From 29d9d58890af36eadfb6e42ae8bbdb4db6b87d2a Mon Sep 17 00:00:00 2001
From: TC1995 <mariogplayer90@gmail.com>
Date: Sun, 26 Jun 2022 23:19:48 +0200
Subject: [PATCH 1/3] Slight fixes to the 8514/A blitter in line draw,
 rectangle and outline modes.

---
 src/video/vid_8514a.c | 213 ++++++++++++++++++++++++++++++------------
 1 file changed, 151 insertions(+), 62 deletions(-)

diff --git a/src/video/vid_8514a.c b/src/video/vid_8514a.c
index 07faf1c49..566cd3739 100644
--- a/src/video/vid_8514a.c
+++ b/src/video/vid_8514a.c
@@ -63,7 +63,7 @@ static void ibm8514_short_stroke_start(int count, int cpu_input, uint32_t mix_da
 static void ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, ibm8514_t *dev, int len);
 
 #define READ_PIXTRANS_WORD(cx, n) \
-        if (cmd <= 1) { \
+        if (cmd <= 1 || (cmd == 5)) { \
             temp = dev->vram[((dev->accel.cy * dev->h_disp) + (cx) + (n)) & dev->vram_mask]; \
             temp |= (dev->vram[((dev->accel.cy * dev->h_disp) + (cx) + (n + 1)) & dev->vram_mask] << 8); \
         } else { \
@@ -248,8 +248,8 @@ regular_nibble:
                                 nibble |= 0x01;
                         }
 
-                        if ((and3 == 0) || (dev->accel.cmd & 0x1000) || (dev->accel.cmd & 8)) {
-                            if (dev->accel.cmd & 8) {
+                        if ((and3 == 0) || (dev->accel.cmd & 0x1000) || ((dev->accel.cmd & 8) && ibm8514_cpu_src(dev))) {
+                            if ((dev->accel.cmd & 8) && ibm8514_cpu_src(dev)) {
                                 monoxfer = val;
                             } else
                                 monoxfer = nibble;
@@ -1117,6 +1117,8 @@ ibm8514_short_stroke_start(int count, int cpu_input, uint32_t mix_dat, uint32_t
     ibm8514_accel_start(count, cpu_input, mix_dat, cpu_dat, dev, len);
 }
 
+#define SWAP(a,b) { tmpswap = a; a = b; b = tmpswap; }
+
 static void
 ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat, ibm8514_t *dev, int len)
 {
@@ -1137,6 +1139,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
 	uint32_t old_mix_dat;
 	int and3 = dev->accel.cur_x & 3;
 	uint8_t poly_src = 0;
+	int16_t tmpswap;
 
 	if (dev->accel.cmd & 0x100) {
 		dev->force_busy = 1;
@@ -1242,6 +1245,11 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
 	/*Bit 4 of the Command register is the draw yes bit, which enables writing to memory/reading from memory when enabled.
 	  When this bit is disabled, no writing to memory/reading from memory is allowed. (This bit is almost meaningless on
 	  the NOP command)*/
+    if (dev->accel.cmd == 0x41f0) {
+        pclog("CMD = %d, full = %04x, pixcntl = %d, curx = %d, cury = %d, cl = %d, cr = %d, ct = %d, cb = %d, len = %d, sx = %d, sy = %d, BX = %04X, CX = %02X\n", cmd, dev->accel.cmd, pixcntl, dev->accel.cur_x, dev->accel.cur_y, dev->accel.multifunc[2], dev->accel.multifunc[4], dev->accel.multifunc[1], dev->accel.multifunc[3], dev->accel.maj_axis_pcnt, dev->accel.sx, dev->accel.sy, BX, CH);
+        pclog("\n");
+    }
+
     switch (cmd) {
         case 0: /*NOP (Short Stroke Vectors)*/
         if (dev->accel.ssv_state == 0)
@@ -1302,6 +1310,7 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
 
         case 1: /*Draw line*/
         if (!cpu_input) {
+            dev->accel.xx_count = 0;
             dev->accel.cx = dev->accel.cur_x;
             dev->accel.cy = dev->accel.cur_y;
 
@@ -1315,6 +1324,35 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
             dev->accel.sy = dev->accel.maj_axis_pcnt;
 
             if (ibm8514_cpu_src(dev)) {
+                if (dev->accel.cmd & 2) {
+                    if (dev->accel.cmd & 8) {
+                        if (and3 == 1) {
+                            dev->accel.sy += 4;
+                            if (dev->accel.cmd & 0x20)
+                                dev->accel.cx += 4;
+                            else
+                                dev->accel.cx -= 4;
+                        } else if (and3 == 2) {
+                            dev->accel.sy += 5;
+                           if (dev->accel.cmd & 0x20)
+                                dev->accel.cx += 5;
+                            else
+                                dev->accel.cx -= 5;
+                        } else if (and3 == 3) {
+                            dev->accel.sy += 6;
+                            if (dev->accel.cmd & 0x20)
+                                dev->accel.cx += 6;
+                            else
+                                dev->accel.cx -= 6;
+                        } else {
+                            dev->accel.sy += 3;
+                            if (dev->accel.cmd & 0x20)
+                                dev->accel.cx += 3;
+                            else
+                                dev->accel.cx -= 3;
+                        }
+                    }
+                }
                 dev->data_available = 0;
                 dev->data_available2 = 0;
                 return; /*Wait for data from CPU*/
@@ -1326,6 +1364,9 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
         }
 
         if (dev->accel.cmd & 8) { /*Vector Line*/
+            if (ibm8514_cpu_dest(dev) && cpu_input && (dev->accel.cmd & 2))
+                count >>= 1;
+            dev->accel.xx_count++;
             while (count-- && (dev->accel.sy >= 0)) {
                 if ((dev->accel.cx >= dev->accel.clip_left && dev->accel.cx <= clip_r &&
                     dev->accel.cy >= dev->accel.clip_top && dev->accel.cy <= clip_b)) {
@@ -1361,11 +1402,71 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         old_dest_dat = dest_dat;
                         MIX(mix_dat & mix_mask, dest_dat, src_dat);
                         dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
-                        if (ibm8514_cpu_src(dev) || !cpu_input) {
-                            if ((dev->accel.cmd & 4) && dev->accel.sy) {
-                                WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
-                            } else if (!(dev->accel.cmd & 4)) {
-                                WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                        if ((dev->accel.cmd & 2) && ibm8514_cpu_src(dev)) {
+                            if (and3 == 1) {
+                                if (dev->accel.xx_count >= 2) {
+                                    if ((dev->accel.cmd & 4) && dev->accel.sy) {
+                                        WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                    } else if (!(dev->accel.cmd & 4)) {
+                                        WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                    }
+                                }
+                            } else if (and3 == 2) {
+                                if (dev->accel.xx_count == 2) {
+                                    if (count <= 2) {
+                                        if ((dev->accel.cmd & 4) && dev->accel.sy) {
+                                            WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                        } else if (!(dev->accel.cmd & 4)) {
+                                            WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                        }
+                                    }
+                                } else if (dev->accel.xx_count >= 3) {
+                                    if ((dev->accel.cmd & 4) && dev->accel.sy) {
+                                        WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                    } else if (!(dev->accel.cmd & 4)) {
+                                        WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                    }
+                                }
+                            } else if (and3 == 3) {
+                                if (dev->accel.xx_count == 2) {
+                                    if (count <= 1) {
+                                        if ((dev->accel.cmd & 4) && dev->accel.sy) {
+                                            WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                        } else if (!(dev->accel.cmd & 4)) {
+                                            WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                        }
+                                    }
+                                } else if (dev->accel.xx_count >= 3) {
+                                    if ((dev->accel.cmd & 4) && dev->accel.sy) {
+                                        WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                    } else if (!(dev->accel.cmd & 4)) {
+                                        WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                    }
+                                }
+                            } else {
+                                if (dev->accel.xx_count == 1) {
+                                    if (!count) {
+                                        if ((dev->accel.cmd & 4) && dev->accel.sy) {
+                                            WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                        } else if (!(dev->accel.cmd & 4)) {
+                                            WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                        }
+                                    }
+                                } else if (dev->accel.xx_count >= 2) {
+                                    if ((dev->accel.cmd & 4) && dev->accel.sy) {
+                                        WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                    } else if (!(dev->accel.cmd & 4)) {
+                                        WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                    }
+                                }
+                            }
+                        } else {
+                            if (ibm8514_cpu_src(dev) || !cpu_input) {
+                                if ((dev->accel.cmd & 4) && dev->accel.sy) {
+                                    WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                } else if (!(dev->accel.cmd & 4)) {
+                                    WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
+                                }
                             }
                         }
                     }
@@ -1576,15 +1677,14 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
             dev->accel.sy = dev->accel.multifunc[0] & 0x7ff;
 
             dev->accel.cx = dev->accel.cur_x & 0x3ff;
-            dev->accel.cy = dev->accel.cur_y & 0x3ff;
-
             if (dev->accel.cur_x & 0x400)
                 dev->accel.cx |= ~0x3ff;
+            dev->accel.cy = dev->accel.cur_y & 0x3ff;
             if (dev->accel.cur_y & 0x400)
                 dev->accel.cy |= ~0x3ff;
 
-            dev->accel.fill_state = 0;
             dev->accel.dest = dev->accel.cy * dev->h_disp;
+            dev->accel.fill_state = 0;
 
             if (cmd == 4)
                 dev->accel.cmd |= 2;
@@ -1644,6 +1744,13 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
                         dev->accel.input = 1;
                         dev->accel.newdest_in = (dev->accel.cy + 1) * dev->h_disp;
                     }
+                } else if (dev->accel.cmd & 2) {
+                    if (dev->accel.cmd & 8) {
+                        dev->accel.sx += and3;
+                        dev->accel.nibbleset = (uint8_t *)calloc(1, (dev->accel.sx >> 3) + 1);
+                        dev->accel.writemono = (uint8_t *)calloc(1, (dev->accel.sx >> 3) + 1);
+                        dev->accel.sys_cnt = (dev->accel.sx >> 3) + 1;
+                    }
                 }
 				dev->data_available = 1;
 				dev->data_available2 = 1;
@@ -1654,25 +1761,12 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
         if (dev->accel.cmd & 2) {
             if (cpu_input) {
 rect_fill_pix:
-                if (dev->accel.cmd & 8) {
+                if ((dev->accel.cmd & 8) && ibm8514_cpu_src(dev)) {
                     dev->accel.xx_count++;
                     while (count-- && (dev->accel.sy >= 0)) {
                         if ((dev->accel.cx >= dev->accel.clip_left && dev->accel.cx <= clip_r &&
                             dev->accel.cy >= dev->accel.clip_top && dev->accel.cy <= clip_b)) {
-                            if (ibm8514_cpu_dest(dev) && (pixcntl == 0)) {
-                                mix_dat = mix_mask; /* Mix data = forced to foreground register. */
-                            } else if (ibm8514_cpu_dest(dev) && (pixcntl == 3)) {
-                                /* Mix data = current video memory value. */
-                                READ(dev->accel.dest + dev->accel.cx, mix_dat);
-                                mix_dat = ((mix_dat & rd_mask) == rd_mask);
-                                mix_dat = mix_dat ? mix_mask : 0;
-                            }
-
-                            if (ibm8514_cpu_dest(dev)) {
-                                READ(dev->accel.dest + dev->accel.cx, src_dat);
-                                if (pixcntl == 3)
-                                    src_dat = ((src_dat & rd_mask) == rd_mask);
-                            } else switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
+                            switch ((mix_dat & mix_mask) ? frgd_mix : bkgd_mix) {
                                 case 0: src_dat = bkgd_color; break;
                                 case 1: src_dat = frgd_color; break;
                                 case 2: src_dat = cpu_dat & 0xff; break;
@@ -2322,7 +2416,7 @@ rect_fill:
                         }
                     }
                 } else {
-                    if (dev->accel.multifunc[0x0a] & 4) {
+                    if (dev->accel.multifunc[0x0a] & 6) {
                         while (count-- && dev->accel.sy >= 0) {
                             if ((dev->accel.cx >= dev->accel.clip_left && dev->accel.cx <= clip_r &&
                                 dev->accel.cy >= dev->accel.clip_top && dev->accel.cy <= clip_b)) {
@@ -2340,8 +2434,9 @@ rect_fill:
                                     poly_src = ((poly_src & rd_mask_polygon) == rd_mask_polygon);
                                 }
 
-                                if (poly_src)
+                                if (poly_src) {
                                     dev->accel.fill_state = !dev->accel.fill_state;
+                                }
 
                                 if (dev->accel.fill_state) {
                                     READ(dev->accel.dest + dev->accel.cx, dest_dat);
@@ -2364,20 +2459,22 @@ rect_fill:
                             mix_dat <<= 1;
                             mix_dat |= 1;
 
-                            if (dev->accel.cmd & 0x20)
+                            if (dev->accel.cmd & 0x20) {
                                 dev->accel.cx++;
-                            else
+                            } else {
                                 dev->accel.cx--;
+                            }
 
                             dev->accel.sx--;
                             if (dev->accel.sx < 0) {
-                                dev->accel.fill_state = 0;
                                 dev->accel.sx = dev->accel.maj_axis_pcnt & 0x7ff;
+                                dev->accel.fill_state = 0;
 
                                 if (dev->accel.cmd & 0x20) {
                                     dev->accel.cx -= (dev->accel.sx) + 1;
-                                } else
+                                } else {
                                     dev->accel.cx += (dev->accel.sx) + 1;
+                                }
 
                                 if (dev->accel.cmd & 0x80)
                                     dev->accel.cy++;
@@ -2464,22 +2561,18 @@ rect_fill:
             dev->accel.cx = dev->accel.cur_x;
             dev->accel.cy = dev->accel.cur_y;
 
-            if (dev->accel.cur_x & 0x400) {
-                if (dev->accel.cx >= 1024) {
-                    dev->accel.cx = 0;
-                } else
-                    dev->accel.cx |= ~0x3ff;
-            }
-
-            if (dev->accel.cur_y & 0x400) {
-                if (dev->accel.cy >= 1024)
-                    dev->accel.cy = 1;
-                else
-                    dev->accel.cy |= ~0x3ff;
-            }
-
             dev->accel.sy = dev->accel.maj_axis_pcnt;
 
+            if (dev->accel.cx < dev->accel.clip_left)
+                dev->accel.cx = dev->accel.clip_left;
+            else if (dev->accel.cx > clip_r)
+                dev->accel.cx = dev->accel.clip_left;
+
+            dev->accel.sx = dev->accel.sy;
+            dev->accel.dx = dev->accel.destx_distp;
+            dev->accel.dy = dev->accel.desty_axstp;
+            dev->accel.err = dev->accel.err_term;
+
             if (ibm8514_cpu_src(dev)) {
                 dev->data_available = 0;
                 dev->data_available2 = 0;
@@ -2527,7 +2620,6 @@ rect_fill:
                         old_dest_dat = dest_dat;
                         MIX(mix_dat & mix_mask, dest_dat, src_dat);
                         dest_dat = (dest_dat & wrt_mask) | (old_dest_dat & ~wrt_mask);
-
                         if ((dev->accel.cmd & 4) && dev->accel.sy) {
                             WRITE((dev->accel.cy * dev->h_disp) + dev->accel.cx, dest_dat);
                         } else if (!(dev->accel.cmd & 4)) {
@@ -2556,8 +2648,6 @@ rect_fill:
 
                 dev->accel.sy--;
             }
-            dev->accel.cur_x = dev->accel.cx;
-            dev->accel.cur_y = dev->accel.cy;
         } else { /*Bresenham*/
             if (pixcntl == 1) {
                 dev->accel.temp_cnt = 8;
@@ -2636,15 +2726,13 @@ rect_fill:
                 }
             } else {
                 while (count-- && (dev->accel.sy >= 0)) {
-                    if (dev->accel.cur_x >= 1024) {
-                        dev->accel.cx = 0;
-                    }
-                    if (dev->accel.cur_y >= 1024) {
-                        dev->accel.cy = 1;
-                    }
+                    if (dev->accel.cx < dev->accel.clip_left)
+                        dev->accel.cx = dev->accel.clip_left;
+                    else if (dev->accel.cx > clip_r)
+                        dev->accel.cx = dev->accel.clip_left;
 
-                    if ((dev->accel.cx >= dev->accel.clip_left && dev->accel.cx <= clip_r &&
-                        dev->accel.cy >= dev->accel.clip_top && dev->accel.cy <= clip_b)) {
+                    if (((dev->accel.cx) >= dev->accel.clip_left && (dev->accel.cx) <= clip_r &&
+                        (dev->accel.cy) >= dev->accel.clip_top && (dev->accel.cy) <= clip_b)) {
                         if (ibm8514_cpu_dest(dev) && (pixcntl == 0)) {
                             mix_dat = mix_mask; /* Mix data = forced to foreground register. */
                         } else if (ibm8514_cpu_dest(dev) && (pixcntl == 3)) {
@@ -2684,6 +2772,7 @@ rect_fill:
                             }
                         }
                     }
+
                     mix_dat <<= 1;
                     mix_dat |= 1;
                     cpu_dat >>= 8;
@@ -2692,8 +2781,8 @@ rect_fill:
                         break;
                     }
 
-                    if (dev->accel.err_term >= dev->accel.maj_axis_pcnt) {
-                        dev->accel.err_term += dev->accel.destx_distp;
+                    if (dev->accel.err >= dev->accel.sx) {
+                        dev->accel.err += dev->accel.dx;
                         /*Step minor axis*/
                         switch (dev->accel.cmd & 0xe0) {
                             case 0x00: dev->accel.cy--; break;
@@ -2706,7 +2795,7 @@ rect_fill:
                             case 0xe0: dev->accel.cx++; break;
                         }
                     } else
-                        dev->accel.err_term += dev->accel.desty_axstp;
+                        dev->accel.err += dev->accel.dy;
 
                     /*Step major axis*/
                     switch (dev->accel.cmd & 0xe0) {
@@ -2723,9 +2812,9 @@ rect_fill:
                     dev->accel.sy--;
                 }
             }
-            dev->accel.cur_x = dev->accel.cx;
-            dev->accel.cur_y = dev->accel.cy;
         }
+        dev->accel.cur_x = dev->accel.cx;
+        dev->accel.cur_y = dev->accel.cy;
         break;
 
         case 6: /*BitBlt*/

From 061c756a18872d6c3dc8b510326c6ceefaf447e5 Mon Sep 17 00:00:00 2001
From: TC1995 <mariogplayer90@gmail.com>
Date: Sun, 26 Jun 2022 23:22:32 +0200
Subject: [PATCH 2/3] Remove excess log.

---
 src/video/vid_8514a.c | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/src/video/vid_8514a.c b/src/video/vid_8514a.c
index 566cd3739..85607f004 100644
--- a/src/video/vid_8514a.c
+++ b/src/video/vid_8514a.c
@@ -1245,11 +1245,6 @@ ibm8514_accel_start(int count, int cpu_input, uint32_t mix_dat, uint32_t cpu_dat
 	/*Bit 4 of the Command register is the draw yes bit, which enables writing to memory/reading from memory when enabled.
 	  When this bit is disabled, no writing to memory/reading from memory is allowed. (This bit is almost meaningless on
 	  the NOP command)*/
-    if (dev->accel.cmd == 0x41f0) {
-        pclog("CMD = %d, full = %04x, pixcntl = %d, curx = %d, cury = %d, cl = %d, cr = %d, ct = %d, cb = %d, len = %d, sx = %d, sy = %d, BX = %04X, CX = %02X\n", cmd, dev->accel.cmd, pixcntl, dev->accel.cur_x, dev->accel.cur_y, dev->accel.multifunc[2], dev->accel.multifunc[4], dev->accel.multifunc[1], dev->accel.multifunc[3], dev->accel.maj_axis_pcnt, dev->accel.sx, dev->accel.sy, BX, CH);
-        pclog("\n");
-    }
-
     switch (cmd) {
         case 0: /*NOP (Short Stroke Vectors)*/
         if (dev->accel.ssv_state == 0)

From 10f867dce1d4f0e1cc0fbe56af15479cfe159976 Mon Sep 17 00:00:00 2001
From: TC1995 <mariogplayer90@gmail.com>
Date: Mon, 27 Jun 2022 00:08:10 +0200
Subject: [PATCH 3/3] Fix compile.

---
 src/include/86box/vid_8514a.h | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/src/include/86box/vid_8514a.h b/src/include/86box/vid_8514a.h
index af46b166d..0bb1e85cd 100644
--- a/src/include/86box/vid_8514a.h
+++ b/src/include/86box/vid_8514a.h
@@ -63,8 +63,9 @@ typedef struct ibm8514_t
         int sys_cnt, sys_cnt2;
         int temp_cnt;
         int16_t cx, cy;
-        int sx, sy;
-        int dx, dy;
+        int16_t sx, sy;
+        int16_t dx, dy;
+        int16_t err;
         uint32_t src, dest;
         uint32_t newsrc_blt, newdest_blt;
         uint32_t newdest_in, newdest_out;