sse opt

git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@3553 b3059339-0415-0410-9bf9-f77b7e298cf2

1 files changed, 65 insertions, 0 deletions

diff --git a/liba52/imdct.c b/liba52/imdct.c
index 449467f66c..54b94d1fc5 100644
--- a/liba52/imdct.c
+++ b/liba52/imdct.c
@@ -87,6 +87,7 @@ static float __attribute__((aligned(16))) sseW5[128];
 static float __attribute__((aligned(16))) sseW6[256];
 static float __attribute__((aligned(16))) *sseW[7]=
 	{NULL /*sseW0*/,sseW1,sseW2,sseW3,sseW4,sseW5,sseW6};
+static float __attribute__((aligned(16))) sseWindow[256];
 #else
 static complex_t buf[128];
 #endif
@@ -488,15 +489,72 @@ imdct_do_512(sample_t data[],sample_t delay[], sample_t bias)
     window_ptr = imdct_window;
 
     /* Window and convert to real valued signal */
+#ifdef HAVE_SSE
+	asm volatile(
+		"xorl %%edi, %%edi			\n\t"  // 0
+		"xorl %%esi, %%esi			\n\t"  // 0
+		"movss %3, %%xmm2			\n\t"  // bias
+		"shufps $0x00, %%xmm2, %%xmm2		\n\t"  // bias, bias, ...
+		".balign 16				\n\t"
+		"1:					\n\t"
+		"movlps (%0, %%esi), %%xmm0		\n\t" // ? ? A ?
+		"movlps 8(%0, %%esi), %%xmm1		\n\t" // ? ? C ?
+		"movhps -16(%0, %%edi), %%xmm1		\n\t" // ? D C ?
+		"movhps -8(%0, %%edi), %%xmm0		\n\t" // ? B A ?
+		"shufps $0x99, %%xmm1, %%xmm0		\n\t" // D C B A
+		"mulps sseWindow(%%esi), %%xmm0		\n\t"
+		"addps (%2, %%esi), %%xmm0		\n\t"
+		"addps %%xmm2, %%xmm0			\n\t"
+		"movaps %%xmm0, (%1, %%esi)		\n\t"
+		"addl $16, %%esi			\n\t"
+		"subl $16, %%edi			\n\t"
+		"cmpl $512, %%esi			\n\t" 
+		" jb 1b					\n\t"
+		:: "r" (buf+64), "r" (data_ptr), "r" (delay_ptr), "m" (bias)
+		: "%esi", "%edi"
+	);
+	data_ptr+=128;
+	delay_ptr+=128;
+	window_ptr+=128;
+#else
     for(i=0; i< 64; i++) { 
 	*data_ptr++   = -buf[64+i].imag   * *window_ptr++ + *delay_ptr++ + bias; 
 	*data_ptr++   =  buf[64-i-1].real * *window_ptr++ + *delay_ptr++ + bias; 
     }
+#endif
 
+#ifdef HAVE_SSE
+	asm volatile(
+		"movl $1024, %%edi			\n\t"  // 512
+		"xorl %%esi, %%esi			\n\t"  // 0
+		"movss %3, %%xmm2			\n\t"  // bias
+		"shufps $0x00, %%xmm2, %%xmm2		\n\t"  // bias, bias, ...
+		".balign 16				\n\t"
+		"1:					\n\t"
+		"movlps (%0, %%esi), %%xmm0		\n\t" // ? ? ? A
+		"movlps 8(%0, %%esi), %%xmm1		\n\t" // ? ? ? C
+		"movhps -16(%0, %%edi), %%xmm1		\n\t" // D ? ? C
+		"movhps -8(%0, %%edi), %%xmm0		\n\t" // B ? ? A
+		"shufps $0xCC, %%xmm1, %%xmm0		\n\t" // D C B A
+		"mulps 512+sseWindow(%%esi), %%xmm0		\n\t"
+		"addps (%2, %%esi), %%xmm0		\n\t"
+		"addps %%xmm2, %%xmm0			\n\t"
+		"movaps %%xmm0, (%1, %%esi)		\n\t"
+		"addl $16, %%esi			\n\t"
+		"subl $16, %%edi			\n\t"
+		"cmpl $512, %%esi			\n\t" 
+		" jb 1b					\n\t"
+		:: "r" (buf), "r" (data_ptr), "r" (delay_ptr), "m" (bias)
+		: "%esi", "%edi"
+	);
+	data_ptr+=128;
+	window_ptr+=128;
+#else
     for(i=0; i< 64; i++) { 
 	*data_ptr++  = -buf[i].real       * *window_ptr++ + *delay_ptr++ + bias; 
 	*data_ptr++  =  buf[128-i-1].imag * *window_ptr++ + *delay_ptr++ + bias; 
     }
+#endif
 
     /* The trailing edge of the window goes into the delay line */
     delay_ptr = delay;
@@ -710,6 +768,13 @@ void imdct_init (uint32_t mm_accel)
 		}
 	    }
 	}
+
+	for(i=0; i<128; i++)
+	{
+		sseWindow[2*i+0]= -imdct_window[2*i+0];
+		sseWindow[2*i+1]=  imdct_window[2*i+1];
+	}
+
 #endif
 	
 	imdct_512 = imdct_do_512;