|
- ##define REALNAME gemm
- #define ASSEMBLER
- #include "common.h"
-
- #define FETCH ld
- #define STACKSIZE 192
- #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 $9
- #define B $10
- #define C $11
- #define LDC $8
-
- #### Pointer A, B, C ####
- #define AO $12
- #define BO $13
-
- #define CO1 $14
- #define CO2 $15
-
- #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 $25
- #endif
-
-
- PROLOGUE
-
- LDARG LDC, 0($sp)
- 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)
-
- sdc1 $f24, 56($sp)
- sdc1 $f25, 64($sp)
- sdc1 $f26, 72($sp)
- sdc1 $f27, 80($sp)
- sdc1 $f28, 88($sp)
-
- #if defined(TRMMKERNEL)
- sd $23, 96($sp)
- sd $24, 104($sp)
- sd $25, 112($sp)
-
- LDARG OFFSET, STACKSIZE+8($sp)
- #endif
-
- #ifndef __64BIT__
- sdc1 $f20,120($sp)
- sdc1 $f21,128($sp)
- sdc1 $f22,136($sp)
- sdc1 $f23,144($sp)
- #endif
-
- .align 4
- .L2:
- dsra J, N, 1 # NR=2
- ST $f15, 152($sp)
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- neg KK, OFFSET
- #endif
-
- dsll LDC, LDC, ZBASE_SHIFT# LDC*SIZE
- blez J, .L1
- ST $f16, 160($sp)
-
- .L24:
- #if defined(TRMMKERNEL) && defined(LEFT)
- move KK, OFFSET
- #endif
-
- dsra I, M, 2 # MR=8
- move AO, A # Reset A
-
- dsll PREA, K, 1 + ZBASE_SHIFT
- move CO1, C
-
- daddu CO2, C, LDC
- daddu PREA, AO, PREA
-
- blez I, .L22
- daddu C, CO2, LDC
-
- .align 4
- .L241:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 2 + ZBASE_SHIFT
- dsll TEMP, KK, 1 + ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- dsll PREB, K, ZBASE_SHIFT
- MOV C21, C11
- MOV C22, C11
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MOV C31, C11
- MOV C32, C11
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C41, C11
- MOV C42, C11
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- MOV C24, C11
-
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
-
- PLU B3, B1, B1
- PLU B4, B2, B2
- daddu PREB, BO, PREB
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 8 * SIZE(CO1)
- FETCH $0, 0 * SIZE(CO2)
- FETCH $0, 8 * SIZE(CO2)
- #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, 2
- blez L, .L242
- NOP
-
- #else
-
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- dsll PREB, K, ZBASE_SHIFT
- MOV C21, C11
- MOV C22, C11
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MOV C31, C11
- MOV C32, C11
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C41, C11
- MOV C42, C11
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MOV C13, C11
- MOV C14, C11
-
- FETCH $0, 0 * SIZE(CO1)
- MOV C23, C11
- MOV C24, C11
-
- FETCH $0, 0 * SIZE(CO2)
- MOV C33, C11
- MOV C34, C11
-
- MOV C43, C11
- MOV C44, C11
- daddu PREB, BO, PREB
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- FETCH $0, 8 * SIZE(CO1)
- blez L, .L242
- FETCH $0, 8 * SIZE(CO2)
- #endif
-
- .L2410:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F5, F4, 2) # A5 A6
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
-
- FETCH $0, 0 * SIZE(PREB)
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
-
- FETCH $0, 0 * SIZE(PREA)
- 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
- MADPS C24, C24, A2, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
- daddu PREB, PREB, 8 * SIZE
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- gsLQC1(R13, F9, F8, 2) # B1 B2
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
-
- gsLQC1(R12, F1, F0, 4) # A1 A2
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
-
- gsLQC1(R12, F3, F2, 5) # A3 A4
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
-
- FETCH $0, 8 * SIZE(PREA)
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
-
- 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
- MADPS C24, C24, A6, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- gsLQC1(R13, F13, F12, 3) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F5, F4, 6) # A5 A6
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- gsLQC1(R12, F7, F6, 7) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- daddiu BO, BO, 4 * 4 * SIZE # 4KR*4NR
-
- FETCH $0, 16 * SIZE(PREA)
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- daddiu AO, AO, 8 * 4 * SIZE # 4KR*8MR
-
- 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
- MADPS C24, C24, A2, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
-
- FETCH $0, 24 * SIZE(PREA)
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
-
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
- daddu PREA, PREA, 32 * SIZE
-
- MADPS C33, C33, A7, B7
- MADPS C43, C43, A8, B7
-
- MADPS C14, C14, A5, B8
- MADPS C24, C24, A6, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- MADPS C34, C34, A7, B8
- bgtz L, .L2410
- MADPS C44, C44, A8, B8
-
-
- .align 4
- .L242:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L247
- NOP
-
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F5, F4, 2) # A5 A6
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- daddiu BO, BO, 2 * 4 * SIZE # 4KR*4NR
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- daddiu AO, AO, 4 * 4 * SIZE
-
- 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
- MADPS C24, C24, A2, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C12, C12, A5, B6
- MADPS C22, C22, A6, B6
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C31, C31, A7, B5
- MADPS C41, C41, A8, B5
-
- MADPS C32, C32, A7, B6
- MADPS C42, C42, A8, B6
-
- 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
- MADPS C24, C24, A6, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- MADPS C34, C34, A7, B8
- MADPS C44, C44, A8, B8
-
- .align 4
- .L247:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L240
- NOP
-
- 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
- daddiu BO, BO, 1 * 4 * SIZE # 4KR*4NR
-
- MADPS C32, C32, A3, B2
- MADPS C42, C42, A4, B2
- daddiu AO, AO, 2 * 4 * SIZE
-
- 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
- MADPS C24, C24, A2, B4
-
- MADPS C34, C34, A3, B4
- MADPS C44, C44, A4, B4
-
-
- .align 4
- .L240: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C31
- CVTU A4, C41
-
- CVTU A5, C13
- CVTU A6, C23
-
- CVTU A7, C33
- CVTU A8, C43
-
- CVTU B1, C12
- CVTU B2, C22
-
- CVTU B3, C32
- CVTU B4, C42
-
- CVTU B5, C14
- CVTU B6, C24
-
- CVTU B7, C34
- CVTU B8, C44
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- SUB C12, C12, B1
- SUB C22, C22, B2
- SUB C32, C32, B3
- SUB C42, C42, B4
- ADD C14, B5, C14
- ADD C24, B6, C24
- ADD C34, B7, C34
- ADD C44, B8, C44
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- LD C13, 0 * SIZE(CO2)
- LD C23, 2 * SIZE(CO2)
- LD C33, 4 * SIZE(CO2)
- LD C43, 6 * SIZE(CO2)
- LD C11, 1 * SIZE(CO2)
- LD C21, 3 * SIZE(CO2)
- LD C31, 5 * SIZE(CO2)
- LD C41, 7 * SIZE(CO2)
-
- MADD C13, C13, C12, A1
- MADD C23, C23, C22, A1
-
- MADD C33, C33, C32, A1
- ST B1, 0 * SIZE(CO1)
-
- MADD C43, C43, C42, A1
- ST B3, 2 * SIZE(CO1)
-
- MADD C11, C11, C14, A1
- ST B5, 4 * SIZE(CO1)
-
- MADD C21, C21, C24, A1
- ST B7, 6 * SIZE(CO1)
-
- MADD C31, C31, C34, A1
- ST B2, 1 * SIZE(CO1)
-
- MADD C41, C41, C44, A1
- ST B4, 3 * SIZE(CO1)
-
- NMSUB C13, C13, C14, A2
- ST B6, 5 * SIZE(CO1)
-
- NMSUB C23, C23, C24, A2
- ST B8, 7 * SIZE(CO1)
-
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- SUB C13, A5, C13 # ad'+'cb
- SUB C23, A6, C23
- SUB C33, A7, C33
- SUB C43, A8, C43
- ADD C12, B1, C12
- ADD C22, B2, C22
- ADD C32, B3, C32
- ADD C42, B4, C42
- SUB C14, B5, C14
- SUB C24, B6, C24
- SUB C34, B7, C34
- SUB C44, B8, C44
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- LD C13, 0 * SIZE(CO2)
- LD C23, 2 * SIZE(CO2)
- LD C33, 4 * SIZE(CO2)
- LD C43, 6 * SIZE(CO2)
- LD C11, 1 * SIZE(CO2)
- LD C21, 3 * SIZE(CO2)
- LD C31, 5 * SIZE(CO2)
- LD C41, 7 * SIZE(CO2)
-
- MADD C13, C13, C12, A1
- MADD C23, C23, C22, A1
-
- MADD C33, C33, C32, A1
- ST B1, 0 * SIZE(CO1)
-
- MADD C43, C43, C42, A1
- ST B3, 2 * SIZE(CO1)
-
- MADD C11, C11, C14, A1
- ST B5, 4 * SIZE(CO1)
-
- MADD C21, C21, C24, A1
- ST B7, 6 * SIZE(CO1)
-
- MADD C31, C31, C34, A1
- ST B2, 1 * SIZE(CO1)
-
- MADD C41, C41, C44, A1
- ST B4, 3 * SIZE(CO1)
-
- NMSUB C13, C13, C14, A2
- ST B6, 5 * SIZE(CO1)
-
- NMSUB C23, C23, C24, A2
- ST B8, 7 * SIZE(CO1)
-
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- SUB C13, C13, A5 # ad'+'cb
- SUB C23, C23, A6
- SUB C33, C33, A7
- SUB C43, C43, A8
- ADD C12, B1, C12
- ADD C22, B2, C22
- ADD C32, B3, C32
- ADD C42, B4, C42
- SUB C14, C14, B5
- SUB C24, C24, B6
- SUB C34, C34, B7
- SUB C44, C44, B8
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- LD C13, 0 * SIZE(CO2)
- LD C23, 2 * SIZE(CO2)
- LD C33, 4 * SIZE(CO2)
- LD C43, 6 * SIZE(CO2)
- LD C11, 1 * SIZE(CO2)
- LD C21, 3 * SIZE(CO2)
- LD C31, 5 * SIZE(CO2)
- LD C41, 7 * SIZE(CO2)
-
- MADD C13, C13, C12, A1
- MADD C23, C23, C22, A1
-
- MADD C33, C33, C32, A1
- ST B1, 0 * SIZE(CO1)
-
- MADD C43, C43, C42, A1
- ST B3, 2 * SIZE(CO1)
-
- MADD C11, C11, C14, A1
- ST B5, 4 * SIZE(CO1)
-
- MADD C21, C21, C24, A1
- ST B7, 6 * SIZE(CO1)
-
- MADD C31, C31, C34, A1
- ST B2, 1 * SIZE(CO1)
-
- MADD C41, C41, C44, A1
- ST B4, 3 * SIZE(CO1)
-
- NMSUB C13, C13, C14, A2
- ST B6, 5 * SIZE(CO1)
-
- NMSUB C23, C23, C24, A2
- ST B8, 7 * SIZE(CO1)
-
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- SUB C12, C12, B1
- SUB C22, C22, B2
- SUB C32, C32, B3
- SUB C42, C42, B4
- ADD C14, B5, C14
- ADD C24, B6, C24
- ADD C34, B7, C34
- ADD C44, B8, C44
- NEG C13, C13
- NEG C23, C23
- NEG C33, C33
- NEG C43, C43
- NEG C14, C14
- NEG C24, C24
- NEG C34, C34
- NEG C44, C44
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- LD C13, 0 * SIZE(CO2)
- LD C43, 6 * SIZE(CO2)
- LD C23, 2 * SIZE(CO2)
- LD C33, 4 * SIZE(CO2)
- LD C11, 1 * SIZE(CO2)
- LD C21, 3 * SIZE(CO2)
- LD C31, 5 * SIZE(CO2)
- LD C41, 7 * SIZE(CO2)
-
- MADD C13, C13, C12, A1
- ST B1, 0 * SIZE(CO1)
-
- MADD C23, C23, C22, A1
- ST B3, 2 * SIZE(CO1)
-
- MADD C33, C33, C32, A1
- ST B5, 4 * SIZE(CO1)
-
- MADD C43, C43, C42, A1
- ST B7, 6 * SIZE(CO1)
-
- MADD C11, C11, C14, A1
- ST B2, 1 * SIZE(CO1)
-
- MADD C21, C21, C24, A1
- ST B4, 3 * SIZE(CO1)
-
- MADD C31, C31, C34, A1
- ST B6, 5 * SIZE(CO1)
-
- MADD C41, C41, C44, A1
- ST B8, 7 * SIZE(CO1)
-
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C31
- CVTU A4, C41
-
- CVTU A5, C13
- CVTU A6, C23
-
- CVTU A7, C33
- CVTU A8, C43
-
- CVTU B1, C12
- CVTU B2, C22
-
- CVTU B3, C32
- CVTU B4, C42
-
- CVTU B5, C14
- CVTU B6, C24
-
- CVTU B7, C34
- CVTU B8, C44
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- SUB C41, C41, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- # LD A2, 0 * SIZE(A) # load alpha_i
- ADD C33, A7, C33
- ADD C43, A8, C43
- SUB C12, C12, B1
- SUB C22, C22, B2
- SUB C32, C32, B3
- SUB C42, C42, B4
- ADD C14, B5, C14
- ADD C24, B6, C24
- ADD C34, B7, C34
- ADD C44, B8, C44
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- MUL C13, C12, A1
- MUL C23, C22, A1
-
- ST B3, 2 * SIZE(CO1)
- MUL C33, C32, A1
- MUL C43, C42, A1
-
- ST B5, 4 * SIZE(CO1)
- MUL C11, C14, A1
- MUL C21, C24, A1
-
- ST B7, 6 * SIZE(CO1)
- MUL C31, C34, A1
- MUL C41, C44, A1
-
- ST B2, 1 * SIZE(CO1)
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
-
- ST B4, 3 * SIZE(CO1)
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- ST B6, 5 * SIZE(CO1)
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- ST B8, 7 * SIZE(CO1)
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- SUB C13, A5, C13 # ad'+'cb
- SUB C23, A6, C23
- SUB C33, A7, C33
- SUB C43, A8, C43
- ADD C12, B1, C12
- ADD C22, B2, C22
- ADD C32, B3, C32
- ADD C42, B4, C42
- SUB C14, B5, C14
- SUB C24, B6, C24
- SUB C34, B7, C34
- SUB C44, B8, C44
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- MUL C13, C12, A1
- MUL C23, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- MUL C33, C32, A1
- MUL C43, C42, A1
-
- ST B3, 2 * SIZE(CO1)
- MUL C11, C14, A1
- MUL C21, C24, A1
-
- ST B5, 4 * SIZE(CO1)
- MUL C31, C34, A1
- MUL C41, C44, A1
-
- ST B7, 6 * SIZE(CO1)
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
-
- ST B2, 1 * SIZE(CO1)
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- ST B4, 3 * SIZE(CO1)
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- ST B6, 5 * SIZE(CO1)
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST B8, 7 * SIZE(CO1)
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- SUB C13, C13, A5 # ad'+'cb
- SUB C23, C23, A6
- SUB C33, C33, A7
- SUB C43, C43, A8
- ADD C12, B1, C12
- ADD C22, B2, C22
- ADD C32, B3, C32
- ADD C42, B4, C42
- SUB C14, C14, B5
- SUB C24, C24, B6
-
- SUB C34, C34, B7
- SUB C44, C44, B8
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- MUL C13, C12, A1
- MUL C23, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- MUL C33, C32, A1
- MUL C43, C42, A1
-
- ST B3, 2 * SIZE(CO1)
- MUL C11, C14, A1
- MUL C21, C24, A1
-
- ST B5, 4 * SIZE(CO1)
- MUL C31, C34, A1
- MUL C41, C44, A1
-
- ST B7, 6 * SIZE(CO1)
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
-
- ST B2, 1 * SIZE(CO1)
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- ST B4, 3 * SIZE(CO1)
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- ST B6, 5 * SIZE(CO1)
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST B8, 7 * SIZE(CO1)
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
-
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- SUB C12, C12, B1
- SUB C22, C22, B2
- SUB C32, C32, B3
- SUB C42, C42, B4
- ADD C14, B5, C14
- ADD C24, B6, C24
- ADD C34, B7, C34
- ADD C44, B8, C44
-
- NEG C13, C13
- NEG C23, C23
- NEG C33, C33
- NEG C43, C43
- NEG C14, C14
- NEG C24, C24
- NEG C34, C34
- NEG C44, C44
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- MUL C13, C12, A1
- MUL C23, C22, A1
-
- ST B3, 2 * SIZE(CO1)
- MUL C33, C32, A1
- MUL C43, C42, A1
-
- ST B5, 4 * SIZE(CO1)
- MUL C11, C14, A1
- MUL C21, C24, A1
-
- ST B7, 6 * SIZE(CO1)
- MUL C31, C34, A1
- MUL C41, C44, A1
-
- ST B2, 1 * SIZE(CO1)
- NMSUB C13, C13, C14, A2
- NMSUB C23, C23, C24, A2
-
- ST B4, 3 * SIZE(CO1)
- NMSUB C33, C33, C34, A2
- NMSUB C43, C43, C44, A2
-
- ST B6, 5 * SIZE(CO1)
- MADD C11, C11, C12, A2
- MADD C21, C21, C22, A2
-
- ST B8, 7 * SIZE(CO1)
- MADD C31, C31, C32, A2
- MADD C41, C41, C42, A2
-
- ST C13, 0 * SIZE(CO2)
- ST C23, 2 * SIZE(CO2)
- ST C33, 4 * SIZE(CO2)
- ST C43, 6 * SIZE(CO2)
- ST C11, 1 * SIZE(CO2)
- ST C21, 3 * SIZE(CO2)
- ST C31, 5 * SIZE(CO2)
- ST C41, 7 * SIZE(CO2)
- #endif
-
-
- #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 + ZBASE_SHIFT
- dsll TEMP, TEMP, 1 + ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 4
- #endif
-
- #endif
- daddiu CO1, CO1, 8 * SIZE
- bgtz I, .L241
- daddiu CO2, CO2, 8 * SIZE
-
- .align 4
- .L22:
- andi I, M, 2 # MR=4
- blez I, .L21
- NOP
-
- .align 4
- .L221:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll TEMP, KK, 1 + ZBASE_SHIFT # NR=2
-
- daddu AO, AO, TEMP
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- MOV C21, C11
- MOV C22, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
-
- FETCH $0, 8 * SIZE(CO1)
- MOV C24, C11
-
- FETCH $0, 0 * SIZE(CO2)
- FETCH $0, 8 * SIZE(CO2)
-
- PLU B3, B1, B1
- PLU B4, B2, B2
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 2 # MR=2
- #else
- daddiu TEMP, KK, 2 # NR=2
- #endif
- dsra L, TEMP, 2
- blez L, .L222
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
-
- MOV C21, C11
- MOV C22, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C14, C11
-
- MOV C23, C11
- FETCH $0, 0 * SIZE(CO1)
-
- FETCH $0, 8 * SIZE(CO1)
- MOV C24, C11
-
- FETCH $0, 0 * SIZE(CO2)
- FETCH $0, 8 * SIZE(CO2)
-
- PLU B3, B1, B1
- blez L, .L222
- PLU B4, B2, B2
- #endif
-
- .L2210:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C14, C14, A1, B4
- MADPS C24, C24, A2, B4
-
- gsLQC1(R12, F5, F4, 2) # A5 A6
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- gsLQC1(R13, F9, F8, 2) # B1 B2
- MADPS C11, C11, A3, B5
- MADPS C21, C21, A4, B5
-
- MADPS C12, C12, A3, B6
- MADPS C22, C22, A4, B6
-
- MADPS C13, C13, A3, B7
- MADPS C23, C23, A4, B7
-
- MADPS C14, C14, A3, B8
- MADPS C24, C24, A4, B8
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- gsLQC1(R13, F13, F12, 3) # B3 B4
- MADPS C11, C11, A5, B1
- MADPS C21, C21, A6, B1
-
- MADPS C12, C12, A5, B2
- MADPS C22, C22, A6, B2
- daddiu BO, BO, 4 * 4 * SIZE # 4KR*4NR
-
- daddiu AO, AO, 4 * 4 * SIZE # 4KR*8MR
- MADPS C13, C13, A5, B3
- MADPS C23, C23, A6, B3
-
- MADPS C14, C14, A5, B4
- MADPS C24, C24, A6, B4
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A7, B5
- MADPS C21, C21, A8, B5
-
- MADPS C12, C12, A7, B6
- MADPS C22, C22, A8, B6
-
- MADPS C13, C13, A7, B7
- MADPS C23, C23, A8, B7
-
- MADPS C14, C14, A7, B8
- MADPS C24, C24, A8, B8
-
- PLU B3, B1, B1
- bgtz L, .L2210
- PLU B4, B2, B2
-
-
- .align 4
- .L222:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L227
- NOP
-
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- 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
- daddiu BO, BO, 2 * 4 * SIZE
-
- daddiu AO, AO, 2 * 4 * SIZE
- MADPS C11, C11, A3, B5
- MADPS C21, C21, A4, B5
- gsLQC1(R13, F9, F8, 0) # A1 A2
-
- MADPS C12, C12, A3, B6
- MADPS C22, C22, A4, B6
- gsLQC1(R12, F1, F0, 0) # A1 A2
-
- 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
-
-
- .align 4
- .L227:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L220
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- daddiu BO, BO, 4 * SIZE
- daddiu AO, AO, 4 * SIZE
-
- MADPS C12, C12, A1, B2
- MADPS C22, C22, A2, B2
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C14, C14, A1, B4
- MADPS C24, C24, A2, B4
-
- .align 4
- .L220: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C13
- CVTU A4, C23
-
- CVTU A5, C12
- CVTU A6, C22
-
- CVTU A7, C14
- CVTU A8, C24
-
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- SUB C22, C22, A6
- ADD C14, A7, C14
- ADD C24, A8, C24
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B7, 2 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
- LD B8, 3 * SIZE(CO2)
-
- MADD B5, B5, C12, A1
- MADD B7, B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MADD B6, B6, C14, A1
- MADD B8, B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, A3, C13 # ad'+'cb
- SUB C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C12, A5, C12
- ADD C22, A6, C22
- SUB C14, A7, C14
- SUB C24, A8, C24
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B7, 2 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
- LD B8, 3 * SIZE(CO2)
-
- MADD B5, B5, C12, A1
- MADD B7, B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MADD B6, B6, C14, A1
- MADD B8, B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, C13, A3 # ad'+'cb
- SUB C23, C23, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C12, A5, C12
- ADD C22, A6, C22
- SUB C14, C14, A7
- SUB C24, C24, A8
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B7, 2 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
- LD B8, 3 * SIZE(CO2)
-
- MADD B5, B5, C12, A1
- MADD B7, B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MADD B6, B6, C14, A1
- MADD B8, B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
-
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- SUB C22, C22, A6
- ADD C14, A7, C14
- ADD C24, A8, C24
- NEG C13, C13
- NEG C23, C23
- NEG C14, C14
- NEG C24, C24
-
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B7, 2 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
- LD B8, 3 * SIZE(CO2)
-
- MADD B5, B5, C12, A1
- MADD B7, B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MADD B6, B6, C14, A1
- MADD B8, B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C13
- CVTU A4, C23
-
- CVTU A5, C12
- CVTU A6, C22
-
- CVTU A7, C14
- CVTU A8, C24
-
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- SUB C22, C22, A6
- ADD C14, A7, C14
- ADD C24, A8, C24
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
-
- MUL B5, C12, A1
- MUL B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MUL B6, C14, A1
- MUL B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, A3, C13 # ad'+'cb
- SUB C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C12, A5, C12
- ADD C22, A6, C22
- SUB C14, A7, C14
- SUB C24, A8, C24
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- MUL B5, C12, A1
- MUL B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MUL B6, C14, A1
- MUL B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, C13, A3 # ad'+'cb
- SUB C23, C23, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C12, A5, C12
- ADD C22, A6, C22
- SUB C14, C14, A7
- SUB C24, C24, A8
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- MUL B5, C12, A1
- MUL B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MUL B6, C14, A1
- MUL B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
-
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- SUB C22, C22, A6
- ADD C14, A7, C14
- ADD C24, A8, C24
- NEG C13, C13
- NEG C23, C23
- NEG C14, C14
- NEG C24, C24
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- MUL B5, C12, A1
- MUL B7, C22, A1
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- MUL B6, C14, A1
- MUL B8, C24, A1
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- NMSUB B7, B7, C24, A2
-
- MADD B6, B6, C12, A2
- MADD B8, B8, C22, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B7, 2 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- ST B8, 3 * SIZE(CO2)
- #endif
-
- #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 TEMP, TEMP, 1 + ZBASE_SHIFT
-
- daddu AO, AO, TEMP
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 2
- #endif
-
- #endif
- daddiu CO1, CO1, 4 * SIZE
- daddiu CO2, CO2, 4 * SIZE
-
-
- .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
- #else
- dsll L, KK, ZBASE_SHIFT # MR=1
- dsll TEMP, KK, 1 + ZBASE_SHIFT # NR=2
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C14, C11
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 0 * SIZE(CO2)
-
- PLU B3, B1, B1
- PLU B4, B2, B2
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 1 # MR=1
- #else
- daddiu TEMP, KK, 2 # NR=2
- #endif
- dsra L, TEMP, 2
- blez L, .L212
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C12, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C14, C11
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 0 * SIZE(CO2)
-
- PLU B3, B1, B1
- blez L, .L212
- PLU B4, B2, B2
- #endif
-
- .L2110:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C12, C12, A1, B2
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- gsLQC1(R13, F9, F8, 2) # B1 B2
- MADPS C11, C11, A2, B5
- MADPS C12, C12, A2, B6
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C13, C13, A2, B7
- MADPS C14, C14, A2, B8
-
- PLU B3, B1, B1
- PLU B4, B2, B2
-
- gsLQC1(R13, F13, F12, 3) # B3 B4
- MADPS C11, C11, A3, B1
- MADPS C12, C12, A3, B2
- daddiu BO, BO, 4 * 4 * SIZE # 4KR*4NR
-
- daddiu AO, AO, 2 * 4 * SIZE # 4KR*8MR
- MADPS C13, C13, A3, B3
- MADPS C14, C14, A3, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A4, B5
- MADPS C12, C12, A4, B6
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C13, C13, A4, B7
- MADPS C14, C14, A4, B8
-
- PLU B3, B1, B1
- bgtz L, .L2110
- PLU B4, B2, B2
-
-
- .align 4
- .L212:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L217
- NOP
-
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C12, C12, A1, B2
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
- daddiu BO, BO, 2 * 4 * SIZE
-
- MADPS C11, C11, A2, B5
- MADPS C12, C12, A2, B6
- daddiu AO, AO, 4 * SIZE
-
- MADPS C13, C13, A2, B7
- MADPS C14, C14, A2, B8
-
- gsLQC1(R12, F1, F0, 0) # A5 A6
- gsLQC1(R13, F9, F8, 0) # B1 B2
- PLU B3, B1, B1
- PLU B4, B2, B2
-
-
- .align 4
- .L217:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L210
- NOP
-
- MADPS C11, C11, A1, B1
- daddiu BO, BO, 4 * SIZE
- MADPS C12, C12, A1, B2
- daddiu AO, AO, 2 * SIZE
-
- MADPS C13, C13, A1, B3
- MADPS C14, C14, A1, B4
-
- .align 4
- .L210: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A3, C13
- CVTU A5, C12
- CVTU A7, C14
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- ADD C14, A7, C14
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
-
- MADD B5, B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MADD B6, B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C12, A5, C12
- SUB C14, A7, C14
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
-
- MADD B5, B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MADD B6, B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, C13, A3 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C12, A5, C12
- SUB C14, C14, A7
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
-
- MADD B5, B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MADD B6, B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- ADD C14, A7, C14
- NEG C13, C13
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- NEG C14, C14
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- LD B5, 0 * SIZE(CO2)
- LD B6, 1 * SIZE(CO2)
-
- MADD B5, B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MADD B6, B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A3, C13
- CVTU A5, C12
- CVTU A7, C14
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- ADD C14, A7, C14
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- MUL B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MUL B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C12, A5, C12
- SUB C14, A7, C14
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- MUL B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MUL B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
-
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, C13, A3 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- ADD C12, A5, C12
- SUB C14, C14, A7
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- MUL B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MUL B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- SUB C12, C12, A5
- ADD C14, A7, C14
- NEG C13, C13
- NEG C14, C14
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- MUL B5, C12, A4
- ST B1, 0 * SIZE(CO1)
- MUL B6, C14, A4
- ST B2, 1 * SIZE(CO1)
-
- NMSUB B5, B5, C14, A2
- MADD B6, B6, C12, A2
-
- ST B5, 0 * SIZE(CO2)
- ST B6, 1 * SIZE(CO2)
- #endif
-
-
- #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, ZBASE_SHIFT
- dsll TEMP, TEMP, 1 + ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 1
- #endif
-
- #endif
- daddiu CO1, CO1, 2 * SIZE
- daddiu CO2, CO2, 2 * SIZE
-
-
- .align 4
- .L20:
- daddiu J, J, -1
- move B, BO
-
- #if defined(TRMMKERNEL) && !defined(LEFT)
- daddiu KK, KK, 2
- #endif
-
- bgtz J, .L24
- NOP
-
-
- .align 4
- .L1:
- andi J, N, 1
- blez J, .L999
- NOP
-
- .L14:
- dsra I, M, 2 # MR=8
- move AO, A # Reset A
-
- #if defined(TRMMKERNEL) && defined(LEFT)
- move KK, OFFSET
- #endif
-
- move CO1, C
- blez I, .L12
- daddu C, CO1, LDC
-
- .align 4
- .L141:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 2 + ZBASE_SHIFT
- dsll TEMP, KK, ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C21, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C31, C11
- MOV C41, C11
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MOV C13, C11
- MOV C23, C11
-
- FETCH $0, 0 * SIZE(CO1)
- MOV C33, C11
- MOV C43, C11
-
- FETCH $0, 8 * SIZE(CO1)
- PLU B3, B1, B1
- PLU B4, B2, B2
- #if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
- dsubu TEMP, K, KK
- #elif defined(LEFT)
- daddiu TEMP, KK, 4 # define Mr=4
- #else
- daddiu TEMP, KK, 1 # define NR=1
- #endif
- dsra L, TEMP, 2
- blez L, .L142
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C21, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C31, C11
- MOV C41, C11
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MOV C13, C11
- MOV C23, C11
-
- FETCH $0, 0 * SIZE(CO1)
- MOV C33, C11
- MOV C43, C11
-
- FETCH $0, 8 * SIZE(CO1)
- PLU B3, B1, B1
- blez L, .L142
- PLU B4, B2, B2
- #endif
-
- .L1410:
- daddiu L, L, -1
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
- gsLQC1(R13, F13, F12, 1) # B3 B4
-
- gsLQC1(R12, F1, F0, 4) # A1 A2
- MADPS C11, C11, A5, B2
- MADPS C21, C21, A6, B2
-
- gsLQC1(R12, F3, F2, 5) # A3 A4
- MADPS C31, C31, A7, B2
- MADPS C41, C41, A8, B2
- daddiu BO, BO, 2 * 4 * SIZE # 4KR*4NR
-
- MADPS C13, C13, A5, B4
- MADPS C23, C23, A6, B4
-
- MADPS C33, C33, A7, B4
- MADPS C43, C43, A8, B4
-
- PLU B7, B5, B5
- PLU B8, B6, B6
-
- MADPS C11, C11, A1, B5
- MADPS C21, C21, A2, B5
- gsLQC1(R12, F5, F4, 6) # A5 A6
-
- gsLQC1(R12, F7, F6, 7) # A7 A8
- MADPS C31, C31, A3, B5
- MADPS C41, C41, A4, B5
-
- daddiu AO, AO, 8 * 4 * SIZE # 4KR*8MR
- MADPS C13, C13, A1, B7
- MADPS C23, C23, A2, B7
-
- MADPS C33, C33, A3, B7
- MADPS C43, C43, A4, B7
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C11, C11, A5, B6
- MADPS C21, C21, A6, B6
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C31, C31, A7, B6
- MADPS C41, C41, A8, B6
-
- MADPS C13, C13, A5, B8
- MADPS C23, C23, A6, B8
-
- MADPS C33, C33, A7, B8
- MADPS C43, C43, A8, B8
-
- PLU B3, B1, B1
- bgtz L, .L1410
- PLU B4, B2, B2
-
-
- .align 4
- .L142:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L147
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- gsLQC1(R12, F5, F4, 2) # A5 A6
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- daddiu AO, AO, 4 * 4 * SIZE # 4KR*8MR
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
- gsLQC1(R13, F13, F8, 1) # B3 B4
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C11, C11, A5, B2
- MADPS C21, C21, A6, B2
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C31, C31, A7, B2
- MADPS C41, C41, A8, B2
- daddiu BO, BO, 4 * SIZE # 4KR*4NR
-
- MADPS C13, C13, A5, B4
- MADPS C23, C23, A6, B4
-
- MADPS C33, C33, A7, B4
- MADPS C43, C43, A8, B4
- PLU B3, B1, B1
-
-
- .align 4
- .L147:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L140
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- daddiu BO, BO, 2 * SIZE
-
- MADPS C31, C31, A3, B1
- MADPS C41, C41, A4, B1
- daddiu AO, AO, 2 * 4 * SIZE
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- MADPS C33, C33, A3, B3
- MADPS C43, C43, A4, B3
-
-
- .align 4
- .L140: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C31
- CVTU A4, C41
-
- CVTU A5, C13
- CVTU A6, C23
-
- CVTU A7, C33
- CVTU A8, C43
-
- CVTU B1, C12
- CVTU B2, C22
-
- CVTU B3, C32
- CVTU B4, C42
-
- CVTU B5, C14
- CVTU B6, C24
-
- CVTU B7, C34
- CVTU B8, C44
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- SUB C13, A5, C13 # ad'+'cb
- SUB C23, A6, C23
- SUB C33, A7, C33
- SUB C43, A8, C43
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- SUB C13, C13, A5 # ad'+'cb
- SUB C23, C23, A6
- SUB C33, C33, A7
- SUB C43, C43, A8
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # AC'+'BD
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # LOAD ALPHA_R
- # LD A1, 0 * SIZE(A) # LOAD ALPHA_R
- SUB C41, C41, A4
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # LOAD ALPHA_I
-
- ADD C13, A5, C13 # AD'+'CB
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- NEG C13, C13 # AD'+'CB
- NEG C23, C23
- NEG C33, C33
- NEG C43, C43
-
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B5, 4 * SIZE(CO1)
- LD B7, 6 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
- LD B6, 5 * SIZE(CO1)
- LD B8, 7 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = ALPHA_R
- MADD B3, B3, C21, A1
- MADD B5, B5, C31, A1
- MADD B7, B7, C41, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- MADD B6, B6, C33, A1
- MADD B8, B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = ALPHA_I
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C31
- CVTU A4, C41
-
- CVTU A5, C13
- CVTU A6, C23
-
- CVTU A7, C33
- CVTU A8, C43
-
- CVTU B1, C12
- CVTU B2, C22
-
- CVTU B3, C32
- CVTU B4, C42
-
- CVTU B5, C14
- CVTU B6, C24
-
- CVTU B7, C34
- CVTU B8, C44
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- # LD A1, 0 * SIZE(A) # load alpha_r
- SUB C31, C31, A3
- LD A1, 152($sp) # load alpha_r
- SUB C41, C41, A4
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
- ADD C13, A5, C13 # ad'+'cb
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
- SUB C13, A5, C13 # ad'+'cb
- SUB C23, A6, C23
- SUB C33, A7, C33
- SUB C43, A8, C43
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- # LD A1, 0 * SIZE(A) # load alpha_r
- ADD C31, A3, C31
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- ADD C41, A4, C41
- LD A2, 160($sp) # load alpha_i
- SUB C13, C13, A5 # ad'+'cb
- SUB C23, C23, A6
- SUB C33, C33, A7
- SUB C43, C43, A8
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # AC'+'BD
- SUB C21, C21, A2
- SUB C31, C31, A3
- LD A1, 152($sp) # LOAD ALPHA_R
- # LD A1, 0 * SIZE(A) # LOAD ALPHA_R
- SUB C41, C41, A4
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # LOAD ALPHA_I
-
- ADD C13, A5, C13 # AD'+'CB
- ADD C23, A6, C23
- ADD C33, A7, C33
- ADD C43, A8, C43
- NEG C13, C13 # AD'+'CB
- NEG C23, C23
- NEG C33, C33
- NEG C43, C43
-
- MUL B1, C11, A1 # A1 = ALPHA_R
- MUL B3, C21, A1
- MUL B5, C31, A1
- MUL B7, C41, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- MUL B6, C33, A1
- MUL B8, C43, A1
- NMSUB B1, B1, C13, A2 # A2 = ALPHA_I
- NMSUB B3, B3, C23, A2
- NMSUB B5, B5, C33, A2
- NMSUB B7, B7, C43, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
- MADD B6, B6, C31, A2
- MADD B8, B8, C41, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B5, 4 * SIZE(CO1)
- ST B7, 6 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- ST B6, 5 * SIZE(CO1)
- ST B8, 7 * SIZE(CO1)
- #endif
-
-
- #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 + ZBASE_SHIFT
- dsll TEMP, TEMP, ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 4
- #endif
-
- #endif
- bgtz I, .L141
- daddiu CO1, CO1, 8 * SIZE
-
- .align 4
- .L12:
- andi I, M, 2 # MR=4
- blez I, .L11
- NOP
-
- .align 4
- .L121:
- #if defined(TRMMKERNEL)
- #if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
- move BO, B
- #else
- dsll L, KK, 1 + ZBASE_SHIFT
- dsll TEMP, KK, ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, B, TEMP
- #endif
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C21, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C23, C11
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 8 * SIZE(CO1)
-
- PLU B3, B1, B1
- 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, 1
- #endif
- dsra L, TEMP, 2
- blez L, .L122
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- MOV C21, C11
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
- MOV C23, C11
-
- FETCH $0, 0 * SIZE(CO1)
- FETCH $0, 8 * SIZE(CO1)
-
- PLU B3, B1, B1
- blez L, .L122
- PLU B4, B2, B2
- #endif
-
- .L1210:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- gsLQC1(R12, F5, F4, 2) # A5 A6
- PLU B7, B5, B5
- PLU B8, B6, B6
- daddiu BO, BO, 2 * 4 * SIZE # 4KR*4NR
-
- MADPS C11, C11, A3, B2
- MADPS C21, C21, A4, B2
-
- gsLQC1(R12, F7, F6, 3) # A7 A8
- MADPS C13, C13, A3, B4
- MADPS C23, C23, A4, B4
-
- MADPS C11, C11, A5, B5
- MADPS C21, C21, A6, B5
- daddiu AO, AO, 4 * 4 * SIZE # 4KR*8MR
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C13, C13, A5, B7
- MADPS C23, C23, A6, B7
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C11, C11, A7, B6
- MADPS C21, C21, A8, B6
-
- MADPS C13, C13, A7, B8
- MADPS C23, C23, A8, B8
-
- PLU B3, B1, B1
- bgtz L, .L1210
- PLU B4, B2, B2
-
-
- .align 4
- .L122:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L127
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- PLU B7, B5, B5
- daddiu BO, BO, 1 * 4 * SIZE
-
- daddiu AO, AO, 2 * 4 * SIZE
- MADPS C11, C11, A3, B2
- MADPS C21, C21, A4, B2
-
- MADPS C13, C13, A3, B4
- MADPS C23, C23, A4, B4
-
- gsLQC1(R13, F9, F8, 0)
- gsLQC1(R12, F1, F0, 0)
- PLU B3, B1, B1
-
- .align 4
- .L127:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L120
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C21, C21, A2, B1
- daddiu BO, BO, 2 * SIZE
- daddiu AO, AO, 4 * SIZE
-
- MADPS C13, C13, A1, B3
- MADPS C23, C23, A2, B3
-
- .align 4
- .L120: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C13
- CVTU A4, C23
-
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, A3, C13 # ad'+'cb
- SUB C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, C13, A3 # ad'+'cb
- SUB C23, C23, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- NEG C13, C13 # ad'+'cb
- NEG C23, C23
-
- LD B1, 0 * SIZE(CO1)
- LD B3, 2 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
- LD B4, 3 * SIZE(CO1)
-
- MADD B1, B1, C11, A1 # A1 = alpha_r
- MADD B3, B3, C21, A1
- MADD B2, B2, C13, A1
- MADD B4, B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A2, C21
-
- CVTU A3, C13
- CVTU A4, C23
-
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
-
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, A3, C13 # ad'+'cb
- SUB C23, A4, C23
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_r
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- ADD C21, A2, C21
- SUB C13, C13, A3 # ad'+'cb
- SUB C23, C23, A4
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A1, 152($sp) # load alpha_r
- # LD A2, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- SUB C21, C21, A2
- ADD C13, A3, C13 # ad'+'cb
- ADD C23, A4, C23
- LD A1, 152($sp) # load alpha_r
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A2, 160($sp)
- # LD A2, 0 * SIZE(A) # load alpha_i
- NEG C13, C13 # ad'+'cb
- NEG C23, C23
-
- MUL B1, C11, A1 # A1 = alpha_r
- MUL B3, C21, A1
- MUL B2, C13, A1
- MUL B4, C23, A1
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- NMSUB B3, B3, C23, A2
- MADD B2, B2, C11, A2
- MADD B4, B4, C21, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B3, 2 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- ST B4, 3 * SIZE(CO1)
- #endif
- #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 + ZBASE_SHIFT
- dsll TEMP, TEMP, ZBASE_SHIFT
-
- daddu AO, AO, L
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 2
- #endif
-
- #endif
- daddiu CO1, CO1, 4 * SIZE
- daddiu CO2, CO2, 4 * SIZE
-
-
- .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 TEMP, KK, ZBASE_SHIFT
-
- daddu AO, AO, TEMP
- daddu BO, B, TEMP
- #endif
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
-
- FETCH $0, 0 * SIZE(CO1)
-
- PLU B3, B1, B1
- PLU B4, B2, B2
- #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, 2
- blez L, .L112
- NOP
-
- #else
- move BO, B # Reset B
- dsra L, K, 2 # UnRoll K=64
-
- MTC $0, C11 # CLEAR REAULTS REGISTERS
- gsLQC1(R13, F9, F8, 0) # B1 B2
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MOV C13, C11
-
- FETCH $0, 0 * SIZE(CO1)
-
- PLU B3, B1, B1
- blez L, .L112
- PLU B4, B2, B2
- #endif
-
- .L1110:
- daddiu L, L, -1
- gsLQC1(R13, F13, F12, 1) # B3 B4
- MADPS C11, C11, A1, B1
-
- gsLQC1(R12, F3, F2, 1) # A3 A4
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 2 * 4 * SIZE # 4KR*4NR
-
- PLU B7, B5, B5
- PLU B8, B6, B6
- daddiu AO, AO, 2 * 4 * SIZE # 4KR*8MR
-
- MADPS C11, C11, A2, B2
- MADPS C13, C13, A2, B4
-
- MADPS C11, C11, A3, B5
- MADPS C13, C13, A3, B7
-
- gsLQC1(R13, F9, F8, 0) # B1 B2
- MADPS C11, C11, A4, B6
-
- gsLQC1(R12, F1, F0, 0) # A1 A2
- MADPS C13, C13, A4, B8
-
- PLU B3, B1, B1
- bgtz L, .L1110
- PLU B4, B2, B2
-
-
- .align 4
- .L112:
- #ifndef TRMMKERNEL
- andi L, K, 2
- #else
- andi L, TEMP, 2
- #endif
- blez L, .L117
- NOP
-
- MADPS C11, C11, A1, B1
- MADPS C13, C13, A1, B3
- daddiu BO, BO, 4 * SIZE
- daddiu AO, AO, 4 * SIZE
-
- MADPS C11, C11, A2, B2
- MADPS C13, C13, A2, B4
-
- gsLQC1(R13, F9, F8, 0)
- gsLQC1(R12, F1, F0, 0)
- PLU B3, B1, B1
-
-
- .align 4
- .L117:
- #ifndef TRMMKERNEL
- andi L, K, 1
- #else
- andi L, TEMP, 1
- #endif
- blez L, .L110
- NOP
-
- daddiu BO, BO, 2 * SIZE
- daddiu AO, AO, 2 * SIZE
-
- MADPS C11, C11, A1, B1
- MADPS C13, C13, A1, B3
-
-
- .align 4
- .L110: # Write Back
- #ifndef TRMMKERNEL
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A3, C13
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, A3, C13 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, C13, A3 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- NEG C13, C13
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp)
-
- LD B1, 0 * SIZE(CO1)
- LD B2, 1 * SIZE(CO1)
-
- MADD B1, B1, C11, A4 # A1 = alpha_r
- MADD B2, B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #else
- daddiu I, I, -1
- CVTU A1, C11
- CVTU A3, C13
-
- #if defined(NN) || defined(NT) || defined(TN) || defined(TT)
- /* (a + bi) * (c + di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- # LD A1, 0 * SIZE(A) # load alpha_r
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
- # LD A2, 0 * SIZE(A) # load alpha_i
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(NR) || defined(NC) || defined(TR) || defined(TC)
- /* (a + bi) * (c - di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, A3, C13 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(RN) || defined(RT) || defined(CN) || defined(CT)
- /* (a - bi) * (c + di) */
- ADD C11, A1, C11 # ac'+'bd
- SUB C13, C13, A3 # ad'+'cb
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp) # load alpha_i
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
- #if defined(RR) || defined(RC) || defined(CR) || defined(CC)
- /* (a - bi) * (c - di) */
- SUB C11, C11, A1 # ac'+'bd
- ADD C13, A3, C13 # ad'+'cb
- NEG C13, C13
- LD A4, 152($sp) # load alpha_r
- LD A2, 160($sp)
-
- MUL B1, C11, A4 # A1 = alpha_r
- MUL B2, C13, A4
- NMSUB B1, B1, C13, A2 # A2 = alpha_i
- MADD B2, B2, C11, A2
-
- ST B1, 0 * SIZE(CO1)
- ST B2, 1 * SIZE(CO1)
- #endif
-
-
- #if ( defined(LEFT) && defined(TRANSA)) || \
- (!defined(LEFT) && !defined(TRANSA))
- dsubu TEMP, K, KK
- #ifdef LEFT
- daddiu TEMP, TEMP, -1
- #else
- daddiu TEMP, TEMP, -1
- #endif
-
- dsll TEMP, TEMP, ZBASE_SHIFT
-
- daddu AO, AO, TEMP
- daddu BO, BO, TEMP
- #endif
-
- #ifdef LEFT
- daddiu KK, KK, 1
- #endif
-
- #endif
- daddiu CO1, CO1, 2 * SIZE
- daddiu CO2, CO2, 2 * SIZE
-
-
- .align 4
- .L10:
- move B, BO
- #if defined(TRMMKERNEL) && !defined(LEFT)
- daddiu KK, KK, 1
- #endif
-
- .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)
-
- ldc1 $f24, 56($sp)
- ldc1 $f25, 64($sp)
- ldc1 $f26, 72($sp)
- ldc1 $f27, 80($sp)
- ldc1 $f28, 88($sp)
-
- #if defined(TRMMKERNEL)
- ld $23, 96($sp)
- ld $24, 104($sp)
- ld $25, 112($sp)
- #endif
-
- #ifndef __64BIT__
- ldc1 $f20,120($sp)
- ldc1 $f21,128($sp)
- ldc1 $f22,136($sp)
- ldc1 $f23,144($sp)
- #endif
-
- daddiu $sp,$sp,STACKSIZE
- j $31
- nop
-
- EPILOGUE
|