From 5a61b0af86934869ec2a52c643b9e1c62d3fb39f Mon Sep 17 00:00:00 2001
From: pagefault <>
Date: Thu, 6 Sep 2001 15:35:26 +0000
Subject: [PATCH] Code updates

---
 zsnes/src/cpu/dspproc.asm  | 40 +++++++++++++++++++++++++++++++++++++-
 zsnes/src/zip/fir_proc.cpp | 19 +++++++++++-------
 2 files changed, 51 insertions(+), 8 deletions(-)

diff --git a/zsnes/src/cpu/dspproc.asm b/zsnes/src/cpu/dspproc.asm
index 8c8a3ec3..937fd7fb 100644
--- a/zsnes/src/cpu/dspproc.asm
+++ b/zsnes/src/cpu/dspproc.asm
@@ -1764,7 +1764,8 @@ section .text
     mov [prev0],edx
 %endmacro
 
-EXTSYM fir_downsample
+;EXTSYM fir_downsample
+EXTSYM fir_lut_co
 
 %macro ProcessDynamicLowPass 0
     mov ecx,[curvoice]
@@ -1948,6 +1949,42 @@ ALIGN16
     mov [ebx+20*4],eax
     mov [ebx+25*4],eax
 
+%if 1
+    mov ecx,16
+    sub edx,0F80000h
+    shr edx,15
+    and edx,7FFF0h
+    add edx,[fir_lut_co]
+    movq mm2,[edx]
+    movq mm3,[edx+8]
+%%DLPF_fir_loop
+    movq mm0,[ebx]
+    packssdw mm0,[ebx+8]
+    movq mm1,[ebx+16]
+    packssdw mm1,[ebx+24]
+    pmaddwd mm0,mm2
+    pmaddwd mm1,mm3
+    paddd mm0,mm1
+    movq [fir_tmp],mm0
+    mov edx,[fir_tmp]
+    add edx,[fir_tmp+4]
+    sar edx,14
+    cmp edx,32767
+    jle %%DLPF_fir_clip
+    mov edx,32767
+%%DLPF_fir_clip
+    cmp edx,-32768
+    jge %%DLPF_fir_clip2
+    mov edx,-32768
+%%DLPF_fir_clip2
+    mov [edi],dx
+    add edi,2
+    add ebx,4
+    dec ecx
+    jnz %%DLPF_fir_loop
+    emms
+
+%else
 
     push edi
     push ebx
@@ -1955,6 +1992,7 @@ ALIGN16
     call fir_downsample
     add esp,+12
     add edi,32
+%endif
     ret
 %endmacro
 
diff --git a/zsnes/src/zip/fir_proc.cpp b/zsnes/src/zip/fir_proc.cpp
index d4e341f6..99f04fda 100644
--- a/zsnes/src/zip/fir_proc.cpp
+++ b/zsnes/src/zip/fir_proc.cpp
@@ -70,7 +70,7 @@ inline float log2(float f) {
 // cutoff (1.0 == pi/2)
 #define WFIR_CUTOFF			0.90f
 #define WFIR_CUTOFFBITS		12
-#define WFIR_CUTOFFLEN		(1L<<(WFIR_CUTOFFBITS))
+#define WFIR_CUTOFFLEN		((1L<<(WFIR_CUTOFFBITS))+1)
 // wfir type
 #define WFIR_HANN			0
 #define WFIR_HAMMING		1
@@ -139,11 +139,11 @@ class CzWINDOWEDFIR
 			return (float)(_LWc*_LSi);
 		}
 		static signed short lut[WFIR_LUTLEN*WFIR_WIDTH];
-		static signed int lut_co[WFIR_CUTOFFLEN*WFIR_WIDTH];
+		static signed short lut_co[WFIR_CUTOFFLEN*WFIR_WIDTH];
 };
 
 signed short CzWINDOWEDFIR::lut[WFIR_LUTLEN*WFIR_WIDTH];
-signed int CzWINDOWEDFIR::lut_co[WFIR_CUTOFFLEN*WFIR_WIDTH];
+signed short CzWINDOWEDFIR::lut_co[WFIR_CUTOFFLEN*WFIR_WIDTH];
 
 CzWINDOWEDFIR::CzWINDOWEDFIR()
 {	int _LPcl;
@@ -177,7 +177,7 @@ CzWINDOWEDFIR::CzWINDOWEDFIR()
 		for( _LCc=0; _LCc<WFIR_WIDTH;_LCc++ )
 		{
 			float _LCoef = (float)floor( 0.5 + _LScale*_LCoefs[_LCc]*_LGain );
-			lut_co[_LIdx+_LCc] = (signed int)( (_LCoef<-_LScale)?-_LScale:((_LCoef>_LScale)?_LScale:_LCoef) );
+			lut_co[_LIdx+_LCc] = (signed short)( (_LCoef<-_LScale)?-_LScale:((_LCoef>_LScale)?_LScale:_LCoef) );
 		}
 	}
 }
@@ -239,6 +239,7 @@ float coef( int _PCnr, float _POfs, float _PCut, int _PWidth, int _PType ) //flo
 CzWINDOWEDFIR sfir;
 
 extern "C" signed short *fir_lut = &CzWINDOWEDFIR::lut[0];
+extern "C" signed short *fir_lut_co = &CzWINDOWEDFIR::lut_co[0];
 
 
 #if 0
@@ -273,8 +274,10 @@ extern "C" int fir_interpolate(unsigned int nPos, int *p)
 
 #endif
 
-#define WFIR_CUTOFFSHIFT	(32-(WFIR_CUTOFFBITS+WFIR_LOG2WIDTH))
-#define WFIR_CUTOFFMASK	((((1L<<(32-WFIR_CUTOFFSHIFT))-1)&~((1L<<WFIR_LOG2WIDTH)-1)))
+#if 0
+
+#define WFIR_CUTOFFSHIFT	(32-(WFIR_CUTOFFBITS+1+WFIR_LOG2WIDTH))
+#define WFIR_CUTOFFMASK	((((1L<<(33-WFIR_CUTOFFSHIFT))-1)&~((1L<<WFIR_LOG2WIDTH)-1)))
 #define WFIR_CUTOFFHALVE	(1L<<(32-(WFIR_CUTOFFBITS+1)))
 
 inline void __fir_downsample(unsigned int freq, signed int *p, signed short *out)
@@ -327,4 +330,6 @@ inline void __fir_downsample(unsigned int freq, signed int *p, signed short *out
 extern "C" void fir_downsample(unsigned int freq, signed int *p, signed short *out)
 {
 	__fir_downsample(freq, p, out);
-}
\ No newline at end of file
+}
+
+#endif
\ No newline at end of file