From db1ee7a2df5570b900a57bb3438eabf9c80eddff Mon Sep 17 00:00:00 2001
From: kub <derkub@gmail.com>
Date: Sun, 28 Jan 2024 19:23:25 +0000
Subject: [PATCH] core, handle background color DMA (arm asm)

---
 pico/draw.c     | 66 +++++++++++++++++++++++++------------------------
 pico/draw_arm.S | 10 ++++++--
 pico/pico.h     |  1 +
 3 files changed, 43 insertions(+), 34 deletions(-)

diff --git a/pico/draw.c b/pico/draw.c
index 40903e74..3584b6a2 100644
--- a/pico/draw.c
+++ b/pico/draw.c
@@ -1607,46 +1607,46 @@ static int BgcDMAlen, BgcDMAoffs;
 static
 #endif
 // handle DMA to background color
-int BgcDMA(u16 *pd, int len, struct PicoEState *est)
+void BgcDMA(struct PicoEState *est)
 {
+  u16 *pd=est->DrawLineDest;
+  int len = (est->Pico->video.reg[12]&1) ? 320 : 256;
+  // TODO for now handles the line as all background.
   int xl = (len == 320 ? 38 : 33); // DMA slots during HSYNC
   int upscale = (est->rendstatus & PDRAW_SOFTSCALE) && len < 320;
+  u16 *q = upscale ? DefOutBuff : pd;
+  int i, l = len;
+  u16 t;
 
-  if (BgcDMAlen > 0) {
-    // BG color DMA under way. TODO for now handles the line as all background.
-    int i, l = len;
-    u16 *q = upscale ? DefOutBuff : pd;
-    u16 t;
-
-    if ((est->rendstatus & PDRAW_BORDER_32) && !upscale)
-      q += (320-len) / 2;
+  if ((est->rendstatus & PDRAW_BORDER_32) && !upscale)
+    q += (320-len) / 2;
 
-    BgcDMAlen -= (l>>1)+xl;
-    if (BgcDMAlen < 0)
-      // partial line
-      l += 2*BgcDMAlen;
+  BgcDMAlen -= ((l-BgcDMAoffs)>>1)+xl;
+  if (BgcDMAlen <= 0) {
+    // partial line
+    l += 2*BgcDMAlen;
+    est->rendstatus &= ~PDRAW_BGC_DMA;
+  }
 
-    for (i = 0; i < l; i += 2) {
-      // TODO use ps to overwrite only real bg pixels
-      t = BgcDMAbase[BgcDMAsrc++ & BgcDMAmask];
-      q[i] = q[i+1] = PXCONV(t);
-    }
-    BgcDMAsrc += xl; // HSYNC DMA
+  for (i = BgcDMAoffs; i < l; i += 2) {
+    // TODO use ps to overwrite only real bg pixels
+    t = BgcDMAbase[BgcDMAsrc++ & BgcDMAmask];
+    q[i] = q[i+1] = PXCONV(t);
+  }
+  BgcDMAsrc += xl; // HSYNC DMA
+  BgcDMAoffs = 0;
 
-    t = est->HighPal[Pico.video.reg[7] & 0x3f];
-    while (i < len) q[i++] = t; // fill partial line with BG
+  t = est->HighPal[Pico.video.reg[7] & 0x3f];
+  while (i < len) q[i++] = t; // fill partial line with BG
 
-    if (upscale) {
-      switch (PicoIn.filter) {
-      case 3: h_upscale_bl4_4_5(pd, 320, q, 256, len, f_nop); break;
-      case 2: h_upscale_bl2_4_5(pd, 320, q, 256, len, f_nop); break;
-      case 1: h_upscale_snn_4_5(pd, 320, q, 256, len, f_nop); break;
-      default: h_upscale_nn_4_5(pd, 320, q, 256, len, f_nop); break;
-      }
+  if (upscale) {
+    switch (PicoIn.filter) {
+    case 3: h_upscale_bl4_4_5(pd, 320, q, 256, len, f_nop); break;
+    case 2: h_upscale_bl2_4_5(pd, 320, q, 256, len, f_nop); break;
+    case 1: h_upscale_snn_4_5(pd, 320, q, 256, len, f_nop); break;
+    default: h_upscale_nn_4_5(pd, 320, q, 256, len, f_nop); break;
     }
-    return 1;
   }
-  return 0;
 }
 
 // --------------------------------------------
@@ -1747,8 +1747,8 @@ void FinalizeLine555(int sh, int line, struct PicoEState *est)
   else if ((PicoIn.AHW & PAHW_SMS) && (est->Pico->video.reg[0] & 0x20))
     len -= 8, ps += 8;
 
-  if (BgcDMA(pd, len, est))
-    return;
+  if (est->rendstatus & PDRAW_BGC_DMA)
+    return BgcDMA(est);
 
   if ((est->rendstatus & PDRAW_SOFTSCALE) && len < 320) {
     if (len >= 240 && len <= 256) {
@@ -2193,6 +2193,8 @@ void PicoDrawBgcDMA(u16 *base, u32 source, u32 mask, int dlen, int sl)
     BgcDMAlen -= (len>>1)+xl;
     BgcDMAoffs = 0;
   }
+  if (BgcDMAlen > 0)
+    est->rendstatus |= PDRAW_BGC_DMA;
 }
 
 // also works for fast renderer
diff --git a/pico/draw_arm.S b/pico/draw_arm.S
index e3ef3e28..05f8ba3d 100644
--- a/pico/draw_arm.S
+++ b/pico/draw_arm.S
@@ -20,6 +20,7 @@
 .equ PDRAW_SHHI_DONE,     (1<<7)
 .equ PDRAW_BORDER_32,     (1<<9)
 .equ PDRAW_32X_SCALE,     (1<<12)
+.equ PDRAW_BGC_DMA,       (1<<14)
 .equ PDRAW_SOFTSCALE,     (1<<15)
 
 @ helpers
@@ -1632,10 +1633,16 @@ PicoDoHighPal555_end:
 FinalizeLine555:
     stmfd   sp!, {r4-r11,lr}
     mov     r11,r2               @ est
-    ldr     r8, [r11, #OFS_EST_Pico]
+    ldr     r4, [r11, #OFS_EST_rendstatus]
 
     bl      PicoDrawUpdateHighPal
 
+    tst     r4, #PDRAW_BGC_DMA
+    movne   r0, r11
+    ldmnefd sp!, {r4-r11,lr}
+    bne     BgcDMA
+
+    ldr     r8, [r11, #OFS_EST_Pico]
     add     r3, r11, #OFS_EST_HighPal
 
     mov     lr, #0xff
@@ -1644,7 +1651,6 @@ FinalizeLine555:
     ldr     r5, [r11, #OFS_EST_PicoOpt]
     ldr     r1, [r11, #OFS_EST_HighCol]
     ldr     r0, [r11, #OFS_EST_DrawLineDest]
-    ldr     r4, [r11, #OFS_EST_rendstatus]
     ldr     r7, [r5, #OFS_PicoIn_AHW-OFS_PicoIn_opt]
     ldrb    r12,[r8, #OFS_Pico_video_reg+12]
     ldrb    r6, [r8, #OFS_Pico_video_reg+0]
diff --git a/pico/pico.h b/pico/pico.h
index 2db502d5..6759f73d 100644
--- a/pico/pico.h
+++ b/pico/pico.h
@@ -235,6 +235,7 @@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est);
 #define PDRAW_30_ROWS      (1<<11) // 30 rows mode (240 lines)
 #define PDRAW_32X_SCALE    (1<<12) // scale CLUT layer for 32X
 #define PDRAW_SMS_BLANK_1  (1<<13) // 1st column blanked
+#define PDRAW_BGC_DMA      (1<<14) // in background color DMA
 #define PDRAW_SOFTSCALE    (1<<15) // H32 upscaling
 #define PDRAW_SYNC_NEEDED  (1<<16) // redraw needed
 #define PDRAW_SYNC_NEXT    (1<<17) // redraw next frame
-- 
2.39.5