updated haswell optimized sgmv_n kernel

wernsaar · wernsaar · commit 7fa7ea3e1e73 · 2014-08-05T08:04:47.000+02:00
diff --git a/kernel/x86_64/sgemv_n_microk_haswell-2.c b/kernel/x86_64/sgemv_n_microk_haswell-2.c
@@ -35,6 +35,7 @@ static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
 
 	__asm__  __volatile__
 	(
+	"vzeroupper			 \n\t"
 	"vbroadcastss    (%2), %%ymm12	 \n\t"	// x0 
 	"vbroadcastss   4(%2), %%ymm13	 \n\t"	// x1 
 	"vbroadcastss   8(%2), %%ymm14	 \n\t"	// x2 
@@ -64,6 +65,7 @@ static void sgemv_kernel_16x4( long n, float **ap, float *x, float *y)
         "addq		$16, %0	  	 	      \n\t"
 	"subq	        $16, %1			      \n\t"		
 	"jnz		.L01LOOP%=		      \n\t"
+	"vzeroupper			 \n\t"
 
 	:
         : 

Original file line number	Diff line number	Diff line change
`@@ -35,6 +35,7 @@ static void sgemv_kernel_16x4( long n, float *ap, float x, float *y)`
`35`	`35`
`36`	`36`	`__asm__ __volatile__`
`37`	`37`	`(`
	`38`	`+ "vzeroupper \n\t"`
`38`	`39`	`"vbroadcastss (%2), %%ymm12 \n\t" // x0`
`39`	`40`	`"vbroadcastss 4(%2), %%ymm13 \n\t" // x1`
`40`	`41`	`"vbroadcastss 8(%2), %%ymm14 \n\t" // x2`
`@@ -64,6 +65,7 @@ static void sgemv_kernel_16x4( long n, float *ap, float x, float *y)`
`64`	`65`	`"addq $16, %0 \n\t"`
`65`	`66`	`"subq $16, %1 \n\t"`
`66`	`67`	`"jnz .L01LOOP%= \n\t"`
	`68`	`+ "vzeroupper \n\t"`
`67`	`69`
`68`	`70`	`:`
`69`	`71`	`:`