#include "vm.h"
#include "emu.h"
#include "fm7_display.h"
-#if defined(_OPENMP)
-#include <omp.h>
-#endif
-
-uint8_t DISPLAY::read_vram_l4_400l(uint32_t addr, uint32_t offset)
-{
#if defined(_FM77L4)
- if(addr < 0x8000) {
- if(workram) {
- uint32_t raddr = addr & 0x3fff;
- if((multimode_accessmask & 0x04) == 0) {
- return gvram[0x8000 + (raddr + offset) & 0x7fff];
- }
- return 0xff;
- }
- pagemod = addr & 0x4000;
- return gvram[((addr + offset) & mask) | pagemod];
- } else if(addr < 0x9800) {
- return textvram[addr & 0x0fff];
- } else { // $9800-$bfff
- return subrom_l4[addr - 0x9800];
- }
+#include "../hd46505.h"
#endif
- return 0xff;
-}
-
-void DISPLAY::write_vram_l4_400l(uint32_t addr, uint32_t offset, uint32_t data)
-{
-#if defined(_FM77L4)
- if(addr < 0x8000) {
- if(workram) {
- uint32_t raddr = addr & 0x3fff;
- if((multimode_accessmask & 0x04) == 0) {
- gvram[0x8000 + (raddr + offset) & 0x7fff] = (uint8_t)data;
- }
- return;
- }
- pagemod = addr & 0x4000;
- gvram[((addr + offset) & mask) | pagemod] = (uint8_t)data;
- } else if(addr < 0x9800) {
- textvram[addr & 0x0fff] = (uint8_t)data;
- } else { // $9800-$bfff
- //return subrom_l4[addr - 0x9800];
- }
- return;
-#endif
-}
+extern config_t config;
+namespace FM7 {
void DISPLAY::draw_screen()
{
//#endif
}
-extern config_t config;
void DISPLAY::draw_screen2()
{
int y;
if(!(vram_wrote_shadow | ff)) return;
vram_wrote_shadow = false;
if(display_mode == DISPLAY_MODE_8_200L) {
+ _render_command_data_t cmd;
+ uint32_t yoff_d = 0;
int ii;
yoff = 0;
-#ifdef USE_GREEN_DISPLAY
- if((config.dipswitch & FM7_DIPSW_GREEN_DISPLAY) != 0) {
- // Green display had only connected to FM-8, FM-7/NEW7 and FM-77.
- for(y = 0; y < 200; y += 8) {
- for(yy = 0; yy < 8; yy++) {
- if(!(vram_draw_table[y + yy] | ff)) continue;
- vram_draw_table[y + yy] = false;
-#if !defined(FIXED_FRAMEBUFFER_SIZE)
- p = emu->get_screen_buffer(y + yy);
- p2 = NULL;
+#if defined(USE_GREEN_DISPLAY)
+ if(use_green_monitor) {
+ cmd.palette = dpalette_pixel_green;
+ } else {
+ cmd.palette = dpalette_pixel;
+ }
#else
- p = emu->get_screen_buffer((y + yy) * 2);
- p2 = emu->get_screen_buffer((y + yy) * 2 + 1);
-#endif
- if(p == NULL) continue;
- yoff = (y + yy) * 80;
- {
- for(x = 0; x < 10; x++) {
- for(ii = 0; ii < 8; ii++) {
- GETVRAM_8_200L_GREEN(yoff + ii, p, p2, false, scan_line);
-#if defined(FIXED_FRAMEBUFFER_SIZE)
- p2 += 8;
-#endif
- p += 8;
- }
- yoff += 8;
- }
- }
- }
- }
- if(ff) force_update = false;
- return;
+ cmd.palette = dpalette_pixel;
+#endif
+ for(int i = 0; i < 3; i++) {
+ cmd.data[i] = gvram_shadow;
+ cmd.baseaddress[i] = i * 0x4000;
+ cmd.voffset[i] = yoff;
+ cmd.is_render[i] = false;
}
-#endif
+ if(!multimode_dispflags[0]) cmd.is_render[0] = true;
+ if(!multimode_dispflags[1]) cmd.is_render[1] = true;
+ if(!multimode_dispflags[2]) cmd.is_render[2] = true;
+ cmd.bit_trans_table[0] = (_bit_trans_table_t*)(&(bit_trans_table_2[0][0])); // B
+ cmd.bit_trans_table[1] = (_bit_trans_table_t*)(&(bit_trans_table_1[0][0])); // R
+ cmd.bit_trans_table[2] = (_bit_trans_table_t*)(&(bit_trans_table_0[0][0])); // G
+ cmd.xzoom = 1;
+ cmd.addrmask = 0x3fff;
+ cmd.addrmask2 = 0x3fff;
+ cmd.begin_pos = 0;
+ cmd.shift = 5;
+ cmd.render_width = 80;
for(y = 0; y < 200; y += 8) {
for(yy = 0; yy < 8; yy++) {
#endif
if(p == NULL) continue;
yoff = (y + yy) * 80;
+ for(int i = 0; i < 3; i++) {
+ cmd.voffset[i] = yoff;
+ }
+
# if defined(_FM77AV40EX) || defined(_FM77AV40SX)
+ int dpage;
+ dpage = vram_display_block;
+ bool window_inv = false;
if(window_opened && (wy_low <= (y + yy)) && (wy_high > (y + yy))) {
- for(x = 0; x < 80; x++) {
- if((x >= wx_begin) && (x < wx_end)) {
- GETVRAM_8_200L(yoff, p, p2, true, scan_line);
- } else {
- GETVRAM_8_200L(yoff, p, p2, false, scan_line);
+ if((wx_begin > 0) && (wx_begin < wx_end) && (wx_begin < 80)) {
+ // Window : left
+ cmd.begin_pos = 0;
+ window_inv = false;
+ int _wend = wx_end;
+ if(_wend >= 80) _wend = 80;
+ cmd.render_width = wx_begin;
+ yoff_d = (dpage != 0) ? 0x18000 : 0x00000;
+#if defined(_FM77AV_VARIANTS)
+ if(display_page_bak == 1) yoff_d += 0xc000;
+#endif
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x4000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, p, p2, scan_line);
+
+ // Center
+ cmd.begin_pos = wx_begin;
+ cmd.render_width = _wend - wx_begin;
+ yoff_d = (dpage != 0) ? 0x00000 : 0x18000;
+#if defined(_FM77AV_VARIANTS)
+ if(display_page_bak == 1) yoff_d += 0xc000;
+#endif
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x4000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), &(p2[cmd.begin_pos * 8]) , scan_line);
+ // Right
+ if(wx_end < 80) {
+ cmd.begin_pos = wx_end;
+ cmd.render_width = 80 - wx_end;
+ yoff_d = (dpage != 0) ? 0x18000 : 0x00000;
+#if defined(_FM77AV_VARIANTS)
+ if(display_page_bak == 1) yoff_d += 0xc000;
+#endif
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x4000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), &(p2[cmd.begin_pos * 8]), scan_line);
}
#if defined(FIXED_FRAMEBUFFER_SIZE)
- p2 += 8;
+ //CopyDrawnData(p, p2, 80, scan_line);
#endif
- p += 8;
- yoff++;
+ continue;
+ } else if((wx_begin <= 0) && (wx_begin < wx_end) && (wx_end >= 0)) {
+ // Left
+ cmd.begin_pos = 0;
+ cmd.render_width = wx_end;
+ yoff_d = (dpage != 0) ? 0x00000 : 0x18000;
+#if defined(_FM77AV_VARIANTS)
+ if(display_page_bak == 1) yoff_d += 0xc000;
+#endif
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x4000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ if(cmd.render_width > 0) Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), &(p2[cmd.begin_pos * 8]), scan_line);
+ // Right
+ if(wx_end < 80) {
+ cmd.begin_pos = wx_end;
+ cmd.render_width = 80 - wx_end;
+ yoff_d = (dpage != 0) ? 0x18000 : 0x00000;
+#if defined(_FM77AV_VARIANTS)
+ if(display_page_bak == 1) yoff_d += 0xc000;
+#endif
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x4000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), &(p2[cmd.begin_pos * 8]), scan_line);
+ }
+#if defined(FIXED_FRAMEBUFFER_SIZE)
+// CopyDrawnData(p, p2, 80, scan_line);
+#endif
+ continue;
}
- } else
-# endif
- {
- for(x = 0; x < 10; x++) {
- for(ii = 0; ii < 8; ii++) {
- GETVRAM_8_200L(yoff + ii, p, p2, false, scan_line);
+ }
+#endif
+ //cmd.begin_pos = 0;
+ //cmd.render_width = 80;
+# if defined(_FM77AV40EX) || defined(_FM77AV40SX)
+ yoff_d = (dpage != 0) ? 0x18000 : 0x00000;
+#else
+// yoff_d = 0;
+#endif
+#if defined(_FM77AV_VARIANTS)
+ if(display_page_bak == 1) yoff_d += 0xc000;
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x4000);
+ }
+#else
+// for(int i = 0; i < 3; i++) {
+// cmd.baseaddress[i] = i * 0x4000;
+// }
+#endif
+
+ Render8Colors_Line(&cmd, p, p2, scan_line);
#if defined(FIXED_FRAMEBUFFER_SIZE)
- p2 += 8;
+ //CopyDrawnData(p, p2, 80, scan_line);
#endif
+ }
+ }
+ if(ff) force_update = false;
+ return;
+ }
+#if defined(_FM77L4)
+ if(display_mode == DISPLAY_MODE_1_400L) {
+ int ii;
+ uint8_t *regs = l4crtc->get_regs();
+ cursor_start = (int)(regs[10] & 0x1f);
+ cursor_end = (int)(regs[11] & 0x1f);
+ cursor_type = (int)((regs[10] & 0x60) >> 5);
+ text_xmax = (int)((uint16_t)regs[1] << 1);
+ text_lines = (int)((regs[9] & 0x1f) + 1);
+ text_ymax = (int)(regs[6] & 0x7f);
+ yoff = 0;
+ // Green display had only connected to FM-8, FM-7/NEW7 and FM-77.
+ for(y = 0; y < 400; y += 8) {
+ bool renderf = false;
+ uint32_t naddr;
+ uint8_t bitcode;
+ uint8_t charcode;
+ uint8_t attr_code;
+ scrntype_t on_color;
+ int xlim, ylim;
+ bool do_green;
+ if((y & 0x0f) == 0) {
+ for(yy = 0; yy < 16; yy++) renderf |= vram_draw_table[y + yy];
+ renderf = renderf | ff;
+ if(renderf) {
+ for(yy = 0; yy < 16; yy++) vram_draw_table[y + yy] = true;
+ }
+ }
+ if(use_green_monitor) {
+ for(yy = 0; yy < 8; yy++) {
+ if(!(vram_draw_table[y + yy] | ff)) continue;
+ vram_draw_table[y + yy] = false;
+ p = emu->get_screen_buffer(y + yy);
+ if(p == NULL) continue;
+ yoff = (y + yy) * 80;
+ for(x = 0; x < 10; x++) {
+ for(ii = 0; ii < 8; ii++) {
+ GETVRAM_1_400L_GREEN(yoff + ii, p);
+ p += 8;
+ }
+ yoff += 8;
+ }
+ }
+ do_green = true;
+ } else {
+ for(yy = 0; yy < 8; yy++) {
+ if(!(vram_draw_table[y + yy] | ff)) continue;
+ vram_draw_table[y + yy] = false;
+ p = emu->get_screen_buffer(y + yy);
+ if(p == NULL) continue;
+ yoff = (y + yy) * 80;
+ for(x = 0; x < 10; x++) {
+ for(ii = 0; ii < 8; ii++) {
+ GETVRAM_1_400L(yoff + ii, p);
p += 8;
}
yoff += 8;
}
}
+ do_green = false;
+ }
+ // Draw Text
+ if(renderf) {
+ bool reverse;
+ bool display_char;
+ int raster;
+ bool cursor_rev;
+ uint8_t bitdata;
+ if(text_width40) {
+ xlim = 40;
+ } else {
+ xlim = 80;
+ }
+
+ for(x = 0; x < xlim; x++) {
+ naddr = (text_start_addr.w.l + ((y / text_lines) * text_xmax + x) * 2) & 0x0ffe;
+ charcode = text_vram[naddr];
+ attr_code = text_vram[naddr + 1];
+
+ on_color = GETVRAM_TEXTCOLOR(attr_code, do_green);
+
+ display_char = ((attr_code & 0x10) == 0);
+ reverse = ((attr_code & 0x08) != 0);
+
+ for(yy = 0; yy < 16; yy++) {
+ raster = y % text_lines;
+ bitdata = 0x00;
+ p = emu->get_screen_buffer(y + yy);
+ if(p == NULL) continue;
+ if((raster < 16) && (display_char || text_blink)) {
+ bitdata = subsys_cg_l4[(uint32_t)charcode * 16 + (uint32_t)raster];
+ }
+ cursor_rev = false;
+ if((naddr == (uint32_t)(cursor_addr.w.l)) && (cursor_type != 1) &&
+ (text_blink || (cursor_type == 0))) {
+ if((raster >= cursor_start) && (raster <= cursor_end)) {
+ cursor_rev = true;
+ }
+ }
+ bitdata = GETVRAM_TEXTPIX(bitdata, reverse, cursor_rev);
+ if(bitdata != 0) {
+ if(text_width40) {
+ scrntype_t *pp = &(p[x * 2]);
+ for(ii = 0; ii < 8; ii++) {
+ if((bitdata & 0x80) != 0) {
+ p[0] = on_color;
+ p[1] = on_color;
+ }
+ bitdata <<= 1;
+ p += 2;
+ }
+ } else {
+ scrntype_t *pp = &(p[x * 2]);
+ for(ii = 0; ii < 8; ii++) {
+ if((bitdata & 0x80) != 0) {
+ p[0] = on_color;
+ }
+ bitdata <<= 1;
+ p += 1;
+ }
+ }
+ }
+ }
+ }
}
-
}
if(ff) force_update = false;
return;
}
+#endif
# if defined(_FM77AV_VARIANTS)
if(display_mode == DISPLAY_MODE_4096) {
uint32_t mask = 0;
}
# if defined(_FM77AV40) || defined(_FM77AV40EX) || defined(_FM77AV40SX)
else if(display_mode == DISPLAY_MODE_8_400L) {
+ _render_command_data_t cmd;
int ii;
yoff = 0;
- //rgbmask = ~multimode_dispmask;
+ cmd.palette = dpalette_pixel;
+ for(int i = 0; i < 3; i++) {
+ cmd.data[i] = gvram_shadow;
+ cmd.baseaddress[i] = i * 0x8000;
+ cmd.voffset[i] = yoff;
+ cmd.is_render[i] = false;
+ }
+ if(!multimode_dispflags[0]) cmd.is_render[0] = true;
+ if(!multimode_dispflags[1]) cmd.is_render[1] = true;
+ if(!multimode_dispflags[2]) cmd.is_render[2] = true;
+ cmd.bit_trans_table[0] = (_bit_trans_table_t*)(&(bit_trans_table_2[0][0])); // B
+ cmd.bit_trans_table[1] = (_bit_trans_table_t*)(&(bit_trans_table_1[0][0])); // R
+ cmd.bit_trans_table[2] = (_bit_trans_table_t*)(&(bit_trans_table_0[0][0])); // G
+ cmd.xzoom = 1;
+ cmd.addrmask = 0x7fff;
+ cmd.addrmask2 = 0x7fff;
+ cmd.begin_pos = 0;
+ cmd.shift = 5;
+ cmd.render_width = 80;
for(y = 0; y < 400; y += 8) {
for(yy = 0; yy < 8; yy++) {
if(!(vram_draw_table[y + yy] | ff)) continue;
if(p == NULL) continue;
pp = p;
yoff = (y + yy) * 80;
+ for(int i = 0; i < 3; i++) {
+ cmd.voffset[i] = yoff;
+ }
+ int dpage;
+ bool window_inv = false;
+ uint32_t yoff_d;
+ dpage = vram_display_block;
# if defined(_FM77AV40EX) || defined(_FM77AV40SX)
- if(window_opened && (wy_low <= (y + yy)) && (wy_high > (y + yy))) {
- for(x = 0; x < 80; x++) {
- if((x >= wx_begin) && (x < wx_end)) {
- GETVRAM_8_400L(yoff, p, true);
- } else {
- GETVRAM_8_400L(yoff, p, false);
+ if(window_opened && (wy_low <= (y + yy)) && (wy_high > (y + yy))) {
+ if((wx_begin > 0) && (wx_begin < wx_end) && (wx_begin < 80)) {
+ // Window : left
+ cmd.begin_pos = 0;
+ window_inv = false;
+ int _wend = wx_end;
+ if(_wend >= 80) _wend = 80;
+ cmd.render_width = wx_begin;
+ yoff_d = (dpage != 0) ? 0x18000 : 0x00000;
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x8000);
}
- p += 8;
- yoff++;
- }
- } else
-# endif
- for(x = 0; x < 10; x++) {
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, p, NULL, false);
- for(ii = 0; ii < 8; ii++) {
- GETVRAM_8_400L(yoff + ii, p);
- p += 8;
+ // Center
+ cmd.begin_pos = wx_begin;
+ cmd.render_width = _wend - wx_begin;
+ yoff_d = (dpage != 0) ? 0x00000 : 0x18000;
+ if(display_page_bak == 1) yoff_d += 0xc000;
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x8000);
}
- yoff += 8;
- }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), NULL, false);
+ // Right
+ if(wx_end < 80) {
+ cmd.begin_pos = wx_end;
+ cmd.render_width = 80 - wx_end;
+ yoff_d = (dpage != 0) ? 0x18000 : 0x00000;
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x8000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), NULL, false);
+ }
+ continue;
+ } else if((wx_begin <= 0) && (wx_begin < wx_end) && (wx_end >= 0)) {
+ // Left
+ cmd.begin_pos = 0;
+ cmd.render_width = wx_end;
+ yoff_d = (dpage != 0) ? 0x00000 : 0x18000;
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x8000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ if(cmd.render_width > 0) Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), NULL, false);
+ // Right
+ if(wx_end < 80) {
+ cmd.begin_pos = wx_end;
+ cmd.render_width = 80 - wx_end;
+ yoff_d = (dpage != 0) ? 0x18000 : 0x00000;
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x8000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), NULL, false);
+ }
+ continue;
+ }
+ }
+# endif
+ // Not Opened
+ cmd.begin_pos = 0;
+ cmd.render_width = 80;
+ yoff_d = (dpage != 0) ? 0x18000 : 0x00000;
+ for(int i = 0; i < 3; i++) {
+ cmd.baseaddress[i] = yoff_d + (i * 0x8000);
+ }
+ if(cmd.render_width > 0) {
+ if(cmd.render_width > 80) cmd.render_width = 80;
+ }
+ Render8Colors_Line(&cmd, &(p[cmd.begin_pos * 8]), NULL, false);
}
}
if(ff) force_update = false;
screen_update_flag = false;
}
-void DISPLAY::GETVRAM_8_200L(int yoff, scrntype_t *p,
- scrntype_t *px,
- bool window_inv,
- bool scan_line)
+void DISPLAY::CopyDrawnData(scrntype_t* src, scrntype_t* dst, int width, bool scan_line)
{
- uint8_t b, r, g;
- uint32_t yoff_d;
-#if defined(_FM77AV40EX) || defined(_FM77AV40SX)
- int dpage = vram_display_block;
+ if(dst == NULL) return;
+ if(src == NULL) return;
+#if defined(_RGB555) || defined(_RGBA565)
+ static const int shift_factor = 2;
+#else // 24bit
+ static const int shift_factor = 3;
#endif
- if(p == NULL) return;
- yoff_d = 0;
- yoff_d = (yoff + yoff_d) & 0x3fff;
-
-#if defined(_FM77AV40EX) || defined(_FM77AV40SX)
- if(window_inv) {
- if(dpage == 0) {
- dpage = 1;
- } else {
- dpage = 0;
+ scrntype_vec8_t* vsrc = (scrntype_vec8_t*)__builtin_assume_aligned(src, sizeof(scrntype_vec8_t));
+ scrntype_vec8_t* vdst = (scrntype_vec8_t*)__builtin_assume_aligned(dst, sizeof(scrntype_vec8_t));
+ __DECL_ALIGNED(32) scrntype_vec8_t tmp_dd;
+ __DECL_ALIGNED(32) scrntype_vec8_t sline;
+
+ if(scan_line) {
+__DECL_VECTORIZED_LOOP
+ for(int i = 0; i < 8; i++) {
+ sline.w[i] = (scrntype_t)RGBA_COLOR(31, 31, 31, 255);
+ }
+__DECL_VECTORIZED_LOOP
+ for(int i = 0; i < width; i++) {
+ tmp_dd.v = vsrc[i].v;
+ tmp_dd.v = tmp_dd.v >> shift_factor;
+ tmp_dd.v = tmp_dd.v & sline.v;
+ vdst[i].v = tmp_dd.v;
+ }
+ } else {
+__DECL_VECTORIZED_LOOP
+ for(int i = 0; i < width; i++) {
+ tmp_dd.v = vsrc[i].v;
+ vdst[i].v = tmp_dd.v;
}
}
- if(dpage != 0) yoff_d += 0x18000;
-#endif
- b = r = g = 0;
-#if defined(_FM77AV_VARIANTS)
- if(display_page_bak == 1) yoff_d += 0xc000;
-#endif
- if(!multimode_dispflags[0]) b = gvram_shadow[yoff_d + 0x00000];
- if(!multimode_dispflags[1]) r = gvram_shadow[yoff_d + 0x04000];
- if(!multimode_dispflags[2]) g = gvram_shadow[yoff_d + 0x08000];
+}
- uint16_t *pg = &(bit_trans_table_0[g][0]);
- uint16_t *pr = &(bit_trans_table_1[r][0]);
- uint16_t *pb = &(bit_trans_table_2[b][0]);
- uint16_t tmp_d[8];
- scrntype_t tmp_dd[8];
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- tmp_d[i] = pr[i];
- tmp_d[i] = tmp_d[i] | pg[i];
- tmp_d[i] = tmp_d[i] | pb[i];
- tmp_d[i] = tmp_d[i] >> 5;
- }
+#if defined(_FM77L4)
+scrntype_t DISPLAY::GETVRAM_TEXTCOLOR(uint8_t attr, bool do_green)
+{
+ int color = attr & 0x07;
+ int r, g, b;
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- tmp_dd[i] = dpalette_pixel[tmp_d[i]];
- }
-#if defined(FIXED_FRAMEBUFFER_SIZE)
- if(scan_line) {
-/* Fancy scanline */
- #if defined(_RGB555) || defined(_RGBA565)
- static const int shift_factor = 2;
- #else // 24bit
- static const int shift_factor = 3;
- #endif
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- p[i] = tmp_dd[i];
- tmp_dd[i] = (tmp_dd[i] >> shift_factor) & (const scrntype_t)RGBA_COLOR(31, 31, 31, 255);;
- px[i] = tmp_dd[i];
+ static const int green_g_table[16] = {0, 24, 48, 64, 80, 96, 112, 128,
+ 140, 155, 175, 186, 210, 220, 240, 255};
+ if(do_green) {
+ if((attr & 0x20) != 0) color += 8;
+ r = b = 0;
+ g = green_g_table[color];
+ if(color >= 10) {
+ r = (color - 9) * 16;
+ b = (color - 9) * 16;
}
} else {
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- p[i] = tmp_dd[i];
- px[i] = tmp_dd[i];
+ if((attr & 0x20) != 0) {
+ g = ((color & 4) != 0) ? 255 : 0;
+ r = ((color & 2) != 0) ? 255 : 0;
+ b = ((color & 1) != 0) ? 255 : 0;
+ } else {
+ g = ((color & 4) != 0) ? 128 : 0;
+ r = ((color & 2) != 0) ? 128 : 0;
+ b = ((color & 1) != 0) ? 128 : 0;
}
}
-#else
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- p[i] = tmp_dd[i];
+ return RGBA_COLOR(r, g, b, 255);
+}
+
+uint8_t DISPLAY::GETVRAM_TEXTPIX(uint8_t bitdata, bool reverse, bool cursor_rev)
+{
+ uint8_t ret = bitdata;
+ if(reverse) {
+ ret = (uint8_t)(~ret);
}
-#endif
+ if(cursor_rev) {
+ ret = (uint8_t)(~ret);
+ }
+ return ret;
}
-#if defined(USE_GREEN_DISPLAY)
-void DISPLAY::GETVRAM_8_200L_GREEN(int yoff, scrntype_t *p,
- scrntype_t *px,
- bool window_inv,
- bool scan_line)
+void DISPLAY::GETVRAM_1_400L(int yoff, scrntype_t *p)
{
- uint8_t b, r, g;
+ uint8_t pixel;
uint32_t yoff_d;
-#if defined(_FM77AV40EX) || defined(_FM77AV40SX)
- int dpage = vram_display_block;
-#endif
if(p == NULL) return;
- yoff_d = 0;
- yoff_d = (yoff + yoff_d) & 0x3fff;
-
- b = r = g = 0;
- if(!multimode_dispflags[0]) b = gvram_shadow[yoff_d + 0x00000];
- if(!multimode_dispflags[1]) r = gvram_shadow[yoff_d + 0x04000];
- if(!multimode_dispflags[2]) g = gvram_shadow[yoff_d + 0x08000];
-
- uint16_t *pg = &(bit_trans_table_0[g][0]);
- uint16_t *pr = &(bit_trans_table_1[r][0]);
- uint16_t *pb = &(bit_trans_table_2[b][0]);
- uint16_t tmp_d[8];
- scrntype_t tmp_dd[8];
+ yoff_d = yoff & 0x7fff;
+ pixel = gvram_shadow[yoff_d];
+ uint16_vec8_t *ppx = (uint16_vec8_t *)__builtin_assume_aligned(&(bit_trans_table_0[pixel][0]), 16);
+ __DECL_ALIGNED(16) uint16_vec8_t tmp_d;
+ __DECL_ALIGNED(32) scrntype_vec8_t tmp_dd;
+ scrntype_vec8_t *vp = (scrntype_vec8_t *)__builtin_assume_aligned(p, sizeof(scrntype_vec8_t));
+ tmp_d.v = ppx->v;
+ tmp_d.v = tmp_d.v >> 5;
+
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
- tmp_d[i] = pr[i];
- tmp_d[i] = tmp_d[i] | pg[i];
- tmp_d[i] = tmp_d[i] | pb[i];
- tmp_d[i] = tmp_d[i] >> 5;
+ tmp_dd.w[i] = dpalette_pixel[tmp_d.w[i]];
}
+ vp->v = tmp_dd.v;
+}
+
+void DISPLAY::GETVRAM_1_400L_GREEN(int yoff, scrntype_t *p)
+{
+ uint8_t pixel;
+ uint32_t yoff_d;
+ if(p == NULL) return;
+ yoff_d = yoff & 0x7fff;
+ pixel = gvram_shadow[yoff_d];
+ uint16_vec8_t *ppx = (uint16_vec8_t *)__builtin_assume_aligned(&(bit_trans_table_0[pixel][0]), 16);
+ __DECL_ALIGNED(16) uint16_vec8_t tmp_d;
+ __DECL_ALIGNED(32) scrntype_vec8_t tmp_dd;
+ scrntype_vec8_t *vp = (scrntype_vec8_t *)__builtin_assume_aligned(p, sizeof(scrntype_vec8_t));
+
+ tmp_d.v = ppx->v;
+ tmp_d.v = tmp_d.v >> 5;
+
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
- tmp_dd[i] = dpalette_pixel_green[tmp_d[i]];
- }
-#if defined(FIXED_FRAMEBUFFER_SIZE)
- if(scan_line) {
-/* Fancy scanline */
- #if defined(_RGB555) || defined(_RGBA565)
- static const int shift_factor = 2;
- #else // 24bit
- static const int shift_factor = 3;
- #endif
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- p[i] = tmp_dd[i];
- tmp_dd[i] = (tmp_dd[i] >> shift_factor) & (const scrntype_t)RGBA_COLOR(31, 31, 31, 255);;
- px[i] = tmp_dd[i];
- }
- } else {
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- p[i] = tmp_dd[i];
- px[i] = tmp_dd[i];
- }
- }
-#else
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- p[i] = tmp_dd[i];
+ tmp_dd.w[i] = dpalette_pixel_green[tmp_d.w[i]];
}
-#endif
+ vp->v = tmp_dd.v;
+
}
#endif
+
#if defined(_FM77AV_VARIANTS)
void DISPLAY::GETVRAM_4096(int yoff, scrntype_t *p, scrntype_t *px,
uint32_t mask,
{
uint32_t b3, r3, g3;
uint8_t bb[4], rr[4], gg[4];
- uint16_t pixels[8];
- const uint16_t __masks[8] = {(uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask};
+ __DECL_ALIGNED(16) uint16_vec8_t pixels;
+ __DECL_ALIGNED(16) const uint16_t __masks[8] = {(uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask, (uint16_t)mask};
scrntype_t b, r, g;
uint32_t idx;;
scrntype_t pixel;
bb[1] = gvram_shadow[yoff_d1 + 0x02000];
rr[0] = gvram_shadow[yoff_d1 + 0x04000];
rr[1] = gvram_shadow[yoff_d1 + 0x06000];
-
gg[0] = gvram_shadow[yoff_d1 + 0x08000];
gg[1] = gvram_shadow[yoff_d1 + 0x0a000];
bb[2] = gvram_shadow[yoff_d2 + 0x0c000];
bb[3] = gvram_shadow[yoff_d2 + 0x0e000];
-
rr[2] = gvram_shadow[yoff_d2 + 0x10000];
rr[3] = gvram_shadow[yoff_d2 + 0x12000];
gg[2] = gvram_shadow[yoff_d2 + 0x14000];
gg[3] = gvram_shadow[yoff_d2 + 0x16000];
- uint16_t tmp_g[8], tmp_r[8], tmp_b[8];
uint16_t *p0, *p1, *p2, *p3;
#if !defined(FIXED_FRAMEBUFFER_SIZE)
- scrntype_t tmp_dd[8];
+ __DECL_ALIGNED(sizeof(scrntype_t) * 8) scrntype_t tmp_dd[8];
#else
- scrntype_t tmp_dd[16];
+ __DECL_ALIGNED(sizeof(scrntype_t) * 8) scrntype_t tmp_dd[16];
#endif
+ __DECL_ALIGNED(16) uint16_vec8_t tmp_g, tmp_r, tmp_b;
+ __v8hi *vp0, *vp1, *vp2, *vp3;
// G
- p0 = &(bit_trans_table_0[gg[0]][0]);
- p1 = &(bit_trans_table_1[gg[1]][0]);
- p2 = &(bit_trans_table_2[gg[2]][0]);
- p3 = &(bit_trans_table_3[gg[3]][0]);
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- tmp_g[i] = p0[i];
- tmp_g[i] = tmp_g[i] | p1[i];
- tmp_g[i] = tmp_g[i] | p2[i];
- tmp_g[i] = tmp_g[i] | p3[i];
- }
+ vp0 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_0[gg[0]][0]), 16);
+ vp1 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_1[gg[1]][0]), 16);
+ vp2 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_2[gg[2]][0]), 16);
+ vp3 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_3[gg[3]][0]), 16);
+ tmp_g.v = *vp0;
+ tmp_g.v = tmp_g.v | *vp1;
+ tmp_g.v = tmp_g.v | *vp2;
+ tmp_g.v = tmp_g.v | *vp3;
// R
- p0 = &(bit_trans_table_0[rr[0]][0]);
- p1 = &(bit_trans_table_1[rr[1]][0]);
- p2 = &(bit_trans_table_2[rr[2]][0]);
- p3 = &(bit_trans_table_3[rr[3]][0]);
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- tmp_r[i] = p0[i];
- tmp_r[i] = tmp_r[i] | p1[i];
- tmp_r[i] = tmp_r[i] | p2[i];
- tmp_r[i] = tmp_r[i] | p3[i];
- }
+ vp0 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_0[rr[0]][0]), 16);
+ vp1 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_1[rr[1]][0]), 16);
+ vp2 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_2[rr[2]][0]), 16);
+ vp3 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_3[rr[3]][0]), 16);
+ tmp_r.v = *vp0;
+ tmp_r.v = tmp_r.v | *vp1;
+ tmp_r.v = tmp_r.v | *vp2;
+ tmp_r.v = tmp_r.v | *vp3;
+
// B
- p0 = &(bit_trans_table_0[bb[0]][0]);
- p1 = &(bit_trans_table_1[bb[1]][0]);
- p2 = &(bit_trans_table_2[bb[2]][0]);
- p3 = &(bit_trans_table_3[bb[3]][0]);
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- tmp_b[i] = p0[i];
- tmp_b[i] = tmp_b[i] | p1[i];
- tmp_b[i] = tmp_b[i] | p2[i];
- tmp_b[i] = tmp_b[i] | p3[i];
- }
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- tmp_g[i] <<= 4;
- pixels[i] = tmp_b[i] >> 4;
- pixels[i] = pixels[i] | tmp_r[i];
- pixels[i] = pixels[i] | tmp_g[i];
- pixels[i] = pixels[i] & __masks[i];
- }
- //for(int i = 0; i < 8; i++) {
- // pixels[i] = pixels[i] & mask;
- //}
+ vp0 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_0[bb[0]][0]), 16);
+ vp1 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_1[bb[1]][0]), 16);
+ vp2 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_2[bb[2]][0]), 16);
+ vp3 = (__v8hi*)__builtin_assume_aligned(&(bit_trans_table_3[bb[3]][0]), 16);
+ tmp_b.v = *vp0;
+ tmp_b.v = tmp_b.v | *vp1;
+ tmp_b.v = tmp_b.v | *vp2;
+ tmp_b.v = tmp_b.v | *vp3;
+
+ __v8hi *mp = (__v8hi*)__masks;
+ tmp_g.v = tmp_g.v << 4;
+ tmp_b.v = tmp_b.v >> 4;
+ pixels.v = tmp_b.v;
+ pixels.v = pixels.v | tmp_r.v;
+ pixels.v = pixels.v | tmp_g.v;
+ pixels.v = pixels.v & *mp;
+
+
+ scrntype_vec8_t *vp = (scrntype_vec8_t*)__builtin_assume_aligned(p, sizeof(scrntype_vec8_t));
+ scrntype_vec8_t *dp = (scrntype_vec8_t*)__builtin_assume_aligned(tmp_dd, sizeof(scrntype_vec8_t));
#if !defined(FIXED_FRAMEBUFFER_SIZE)
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
tmp_dd[i] = analog_palette_pixel[pixels[i]];
- p[i] = tmp_dd[i];
}
+ vp->v = dp->v;
#else
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
- tmp_dd[i * 2] = tmp_dd[i * 2 + 1] = analog_palette_pixel[pixels[i]];;
+ tmp_dd[i * 2] = tmp_dd[i * 2 + 1] = analog_palette_pixel[pixels.w[i]];;
}
+ scrntype_vec8_t *vpx = (scrntype_vec8_t*)__builtin_assume_aligned(px, sizeof(scrntype_vec8_t));
+ __DECL_ALIGNED(32) scrntype_vec8_t vmask;
__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 16; i++) {
- p[i] = tmp_dd[i];
+ for(int i = 0; i < 2; i++) {
+ vp[i].v = dp[i].v;
}
if(scan_line) {
/* Fancy scanline */
__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 16; i++) {
+ for(int i = 0; i < 2; i++) {
#if defined(_RGB888) || defined(_RGBA888)
- tmp_dd[i] = tmp_dd[i] >> 3;
+ dp[i].v = dp[i].v >> 3;
#elif defined(_RGB555)
- tmp_dd[i] = tmp_dd[i] >> 2;
+ dp[i].v = dp[i].v >> 2;
#elif defined(_RGB565)
- tmp_dd[i] = tmp_dd[i] >> 2;
+ dp[i].v = dp[i].v >> 2;
#endif
}
__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 16; i++) {
- tmp_dd[i] = tmp_dd[i] & (const scrntype_t)RGBA_COLOR(31, 31, 31, 255);
- px[i] = tmp_dd[i];
+ for(int i = 0; i < 8; i++) {
+ vmask.w[i] = (const scrntype_t)RGBA_COLOR(31, 31, 31, 255);
+ }
+__DECL_VECTORIZED_LOOP
+ for(int i = 0; i < 2; i++) {
+ dp[i].v = dp[i].v & vmask.v;
+ vpx[i].v = dp[i].v;
}
} else {
__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 16; i++) {
- px[i] = tmp_dd[i];
+ for(int i = 0; i < 2; i++) {
+ vpx[i].v = dp[i].v;
}
}
-#endif
-
+#endif
}
#endif
#if defined(_FM77AV40) || defined(_FM77AV40EX) || defined(_FM77AV40SX)
-void DISPLAY::GETVRAM_8_400L(int yoff, scrntype_t *p,
- bool window_inv)
-{
- uint8_t b, r, g;
- uint32_t dot;
- uint32_t yoff_d;
-# if defined(_FM77AV40EX) || defined(_FM77AV40SX)
- int dpage = vram_display_block;
-# endif
- if(p == NULL) return;
- yoff_d = yoff;
-# if defined(_FM77AV40EX) || defined(_FM77AV40SX)
- if(window_inv) {
- if(dpage == 0) {
- dpage = 1;
- } else {
- dpage = 0;
- }
- }
- if(dpage != 0) yoff_d += 0x18000;
-# endif
- b = r = g = 0;
- if(!multimode_dispflags[0]) b = gvram_shadow[yoff_d + 0x00000];
- if(!multimode_dispflags[1]) r = gvram_shadow[yoff_d + 0x08000];
- if(!multimode_dispflags[2]) g = gvram_shadow[yoff_d + 0x10000];
-
- uint16_t *pg = &(bit_trans_table_0[g][0]);
- uint16_t *pr = &(bit_trans_table_1[r][0]);
- uint16_t *pb = &(bit_trans_table_2[b][0]);
- uint16_t tmp_d[8];
-
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- tmp_d[i] = pg[i] | pr[i] | pb[i];
- tmp_d[i] = tmp_d[i] >> 5;
- }
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- p[i] = dpalette_pixel[tmp_d[i]];
- }
-}
void DISPLAY::GETVRAM_256k(int yoff, scrntype_t *p, scrntype_t *px, bool scan_line)
{
yoff_d2 = yoff;
uint8_t bb[8], rr[8], gg[8];
- uint16_t *p0, *p1, *p2, *p3, *p4, *p5;
- uint32_t _btmp[8], _rtmp[8], _gtmp[8];
+
+ __DECL_ALIGNED(16) uint16_vec8_t _btmp;
+ __DECL_ALIGNED(16) uint16_vec8_t _rtmp;
+ __DECL_ALIGNED(16) uint16_vec8_t _gtmp;
+ uint16_vec8_t *vp0, *vp1, *vp2, *vp3, *vp4, *vp5;
#if !defined(FIXED_FRAMEBUFFER_SIZE)
- scrntype_t tmp_dd[8];
+ __DECL_ALIGNED(sizeof(scrntype_t) * 8) scrntype_t tmp_dd[8];
#else
- scrntype_t tmp_dd[16];
+ __DECL_ALIGNED(sizeof(scrntype_t) * 8) scrntype_t tmp_dd[16];
#endif
// if(mask & 0x01) {
if(!multimode_dispflags[0]) {
bb[4] = gvram_shadow[yoff_d1 + 0x18000];
bb[5] = gvram_shadow[yoff_d1 + 0x1a000];
- p0 = &(bit_trans_table_0[bb[0]][0]);
- p1 = &(bit_trans_table_1[bb[1]][0]);
- p2 = &(bit_trans_table_2[bb[2]][0]);
- p3 = &(bit_trans_table_3[bb[3]][0]);
- p4 = &(bit_trans_table_4[bb[4]][0]);
- p5 = &(bit_trans_table_5[bb[5]][0]);
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- _btmp[i] = p0[i];
- _btmp[i] = _btmp[i] | p1[i];
- _btmp[i] = _btmp[i] | p2[i];
- _btmp[i] = _btmp[i] | p3[i];
- _btmp[i] = _btmp[i] | p4[i];
- _btmp[i] = _btmp[i] | p5[i];
- }
+ vp0 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_0[bb[0]][0]), 16);
+ vp1 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_1[bb[1]][0]), 16);
+ vp2 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_2[bb[2]][0]), 16);
+ vp3 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_3[bb[3]][0]), 16);
+ vp4 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_4[bb[4]][0]), 16);
+ vp5 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_5[bb[5]][0]), 16);
+ _btmp.v = vp0->v;
+ _btmp.v = _btmp.v | vp1->v;
+ _btmp.v = _btmp.v | vp2->v;
+ _btmp.v = _btmp.v | vp3->v;
+ _btmp.v = _btmp.v | vp4->v;
+ _btmp.v = _btmp.v | vp5->v;
} else {
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
- _btmp[i] = 0;
+ _btmp.w[i] = 0;
}
}
if(!multimode_dispflags[1]) {
rr[4] = gvram_shadow[yoff_d1 + 0x1c000];
rr[5] = gvram_shadow[yoff_d1 + 0x1e000];
- p0 = &(bit_trans_table_0[rr[0]][0]);
- p1 = &(bit_trans_table_1[rr[1]][0]);
- p2 = &(bit_trans_table_2[rr[2]][0]);
- p3 = &(bit_trans_table_3[rr[3]][0]);
- p4 = &(bit_trans_table_4[rr[4]][0]);
- p5 = &(bit_trans_table_5[rr[5]][0]);
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- _rtmp[i] = p0[i];
- _rtmp[i] = _rtmp[i] | p1[i];
- _rtmp[i] = _rtmp[i] | p2[i];
- _rtmp[i] = _rtmp[i] | p3[i];
- _rtmp[i] = _rtmp[i] | p4[i];
- _rtmp[i] = _rtmp[i] | p5[i];
- }
+ vp0 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_0[rr[0]][0]), 16);
+ vp1 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_1[rr[1]][0]), 16);
+ vp2 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_2[rr[2]][0]), 16);
+ vp3 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_3[rr[3]][0]), 16);
+ vp4 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_4[rr[4]][0]), 16);
+ vp5 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_5[rr[5]][0]), 16);
+ _rtmp.v = vp0->v;
+ _rtmp.v = _rtmp.v | vp1->v;
+ _rtmp.v = _rtmp.v | vp2->v;
+ _rtmp.v = _rtmp.v | vp3->v;
+ _rtmp.v = _rtmp.v | vp4->v;
+ _rtmp.v = _rtmp.v | vp5->v;
} else {
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
- _rtmp[i] = 0;
+ _rtmp.w[i] = 0;
}
}
if(!multimode_dispflags[2]) {
gg[4] = gvram_shadow[yoff_d1 + 0x20000];
gg[5] = gvram_shadow[yoff_d1 + 0x22000];
- p0 = &(bit_trans_table_0[gg[0]][0]);
- p1 = &(bit_trans_table_1[gg[1]][0]);
- p2 = &(bit_trans_table_2[gg[2]][0]);
- p3 = &(bit_trans_table_3[gg[3]][0]);
- p4 = &(bit_trans_table_4[gg[4]][0]);
- p5 = &(bit_trans_table_5[gg[5]][0]);
-__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 8; i++) {
- _gtmp[i] = p0[i];
- _gtmp[i] = _gtmp[i] | p1[i];
- _gtmp[i] = _gtmp[i] | p2[i];
- _gtmp[i] = _gtmp[i] | p3[i];
- _gtmp[i] = _gtmp[i] | p4[i];
- _gtmp[i] = _gtmp[i] | p5[i];
- }
+ vp0 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_0[gg[0]][0]), 16);
+ vp1 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_1[gg[1]][0]), 16);
+ vp2 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_2[gg[2]][0]), 16);
+ vp3 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_3[gg[3]][0]), 16);
+ vp4 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_4[gg[4]][0]), 16);
+ vp5 = (uint16_vec8_t*)__builtin_assume_aligned(&(bit_trans_table_5[gg[5]][0]), 16);
+ _gtmp.v = vp0->v;
+ _gtmp.v = _gtmp.v | vp1->v;
+ _gtmp.v = _gtmp.v | vp2->v;
+ _gtmp.v = _gtmp.v | vp3->v;
+ _gtmp.v = _gtmp.v | vp4->v;
+ _gtmp.v = _gtmp.v | vp5->v;
} else {
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
- _gtmp[i] = 0;
+ _gtmp.w[i] = 0;
}
}
+
+ scrntype_vec8_t* vpp = (scrntype_vec8_t*)__builtin_assume_aligned(p, sizeof(scrntype_vec8_t));
+ scrntype_vec8_t* dp = (scrntype_vec8_t*)tmp_dd;
#if !defined(FIXED_FRAMEBUFFER_SIZE)
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
- tmp_dd[i] = RGB_COLOR(_rtmp[i], _gtmp[i], _btmp[i]);
- p[i] = tmp_dd[i];
+ tmp_dd[i] = RGB_COLOR(_rtmp.w[i], _gtmp.w[i], _btmp.w[i]);
}
+ vpp->v = dp->v;
#else
__DECL_VECTORIZED_LOOP
for(int i = 0; i < 8; i++) {
- tmp_dd[i * 2] = tmp_dd[i * 2 + 1] = RGB_COLOR(_rtmp[i], _gtmp[i], _btmp[i]);
+ tmp_dd[i * 2] = tmp_dd[i * 2 + 1] = RGB_COLOR(_rtmp.w[i], _gtmp.w[i], _btmp.w[i]);
}
+
__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 16; i++) {
- p[i] = tmp_dd[i];
+ for(int i = 0; i < 2; i++) {
+ vpp[i].v = dp[i].v;
}
+ scrntype_vec8_t* vpx = (scrntype_vec8_t*)__builtin_assume_aligned(px, sizeof(scrntype_vec8_t));
if(scan_line) {
/* Fancy scanline */
__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 16; i++) {
+ for(int i = 0; i < 2; i++) {
#if defined(_RGB888) || defined(_RGBA888)
- tmp_dd[i] = tmp_dd[i] >> 3;
+ dp[i].v = dp[i].v >> 3;
#elif defined(_RGB555)
- tmp_dd[i] = tmp_dd[i] >> 2;
+ dp[i].v = dp[i].v >> 2;
#elif defined(_RGB565)
- tmp_dd[i] = tmp_dd[i] >> 2;
+ dp[i].v = dp[i].v >> 2;
#endif
}
+ __DECL_ALIGNED(32) scrntype_vec8_t scanline_data;
__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 16; i++) {
- tmp_dd[i] = tmp_dd[i] & RGBA_COLOR(31, 31, 31, 256);
- px[i] = tmp_dd[i];
+ for(int i = 0; i < 8; i++) {
+ scanline_data.w[i] = RGBA_COLOR(31, 31, 31, 255);
}
- } else {
__DECL_VECTORIZED_LOOP
- for(int i = 0; i < 16; i++) {
- px[i] = tmp_dd[i];
+ for(int i = 0; i < 2; i++) {
+ dp[i].v = dp[i].v & scanline_data.v;
+ vpx[i].v = dp[i].v;
+ }
+ } else {
+ for(int i = 0; i < 2; i++) {
+ vpx[i].v = dp[i].v;
}
}
#endif
}
#endif
+
+}
OSDN Git Service
