+ convert_pal555(Pico32x.vdp_regs[0] & P32XV_PRI);
+
+ if ((Pico32x.vdp_regs[0] & P32XV_Mx) == 1) { // Packed Pixel Mode
+ unsigned char *p32xb = (void *)p32x;
+ if (Pico32x.vdp_regs[2 / 2] & P32XV_SFT)
+ p32xb++;
+ do_line_pp(pd, p32xb, pmd,);
+ }
+ else { // Run Length Mode
+ do_line_rl(pd, p32x, pmd,);
+ }
+}
+
+#define MD_LAYER_CODE \
+ *dst = palmd[*pmd]
+
+#define PICOSCAN_PRE \
+ PicoScan32xBegin(l + (lines_sft_offs & 0xff)); \
+ dst = Pico.est.DrawLineDest; \
+
+#define PICOSCAN_POST \
+ PicoScan32xEnd(l + (lines_sft_offs & 0xff)); \
+
+#define make_do_loop(name, pre_code, post_code, md_code) \
+/* Direct Color Mode */ \
+static void do_loop_dc##name(unsigned short *dst, \
+ unsigned short *dram, int lines_sft_offs, int mdbg) \
+{ \
+ int inv_bit = (Pico32x.vdp_regs[0] & P32XV_PRI) ? 0x8000 : 0; \
+ unsigned char *pmd = Pico.est.Draw2FB + \
+ 328 * (lines_sft_offs & 0xff) + 8; \
+ unsigned short *palmd = Pico.est.HighPal; \
+ unsigned short *p32x; \
+ int lines = lines_sft_offs >> 16; \
+ int l; \
+ (void)palmd; \
+ for (l = 0; l < lines; l++, pmd += 8) { \
+ pre_code; \
+ p32x = dram + dram[l]; \
+ do_line_dc(dst, p32x, pmd, inv_bit, md_code); \
+ post_code; \
+ } \
+} \
+ \
+/* Packed Pixel Mode */ \
+static void do_loop_pp##name(unsigned short *dst, \
+ unsigned short *dram, int lines_sft_offs, int mdbg) \
+{ \
+ unsigned short *pal = Pico32xMem->pal_native; \
+ unsigned char *pmd = Pico.est.Draw2FB + \
+ 328 * (lines_sft_offs & 0xff) + 8; \
+ unsigned short *palmd = Pico.est.HighPal; \
+ unsigned char *p32x; \
+ int lines = lines_sft_offs >> 16; \
+ int l; \
+ (void)palmd; \
+ for (l = 0; l < lines; l++, pmd += 8) { \
+ pre_code; \
+ p32x = (void *)(dram + dram[l]); \
+ p32x += (lines_sft_offs >> 8) & 1; \
+ do_line_pp(dst, p32x, pmd, md_code); \
+ post_code; \
+ } \
+} \
+ \
+/* Run Length Mode */ \
+static void do_loop_rl##name(unsigned short *dst, \
+ unsigned short *dram, int lines_sft_offs, int mdbg) \
+{ \
+ unsigned short *pal = Pico32xMem->pal_native; \
+ unsigned char *pmd = Pico.est.Draw2FB + \
+ 328 * (lines_sft_offs & 0xff) + 8; \
+ unsigned short *palmd = Pico.est.HighPal; \
+ unsigned short *p32x; \
+ int lines = lines_sft_offs >> 16; \
+ int l; \
+ (void)palmd; \
+ for (l = 0; l < lines; l++, pmd += 8) { \
+ pre_code; \
+ p32x = dram + dram[l]; \
+ do_line_rl(dst, p32x, pmd, md_code); \
+ post_code; \
+ } \
+}
+
+#ifdef _ASM_32X_DRAW
+#undef make_do_loop
+#define make_do_loop(name, pre_code, post_code, md_code) \
+extern void do_loop_dc##name(unsigned short *dst, \
+ unsigned short *dram, int lines_offs, int mdbg); \
+extern void do_loop_pp##name(unsigned short *dst, \
+ unsigned short *dram, int lines_offs, int mdbg); \
+extern void do_loop_rl##name(unsigned short *dst, \
+ unsigned short *dram, int lines_offs, int mdbg);
+#endif
+
+make_do_loop(,,,)
+make_do_loop(_md, , , MD_LAYER_CODE)
+make_do_loop(_scan, PICOSCAN_PRE, PICOSCAN_POST, )
+make_do_loop(_scan_md, PICOSCAN_PRE, PICOSCAN_POST, MD_LAYER_CODE)
+
+typedef void (*do_loop_func)(unsigned short *dst, unsigned short *dram, int lines, int mdbg);
+enum { DO_LOOP, DO_LOOP_MD, DO_LOOP_SCAN, DO_LOOP_MD_SCAN };
+
+static const do_loop_func do_loop_dc_f[] = { do_loop_dc, do_loop_dc_md, do_loop_dc_scan, do_loop_dc_scan_md };
+static const do_loop_func do_loop_pp_f[] = { do_loop_pp, do_loop_pp_md, do_loop_pp_scan, do_loop_pp_scan_md };
+static const do_loop_func do_loop_rl_f[] = { do_loop_rl, do_loop_rl_md, do_loop_rl_scan, do_loop_rl_scan_md };
+
+void PicoDraw32xLayer(int offs, int lines, int md_bg)
+{
+ int have_scan = PicoScan32xBegin != NULL && PicoScan32xEnd != NULL;
+ const do_loop_func *do_loop;
+ unsigned short *dram;
+ int lines_sft_offs;
+ int which_func;
+
+ Pico.est.DrawLineDest = (char *)DrawLineDestBase + offs * DrawLineDestIncrement;
+ dram = Pico32xMem->dram[Pico32x.vdp_regs[0x0a/2] & P32XV_FS];
+
+ if (Pico32xDrawMode == PDM32X_BOTH) {
+ if (Pico.m.dirtyPal)
+ PicoDrawUpdateHighPal();
+ }
+
+ if ((Pico32x.vdp_regs[0] & P32XV_Mx) == 2)
+ {
+ // Direct Color Mode
+ do_loop = do_loop_dc_f;
+ goto do_it;
+ }
+
+ if (Pico32x.dirty_pal)
+ convert_pal555(Pico32x.vdp_regs[0] & P32XV_PRI);
+
+ if ((Pico32x.vdp_regs[0] & P32XV_Mx) == 1)
+ {
+ // Packed Pixel Mode
+ do_loop = do_loop_pp_f;
+ }
+ else
+ {
+ // Run Length Mode
+ do_loop = do_loop_rl_f;
+ }
+
+do_it:
+ if (Pico32xDrawMode == PDM32X_BOTH)
+ which_func = have_scan ? DO_LOOP_MD_SCAN : DO_LOOP_MD;
+ else
+ which_func = have_scan ? DO_LOOP_SCAN : DO_LOOP;
+ lines_sft_offs = (lines << 16) | offs;
+ if (Pico32x.vdp_regs[2 / 2] & P32XV_SFT)
+ lines_sft_offs |= 1 << 8;
+
+ do_loop[which_func](Pico.est.DrawLineDest, dram, lines_sft_offs, md_bg);
+}
+
+// mostly unused, games tend to keep 32X layer on
+void PicoDraw32xLayerMdOnly(int offs, int lines)
+{
+ int have_scan = PicoScan32xBegin != NULL && PicoScan32xEnd != NULL;
+ unsigned short *dst = (void *)((char *)DrawLineDestBase + offs * DrawLineDestIncrement);
+ unsigned char *pmd = Pico.est.Draw2FB + 328 * offs + 8;
+ unsigned short *pal = Pico.est.HighPal;
+ int poffs = 0, plen = 320;
+ int l, p;
+
+ if (!(Pico.video.reg[12] & 1)) {
+ // 32col mode
+ poffs = 32;
+ plen = 256;
+ }
+
+ if (Pico.m.dirtyPal)
+ PicoDrawUpdateHighPal();
+
+ dst += poffs;
+ for (l = 0; l < lines; l++) {
+ if (have_scan) {
+ PicoScan32xBegin(l + offs);
+ dst = (unsigned short *)Pico.est.DrawLineDest + poffs;
+ }
+ for (p = 0; p < plen; p += 4) {
+ dst[p + 0] = pal[*pmd++];
+ dst[p + 1] = pal[*pmd++];
+ dst[p + 2] = pal[*pmd++];
+ dst[p + 3] = pal[*pmd++];