD	a3,  2 * SIZE(AO)
+	LD	b3,  2 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+
+	LD	a7,  3 * SIZE(AO)
+	LD	b7,  3 * SIZE(BO)
+
+	MADD	t11, t11, a3, b3
+
+	daddiu	AO, AO,  4 * SIZE			#	AO += 1mr*4kr	
+	daddiu	BO, BO,  4 * SIZE			#	BP += 1nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+
+	MADD	t11, t11, a7, b7
+
+	daddiu	L, L, -1
+	bgtz	L, .L92
+	NOP
+
+
+.L95:
+	andi	L, KK, 3					#	deal with kc remainder part
+	blez	L, .L98
+	NOP
+
+	.align	3
+.L96:
+	MADD	t11, t11, a1, b1
+
+	daddiu	AO, AO, 1 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 1 * SIZE			#	BP += 1nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L96
+	NOP
+
+
+	.align 3
+.L98:									#	.L98 always deal with the trigular data part
+ 	LD	b1,  0 * SIZE(AO)				#	for RN & RT A is the result matrix
+
+	SUB	t11, b1, t11
+
+
+	LD	b1,  0 * SIZE(BO)				#	BO point to the beginning of the trigular data part of Bj
+	MUL	t11, b1, t11
+
+
+	ST	t11,  0 * SIZE(AO)				#	update packed blockA for follow-up compute
+
+	ST	t11,  0 * SIZE(CO1)				#	write back results
+	
+
+	daddiu	CO1, CO1, 1 * SIZE			#	fixed address
+
+	dsubu	TEMP, K, KK					#	temp = kc - retangular data length of every panel
+	dsll	L,    TEMP, BASE_SHIFT
+	dsll	TEMP, TEMP, BASE_SHIFT		#	nr=1
+	daddu	AO, AO, L					#	move AO to the end of this panel. also the beginning of next panel
+	daddu	BO, BO, TEMP				#	move BO to the end of this panel
+
+
+	.align 3
+.L79:
+	move	B,  BO
+	daddiu	KK, KK, 1
+
+
+	.align 3
+
+
+.L999:
+	LDARG	$16,   0($sp)
+	LDARG	$17,   8($sp)
+	LDARG	$18,  16($sp)
+	LDARG	$19,  24($sp)
+	LDARG	$20,  32($sp)
+	LDARG	$21,  40($sp)
+	ldc1	$f24, 48($sp)
+	ldc1	$f25, 56($sp)
+	ldc1	$f26, 64($sp)
+	ldc1	$f27, 72($sp)
+	ldc1	$f28, 80($sp)
+
+	LDARG	$22,  88($sp)
+	LDARG	$23,  96($sp)
+	LDARG	$24, 104($sp)
+	LDARG	$25, 112($sp)
+
+#ifndef __64BIT__
+	ldc1	$f20,112($sp)
+	ldc1	$f21,120($sp)
+	ldc1	$f22,128($sp)
+	ldc1	$f23,136($sp)
+#endif
+
+	j	$31
+	daddiu	$sp, $sp, 144
+
+	EPILOGUE
diff --git a/kernel/mips64/trsm_kernel_RT_loongson3a.S b/kernel/mips64/trsm_kernel_RT_loongson3a.S
new file mode 100644
index 000000000..cf20cf9e0
--- /dev/null
+++ b/kernel/mips64/trsm_kernel_RT_loongson3a.S
@@ -0,0 +1,1958 @@
+#define REALNAME ASMNAME
+
+#define ASSEMBLER
+#include "common.h"
+
+
+#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 OFFSET	$22
+#define KK	$23
+#define TEMP	$24
+#define AORIG	$25
+
+#define a1	$f0
+#define a2	$f1
+#define a3	$f26
+#define a4	$f27
+
+#define a5	$f28
+#define a6	$f29
+#define	a7	$f30
+#define	a8	$f31
+
+#define b1	$f2
+#define b2	$f3
+#define b3	$f4
+#define b4	$f5
+
+#define b5	$f6
+#define b6	$f7
+#define b7	$f8
+#define b8	$f9
+
+#define t11	$f10
+#define t21	$f11
+#define t31	$f12
+#define	t41	$f13
+
+#define t12	$f14
+#define	t22	$f15
+#define t32	$f16
+#define	t42	$f17
+
+#define	t13	$f18
+#define	t23	$f19
+#define	t33	$f20
+#define	t43	$f21
+
+#define	t14	$f22
+#define	t24	$f23
+#define	t34	$f24
+#define t44	$f25
+
+	PROLOGUE
+	
+	daddiu	$sp, $sp, -144
+
+	SDARG	$16,   0($sp)
+	SDARG	$17,   8($sp)
+	SDARG	$18,  16($sp)
+	SDARG	$19,  24($sp)
+	SDARG	$20,  32($sp)
+	SDARG	$21,  40($sp)
+	sdc1	$f24, 48($sp)
+	sdc1	$f25, 56($sp)
+	sdc1	$f26, 64($sp)
+	sdc1	$f27, 72($sp)
+	sdc1	$f28, 80($sp)
+
+	SDARG	$22,  88($sp)
+	SDARG	$23,  96($sp)
+	SDARG	$24, 104($sp)
+	SDARG	$25, 112($sp)
+
+#ifndef __64BIT__
+	sdc1	$f20,112($sp)
+	sdc1	$f21,120($sp)
+	sdc1	$f22,128($sp)
+	sdc1	$f23,136($sp)
+#endif
+
+
+	.align	3								#	RT compute from right to left
+	LDARG	OFFSET, 144($sp)				#	get the last parameter
+	dsll	LDC, LDC, BASE_SHIFT			#	LDC * data_Byte
+
+	mult	N, K
+	mflo	TEMP
+
+	dsll	TEMP, TEMP, BASE_SHIFT			#	B Representative triangle matrix!!!
+	daddu	B, B, TEMP						#	B point to the end of sb
+											#	Be carefull B has no effeck of mc!!
+	mult	N, LDC
+	mflo	TEMP
+	daddu	C, C, TEMP						#	C point to the last colum of blockB 
+
+	dsubu	KK, K, OFFSET					#	KC-KK is the length of rectangular data part of Bj
+
+	andi	J,  N, 1
+	blez	J, .L30
+	nop
+
+	dsll	TEMP, K, BASE_SHIFT
+	dsubu	B, B, TEMP						#	move B to the beginning address of Bj
+
+	dsubu	C, C, LDC
+
+	move	CO1, C
+
+	move	AORIG, A
+	
+	dsra	I,  M, 2
+	blez	I, .L80
+	NOP
+
+.L31:										#	mr=4,nr=1
+	dsll	L,    KK, 2 + BASE_SHIFT		#	mr=4	
+	dsll	TEMP, KK, BASE_SHIFT			#	nr=1
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 4 results registers
+	MOV	t21, t11
+	MOV	t31, t11
+	MOV	t41, t11
+
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L35
+	NOP
+	
+	.align	3
+
+.L32:
+	LD	a5,  4 * SIZE(AO)
+	LD	a6,  5 * SIZE(AO)
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b5,  1 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	LD	a1,  8 * SIZE(AO)
+	LD	a2,  9 * SIZE(AO)
+	LD	a3,  10 * SIZE(AO)
+	LD	a4,  11 * SIZE(AO)
+
+	LD	b3,  2 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t31, t31, a7, b5
+	MADD	t41, t41, a8, b5
+
+	LD	a5,  12 * SIZE(AO)
+	LD	a6,  13 * SIZE(AO)
+	LD	a7,  14 * SIZE(AO)
+	LD	a8,  15 * SIZE(AO)
+
+	LD	b7,  3 * SIZE(BO)
+
+	MADD	t11, t11, a1, b3
+	MADD	t21, t21, a2, b3
+	MADD	t31, t31, a3, b3
+	MADD	t41, t41, a4, b3
+
+	daddiu	AO, AO, 16 * SIZE			#	AO += 4mr*4kr	
+	daddiu	BO, BO,  4 * SIZE			#	BO += 1nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	MADD	t11, t11, a5, b7
+	MADD	t21, t21, a6, b7
+	MADD	t31, t31, a7, b7
+	MADD	t41, t41, a8, b7
+
+	daddiu	L, L, -1
+	bgtz	L, .L32
+	NOP
+
+
+	.align 3
+
+.L35:
+	andi	L, TEMP, 3
+	blez	L, .L38
+	NOP
+	.align	3
+
+.L36:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	daddiu	AO, AO, 4 * SIZE			#	AO += 4mr	
+	daddiu	BO, BO, 1 * SIZE			#	BO += 2nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L36
+	NOP
+
+
+	.align
+.L38:
+	daddiu	TEMP, KK,  -1				#	deal with the triangular data part
+	dsll	L,    TEMP, 2 + BASE_SHIFT
+	dsll	TEMP, TEMP, BASE_SHIFT		#	nr=1
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+	LD	b2,  1 * SIZE(AO)
+	LD	b3,  2 * SIZE(AO)
+	LD	b4,  3 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+	SUB	t31, b3, t31
+	SUB	t41, b4, t41
+
+
+	LD	b2,   0 * SIZE(BO)
+	MUL	t11, b2, t11
+	MUL	t21, b2, t21
+	MUL	t31, b2, t31
+	MUL	t41, b2, t41
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+	ST	t21,  1 * SIZE(AO)
+	ST	t31,  2 * SIZE(AO)
+	ST	t41,  3 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	ST	t21,  1 * SIZE(CO1)
+	ST	t31,  2 * SIZE(CO1)
+	ST	t41,  3 * SIZE(CO1)
+	
+
+	daddiu	CO1, CO1, 4 * SIZE			#	fixed pointer
+
+	dsll	TEMP, K, 2 + BASE_SHIFT		
+	daddu	AORIG, AORIG, TEMP			# move to next panel Ai
+
+	daddiu	I, I, -1
+	bgtz	I, .L31
+	NOP
+
+
+	.align 3
+.L80:
+	andi	I, M, 2
+	blez	I, .L90
+	nop
+
+	dsll	L,    KK, 1 + BASE_SHIFT		#	mr=2	
+	dsll	TEMP, KK, BASE_SHIFT			#	nr=1
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 4 results registers
+	MOV	t21, t11
+
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L85
+	NOP
+	
+	.align	3
+
+.L82:
+	LD	a5,  2 * SIZE(AO)
+	LD	a6,  3 * SIZE(AO)
+
+	LD	b5,  1 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	LD	a3,  4 * SIZE(AO)
+	LD	a4,  5 * SIZE(AO)
+
+	LD	b3,  2 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b7,  3 * SIZE(BO)
+
+	MADD	t11, t11, a3, b3
+	MADD	t21, t21, a4, b3
+
+	daddiu	AO, AO, 16 * SIZE			#	AO += 4mr*4kr	
+	daddiu	BO, BO,  4 * SIZE			#	BO += 1nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	MADD	t11, t11, a7, b7
+	MADD	t21, t21, a8, b7
+
+	daddiu	L, L, -1
+	bgtz	L, .L82
+	NOP
+
+
+	.align 3
+
+.L85:
+	andi	L, TEMP, 3
+	blez	L, .L88
+	NOP
+	.align	3
+
+.L86:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	daddiu	AO, AO, 2 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 1 * SIZE			#	BO += 1nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L86
+	NOP
+
+
+	.align
+.L88:
+	daddiu	TEMP, KK,  -1				#	deal with the triangular data part
+	dsll	L,    TEMP, 1 + BASE_SHIFT
+	dsll	TEMP, TEMP, BASE_SHIFT		#	nr=1
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+	LD	b2,  1 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+
+
+	LD	b2,   0 * SIZE(BO)
+	MUL	t11, b2, t11
+	MUL	t21, b2, t21
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+	ST	t21,  1 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	ST	t21,  1 * SIZE(CO1)
+	
+
+	daddiu	CO1, CO1, 2 * SIZE			#	fixed pointer
+
+	dsll	TEMP, K, 1 + BASE_SHIFT		
+	daddu	AORIG, AORIG, TEMP			# move to next panel Ai
+
+
+	.align 3
+.L90:
+	andi	I, M, 1
+	blez	I, .L39
+	nop
+
+	dsll	L,    KK, BASE_SHIFT			#	mr=1	
+	dsll	TEMP, KK, BASE_SHIFT			#	nr=1
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 4 results registers
+
+	LD	a1,  0 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L95
+	NOP
+	
+	.align	3
+
+.L92:
+	LD	a5,  1 * SIZE(AO)
+	LD	b5,  1 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+
+	LD	a3,  2 * SIZE(AO)
+	LD	b3,  2 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+
+	LD	a7,  3 * SIZE(AO)
+	LD	b7,  3 * SIZE(BO)
+
+	MADD	t11, t11, a3, b3
+
+	daddiu	AO, AO,  4 * SIZE			#	AO += 1mr*4kr	
+	daddiu	BO, BO,  4 * SIZE			#	BO += 1nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+
+	MADD	t11, t11, a7, b7
+
+	daddiu	L, L, -1
+	bgtz	L, .L92
+	NOP
+
+
+	.align 3
+
+.L95:
+	andi	L, TEMP, 3
+	blez	L, .L98
+	NOP
+	.align	3
+
+.L96:
+	MADD	t11, t11, a1, b1
+
+	daddiu	AO, AO, 1 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 1 * SIZE			#	BO += 1nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	b1,  0 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L96
+	NOP
+
+
+	.align
+.L98:
+	daddiu	TEMP, KK,  -1				#	deal with the triangular data part
+	dsll	L,    TEMP, BASE_SHIFT
+	dsll	TEMP, TEMP, BASE_SHIFT		#	nr=1
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+
+	SUB	t11, b1, t11
+
+
+	LD	b2,   0 * SIZE(BO)
+	MUL	t11, b2, t11
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	
+
+	daddiu	CO1, CO1, 1 * SIZE			#	fixed pointer
+
+	dsll	TEMP, K, BASE_SHIFT		
+	daddu	AORIG, AORIG, TEMP			# move to next panel Ai
+
+
+.L39:
+	daddiu	KK, KK, -1						#	rectangular data length increased by 1
+
+
+	.align 3
+.L30:										#	nr=2
+	andi	J,  N, 2
+	blez	J, .L50
+	nop
+
+	dsll	TEMP, K, 1 + BASE_SHIFT			#	Kc*2nr	move B to the beginning address of Bj
+	dsubu	B, B, TEMP
+
+	dsll	TEMP, LDC, 1					#	C 
+	dsubu	C, C, TEMP
+
+	move	CO1, C
+	daddu	CO2, C,   LDC
+
+	move	AORIG, A
+
+	dsra	I,  M, 2
+	blez	I, .L60
+	NOP
+
+.L51:										#	mr=4,nr=2
+	dsll	L,    KK, 2 + BASE_SHIFT		#	mr=4	
+	dsll	TEMP, KK, 1 + BASE_SHIFT		#	nr=2
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 8 results registers
+	MOV	t21, t11
+	MOV	t31, t11
+	MOV	t41, t11
+	MOV	t12, t11
+	MOV	t22, t11
+	MOV	t32, t11
+	MOV	t42, t11
+
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L55
+	NOP
+	
+	.align	3
+
+.L52:
+	LD	a5,  4 * SIZE(AO)
+	LD	a6,  5 * SIZE(AO)
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b5,  2 * SIZE(BO)
+	LD	b6,  3 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	LD	a1,  8 * SIZE(AO)
+	LD	a2,  9 * SIZE(AO)
+	LD	a3,  10 * SIZE(AO)
+	LD	a4,  11 * SIZE(AO)
+
+	LD	b3,  4 * SIZE(BO)
+	LD	b4,  5 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t31, t31, a7, b5
+	MADD	t41, t41, a8, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+	MADD	t32, t32, a7, b6
+	MADD	t42, t42, a8, b6
+
+	LD	a5,  12 * SIZE(AO)
+	LD	a6,  13 * SIZE(AO)
+	LD	a7,  14 * SIZE(AO)
+	LD	a8,  15 * SIZE(AO)
+
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a1, b3
+	MADD	t21, t21, a2, b3
+	MADD	t31, t31, a3, b3
+	MADD	t41, t41, a4, b3
+
+	MADD	t12, t12, a1, b4
+	MADD	t22, t22, a2, b4
+	MADD	t32, t32, a3, b4
+	MADD	t42, t42, a4, b4
+
+	daddiu	AO, AO, 16 * SIZE			#	AO += 4mr*4kr	
+	daddiu	BO, BO,  8 * SIZE			#	BO += 2nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	MADD	t11, t11, a5, b7
+	MADD	t21, t21, a6, b7
+	MADD	t31, t31, a7, b7
+	MADD	t41, t41, a8, b7
+
+	MADD	t12, t12, a5, b8
+	MADD	t22, t22, a6, b8
+	MADD	t32, t32, a7, b8
+	MADD	t42, t42, a8, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L52
+	NOP
+
+
+	.align 3
+
+.L55:
+	andi	L, TEMP, 3
+	blez	L, .L58
+	NOP
+	.align	3
+
+.L56:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	daddiu	AO, AO, 4 * SIZE			#	AO += 4mr	
+	daddiu	BO, BO, 2 * SIZE			#	BO += 2nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L56
+	NOP
+
+
+	.align
+.L58:
+	daddiu	TEMP, KK,  -2				#	deal with the triangular data part
+	dsll	L,    TEMP, 2 + BASE_SHIFT
+	dsll	TEMP, TEMP, 1 + BASE_SHIFT	#	nr=2
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+	LD	b2,  1 * SIZE(AO)
+	LD	b3,  2 * SIZE(AO)
+	LD	b4,  3 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+	SUB	t31, b3, t31
+	SUB	t41, b4, t41
+
+ 	LD	b5,  4 * SIZE(AO)
+	LD	b6,  5 * SIZE(AO)
+	LD	b7,  6 * SIZE(AO)
+	LD	b8,  7 * SIZE(AO)
+
+	SUB	t12, b5, t12
+	SUB	t22, b6, t22
+	SUB	t32, b7, t32
+	SUB	t42, b8, t42
+
+
+	LD	b8,   3 * SIZE(BO)
+	LD	b1,   2 * SIZE(BO)
+	MUL	t12, b8, t12
+	MUL	t22, b8, t22
+	MUL	t32, b8, t32
+	MUL	t42, b8, t42
+	NMSUB	t11, t11, b1, t12
+	NMSUB	t21, t21, b1, t22
+	NMSUB	t31, t31, b1, t32
+	NMSUB	t41, t41, b1, t42
+
+
+	LD	b2,   0 * SIZE(BO)
+	MUL	t11, b2, t11
+	MUL	t21, b2, t21
+	MUL	t31, b2, t31
+	MUL	t41, b2, t41
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+	ST	t21,  1 * SIZE(AO)
+	ST	t31,  2 * SIZE(AO)
+	ST	t41,  3 * SIZE(AO)
+
+	ST	t12,  4 * SIZE(AO)
+	ST	t22,  5 * SIZE(AO)
+	ST	t32,  6 * SIZE(AO)
+	ST	t42,  7 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	ST	t21,  1 * SIZE(CO1)
+	ST	t31,  2 * SIZE(CO1)
+	ST	t41,  3 * SIZE(CO1)
+	
+	ST	t12,  0 * SIZE(CO2)
+	ST	t22,  1 * SIZE(CO2)
+	ST	t32,  2 * SIZE(CO2)
+	ST	t42,  3 * SIZE(CO2)
+	
+	daddiu	CO1, CO1, 4 * SIZE			#	fixed pointer
+	daddiu	CO2, CO2, 4 * SIZE
+
+	dsll	TEMP, K, 2 + BASE_SHIFT		
+	daddu	AORIG, AORIG, TEMP			# move to next panel Ai
+
+	daddiu	I, I, -1
+	bgtz	I, .L51
+	NOP
+
+
+
+	.align 3
+.L60:
+	andi	I,  M, 2						#	mr=2
+	blez	I, .L70
+	nop
+
+	dsll	L,    KK, 1 + BASE_SHIFT		#	mr=2	
+	dsll	TEMP, KK, 1 + BASE_SHIFT		#	nr=2
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 8 results registers
+	MOV	t21, t11
+	MOV	t12, t11
+	MOV	t22, t11
+
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L65
+	NOP
+	
+	.align	3
+
+.L62:
+	LD	a5,  2 * SIZE(AO)
+	LD	a6,  3 * SIZE(AO)
+
+	LD	b5,  2 * SIZE(BO)
+	LD	b6,  3 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+
+	LD	a3,  4 * SIZE(AO)
+	LD	a4,  5 * SIZE(AO)
+
+	LD	b3,  4 * SIZE(BO)
+	LD	b4,  5 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a3, b3
+	MADD	t21, t21, a4, b3
+
+	MADD	t12, t12, a3, b4
+	MADD	t22, t22, a4, b4
+
+	daddiu	AO, AO,  8 * SIZE			#	AO += 2mr*4kr	
+	daddiu	BO, BO,  8 * SIZE			#	BO += 2nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	MADD	t11, t11, a7, b7
+	MADD	t21, t21, a8, b7
+
+	MADD	t12, t12, a7, b8
+	MADD	t22, t22, a8, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L62
+	NOP
+
+
+	.align 3
+
+.L65:
+	andi	L, TEMP, 3
+	blez	L, .L68
+	NOP
+	.align	3
+
+.L66:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+
+	daddiu	AO, AO, 2 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 2 * SIZE			#	BO += 2nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L66
+	NOP
+
+
+	.align
+.L68:
+	daddiu	TEMP, KK,  -2				#	deal with the triangular data part
+	dsll	L,    TEMP, 1 + BASE_SHIFT
+	dsll	TEMP, TEMP, 1 + BASE_SHIFT	#	nr=2
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+	LD	b2,  1 * SIZE(AO)
+	LD	b3,  2 * SIZE(AO)
+	LD	b4,  3 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+	SUB	t12, b3, t12
+	SUB	t22, b4, t22
+
+
+	LD	b8,   3 * SIZE(BO)
+	LD	b7,   2 * SIZE(BO)
+	MUL	t12, b8, t12
+	MUL	t22, b8, t22
+	NMSUB	t11, t11, b7, t12
+	NMSUB	t21, t21, b7, t22
+
+
+	LD	b6,   0 * SIZE(BO)
+	MUL	t11, b6, t11
+	MUL	t21, b6, t21
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+	ST	t21,  1 * SIZE(AO)
+	ST	t12,  2 * SIZE(AO)
+	ST	t22,  3 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	ST	t21,  1 * SIZE(CO1)
+	
+	ST	t12,  0 * SIZE(CO2)
+	ST	t22,  1 * SIZE(CO2)
+	
+	daddiu	CO1, CO1, 2 * SIZE			#	fixed pointer
+	daddiu	CO2, CO2, 2 * SIZE
+
+	dsll	TEMP, K, 1 + BASE_SHIFT		#	mr=2
+	daddu	AORIG, AORIG, TEMP			# 	move to next panel Ai
+
+
+
+	.align 3
+.L70:
+	andi	I, M, 1							#	mr=1
+	blez	I, .L59
+	nop
+
+	dsll	L,    KK, BASE_SHIFT			#	mr=1	
+	dsll	TEMP, KK, 1 + BASE_SHIFT		#	nr=2
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 8 results registers
+	MOV	t12, t11
+
+	LD	a1,  0 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L75
+	NOP
+	
+	.align	3
+
+.L72:
+	LD	a5,  1 * SIZE(AO)
+
+	LD	b5,  2 * SIZE(BO)
+	LD	b6,  3 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t12, t12, a1, b2
+
+	LD	a3,  2 * SIZE(AO)
+
+	LD	b3,  4 * SIZE(BO)
+	LD	b4,  5 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t12, t12, a5, b6
+
+	LD	a7,  3 * SIZE(AO)
+
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a3, b3
+	MADD	t12, t12, a3, b4
+
+	daddiu	AO, AO,  4 * SIZE			#	AO += 1mr*4kr	
+	daddiu	BO, BO,  8 * SIZE			#	BO += 2nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	MADD	t11, t11, a7, b7
+	MADD	t12, t12, a7, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L72
+	NOP
+
+
+	.align 3
+
+.L75:
+	andi	L, TEMP, 3
+	blez	L, .L78
+	NOP
+	.align	3
+
+.L76:
+	MADD	t11, t11, a1, b1
+	MADD	t12, t12, a1, b2
+
+	daddiu	AO, AO, 1 * SIZE			#	AO += 1mr	
+	daddiu	BO, BO, 2 * SIZE			#	BO += 2nr
+	
+	LD	a1,  0 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L76
+	NOP
+
+
+	.align
+.L78:
+	daddiu	TEMP, KK,  -2				#	deal with the triangular data part
+	dsll	L,    TEMP, BASE_SHIFT
+	dsll	TEMP, TEMP, 1 + BASE_SHIFT	#	nr=2
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+	LD	b2,  1 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t12, b2, t12
+
+
+	LD	b8,   3 * SIZE(BO)
+	LD	b7,   2 * SIZE(BO)
+	MUL	t12, b8, t12
+	NMSUB	t11, t11, b7, t12
+
+
+	LD	b6,   0 * SIZE(BO)
+	MUL	t11, b6, t11
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+	ST	t12,  1 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	ST	t12,  0 * SIZE(CO2)
+	
+	daddiu	CO1, CO1, 1 * SIZE			#	fixed pointer
+	daddiu	CO2, CO2, 1 * SIZE
+
+	dsll	TEMP, K, BASE_SHIFT		#	mr=2
+	daddu	AORIG, AORIG, TEMP			# 	move to next panel Ai
+
+
+.L59:
+	daddiu	KK, KK, -2						#	rectangular data length increased by 2
+
+
+
+	.align 3
+.L50:
+	dsra	J,  N, 2						#	J = NC/4 
+	blez	J, .L999
+	NOP
+
+.L10:
+	dsll	TEMP, K, 2 + BASE_SHIFT
+	dsubu	B, B, TEMP						#	move B to the beginning address of Bj	
+
+	dsll	TEMP, LDC, 2
+	dsubu	C, C, TEMP						#	move C to the beginning address of Cj
+
+	daddiu	J, J, -1
+
+	move	CO1, C
+	daddu	CO2, C,   LDC
+	daddu	CO3, CO2, LDC
+	daddu	CO4, CO3, LDC
+	
+	move	AORIG, A						#	reset A
+
+	dsra	I,  M, 2						#	I=MC/4
+	blez	I, .L20
+	NOP
+
+	.align 3
+.L11:
+	dsll	L,    KK, 2 + BASE_SHIFT		#	mr=4	
+	dsll	TEMP, KK, 2 + BASE_SHIFT		#	nr=4
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 16 results registers
+	MOV	t21, t11
+	MOV	t31, t11
+	MOV	t41, t11
+	MOV	t12, t11
+	MOV	t22, t11
+	MOV	t32, t11
+	MOV	t42, t11
+	MOV	t13, t11
+	MOV	t23, t11
+	MOV	t33, t11
+	MOV	t43, t11
+	MOV	t14, t11
+	MOV	t24, t11
+	MOV	t34, t11
+	MOV	t44, t11
+
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L15
+	NOP
+	
+	.align	3
+
+.L12:
+	LD	a5,  4 * SIZE(AO)
+	LD	a6,  5 * SIZE(AO)
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b5,  4 * SIZE(BO)
+	LD	b6,  5 * SIZE(BO)
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+	MADD	t33, t33, a3, b3
+	MADD	t43, t43, a4, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+	MADD	t34, t34, a3, b4
+	MADD	t44, t44, a4, b4			#	fisrt 	
+
+	LD	a1,  8 * SIZE(AO)
+	LD	a2,  9 * SIZE(AO)
+	LD	a3,  10 * SIZE(AO)
+	LD	a4,  11 * SIZE(AO)
+
+	LD	b1,  8 * SIZE(BO)
+	LD	b2,  9 * SIZE(BO)
+	LD	b3,  10 * SIZE(BO)
+	LD	b4,  11 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t31, t31, a7, b5
+	MADD	t41, t41, a8, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+	MADD	t32, t32, a7, b6
+	MADD	t42, t42, a8, b6
+
+	MADD	t13, t13, a5, b7
+	MADD	t23, t23, a6, b7
+	MADD	t33, t33, a7, b7
+	MADD	t43, t43, a8, b7
+
+	MADD	t14, t14, a5, b8
+	MADD	t24, t24, a6, b8
+	MADD	t34, t34, a7, b8
+	MADD	t44, t44, a8, b8			#	second
+
+	LD	a5,  12 * SIZE(AO)
+	LD	a6,  13 * SIZE(AO)
+	LD	a7,  14 * SIZE(AO)
+	LD	a8,  15 * SIZE(AO)
+
+	LD	b5,  12 * SIZE(BO)
+	LD	b6,  13 * SIZE(BO)
+	LD	b7,  14 * SIZE(BO)
+	LD	b8,  15 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+	MADD	t33, t33, a3, b3
+	MADD	t43, t43, a4, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+	MADD	t34, t34, a3, b4
+	MADD	t44, t44, a4, b4			#	third
+
+	daddiu	AO, AO, 16 * SIZE			#	AO += 4mr*4kr	
+	daddiu	BO, BO, 16 * SIZE			#	BO += 4nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+	MADD	t31, t31, a7, b5
+	MADD	t41, t41, a8, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+	MADD	t32, t32, a7, b6
+	MADD	t42, t42, a8, b6
+
+	MADD	t13, t13, a5, b7
+	MADD	t23, t23, a6, b7
+	MADD	t33, t33, a7, b7
+	MADD	t43, t43, a8, b7
+
+	MADD	t14, t14, a5, b8
+	MADD	t24, t24, a6, b8
+	MADD	t34, t34, a7, b8
+	MADD	t44, t44, a8, b8			#	fouth
+
+	daddiu	L, L, -1
+	bgtz	L, .L12
+	NOP
+
+
+	.align 3
+
+.L15:
+	andi	L, TEMP, 3
+	blez	L, .L18
+	NOP
+	.align	3
+
+.L16:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+	MADD	t31, t31, a3, b1
+	MADD	t41, t41, a4, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+	MADD	t32, t32, a3, b2
+	MADD	t42, t42, a4, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+	MADD	t33, t33, a3, b3
+	MADD	t43, t43, a4, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+	MADD	t34, t34, a3, b4
+	MADD	t44, t44, a4, b4			#	third
+
+	daddiu	AO, AO, 4 * SIZE			#	AO += 4mr	
+	daddiu	BO, BO, 4 * SIZE			#	BO += 4nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+	LD	a3,  2 * SIZE(AO)
+	LD	a4,  3 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L16
+	NOP
+
+
+	.align
+.L18:
+	daddiu	TEMP, KK, -4				#	deal with the triangular data part
+	dsll	L,    TEMP, 2 + BASE_SHIFT
+	dsll	TEMP, TEMP, 2 + BASE_SHIFT
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+	LD	b2,  1 * SIZE(AO)
+	LD	b3,  2 * SIZE(AO)
+	LD	b4,  3 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+	SUB	t31, b3, t31
+	SUB	t41, b4, t41
+
+ 	LD	b5,  4 * SIZE(AO)
+	LD	b6,  5 * SIZE(AO)
+	LD	b7,  6 * SIZE(AO)
+	LD	b8,  7 * SIZE(AO)
+
+	SUB	t12, b5, t12
+	SUB	t22, b6, t22
+	SUB	t32, b7, t32
+	SUB	t42, b8, t42
+
+ 	LD	b1,  8 * SIZE(AO)
+	LD	b2,  9 * SIZE(AO)
+	LD	b3, 10 * SIZE(AO)
+	LD	b4, 11 * SIZE(AO)
+
+	SUB	t13, b1, t13
+	SUB	t23, b2, t23
+	SUB	t33, b3, t33
+	SUB	t43, b4, t43
+
+ 	LD	b5, 12 * SIZE(AO)
+	LD	b6, 13 * SIZE(AO)
+	LD	b7, 14 * SIZE(AO)
+	LD	b8, 15 * SIZE(AO)
+
+	SUB	t14, b5, t14
+	SUB	t24, b6, t24
+	SUB	t34, b7, t34
+	SUB	t44, b8, t44
+
+
+	LD	b1, 15 * SIZE(BO)
+	LD	b2, 14 * SIZE(BO)
+	LD	b3, 13 * SIZE(BO)
+	LD	b4, 12 * SIZE(BO)
+	MUL	t14, b1, t14
+	MUL	t24, b1, t24
+	MUL	t34, b1, t34
+	MUL	t44, b1, t44
+	NMSUB	t13, t13, b2, t14
+	NMSUB	t23, t23, b2, t24
+	NMSUB	t33, t33, b2, t34
+	NMSUB	t43, t43, b2, t44
+	NMSUB	t12, t12, b3, t14
+	NMSUB	t22, t22, b3, t24
+	NMSUB	t32, t32, b3, t34
+	NMSUB	t42, t42, b3, t44
+	NMSUB	t11, t11, b4, t14
+	NMSUB	t21, t21, b4, t24
+	NMSUB	t31, t31, b4, t34
+	NMSUB	t41, t41, b4, t44
+
+	
+	LD	b5,  10 * SIZE(BO)
+	LD	b6,   9 * SIZE(BO)
+	LD	b7,   8 * SIZE(BO)
+	MUL	t13, b5, t13
+	MUL	t23, b5, t23
+	MUL	t33, b5, t33
+	MUL	t43, b5, t43
+	NMSUB	t12, t12, b6, t13
+	NMSUB	t22, t22, b6, t23
+	NMSUB	t32, t32, b6, t33
+	NMSUB	t42, t42, b6, t43
+	NMSUB	t11, t11, b7, t13
+	NMSUB	t21, t21, b7, t23
+	NMSUB	t31, t31, b7, t33
+	NMSUB	t41, t41, b7, t43
+
+
+	LD	b8,   5 * SIZE(BO)
+	LD	b1,   4 * SIZE(BO)
+	MUL	t12, b8, t12
+	MUL	t22, b8, t22
+	MUL	t32, b8, t32
+	MUL	t42, b8, t42
+	NMSUB	t11, t11, b1, t12
+	NMSUB	t21, t21, b1, t22
+	NMSUB	t31, t31, b1, t32
+	NMSUB	t41, t41, b1, t42
+
+
+	LD	b2,   0 * SIZE(BO)
+	MUL	t11, b2, t11
+	MUL	t21, b2, t21
+	MUL	t31, b2, t31
+	MUL	t41, b2, t41
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+	ST	t21,  1 * SIZE(AO)
+	ST	t31,  2 * SIZE(AO)
+	ST	t41,  3 * SIZE(AO)
+
+	ST	t12,  4 * SIZE(AO)
+	ST	t22,  5 * SIZE(AO)
+	ST	t32,  6 * SIZE(AO)
+	ST	t42,  7 * SIZE(AO)
+
+	ST	t13,  8 * SIZE(AO)
+	ST	t23,  9 * SIZE(AO)
+	ST	t33, 10 * SIZE(AO)
+	ST	t43, 11 * SIZE(AO)
+
+	ST	t14, 12 * SIZE(AO)
+	ST	t24, 13 * SIZE(AO)
+	ST	t34, 14 * SIZE(AO)
+	ST	t44, 15 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	ST	t21,  1 * SIZE(CO1)
+	ST	t31,  2 * SIZE(CO1)
+	ST	t41,  3 * SIZE(CO1)
+	
+	ST	t12,  0 * SIZE(CO2)
+	ST	t22,  1 * SIZE(CO2)
+	ST	t32,  2 * SIZE(CO2)
+	ST	t42,  3 * SIZE(CO2)
+	
+	ST	t13,  0 * SIZE(CO3)
+	ST	t23,  1 * SIZE(CO3)
+	ST	t33,  2 * SIZE(CO3)
+	ST	t43,  3 * SIZE(CO3)
+	
+	ST	t14,  0 * SIZE(CO4)
+	ST	t24,  1 * SIZE(CO4)
+	ST	t34,  2 * SIZE(CO4)
+	ST	t44,  3 * SIZE(CO4)
+
+	daddiu	CO1, CO1, 4 * SIZE			#	fixed pointer
+	daddiu	CO2, CO2, 4 * SIZE
+	daddiu	CO3, CO3, 4 * SIZE
+	daddiu	CO4, CO4, 4 * SIZE
+
+	dsll	TEMP, K, 2 + BASE_SHIFT		
+	daddu	AORIG, AORIG, TEMP			# move to next panel Ai
+
+	daddiu	I, I, -1
+	bgtz	I, .L11
+	NOP
+
+	.align 3
+.L20:
+	andi	I,  M, 2						#	mr=2
+	blez	I, .L40
+	NOP
+
+	dsll	L,    KK, 1 + BASE_SHIFT		#	mr=2	
+	dsll	TEMP, KK, 2 + BASE_SHIFT		#	nr=4
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 8 results registers
+	MOV	t21, t11
+	MOV	t12, t11
+	MOV	t22, t11
+	MOV	t13, t11
+	MOV	t23, t11
+	MOV	t14, t11
+	MOV	t24, t11
+
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L25
+	NOP
+	
+	.align	3
+
+.L22:
+	LD	a5,  2 * SIZE(AO)
+	LD	a6,  3 * SIZE(AO)
+
+	LD	b5,  4 * SIZE(BO)
+	LD	b6,  5 * SIZE(BO)
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+
+	LD	a3,  4 * SIZE(AO)
+	LD	a4,  5 * SIZE(AO)
+
+	LD	b1,  8 * SIZE(BO)
+	LD	b2,  9 * SIZE(BO)
+	LD	b3,  10 * SIZE(BO)
+	LD	b4,  11 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t21, t21, a6, b5
+
+	MADD	t12, t12, a5, b6
+	MADD	t22, t22, a6, b6
+
+	MADD	t13, t13, a5, b7
+	MADD	t23, t23, a6, b7
+
+	MADD	t14, t14, a5, b8
+	MADD	t24, t24, a6, b8
+
+	LD	a7,  6 * SIZE(AO)
+	LD	a8,  7 * SIZE(AO)
+
+	LD	b5,  12 * SIZE(BO)
+	LD	b6,  13 * SIZE(BO)
+	LD	b7,  14 * SIZE(BO)
+	LD	b8,  15 * SIZE(BO)
+
+	MADD	t11, t11, a3, b1
+	MADD	t21, t21, a4, b1
+
+	MADD	t12, t12, a3, b2
+	MADD	t22, t22, a4, b2
+
+	MADD	t13, t13, a3, b3
+	MADD	t23, t23, a4, b3
+
+	MADD	t14, t14, a3, b4
+	MADD	t24, t24, a4, b4
+
+	daddiu	AO, AO, 8 * SIZE			#	AO += 2mr*4kr	
+	daddiu	BO, BO, 16 * SIZE			#	BO += 4nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	MADD	t11, t11, a7, b5
+	MADD	t21, t21, a8, b5
+
+	MADD	t12, t12, a7, b6
+	MADD	t22, t22, a8, b6
+
+	MADD	t13, t13, a7, b7
+	MADD	t23, t23, a8, b7
+
+	MADD	t14, t14, a7, b8
+	MADD	t24, t24, a8, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L22
+	NOP
+
+
+	.align 3
+
+.L25:
+	andi	L, TEMP, 3
+	blez	L, .L28
+	NOP
+	.align	3
+
+.L26:
+	MADD	t11, t11, a1, b1
+	MADD	t21, t21, a2, b1
+
+	MADD	t12, t12, a1, b2
+	MADD	t22, t22, a2, b2
+
+	MADD	t13, t13, a1, b3
+	MADD	t23, t23, a2, b3
+
+	MADD	t14, t14, a1, b4
+	MADD	t24, t24, a2, b4
+
+	daddiu	AO, AO, 2 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 4 * SIZE			#	BO += 4nr
+	
+	LD	a1,  0 * SIZE(AO)
+	LD	a2,  1 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L26
+	NOP
+
+
+	.align
+.L28:
+	daddiu	TEMP, KK, -4				#	deal with the triangular data part
+	dsll	L,    TEMP, 1 + BASE_SHIFT	#	mr=2
+	dsll	TEMP, TEMP, 2 + BASE_SHIFT
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+	LD	b2,  1 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t21, b2, t21
+
+ 	LD	b5,  2 * SIZE(AO)
+	LD	b6,  3 * SIZE(AO)
+
+	SUB	t12, b5, t12
+	SUB	t22, b6, t22
+
+	LD	b3,  4 * SIZE(AO)
+	LD	b4,  5 * SIZE(AO)
+
+	SUB	t13, b3, t13
+	SUB	t23, b4, t23
+
+	LD	b7,  6 * SIZE(AO)
+	LD	b8,  7 * SIZE(AO)
+
+	SUB	t14, b7, t14
+	SUB	t24, b8, t24
+
+
+	LD	b1, 15 * SIZE(BO)
+	LD	b2, 14 * SIZE(BO)
+	LD	b3, 13 * SIZE(BO)
+	LD	b4, 12 * SIZE(BO)
+	MUL	t14, b1, t14
+	MUL	t24, b1, t24
+	NMSUB	t13, t13, b2, t14
+	NMSUB	t23, t23, b2, t24
+	NMSUB	t12, t12, b3, t14
+	NMSUB	t22, t22, b3, t24
+	NMSUB	t11, t11, b4, t14
+	NMSUB	t21, t21, b4, t24
+
+	
+	LD	b5,  10 * SIZE(BO)
+	LD	b6,   9 * SIZE(BO)
+	LD	b7,   8 * SIZE(BO)
+	MUL	t13, b5, t13
+	MUL	t23, b5, t23
+	NMSUB	t12, t12, b6, t13
+	NMSUB	t22, t22, b6, t23
+	NMSUB	t11, t11, b7, t13
+	NMSUB	t21, t21, b7, t23
+
+
+	LD	b8,   5 * SIZE(BO)
+	LD	b1,   4 * SIZE(BO)
+	MUL	t12, b8, t12
+	MUL	t22, b8, t22
+	NMSUB	t11, t11, b1, t12
+	NMSUB	t21, t21, b1, t22
+
+
+	LD	b2,   0 * SIZE(BO)
+	MUL	t11, b2, t11
+	MUL	t21, b2, t21
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+	ST	t21,  1 * SIZE(AO)
+
+	ST	t12,  2 * SIZE(AO)
+	ST	t22,  3 * SIZE(AO)
+
+	ST	t13,  4 * SIZE(AO)
+	ST	t23,  5 * SIZE(AO)
+
+	ST	t14,  6 * SIZE(AO)
+	ST	t24,  7 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	ST	t21,  1 * SIZE(CO1)
+	
+	ST	t12,  0 * SIZE(CO2)
+	ST	t22,  1 * SIZE(CO2)
+	
+	ST	t13,  0 * SIZE(CO3)
+	ST	t23,  1 * SIZE(CO3)
+	
+	ST	t14,  0 * SIZE(CO4)
+	ST	t24,  1 * SIZE(CO4)
+
+	daddiu	CO1, CO1, 2 * SIZE			#	fixed pointer
+	daddiu	CO2, CO2, 2 * SIZE
+	daddiu	CO3, CO3, 2 * SIZE
+	daddiu	CO4, CO4, 2 * SIZE
+
+	dsll	TEMP, K, 1 + BASE_SHIFT		#	mr=2
+	daddu	AORIG, AORIG, TEMP			# 	move to next panel Ai
+
+
+	.align 3
+.L40:
+	andi	I,  M, 1
+	blez	I, .L29
+	NOP
+
+	dsll	L,    KK, BASE_SHIFT			#	mr=1	
+	dsll	TEMP, KK, 2 + BASE_SHIFT		#	nr=4
+	daddu	AO, AORIG, L					
+	daddu	BO, B, TEMP						#	BO point to the retangular data part,also reset BO
+	dsubu	TEMP, K, KK						#	temp = the length of rectangular data part 
+
+	MTC	$0,  t11							#	clear 4 results registers
+	MOV	t12, t11
+	MOV	t13, t11
+	MOV	t14, t11
+
+	LD	a1,  0 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	dsra	L,  TEMP, 2						#	L=(KC-offset)/4
+	blez	L, .L45
+	NOP
+	
+	.align	3
+
+.L42:
+	LD	a5,  1 * SIZE(AO)
+
+	LD	b5,  4 * SIZE(BO)
+	LD	b6,  5 * SIZE(BO)
+	LD	b7,  6 * SIZE(BO)
+	LD	b8,  7 * SIZE(BO)
+
+	MADD	t11, t11, a1, b1
+	MADD	t12, t12, a1, b2
+	MADD	t13, t13, a1, b3
+	MADD	t14, t14, a1, b4
+
+	LD	a3,  2 * SIZE(AO)
+
+	LD	b1,  8 * SIZE(BO)
+	LD	b2,  9 * SIZE(BO)
+	LD	b3,  10 * SIZE(BO)
+	LD	b4,  11 * SIZE(BO)
+
+	MADD	t11, t11, a5, b5
+	MADD	t12, t12, a5, b6
+	MADD	t13, t13, a5, b7
+	MADD	t14, t14, a5, b8
+
+	LD	a7,  3 * SIZE(AO)
+
+	LD	b5,  12 * SIZE(BO)
+	LD	b6,  13 * SIZE(BO)
+	LD	b7,  14 * SIZE(BO)
+	LD	b8,  15 * SIZE(BO)
+
+	MADD	t11, t11, a3, b1
+	MADD	t12, t12, a3, b2
+	MADD	t13, t13, a3, b3
+	MADD	t14, t14, a3, b4
+
+	daddiu	AO, AO, 4 * SIZE			#	AO += 1mr*4kr	
+	daddiu	BO, BO, 16 * SIZE			#	BO += 4nr*4kr
+	
+	LD	a1,  0 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	MADD	t11, t11, a7, b5
+	MADD	t12, t12, a7, b6
+	MADD	t13, t13, a7, b7
+	MADD	t14, t14, a7, b8
+
+	daddiu	L, L, -1
+	bgtz	L, .L42
+	NOP
+
+
+	.align 3
+
+.L45:
+	andi	L, TEMP, 3
+	blez	L, .L48
+	NOP
+	.align	3
+
+.L46:
+	MADD	t11, t11, a1, b1
+	MADD	t12, t12, a1, b2
+	MADD	t13, t13, a1, b3
+	MADD	t14, t14, a1, b4
+
+	daddiu	AO, AO, 1 * SIZE			#	AO += 2mr	
+	daddiu	BO, BO, 4 * SIZE			#	BO += 4nr
+	
+	LD	a1,  0 * SIZE(AO)
+
+	LD	b1,  0 * SIZE(BO)
+	LD	b2,  1 * SIZE(BO)
+	LD	b3,  2 * SIZE(BO)
+	LD	b4,  3 * SIZE(BO)
+
+	daddiu	L, L, -1
+	bgtz	L, .L46
+	NOP
+
+
+	.align
+.L48:
+	daddiu	TEMP, KK, -4				#	deal with the triangular data part
+	dsll	L,    TEMP, BASE_SHIFT		#	mr=1
+	dsll	TEMP, TEMP, 2 + BASE_SHIFT
+	daddu	AO, AORIG, L
+	daddu	BO, B, TEMP					#	BO point to the trigular data part
+
+ 	LD	b1,  0 * SIZE(AO)				#	fixed results
+ 	LD	b5,  1 * SIZE(AO)
+	LD	b3,  2 * SIZE(AO)
+	LD	b7,  3 * SIZE(AO)
+
+	SUB	t11, b1, t11
+	SUB	t12, b5, t12
+	SUB	t13, b3, t13
+	SUB	t14, b7, t14
+
+
+	LD	b1, 15 * SIZE(BO)
+	LD	b2, 14 * SIZE(BO)
+	LD	b3, 13 * SIZE(BO)
+	LD	b4, 12 * SIZE(BO)
+	MUL	t14, b1, t14
+	NMSUB	t13, t13, b2, t14
+	NMSUB	t12, t12, b3, t14
+	NMSUB	t11, t11, b4, t14
+
+	
+	LD	b5,  10 * SIZE(BO)
+	LD	b6,   9 * SIZE(BO)
+	LD	b7,   8 * SIZE(BO)
+	MUL	t13, b5, t13
+	NMSUB	t12, t12, b6, t13
+	NMSUB	t11, t11, b7, t13
+
+
+	LD	b8,   5 * SIZE(BO)
+	LD	b1,   4 * SIZE(BO)
+	MUL	t12, b8, t12
+	NMSUB	t11, t11, b1, t12
+
+
+	LD	b2,   0 * SIZE(BO)
+	MUL	t11, b2, t11
+
+
+	ST	t11,  0 * SIZE(AO)				# updata packed A
+	ST	t12,  1 * SIZE(AO)
+	ST	t13,  2 * SIZE(AO)
+	ST	t14,  3 * SIZE(AO)
+
+	ST	t11,  0 * SIZE(CO1)				#	write back
+	ST	t12,  0 * SIZE(CO2)
+	ST	t13,  0 * SIZE(CO3)
+	ST	t14,  0 * SIZE(CO4)
+
+	daddiu	CO1, CO1, 1 * SIZE			#	fixed pointer
+	daddiu	CO2, CO2, 1 * SIZE
+	daddiu	CO3, CO3, 1 * SIZE
+	daddiu	CO4, CO4, 1 * SIZE
+
+	dsll	TEMP, K, BASE_SHIFT			#	mr=2
+	daddu	AORIG, AORIG, TEMP			# 	move to next panel Ai
+
+
+.L29:
+	daddiu	KK, KK, -4					#	rectangular data part increased by 4
+	bgtz	J, .L10
+	NOP
+
+
+
+	.align 3
+
+
+.L999:
+	LDARG	$16,   0($sp)
+	LDARG	$17,   8($sp)
+	LDARG	$18,  16($sp)
+	LDARG	$19,  24($sp)
+	LDARG	$20,  32($sp)
+	LDARG	$21,  40($sp)
+	ldc1	$f24, 48($sp)
+	ldc1	$f25, 56($sp)
+	ldc1	$f26, 64($sp)
+	ldc1	$f27, 72($sp)
+	ldc1	$f28, 80($sp)
+
+	LDARG	$22,  88($sp)
+	LDARG	$23,  96($sp)
+	LDARG	$24, 104($sp)
+	LDARG	$25, 112($sp)
+
+#ifndef __64BIT__
+	ldc1	$f20,112($sp)
+	ldc1	$f21,120($sp)
+	ldc1	$f22,128($sp)
+	ldc1	$f23,136($sp)
+#endif
+
+	j	$31
+	daddiu	$sp, $sp, 144
+
+	EPILOGUE