Use raster(logic) operation, or ROP, function on LUNA frame buffer.

It makes 4bpp wscons putchar ~20% faster.

ok @miod

2 files changed, 89 insertions, 40 deletions

diff --git a/sys/arch/luna88k/dev/omrasops.c b/sys/arch/luna88k/dev/omrasops.c
index 889f68fdc9a..2e432475dcf 100644
--- a/sys/arch/luna88k/dev/omrasops.c
+++ b/sys/arch/luna88k/dev/omrasops.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: omrasops.c,v 1.11 2014/01/02 15:30:34 aoyama Exp $ */
+/* $OpenBSD: omrasops.c,v 1.12 2014/09/29 13:57:35 aoyama Exp $ */
 /* $NetBSD: omrasops.c,v 1.1 2000/01/05 08:48:56 nisimura Exp $ */
 
 /*-
@@ -175,8 +175,15 @@ om4_putchar(void *cookie, int row, int startcol, u_int uc, long attr)
 	width = ri->ri_font->fontwidth + align;
 	lmask = ALL1BITS >> align;
 	rmask = ALL1BITS << (-width & ALIGNMASK);
+
+	/* select all planes for later ROP function target */ 
+	*(volatile u_int32_t *)OMFB_PLANEMASK = 0xff;
+
 	if (width <= BLITWIDTH) {
 		lmask &= rmask;
+		/* set lmask as ROP mask value, with THROUGH mode */
+		((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_THROUGH] = lmask;
+
 		while (height > 0) {
 			glyph = 0;
 			for (i = ri->ri_font->stride; i != 0; i--)
@@ -187,20 +194,22 @@ om4_putchar(void *cookie, int row, int startcol, u_int uc, long attr)
 
 			fgpat = (fg & 0x01) ? glyph : 0;
 			bgpat = (bg & 0x01) ? glyphbg : 0;
-			*P0(p) = (*P0(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+			*P0(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x02) ? glyph : 0;
 			bgpat = (bg & 0x02) ? glyphbg : 0;
-			*P1(p) = (*P1(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+			*P1(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x04) ? glyph : 0;
 			bgpat = (bg & 0x04) ? glyphbg : 0;
-			*P2(p) = (*P2(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+			*P2(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x08) ? glyph : 0;
 			bgpat = (bg & 0x08) ? glyphbg : 0;
-			*P3(p) = (*P3(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+			*P3(p) = (fgpat | bgpat);
 
 			p += scanspan;
 			height--;
 		}
+		/* reset mask value */
+		((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_THROUGH] = ALL1BITS;
 	} else {
 		u_int8_t *q = p;
 		u_int32_t lhalf, rhalf;
@@ -213,42 +222,52 @@ om4_putchar(void *cookie, int row, int startcol, u_int uc, long attr)
 			glyph <<= (4 - ri->ri_font->stride) * NBBY;
 			lhalf = (glyph >> align);
 			lhalfbg = lhalf ^ ALL1BITS;
+			/* set lmask as ROP mask value, with THROUGH mode */
+			((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_THROUGH]
+				= lmask;
 
 			fgpat = (fg & 0x01) ? lhalf : 0;
 			bgpat = (bg & 0x01) ? lhalfbg : 0;
-			*P0(p) = (*P0(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+			*P0(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x02) ? lhalf : 0;
 			bgpat = (bg & 0x02) ? lhalfbg : 0;
-			*P1(p) = (*P1(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+			*P1(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x04) ? lhalf : 0;
 			bgpat = (bg & 0x04) ? lhalfbg : 0;
-			*P2(p) = (*P2(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+			*P2(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x08) ? lhalf : 0;
 			bgpat = (bg & 0x08) ? lhalfbg : 0;
-			*P3(p) = (*P3(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+			*P3(p) = (fgpat | bgpat);
 
 			p += BYTESDONE;
 
 			rhalf = (glyph << (BLITWIDTH - align));
 			rhalfbg = rhalf ^ ALL1BITS;
+			/* set rmask as ROP mask value, with THROUGH mode */
+			((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_THROUGH]
+				= rmask;
 
 			fgpat = (fg & 0x01) ? rhalf : 0;
 			bgpat = (bg & 0x01) ? rhalfbg : 0;
-			*P0(p) = ((fgpat | bgpat) & rmask) | (*P0(p) & ~rmask);
+			*P0(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x02) ? rhalf : 0;
 			bgpat = (bg & 0x02) ? rhalfbg : 0;
-			*P1(p) = ((fgpat | bgpat) & rmask) | (*P1(p) & ~rmask);
+			*P1(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x04) ? rhalf : 0;
 			bgpat = (bg & 0x04) ? rhalfbg : 0;
-			*P2(p) = ((fgpat | bgpat) & rmask) | (*P2(p) & ~rmask);
+			*P2(p) = (fgpat | bgpat);
 			fgpat = (fg & 0x08) ? rhalf : 0;
 			bgpat = (bg & 0x08) ? rhalfbg : 0;
-			*P3(p) = ((fgpat | bgpat) & rmask) | (*P3(p) & ~rmask);
+			*P3(p) = (fgpat | bgpat);
 
 			p = (q += scanspan);
 			height--;
 		}
+		/* reset mask value */
+		((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_THROUGH] = ALL1BITS;
 	}
+	/* select plane #0 only; XXX need this ? */
+	*(volatile u_int32_t *)OMFB_PLANEMASK = 0x01;
 
 	return 0;
 }
@@ -408,7 +427,7 @@ om4_cursor(void *cookie, int on, int row, int col)
 	struct rasops_info *ri = cookie;
 	u_int8_t *p;
 	int scanspan, startx, height, width, align, y;
-	u_int32_t lmask, rmask, image;
+	u_int32_t lmask, rmask;
 
 	if (!on) {
 		/* make sure it's on */
@@ -433,48 +452,45 @@ om4_cursor(void *cookie, int on, int row, int col)
 	width = ri->ri_font->fontwidth + align;
 	lmask = ALL1BITS >> align;
 	rmask = ALL1BITS << (-width & ALIGNMASK);
+
+	/* select all planes for later ROP function target */ 
+	*(volatile u_int32_t *)OMFB_PLANEMASK = 0xff;
+
 	if (width <= BLITWIDTH) {
 		lmask &= rmask;
+		/* set lmask as ROP mask value, with INV2 mode */
+		((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_INV2] = lmask;
+
 		while (height > 0) {
-			image = *P0(p);
-			*P0(p) = (image & ~lmask) | ((image ^ ALL1BITS) & lmask);
-			image = *P1(p);
-			*P1(p) = (image & ~lmask) | ((image ^ ALL1BITS) & lmask);
-			image = *P2(p);
-			*P2(p) = (image & ~lmask) | ((image ^ ALL1BITS) & lmask);
-			image = *P3(p);
-			*P3(p) = (image & ~lmask) | ((image ^ ALL1BITS) & lmask);
+			*W(p) = ALL1BITS;
 			p += scanspan;
 			height--;
 		}
+		/* reset mask value */
+		((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_THROUGH] = ALL1BITS;
 	} else {
 		u_int8_t *q = p;
 
 		while (height > 0) {
-			image = *P0(p);
-			*P0(p) = (image & ~lmask) | ((image ^ ALL1BITS) & lmask);
-			image = *P1(p);
-			*P1(p) = (image & ~lmask) | ((image ^ ALL1BITS) & lmask);
-			image = *P2(p);
-			*P2(p) = (image & ~lmask) | ((image ^ ALL1BITS) & lmask);
-			image = *P3(p);
-			*P3(p) = (image & ~lmask) | ((image ^ ALL1BITS) & lmask);
+			/* set lmask as ROP mask value, with INV2 mode */
+			((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_INV2] = lmask;
+			*W(p) = ALL1BITS;
 
 			p += BYTESDONE;
 
-			image = *P0(p);
-			*P0(p) = ((image ^ ALL1BITS) & rmask) | (image & ~rmask);
-			image = *P1(p);
-			*P1(p) = ((image ^ ALL1BITS) & rmask) | (image & ~rmask);
-			image = *P2(p);
-			*P2(p) = ((image ^ ALL1BITS) & rmask) | (image & ~rmask);
-			image = *P3(p);
-			*P3(p) = ((image ^ ALL1BITS) & rmask) | (image & ~rmask);
+			/* set rmask as ROP mask value, with INV2 mode */
+			((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_INV2] = rmask;
+			*W(p) = ALL1BITS;
 
 			p = (q += scanspan);
 			height--;
 		}
+		/* reset mask value */
+		((volatile u_int32_t *)OMFB_ROPFUNC)[ROP_THROUGH] = ALL1BITS;
 	}
+	/* select plane #0 only; XXX need this ? */
+	*(volatile u_int32_t *)OMFB_PLANEMASK = 0x01;
+
 	ri->ri_flg ^= RI_CURSOR;
 
 	return 0;
diff --git a/sys/arch/luna88k/dev/omrasops.h b/sys/arch/luna88k/dev/omrasops.h
index dd5d9513538..4d6bd52dd77 100644
--- a/sys/arch/luna88k/dev/omrasops.h
+++ b/sys/arch/luna88k/dev/omrasops.h
@@ -18,14 +18,17 @@
  * Base addresses of LUNA's frame buffer
  * XXX: We consider only 1bpp and 4bpp for now
  */
+
+#define OMFB_PLANEMASK	0xB1040000	/* BMSEL register */
 #define OMFB_FB_WADDR   0xB1080008      /* common plane */
 #define OMFB_FB_RADDR   0xB10C0008      /* plane #0 */
+#define OMFB_ROPFUNC	0xB12C0000	/* common ROP function */
 
 /*
  * Helper macros
  */
-#define W(addr) ((u_int32_t *)(addr))
-#define R(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x40000))
+#define W(addr)  ((u_int32_t *)(addr))
+#define R(addr)  ((u_int32_t *)((u_int8_t *)(addr) +  0x40000))
 #define P0(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0x40000))
 #define P1(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0x80000))
 #define P2(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0xC0000))
@@ -36,3 +39,33 @@
  */
 #define RR_CLEAR        0x0
 #define RR_COPY         0x3
+
+/*
+ * ROP function
+ *
+ * LUNA's frame buffer uses Hitach HM53462 video RAM, which has raster
+ * (logic) operation, or ROP, function.  To use ROP function on LUNA, write
+ * a 32bit `mask' value to the specified address corresponding to each ROP
+ * logic.
+ *
+ * D: the data writing to the video RAM
+ * M: the data already stored on the video RAM
+ */
+
+/* operation		index	the video RAM contents will be */
+#define ROP_ZERO	 0	/* all 0	*/
+#define ROP_AND1	 1	/* D & M	*/ 
+#define ROP_AND2	 2	/* ~D & M	*/
+/* Not used on LUNA	 3			*/
+#define ROP_AND3	 4	/* D & ~M	*/
+#define ROP_THROUGH	 5	/* D		*/
+#define ROP_EOR		 6	/* (~D & M) | (D & ~M)	*/
+#define ROP_OR1		 7	/* D | M	*/
+#define ROP_NOR		 8	/* ~D | ~M	*/
+#define ROP_ENOR	 9	/* (D & M) | (~D & ~M)	*/
+#define ROP_INV1	10	/* ~D		*/
+#define ROP_OR2		11	/* ~D | M	*/
+#define ROP_INV2	12	/* ~M		*/
+#define ROP_OR3		13	/* D | ~M	*/
+#define ROP_NAND	14	/* ~D | ~M	*/
+#define ROP_ONE		15	/* all 1	*/