#define R16 16
#define R17 17
+#if defined(TRMMKERNEL)
+#define OFFSET $23
+#define KK $24
+#define TEMP $25
+#endif
+
# .text
# .align 2
## .globl gemm
.L4:
dsra J, N, 2 # NR=4
dsll LDC, LDC, BASE_SHIFT# LDC*SIZE
+
+#if defined(TRMMKERNEL)
+ LD OFFSET, 192($fp)
+#endif
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+ neg KK, OFFSET
+#endif
+
blez J, .L2
ST ALPHA, 152($fp)
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
PLU B4, B2, B2
blez L, .L482
FETCH $0, 4 * SIZE(CO4)
+#endif
.L4810:
daddiu L, L, -1
.align 4
.L482:
+#ifndef TRMMKERNEL
andi L, K, 32
+#else
+ andi L, TEMP, 32
+#endif
blez L, .L483
NOP
.align 4
.L483:
+#ifndef TRMMKERNEL
andi L, K, 16
+#else
+ andi L, TEMP, 16
+#endif
blez L, .L484
NOP
.align 4
.L484:
+#ifndef TRMMKERNEL
andi L, K, 8
+#else
+ andi L, TEMP, 8
+#endif
blez L, .L485
NOP
.align 4
.L485:
+#ifndef TRMMKERNEL
andi L, K, 4
+#else
+ andi L, TEMP, 4
+#endif
blez L, .L486
NOP
.align 4
.L486:
+#ifndef TRMMKERNEL
andi L, K, 2
+#else
+ andi L, TEMP, 2
+#endif
blez L, .L487
NOP
.align 4
.L487:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L480
LD ALPHA, 152($fp)
.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
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
+#endif
.align 4
.L44:
.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
FETCH $0, 0 * SIZE(CO3)
MOV C44, C11
-
PLU B3, B1, B1
- FETCH $0, 0 * SIZE(CO4)
- PLU B4, B2, B2
+ FETCH $0, 0 * SIZE(CO4)
blez L, .L442
- NOP
+ PLU B4, B2, B2
+#endif
.L4410: #
daddiu L, L, -1
.align 4
.L442:
+#ifndef TRMMKERNEL
andi L, K, 2
+#else
+ andi L, TEMP, 2
+#endif
blez L, .L443
NOP
.align 4
.L443:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L440
LD ALPHA, 152($fp)
.align 4
.L440:
+#ifndef TRMMKERNEL
CVTU A1, C13 # A1=C13.upper=c12
LD B1, 1 * SIZE(CO1)
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:
.align 4
.L421:
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=4
-
+#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
FETCH $0, 0 * SIZE(CO3)
MOV C44, C11
-
PLU B3, B1, B1
- FETCH $0, 0 * SIZE(CO4)
+ 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
-.L4210:
- daddiu L, L, -1
- MADPS C11, C11, A1, B1
- MADPS C12, C12, A1, B2
- gsLQC1(R13, F13, F12, 1) # B3 B4
+#else
+ move BO, B # Reset B
+ dsra L, K, 2 # UnRoll K=4
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
- gsLQC1(R12, F3, F2, 1) # B1 B2
+ MTC $0, C11 # CLEAR REAULTS REGISTERS
+ MOV C12, C11
- PLU B7, B5, B5
- PLU B8, B6, B6
+ MOV C21, C11
+ MOV C22, C11
+
+ MOV C31, C11
+ MOV C32, C11
+ gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A2, B5
+ 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 $0, 0 * SIZE(CO4)
+ blez L, .L422
+ PLU B4, B2, B2
+#endif
+
+.L4210:
+ daddiu L, L, -1
+ MADPS C11, C11, A1, B1
+ MADPS C12, C12, A1, B2
+ gsLQC1(R13, F13, F12, 1) # B3 B4
+
+ MADPS C13, C13, A1, B3
+ MADPS C14, C14, A1, B4
+ gsLQC1(R12, F3, F2, 1) # B1 B2
+
+ PLU B7, B5, B5
+ PLU B8, B6, B6
+
+ MADPS C11, C11, A2, B5
MADPS C12, C12, A2, B6
daddiu AO, AO, 8 * SIZE # 4KR*2MR
gsLQC1(R13, F9, F8, 2) # B1 B2
.align 4
.L422:
+#ifndef TRMMKERNEL
andi L, K, 2
+#else
+ andi L, TEMP, 2
+#endif
blez L, .L423
NOP
PLU B4, B2, B2
.L423:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L420
LD ALPHA, 152($fp)
.align 4
.L420:
+#ifndef TRMMKERNEL
CVTU A1, C13 # A1=C13.upper=c12
LD B1, 1 * SIZE(CO1)
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
.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 $0, C11 # CLEAR REAULTS REGISTERS
+ MOV C12, C11
+ LD B1, 0 * SIZE(BO)
+
+ MOV C21, C11
+ MOV C22, C11
+ LD A1, 0 * SIZE(AO)
+
+ MOV C31, C11
+ MOV C32, C11
+ LD B2, 1 * SIZE(BO)
+
+ MOV C41, C11
+ MOV C42, C11
+ LD B3, 2 * SIZE(BO)
+
+ MOV C13, C11
+ MOV C14, C11
+ LD B4, 3 * SIZE(BO)
+
+ MOV C23, C11
+ MOV C24, C11
+
+ MOV C33, C11
+ MOV C34, C11
+
+ MOV C43, C11
+ MOV C44, C11
+#if (defined(LEFT) && !defined(TRANSA))||\
+ (!defined(LEFT) && defined(TRANSA))
+ dsubu TEMP, K, KK
+#elif defined(LEFT)
+ daddiu TEMP, KK, 1
+#else
+ daddiu TEMP, KK, 4
+#endif
+ dsra L, TEMP, 2
+ blez L, .L412
+
+#else
move BO, B # Reset B
dsra L, K, 2 # UnRoll K=4
MOV C43, C11
blez L, .L412
MOV C44, C11
+#endif
.L4110:
daddiu L, L, -1
LD B4, 3 * SIZE(BO)
.L412:
+#ifndef TRMMKERNEL
andi L, K, 2
+#else
+ andi L, TEMP, 2
+#endif
blez L, .L413
NOP
LD B4, 3 * SIZE(BO)
.L413:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L410
LD ALPHA, 152($fp)
.align 4
.L410:
+#ifndef TRMMKERNEL
LD A5, 0 * SIZE(CO1)
LD A6, 0 * SIZE(CO2)
LD A7, 0 * SIZE(CO3)
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
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, 2 + BASE_SHIFT
+
+ daddu AO, AO, L
+ daddu BO, B, TEMP
+#endif
+ MTC $0, C11 # CLEAR REAULTS REGISTERS
+ LD A1, 0 * SIZE(AO)
+
+ MOV C12, C11
+ LD A2, 1 * SIZE(AO)
+
+ MOV C21, C11
+ LD A3, 2 * SIZE(AO)
+
+ MOV C22, C11
+ LD A4, 3 * SIZE(AO)
+
+ MOV C31, C11
+ LD A5, 4 * SIZE(AO)
+
+ MOV C32, C11
+ LD A6, 5 * SIZE(AO)
+
+ MOV C41, C11
+ LD B1, 0 * SIZE(BO)
+
+ MOV C42, C11
+ LD B2, 1 * SIZE(BO)
+
+ MOV C13, C11
+ LD A7, 6 * SIZE(AO)
+
+ MOV C14, C11
+ LD A8, 7 * SIZE(AO)
+
+ MOV C23, C11
+ MOV C24, C11
+
+ MOV C33, C11
+ MOV C34, C11
+
+ MOV C43, C11
+ MOV C44, C11
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ dsubu TEMP, K, KK
+#elif defined(LEFT)
+ daddiu TEMP, KK, 8
+#else
+ daddiu TEMP, KK, 2
+#endif
+ dsra L, TEMP, 1
+ blez L, .L282
+ NOP
+
+#else
move BO, B # Reset B
dsra L, K, 1 # UnRoll K=4
MOV C43, C11
blez L, .L282
MOV C44, C11
-
+#endif
.align 4
.L2810:
.align 4
.L282:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L280
LD ALPHA, 152($fp)
.align 4
.L280: # Write Back
+#ifndef TRMMKERNEL
daddiu I, I, -1
LD A1, 0 * SIZE(CO1)
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
.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 $0, C11 # CLEAR REAULTS REGISTERS
+ MOV C12, C11
+ LD A1, 0 * SIZE(AO)
+
+ MOV C21, C11
+ MOV C22, C11
+ LD A2, 1 * SIZE(AO)
+
+ MOV C31, C11
+ MOV C32, C11
+ LD A3, 2 * SIZE(AO)
+
+ MOV C41, C11
+ MOV C42, C11
+ LD A4, 3 * SIZE(AO)
+
+ MOV C13, C11
+ MOV C14, C11
+ LD B1, 0 * SIZE(BO)
+
+ MOV C23, C11
+ MOV C24, C11
+ LD B2, 1 * SIZE(BO)
+
+ MOV C33, C11
+ MOV C34, C11
+
+ MOV C43, C11
+ blez L, .L242
+ MOV C44, C11
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ dsubu TEMP, K, KK
+#elif defined(LEFT)
+ daddiu TEMP, KK, 4
+#else
+ daddiu TEMP, KK, 2
+#endif
+ dsra L, TEMP, 1
+ blez L, .L242
+ NOP
+
+#else
move BO, B # Reset B
dsra L, K, 1 # UnRoll K=4
MOV C43, C11
blez L, .L242
MOV C44, C11
-
+#endif
.align 4
.L2410:
.align 4
.L242:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L240
LD ALPHA, 152($fp)
.align 4
.L240: # Write Back
+#ifndef TRMMKERNEL
LD A1, 0 * SIZE(CO1)
LD A2, 1 * SIZE(CO1)
LD A3, 2 * SIZE(CO1)
daddiu CO1, CO1, 4 * SIZE
daddiu CO2, CO2, 4 * SIZE
+#else
- .align 4
-.L22:
- andi I, M, 2
- blez I, .L21
- NOP
+ MUL A1, C11, ALPHA
+ MUL A2, C21, ALPHA
+ MUL A3, C31, ALPHA
+ MUL A4, C41, ALPHA
- .align 4
-.L221:
- move BO, B # Reset B
- dsra L, K, 1 # UnRoll K=4
+ MUL B1, C12, ALPHA
+ ST A1, 0 * SIZE(CO1)
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- LD A1, 0 * SIZE(AO)
+ 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 $0, C11 # CLEAR REAULTS REGISTERS
+ MOV C12, C11
+ LD A1, 0 * SIZE(AO)
MOV C21, C11
MOV C22, C11
MOV C43, C11
blez L, .L222
MOV C44, C11
+#endif
.align 4
.align 4
.L222:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L220
LD ALPHA, 152($fp)
.align 4
.L220: # Write Back
+#ifndef TRMMKERNEL
LD A1, 0 * SIZE(CO1)
LD A2, 1 * SIZE(CO1)
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:
.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 $0, C11 # CLEAR REAULTS REGISTERS
+ MOV C12, C11
+ LD A1, 0 * SIZE(AO)
+
+ MOV C21, C11
+ MOV C22, C11
+
+ MOV C31, C11
+ MOV C32, C11
+ LD B1, 0 * SIZE(BO)
+
+ MOV C41, C11
+ MOV C42, C11
+ LD B2, 1 * SIZE(BO)
+
+ MOV C43, C11
+ MOV C44, C11
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ dsubu TEMP, K, KK
+#elif defined(LEFT)
+ daddiu TEMP, KK, 1
+#else
+ daddiu TEMP, KK, 2
+#endif
+ dsra L, TEMP, 1
+ blez L, .L212
+ NOP
+
+#else
move BO, B # Reset B
dsra L, K, 1 # UnRoll K=4
MOV C43, C11
blez L, .L212
MOV C44, C11
-
+#endif
.align 4
.L2110:
.align 4
.L212:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L210
LD ALPHA, 152($fp)
.align 4
.L210: # Write Back
+#ifndef TRMMKERNEL
LD A1, 0 * SIZE(CO1)
MADD A1, A1, C11, ALPHA
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
- 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 $0, C11 # CLEAR REAULTS REGISTERS
+ LD A1, 0 * SIZE(AO)
+
+ MOV C12, C11
+ LD A2, 1 * SIZE(AO)
+
+ MOV C21, C11
+ LD A3, 2 * SIZE(AO)
+
+ MOV C22, C11
+ LD A4, 3 * SIZE(AO)
+
+ MOV C31, C11
+ LD A5, 4 * SIZE(AO)
+
+ MOV C32, C11
+ LD A6, 5 * SIZE(AO)
+
+ MOV C41, C11
+ LD B1, 0 * SIZE(BO)
+
+ MOV C42, C11
+ LD A7, 6 * SIZE(AO)
+
+ MOV C13, C11
+ LD A8, 7 * SIZE(AO)
+
+ MOV C14, C11
+
+ MOV C23, C11
+ MOV C24, C11
+
+ MOV C33, C11
+ MOV C34, C11
+
+ MOV C43, C11
+ MOV C44, C11
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ dsubu TEMP, K, KK
+#elif defined(LEFT)
+ daddiu TEMP, KK, 8
+#else
+ daddiu TEMP, KK, 1
+#endif
+ dsra L, TEMP, 1
+ blez L, .L182
+ NOP
+
+#else
move BO, B # Reset B
dsra L, K, 1 # UnRoll K=4
MOV C43, C11
blez L, .L182
MOV C44, C11
+#endif
.align 4
.align 4
.L182:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L180
LD ALPHA, 152($fp)
.align 4
.L180: # Write Back
+#ifndef TRMMKERNEL
daddiu I, I, -1
LD A1, 0 * 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:
.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 $0, C11 # CLEAR REAULTS REGISTERS
+ MOV C12, C11
+ LD A1, 0 * SIZE(AO)
+
+ MOV C21, C11
+ MOV C22, C11
+ LD A2, 1 * SIZE(AO)
+
+ MOV C31, C11
+ MOV C32, C11
+ LD A3, 2 * SIZE(AO)
+
+ MOV C41, C11
+ MOV C42, C11
+ LD A4, 3 * SIZE(AO)
+
+ MOV C13, C11
+ MOV C14, C11
+ LD B1, 0 * SIZE(BO)
+
+ MOV C23, C11
+ MOV C24, C11
+
+ MOV C33, C11
+ MOV C34, C11
+
+ MOV C43, C11
+ MOV C44, C11
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ dsubu TEMP, K, KK
+#elif defined(LEFT)
+ daddiu TEMP, KK, 4
+#else
+ daddiu TEMP, KK, 1
+#endif
+ dsra L, TEMP, 1
+ blez L, .L142
+ NOP
+
+#else
move BO, B # Reset B
dsra L, K, 1 # UnRoll K=4
MOV C43, C11
blez L, .L142
MOV C44, C11
-
+#endif
.align 4
.L1410:
.align 4
.L142:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L140
LD ALPHA, 152($fp)
.align 4
.L140: # Write Back
+#ifndef TRMMKERNEL
LD A1, 0 * SIZE(C)
LD A2, 1 * SIZE(C)
LD A3, 2 * 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:
.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 $0, C11 # CLEAR REAULTS REGISTERS
+ MOV C12, C11
+ LD A1, 0 * SIZE(AO)
+
+ MOV C21, C11
+ MOV C22, C11
+ LD A2, 1 * SIZE(AO)
+
+ MOV C31, C11
+ MOV C32, C11
+ LD B1, 0 * SIZE(BO)
+
+ MOV C41, C11
+ MOV C42, C11
+
+ MOV C43, C11
+ MOV C44, C11
+#if (defined(LEFT) && !defined(TRANSA)) ||\
+ (!defined(LEFT) && defined(TRANSA))
+ dsubu TEMP, K, KK
+#elif defined(LEFT)
+ daddiu TEMP, KK, 2
+#else
+ daddiu TEMP, KK, 1
+#endif
+ dsra L, TEMP, 1
+ blez L, .L122
+ NOP
+
+#else
move BO, B # Reset B
dsra L, K, 1 # UnRoll K=4
MOV C43, C11
blez L, .L122
MOV C44, C11
-
+#endif
.align 4
.L1210:
.align 4
.L122:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L120
LD ALPHA, 152($fp)
.align 4
.L120: # Write Back
+#ifndef TRMMKERNEL
LD A1, 0 * SIZE(C)
LD A2, 1 * 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:
.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 $0, C11 # CLEAR REAULTS REGISTERS
+ MOV C12, C11
+ LD A1, 0 * SIZE(AO)
+
+ MOV C21, C11
+ MOV C22, C11
+ LD B1, 0 * SIZE(BO)
+
+ MOV C31, C11
+ MOV C32, C11
+#if (defined(LEFT) && !defined(TRANSA))||\
+ (!defined(LEFT) && defined(TRANSA))
+ dsubu TEMP, K, KK
+#elif defined(LEFT)
+ daddiu TEMP, KK, 1
+#else
+ daddiu TEMP, KK, 1
+#endif
+ dsra L, TEMP, 1
+ blez L, .L112
+ NOP
+
+#else
move BO, B # Reset B
dsra L, K, 1 # UnRoll K=4
MOV C31, C11
blez L, .L112
MOV C32, C11
-
+#endif
.align 4
.align 4
.L112:
+#ifndef TRMMKERNEL
andi L, K, 1
+#else
+ andi L, TEMP, 1
+#endif
blez L, .L110
LD ALPHA, 152($fp)
.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 $16, 0($fp)
ld $17, 8($fp)
projects / platform / upstream / openblas.git / commitdiff