From 39715ca6bf65d2e2dd889cdef4b39d584464d9e7 Mon Sep 17 00:00:00 2001
From: Peter Fors <peter.fors@mindkiller.com>
Date: Sun, 6 Apr 2025 12:27:12 +0200
Subject: added more mappers (buggy)

---
 ppu.c | 347 ++++++++++++++++++++++++++++++++++++++++--------------------------
 1 file changed, 210 insertions(+), 137 deletions(-)

(limited to 'ppu.c')

diff --git a/ppu.c b/ppu.c
index db78b91..da360e0 100644
--- a/ppu.c
+++ b/ppu.c
@@ -124,8 +124,8 @@ static inline uint8_t ppu_read(struct nes_state *state, uint32_t offset) {
 	return result;
 }
 
-__attribute__((hot))
-static void ppu_evaluate_sprites(struct nes_state *state) {
+__attribute__((always_inline, hot))
+static inline void ppu_evaluate_sprites(struct nes_state *state) {
 	struct ppu_state *ppu = &state->ppu;
 	uint8_t sprite_height = (ppu->reg_ctrl & 0x20) ? 16 : 8;
 	uint8_t n = 0;
@@ -158,8 +158,8 @@ static void ppu_evaluate_sprites(struct nes_state *state) {
 	ppu->sprite_count = n;
 }
 
-__attribute__((hot))
-static void ppu_fetch_sprite_patterns(struct nes_state *state) {
+__attribute__((always_inline, hot))
+static inline void ppu_fetch_sprite_patterns(struct nes_state *state) {
 	struct ppu_state *ppu = &state->ppu;
 	uint32_t addr;
 	uint32_t bank;
@@ -201,16 +201,14 @@ static void ppu_fetch_sprite_patterns(struct nes_state *state) {
 	}
 }
 
-__attribute__((hot))
-static void ppu_render_pixel(struct nes_state *state) {
+__attribute__((always_inline, hot))
+static inline void ppu_render_pixel(struct nes_state *state) {
 	struct ppu_state *ppu = &state->ppu;
 
 	uint32_t x = ppu->dot - 1;
 	uint32_t y = ppu->scanline;
 
-	// Fine X shift mask
-	// static const uint16_t fine_shift[8] = { 0x8000, 0x4000, 0x2000, 0x1000, 0x0800, 0x0400, 0x0200, 0x0100 };
-	uint16_t bit = 0x8000 >> ppu->fine_x;//fine_shift[ppu->fine_x];
+	uint16_t bit = 0x8000 >> ppu->fine_x;
 
 	uint8_t bg_pixel = 0;
 	uint8_t bg_palette = 0;
@@ -219,7 +217,6 @@ static void ppu_render_pixel(struct nes_state *state) {
 	uint8_t sp_prio = 0;
 	uint8_t sp_zero = 0;
 
-#if 1		// TODO(peter): Decide what I prefer, masking away unlikely path, or LIKELY hint to the compiler
 	uint8_t bg_mask = (ppu->reg_mask & 0x08) ? 0xff : 0x00;
 	uint8_t sp_mask = (ppu->reg_mask & 0x10) ? 0xff : 0x00;
 
@@ -247,36 +244,7 @@ static void ppu_render_pixel(struct nes_state *state) {
 		sp_zero    = (ppu->sprite_indexes[i] == 0);
 		break;
 	}
-#else
-	// Background fetch
-	if(LIKELY(ppu->reg_mask & 0x08)) {
-		uint8_t p0 = !!(ppu->bg_shift_pattern_low & bit);
-		uint8_t p1 = !!(ppu->bg_shift_pattern_high & bit);
-		bg_pixel = (p1 << 1) | p0;
-
-		uint8_t a0 = !!(ppu->bg_shift_attrib_low & bit);
-		uint8_t a1 = !!(ppu->bg_shift_attrib_high & bit);
-		bg_palette = (a1 << 1) | a0;
-	}
-
-	// Sprite fetch
-	if(LIKELY(ppu->reg_mask & 0x10)) {
-		for(uint8_t i = 0; i < ppu->sprite_count; i++) {
-			if(ppu->sprite_positions[i]) continue;
-
-			uint8_t lo = ppu->sprite_shift_lo[i];
-			uint8_t hi = ppu->sprite_shift_hi[i];
-			sp_pixel = ((hi & 0x80) >> 6) | ((lo & 0x80) >> 7);
-
-			if(!sp_pixel) continue;
 
-			sp_palette = ppu->secondary_oam[i * 4 + 2] & 3;
-			sp_prio    = ppu->sprite_priorities[i];
-			sp_zero    = (ppu->sprite_indexes[i] == 0);
-			break;
-		}
-	}
-#endif
 	// Final pixel composition
 	uint8_t palette_index = 0;
 	uint8_t bg_index = (bg_palette << 2) + bg_pixel;
@@ -309,117 +277,222 @@ static void ppu_tick(struct nes_state *state) {
 	for(uint32_t ppu_loops = 0; ppu_loops < 3; ++ppu_loops) {
 
 		if(LIKELY(rendering)) {
+			switch(scanline) {
+				case 0 ... 239: {
+					switch(dot) {
+						case 1:
+						    __attribute__((fallthrough));
+
+						case 2 ... 256:		// fallthrough: this is 1->256
+							ppu_render_pixel(state);
+
+							if(UNLIKELY(dot == 256)) {
+								if((ppu->vram_addr & 0x7000) != 0x7000) {
+									ppu->vram_addr += 0x1000;
+								} else {
+									ppu->vram_addr &= ~0x7000;
+									uint32_t y = (ppu->vram_addr & 0x03e0) >> 5;
+									if(y == 29) {
+										y = 0;
+										ppu->vram_addr ^= 0x0800;
+									} else if(y == 31) {
+										y = 0;
+									} else {
+										y++;
+									}
+									ppu->vram_addr = (ppu->vram_addr & ~0x03e0) | (y << 5);
+								}
+							}
+							__attribute__((fallthrough));
+
+						case 321 ... 336: {	// fallthrough: the code below has to run  1->256 + 321->336
+							// Rendering and tile fetch goes here
+							if(ppu->reg_mask & 0x10) {
+								for(uint32_t i = 0; i < ppu->sprite_count; i++) {
+									if(ppu->sprite_positions[i] > 0) {
+										ppu->sprite_positions[i]--;
+									} else {
+										ppu->sprite_shift_lo[i] <<= 1;
+										ppu->sprite_shift_hi[i] <<= 1;
+									}
+								}
+							}
 
-			if(ppu->even_frame && dot == 0) {
-				// call mapper_tick here.
-				ppu->dot++;
-			}
+							ppu->bg_shift_pattern_low <<= 1;
+							ppu->bg_shift_pattern_high <<= 1;
+							ppu->bg_shift_attrib_low <<= 1;
+							ppu->bg_shift_attrib_high <<= 1;
+
+							switch(dot % 8) {
+								case 1: {
+									uint32_t nt_addr = 0x2000 | (ppu->vram_addr & 0x0fff);
+									ppu->bg_next_tile_id = state->mapper.ciram_read(state, nt_addr);
+									break;
+								}
+								case 3: {
+									uint32_t attr_addr = 0x23c0 | (ppu->vram_addr & 0x0c00) | ((ppu->vram_addr >> 4) & 0x38) | ((ppu->vram_addr >> 2) & 0x07);
+									uint8_t attr = state->mapper.ciram_read(state, attr_addr & 0x0fff);
+									uint8_t shift = ((ppu->vram_addr >> 4) & 4) | (ppu->vram_addr & 2);
+									ppu->bg_next_tile_attrib = (attr >> shift) & 3;
+									break;
+								}
+								case 5: {
+									uint32_t base = (ppu->reg_ctrl & 0x10) ? 0x1000 : 0x0000;
+									uint32_t tile = ppu->bg_next_tile_id;
+									uint32_t fine_y = (ppu->vram_addr >> 12) & 7;
+									uint32_t addr_lsb = (base + tile * 16 + fine_y) & 0x1fff;
+									ppu->bg_next_tile_lsb = state->mapper.chr_read(state, addr_lsb);
+									break;
+								}
+								case 7: {
+									uint32_t base = (ppu->reg_ctrl & 0x10) ? 0x1000 : 0x0000;
+									uint32_t tile = ppu->bg_next_tile_id;
+									uint32_t fine_y = (ppu->vram_addr >> 12) & 7;
+									uint32_t addr_msb = (base + tile * 16 + fine_y + 8) & 0x1fff;
+									ppu->bg_next_tile_msb =  state->mapper.chr_read(state, addr_msb);
+									break;
+								}
+								case 0: {
+									ppu->bg_shift_pattern_low = (ppu->bg_shift_pattern_low & 0xff00) | ppu->bg_next_tile_lsb;
+									ppu->bg_shift_pattern_high = (ppu->bg_shift_pattern_high & 0xff00) | ppu->bg_next_tile_msb;
+
+									uint8_t a = ppu->bg_next_tile_attrib;
+									ppu->bg_shift_attrib_low = (ppu->bg_shift_attrib_low & 0xff00) | ((a & 1) ? 0xff : 0x00);
+									ppu->bg_shift_attrib_high = (ppu->bg_shift_attrib_high & 0xff00) | ((a & 2) ? 0xff : 0x00);
+
+									if((ppu->vram_addr & 0x001f) == 31) {
+										ppu->vram_addr &= ~0x001f;
+										ppu->vram_addr ^= 0x0400;
+									} else {
+										ppu->vram_addr++;
+									}
+
+									break;
+								}
+							}
+						} break;
 
-			if(scanline < 240 && dot >= 1 && dot <= 256) {
-				ppu_render_pixel(state);
-			}
+						case 257: {			// Sprite evaluation and horizontal vram transfer
+							ppu->vram_addr = (ppu->vram_addr & ~0x041f) | (ppu->temp_addr & 0x041f);
+							ppu_evaluate_sprites(state);
+						} break;
 
-			if(scanline < 240 || scanline == 261) {
-				if((dot >= 1 && dot <= 256) || (dot >= 321 && dot <= 336)) {
+						case 340: {			// sprite fetch pattern
+							ppu_fetch_sprite_patterns(state);
+						} break;
+					}
 
-					if(ppu->reg_mask & 0x10) {
-						for(uint32_t i = 0; i < ppu->sprite_count; i++) {
-							if(ppu->sprite_positions[i] > 0) {
-								ppu->sprite_positions[i]--;
-							} else {
-								ppu->sprite_shift_lo[i] <<= 1;
-								ppu->sprite_shift_hi[i] <<= 1;
+				} break;
+
+				case 261: {
+					switch(dot) {
+						case 1:
+							// Clear vblank, sprite 0 hit, etc.
+							__attribute__((fallthrough));
+
+						case 2 ... 256:
+							if(UNLIKELY(dot == 256)) {
+								if((ppu->vram_addr & 0x7000) != 0x7000) {
+									ppu->vram_addr += 0x1000;
+								} else {
+									ppu->vram_addr &= ~0x7000;
+									uint32_t y = (ppu->vram_addr & 0x03e0) >> 5;
+									if(y == 29) {
+										y = 0;
+										ppu->vram_addr ^= 0x0800;
+									} else if(y == 31) {
+										y = 0;
+									} else {
+										y++;
+									}
+									ppu->vram_addr = (ppu->vram_addr & ~0x03e0) | (y << 5);
+								}
 							}
-						}
-					}
 
-					ppu->bg_shift_pattern_low <<= 1;
-					ppu->bg_shift_pattern_high <<= 1;
-					ppu->bg_shift_attrib_low <<= 1;
-					ppu->bg_shift_attrib_high <<= 1;
-
-					switch(dot % 8) {
-						case 1: {
-							uint32_t nt_addr = 0x2000 | (ppu->vram_addr & 0x0fff);
-							ppu->bg_next_tile_id = state->mapper.ciram_read(state, nt_addr);
-							break;
-						}
-						case 3: {
-							uint32_t attr_addr = 0x23c0 | (ppu->vram_addr & 0x0c00) | ((ppu->vram_addr >> 4) & 0x38) | ((ppu->vram_addr >> 2) & 0x07);
-							uint8_t attr = state->mapper.ciram_read(state, attr_addr & 0x0fff);
-							uint8_t shift = ((ppu->vram_addr >> 4) & 4) | (ppu->vram_addr & 2);
-							ppu->bg_next_tile_attrib = (attr >> shift) & 3;
-							break;
-						}
-						case 5: {
-							uint32_t base = (ppu->reg_ctrl & 0x10) ? 0x1000 : 0x0000;
-							uint32_t tile = ppu->bg_next_tile_id;
-							uint32_t fine_y = (ppu->vram_addr >> 12) & 7;
-							uint32_t addr_lsb = (base + tile * 16 + fine_y) & 0x1fff;
-							ppu->bg_next_tile_lsb = state->mapper.chr_read(state, addr_lsb);
-							break;
-						}
-						case 7: {
-							uint32_t base = (ppu->reg_ctrl & 0x10) ? 0x1000 : 0x0000;
-							uint32_t tile = ppu->bg_next_tile_id;
-							uint32_t fine_y = (ppu->vram_addr >> 12) & 7;
-							uint32_t addr_msb = (base + tile * 16 + fine_y + 8) & 0x1fff;
-							ppu->bg_next_tile_msb =  state->mapper.chr_read(state, addr_msb);
-							break;
-						}
-						case 0: {
-							ppu->bg_shift_pattern_low = (ppu->bg_shift_pattern_low & 0xff00) | ppu->bg_next_tile_lsb;
-							ppu->bg_shift_pattern_high = (ppu->bg_shift_pattern_high & 0xff00) | ppu->bg_next_tile_msb;
-
-							uint8_t a = ppu->bg_next_tile_attrib;
-							ppu->bg_shift_attrib_low = (ppu->bg_shift_attrib_low & 0xff00) | ((a & 1) ? 0xff : 0x00);
-							ppu->bg_shift_attrib_high = (ppu->bg_shift_attrib_high & 0xff00) | ((a & 2) ? 0xff : 0x00);
-
-							if((ppu->vram_addr & 0x001f) == 31) {
-								ppu->vram_addr &= ~0x001f;
-								ppu->vram_addr ^= 0x0400;
-							} else {
-								ppu->vram_addr++;
+							__attribute__((fallthrough));
+						case 321 ... 336: {		// Rendering and tile fetch
+							if(ppu->reg_mask & 0x10) {
+								for(uint32_t i = 0; i < ppu->sprite_count; i++) {
+									if(ppu->sprite_positions[i] > 0) {
+										ppu->sprite_positions[i]--;
+									} else {
+										ppu->sprite_shift_lo[i] <<= 1;
+										ppu->sprite_shift_hi[i] <<= 1;
+									}
+								}
 							}
 
-							break;
-						}
-					}
-				}
+							ppu->bg_shift_pattern_low <<= 1;
+							ppu->bg_shift_pattern_high <<= 1;
+							ppu->bg_shift_attrib_low <<= 1;
+							ppu->bg_shift_attrib_high <<= 1;
+
+							switch(dot % 8) {
+								case 1: {
+									uint32_t nt_addr = 0x2000 | (ppu->vram_addr & 0x0fff);
+									ppu->bg_next_tile_id = state->mapper.ciram_read(state, nt_addr);
+									break;
+								}
+								case 3: {
+									uint32_t attr_addr = 0x23c0 | (ppu->vram_addr & 0x0c00) | ((ppu->vram_addr >> 4) & 0x38) | ((ppu->vram_addr >> 2) & 0x07);
+									uint8_t attr = state->mapper.ciram_read(state, attr_addr & 0x0fff);
+									uint8_t shift = ((ppu->vram_addr >> 4) & 4) | (ppu->vram_addr & 2);
+									ppu->bg_next_tile_attrib = (attr >> shift) & 3;
+									break;
+								}
+								case 5: {
+									uint32_t base = (ppu->reg_ctrl & 0x10) ? 0x1000 : 0x0000;
+									uint32_t tile = ppu->bg_next_tile_id;
+									uint32_t fine_y = (ppu->vram_addr >> 12) & 7;
+									uint32_t addr_lsb = (base + tile * 16 + fine_y) & 0x1fff;
+									ppu->bg_next_tile_lsb = state->mapper.chr_read(state, addr_lsb);
+									break;
+								}
+								case 7: {
+									uint32_t base = (ppu->reg_ctrl & 0x10) ? 0x1000 : 0x0000;
+									uint32_t tile = ppu->bg_next_tile_id;
+									uint32_t fine_y = (ppu->vram_addr >> 12) & 7;
+									uint32_t addr_msb = (base + tile * 16 + fine_y + 8) & 0x1fff;
+									ppu->bg_next_tile_msb =  state->mapper.chr_read(state, addr_msb);
+									break;
+								}
+								case 0: {
+									ppu->bg_shift_pattern_low = (ppu->bg_shift_pattern_low & 0xff00) | ppu->bg_next_tile_lsb;
+									ppu->bg_shift_pattern_high = (ppu->bg_shift_pattern_high & 0xff00) | ppu->bg_next_tile_msb;
+
+									uint8_t a = ppu->bg_next_tile_attrib;
+									ppu->bg_shift_attrib_low = (ppu->bg_shift_attrib_low & 0xff00) | ((a & 1) ? 0xff : 0x00);
+									ppu->bg_shift_attrib_high = (ppu->bg_shift_attrib_high & 0xff00) | ((a & 2) ? 0xff : 0x00);
+
+									if((ppu->vram_addr & 0x001f) == 31) {
+										ppu->vram_addr &= ~0x001f;
+										ppu->vram_addr ^= 0x0400;
+									} else {
+										ppu->vram_addr++;
+									}
+
+									break;
+								}
+							}
+						} break;
 
-				if(dot == 256) {
-					if((ppu->vram_addr & 0x7000) != 0x7000) {
-						ppu->vram_addr += 0x1000;
-					} else {
-						ppu->vram_addr &= ~0x7000;
-						uint32_t y = (ppu->vram_addr & 0x03e0) >> 5;
-						if(y == 29) {
-							y = 0;
-							ppu->vram_addr ^= 0x0800;
-						} else if(y == 31) {
-							y = 0;
-						} else {
-							y++;
-						}
-						ppu->vram_addr = (ppu->vram_addr & ~0x03e0) | (y << 5);
-					}
-				}
+						case 257: {					// Sprite evaluation and horizontal vram transfer
+							ppu->vram_addr = (ppu->vram_addr & ~0x041f) | (ppu->temp_addr & 0x041f);
+							ppu_evaluate_sprites(state);
 
-				if(dot == 257) {
-					ppu->vram_addr = (ppu->vram_addr & ~0x041f) | (ppu->temp_addr & 0x041f);
-				}
+						} break;
 
-				if(UNLIKELY(scanline == 261) && dot >= 280 && dot <= 304) {
-					ppu->vram_addr = (ppu->vram_addr & ~0x7be0) | (ppu->temp_addr & 0x7be0);
-				}
+						case 280 ... 304: {		// Vertical vram transfer
+							ppu->vram_addr = (ppu->vram_addr & ~0x7be0) | (ppu->temp_addr & 0x7be0);
+						} break;
 
-				if(dot == 257 && LIKELY(scanline < 240)) {
-					ppu_evaluate_sprites(state);
-				}
+						case 340: {					// Sprite pattern fetch
+							ppu_fetch_sprite_patterns(state);
+						} break;
 
-				if(dot == 340 && (LIKELY(scanline < 240) || UNLIKELY(scanline == 261))) {
-					ppu_fetch_sprite_patterns(state);
-				}
+					}
+
+				} break;
 			}
 		}
 
-- 
cgit v1.2.3