From 5ee45932e0034f8148653b2dea9214f745e630df Mon Sep 17 00:00:00 2001
From: pagefault <>
Date: Thu, 6 Sep 2001 14:47:07 +0000
Subject: [PATCH] Updated 8-point interpolation

---
 zsnes/src/cpu/dspproc.asm  | 32 ++++++++++++++++++++++++++++++++
 zsnes/src/zip/fir_proc.cpp | 23 ++++++++++++-----------
 2 files changed, 44 insertions(+), 11 deletions(-)

diff --git a/zsnes/src/cpu/dspproc.asm b/zsnes/src/cpu/dspproc.asm
index c65f6a8e..9d49d3e6 100644
--- a/zsnes/src/cpu/dspproc.asm
+++ b/zsnes/src/cpu/dspproc.asm
@@ -3513,6 +3513,11 @@ WaveIndex  times 8 dd 0
 ;%%DontFilter1                  ;
 
 EXTSYM fir_interpolate
+EXTSYM fir_lut
+
+section .data
+NEWSYM fir_tmp,dd 0,0
+section .text
 
 %macro DSPInterpolate 1
 
@@ -3558,10 +3563,37 @@ EXTSYM fir_interpolate
     jmp %%end
 
 %%fir_interpolate
+%if 1
+    xor eax,eax
+    mov ebx,[BRRPlace0+%1*8]
+    mov al,[BRRPlace0+%1*8+3]
+    shl eax,2
+    and ebx,0FFFFFFh
+    add ebx,1000h
+    shr ebx,9
+    and ebx,0FFF0h
+    add ebx,[fir_lut]
+    movq mm0,[eax+PSampleBuf+(%1*26*4)]
+    packssdw mm0,[eax+PSampleBuf+(%1*26*4)+8]
+    movq mm1,[eax+PSampleBuf+(%1*26*4)+16]
+    packssdw mm1,[eax+PSampleBuf+(%1*26*4)+24]
+    movq mm2,[ebx]
+    movq mm3,[ebx+8]
+    pmaddwd mm0,mm2
+    pmaddwd mm1,mm3
+    paddd mm0,mm1
+    movq [fir_tmp],mm0
+    emms
+    mov eax,[fir_tmp]
+    add eax,[fir_tmp+4]
+    sar eax,14
+
+%else
     push dword PSampleBuf+(%1*26*4)
     push dword [BRRPlace0+%1*8]
     call fir_interpolate
     add esp,+8
+%endif
     cmp eax,32767
     jle %%clip1
     mov eax,32767
diff --git a/zsnes/src/zip/fir_proc.cpp b/zsnes/src/zip/fir_proc.cpp
index d77e66e0..860e2726 100644
--- a/zsnes/src/zip/fir_proc.cpp
+++ b/zsnes/src/zip/fir_proc.cpp
@@ -9,11 +9,8 @@
  *          
 */
 
-#ifdef __LINUX__
-#include "../gblhdr.h"
-#else
 #include <math.h>
-#endif
+#include <stdio.h>
 
 #ifndef log2
 inline float log2(float f) {
@@ -55,7 +52,7 @@ inline float log2(float f) {
   ------------------------------------------------------------------------------------------------
 */
 // quantizer scale of window coefs
-#define WFIR_QUANTBITS		15
+#define WFIR_QUANTBITS		14
 #define WFIR_QUANTSCALE		(1L<<WFIR_QUANTBITS)
 #define WFIR_8SHIFT			(WFIR_QUANTBITS-8)
 #define WFIR_16BITSHIFT		(WFIR_QUANTBITS)
@@ -137,11 +134,11 @@ class CzWINDOWEDFIR
 			}
 			return (float)(_LWc*_LSi);
 		}
-		static signed int lut[WFIR_LUTLEN*WFIR_WIDTH];
+		static signed short lut[WFIR_LUTLEN*WFIR_WIDTH];
 		static signed int lut_co[WFIR_CUTOFFLEN*WFIR_WIDTH];
 };
 
-signed int CzWINDOWEDFIR::lut[WFIR_LUTLEN*WFIR_WIDTH];
+signed short CzWINDOWEDFIR::lut[WFIR_LUTLEN*WFIR_WIDTH];
 signed int CzWINDOWEDFIR::lut_co[WFIR_CUTOFFLEN*WFIR_WIDTH];
 
 CzWINDOWEDFIR::CzWINDOWEDFIR()
@@ -161,7 +158,7 @@ CzWINDOWEDFIR::CzWINDOWEDFIR()
 		_LGain = 1.0f/_LGain;
 		for( _LCc=0;_LCc<WFIR_WIDTH;_LCc++ )
 		{	float _LCoef = (float)floor( 0.5 + _LScale*_LCoefs[_LCc]*_LGain );
-			lut[_LIdx+_LCc] = (signed int)( (_LCoef<-_LScale)?-_LScale:((_LCoef>_LScale)?_LScale:_LCoef) );
+			lut[_LIdx+_LCc] = (signed short)( (_LCoef<-_LScale)?-_LScale:((_LCoef>_LScale)?_LScale:_LCoef) );
 		}
 	}
 	for( _LPcl=0;_LPcl<WFIR_CUTOFFLEN;_LPcl++ )
@@ -260,9 +257,13 @@ inline int __fir_interpolate(unsigned int nPos, int *p)
 	return vol;
 }
 
-extern "C" int fir_interpolate(unsigned int nPos, int *p)
-{
-	return __fir_interpolate(nPos, p);
+extern "C" {
+	signed short *fir_lut = &CzWINDOWEDFIR::lut[0];
+
+	int fir_interpolate(unsigned int nPos, int *p)
+	{
+		return __fir_interpolate(nPos, p);
+	}
 }
 
 #define WFIR_CUTOFFSHIFT	(32-(WFIR_CUTOFFBITS+WFIR_LOG2WIDTH))