Merge branch 'loongson3a' of github.com:xianyi/OpenBLAS into loongson3a

diff --git a/kernel/mips64/KERNEL b/kernel/mips64/KERNEL
index ebb447b..a14b1cb 100644 (file)
--- a/kernel/mips64/KERNEL
+++ b/kernel/mips64/KERNEL
@@ -128,10 +128,21 @@ CTRSMKERNEL_LT    =  ztrsm_kernel_LT.S
 CTRSMKERNEL_RN =  ztrsm_kernel_LT.S
 CTRSMKERNEL_RT =  ztrsm_kernel_RT.S
 
+ifndef ZTRSMKERNEL_LN
 ZTRSMKERNEL_LN =  ztrsm_kernel_LT.S
+endif
+
+ifndef ZTRSMKERNEL_LT
 ZTRSMKERNEL_LT =  ztrsm_kernel_LT.S
+endif
+
+ifndef ZTRSMKERNEL_RN
 ZTRSMKERNEL_RN =  ztrsm_kernel_LT.S
+endif
+
+ifndef ZTRSMKERNEL_RT
 ZTRSMKERNEL_RT =  ztrsm_kernel_RT.S
+endif
 
 CGEMM3MKERNEL    =  zgemm3m_kernel.S
 ZGEMM3MKERNEL    =  zgemm3m_kernel.S

diff --git a/kernel/mips64/KERNEL.LOONGSON3A b/kernel/mips64/KERNEL.LOONGSON3A
index 94c8b1b..706f481 100644 (file)
--- a/kernel/mips64/KERNEL.LOONGSON3A
+++ b/kernel/mips64/KERNEL.LOONGSON3A
@@ -28,3 +28,12 @@ DTRSMKERNEL_LN       = ../generic/trsm_kernel_LN.c
 DTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
 DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
 DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
+
+
+
+

diff --git a/kernel/mips64/zgemm_kernel_loongson3a.S b/kernel/mips64/zgemm_kernel_loongson3a.S
index 4960367..4cc3966 100644 (file)
--- a/kernel/mips64/zgemm_kernel_loongson3a.S
+++ b/kernel/mips64/zgemm_kernel_loongson3a.S
@@ -1,12 +1,10 @@
 #define ASSEMBLER
 #include "common.h"
 
-
 #define FETCH  ld
 #define gsLQC1(base,fq,ft,offset) .word(0x32<<26|base<<21|ft<<16|0x1<<15|offset<<6|0x1<<5|fq)
 #define gsSQC1(base,fq,ft,offset) .word(0x3A<<26|base<<21|ft<<16|0x1<<15|offset<<6|0x1<<5|fq)
 
-
 #define STACKSIZE      160
 #define M      $4
 #define        N       $5
@@ -116,12 +114,12 @@
 ##     MADD3   a*d
 ##     MADD4   d*b
 ##################################
-####if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
+#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
 #define MADD1    MADD
 #define MADD2    MADD
 #define MADD3    MADD
 #define MADD4    NMSUB
-###endif
+#endif
 
 #if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
 #define MADD1    MADD
@@ -175,6 +173,9 @@
 
        dsra    J,  N, 1                                                        #       J=N/2
        ST      ALPHA_R, 128($sp)                                               #       store alpha_r & alpha_i
+#if defined(TRMMKERNEL) && !defined(LEFT)
+       neg     KK, OFFSET
+#endif
 
        dsll    LDC, LDC, ZBASE_SHIFT                           #       LDC*SIZE*COMPSIZE
        blez    J, .L20
@@ -183,6 +184,10 @@
 
        .align  5
 .L10:
+#if defined(TRMMKERNEL) &&  defined(LEFT)
+       move    KK, OFFSET
+#endif
+
        daddiu  J, J, -1                                                        
        dsra    I,  M, 1                                                        #       I=M/2
 
@@ -193,12 +198,66 @@
        daddu   CO2, C,   LDC
 
        move    AO, A                                                           #       Reset AO
-       daddu   PREB, PREB, B                                           #       PREA=A+panel size
-
        blez    I, .L30
        daddu   PREA, PREA, A                                           #       PREA=A+panel size
 
 .L11:
+#if defined(TRMMKERNEL)
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+       move    BO,  B
+#else
+       dsll    L,    KK, 1 + ZBASE_SHIFT                       #       MR=NR=2
+       dsll    TEMP, KK, 1 + ZBASE_SHIFT
+
+       daddu   AO, AO, L
+       daddu   BO, B,  TEMP
+#endif
+       MTC     $0,  c11                                                                #       Clear results regs
+       MOV     c12, c11
+       gsLQC1(R12, F1, F0, 0)                                          #       R:a1    I:a2
+
+       MOV     c13, c11
+       MOV     c14, c11
+       gsLQC1(R13, F5, F4, 0)                                          #       R:b1    I:b2
+
+       MOV     c21, c11
+       MOV     c22, c11
+       gsLQC1(R12, F3, F2, 1)                                          #       R:a3    I:a4
+       
+       MOV     c23, c11
+       MOV     c24, c11
+       gsLQC1(R13, F7, F6, 1)                                          #       R:b2    I:b3
+
+       FETCH   $0, 0 * SIZE(CO2)
+       MOV     c31, c11
+       MOV     c32, c11
+
+       FETCH   $0, 0 * SIZE(CO1)
+       MOV     c33, c11
+       MOV     c34, c11
+       
+       FETCH   $0, 4 * SIZE(CO2)
+       MOV     c41, c11
+       MOV     c42, c11
+
+       FETCH   $0, 4 * SIZE(CO1)
+       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, 2
+       daddu   PREB, PREB, B                                           #       PREA=A+panel size
+       blez    L, .L15
+       NOP
+
+#else
+
        dsra    L,  K, 2                                                        #       Unroll K 4 times        
        move    BO,  B
 
@@ -218,18 +277,25 @@
        MOV     c24, c11
        gsLQC1(R13, F7, F6, 1)                                          #       R:b2    I:b3
 
+       FETCH   $0, 0 * SIZE(CO2)
        MOV     c31, c11
        MOV     c32, c11
 
+       FETCH   $0, 0 * SIZE(CO1)
        MOV     c33, c11
        MOV     c34, c11
        
+       FETCH   $0, 4 * SIZE(CO2)
        MOV     c41, c11
        MOV     c42, c11
 
+       FETCH   $0, 4 * SIZE(CO1)
        MOV     c43, c11
+
+       daddu   PREB, PREB, B                                           #       PREA=A+panel size
        blez    L, .L15
        MOV     c44, c11
+#endif
 
        .align  5
 
@@ -361,8 +427,13 @@
        .align  5
 
 .L15:
+#ifndef        TRMMKERNEL
        andi    L,  K, 3
        LD      ALPHA_R, 128($sp)
+#else
+       andi    L,  TEMP, 3
+       LD      ALPHA_R, 128($sp)
+#endif
        blez    L, .L18
        LD      ALPHA_I, 136($sp)
 
@@ -408,7 +479,7 @@
        NOP
 
 .L18:
-
+#ifndef TRMMKERNEL
        ADD     c11, c14, c11
        LD      a1,  0 * SIZE(CO1)
        ADD     c12, c13, c12
@@ -458,20 +529,75 @@
        ST      b3, 2 * SIZE(CO2)
        ST      b4, 3 * SIZE(CO2)
 
-       FETCH   $0, 4 * SIZE(CO2)
-       FETCH   $0, 4 * SIZE(CO1)
-       FETCH   $0, 8 * SIZE(CO2)
-       FETCH   $0, 8 * SIZE(CO1)
-       FETCH   $0, 12 * SIZE(CO2)
-       FETCH   $0, 12 * SIZE(CO1)
-       FETCH   $0, 16 * SIZE(CO2)
-       FETCH   $0, 16 * SIZE(CO1)
+#else
+       ADD     c11, c14, c11
+       ADD     c12, c13, c12
+       ADD     c21, c24, c21
+       ADD     c22, c23, c22
 
+       ADD     c31, c34, c31
+       ADD     c32, c33, c32
+       ADD     c41, c44, c41
+       ADD     c42, c43, c42
+
+       daddiu  I, I, -1
+       MUL a1, ALPHA_R, c11
+       MUL a2, ALPHA_R, c12
+       MUL     b1, ALPHA_R, c21
+       MUL     b2, ALPHA_R, c22
+
+       NMSUB   a1, a1, ALPHA_I, c12
+       MADD    a2, a2, ALPHA_I, c11
+       NMSUB   b1, b1, ALPHA_I, c22
+       MADD    b2, b2, ALPHA_I, c21
+
+       MUL     a3, ALPHA_R, c31
+       MUL     a4, ALPHA_R, c32
+       MUL     b3, ALPHA_R, c41
+       MUL     b4, ALPHA_R, c42
+
+       NMSUB   a3, a3, ALPHA_I, c32
+       MADD    a4, a4, ALPHA_I, c31
+       NMSUB   b3, b3, ALPHA_I, c42
+       MADD    b4, b4, ALPHA_I, c41
+
+       ST      a1, 0 * SIZE(CO1)
+       ST      a2, 1 * SIZE(CO1)
+       ST      b1, 2 * SIZE(CO1)
+       ST      b2, 3 * SIZE(CO1)
+
+       ST      a3, 0 * SIZE(CO2)
+       ST      a4, 1 * SIZE(CO2)
+       ST      b3, 2 * SIZE(CO2)
+       ST      b4, 3 * SIZE(CO2)
+
+#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 + ZBASE_SHIFT
+       dsll    TEMP, TEMP, 1 + ZBASE_SHIFT
+
+       daddu   AO, AO, L
+       daddu   BO, BO, TEMP
+#endif
+
+#ifdef LEFT
+       daddiu  KK, KK, 2
+#endif
+#endif
+
+       dsll    PREB, K, 1 + ZBASE_SHIFT                                #       PREA=K*2*2^4
        daddiu  CO1,CO1, 4 * SIZE 
        bgtz    I, .L11
        daddiu  CO2,CO2, 4 * SIZE 
 
-
+       .align  5
 .L30:
        andi    I, M, 1
        daddu   C,   C, LDC                                             #       Change C to next panel
@@ -480,8 +606,16 @@
        blez    I, .L19
        daddu   C,   C, LDC                                             #       Change C to next panel
 
-       dsra    L,  K, 2                                                        #       Unroll K 4 times        
+#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
 
        gsLQC1(R12, F1, F0, 0)                                          #       R:a1    I:a2
        MTC     $0,  c11                                                                #       Clear results regs
@@ -496,6 +630,43 @@
        MOV     c32, c11
 
        FETCH   $0, 0 * SIZE(PREB)
+       MOV     c33, c11
+       MOV     c34, c11
+
+       FETCH   $0, 0 * SIZE(CO1)
+       FETCH   $0, 0 * SIZE(CO2)
+       FETCH   $0, 4 * SIZE(CO1)
+       FETCH   $0, 4 * SIZE(CO2)
+
+#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, .L35
+       NOP
+
+#else
+
+       gsLQC1(R12, F1, F0, 0)                                          #       R:a1    I:a2
+       dsra    L,  K, 2                                                        #       Unroll K 4 times        
+       move    BO,  B
+
+       gsLQC1(R13, F5, F4, 0)                                          #       R:b1    I:b2
+       MTC     $0,  c11                                                                #       Clear results regs
+       MOV     c12, c11
+       
+       gsLQC1(R13, F7, F6, 1)                                          #       R:b2    I:b3
+       MOV     c13, c11
+       MOV     c14, c11
+
+       FETCH   $0, 0 * SIZE(PREB)
+       MOV     c31, c11
+       MOV     c32, c11
+
        FETCH   $0, 0 * SIZE(CO1)
        FETCH   $0, 0 * SIZE(CO2)
        FETCH   $0, 4 * SIZE(CO1)
@@ -504,6 +675,7 @@
        MOV     c33, c11
        blez    L, .L35
        MOV     c34, c11
+#endif
 
        .align  5
 
@@ -582,15 +754,18 @@
 
 
 .L35:
+#ifndef TRMMKERNEL
        andi    L,  K, 3
        LD      ALPHA_R, 128($sp)
-       NOP
+#else
+       andi    L,  TEMP, 3
+       LD      ALPHA_R, 128($sp)
+#endif
        blez    L, .L38
        LD      ALPHA_I, 136($sp)
        .align  5
 
 .L36:
-
        daddiu  L, L, -1
        MADD1   c11, c11, a1, b1                                        #       axc             A1xB1   
        MADD3   c13, c13, a1, b2                                        #       axd
@@ -615,6 +790,7 @@
        gsLQC1(R13, F7, F6, 1)                                          #       R:b2    I:b3
 
 .L38:
+#ifndef        TRMMKERNEL
        ADD     c11, c14, c11
        LD      a1,  0 * SIZE(CO1)
        ADD     c12, c13, c12
@@ -645,10 +821,60 @@
 
        daddiu  CO1,CO1, 2 * SIZE 
        daddiu  CO2,CO2, 2 * SIZE 
+#else
+       ADD     c11, c14, c11
+       ADD     c12, c13, c12
+
+       ADD     c31, c34, c31
+       ADD     c32, c33, c32
+
+       MUL     a1, ALPHA_R, c11
+       MUL     a2, ALPHA_R, c12
+       MUL     a3, ALPHA_R, c31
+       MUL     a4, ALPHA_R, c32
+
+       NMSUB   a1, a1, ALPHA_I, c12
+       MADD    a2, a2, ALPHA_I, c11
+
+       NMSUB   a3, a3, ALPHA_I, c32
+       MADD    a4, a4, ALPHA_I, c31
+
+       ST      a1, 0 * SIZE(CO1)
+       ST      a2, 1 * SIZE(CO1)
+
+       ST      a3, 0 * SIZE(CO2)
+       ST      a4, 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, -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
 
        .align 5
 
 .L19:
+#if defined(TRMMKERNEL) && !defined(LEFT)
+       daddiu  KK, KK, 2
+#endif
+
        bgtz    J, .L10
        move    B, BO
 
@@ -662,11 +888,56 @@
        dsra    I,  M, 1                                                        #       I=M/2
        move    CO1, C
 
+#if defined(TRMMKERNEL) &&  defined(LEFT)
+       move    KK, OFFSET
+#endif
+
        move    AO, A                                                           #       Reset AO
        blez    I, .L29
        daddu   PREA, PREA, A
 
 .L21:
+#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
+       gsLQC1(R12, F1, F0, 0)                                          #       R:a1    I:a2
+       MTC     $0,  c11                                                                #       Clear results regs
+       MOV     c12, c11
+
+       gsLQC1(R13, F5, F4, 0)                                          #       R:b1    I:b2
+       MOV     c13, c11
+       MOV     c14, c11
+
+       gsLQC1(R12, F3, F2, 1)                                          #       R:a3    I:a4
+       MOV     c21, c11
+       MOV     c22, c11
+
+       FETCH   $0, 0 * SIZE(PREA)
+       MOV     c23, c11
+       MOV     c24, c11
+
+       FETCH   $0, 0 * SIZE(CO1)
+       FETCH   $0, 4 * SIZE(CO1)
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+       dsubu   TEMP, K, KK
+#elif defined(LEFT)
+       daddiu  TEMP, KK, 2                                             #       define Mr=2
+#else
+       daddiu  TEMP, KK, 1                                             #       define  NR=1
+#endif
+       dsra    L,  TEMP, 2
+       blez    L, .L25
+       NOP
+
+#else
        dsra    L,  K, 2                                                        #       Unroll K 4 times        
        move    BO,  B
 
@@ -691,8 +962,9 @@
 
        blez    L, .L25
        NOP
+#endif
 
-       .align  3
+       .align  5
 
 .L22:
        gsLQC1(R12, F9, F8, 2)                                          #       Unroll K=1      
@@ -766,15 +1038,18 @@
 
 
 .L25:
+#ifndef TRMMKERNEL
        andi    L,  K, 3
        LD      ALPHA_R, 128($sp)
-
+#else
+       andi    L, TEMP, 3
+       LD      ALPHA_R, 128($sp)
+#endif
        blez    L, .L28
        LD      ALPHA_I, 136($sp)
        .align  3
 
 .L26:
-
        daddiu  L, L, -1
        MADD1   c11, c11, a1, b1                                        #       axc             A1xB1   
        MADD3   c13, c13, a1, b2                                        #       axd
@@ -799,6 +1074,7 @@
        FETCH   $0, 0 * SIZE(PREA)
 
 .L28:
+#ifndef        TRMMKERNEL
        ADD     c11, c14, c11
        LD      a1,  0 * SIZE(CO1)
        ADD     c12, c13, c12
@@ -824,6 +1100,48 @@
        ST      b1, 2 * SIZE(CO1)
        ST      b2, 3 * SIZE(CO1)
 
+#else
+       ADD     c11, c14, c11
+       ADD     c12, c13, c12
+       ADD     c21, c24, c21
+       ADD     c22, c23, c22
+
+       daddiu  I, I, -1
+       MUL     a1, ALPHA_R, c11
+       MUL     a2, ALPHA_R, c12
+       MUL     b1, ALPHA_R, c21
+       MUL     b2, ALPHA_R, c22
+
+       NMSUB   a1, a1, ALPHA_I, c12
+       MADD    a2, a2, ALPHA_I, c11
+       NMSUB   b1, b1, ALPHA_I, c22
+       MADD    b2, b2, ALPHA_I, c21
+
+       ST      a1, 0 * SIZE(CO1)
+       ST      a2, 1 * SIZE(CO1)
+       ST      b1, 2 * SIZE(CO1)
+       ST      b2, 3 * SIZE(CO1)
+
+#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 
        bgtz    I, .L21
        NOP
@@ -833,6 +1151,39 @@
        blez    I, .L999
        NOP
 
+#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
+
+       gsLQC1(R12, F1, F0, 0)                                          #       R:a1    I:a2
+       MTC     $0,  c11                                                                #       Clear results regs
+       MOV     c12, c11
+
+       gsLQC1(R13, F5, F4, 0)                                          #       R:b1    I:b2
+       MOV     c13, c11
+       MOV     c14, c11
+
+       FETCH   $0, 0 * SIZE(PREA)
+       FETCH   $0, 4 * SIZE(PREA)
+
+#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, .L45
+       NOP
+
+#else
        dsra    L,  K, 2                                                        #       Unroll K 4 times        
        move    BO,  B
 
@@ -848,6 +1199,7 @@
        FETCH   $0, 4 * SIZE(PREA)
        blez    L, .L45
        NOP
+#endif
 
        .align  3
 
@@ -892,8 +1244,13 @@
        .align 5
 
 .L45:
+#ifndef        TRMMKERNEL
        andi    L,  K, 3
        LD      ALPHA_R, 128($sp)
+#else
+       andi    L,  TEMP, 3
+       LD      ALPHA_R, 128($sp)
+#endif
        blez    L, .L48
        LD      ALPHA_I, 136($sp)
 
@@ -914,6 +1271,7 @@
        NOP
 
 .L48:
+#ifndef        TRMMKERNEL
        ADD     c11, c14, c11
        ADD     c12, c13, c12
 
@@ -929,7 +1287,40 @@
        ST      a1, 0 * SIZE(CO1)
        ST      a2, 1 * SIZE(CO1)
 
+#else
+       ADD     c11, c14, c11
+       ADD     c12, c13, c12
+
+       MUL     a1, ALPHA_R, c11
+       MUL     a2, ALPHA_R, c12
+
+       NMSUB   a1, a1, ALPHA_I, c12
+       MADD    a2, a2, ALPHA_I, c11
+
+       ST      a1, 0 * SIZE(CO1)
+       ST      a2, 1 * SIZE(CO1)
+
+#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
+
        daddiu  CO1,CO1, 2 * SIZE 
+#endif
 
 
 

diff --git a/param.h b/param.h
index b7f0d66..cab3e68 100644 (file)
--- a/param.h
+++ b/param.h
@@ -1500,7 +1500,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SGEMM_DEFAULT_Q 116
 #define DGEMM_DEFAULT_Q 116
 #define CGEMM_DEFAULT_Q 144
-#define ZGEMM_DEFAULT_Q  60
+#define ZGEMM_DEFAULT_Q 80
 
 #define SGEMM_DEFAULT_R 1000
 #define DGEMM_DEFAULT_R 1000