diff options
Diffstat (limited to 'kernel/x86/zgemm_kernel_2x2_sse3.S')
| -rw-r--r-- | kernel/x86/zgemm_kernel_2x2_sse3.S | 46 |
1 files changed, 23 insertions, 23 deletions
diff --git a/kernel/x86/zgemm_kernel_2x2_sse3.S b/kernel/x86/zgemm_kernel_2x2_sse3.S index 23afa8f21..4bca5ff68 100644 --- a/kernel/x86/zgemm_kernel_2x2_sse3.S +++ b/kernel/x86/zgemm_kernel_2x2_sse3.S @@ -41,7 +41,7 @@ #define STACK 16 #define ARGS 0 - + #define STACK_M 4 + STACK + ARGS(%esi) #define STACK_N 8 + STACK + ARGS(%esi) #define STACK_K 12 + STACK + ARGS(%esi) @@ -268,7 +268,7 @@ movss %xmm4, KK #ifndef LEFT negl KK -#endif +#endif #endif sall $ZBASE_SHIFT, LDC @@ -281,7 +281,7 @@ #if defined(TRMMKERNEL) && defined(LEFT) movl OFFSET, %eax movl %eax, KK -#endif +#endif /* Copying to Sub Buffer */ leal BUFFER, %ecx @@ -360,7 +360,7 @@ leal (, %eax, 8), %eax leal (AA, %eax, 2), AA leal (BB, %eax, 4), BB -#endif +#endif movaps 0 * SIZE(AA), %xmm0 pxor %xmm4, %xmm4 @@ -379,7 +379,7 @@ #elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movl K, %eax subl KK, %eax - movl %eax, KKK + movl %eax, KKK #else movl KK, %eax #ifdef LEFT @@ -395,7 +395,7 @@ andl $-8, %eax sall $4, %eax je .L15 -.L1X: +.L1X: KERNEL1(32 * 0) KERNEL2(32 * 0) KERNEL3(32 * 0) @@ -588,7 +588,7 @@ jne .L11 ALIGN_4 #endif - + .L15: #ifndef TRMMKERNEL movl K, %eax @@ -714,7 +714,7 @@ leal (, %eax, 8), %eax leal (AA, %eax, 1), AA leal (BB, %eax, 4), BB -#endif +#endif movddup 0 * SIZE(AA), %xmm0 pxor %xmm4, %xmm4 @@ -728,7 +728,7 @@ #elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movl K, %eax subl KK, %eax - movl %eax, KKK + movl %eax, KKK #else movl KK, %eax #ifdef LEFT @@ -822,7 +822,7 @@ decl %eax jne .L41 ALIGN_4 - + .L42: #ifndef TRMMKERNEL movl K, %eax @@ -859,12 +859,12 @@ movhlps %xmm6, %xmm5 #if defined(NR) || defined(NC) || defined(TR) || defined(TC) || \ - defined(RR) || defined(RC) || defined(CR) || defined(CC) + defined(RR) || defined(RC) || defined(CR) || defined(CC) cmpeqps %xmm7, %xmm7 pslld $31, %xmm7 xorps %xmm7, %xmm5 -#endif - +#endif + #if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \ defined(NR) || defined(NC) || defined(TR) || defined(TC) shufps $0xb1, %xmm5, %xmm5 @@ -934,7 +934,7 @@ #if defined(TRMMKERNEL) && defined(LEFT) movl OFFSET, %eax movl %eax, KK -#endif +#endif /* Copying to Sub Buffer */ leal BUFFER, %ecx @@ -1009,7 +1009,7 @@ leal (, %eax, 8), %eax leal (AA, %eax, 2), AA leal (BB, %eax, 2), BB -#endif +#endif movaps 0 * SIZE(AA), %xmm0 pxor %xmm4, %xmm4 @@ -1029,7 +1029,7 @@ #elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movl K, %eax subl KK, %eax - movl %eax, KKK + movl %eax, KKK #else movl KK, %eax #ifdef LEFT @@ -1107,7 +1107,7 @@ decl %eax jne .L111 ALIGN_4 - + .L112: #ifndef TRMMKERNEL movl K, %eax @@ -1208,7 +1208,7 @@ leal (, %eax, 8), %eax leal (AA, %eax, 1), AA leal (BB, %eax, 2), BB -#endif +#endif movddup 0 * SIZE(AA), %xmm0 pxor %xmm4, %xmm4 @@ -1222,7 +1222,7 @@ #elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movl K, %eax subl KK, %eax - movl %eax, KKK + movl %eax, KKK #else movl KK, %eax #ifdef LEFT @@ -1284,7 +1284,7 @@ decl %eax jne .L141 ALIGN_4 - + .L142: #ifndef TRMMKERNEL movl K, %eax @@ -1317,12 +1317,12 @@ movhlps %xmm4, %xmm5 #if defined(NR) || defined(NC) || defined(TR) || defined(TC) || \ - defined(RR) || defined(RC) || defined(CR) || defined(CC) + defined(RR) || defined(RC) || defined(CR) || defined(CC) cmpeqps %xmm7, %xmm7 pslld $31, %xmm7 xorps %xmm7, %xmm5 -#endif - +#endif + #if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \ defined(NR) || defined(NC) || defined(TR) || defined(TC) shufps $0xb1, %xmm5, %xmm5 |