asm optimization for 8bpp (no mmx)

2026-01-06 22:04:12 +01:00 · 2000-01-21 21:04:20 +00:00
parent 34cd5125e3
commit 58ca0c2824
5 changed files with 279 additions and 115 deletions
--- a/wrlib/bench.h
+++ b/wrlib/bench.h
@@ -5,7 +5,7 @@ inline static void
 cycle_bench(int start) 
 {
    static volatile unsigned int start_cycle;
-    
+
    if (start) {
 	asm volatile
 	    ("pushl %%edx	\n\t"
@@ -24,7 +24,7 @@ cycle_bench(int start)
 	     : "=m" (start_cycle) : : "eax", "edx");
    } else {
 	volatile int end;
-	
+
 	asm volatile
 	    ("pushl %%edx	\n\t"
 	     "clc            	\n\t"
--- a/wrlib/convert.c
+++ b/wrlib/convert.c
@@ -41,6 +41,19 @@ extern Pixmap R_CreateXImageMappedPixmap(RContext *context, RXImage *ximage);
 #endif


+#ifdef ASM_X86
+extern void x86_PseudoColor_to_8(unsigned char *image,
+				 unsigned char *ximage, 
+				 char *err, char *nerr,
+				 short *ctable,
+				 int dr, int dg, int db,
+				 unsigned long *pixels,
+				 int cpc, 
+				 int width, int height,
+				 int bytesPerPixel,
+				 int line_offset);
+#endif /* ASM_X86 */
+
 #ifdef ASM_X86_MMX

 extern int x86_check_mmx();
@@ -59,7 +72,7 @@ extern void x86_mmx_TrueColor_32_to_16(unsigned char *image,



-#endif /* X86_ASM */
+#endif /* ASM_X86_MMX */



@@ -310,7 +323,8 @@ image2TrueColor(RContext *ctx, RImage *image)
 	    }
 	    memset(err, 0, 8*(image->width+3));
 	    memset(nerr, 0, 8*(image->width+3));
-	    
+
+	    puts("USING MMX");
 	    x86_mmx_TrueColor_32_to_16(image->data, 
 				       (unsigned short*)ximg->image->data, 
 				       err+8, nerr+8,
@@ -414,6 +428,7 @@ convertPseudoColor_to_8(RXImage *ximg, RImage *image,
 	    r = (rer*3)/8;
 	    g = (ger*3)/8;
 	    b = (ber*3)/8;
+
 	    /* x+1, y */
 	    err[x+3*1]+=r;
 	    err[x+1+3*1]+=g;
@@ -499,8 +514,8 @@ image2PseudoColor(RContext *ctx, RImage *image)
 #ifdef DEBUG
        printf("pseudo color dithering with %d colors per channel\n", cpc);
 #endif
-	err = malloc(3*(image->width+2));
-	nerr = malloc(3*(image->width+2));
+	err = malloc(4*(image->width+3));
+	nerr = malloc(4*(image->width+3));
 	if (!err || !nerr) {
 	    if (nerr)
 		free(nerr);
@@ -508,12 +523,22 @@ image2PseudoColor(RContext *ctx, RImage *image)
 	    RDestroyXImage(ctx, ximg);
 	    return NULL;
 	}
-	memset(err, 0, 3*(image->width+3));
-	memset(nerr, 0, 3*(image->width+3));
-  
+	memset(err, 0, 4*(image->width+3));
+	memset(nerr, 0, 4*(image->width+3));
+
+#ifdef ASM_X86
+	x86_PseudoColor_to_8(image->data, ximg->image->data,
+			     err+4, nerr+4,
+			     rtable,
+			     dr, dg, db, ctx->pixels, cpc,
+			     image->width, image->height,
+			     channels, 
+			     ximg->image->bytes_per_line - image->width);
+#else
 	convertPseudoColor_to_8(ximg, image, err+4, nerr+4,
 				rtable,	gtable,	btable,
 				dr, dg, db, ctx->pixels, cpc);
+#endif

 	free(err);
 	free(nerr);
--- a/wrlib/gradient.c
+++ b/wrlib/gradient.c
@@ -231,8 +231,9 @@ renderDGradient(unsigned width, unsigned height, int r0, int g0, int b0,
 		int rf, int gf, int bf)
 {
    RImage *image, *tmp;
-    unsigned long a;
    int i, j, offset;
+    float a;
+    char *ptr;

    if (width == 1)
        return renderVGradient(width, height, r0, g0, b0, rf, gf, bf);
@@ -249,15 +250,18 @@ renderDGradient(unsigned width, unsigned height, int r0, int g0, int b0,
        RDestroyImage(image);
        return NULL;
    }
-
-    a = (((width - 1)<<16) / (height - 1));
+    
+    ptr = tmp->data;

    width *= 3;
+    a = ((float)(width - 1))/((float)(height - 1));
+
    /* copy the first line to the other lines with corresponding offset */
-    for (i=0, j=0, offset = 0; i<height; i++, j+= width) {
-        offset += a;
-	memcpy(&(image->data[j]), &(tmp->data[(offset>>16)*3]), width);
+    for (i=0, j=0, offset=0; i<height; i++, j += width) {
+        offset = (int)(a*i+0.5)*3;
+	memcpy(&(image->data[j]), &ptr[offset], width);
    }
+
    RDestroyImage(tmp);
    return image;
 }
@@ -428,7 +432,8 @@ renderMDGradient(unsigned width, unsigned height, RColor **colors, int count)
 {
    RImage *image, *tmp;
    float a;
-    int i, offset;
+    int i, offset, j;
+    unsigned char *ptr;

    assert(count > 2);

@@ -459,13 +464,15 @@ renderMDGradient(unsigned width, unsigned height, RColor **colors, int count)
        RDestroyImage(image);
        return NULL;
    }
+    ptr = tmp->data;
+

    a = ((float)(width - 1))/((float)(height - 1));

    /* copy the first line to the other lines with corresponding offset */
-    for (i=0; i<height; i++) {
+    for (i=0, j=0, offset=0; i<height; i++, j += width) {
        offset = (int)(a*i+0.5)*3;
-        memcpy(&(image->data[i*width*3]), &(tmp->data[offset]), width*3);
+	memcpy(&(image->data[j]), &ptr[offset], width);
    }
    RDestroyImage(tmp);
    return image;
--- a/wrlib/view.c
+++ b/wrlib/view.c
@@ -49,10 +49,13 @@ int main(int argc, char **argv)
 	img = RScaleImage(img, img->width*atof(argv[2]), 
 			  img->height*atof(argv[2]));
    }
+    {int i;
+    for (i=0; i<10;i++)
    if (!RConvertImage(ctx, img, &pix)) {
 	puts(RMessageForError(RErrorCode));
 	exit(1);
    }
+    }
    
    printf("%ix%i\n", img->width, img->height);
 	   
--- a/wrlib/x86_specific.c
+++ b/wrlib/x86_specific.c
@@ -61,12 +61,11 @@ x86_check_mmx()
         "jmp .noPop		\n"

 	 ".NotMMX:		\n"
-	 ".Bye:			\n"
 	 ".NotPentium:		\n"
         "popal			\n"
         ".noPop:		\n"

-	 : "=rm" (result));
+	 : "=m" (result));

    return result;
 }
@@ -75,7 +74,7 @@ x86_check_mmx()
 /*
 * TODO:
 * 		32/8	24/8	32/16	24/16	32/24	24/24
- * PPlain	YES	YES	
+ * PPlain	YES	YES
 * MMX				DONE
 *
 * 
@@ -177,7 +176,7 @@ x86_mmx_TrueColor_32_to_16(unsigned char *image, // 8
 	 "movl 16(%ebp), %eax		\n" //
 	 "movl %eax, -32(%ebp)		\n" // [-32] = eerr

-
+	 ".align 16			\n"
 ".LoopXa:				\n"

 	 // calculate errors and pixel components
@@ -192,19 +191,19 @@ x86_mmx_TrueColor_32_to_16(unsigned char *image, // 8
 	 "movzwl -24(%ebp), %ecx	\n" // ecx = pixel.red
 	 "movl 24(%ebp), %edi		\n" // edi = rtable
 	 "leal (%edi, %ecx, 2), %eax	\n" // eax = &rtable[pixel.red]
-	 "movl (%eax), %edx		\n" // edx = rtable[pixel.red]
+	 "movw (%eax), %dx		\n" // dx = rtable[pixel.red]
 	 "movw %dx, -16(%ebp)		\n" // save rr

 	 "movzwl -22(%ebp), %ecx	\n" // ecx = pixel.green
 	 "movl 28(%ebp), %edi		\n" // edi = gtable
 	 "leal (%edi, %ecx, 2), %eax	\n" // eax = &gtable[pixel.green]
-	 "movl (%eax), %edx		\n" // ebx = gtable[pixel.green]
+	 "movw (%eax), %dx		\n" // dx = gtable[pixel.green]
 	 "movw %dx, -14(%ebp)		\n" // save gg

 	 "movzwl -20(%ebp), %ecx	\n" // ecx = pixel.blue
 	 "movl 32(%ebp), %edi		\n" // ebx = btable
 	 "leal (%edi, %ecx, 2), %eax	\n" // eax = &btable[pixel.blue]
-	 "movl (%eax), %edx		\n" // ecx = btable[pixel.blue]
+	 "movw (%eax), %dx		\n" // dx = btable[pixel.blue]
 	 "movw %dx, -12(%ebp)		\n" // save bb

 	 "movw $0, -10(%ebp)		\n" // save dummy aa
@@ -303,133 +302,263 @@ x86_mmx_TrueColor_32_to_16(unsigned char *image, // 8

 #endif /* ASM_X86_MMX */

-#if 0
-
-    /* convert and dither the image to XImage */
-    for (y=0; y<image->height; y++) {
-	nerr[0] = 0;
-	nerr[1] = 0;
-	nerr[2] = 0;
-	for (x=0; x<image->width*3; x+=3, ptr+=channels) {
-
-	    /* reduce pixel */
-	    pixel = *ptr + err[x];
-	    if (pixel<0) pixel=0; else if (pixel>0xff) pixel=0xff;
-	    r = rtable[pixel];
-	    /* calc error */
-	    rer = pixel - r*dr;
-
-	    /* reduce pixel */
-	    pixel = *(ptr+1) + err[x+1];
-	    if (pixel<0) pixel=0; else if (pixel>0xff) pixel=0xff;
-	    g = gtable[pixel];
-	    /* calc error */
-	    ger = pixel - g*dg;
-
-	    /* reduce pixel */
-	    pixel = *(ptr+2) + err[x+2];
-	    if (pixel<0) pixel=0; else if (pixel>0xff) pixel=0xff;
-	    b = btable[pixel];
-	    /* calc error */
-	    ber = pixel - b*db;
-
-	    *optr++ = pixels[r*cpcpc + g*cpc + b];
-
-	    /* distribute error */
-	    r = (rer*3)/8;
-	    g = (ger*3)/8;
-	    b = (ber*3)/8;
-	    /* x+1, y */
-	    err[x+3*1]+=r;
-	    err[x+1+3*1]+=g;
-	    err[x+2+3*1]+=b;
-	    /* x, y+1 */
-	    nerr[x]+=r;
-	    nerr[x+1]+=g;
-	    nerr[x+2]+=b;
-	    /* x+1, y+1 */
-	    nerr[x+3*1]=rer-2*r;
-	    nerr[x+1+3*1]=ger-2*g;
-	    nerr[x+2+3*1]=ber-2*b;
-	}
-	/* skip to next line */
-	terr = err;
-	err = nerr;
-	nerr = terr;
-	
-	optr += ximg->image->bytes_per_line - image->width;
-    }
-}
-#endif


 void
-x86_PseudoColor_32_to_8(unsigned char *image, // 8
-			unsigned char *ximage, // 12
-			char *err, // 16
-			char *nerr, // 20
-			short *rtable, // 24
-			short *gtable, // 28
-			short *btable, // 32
-			int dr, // 36
-			int dg, // 40
-			int db, // 44
-			unsigned long *pixels, // 48
-			int cpc, // 52
-			int width, // 56
-			int height, // 60
-			int line_offset) // 64
+x86_PseudoColor_to_8(unsigned char *image, // 8
+		     unsigned char *ximage, // 12
+		     char *err, // 16
+		     char *nerr, // 20
+		     short *ctable, // 24
+		     int dr, // 28
+		     int dg, // 32
+		     int db, // 36
+		     unsigned long *pixels, // 40
+		     int cpc, // 44
+		     int width, // 48
+		     int height, // 52
+		     int bytesPerPixel, // 56
+		     int line_offset) // 60
 {
+    /*
+     * int x; -4
+     * int cpcpc; -8
+     * 
+     * int rr; -12
+     * int gg; -16
+     * int bb; -20
+     * 
+     * char ndr; -21
+     * char ndg; -22
+     * char ndb; -23
+     * 
+     * char *err; -32
+     * char *nerr; -36
+     * 
+     */
    asm volatile
 	(
-	 "andl $-8, %ebp		\n"
 	 "subl $128, %esp		\n" // alloc some stack space
 	 "pushal       			\n"

+	 "movl 44(%ebp), %eax		\n"
+	 "mulb 44(%ebp)			\n"
+	 "movl %eax, -8(%ebp)		\n" // cpcpc = cpc*cpc
+	 
+	 // eax will always be <= 0xffff
+
 	 // process 1 pixel / cycle, each component treated as 16bit
 	 "movl 8(%ebp), %esi		\n" // esi = image->data

 ".LoopYb:				\n"
-	 "movl 56(%ebp), %eax		\n"
-	 "movl %eax, -4(%ebp)		\n" // x = width
+	 "movl 48(%ebp), %ecx		\n"
+	 "movl %ecx, -4(%ebp)		\n" // x = width

-	 "movl 60(%ebp), %eax		\n"
-	 "decl %eax			\n" // y--
-	 "movl %eax, 64(%ebp)		\n"
+	 "movl 52(%ebp), %ecx		\n"
+	 "decl %ecx			\n" // y--
+	 "movl %ecx, 52(%ebp)		\n"
 	 "js .Endb			\n" // if y < 0, goto end
-	 "andl $1, %eax			\n"
+	 "andl $1, %ecx			\n"
 	 "jz .LoopY_1b			\n" // if (y&1) goto LoopY_1

 ".LoopY_0b:				\n"

 	 "movl 16(%ebp), %ebx		\n" // ebx = err
-	 "movl %ebx, -36(%ebp)		\n" // [-36] = err
-	 "movl 20(%ebp), %eax		\n" //
-	 "movl %eax, -32(%ebp)		\n" // [-32] = nerr
+//	 "movl %ebx, -36(%ebp)		\n" // [-36] = err
+	 "movl 20(%ebp), %ecx		\n" //
+	 "movl %ecx, -32(%ebp)		\n" // [-32] = nerr

-	 "movl $0, -32(%ebp)		\n" // init error of nerr[0] to 0
+	 "movl $0, (%ecx)		\n" // init error of nerr[0] to 0

 	 "jmp .LoopXb			\n"

 ".LoopY_1b:				\n"

 	 "movl 20(%ebp), %ebx		\n" // ebx = nerr
-	 "movl %ebx, -36(%ebp)		\n" // [-36] = nerr
-	 "movl 16(%ebp), %eax		\n" //
-	 "movl %eax, -32(%ebp)		\n" // [-32] = err
+//	 "movl %ebx, -36(%ebp)		\n" // [-36] = nerr
+	 "movl 16(%ebp), %ecx		\n" //
+	 "movl %ecx, -32(%ebp)		\n" // [-32] = err

-	 "movl $0, -32(%ebp)		\n" // init error of nerr[0] to 0
+	 "movl $0, (%ecx)		\n" // init error of nerr[0] to 0

+	 
+	 ".align 16			\n"
 ".LoopXb:				\n"
 	 
-	 "movl (%esi), %edx		\n" // fetch a pixel

-//	 "movl  				\n"
+	 "movl 24(%ebp), %edi		\n" // edi = ctable
+	 "xorl %edx, %edx		\n" // zero the upper word on edx
+
+	 // RED
+
+	 // depends on ebx==err, esi==image->data, edi
+	 "movzbw (%esi), %dx		\n" // dx = image->data[0]
+	 "movsbw (%ebx), %ax		\n" // ax = error[0]
+	 "addw %ax, %dx			\n" // pixel.red = data[0] + error[0]
+
+	 "testb %dh, %dh		\n" // test if pixel.red < 0 or > 255
+	 "jz .OKRb			\n" // 0 <= pixel.red <= 255
+	 "js .NEGRb			\n" // pixel.red < 0
+	 "movw $0xff, %dx		\n" // pixel.red > 255
+	 "jmp .OKRb			\n"
+".NEGRb:				\n"
+	 "xorw %dx, %dx			\n"
+".OKRb:					\n"
+	 "leal (%edi, %edx, 2), %ecx	\n" // ecx = &ctable[pixel.red]
+	 "movl (%ecx), %eax		\n" // ax = ctable[pixel.red]
+	 "movw %ax, -12(%ebp)		\n" // save rr
+
+	 "mulb 28(%ebp)			\n" // ax = rr*dr
+	 "subw %ax, %dx			\n" // rer = dx = dx - rr*dr
+
+	 "movswl %dx, %eax		\n" // save rer
+
+	 // distribute error
+	 "leal (, %eax, 8), %ecx	\n"
+	 "subw %dx, %cx			\n" // cx = rer * 7
+	 "sarw $4, %cx			\n" // cx = rer * 7 / 16
+	 "addb %cl, 4(%ebx)		\n" // err[x+1] += rer * 7 / 16
 	 
+	 "movl -32(%ebp), %ecx		\n" // ecx = nerr
+	 
+	 "leaw (%eax, %eax, 4), %dx	\n" // dx = rer * 5
+	 "sarw $4, %dx			\n" // dx = rer * 5 / 16
+	 "addb %dl, (%ecx)		\n" // nerr[x] += rer * 5 / 16

+	 "leaw (%eax, %eax, 2), %dx	\n" // dx = rer * 3
+	 "sarw $4, %dx			\n" // dx = rer * 3 / 16
+	 "addb %dl, -4(%ecx)		\n" // nerr[x-1] += rer * 3 / 16
+
+	 "sarw $4, %ax			\n" // ax = rer / 16
+	 "movb %al, 4(%ecx)		\n" // nerr[x+1] = rer / 16
+
+	 
+	 // GREEN
+
+	 // depends on ebx, esi, edi
+	 "movzbw 1(%esi), %dx		\n" // dx = image->data[1]
+	 "movsbw 1(%ebx), %ax		\n" // ax = error[1]
+	 "addw %ax, %dx			\n" // pixel.grn = data[1] + error[1]
+
+	 "testb %dh, %dh		\n" // test if pixel.grn < 0 or > 255
+	 "jz .OKGb			\n" // 0 <= pixel.grn <= 255
+	 "js .NEGGb			\n" // pixel.grn < 0
+	 "movw $0xff, %dx		\n" // pixel.grn > 255
+	 "jmp .OKGb			\n"
+".NEGGb:				\n"
+	 "xorw %dx, %dx			\n"
+".OKGb:					\n"
+	 "leal (%edi, %edx, 2), %ecx	\n" // ecx = &ctable[pixel.grn]
+	 "movw (%ecx), %ax		\n" // ax = ctable[pixel.grn]
+	 "movw %ax, -16(%ebp)		\n" // save gg
+
+	 "mulb 28(%ebp)			\n" // ax = gg*dg
+	 "subw %ax, %dx			\n" // ger = dx = dx - gg*dg
+
+	 "movswl %dx, %eax		\n" // save ger
+
+	 // distribute error
+
+	 "leal (, %eax, 8), %ecx	\n"
+	 "subw %dx, %cx			\n" // cx = ger * 7
+	 "sarw $4, %cx			\n" // cx = ger * 7 / 16
+	 "addb %cl, 5(%ebx)		\n" // err[x+1] += ger * 7 / 16
+	 
+	 "movl -32(%ebp), %ecx		\n" // ecx = nerr
+	 
+	 "leaw (%eax, %eax, 4), %dx	\n" // dx = ger * 5
+	 "sarw $4, %dx			\n" // dx = ger * 5 / 16
+	 "addb %dl, 1(%ecx)		\n" // nerr[x] += ger * 5 / 16
+
+	 "leaw (%eax, %eax, 2), %dx	\n" // dx = ger * 3
+	 "sarw $4, %dx			\n" // dx = ger * 3 / 16
+	 "addb %dl, -3(%ecx)		\n" // nerr[x-1] += ger * 3 / 16
+
+	 "sarw $4, %ax			\n" // ax = ger / 16
+	 "movb %al, 5(%ecx)		\n" // nerr[x+1] = ger / 16
+
+	 
+	 // BLUE
+
+	 // depends on ebx, esi
+	 "movzbw 2(%esi), %dx		\n" // dx = image->data[2]
+	 "movsbw 2(%ebx), %ax		\n" // ax = error[2]
+	 "addw %ax, %dx			\n" // pixel.grn = data[2] + error[2]
+	 
+	 "testb %dh, %dh		\n" // test if pixel.blu < 0 or > 255
+	 "jz .OKBb			\n" // 0 <= pixel.blu <= 255
+	 "js .NEGBb			\n" // pixel.blu < 0
+	 "movw $0xff, %dx		\n" // pixel.blu > 255
+	 "jmp .OKBb			\n"
+".NEGBb:				\n"
+	 "xorw %dx, %dx			\n"
+".OKBb:					\n"
+	 "leal (%edi, %edx, 2), %ecx	\n" // ecx = &ctable[pixel.blu]
+	 "movw (%ecx), %ax		\n" // ax = ctable[pixel.blu]
+	 "movw %ax, -20(%ebp)		\n" // save bb
+
+	 "mulb 28(%ebp)			\n" // ax = bb*db
+	 "subw %ax, %dx			\n" // ber = dx = dx - bb*db
+	 "movswl %dx, %eax		\n" // save ber
+	 
+	 // distribute error
+	 "leal (, %eax, 8), %ecx	\n"
+	 "subw %dx, %cx			\n" // cx = ber * 7
+	 "sarw $4, %cx			\n" // cx = ber * 7 / 16
+	 "addb %cl, 6(%ebx)		\n" // err[x+1] += ber * 7 / 16
+	 
+	 "movl -32(%ebp), %ecx		\n" // ecx = nerr
+	 
+	 "leaw (%eax, %eax, 4), %dx	\n" // dx = ber * 5
+	 "sarw $4, %dx			\n" // dx = ber * 5 / 16
+	 "addb %dl, 2(%ecx)		\n" // nerr[x] += ber * 5 / 16
+
+	 "leaw (%eax, %eax, 2), %dx	\n" // dx = ber * 3
+	 "sarw $4, %dx			\n" // dx = ber * 3 / 16
+	 "addb %dl, -4(%ecx)		\n" // nerr[x-1] += ber * 3 / 16
+
+	 "sarw $4, %ax			\n" // ax = ber / 16
+	 "movb %al, 6(%ecx)		\n" // nerr[x+1] = ber / 16
+
+	 "andl $0xffff, %eax		\n"
+	 // depends on eax & 0xffff0000 == 0
+	 // calculate the index of the value of the pixel
+	 "movw -12(%ebp), %ax		\n" // ax = rr
+	 "mulb -8(%ebp)			\n" // ax = cpcpc*rr
+	 "movw %ax, %cx			\n"
+	 "movw -16(%ebp), %ax		\n" // ax = gg
+	 "mulb 44(%ebp)			\n" // ax = cpc*gg
+	 "addw %cx, %ax			\n" // ax = cpc*gg + cpcpc*rr
+	 "addw -20(%ebp), %ax		\n" // ax = cpcpc*rr + cpc*gg + bb
+
+	 "movl 40(%ebp), %ecx		\n"
+	 "leal (%ecx, %eax, 4), %edx	\n"
+	 "movb (%edx), %cl		\n" // cl = pixels[ax]
+
+	 // store the pixel
+	 "movl 12(%ebp), %eax		\n"
+	 "movb %cl, (%eax)		\n" // *ximage = cl
+	 "incl 12(%ebp)			\n" // ximage++
+
+	 // prepare for next iteration on X
+	 
+	 "addl $4, -32(%ebp)		\n" // nerr += 4
+	 "addl $4, %ebx			\n" // err += 4
+
+	 "addl 56(%ebp), %esi		\n" // image->data += bpp
+
+	 "decl -4(%ebp)			\n" // x--
+	 "jnz .LoopXb			\n" // if x>0, goto .LoopX
+
+	 
+	 "movl 60(%ebp), %eax		\n"
+	 "addl %eax, 12(%ebp)		\n" // add extra offset to ximage
+	 
+	 "jmp .LoopYb			\n"

 ".Endb:					\n"
 	 
+	 "emms				\n"
 	 "popal				\n"
 	 );