From b6e4b5ac8c4c85693f953ad920bee4451236d500 Mon Sep 17 00:00:00 2001
From: Vince Weaver <vince@deater.net>
Date: Fri, 23 Feb 2018 01:03:18 -0500
Subject: [PATCH] chiptune_ update krw decode

lz4 is taking 0.6s!
---
 gr-sim/krw_decode.c | 374 +++++++++++++++++++++++---------------------
 1 file changed, 192 insertions(+), 182 deletions(-)

diff --git a/gr-sim/krw_decode.c b/gr-sim/krw_decode.c
index 01e0d0e0..b967da9c 100644
--- a/gr-sim/krw_decode.c
+++ b/gr-sim/krw_decode.c
@@ -8,13 +8,14 @@ static unsigned char input[MAX_INPUT];
 
 #define LZ4_BUFFER	0x2000
 #define ORGOFFSET	0x6000
-
-
-#define src	0x0
-#define dst	0x2
 #define end	0x4
-#define count	0x6
-#define delta	0x8
+
+#define LZ4_SRC 0x00
+#define LZ4_DST 0x02
+#define LZ4_END 0x04
+#define COUNT   0x06
+#define DELTA   0x08
+
 
 #define CH	0x24
 #define CV	0x25
@@ -22,249 +23,221 @@ static unsigned char input[MAX_INPUT];
 #define	OUTL	0xFE
 #define OUTH	0xFF
 
+static int cycles;
+
 static void getsrc(void) {
 //getsrc:
-	a=ram[y_indirect(src,y)];		// lda 	(src), y
-
-	printf("LOAD %02X%02X: %02X\n",ram[src+1],ram[src],a);
-
-	ram[src]++;				//inc	src
-	if (ram[src]!=0) goto done_getsrc;	//bne	+
-	ram[src+1]++;				//inc	src+1
-done_getsrc: ;
-
+	a=ram[y_indirect(LZ4_SRC,y)];	cycles+=5;	// lda 	(LZ4_SRC), y
+	ram[LZ4_SRC]++;			cycles+=5;	//inc	LZ4_SRC
+					cycles+=2;
+	if (ram[LZ4_SRC]!=0) {
+					cycles+=1;
+		goto done_getsrc;			//bne	+
+	}
+	ram[LZ4_SRC+1]++;		cycles+=5;	//inc	LZ4_SRC+1
+done_getsrc:
+	cycles+=6;
 	//+	rts
 }
 
 void buildcount(void) {
-	printf("\tBUILDCOUNT: A=0x%x\n",a);
 //buildcount:
-	x=1;					// ??		// ldx	#1
-	ram[count+1]=x; 			// ??		// stx	count+1
-	cmp(0xf);				// if 15, more complicated // cmp	#$0f
-	if (z==0) goto done_buildcount;		// otherwise A is count // bne	++
+	x=1;		cycles+=2;		// ??		// ldx	#1
+	ram[COUNT+1]=x; cycles+=3;		// ??		// stx	COUNT+1
+	cmp(0xf);	cycles+=2;		// if 15, more complicated // cmp	#$0f
+	cycles+=2;
+	if (z==0) {
+		cycles+=1;
+		goto done_buildcount;		// otherwise A is count // bne	++
+	}
 buildcount_loop:
-	ram[count]=a;				//-	sta	count
-//	printf("MBC ");
-	getsrc();				//jsr	getsrc
-	printf("\tADDITIONAL BUILDCOUNT 0x%x, adding 0x%x\n",a,ram[count]);
-	x=a;					//tax
-	c=0;					//clc
-	adc(ram[count]);			//adc	count
-	printf("\tGOT 0x%x c=%d\n",a,c);
-	if (c==0) goto skip_buildcount;		// bcc	+
-	ram[count+1]++;				//inc	count+1
+	ram[COUNT]=a;	cycles+=3;		//-	sta	count
+	getsrc();	cycles+=6;		//jsr	getsrc
+	x=a;		cycles+=2;		//tax
+	c=0;		cycles+=2;		//clc
+	adc(ram[COUNT]);cycles+=3;		//adc	COUNT
+	cycles+=2;
+	if (c==0) {
+		cycles+=1;
+		goto skip_buildcount;		// bcc	+
+	}
+	ram[COUNT+1]++;	cycles+=5;		//inc	COUNT+1
 skip_buildcount:
-	printf("\tUPDATED COUNT %02X%02X\n",ram[count+1],a);
-	x++;					// check if x is 255	//+	inx
-	if (x==0) goto buildcount_loop;		// if so, add in next byte //beq	-
-done_buildcount: ;				//++	rts
-	printf("\tBUILDCOUNT= %02X%02X r[c+1]=%02X r[c]=%02X a=%02X x=%02X\n",
-		ram[count+1],a,ram[count+1],ram[count],a,x);
+	x++;		cycles+=2;		// check if x is 255	//+	inx
+	if (x==0) {
+		cycles+=1;
+		goto buildcount_loop;		// if so, add in next byte //beq	-
+	}
+done_buildcount:
+	cycles+=6;				//++	rts
+
 }
 
 
 static void putdst(void) {
-//	printf("PUTADDR=%04X\n",y_indirect(dst,y));
-						// putdst:
-	ram[y_indirect(dst,y)]=a;		//	sta 	(dst), y
-	if (y!=0) printf("ERROR ERROR ERROR ERROR ERROR\n");
-	printf("\t\tPUT: %02X%02X = %02X\n",ram[dst+1],ram[dst],a);
-	ram[dst]++;				//	inc	dst
-	if (ram[dst]!=0) goto putdst_end;	//	bne	+
-	ram[dst+1]++;				//	inc	dst+1
+							// putdst:
+	ram[y_indirect(LZ4_DST,y)]=a; cycles+=6;	//	sta 	(LZ4_DST), y
+	ram[LZ4_DST]++;		cycles+=5;		//	inc	LZ4_DST
+	cycles+=2;
+	if (ram[LZ4_DST]!=0) {
+		cycles+=1;
+		goto putdst_end;			//	bne	+
+	}
+	ram[LZ4_DST+1]++;	cycles+=5;		//	inc	LZ4_DST+1
 putdst_end:;
-						//+	rts
+	cycles+=6;					//+	rts
 
 }
 
 static void getput(void) {
 						// getput:
-	printf("GP ");
+
 	getsrc();				// jsr	getsrc
+	cycles+=6;
 	putdst();				// ; fallthrough
 }
 
-// 202 -> 201 -> 100 -> 1FF
-// 201 -> 100 -> 1FF
-// 200 -> 2ff -> 2fe
-// 1ff -> 1fe
 
 static void docopy(void) {
-	printf("\tDOCOPY ENTRY: %02X%02X\n",ram[count+1],x);
 						// docopy:
 docopy_label:
-	printf("\tDOCOPY %02X%02X: ",ram[count+1],x);
-	getput();				// jsr	getput
-	x--;					// dex
-	if (x!=0) goto docopy_label;		// bne	docopy
-	ram[count+1]--;				// dec	count+1
-	if (ram[count+1]!=0) goto docopy_label;	//bne	docopy
+	getput();	cycles+=6;		// jsr	getput
+	x--;		cycles+=2;		// dex
+	cycles+=2;
+	if (x!=0) {
+		cycles+=1;
+		goto docopy_label;		// bne	docopy
+	}
+	ram[COUNT+1]--;		cycles+=5;	// dec	COUNT+1
+	cycles+=2;
+	if (ram[COUNT+1]!=0) {
+		cycles++;
+		goto docopy_label;	//bne	docopy
+	}
+	cycles+=6;
 						//rts
 }
 
 
-
+#define orgoff	0x6000
 
 
 int lz4_decode(void) {
 
 	FILE *fff;
 
-	int size;
-	short orgoff,paksize,pakoff;
+	cycles=0;
 
 	//LZ4 data decompressor for Apple II
 	//Peter Ferrie (peter.ferrie@gmail.com)
+// lz4_decode:
 
-	//init	=	0 ;set to 1 if you know the values
-	//hiunp	=	0 ;unpacker entirely in high memory
-	//hipak	=	0 ;packed data entirely in high memory (requires hiunp)
+	a=ram[LZ4_SRC];		cycles+=3;	// lda     LZ4_SRC
+	c=0;			cycles+=2;	// clc
+	adc(ram[LZ4_END]);	cycles+=3;	// adc     LZ4_END
+	ram[LZ4_END]=a;		cycles+=3;	// sta     LZ4_END
+        a=ram[LZ4_SRC+1];	cycles+=3;	// lda     LZ4_SRC+1
+	adc(ram[LZ4_END+1]);	cycles+=3;	// adc     LZ4_END+1
+	ram[LZ4_END+1]=a;	cycles+=3;	// sta     LZ4_END+1
 
-	//oep = 0; //first unpacked byte to run, you must set this by yourself
-	orgoff = ORGOFFSET; //offset of first unpacked byte, you must set this by yourself
-	paksize	= size-0xb-8;
-			// minus 4 for checksum at end
-			// not sure what other 4 is from?
-			// block checksum? though had that disabled?
+	a=high(orgoff);		cycles+=2;	// lda     #>orgoff                ; original unpacked data offset
+        ram[LZ4_DST+1]=a;	cycles+=3;	// sta     LZ4_DST+1
+	a=low(orgoff);		cycles+=2;	// lda     #<orgoff
+	ram[LZ4_DST]=a;		cycles+=3;	// sta     LZ4_DST
 
-		//size of packed data, you must set this by yourself if hiunp=0
-	pakoff = LZ4_BUFFER+11; // 11 byte offset to data?
-
-
-//LCBANK2	=	$c083
-//MOVE	=	$fe2c
-
-	a=(pakoff&0xff);	//lda	#<pakoff ;packed data offset
-	ram[src]=a;		//sta	src
-	a=(pakoff+paksize)&0xff;//lda	#<(pakoff+paksize) ;packed data size
-	ram[end]=a;		// sta	end
-	a=(pakoff>>8);		//lda	#>pakoff
-	ram[src+1]=a;		//sta	src+1
-	a=(pakoff+paksize)>>8;	//lda	#>(pakoff+paksize)
-	ram[end+1]=a;		//	sta	end+1
-	a=(orgoff>>8);		//lda	#>orgoff ;original unpacked data offset
-	ram[dst+1]=a;		//sta	dst+1
-	a=(orgoff&0xff);	//lda	#<orgoff
-	ram[dst]=a;		// sta	dst
-
-	printf("packed size: raw=%x, adj=%x\n",size,paksize);
-	printf("packed addr: %02X%02X\n",ram[src+1],ram[src]);
+//	printf("packed size: raw=%x, adj=%x\n",size,paksize);
+	printf("packed addr: %02X%02X\n",ram[LZ4_SRC+1],ram[LZ4_SRC]);
 	printf("packed end : %02X%02X\n",ram[end+1],ram[end]);
-	printf("dest addr  : %02X%02X\n",ram[dst+1],ram[dst]);
+	printf("dest addr  : %02X%02X\n",ram[LZ4_DST+1],ram[LZ4_DST]);
 
-// https://github.com/lz4/lz4/wiki/lz4_Frame_format.md
-
-// Should: check for magic number 04 22 4d 18
-//	FLG: 64 in our case (01=version, block.index=1, block.checksum=0
-//		size=0, checksum=1, reserved
-//	MAX Blocksize: 40 (64kB)
-//	HEADER CHECKSUM: a7
-//	BLOCK HEADER: 4 bytes (le)  If highest bit set, uncompressed!
-
-
-//unpack:				//;unpacker entrypoint
-//	goto unpmain;			// jmp	unpmain
-
-//unpack:
-	y=0;				// used for offset	//ldy	#0
+//unpmain:
+	y=0;		cycles+=2;	// used for offset	//ldy	#0
 
 parsetoken:
-	printf("LOAD TOKEN: ");
-	getsrc();						// jsr	getsrc
+	getsrc();	cycles+=6;	// jsr	getsrc
 					// get token
-	pha();				// save for later	// pha
-	lsr();				// num literals in top 4	// lsr
-	lsr();							// lsr
-	lsr();							// lsr
-	lsr();							// lsr
-	if (a==0) goto copymatches; 	// if zero, no literals // beq	copymatches
+	pha();		cycles+=3;	// save for later	// pha
+	lsr();		cycles+=2;	// num literals in top 4// lsr
+	lsr();		cycles+=2;				// lsr
+	lsr();		cycles+=2;				// lsr
+	lsr();		cycles+=2;				// lsr
+	cycles+=2;
+	if (a==0) {
+		cycles+=1;
+		goto copymatches; 	// if zero, no literals // beq	copymatches
+	}
 
-	buildcount();			// otherwise, build the count // jsr	buildcount
+	buildcount();	cycles+=6;	// otherwise, build the count // jsr	buildcount
 
-	x=a;				// tax
-	docopy();			// jsr	docopy
-	a=ram[src];			// lda	src
-	cmp(ram[end]);			// cmp	end
-	a=ram[src+1];			// lda	src+1
+	x=a;			cycles+=2;	// tax
+	docopy();		cycles+=6;	// jsr	docopy
+	a=ram[LZ4_SRC];		cycles+=3;	// lda	LZ4_SRC
+	cmp(ram[end]);		cycles+=3;	// cmp	end
+	a=ram[LZ4_SRC+1];	cycles+=3;	// lda	LZ4_SRC+1
 
-	sbc(ram[end+1]);		// sbc	end+1
+	sbc(ram[end+1]);	cycles+=3;	// sbc	end+1
+	cycles+=2;
 	if (c) {
 		printf("Done!\n");
-		printf("src        : %02X%02X\n",ram[src+1],ram[src]);
+		printf("src        : %02X%02X\n",ram[LZ4_SRC+1],ram[LZ4_SRC]);
 		printf("packed end : %02X%02X\n",ram[end+1],ram[end]);
+		cycles+=1;
 		goto done;		// bcs	done
 	}
 
 copymatches:
-	printf("\tDELTAL ");
-	getsrc();					// jsr	getsrc
-	ram[delta]=a;					// sta	delta
-	printf("\tDELTAH ");
-	getsrc();					// jsr	getsrc
-	ram[delta+1]=a;					// sta	delta+1
-	printf("\tDELTA is %02X%02X\n",ram[delta+1],ram[delta]);
-	pla();				// restore token	// pla
-	a=a&0xf;			// get bottom 4 bits	// and	#$0f
-	buildcount();			// jsr	buildcount
+	getsrc();		cycles+=6;	// jsr	getsrc
+	ram[DELTA]=a;		cycles+=3;	// sta	DELTA
+	getsrc();		cycles+=6;	// jsr	getsrc
+	ram[DELTA+1]=a;		cycles+=3;	// sta	DELTA+1
+	pla();			cycles+=3;	// restore token	// pla
+	a=a&0xf;		cycles+=2;	// get bottom 4 bits	// and	#$0f
+	buildcount();		cycles+=6;	// jsr	buildcount
 
-	c=0;				// clc
-	adc(4);				// adc	#4
-	x=a;				// tax
-	if (x==0) goto copy_skip;	//BUGFIX // beq  +
-	if (c==0) goto copy_skip;	// bcc	+
-	ram[count+1]++;			// inc	count+1
+	c=0;			cycles+=2;	// clc
+	adc(4);			cycles+=2;	// adc	#4
+	x=a;			cycles+=2;	// tax
+	cycles+=2;
+	if (x==0) {
+		cycles+=1;
+		goto copy_skip;	//BUGFIX // beq  +
+	}
+	cycles+=2;
+	if (c==0) {
+		cycles+=1;
+		goto copy_skip;	// bcc	+
+	}
+	ram[COUNT+1]++;		cycles+=5;	// inc	count+1
 copy_skip:
-	a=ram[src+1];			//+	lda	src+1
-	pha();				// pha
-	a=ram[src];			// lda	src
-	pha();				// pha
-	printf("\tSAVED SRC: %02X%02X\n",ram[src+1],ram[src]);
-//	printf("CALCULATING: DST %02X%02X - DELTA %02X%02X\n",ram[dst+1],ram[dst],ram[delta+1],ram[delta]);
-	c=1;				// sec
-	a=ram[dst];			// lda	dst
-	sbc(ram[delta]);		// sbc	delta
-	ram[src]=a;			// sta	src
-	a=ram[dst+1];			// lda	dst+1
-	sbc(ram[delta+1]);		// sbc	delta+1
-	ram[src+1]=a;			// sta	src+1
-	printf("\tNEW SRC: %02X%02X\n",ram[src+1],ram[src]);
-	docopy();			// jsr	docopy
-	pla();				// pla
-	ram[src]=a;			// sta	src
-	pla();				// pla
-	ram[src+1]=a;			// sta	src+1
-	printf("\tRESTORED SRC: %02X%02X\n",ram[src+1],ram[src]);
+	a=ram[LZ4_SRC+1];	cycles+=3;	//+	lda	src+1
+	pha();			cycles+=3;	// pha
+	a=ram[LZ4_SRC];		cycles+=3;	// lda	src
+	pha();			cycles+=3;	// pha
+	c=1;			cycles+=2;	// sec
+	a=ram[LZ4_DST];		cycles+=3;	// lda	LZ4_DST
+	sbc(ram[DELTA]);	cycles+=3;	// sbc	DELTA
+	ram[LZ4_SRC]=a;		cycles+=3;	// sta	LZ4_SRC
+	a=ram[LZ4_DST+1];	cycles+=3;	// lda	LZ4_DST+1
+	sbc(ram[DELTA+1]);	cycles+=3;	// sbc	DELTA+1
+	ram[LZ4_SRC+1]=a;	cycles+=3;	// sta	LZ4_SRC+1
+
+	docopy();		cycles+=6;	// jsr	docopy
+	pla();			cycles+=3;	// pla
+	ram[LZ4_SRC]=a;		cycles+=3;	// sta	LZ4_SRC
+	pla();			cycles+=3;	// pla
+	ram[LZ4_SRC+1]=a;	cycles+=3;	// sta	LZ4_SRC+1
+	cycles+=3;
 	goto parsetoken;		// jmp	parsetoken
 
 done:
+	cycles+=3;
 	pla();				// pla
+	cycles+=6;			// rts
 
-
-	int out_size=(ram[dst+1]<<8)+ram[dst];
+	int out_size=(ram[LZ4_DST+1]<<8)+ram[LZ4_DST];
 	out_size-=ORGOFFSET;
 
-	printf("dest addr  : %02X%02X\n",ram[dst+1],ram[dst]);
-
-	int i,j,addr,temp;
-	addr=ORGOFFSET;
-
-	printf("\n");
-	for(i=0;i<256;i++) {
-		if (i%16==0) printf("%04X: ",addr+i);
-		printf("%02X ",ram[addr+i]);
-		if (i%16==15) {
-			for(j=0;j<16;j++) {
-				temp=ram[((addr+i)&0xfff0)+j];
-				if ((temp<' ') || (temp>127)) printf(".");
-				else printf("%c",temp);
-			}
-			printf("\n");
-		}
-	}
-
-	printf("\n");
-					// rts
+	printf("dest addr  : %02X%02X\n",ram[LZ4_DST+1],ram[LZ4_DST]);
 
 	fff=fopen("out.out","w");
 	if (fff==NULL) {
@@ -278,6 +251,9 @@ done:
 
 	fclose(fff);
 
+	printf("Cycles: %d\t %lfs\n",cycles,(double)cycles/1023000.0);
+	printf("\t50Hz\t %lfs\n",(double)1/50.0);
+
 	return 0;
 }
 
@@ -362,6 +338,40 @@ int main(int argc, char **argv) {
 	a=23;
 	print_header_info();
 
+	y=0;
+	a=high(LZ4_BUFFER+3);
+	ram[LZ4_SRC+1]=a;
+	a=low(LZ4_BUFFER+3);
+	ram[LZ4_SRC]=a;
+
+	a=ram[y_indirect(LZ4_SRC,y)];
+	c=0;
+	adc(ram[LZ4_SRC]);
+	ram[LZ4_SRC]=a;
+	a=ram[LZ4_SRC+1];
+	adc(0);
+	ram[LZ4_SRC+1]=a;
+
+// next_subsong:
+
+	y=0;
+
+	a=ram[y_indirect(LZ4_SRC,y)];
+	ram[LZ4_END]=a;
+	y++;
+	a=ram[y_indirect(LZ4_SRC,y)];
+	ram[LZ4_END+1]=a;
+	y++;
+
+	a=2;
+	c=0;
+	adc(ram[LZ4_SRC]);
+	ram[LZ4_SRC]=a;
+	a=(ram[LZ4_SRC+1]);
+	adc(0);
+	ram[LZ4_SRC+1]=a;
+
+	lz4_decode();
 
 	return 0;
 }