7797 lines
134 KiB
ArmAsm
7797 lines
134 KiB
ArmAsm
#define REALNAME ASMNAME
|
|
#define ASSEMBLER
|
|
#include "common.h"
|
|
|
|
#define FETCH ld
|
|
#define STACKSIZE 160
|
|
#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)
|
|
|
|
|
|
##### Parameter registers ####
|
|
|
|
#define M $4
|
|
#define N $5
|
|
#define K $6
|
|
#define A $8
|
|
#define B $9
|
|
#define C $10
|
|
#define LDC $11
|
|
|
|
#### Pointer A, B, C ####
|
|
#define AO $12
|
|
#define BO $13
|
|
|
|
#define CO1 $14
|
|
#define CO2 $15
|
|
#define CO3 $16
|
|
#define CO4 $17
|
|
|
|
#define PREA $18
|
|
#define PREB $19
|
|
|
|
#### Used registers ####
|
|
#define A1 $f0
|
|
#define A2 $f1
|
|
#define A3 $f2
|
|
#define A4 $f3
|
|
#define A5 $f4
|
|
#define A6 $f5
|
|
#define A7 $f6
|
|
#define A8 $f7
|
|
|
|
#define B1 $f8
|
|
#define B2 $f9
|
|
#define B3 $f10
|
|
#define B4 $f11
|
|
#define B5 $f12
|
|
#define B6 $f13
|
|
#define B7 $f14
|
|
#define B8 $f15
|
|
|
|
#define C11 $f16
|
|
#define C12 $f17
|
|
#define C21 $f18
|
|
#define C22 $f19
|
|
#define C31 $f20
|
|
#define C32 $f21
|
|
#define C41 $f22
|
|
#define C42 $f23
|
|
#define C13 $f24
|
|
#define C14 $f25
|
|
#define C23 $f26
|
|
#define C24 $f27
|
|
#define C33 $f28
|
|
#define C34 $f29
|
|
#define C43 $f30
|
|
#define C44 $f31
|
|
|
|
#define I $2
|
|
#define J $3
|
|
#define L $7
|
|
|
|
#### Alpha register ####
|
|
#define ALPHA $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
|
|
|
|
#define R12 12
|
|
#define R13 13
|
|
|
|
#define R14 14
|
|
#define R15 15
|
|
#define R16 16
|
|
#define R17 17
|
|
|
|
#if defined(TRMMKERNEL)
|
|
#define OFFSET $23
|
|
#define KK $24
|
|
#define TEMP $25
|
|
#endif
|
|
|
|
# .text
|
|
# .align 2
|
|
## .globl gemm
|
|
# .set nomips16
|
|
# .ent gemm
|
|
# .type gemm, @function
|
|
#gemm:
|
|
# .frame $sp,STACKSIZE,$31 # vars= 48, regs= 1/0, args= 0, gp= 0
|
|
# .mask 0x40000000,-8
|
|
# .fmask 0x00000000,0
|
|
# .set noreorder
|
|
# .set nomacro
|
|
|
|
|
|
PROLOGUE
|
|
|
|
daddiu $sp,$sp,-STACKSIZE
|
|
|
|
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)
|
|
|
|
ST $f24, 56($sp)
|
|
ST $f25, 64($sp)
|
|
ST $f26, 72($sp)
|
|
ST $f27, 80($sp)
|
|
ST $f28, 88($sp)
|
|
|
|
#if defined(TRMMKERNEL)
|
|
sd $23, 96($sp)
|
|
sd $24, 104($sp)
|
|
sd $25, 112($sp)
|
|
|
|
LDARG OFFSET, 160($sp)
|
|
#endif
|
|
|
|
#ifndef __64BIT__
|
|
ST $f20,120($sp)
|
|
ST $f21,128($sp)
|
|
ST $f22,136($sp)
|
|
ST $f23,144($sp)
|
|
#endif
|
|
|
|
.align 4
|
|
.L4:
|
|
dsra J, N, 2 # NR=4
|
|
dsll LDC, LDC, BASE_SHIFT# LDC*SIZE
|
|
|
|
#if defined(TRMMKERNEL) && !defined(LEFT)
|
|
neg KK, OFFSET
|
|
#endif
|
|
|
|
blez J, .L2
|
|
ST ALPHA, 152($sp)
|
|
|
|
.L48:
|
|
dsra I, M, 3 # MR=8
|
|
dsll PREA, K, BASE_SHIFT
|
|
|
|
move AO, A # Reset A
|
|
move CO1, C
|
|
|
|
daddu CO2, C, LDC
|
|
daddu CO3, CO2, LDC
|
|
|
|
daddu CO4, CO3, LDC
|
|
daddu PREA, A, PREA
|
|
|
|
#if defined(TRMMKERNEL) && defined(LEFT)
|
|
move KK, OFFSET
|
|
#endif
|
|
|
|
blez I, .L44
|
|
daddu C, CO4, LDC
|
|
|
|
.align 4
|
|
.L481:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) ||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B
|
|
#else
|
|
dsll L, KK, 3 + BASE_SHIFT # kk*8mr*datasize
|
|
dsll TEMP, KK, 2 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L # AO point to the data addr
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
|
|
dsll PREB, K, BASE_SHIFT
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MOV C23, C11
|
|
FETCH $0, 0 * SIZE(CO1)
|
|
MOV C24, C11
|
|
FETCH $0, 4 * SIZE(CO1)
|
|
|
|
MOV C33, C11
|
|
FETCH $0, 0 * SIZE(CO2)
|
|
MOV C34, C11
|
|
FETCH $0, 4 * SIZE(CO2)
|
|
|
|
daddu PREB, B, PREB
|
|
MOV C43, C11
|
|
FETCH $0, 0 * SIZE(CO3)
|
|
|
|
MOV C44, C11
|
|
FETCH $0, 4 * SIZE(CO3)
|
|
|
|
PLU B3, B1, B1
|
|
FETCH $0, 0 * SIZE(CO4)
|
|
|
|
PLU B4, B2, B2
|
|
FETCH $0, 4 * SIZE(CO4)
|
|
|
|
#if (defined(LEFT) && !defined(TRANSA)) ||\
|
|
(!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK # TEMP is the length of the data part
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 8
|
|
#else
|
|
daddiu TEMP, KK, 4
|
|
#endif
|
|
dsra L, TEMP, 6
|
|
blez L, .L482
|
|
NOP
|
|
#else
|
|
# GEMM PART
|
|
move BO, B # Reset B
|
|
dsra L, K, 6 # UnRoll K=64
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
|
|
dsll PREB, K, BASE_SHIFT
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MOV C23, C11
|
|
FETCH $0, 0 * SIZE(CO1)
|
|
MOV C24, C11
|
|
FETCH $0, 4 * SIZE(CO1)
|
|
|
|
MOV C33, C11
|
|
FETCH $0, 0 * SIZE(CO2)
|
|
MOV C34, C11
|
|
FETCH $0, 4 * SIZE(CO2)
|
|
|
|
daddu PREB, B, PREB
|
|
MOV C43, C11
|
|
FETCH $0, 0 * SIZE(CO3)
|
|
|
|
MOV C44, C11
|
|
FETCH $0, 4 * SIZE(CO3)
|
|
|
|
PLU B3, B1, B1
|
|
FETCH $0, 0 * SIZE(CO4)
|
|
|
|
PLU B4, B2, B2
|
|
blez L, .L482
|
|
FETCH $0, 4 * SIZE(CO4)
|
|
#endif
|
|
|
|
.L4810:
|
|
daddiu L, L, -1
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
bgtz L, .L4810
|
|
MADPS C44, C44, A8, B8
|
|
|
|
.align 4
|
|
.L482:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 32
|
|
#else
|
|
andi L, TEMP, 32
|
|
#endif
|
|
blez L, .L483
|
|
NOP
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
|
|
.align 4
|
|
.L483:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 16
|
|
#else
|
|
andi L, TEMP, 16
|
|
#endif
|
|
blez L, .L484
|
|
NOP
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
|
|
.align 4
|
|
.L484:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 8
|
|
#else
|
|
andi L, TEMP, 8
|
|
#endif
|
|
blez L, .L485
|
|
NOP
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
|
|
.align 4
|
|
.L485:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 4
|
|
#else
|
|
andi L, TEMP, 4
|
|
#endif
|
|
blez L, .L486
|
|
NOP
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 4) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 5) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 6) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 7) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 32 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 16 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 20 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 24 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 28 * SIZE(PREA)
|
|
daddiu PREA, PREA, 32 * SIZE
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
|
|
|
|
.align 4
|
|
.L486:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 2
|
|
#else
|
|
andi L, TEMP, 2
|
|
#endif
|
|
blez L, .L487
|
|
NOP
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 8 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 16 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
PLU B7, B5, B5
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A2, B4
|
|
PLU B8, B6, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
MADPS C11, C11, A5, B5
|
|
MADPS C21, C21, A6, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C12, C12, A5, B6
|
|
MADPS C22, C22, A6, B6
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C31, C31, A7, B5
|
|
MADPS C41, C41, A8, B5
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C32, C32, A7, B6
|
|
MADPS C42, C42, A8, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C13, C13, A5, B7
|
|
MADPS C23, C23, A6, B7
|
|
daddiu PREB, PREB, 8 * SIZE
|
|
|
|
MADPS C33, C33, A7, B7
|
|
MADPS C43, C43, A8, B7
|
|
|
|
MADPS C14, C14, A5, B8
|
|
PLU B3, B1, B1
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
|
|
MADPS C24, C24, A6, B8
|
|
PLU B4, B2, B2
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C34, C34, A7, B8
|
|
MADPS C44, C44, A8, B8
|
|
daddiu PREA, PREA, 16 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L487:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L480
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
|
|
MADPS C12, C12, A1, B2
|
|
MADPS C22, C22, A2, B2
|
|
|
|
MADPS C31, C31, A3, B1
|
|
MADPS C41, C41, A4, B1
|
|
|
|
MADPS C32, C32, A3, B2
|
|
MADPS C42, C42, A4, B2
|
|
|
|
MADPS C13, C13, A1, B3
|
|
daddiu BO, BO, 4 * SIZE # 4KR*4NR
|
|
MADPS C23, C23, A2, B3
|
|
daddiu AO, AO, 8 * SIZE # 4KR*8MR
|
|
|
|
MADPS C33, C33, A3, B3
|
|
MADPS C43, C43, A4, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
MADPS C24, C24, A2, B4
|
|
|
|
MADPS C34, C34, A3, B4
|
|
MADPS C44, C44, A4, B4
|
|
|
|
|
|
|
|
.align 4
|
|
.L480: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
daddiu I, I, -1
|
|
CVTU A1, C13 # A1=C13.upper=c12
|
|
CVTU A2, C11 # A2=C11.upper=c22
|
|
|
|
CVTU A3, C23 # A3=C23.upper=c14
|
|
LD B1, 1 * SIZE(CO1)
|
|
|
|
CVTU A4, C21 # A4=C21.upper=c24
|
|
LD B2, 1 * SIZE(CO2)
|
|
|
|
CVTU A5, C33 # A5=C33.upper=c16
|
|
LD B3, 3 * SIZE(CO1)
|
|
|
|
CVTU A6, C31 # A6=C31.upper=c26
|
|
LD B4, 3 * SIZE(CO2)
|
|
|
|
CVTU A7, C43 # A7=C43.upper=c18
|
|
LD B5, 5 * SIZE(CO1)
|
|
|
|
CVTU A8, C41 # A8=C41.upper=c28
|
|
LD B6, 5 * SIZE(CO2)
|
|
|
|
MADD A1, B1, A1, ALPHA # c12
|
|
LD B7, 7 * SIZE(CO1)
|
|
|
|
MADD A2, B2, A2, ALPHA # c22
|
|
LD B1, 7 * SIZE(CO2)
|
|
|
|
MADD A3, B3, A3, ALPHA # c14
|
|
LD B2, 0 * SIZE(CO1)
|
|
|
|
MADD A4, B4, A4, ALPHA # c24
|
|
LD B3, 0 * SIZE(CO2)
|
|
|
|
MADD A5, B5, A5, ALPHA # c16
|
|
LD B4, 2 * SIZE(CO1)
|
|
|
|
MADD A6, B6, A6, ALPHA # c26
|
|
LD B5, 2 * SIZE(CO2)
|
|
|
|
MADD A7, B7, A7, ALPHA # c18
|
|
LD B6, 4 * SIZE(CO1)
|
|
|
|
MADD A8, B1, A8, ALPHA # c28
|
|
ST A1, 1 * SIZE(CO1)
|
|
|
|
MADD C11, B2, C11, ALPHA # c12
|
|
LD B7, 4 * SIZE(CO2)
|
|
|
|
MADD C13, B3, C13, ALPHA # c22
|
|
ST A2, 1 * SIZE(CO2)
|
|
|
|
MADD C21, B4, C21, ALPHA # c14
|
|
LD A1, 6 * SIZE(CO1)
|
|
|
|
MADD C23, B5, C23, ALPHA # c24
|
|
ST A3, 3 * SIZE(CO1)
|
|
|
|
MADD C31, B6, C31, ALPHA # c16
|
|
LD A2, 6 * SIZE(CO2)
|
|
|
|
MADD C33, B7, C33, ALPHA # c26
|
|
ST A4, 3 * SIZE(CO2)
|
|
|
|
ST A5, 5 * SIZE(CO1)
|
|
ST A6, 5 * SIZE(CO2)
|
|
ST A7, 7 * SIZE(CO1)
|
|
ST A8, 7 * SIZE(CO2)
|
|
|
|
MADD C41, A1, C41, ALPHA # c18
|
|
ST C11, 0 * SIZE(CO1)
|
|
|
|
MADD C43, A2, C43, ALPHA # c28
|
|
ST C13, 0 * SIZE(CO2)
|
|
|
|
ST C21, 2 * SIZE(CO1)
|
|
ST C23, 2 * SIZE(CO2)
|
|
ST C31, 4 * SIZE(CO1)
|
|
ST C33, 4 * SIZE(CO2)
|
|
ST C41, 6 * SIZE(CO1)
|
|
|
|
CVTU A1, C14 # B1=C12.upper=c42
|
|
ST C43, 6 * SIZE(CO2)
|
|
|
|
CVTU A2, C12 # B2=C14.upper=c32
|
|
LD B1, 1 * SIZE(CO3)
|
|
|
|
CVTU A3, C24 # B3=C22.upper=c44
|
|
LD B2, 1 * SIZE(CO4)
|
|
|
|
CVTU A4, C22 # B4=C24.upper=c34
|
|
LD B3, 3 * SIZE(CO3)
|
|
|
|
CVTU A5, C34 # B5=C32.upper=c46
|
|
LD B4, 3 * SIZE(CO4)
|
|
|
|
CVTU A6, C32 # B6=C24.upper=c36
|
|
LD B5, 5 * SIZE(CO3)
|
|
|
|
CVTU A7, C44 # B7=C42.upper=c48
|
|
LD B6, 5 * SIZE(CO4)
|
|
|
|
CVTU A8, C42 # A1=C44.upper=c38
|
|
LD B7, 7 * SIZE(CO3)
|
|
|
|
MADD A1, B1, A1, ALPHA # c31
|
|
LD C11, 7 * SIZE(CO4)
|
|
|
|
MADD A2, B2, A2, ALPHA
|
|
LD C13, 0 * SIZE(CO3)
|
|
|
|
MADD A3, B3, A3, ALPHA
|
|
LD C21, 0 * SIZE(CO4)
|
|
|
|
MADD A4, B4, A4, ALPHA
|
|
LD C23, 2 * SIZE(CO3)
|
|
|
|
MADD A5, B5, A5, ALPHA
|
|
LD C31, 2 * SIZE(CO4)
|
|
|
|
MADD A6, B6, A6, ALPHA
|
|
LD C33, 4 * SIZE(CO3)
|
|
|
|
MADD A7, B7, A7, ALPHA
|
|
LD C41, 4 * SIZE(CO4)
|
|
|
|
MADD A8, C11, A8, ALPHA
|
|
ST A1, 1 * SIZE(CO3)
|
|
|
|
MADD C12, C13, C12, ALPHA
|
|
LD C43, 6 * SIZE(CO3)
|
|
|
|
MADD C14, C21, C14, ALPHA
|
|
ST A2, 1 * SIZE(CO4)
|
|
|
|
MADD C22, C23, C22, ALPHA
|
|
LD B1, 6 * SIZE(CO4)
|
|
|
|
MADD C24, C31, C24, ALPHA
|
|
ST A3, 3 * SIZE(CO3)
|
|
|
|
MADD C32, C33, C32, ALPHA
|
|
ST A4, 3 * SIZE(CO4)
|
|
|
|
MADD C34, C41, C34, ALPHA
|
|
ST A5, 5 * SIZE(CO3)
|
|
|
|
MADD C42, C43, C42, ALPHA
|
|
ST A6, 5 * SIZE(CO4)
|
|
|
|
ST A7, 7 * SIZE(CO3)
|
|
NOP
|
|
|
|
MADD C44, B1, C44, ALPHA
|
|
ST A8, 7 * SIZE(CO4)
|
|
|
|
ST C12, 0 * SIZE(CO3)
|
|
ST C14, 0 * SIZE(CO4)
|
|
ST C22, 2 * SIZE(CO3)
|
|
ST C24, 2 * SIZE(CO4)
|
|
ST C32, 4 * SIZE(CO3)
|
|
ST C34, 4 * SIZE(CO4)
|
|
ST C42, 6 * SIZE(CO3)
|
|
ST C44, 6 * SIZE(CO4)
|
|
|
|
daddiu CO1, CO1, 8 * SIZE
|
|
daddiu CO2, CO2, 8 * SIZE
|
|
daddiu CO3, CO3, 8 * SIZE
|
|
bgtz I, .L481
|
|
daddiu CO4, CO4, 8 * SIZE
|
|
#else
|
|
daddiu I, I, -1
|
|
CVTU A1, C13 # A1=C13.upper=c12
|
|
CVTU A2, C11 # A2=C11.upper=c22
|
|
CVTU A3, C23 # A3=C23.upper=c14
|
|
CVTU A4, C21 # A4=C21.upper=c24
|
|
CVTU A5, C33 # A5=C33.upper=c16
|
|
CVTU A6, C31 # A6=C31.upper=c26
|
|
CVTU A7, C43 # A7=C43.upper=c18
|
|
CVTU A8, C41 # A8=C41.upper=c28
|
|
|
|
MUL A1, A1, ALPHA # c12
|
|
MUL A2, A2, ALPHA # c22
|
|
MUL A3, A3, ALPHA # c14
|
|
MUL A4, A4, ALPHA # c24
|
|
MUL A5, A5, ALPHA # c16
|
|
MUL A6, A6, ALPHA # c26
|
|
MUL A7, A7, ALPHA # c18
|
|
MUL A8, A8, ALPHA # c28
|
|
|
|
MUL C11, C11, ALPHA # c12
|
|
ST A1, 1 * SIZE(CO1)
|
|
|
|
MUL C13, C13, ALPHA # c22
|
|
ST A2, 1 * SIZE(CO2)
|
|
|
|
MUL C21, C21, ALPHA # c14
|
|
ST A3, 3 * SIZE(CO1)
|
|
|
|
MUL C23, C23, ALPHA # c24
|
|
ST A4, 3 * SIZE(CO2)
|
|
|
|
MUL C31, C31, ALPHA # c16
|
|
ST A5, 5 * SIZE(CO1)
|
|
|
|
MUL C33, C33, ALPHA # c26
|
|
ST A6, 5 * SIZE(CO2)
|
|
|
|
MUL C41, C41, ALPHA # c18
|
|
ST A7, 7 * SIZE(CO1)
|
|
|
|
MUL C43, C43, ALPHA # c28
|
|
ST A8, 7 * SIZE(CO2)
|
|
|
|
CVTU A1, C14 # B1=C12.upper=c42
|
|
ST C11, 0 * SIZE(CO1)
|
|
|
|
CVTU A2, C12 # B2=C14.upper=c32
|
|
ST C13, 0 * SIZE(CO2)
|
|
|
|
CVTU A3, C24 # B3=C22.upper=c44
|
|
ST C21, 2 * SIZE(CO1)
|
|
|
|
CVTU A4, C22 # B4=C24.upper=c34
|
|
ST C23, 2 * SIZE(CO2)
|
|
|
|
CVTU A5, C34 # B5=C32.upper=c46
|
|
ST C31, 4 * SIZE(CO1)
|
|
|
|
CVTU A6, C32 # B6=C24.upper=c36
|
|
ST C33, 4 * SIZE(CO2)
|
|
|
|
CVTU A7, C44 # B7=C42.upper=c48
|
|
ST C41, 6 * SIZE(CO1)
|
|
|
|
CVTU A8, C42 # A1=C44.upper=c38
|
|
ST C43, 6 * SIZE(CO2)
|
|
|
|
MUL A1, A1, ALPHA # c31
|
|
MUL A2, A2, ALPHA
|
|
MUL A3, A3, ALPHA
|
|
MUL A4, A4, ALPHA
|
|
MUL A5, A5, ALPHA
|
|
MUL A6, A6, ALPHA
|
|
MUL A7, A7, ALPHA
|
|
MUL A8, A8, ALPHA
|
|
|
|
MUL C12, C12, ALPHA
|
|
ST A1, 1 * SIZE(CO3)
|
|
|
|
MUL C14, C14, ALPHA
|
|
ST A2, 1 * SIZE(CO4)
|
|
|
|
MUL C22, C22, ALPHA
|
|
ST A3, 3 * SIZE(CO3)
|
|
|
|
MUL C24, C24, ALPHA
|
|
ST A4, 3 * SIZE(CO4)
|
|
|
|
MUL C32, C32, ALPHA
|
|
ST A5, 5 * SIZE(CO3)
|
|
|
|
MUL C34, C34, ALPHA
|
|
ST A6, 5 * SIZE(CO4)
|
|
|
|
MUL C42, C42, ALPHA
|
|
ST A7, 7 * SIZE(CO3)
|
|
|
|
MUL C44, C44, ALPHA
|
|
ST A8, 7 * SIZE(CO4)
|
|
|
|
ST C12, 0 * SIZE(CO3)
|
|
ST C14, 0 * SIZE(CO4)
|
|
ST C22, 2 * SIZE(CO3)
|
|
ST C24, 2 * SIZE(CO4)
|
|
ST C32, 4 * SIZE(CO3)
|
|
ST C34, 4 * SIZE(CO4)
|
|
ST C42, 6 * SIZE(CO3)
|
|
ST C44, 6 * SIZE(CO4)
|
|
|
|
daddiu CO1, CO1, 8 * SIZE
|
|
daddiu CO2, CO2, 8 * SIZE
|
|
daddiu CO3, CO3, 8 * SIZE
|
|
daddiu CO4, CO4, 8 * SIZE
|
|
|
|
#if ( defined(LEFT) && defined(TRANSA)) ||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -8
|
|
#else
|
|
daddiu TEMP, TEMP, -4
|
|
#endif
|
|
dsll L, TEMP, 3 + BASE_SHIFT
|
|
dsll TEMP, TEMP, 2 + BASE_SHIFT
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 8
|
|
#endif
|
|
|
|
bgtz I, .L481
|
|
NOP
|
|
#endif
|
|
|
|
.align 4
|
|
.L44:
|
|
andi I, M, 4 # MR=4
|
|
blez I, .L42
|
|
NOP
|
|
|
|
.align 4
|
|
.L441:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) ||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B # Reset B
|
|
#else
|
|
dsll L, KK, 2 + BASE_SHIFT
|
|
dsll TEMP, KK, 2 + BASE_SHIFT
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
|
|
dsll PREB, K, BASE_SHIFT
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
|
|
MOV C23, C11
|
|
FETCH $0, 0 * SIZE(CO1)
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
FETCH $0, 0 * SIZE(CO2)
|
|
MOV C34, C11
|
|
|
|
daddu PREB, B, PREB
|
|
MOV C43, C11
|
|
FETCH $0, 0 * SIZE(CO3)
|
|
|
|
MOV C44, C11
|
|
PLU B3, B1, B1
|
|
|
|
FETCH $0, 0 * SIZE(CO4)
|
|
PLU B4, B2, B2
|
|
|
|
#if (defined(LEFT) && !defined(TRANSA)) ||\
|
|
(!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddu TEMP, KK, 4
|
|
#else
|
|
daddu TEMP, KK, 4
|
|
#endif
|
|
dsra L, TEMP, 2
|
|
blez L, .L442
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 2 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
|
|
dsll PREB, K, BASE_SHIFT
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
|
|
MOV C23, C11
|
|
FETCH $0, 0 * SIZE(CO1)
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
FETCH $0, 0 * SIZE(CO2)
|
|
MOV C34, C11
|
|
|
|
daddu PREB, B, PREB
|
|
MOV C43, C11
|
|
FETCH $0, 0 * SIZE(CO3)
|
|
|
|
MOV C44, C11
|
|
PLU B3, B1, B1
|
|
|
|
FETCH $0, 0 * SIZE(CO4)
|
|
blez L, .L442
|
|
PLU B4, B2, B2
|
|
#endif
|
|
|
|
.L4410: #
|
|
daddiu L, L, -1
|
|
MADPS C11, C11, A1, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
|
|
MADPS C22, C22, A2, B2
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
MADPS C24, C24, A2, B4
|
|
|
|
PLU B7, B5, B5
|
|
PLU B8, B6, B6
|
|
|
|
MADPS C11, C11, A3, B5
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C21, C21, A4, B5
|
|
gsLQC1(R12, F5, F4, 2) # A5 A6
|
|
|
|
MADPS C12, C12, A3, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C22, C22, A4, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C13, C13, A3, B7
|
|
MADPS C23, C23, A4, B7
|
|
|
|
MADPS C14, C14, A3, B8
|
|
MADPS C24, C24, A4, B8
|
|
|
|
PLU B3, B1, B1
|
|
PLU B4, B2, B2
|
|
|
|
MADPS C11, C11, A5, B1
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
MADPS C21, C21, A6, B1
|
|
gsLQC1(R12, F7, F6, 3) # A7 A8
|
|
|
|
MADPS C12, C12, A5, B2
|
|
FETCH $0, 8 * SIZE(PREB)
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
|
|
MADPS C22, C22, A6, B2
|
|
FETCH $0, 8 * SIZE(PREA)
|
|
daddiu AO, AO, 16 * SIZE # 4KR*4MR
|
|
|
|
MADPS C13, C13, A5, B3
|
|
MADPS C23, C23, A6, B3
|
|
|
|
MADPS C14, C14, A5, B4
|
|
MADPS C24, C24, A6, B4
|
|
|
|
PLU B7, B5, B5
|
|
PLU B8, B6, B6
|
|
|
|
MADPS C11, C11, A7, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C21, C21, A8, B5
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MADPS C12, C12, A7, B6
|
|
FETCH $0, 12 * SIZE(PREB)
|
|
|
|
MADPS C22, C22, A8, B6
|
|
FETCH $0, 12 * SIZE(PREA)
|
|
|
|
MADPS C13, C13, A7, B7
|
|
daddiu PREA, PREA, 16 * SIZE
|
|
MADPS C23, C23, A8, B7
|
|
daddiu PREB, PREB, 16 * SIZE
|
|
|
|
MADPS C14, C14, A7, B8
|
|
MADPS C24, C24, A8, B8
|
|
|
|
PLU B3, B1, B1
|
|
bgtz L, .L4410
|
|
PLU B4, B2, B2
|
|
|
|
.align 4
|
|
.L442:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 2
|
|
#else
|
|
andi L, TEMP, 2
|
|
#endif
|
|
blez L, .L443
|
|
NOP
|
|
|
|
MADPS C11, C11, A1, B1
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C21, C21, A2, B1
|
|
gsLQC1(R12, F3, F2, 1) # A3 A4
|
|
|
|
MADPS C12, C12, A1, B2
|
|
FETCH $0, 0 * SIZE(PREB)
|
|
daddiu BO, BO, 8 * SIZE # 2KR*4NR
|
|
|
|
MADPS C22, C22, A2, B2
|
|
FETCH $0, 0 * SIZE(PREA)
|
|
daddiu AO, AO, 8 * SIZE # 2KR*4MR
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
MADPS C24, C24, A2, B4
|
|
|
|
PLU B7, B5, B5
|
|
PLU B8, B6, B6
|
|
|
|
MADPS C11, C11, A3, B5
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C21, C21, A4, B5
|
|
gsLQC1(R12, F1, F0, 0) # A5 A6
|
|
|
|
MADPS C12, C12, A3, B6
|
|
FETCH $0, 4 * SIZE(PREB)
|
|
|
|
MADPS C22, C22, A4, B6
|
|
FETCH $0, 4 * SIZE(PREA)
|
|
|
|
MADPS C13, C13, A3, B7
|
|
daddiu PREB, PREB, 8
|
|
MADPS C23, C23, A4, B7
|
|
daddiu PREA, PREA, 8
|
|
|
|
MADPS C14, C14, A3, B8
|
|
MADPS C24, C24, A4, B8
|
|
|
|
PLU B3, B1, B1
|
|
PLU B4, B2, B2
|
|
|
|
|
|
.align 4
|
|
.L443:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L440
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C21, C21, A2, B1
|
|
|
|
MADPS C12, C12, A1, B2
|
|
daddiu BO, BO, 4 * SIZE # 1KR*4NR
|
|
MADPS C22, C22, A2, B2
|
|
daddiu AO, AO, 4 * SIZE # 1KR*4MR
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C23, C23, A2, B3
|
|
|
|
MADPS C14, C14, A1, B4
|
|
MADPS C24, C24, A2, B4
|
|
|
|
|
|
.align 4
|
|
.L440:
|
|
#ifndef TRMMKERNEL
|
|
CVTU A1, C13 # A1=C13.upper=c12
|
|
LD B1, 1 * SIZE(CO1)
|
|
|
|
CVTU A2, C11 # A2=C11.upper=c22
|
|
LD B2, 1 * SIZE(CO2)
|
|
|
|
CVTU A3, C23 # A3=C23.upper=c14
|
|
LD B3, 3 * SIZE(CO1)
|
|
|
|
CVTU A4, C21 # A4=C21.upper=c24
|
|
LD B4, 3 * SIZE(CO2)
|
|
|
|
|
|
MADD A1, B1, A1, ALPHA # c12
|
|
LD B5, 0 * SIZE(CO1)
|
|
|
|
MADD A2, B2, A2, ALPHA # c22
|
|
LD B6, 0 * SIZE(CO2)
|
|
|
|
MADD A3, B3, A3, ALPHA # c14
|
|
LD B7, 2 * SIZE(CO1)
|
|
|
|
MADD A4, B4, A4, ALPHA # c24
|
|
LD B1, 2 * SIZE(CO2)
|
|
|
|
MADD C11, B5, C11, ALPHA # c12
|
|
ST A1, 1 * SIZE(CO1)
|
|
|
|
MADD C13, B6, C13, ALPHA # c22
|
|
ST A2, 1 * SIZE(CO2)
|
|
|
|
MADD C21, B7, C21, ALPHA # c14
|
|
ST A3, 3 * SIZE(CO1)
|
|
|
|
MADD C23, B1, C23, ALPHA # c24
|
|
ST A4, 3 * SIZE(CO2)
|
|
|
|
ST C11, 0 * SIZE(CO1)
|
|
ST C13, 0 * SIZE(CO2)
|
|
ST C21, 2 * SIZE(CO1)
|
|
ST C23, 2 * SIZE(CO2)
|
|
|
|
CVTU A1, C14 # B1=C12.upper=c42
|
|
LD B1, 1 * SIZE(CO3)
|
|
|
|
CVTU A2, C12 # B2=C14.upper=c32
|
|
LD B2, 1 * SIZE(CO4)
|
|
|
|
CVTU A3, C24 # B3=C22.upper=c44
|
|
LD B3, 3 * SIZE(CO3)
|
|
|
|
CVTU A4, C22 # B4=C24.upper=c34
|
|
LD B4, 3 * SIZE(CO4)
|
|
|
|
MADD A1, B1, A1, ALPHA # c31
|
|
LD A5, 0 * SIZE(CO3)
|
|
|
|
MADD A2, B2, A2, ALPHA
|
|
LD A6, 0 * SIZE(CO4)
|
|
|
|
MADD A3, B3, A3, ALPHA
|
|
LD A7, 2 * SIZE(CO3)
|
|
|
|
MADD A4, B4, A4, ALPHA
|
|
LD A8, 2 * SIZE(CO4)
|
|
|
|
MADD C12, A5, C12, ALPHA
|
|
ST A1, 1 * SIZE(CO3)
|
|
|
|
MADD C14, A6, C14, ALPHA
|
|
ST A2, 1 * SIZE(CO4)
|
|
|
|
MADD C22, A7, C22, ALPHA
|
|
ST A3, 3 * SIZE(CO3)
|
|
|
|
MADD C24, A8, C24, ALPHA
|
|
ST A4, 3 * SIZE(CO4)
|
|
|
|
ST C12, 0 * SIZE(CO3)
|
|
ST C14, 0 * SIZE(CO4)
|
|
ST C22, 2 * SIZE(CO3)
|
|
ST C24, 2 * SIZE(CO4)
|
|
|
|
daddiu CO1, CO1, 4 * SIZE
|
|
daddiu CO2, CO2, 4 * SIZE
|
|
daddiu CO3, CO3, 4 * SIZE
|
|
daddiu CO4, CO4, 4 * SIZE
|
|
|
|
#else
|
|
CVTU A1, C13 # A1=C13.upper=c12
|
|
CVTU A2, C11 # A2=C11.upper=c22
|
|
CVTU A3, C23 # A3=C23.upper=c14
|
|
CVTU A4, C21 # A4=C21.upper=c24
|
|
|
|
MUL A1, A1, ALPHA # c12
|
|
MUL A2, A2, ALPHA # c22
|
|
MUL A3, A3, ALPHA # c14
|
|
MUL A4, A4, ALPHA # c24
|
|
|
|
MUL C11, C11, ALPHA # c12
|
|
ST A1, 1 * SIZE(CO1)
|
|
|
|
MUL C13, C13, ALPHA # c22
|
|
ST A2, 1 * SIZE(CO2)
|
|
|
|
MUL C21, C21, ALPHA # c14
|
|
ST A3, 3 * SIZE(CO1)
|
|
|
|
MUL C23, C23, ALPHA # c24
|
|
ST A4, 3 * SIZE(CO2)
|
|
|
|
CVTU A5, C14 # B1=C12.upper=c42
|
|
ST C11, 0 * SIZE(CO1)
|
|
|
|
CVTU A6, C12 # B2=C14.upper=c32
|
|
ST C13, 0 * SIZE(CO2)
|
|
|
|
CVTU A7, C24 # B3=C22.upper=c44
|
|
ST C21, 2 * SIZE(CO1)
|
|
|
|
CVTU A8, C22 # B4=C24.upper=c34
|
|
ST C23, 2 * SIZE(CO2)
|
|
|
|
MUL A5, A5, ALPHA # c31
|
|
MUL A6, A6, ALPHA
|
|
MUL A7, A7, ALPHA
|
|
MUL A8, A8, ALPHA
|
|
|
|
MUL C12, C12, ALPHA
|
|
ST A5, 1 * SIZE(CO3)
|
|
|
|
MUL C14, C14, ALPHA
|
|
ST A6, 1 * SIZE(CO4)
|
|
|
|
MUL C22, C22, ALPHA
|
|
ST A7, 3 * SIZE(CO3)
|
|
|
|
MUL C24, C24, ALPHA
|
|
ST A8, 3 * SIZE(CO4)
|
|
|
|
ST C12, 0 * SIZE(CO3)
|
|
ST C14, 0 * SIZE(CO4)
|
|
ST C22, 2 * SIZE(CO3)
|
|
ST C24, 2 * SIZE(CO4)
|
|
|
|
daddiu CO1, CO1, 4 * SIZE
|
|
daddiu CO2, CO2, 4 * SIZE
|
|
daddiu CO3, CO3, 4 * SIZE
|
|
daddiu CO4, CO4, 4 * SIZE
|
|
|
|
#if ( defined(LEFT) && defined(TRANSA))||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -4
|
|
#else
|
|
daddiu TEMP, TEMP, -4
|
|
#endif
|
|
dsll L, TEMP, 2 + BASE_SHIFT
|
|
dsll TEMP, TEMP, 2 + BASE_SHIFT
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 4
|
|
#endif
|
|
#endif
|
|
|
|
.align 4
|
|
.L42:
|
|
andi I, M, 2
|
|
blez I, .L41
|
|
NOP
|
|
|
|
.align 4
|
|
.L421:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) ||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B
|
|
#else
|
|
dsll L, KK, 1 + BASE_SHIFT
|
|
dsll TEMP, KK, 2 + BASE_SHIFT
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
|
|
MOV C23, C11
|
|
FETCH $0, 0 * SIZE(CO1)
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
FETCH $0, 0 * SIZE(CO2)
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
FETCH $0, 0 * SIZE(CO3)
|
|
|
|
MOV C44, C11
|
|
PLU B3, B1, B1
|
|
|
|
FETCH $0, 0 * SIZE(CO4)
|
|
PLU B4, B2, B2
|
|
#if (defined(LEFT) && !defined(TRANSA)) ||\
|
|
(!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 2
|
|
#else
|
|
daddiu TEMP, KK, 4
|
|
#endif
|
|
dsra L, TEMP, 2
|
|
blez L, .L422
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 2 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
gsLQC1(R12, F1, F0, 0) # A1 A2
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
|
|
MOV C23, C11
|
|
FETCH $0, 0 * SIZE(CO1)
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
FETCH $0, 0 * SIZE(CO2)
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
FETCH $0, 0 * SIZE(CO3)
|
|
|
|
MOV C44, C11
|
|
PLU B3, B1, B1
|
|
|
|
FETCH $0, 0 * SIZE(CO4)
|
|
blez L, .L422
|
|
PLU B4, B2, B2
|
|
#endif
|
|
|
|
.L4210:
|
|
daddiu L, L, -1
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C12, C12, A1, B2
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C14, C14, A1, B4
|
|
gsLQC1(R12, F3, F2, 1) # B1 B2
|
|
|
|
PLU B7, B5, B5
|
|
PLU B8, B6, B6
|
|
|
|
MADPS C11, C11, A2, B5
|
|
MADPS C12, C12, A2, B6
|
|
daddiu AO, AO, 8 * SIZE # 4KR*2MR
|
|
gsLQC1(R13, F9, F8, 2) # B1 B2
|
|
|
|
MADPS C13, C13, A2, B7
|
|
MADPS C14, C14, A2, B8
|
|
|
|
PLU B3, B1, B1
|
|
PLU B4, B2, B2
|
|
|
|
MADPS C11, C11, A3, B1
|
|
gsLQC1(R12, F1, F0, 0) # B3 B4
|
|
|
|
MADPS C12, C12, A3, B2
|
|
gsLQC1(R13, F13, F12, 3) # B3 B4
|
|
|
|
daddiu BO, BO, 16 * SIZE # 4KR*4NR
|
|
MADPS C13, C13, A3, B3
|
|
MADPS C14, C14, A3, B4
|
|
|
|
PLU B7, B5, B5
|
|
PLU B8, B6, B6
|
|
|
|
MADPS C11, C11, A4, B5
|
|
MADPS C12, C12, A4, B6
|
|
gsLQC1(R13, F9, F8, 0) # B3 B4
|
|
|
|
MADPS C13, C13, A4, B7
|
|
MADPS C14, C14, A4, B8
|
|
|
|
PLU B3, B1, B1
|
|
bgtz L, .L4210
|
|
PLU B4, B2, B2
|
|
|
|
.align 4
|
|
.L422:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 2
|
|
#else
|
|
andi L, TEMP, 2
|
|
#endif
|
|
blez L, .L423
|
|
NOP
|
|
|
|
daddiu AO, AO, 4 * SIZE # 2KR*2MR
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C12, C12, A1, B2
|
|
gsLQC1(R13, F13, F12, 1) # B3 B4
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C14, C14, A1, B4
|
|
daddiu BO, BO, 8 * SIZE # 2KR*2MR
|
|
|
|
PLU B7, B5, B5
|
|
PLU B8, B6, B6
|
|
|
|
MADPS C11, C11, A2, B5
|
|
MADPS C12, C12, A2, B6
|
|
gsLQC1(R13, F9, F8, 0) # B1 B2
|
|
|
|
MADPS C13, C13, A2, B7
|
|
MADPS C14, C14, A2, B8
|
|
gsLQC1(R12, F1, F0, 0)
|
|
|
|
PLU B3, B1, B1
|
|
PLU B4, B2, B2
|
|
|
|
.L423:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L420
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADPS C11, C11, A1, B1
|
|
MADPS C12, C12, A1, B2
|
|
daddiu BO, BO, 4 * SIZE # 2KR*4NR
|
|
daddiu AO, AO, 2 * SIZE # 2KR*4MR
|
|
|
|
MADPS C13, C13, A1, B3
|
|
MADPS C14, C14, A1, B4
|
|
|
|
.align 4
|
|
.L420:
|
|
#ifndef TRMMKERNEL
|
|
CVTU A1, C13 # A1=C13.upper=c12
|
|
LD B1, 1 * SIZE(CO1)
|
|
|
|
CVTU A2, C11 # A2=C11.upper=c22
|
|
LD B2, 1 * SIZE(CO2)
|
|
|
|
MADD A1, B1, A1, ALPHA # c12
|
|
LD B5, 0 * SIZE(CO1)
|
|
|
|
MADD A2, B2, A2, ALPHA # c22
|
|
LD B6, 0 * SIZE(CO2)
|
|
|
|
MADD C11, B5, C11, ALPHA # c12
|
|
ST A1, 1 * SIZE(CO1)
|
|
|
|
MADD C13, B6, C13, ALPHA # c22
|
|
ST A2, 1 * SIZE(CO2)
|
|
|
|
ST C11, 0 * SIZE(CO1)
|
|
ST C13, 0 * SIZE(CO2)
|
|
|
|
CVTU A1, C14 # B1=C12.upper=c42
|
|
LD B1, 1 * SIZE(CO3)
|
|
|
|
CVTU A2, C12 # B2=C14.upper=c32
|
|
LD B2, 1 * SIZE(CO4)
|
|
|
|
MADD A1, B1, A1, ALPHA # c31
|
|
LD A5, 0 * SIZE(CO3)
|
|
|
|
MADD A2, B2, A2, ALPHA
|
|
LD A6, 0 * SIZE(CO4)
|
|
|
|
MADD C12, A5, C12, ALPHA
|
|
ST A1, 1 * SIZE(CO3)
|
|
|
|
MADD C14, A6, C14, ALPHA
|
|
ST A2, 1 * SIZE(CO4)
|
|
|
|
ST C12, 0 * SIZE(CO3)
|
|
ST C14, 0 * SIZE(CO4)
|
|
|
|
daddiu CO1, CO1, 2 * SIZE
|
|
daddiu CO2, CO2, 2 * SIZE
|
|
daddiu CO3, CO3, 2 * SIZE
|
|
daddiu CO4, CO4, 2 * SIZE
|
|
#else
|
|
CVTU A1, C13 # A1=C13.upper=c12
|
|
CVTU A2, C11 # A2=C11.upper=c22
|
|
|
|
MUL A1, A1, ALPHA # c12
|
|
MUL A2, A2, ALPHA # c22
|
|
|
|
MUL C11, C11, ALPHA # c12
|
|
MUL C13, C13, ALPHA # c22
|
|
|
|
CVTU A3, C14 # B1=C12.upper=c42
|
|
CVTU A4, C12 # B2=C14.upper=c32
|
|
|
|
MUL A3, A3, ALPHA # c31
|
|
ST A1, 1 * SIZE(CO1)
|
|
|
|
MUL A4, A4, ALPHA
|
|
ST A2, 1 * SIZE(CO2)
|
|
|
|
MUL C12, C12, ALPHA
|
|
ST C11, 0 * SIZE(CO1)
|
|
|
|
MUL C14, C14, ALPHA
|
|
ST C13, 0 * SIZE(CO2)
|
|
|
|
ST A3, 1 * SIZE(CO3)
|
|
ST A4, 1 * SIZE(CO4)
|
|
|
|
ST C12, 0 * SIZE(CO3)
|
|
ST C14, 0 * SIZE(CO4)
|
|
|
|
daddiu CO1, CO1, 2 * SIZE
|
|
daddiu CO2, CO2, 2 * SIZE
|
|
daddiu CO3, CO3, 2 * SIZE
|
|
daddiu CO4, CO4, 2 * SIZE
|
|
#if ( defined(LEFT) && defined(TRANSA))||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -2
|
|
#else
|
|
daddiu TEMP, TEMP, -4
|
|
#endif
|
|
dsll L, TEMP, 1 + BASE_SHIFT
|
|
dsll TEMP, TEMP, 2 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 2
|
|
#endif
|
|
#endif
|
|
|
|
|
|
.align 4
|
|
.L41:
|
|
andi I, M, 1
|
|
blez I, .L40
|
|
NOP
|
|
|
|
.align 4
|
|
.L411:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) ||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B
|
|
#else
|
|
dsll L, KK, BASE_SHIFT
|
|
dsll TEMP, KK, 2 + BASE_SHIFT
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD B3, 2 * SIZE(BO)
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
LD B4, 3 * SIZE(BO)
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
MOV C44, C11
|
|
#if (defined(LEFT) && !defined(TRANSA))||\
|
|
(!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 1
|
|
#else
|
|
daddiu TEMP, KK, 4
|
|
#endif
|
|
dsra L, TEMP, 2
|
|
blez L, .L412
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 2 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD B3, 2 * SIZE(BO)
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
LD B4, 3 * SIZE(BO)
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
blez L, .L412
|
|
MOV C44, C11
|
|
#endif
|
|
|
|
.L4110:
|
|
daddiu L, L, -1
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MADD C11, C11, A1, B1
|
|
LD B5, 4 * SIZE(BO)
|
|
|
|
MADD C12, C12, A1, B2
|
|
LD B6, 5 * SIZE(BO)
|
|
|
|
MADD C13, C13, A1, B3
|
|
LD B7, 6 * SIZE(BO)
|
|
|
|
MADD C14, C14, A1, B4
|
|
LD B8, 7 * SIZE(BO)
|
|
|
|
LD A3, 2 * SIZE(AO)
|
|
NOP
|
|
|
|
MADD C11, C11, A2, B5
|
|
LD B1, 8 * SIZE(BO)
|
|
|
|
MADD C12, C12, A2, B6
|
|
LD B2, 9 * SIZE(BO)
|
|
|
|
MADD C13, C13, A2, B7
|
|
LD B3, 10 * SIZE(BO)
|
|
|
|
MADD C14, C14, A2, B8
|
|
LD B4, 11 * SIZE(BO)
|
|
|
|
LD A4, 3 * SIZE(AO)
|
|
daddiu AO, AO, 4 * SIZE
|
|
|
|
MADD C11, C11, A3, B1
|
|
LD B5, 12 * SIZE(BO)
|
|
|
|
MADD C12, C12, A3, B2
|
|
LD B6, 13 * SIZE(BO)
|
|
|
|
MADD C13, C13, A3, B3
|
|
LD B7, 14 * SIZE(BO)
|
|
|
|
MADD C14, C14, A3, B4
|
|
LD B8, 15 * SIZE(BO)
|
|
|
|
LD A1, 0 * SIZE(AO)
|
|
daddiu BO, BO, 16 * SIZE
|
|
|
|
MADD C11, C11, A4, B5
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MADD C12, C12, A4, B6
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MADD C13, C13, A4, B7
|
|
LD B3, 2 * SIZE(BO)
|
|
|
|
MADD C14, C14, A4, B8
|
|
bgtz L, .L4110
|
|
LD B4, 3 * SIZE(BO)
|
|
|
|
.L412:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 2
|
|
#else
|
|
andi L, TEMP, 2
|
|
#endif
|
|
blez L, .L413
|
|
NOP
|
|
|
|
LD A2, 1 * SIZE(AO)
|
|
daddiu AO, AO, 2 * SIZE
|
|
|
|
MADD C11, C11, A1, B1
|
|
LD B5, 4 * SIZE(BO)
|
|
|
|
MADD C12, C12, A1, B2
|
|
LD B6, 5 * SIZE(BO)
|
|
|
|
MADD C13, C13, A1, B3
|
|
LD B7, 6 * SIZE(BO)
|
|
|
|
MADD C14, C14, A1, B4
|
|
LD B8, 7 * SIZE(BO)
|
|
|
|
LD A1, 0 * SIZE(AO)
|
|
daddiu BO, BO, 8 * SIZE
|
|
|
|
MADD C11, C11, A2, B5
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MADD C12, C12, A2, B6
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MADD C13, C13, A2, B7
|
|
LD B3, 2 * SIZE(BO)
|
|
|
|
MADD C14, C14, A2, B8
|
|
LD B4, 3 * SIZE(BO)
|
|
|
|
.L413:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L410
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C11, C11, A1, B1
|
|
MADD C12, C12, A1, B2
|
|
daddiu AO, AO, 1 * SIZE
|
|
MADD C13, C13, A1, B3
|
|
MADD C14, C14, A1, B4
|
|
daddiu BO, BO, 4 * SIZE
|
|
|
|
.align 4
|
|
.L410:
|
|
#ifndef TRMMKERNEL
|
|
LD A5, 0 * SIZE(CO1)
|
|
LD A6, 0 * SIZE(CO2)
|
|
LD A7, 0 * SIZE(CO3)
|
|
LD A8, 0 * SIZE(CO4)
|
|
|
|
MADD A5, A5, C11, ALPHA
|
|
MADD A6, A6, C12, ALPHA
|
|
MADD A7, A7, C13, ALPHA
|
|
MADD A8, A8, C14, ALPHA
|
|
|
|
ST A5, 0 * SIZE(CO1)
|
|
ST A6, 0 * SIZE(CO2)
|
|
ST A7, 0 * SIZE(CO3)
|
|
ST A8, 0 * SIZE(CO4)
|
|
|
|
daddiu CO1, CO1, 1 * SIZE
|
|
daddiu CO2, CO2, 1 * SIZE
|
|
daddiu CO3, CO3, 1 * SIZE
|
|
daddiu CO4, CO4, 1 * SIZE
|
|
#else
|
|
MUL A5, C11, ALPHA
|
|
MUL A6, C12, ALPHA
|
|
MUL A7, C13, ALPHA
|
|
MUL A8, C14, ALPHA
|
|
|
|
ST A5, 0 * SIZE(CO1)
|
|
ST A6, 0 * SIZE(CO2)
|
|
ST A7, 0 * SIZE(CO3)
|
|
ST A8, 0 * SIZE(CO4)
|
|
|
|
daddiu CO1, CO1, 1 * SIZE
|
|
daddiu CO2, CO2, 1 * SIZE
|
|
daddiu CO3, CO3, 1 * SIZE
|
|
daddiu CO4, CO4, 1 * SIZE
|
|
|
|
#if ( defined(LEFT) && defined(TRANSA))||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -1
|
|
#else
|
|
daddiu TEMP, TEMP, -4
|
|
#endif
|
|
|
|
dsll L, TEMP, BASE_SHIFT
|
|
dsll TEMP, TEMP, 2 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 1
|
|
#endif
|
|
#endif
|
|
|
|
.align 4
|
|
.L40:
|
|
#if defined(TRMMKERNEL) && !defined(LEFT)
|
|
daddiu KK, KK, 4
|
|
#endif
|
|
daddiu J, J, -1
|
|
move B, BO
|
|
bgtz J, .L48
|
|
NOP
|
|
|
|
|
|
|
|
.align 4
|
|
.L2: # Nr=2
|
|
andi J, N, 2
|
|
blez J, .L1
|
|
NOP
|
|
|
|
.L28:
|
|
dsra I, M, 3 # MR=8
|
|
|
|
move AO, A # Reset A
|
|
move CO1, C
|
|
|
|
#if defined(TRMMKERNEL) && defined(LEFT)
|
|
move KK, OFFSET
|
|
#endif
|
|
daddu CO2, C, LDC
|
|
blez I, .L24
|
|
daddu C, CO2, LDC
|
|
|
|
.align 4
|
|
.L281:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B
|
|
#else
|
|
dsll L, KK, 3 + BASE_SHIFT
|
|
dsll TEMP, KK, 1 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C12, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MOV C22, C11
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
LD A5, 4 * SIZE(AO)
|
|
|
|
MOV C32, C11
|
|
LD A6, 5 * SIZE(AO)
|
|
|
|
MOV C41, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C42, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C13, C11
|
|
LD A7, 6 * SIZE(AO)
|
|
|
|
MOV C14, C11
|
|
LD A8, 7 * SIZE(AO)
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
MOV C44, C11
|
|
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 8
|
|
#else
|
|
daddiu TEMP, KK, 2
|
|
#endif
|
|
dsra L, TEMP, 1
|
|
blez L, .L282
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 1 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C12, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MOV C22, C11
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
LD A5, 4 * SIZE(AO)
|
|
|
|
MOV C32, C11
|
|
LD A6, 5 * SIZE(AO)
|
|
|
|
MOV C41, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C42, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C13, C11
|
|
LD A7, 6 * SIZE(AO)
|
|
|
|
MOV C14, C11
|
|
LD A8, 7 * SIZE(AO)
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
blez L, .L282
|
|
MOV C44, C11
|
|
#endif
|
|
|
|
.align 4
|
|
.L2810:
|
|
daddiu L, L, -1
|
|
MADD C11, C11, A1, B1
|
|
LD B5, 8 * SIZE(AO)
|
|
|
|
MADD C21, C21, A2, B1
|
|
LD B6, 9 * SIZE(AO)
|
|
|
|
MADD C31, C31, A3, B1
|
|
LD B7, 10 * SIZE(AO)
|
|
|
|
MADD C41, C41, A4, B1
|
|
LD B8, 11 * SIZE(AO)
|
|
|
|
MADD C12, C12, A1, B2
|
|
MADD C22, C22, A2, B2
|
|
LD B3, 2 * SIZE(BO)
|
|
|
|
MADD C32, C32, A3, B2
|
|
MADD C42, C42, A4, B2
|
|
LD B4, 3 * SIZE(BO)
|
|
daddiu BO, BO, 4 * SIZE
|
|
|
|
MADD C13, C13, A5, B1
|
|
MADD C23, C23, A6, B1
|
|
LD A1, 12 * SIZE(AO)
|
|
|
|
MADD C33, C33, A7, B1
|
|
MADD C43, C43, A8, B1
|
|
LD A2, 13 * SIZE(AO)
|
|
|
|
MADD C14, C14, A5, B2
|
|
MADD C24, C24, A6, B2
|
|
LD A3, 14 * SIZE(AO)
|
|
|
|
MADD C34, C34, A7, B2
|
|
MADD C44, C44, A8, B2
|
|
LD A4, 15 * SIZE(AO)
|
|
daddiu AO, AO, 16 * SIZE
|
|
|
|
MADD C11, C11, B5, B3
|
|
LD A5, 4 * SIZE(AO)
|
|
|
|
MADD C21, C21, B6, B3
|
|
LD A6, 5 * SIZE(AO)
|
|
|
|
MADD C13, C13, A1, B3
|
|
MADD C23, C23, A2, B3
|
|
LD A7, 6 * SIZE(AO)
|
|
|
|
MADD C33, C33, A3, B3
|
|
MADD C43, C43, A4, B3
|
|
LD A8, 7 * SIZE(AO)
|
|
|
|
MADD C14, C14, A1, B4
|
|
MADD C24, C24, A2, B4
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MADD C34, C34, A3, B4
|
|
MADD C44, C44, A4, B4
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MADD C31, C31, B7, B3
|
|
MADD C41, C41, B8, B3
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MADD C12, C12, B5, B4
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MADD C22, C22, B6, B4
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
LD A4, 3 * SIZE(AO)
|
|
MADD C32, C32, B7, B4
|
|
bgtz L, .L2810
|
|
MADD C42, C42, B8, B4
|
|
|
|
.align 4
|
|
.L282:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L280
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C13, C13, A5, B1
|
|
MADD C23, C23, A6, B1
|
|
MADD C33, C33, A7, B1
|
|
MADD C43, C43, A8, B1
|
|
MADD C14, C14, A5, B2
|
|
MADD C24, C24, A6, B2
|
|
MADD C34, C34, A7, B2
|
|
MADD C44, C44, A8, B2
|
|
daddiu AO, AO, 8 * SIZE
|
|
|
|
MADD C11, C11, A1, B1
|
|
MADD C21, C21, A2, B1
|
|
MADD C31, C31, A3, B1
|
|
MADD C41, C41, A4, B1
|
|
MADD C12, C12, A1, B2
|
|
MADD C22, C22, A2, B2
|
|
MADD C32, C32, A3, B2
|
|
MADD C42, C42, A4, B2
|
|
daddiu BO, BO, 2 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L280: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
daddiu I, I, -1
|
|
|
|
LD A1, 0 * SIZE(CO1)
|
|
LD A2, 1 * SIZE(CO1)
|
|
LD A3, 2 * SIZE(CO1)
|
|
LD A4, 3 * SIZE(CO1)
|
|
LD A5, 4 * SIZE(CO1)
|
|
LD A6, 5 * SIZE(CO1)
|
|
LD A7, 6 * SIZE(CO1)
|
|
LD A8, 7 * SIZE(CO1)
|
|
|
|
MADD A1, A1, C11, ALPHA
|
|
LD B1, 0 * SIZE(CO2)
|
|
|
|
MADD A2, A2, C21, ALPHA
|
|
LD B2, 1 * SIZE(CO2)
|
|
|
|
MADD A3, A3, C31, ALPHA
|
|
LD B3, 2 * SIZE(CO2)
|
|
|
|
MADD A4, A4, C41, ALPHA
|
|
LD B4, 3 * SIZE(CO2)
|
|
|
|
MADD A5, A5, C13, ALPHA
|
|
LD B5, 4 * SIZE(CO2)
|
|
|
|
MADD A6, A6, C23, ALPHA
|
|
LD B6, 5 * SIZE(CO2)
|
|
|
|
MADD A7, A7, C33, ALPHA
|
|
LD B7, 6 * SIZE(CO2)
|
|
|
|
MADD A8, A8, C43, ALPHA
|
|
LD C11, 7 * SIZE(CO2)
|
|
|
|
MADD B1, B1, C12, ALPHA
|
|
ST A1, 0 * SIZE(CO1)
|
|
|
|
MADD B2, B2, C22, ALPHA
|
|
ST A2, 1 * SIZE(CO1)
|
|
|
|
MADD B3, B3, C32, ALPHA
|
|
ST A3, 2 * SIZE(CO1)
|
|
|
|
MADD B4, B4, C42, ALPHA
|
|
ST A4, 3 * SIZE(CO1)
|
|
|
|
MADD B5, B5, C14, ALPHA
|
|
ST A5, 4 * SIZE(CO1)
|
|
|
|
MADD B6, B6, C24, ALPHA
|
|
ST A6, 5 * SIZE(CO1)
|
|
|
|
MADD B7, B7, C34, ALPHA
|
|
ST A7, 6 * SIZE(CO1)
|
|
|
|
MADD C11, C11, C44, ALPHA
|
|
ST A8, 7 * SIZE(CO1)
|
|
|
|
ST B1, 0 * SIZE(CO2)
|
|
ST B2, 1 * SIZE(CO2)
|
|
ST B3, 2 * SIZE(CO2)
|
|
ST B4, 3 * SIZE(CO2)
|
|
ST B5, 4 * SIZE(CO2)
|
|
ST B6, 5 * SIZE(CO2)
|
|
ST B7, 6 * SIZE(CO2)
|
|
ST C11, 7 * SIZE(CO2)
|
|
|
|
daddiu CO1, CO1, 8 * SIZE
|
|
bgtz I, .L281
|
|
daddiu CO2, CO2, 8 * SIZE
|
|
#else
|
|
daddiu I, I, -1
|
|
|
|
MUL A1, C11, ALPHA
|
|
MUL A2, C21, ALPHA
|
|
MUL A3, C31, ALPHA
|
|
MUL A4, C41, ALPHA
|
|
MUL A5, C13, ALPHA
|
|
MUL A6, C23, ALPHA
|
|
MUL A7, C33, ALPHA
|
|
MUL A8, C43, ALPHA
|
|
|
|
MUL B1, C12, ALPHA
|
|
ST A1, 0 * SIZE(CO1)
|
|
|
|
MUL B2, C22, ALPHA
|
|
ST A2, 1 * SIZE(CO1)
|
|
|
|
MUL B3, C32, ALPHA
|
|
ST A3, 2 * SIZE(CO1)
|
|
|
|
MUL B4, C42, ALPHA
|
|
ST A4, 3 * SIZE(CO1)
|
|
|
|
MUL B5, C14, ALPHA
|
|
ST A5, 4 * SIZE(CO1)
|
|
|
|
MUL B6, C24, ALPHA
|
|
ST A6, 5 * SIZE(CO1)
|
|
|
|
MUL B7, C34, ALPHA
|
|
ST A7, 6 * SIZE(CO1)
|
|
|
|
MUL C11, C44, ALPHA
|
|
ST A8, 7 * SIZE(CO1)
|
|
|
|
ST B1, 0 * SIZE(CO2)
|
|
ST B2, 1 * SIZE(CO2)
|
|
ST B3, 2 * SIZE(CO2)
|
|
ST B4, 3 * SIZE(CO2)
|
|
ST B5, 4 * SIZE(CO2)
|
|
ST B6, 5 * SIZE(CO2)
|
|
ST B7, 6 * SIZE(CO2)
|
|
ST C11, 7 * SIZE(CO2)
|
|
|
|
#if ( defined(LEFT) && defined(TRANSA)) ||(!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -8
|
|
#else
|
|
daddiu TEMP, TEMP, -2
|
|
#endif
|
|
dsll L, TEMP, 3 + BASE_SHIFT
|
|
dsll TEMP, TEMP, 1 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 8
|
|
#endif
|
|
daddiu CO1, CO1, 8 * SIZE
|
|
bgtz I, .L281
|
|
daddiu CO2, CO2, 8 * SIZE
|
|
#endif
|
|
|
|
|
|
.align 4
|
|
.L24:
|
|
andi I, M, 4 # MR=4
|
|
blez I, .L22
|
|
NOP
|
|
|
|
.align 4
|
|
.L241:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B
|
|
#else
|
|
dsll L, KK, 2 + BASE_SHIFT
|
|
dsll TEMP, KK, 1 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
MOV C44, C11
|
|
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 4
|
|
#else
|
|
daddiu TEMP, KK, 2
|
|
#endif
|
|
dsra L, TEMP, 1
|
|
blez L, .L242
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 1 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
blez L, .L242
|
|
MOV C44, C11
|
|
#endif
|
|
|
|
.align 4
|
|
.L2410:
|
|
daddiu L, L, -1
|
|
MADD C11, C11, A1, B1
|
|
LD A5, 4 * SIZE(AO)
|
|
|
|
MADD C21, C21, A2, B1
|
|
LD B3, 2 * SIZE(BO)
|
|
|
|
MADD C31, C31, A3, B1
|
|
LD B4, 3 * SIZE(BO)
|
|
|
|
MADD C41, C41, A4, B1
|
|
LD A6, 5 * SIZE(AO)
|
|
daddiu BO, BO, 4 * SIZE
|
|
|
|
MADD C12, C12, A1, B2
|
|
LD A7, 6 * SIZE(AO)
|
|
|
|
MADD C22, C22, A2, B2
|
|
LD A8, 7 * SIZE(AO)
|
|
daddiu AO, AO, 8 * SIZE
|
|
|
|
MADD C32, C32, A3, B2
|
|
MADD C42, C42, A4, B2
|
|
|
|
MADD C11, C11, A5, B3
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MADD C21, C21, A6, B3
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MADD C31, C31, A7, B3
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MADD C41, C41, A8, B3
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MADD C12, C12, A5, B4
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MADD C22, C22, A6, B4
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MADD C32, C32, A7, B4
|
|
bgtz L, .L2410
|
|
MADD C42, C42, A8, B4
|
|
|
|
.align 4
|
|
.L242:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L240
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C11, C11, A1, B1
|
|
MADD C21, C21, A2, B1
|
|
MADD C31, C31, A3, B1
|
|
MADD C41, C41, A4, B1
|
|
MADD C12, C12, A1, B2
|
|
MADD C22, C22, A2, B2
|
|
MADD C32, C32, A3, B2
|
|
MADD C42, C42, A4, B2
|
|
daddiu AO, AO, 4 * SIZE
|
|
daddiu BO, BO, 2 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L240: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
LD A1, 0 * SIZE(CO1)
|
|
LD A2, 1 * SIZE(CO1)
|
|
LD A3, 2 * SIZE(CO1)
|
|
LD A4, 3 * SIZE(CO1)
|
|
|
|
MADD A1, A1, C11, ALPHA
|
|
LD B1, 0 * SIZE(CO2)
|
|
|
|
MADD A2, A2, C21, ALPHA
|
|
LD B2, 1 * SIZE(CO2)
|
|
|
|
MADD A3, A3, C31, ALPHA
|
|
LD B3, 2 * SIZE(CO2)
|
|
|
|
MADD A4, A4, C41, ALPHA
|
|
LD B4, 3 * SIZE(CO2)
|
|
|
|
MADD B1, B1, C12, ALPHA
|
|
ST A1, 0 * SIZE(CO1)
|
|
|
|
MADD B2, B2, C22, ALPHA
|
|
ST A2, 1 * SIZE(CO1)
|
|
|
|
MADD B3, B3, C32, ALPHA
|
|
ST A3, 2 * SIZE(CO1)
|
|
|
|
MADD B4, B4, C42, ALPHA
|
|
ST A4, 3 * SIZE(CO1)
|
|
|
|
ST B1, 0 * SIZE(CO2)
|
|
ST B2, 1 * SIZE(CO2)
|
|
ST B3, 2 * SIZE(CO2)
|
|
ST B4, 3 * SIZE(CO2)
|
|
|
|
daddiu CO1, CO1, 4 * SIZE
|
|
daddiu CO2, CO2, 4 * SIZE
|
|
#else
|
|
|
|
MUL A1, C11, ALPHA
|
|
MUL A2, C21, ALPHA
|
|
MUL A3, C31, ALPHA
|
|
MUL A4, C41, ALPHA
|
|
|
|
MUL B1, C12, ALPHA
|
|
ST A1, 0 * SIZE(CO1)
|
|
|
|
MUL B2, C22, ALPHA
|
|
ST A2, 1 * SIZE(CO1)
|
|
|
|
MUL B3, C32, ALPHA
|
|
ST A3, 2 * SIZE(CO1)
|
|
|
|
MUL B4, C42, ALPHA
|
|
ST A4, 3 * SIZE(CO1)
|
|
|
|
ST B1, 0 * SIZE(CO2)
|
|
ST B2, 1 * SIZE(CO2)
|
|
ST B3, 2 * SIZE(CO2)
|
|
ST B4, 3 * SIZE(CO2)
|
|
|
|
daddiu CO1, CO1, 4 * SIZE
|
|
daddiu CO2, CO2, 4 * SIZE
|
|
#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -4
|
|
#else
|
|
daddiu TEMP, TEMP, -2
|
|
#endif
|
|
dsll L, TEMP, 2 + BASE_SHIFT
|
|
dsll TEMP, TEMP, 1 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 4
|
|
#endif
|
|
#endif
|
|
|
|
.align 4
|
|
.L22:
|
|
andi I, M, 2
|
|
blez I, .L21
|
|
NOP
|
|
|
|
.align 4
|
|
.L221:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B
|
|
#else
|
|
dsll L, KK, 1 + BASE_SHIFT
|
|
dsll TEMP, KK, 1 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C43, C11
|
|
MOV C44, C11
|
|
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 2
|
|
#else
|
|
daddiu TEMP, KK, 2
|
|
#endif
|
|
dsra L, TEMP, 1
|
|
blez L, .L222
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 1 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C43, C11
|
|
blez L, .L222
|
|
MOV C44, C11
|
|
#endif
|
|
|
|
|
|
.align 4
|
|
.L2210:
|
|
daddiu L, L, -1
|
|
MADD C11, C11, A1, B1
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MADD C21, C21, A2, B1
|
|
LD B3, 2 * SIZE(BO)
|
|
|
|
MADD C12, C12, A1, B2
|
|
LD A4, 3 * SIZE(AO)
|
|
daddiu AO, AO, 4 * SIZE
|
|
|
|
MADD C22, C22, A2, B2
|
|
LD B4, 3 * SIZE(BO)
|
|
daddiu BO, BO, 4 * SIZE
|
|
|
|
MADD C11, C11, A3, B3
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MADD C21, C21, A4, B3
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MADD C12, C12, A3, B4
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MADD C22, C22, A4, B4
|
|
bgtz L, .L2210
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
|
|
.align 4
|
|
.L222:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L220
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C11, C11, A1, B1
|
|
MADD C21, C21, A2, B1
|
|
MADD C12, C12, A1, B2
|
|
MADD C22, C22, A2, B2
|
|
daddiu AO, AO, 2 * SIZE
|
|
daddiu BO, BO, 2 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L220: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
LD A1, 0 * SIZE(CO1)
|
|
LD A2, 1 * SIZE(CO1)
|
|
|
|
MADD A1, A1, C11, ALPHA
|
|
LD B1, 0 * SIZE(CO2)
|
|
|
|
MADD A2, A2, C21, ALPHA
|
|
LD B2, 1 * SIZE(CO2)
|
|
|
|
MADD B1, B1, C12, ALPHA
|
|
ST A1, 0 * SIZE(CO1)
|
|
|
|
MADD B2, B2, C22, ALPHA
|
|
ST A2, 1 * SIZE(CO1)
|
|
|
|
ST B1, 0 * SIZE(CO2)
|
|
ST B2, 1 * SIZE(CO2)
|
|
|
|
daddiu CO1, CO1, 2 * SIZE
|
|
daddiu CO2, CO2, 2 * SIZE
|
|
#else
|
|
|
|
MUL A1, C11, ALPHA
|
|
MUL A2, C21, ALPHA
|
|
MUL B1, C12, ALPHA
|
|
MUL B2, C22, ALPHA
|
|
|
|
ST A1, 0 * SIZE(CO1)
|
|
ST A2, 1 * SIZE(CO1)
|
|
ST B1, 0 * SIZE(CO2)
|
|
ST B2, 1 * SIZE(CO2)
|
|
|
|
daddiu CO1, CO1, 2 * SIZE
|
|
daddiu CO2, CO2, 2 * SIZE
|
|
|
|
#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -2
|
|
#else
|
|
daddiu TEMP, TEMP, -2
|
|
#endif
|
|
dsll L, TEMP, 1 + BASE_SHIFT
|
|
dsll TEMP, TEMP, 1 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddu KK, KK, 2
|
|
#endif
|
|
#endif
|
|
|
|
.align 4
|
|
.L21:
|
|
andi I, M, 1
|
|
blez I, .L20
|
|
NOP
|
|
|
|
.align 4
|
|
.L211:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B # Reset B
|
|
#else
|
|
dsll L, KK, BASE_SHIFT
|
|
dsll TEMP, KK, 1 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C43, C11
|
|
MOV C44, C11
|
|
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 1
|
|
#else
|
|
daddiu TEMP, KK, 2
|
|
#endif
|
|
dsra L, TEMP, 1
|
|
blez L, .L212
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 1 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
MOV C43, C11
|
|
blez L, .L212
|
|
MOV C44, C11
|
|
#endif
|
|
|
|
.align 4
|
|
.L2110:
|
|
daddiu L, L, -1
|
|
MADD C11, C11, A1, B1
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MADD C12, C12, A1, B2
|
|
LD B3, 2 * SIZE(BO)
|
|
|
|
LD B4, 3 * SIZE(BO)
|
|
daddiu AO, AO, 2 * SIZE
|
|
daddiu BO, BO, 4 * SIZE
|
|
|
|
MADD C11, C11, A2, B3
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MADD C12, C12, A2, B4
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
bgtz L, .L2110
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
|
|
.align 4
|
|
.L212:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L210
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C11, C11, A1, B1
|
|
MADD C12, C12, A1, B2
|
|
daddiu AO, AO, 1 * SIZE
|
|
daddiu BO, BO, 2 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L210: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
LD A1, 0 * SIZE(CO1)
|
|
|
|
MADD A1, A1, C11, ALPHA
|
|
LD B1, 0 * SIZE(CO2)
|
|
|
|
MADD B1, B1, C12, ALPHA
|
|
ST A1, 0 * SIZE(CO1)
|
|
|
|
ST B1, 0 * SIZE(CO2)
|
|
|
|
daddiu CO1, CO1, 1 * SIZE
|
|
daddiu CO2, CO2, 1 * SIZE
|
|
#else
|
|
|
|
MUL A1, C11, ALPHA
|
|
MUL B1, C12, ALPHA
|
|
|
|
ST A1, 0 * SIZE(CO1)
|
|
ST B1, 0 * SIZE(CO2)
|
|
|
|
daddiu CO1, CO1, 1 * SIZE
|
|
daddiu CO2, CO2, 1 * SIZE
|
|
#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -1
|
|
#else
|
|
daddiu TEMP, TEMP, -2
|
|
#endif
|
|
dsll L, TEMP, BASE_SHIFT
|
|
dsll TEMP, TEMP, 1 + BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 1
|
|
#endif
|
|
#endif
|
|
|
|
|
|
.align 4
|
|
.L20:
|
|
#if defined(TRMMKERNEL) && !defined(LEFT)
|
|
daddiu KK, KK, 2
|
|
#endif
|
|
move B, BO
|
|
|
|
|
|
|
|
.align 4
|
|
.L1:
|
|
andi J, N, 1
|
|
blez J, .L999
|
|
NOP
|
|
|
|
.L18:
|
|
dsra I, M, 3 # MR=8
|
|
move AO, A # Reset A
|
|
|
|
#if defined(TRMMKERNEL) && defined(LEFT)
|
|
move KK, OFFSET
|
|
#endif
|
|
blez I, .L14
|
|
NOP
|
|
|
|
|
|
.align 4
|
|
.L181:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B # Reset B
|
|
#else
|
|
dsll L, KK, 3 + BASE_SHIFT
|
|
dsll TEMP, KK, BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C12, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MOV C22, C11
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
LD A5, 4 * SIZE(AO)
|
|
|
|
MOV C32, C11
|
|
LD A6, 5 * SIZE(AO)
|
|
|
|
MOV C41, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C42, C11
|
|
LD A7, 6 * SIZE(AO)
|
|
|
|
MOV C13, C11
|
|
LD A8, 7 * SIZE(AO)
|
|
|
|
MOV C14, C11
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
MOV C44, C11
|
|
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 8
|
|
#else
|
|
daddiu TEMP, KK, 1
|
|
#endif
|
|
dsra L, TEMP, 1
|
|
blez L, .L182
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 1 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C12, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MOV C22, C11
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
LD A5, 4 * SIZE(AO)
|
|
|
|
MOV C32, C11
|
|
LD A6, 5 * SIZE(AO)
|
|
|
|
MOV C41, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C42, C11
|
|
LD A7, 6 * SIZE(AO)
|
|
|
|
MOV C13, C11
|
|
LD A8, 7 * SIZE(AO)
|
|
|
|
MOV C14, C11
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
blez L, .L182
|
|
MOV C44, C11
|
|
#endif
|
|
|
|
|
|
.align 4
|
|
.L1810:
|
|
daddiu L, L, -1
|
|
MADD C11, C11, A1, B1
|
|
LD B5, 8 * SIZE(AO)
|
|
|
|
MADD C21, C21, A2, B1
|
|
LD B6, 9 * SIZE(AO)
|
|
|
|
MADD C31, C31, A3, B1
|
|
LD B7, 10 * SIZE(AO)
|
|
|
|
MADD C41, C41, A4, B1
|
|
LD B8, 11 * SIZE(AO)
|
|
|
|
MADD C13, C13, A5, B1
|
|
LD B2, 1 * SIZE(BO)
|
|
daddiu BO, BO, 2 * SIZE
|
|
|
|
MADD C23, C23, A6, B1
|
|
LD A1, 12 * SIZE(AO)
|
|
|
|
MADD C33, C33, A7, B1
|
|
LD A2, 13 * SIZE(AO)
|
|
|
|
MADD C43, C43, A8, B1
|
|
LD A3, 14 * SIZE(AO)
|
|
|
|
LD A4, 15 * SIZE(AO)
|
|
daddiu AO, AO, 16 * SIZE
|
|
|
|
MADD C11, C11, B5, B2
|
|
LD A5, 4 * SIZE(AO)
|
|
|
|
MADD C21, C21, B6, B2
|
|
LD A6, 5 * SIZE(AO)
|
|
|
|
MADD C13, C13, A1, B2
|
|
LD A7, 6 * SIZE(AO)
|
|
|
|
MADD C23, C23, A2, B2
|
|
LD A8, 7 * SIZE(AO)
|
|
|
|
MADD C33, C33, A3, B2
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MADD C43, C43, A4, B2
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MADD C31, C31, B7, B2
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MADD C41, C41, B8, B2
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
bgtz L, .L1810
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
.align 4
|
|
.L182:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L180
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C13, C13, A5, B1
|
|
MADD C23, C23, A6, B1
|
|
MADD C33, C33, A7, B1
|
|
MADD C43, C43, A8, B1
|
|
daddiu AO, AO, 8 * SIZE
|
|
|
|
MADD C11, C11, A1, B1
|
|
MADD C21, C21, A2, B1
|
|
MADD C31, C31, A3, B1
|
|
MADD C41, C41, A4, B1
|
|
daddiu BO, BO, 1 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L180: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
daddiu I, I, -1
|
|
|
|
LD A1, 0 * SIZE(C)
|
|
LD A2, 1 * SIZE(C)
|
|
LD A3, 2 * SIZE(C)
|
|
LD A4, 3 * SIZE(C)
|
|
LD A5, 4 * SIZE(C)
|
|
LD A6, 5 * SIZE(C)
|
|
LD A7, 6 * SIZE(C)
|
|
LD A8, 7 * SIZE(C)
|
|
|
|
MADD A1, A1, C11, ALPHA
|
|
MADD A2, A2, C21, ALPHA
|
|
MADD A3, A3, C31, ALPHA
|
|
MADD A4, A4, C41, ALPHA
|
|
MADD A5, A5, C13, ALPHA
|
|
MADD A6, A6, C23, ALPHA
|
|
MADD A7, A7, C33, ALPHA
|
|
MADD A8, A8, C43, ALPHA
|
|
|
|
ST A1, 0 * SIZE(C)
|
|
ST A2, 1 * SIZE(C)
|
|
ST A3, 2 * SIZE(C)
|
|
ST A4, 3 * SIZE(C)
|
|
ST A5, 4 * SIZE(C)
|
|
ST A6, 5 * SIZE(C)
|
|
ST A7, 6 * SIZE(C)
|
|
ST A8, 7 * SIZE(C)
|
|
|
|
daddiu C, C, 8 * SIZE
|
|
bgtz I, .L181
|
|
NOP
|
|
#else
|
|
daddiu I, I, -1
|
|
|
|
MUL A1, C11, ALPHA
|
|
MUL A2, C21, ALPHA
|
|
MUL A3, C31, ALPHA
|
|
MUL A4, C41, ALPHA
|
|
MUL A5, C13, ALPHA
|
|
MUL A6, C23, ALPHA
|
|
MUL A7, C33, ALPHA
|
|
MUL A8, C43, ALPHA
|
|
|
|
ST A1, 0 * SIZE(C)
|
|
ST A2, 1 * SIZE(C)
|
|
ST A3, 2 * SIZE(C)
|
|
ST A4, 3 * SIZE(C)
|
|
ST A5, 4 * SIZE(C)
|
|
ST A6, 5 * SIZE(C)
|
|
ST A7, 6 * SIZE(C)
|
|
ST A8, 7 * SIZE(C)
|
|
|
|
#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -8
|
|
#else
|
|
daddiu TEMP, TEMP, -1
|
|
#endif
|
|
|
|
dsll L, TEMP, 3 + BASE_SHIFT
|
|
dsll TEMP, TEMP, BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 8
|
|
#endif
|
|
|
|
daddiu C, C, 8 * SIZE
|
|
bgtz I, .L181
|
|
NOP
|
|
#endif
|
|
|
|
.align 4
|
|
.L14:
|
|
andi I, M, 4 # MR=4
|
|
blez I, .L12
|
|
NOP
|
|
|
|
.align 4
|
|
.L141:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B
|
|
#else
|
|
dsll L, KK, 2 + BASE_SHIFT
|
|
dsll TEMP, KK, BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
MOV C44, C11
|
|
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 4
|
|
#else
|
|
daddiu TEMP, KK, 1
|
|
#endif
|
|
dsra L, TEMP, 1
|
|
blez L, .L142
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 1 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
MOV C13, C11
|
|
MOV C14, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C23, C11
|
|
MOV C24, C11
|
|
|
|
MOV C33, C11
|
|
MOV C34, C11
|
|
|
|
MOV C43, C11
|
|
blez L, .L142
|
|
MOV C44, C11
|
|
#endif
|
|
|
|
.align 4
|
|
.L1410:
|
|
daddiu L, L, -1
|
|
MADD C11, C11, A1, B1
|
|
LD A5, 4 * SIZE(AO)
|
|
|
|
MADD C21, C21, A2, B1
|
|
LD B3, 1 * SIZE(BO)
|
|
|
|
MADD C31, C31, A3, B1
|
|
LD A6, 5 * SIZE(AO)
|
|
daddiu BO, BO, 2 * SIZE
|
|
|
|
MADD C41, C41, A4, B1
|
|
LD A7, 6 * SIZE(AO)
|
|
|
|
LD A8, 7 * SIZE(AO)
|
|
daddiu AO, AO, 8 * SIZE
|
|
|
|
|
|
MADD C11, C11, A5, B3
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MADD C21, C21, A6, B3
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MADD C31, C31, A7, B3
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MADD C41, C41, A8, B3
|
|
LD A3, 2 * SIZE(AO)
|
|
|
|
bgtz L, .L1410
|
|
LD A4, 3 * SIZE(AO)
|
|
|
|
.align 4
|
|
.L142:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L140
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C11, C11, A1, B1
|
|
MADD C21, C21, A2, B1
|
|
MADD C31, C31, A3, B1
|
|
MADD C41, C41, A4, B1
|
|
daddiu AO, AO, 4 * SIZE
|
|
daddiu BO, BO, 1 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L140: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
LD A1, 0 * SIZE(C)
|
|
LD A2, 1 * SIZE(C)
|
|
LD A3, 2 * SIZE(C)
|
|
LD A4, 3 * SIZE(C)
|
|
|
|
MADD A1, A1, C11, ALPHA
|
|
MADD A2, A2, C21, ALPHA
|
|
MADD A3, A3, C31, ALPHA
|
|
MADD A4, A4, C41, ALPHA
|
|
|
|
ST A1, 0 * SIZE(C)
|
|
ST A2, 1 * SIZE(C)
|
|
ST A3, 2 * SIZE(C)
|
|
ST A4, 3 * SIZE(C)
|
|
daddiu C, C, 4 * SIZE
|
|
#else
|
|
MUL A1, C11, ALPHA
|
|
MUL A2, C21, ALPHA
|
|
MUL A3, C31, ALPHA
|
|
MUL A4, C41, ALPHA
|
|
|
|
ST A1, 0 * SIZE(C)
|
|
ST A2, 1 * SIZE(C)
|
|
ST A3, 2 * SIZE(C)
|
|
ST A4, 3 * SIZE(C)
|
|
daddiu C, C, 4 * SIZE
|
|
|
|
#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -4
|
|
#else
|
|
daddiu TEMP, TEMP, -1
|
|
#endif
|
|
|
|
dsll L, TEMP, 2 + BASE_SHIFT
|
|
dsll TEMP, TEMP, BASE_SHIFT
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 4
|
|
#endif
|
|
#endif
|
|
|
|
.align 4
|
|
.L12:
|
|
andi I, M, 2
|
|
blez I, .L11
|
|
NOP
|
|
|
|
.align 4
|
|
.L121:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA)) ||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B # Reset B
|
|
#else
|
|
dsll L, KK, 1 + BASE_SHIFT
|
|
dsll TEMP, KK, BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, B, TEMP
|
|
#endif
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
|
|
MOV C43, C11
|
|
MOV C44, C11
|
|
#if (defined(LEFT) && !defined(TRANSA)) ||\
|
|
(!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 2
|
|
#else
|
|
daddiu TEMP, KK, 1
|
|
#endif
|
|
dsra L, TEMP, 1
|
|
blez L, .L122
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 1 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C41, C11
|
|
MOV C42, C11
|
|
|
|
MOV C43, C11
|
|
blez L, .L122
|
|
MOV C44, C11
|
|
#endif
|
|
|
|
.align 4
|
|
.L1210:
|
|
daddiu L, L, -1
|
|
MADD C11, C11, A1, B1
|
|
LD B3, 1 * SIZE(BO)
|
|
|
|
MADD C21, C21, A2, B1
|
|
daddiu BO, BO, 2 * SIZE
|
|
|
|
LD A3, 2 * SIZE(AO)
|
|
LD A4, 3 * SIZE(AO)
|
|
daddiu AO, AO, 4 * SIZE
|
|
|
|
MADD C11, C11, A3, B3
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MADD C21, C21, A4, B3
|
|
LD A1, 0 * SIZE(AO)
|
|
bgtz L, .L1210
|
|
LD A2, 1 * SIZE(AO)
|
|
|
|
|
|
.align 4
|
|
.L122:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L120
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C11, C11, A1, B1
|
|
MADD C21, C21, A2, B1
|
|
daddiu AO, AO, 2 * SIZE
|
|
daddiu BO, BO, 1 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L120: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
LD A1, 0 * SIZE(C)
|
|
LD A2, 1 * SIZE(C)
|
|
|
|
MADD A1, A1, C11, ALPHA
|
|
MADD A2, A2, C21, ALPHA
|
|
|
|
ST A1, 0 * SIZE(C)
|
|
ST A2, 1 * SIZE(C)
|
|
|
|
daddiu C, C, 2 * SIZE
|
|
#else
|
|
MUL A1, C11, ALPHA
|
|
MUL A2, C21, ALPHA
|
|
|
|
ST A1, 0 * SIZE(C)
|
|
ST A2, 1 * SIZE(C)
|
|
|
|
daddiu C, C, 2 * SIZE
|
|
#if ( defined(LEFT) && defined(TRANSA))||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#ifdef LEFT
|
|
daddiu TEMP, TEMP, -2
|
|
#else
|
|
daddiu TEMP, TEMP, -1
|
|
#endif
|
|
dsll L, TEMP, 1 + BASE_SHIFT
|
|
dsll TEMP, TEMP, BASE_SHIFT
|
|
|
|
daddu AO, AO, L
|
|
daddu BO, BO, TEMP
|
|
#endif
|
|
|
|
#ifdef LEFT
|
|
daddiu KK, KK, 2
|
|
#endif
|
|
#endif
|
|
|
|
.align 4
|
|
.L11:
|
|
andi I, M, 1
|
|
blez I, .L10
|
|
NOP
|
|
|
|
.align 4
|
|
.L111:
|
|
#if defined(TRMMKERNEL)
|
|
#if (defined(LEFT) && defined(TRANSA))||\
|
|
(!defined(LEFT) && !defined(TRANSA))
|
|
move BO, B
|
|
#else
|
|
dsll L, KK, BASE_SHIFT
|
|
daddu AO, AO, L
|
|
daddu BO, B, L
|
|
#endif
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C31, C11
|
|
MOV C32, C11
|
|
#if (defined(LEFT) && !defined(TRANSA))||\
|
|
(!defined(LEFT) && defined(TRANSA))
|
|
dsubu TEMP, K, KK
|
|
#elif defined(LEFT)
|
|
daddiu TEMP, KK, 1
|
|
#else
|
|
daddiu TEMP, KK, 1
|
|
#endif
|
|
dsra L, TEMP, 1
|
|
blez L, .L112
|
|
NOP
|
|
|
|
#else
|
|
move BO, B # Reset B
|
|
dsra L, K, 1 # UnRoll K=4
|
|
|
|
MTC $0, C11 # CLEAR REAULTS REGISTERS
|
|
MOV C12, C11
|
|
LD A1, 0 * SIZE(AO)
|
|
|
|
MOV C21, C11
|
|
MOV C22, C11
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
MOV C31, C11
|
|
blez L, .L112
|
|
MOV C32, C11
|
|
#endif
|
|
|
|
|
|
.align 4
|
|
.L1110:
|
|
daddiu L, L, -1
|
|
MADD C11, C11, A1, B1
|
|
|
|
LD A2, 1 * SIZE(AO)
|
|
LD B2, 1 * SIZE(BO)
|
|
|
|
daddiu AO, AO, 2 * SIZE
|
|
daddiu BO, BO, 2 * SIZE
|
|
|
|
MADD C11, C11, A2, B2
|
|
LD A1, 0 * SIZE(AO)
|
|
LD B1, 0 * SIZE(BO)
|
|
|
|
bgtz L, .L1110
|
|
NOP
|
|
|
|
|
|
.align 4
|
|
.L112:
|
|
#ifndef TRMMKERNEL
|
|
andi L, K, 1
|
|
#else
|
|
andi L, TEMP, 1
|
|
#endif
|
|
blez L, .L110
|
|
LD ALPHA, 152($sp)
|
|
|
|
MADD C11, C11, A1, B1
|
|
daddiu AO, AO, 1 * SIZE
|
|
daddiu BO, BO, 1 * SIZE
|
|
|
|
|
|
.align 4
|
|
.L110: # Write Back
|
|
#ifndef TRMMKERNEL
|
|
LD A1, 0 * SIZE(C)
|
|
|
|
MADD A1, A1, C11, ALPHA
|
|
|
|
ST A1, 0 * SIZE(C)
|
|
|
|
daddiu C, C, 1 * SIZE
|
|
#else
|
|
MUL A1, C11, ALPHA
|
|
|
|
ST A1, 0 * SIZE(C)
|
|
|
|
daddiu C, C, 1 * SIZE
|
|
|
|
#endif
|
|
|
|
.align 4
|
|
.L10:
|
|
move B, BO
|
|
NOP
|
|
|
|
.L999:
|
|
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)
|
|
|
|
LD $f24, 56($sp)
|
|
LD $f25, 64($sp)
|
|
LD $f26, 72($sp)
|
|
LD $f27, 80($sp)
|
|
LD $f28, 88($sp)
|
|
|
|
#if defined(TRMMKERNEL)
|
|
ld $23, 96($sp)
|
|
ld $24, 104($sp)
|
|
ld $25, 112($sp)
|
|
#endif
|
|
|
|
#ifndef __64BIT__
|
|
LD $f20,120($sp)
|
|
LD $f21,128($sp)
|
|
LD $f22,136($sp)
|
|
LD $f23,144($sp)
|
|
#endif
|
|
|
|
daddiu $sp,$sp,STACKSIZE
|
|
j $31
|
|
nop
|
|
|
|
EPILOGUE
|
|
# .set macro
|
|
# .set reorder
|
|
# .end gemm
|
|
# .size gemm, .-gemm
|
|
# .ident "GCC: (Debian 4.4.6-6) 4.4.6"
|