Finally back above 3000fps - this was a lot of work

vital@claybabble:/work/current/mknes(master*|u=)$ ./Bench.sh
Performance counter coverage: 100% (no multiplexing - full precision)

        352509230343   instructions                     #    4.81  insn per cycle
                                                        #    0.01  stalled cycles per insn
         73277044838   cycles                           #    5.416 GHz
          2957685039   stalled-cycles-frontend          #    4.04% frontend cycles idle
         70065301653   branches                         #    5.179 G/sec
           297927451   branch-misses                    #    0.43% of all branches

Throughput: 26056.16 MIPS, 5416.36 Mcycles/sec

cycles/frame  mean=  1788990  sd=  2803  relSD=0.157%  n=10
insn/frame    mean=  8606182  sd=     0  relSD=0.000%  n=10
time (ms)     mean= 1352.883  sd= 2.140  relSD=0.158%  n=10
FPS (frames/second) = 3027.61
           ms/frame = 0.330294

2 files changed, 100 insertions, 13 deletions

diff --git a/mknes.c b/mknes.c
index b82fc55..244a032 100644
--- a/mknes.c
+++ b/mknes.c
@@ -58,8 +58,8 @@ static void audio_callback(int16_t *data, size_t frames) { }
 #ifdef BENCHMARK
 // Embed the ROM for benchmarking to eliminate file I/O overhead
 // Uncomment the ROM you want to benchmark:
-// INCBIN_BYTES(benchmark_rom, "data/Life Force (USA).nes");
-INCBIN_BYTES(benchmark_rom, "data/0000/Super Mario Bros. (World) (HVC-SM).nes");
+INCBIN_BYTES(benchmark_rom, "data/Life Force (USA).nes");
+// INCBIN_BYTES(benchmark_rom, "data/0000/Super Mario Bros. (World) (HVC-SM).nes");
 // INCBIN_BYTES(benchmark_rom, "data/0003/Gradius (USA).nes");
 #endif
 
@@ -271,7 +271,7 @@ int main(int argc, char **argv) {
 	// ines2_load(nstate, "data/0000/Excitebike (Japan, USA).nes");
 	// ines2_load(nstate, "data/0000/Ice Climber (USA, Europe, Korea).nes");
 	// ines2_load(nstate, "data/0000/Kung Fu (Japan, USA).nes");
-	// ines2_load(nstate, "data/0000/Super Mario Bros. (World) (HVC-SM).nes");
+	ines2_load(nstate, "data/0000/Super Mario Bros. (World) (HVC-SM).nes");
 	// ines2_load(nstate, "data/Super Mario Bros. (W) (V1.0) [!].nes");
 	// ines2_load(nstate, "data/Super Mario Bros. (JU) [!].nes");
 	// ines2_load(nstate, "data/0000/Urban Champion (World).nes");
@@ -293,7 +293,7 @@ int main(int argc, char **argv) {
 	// ines2_load(nstate, "data/0000/Xevious - The Avenger (USA).zip");
 	// ines2_load(nstate, "data/tv.nes");
 
-	ines2_load(nstate, "data/Life Force (USA).nes");	// 2002
+	// ines2_load(nstate, "data/Life Force (USA).nes");	// 2002
 
 	// ines2_load(nstate, "data/0003/Flipull - An Exciting Cube Game (Japan) (En).zip");
 	// ines2_load(nstate, "data/0003/Friday the 13th (USA).zip");
diff --git a/mknes_ppu.c b/mknes_ppu.c
index 6e8e3ae..249f9f9 100644
--- a/mknes_ppu.c
+++ b/mknes_ppu.c
@@ -109,6 +109,93 @@ static inline void ppu_fetch_sprite_patterns(struct nes_state * restrict state,
 	}
 }
 
+
+#if 0
+// layout reminder
+// struct sprite_data { u8 shift_lo, shift_hi, position, priority, palette; } __attribute__((packed));
+
+__attribute__((always_inline, hot, optimize("no-jump-tables","no-unroll-loops")))
+static inline void ppu_render_pixel(struct nes_state * restrict state, uint32_t x, uint32_t y, uint8_t mask_reg) {
+	struct ppu_state *restrict ppu = &state->ppu;
+
+	// 32-bit temps for bg path; keep struct 16-bit/8-bit
+	const uint32_t s = 15u - (uint32_t)ppu->fine_x;
+	const uint32_t show_bg      = (mask_reg & PPU_MASK_SHOW_BG) != 0;
+	const uint32_t show_sprites = (mask_reg & PPU_MASK_SHOW_SPRITES) != 0;
+	const uint32_t left_bg      = (mask_reg & 0x02) != 0;
+	const uint32_t left_sp      = (mask_reg & 0x04) != 0;
+	const uint32_t x_ge_8       = (x & ~7u) != 0;
+
+	const uint32_t bg_on = show_bg & (left_bg | x_ge_8);
+	const uint32_t sp_on = show_sprites & (left_sp | x_ge_8);
+
+	const uint32_t pat_lo = (uint32_t)ppu->bg_shift_pattern_low;
+	const uint32_t pat_hi = (uint32_t)ppu->bg_shift_pattern_high;
+	const uint32_t att_lo = (uint32_t)ppu->bg_shift_attrib_low;
+	const uint32_t att_hi = (uint32_t)ppu->bg_shift_attrib_high;
+
+	const uint32_t p0 = (pat_lo >> s) & 1u;
+	const uint32_t p1 = (pat_hi >> s) & 1u;
+	const uint32_t a0 = (att_lo >> s) & 1u;
+	const uint32_t a1 = (att_hi >> s) & 1u;
+
+	const uint32_t bg_pixel   = ((p1 << 1) | p0) & -bg_on;
+	const uint32_t bg_palette = ((a1 << 1) | a0) & -bg_on;
+
+	// Sprite resolve: only load fields in this order:
+	// position (branch filter) -> shift bytes (pixel test) -> meta (on hit)
+	uint32_t sp_pixel = 0, sp_palette = 0, sp_prio = 0, sp_zero = 0;
+
+	if (sp_on) {
+		struct sprite_data * restrict s_ptr = ppu->sprites;      // address calc only
+		uint32_t n = ppu->sprite_count;                          // one load
+		if (n) {
+			uint32_t i = 0;
+		find_sprite:
+			// 1) position
+			uint32_t pos = s_ptr->position;                      // load 1
+			if (!pos) {
+					// 2) pixel from shift bytes
+					uint32_t sh = s_ptr->shift_hi;                   // load 2
+					uint32_t sl = s_ptr->shift_lo;                   // load 3
+					uint32_t pix = ((sh & 0x80u) >> 6) | ((sl & 0x80u) >> 7);
+					if (pix) {
+						// 3) only now fetch metadata
+						sp_pixel   = pix;
+						sp_palette = s_ptr->palette;                 // load 4 (only on hit)
+						sp_prio    = s_ptr->priority;                // load 5 (only on hit)
+						sp_zero    = (ppu->sprite_zero_in_range != 0u) & (i == 0u);
+						goto sprite_done;
+					}
+			}
+			// next sprite
+			++s_ptr; ++i;
+			if (i < n) goto find_sprite;
+		}
+	}
+sprite_done: ;
+
+	const uint32_t bg_index = (bg_palette << 2) + bg_pixel;
+	const uint32_t sp_index = (sp_palette << 2) + sp_pixel;
+	const uint32_t selector = ((bg_pixel != 0u) << 1) | (sp_pixel != 0u);
+
+	// Two-way combine with minimal control flow
+	uint32_t palette_index = 0;
+	if (selector == 1u) {
+		palette_index = 0x10u | sp_index;
+	} else if (selector == 2u) {
+		palette_index = bg_index;
+	} else if (selector == 3u) {
+		const uint32_t use_bg = (sp_prio != 0u);
+		palette_index = use_bg ? bg_index : (0x10u | sp_index);
+		if ((ppu->sprite_zero_in_range != 0u) & (sp_zero != 0u) & (x <= 254u)) {
+			ppu->reg_status |= PPU_STATUS_SPRITE_ZERO_HIT;
+		}
+	}
+	state->pixels[y * 256 + x] = ppu->palette[palette_index];
+}
+
+#else
 __attribute__((always_inline, hot, optimize("no-jump-tables", "no-unroll-loops")))
 static inline void ppu_render_pixel(struct nes_state * restrict state, uint32_t x, uint32_t y, uint8_t mask_reg) {
 	struct ppu_state *restrict ppu = &state->ppu;
@@ -159,25 +246,25 @@ sprite_found:
 	}
 no_sprite:
 
-
 	// Final pixel composition
 	uint8_t bg_index = (bg_palette << 2) + bg_pixel;
 	uint8_t sp_index = (sp_palette << 2) + sp_pixel;
 	uint8_t selector = (bg_pixel ? 2 : 0) | (sp_pixel ? 1 : 0);
 	uint8_t palette_index = 0;
 
-	switch(selector) {
-		// case 0: { palette_index = 0;												} break;
-		case 1: { palette_index = 0x10 | sp_index;							} break;
-		case 2: { palette_index = bg_index;										} break;
-		case 3: {
-			palette_index = (sp_prio) ? bg_index : 0x10 | sp_index;
-			ppu->reg_status |= (sp_zero && x < 255) ? PPU_STATUS_SPRITE_ZERO_HIT : 0;
-		} break;		// NOTE(peter): Sprite zero hit!
+	if(selector == 1) {
+		palette_index = 0x10 | sp_index;
+	} else if(selector == 2) {
+		palette_index = bg_index;
+	} else if(selector == 3) {
+		palette_index = (sp_prio) ? bg_index : (0x10 | sp_index);
+		if (sp_zero && x <= 254) ppu->reg_status |= PPU_STATUS_SPRITE_ZERO_HIT;
 	}
 
 	state->pixels[y * 256 + x] = ppu->palette[palette_index];		// NOTE(peter): Add color_emphasis bits (expand palette to 8x).
 }
+#endif
+
 
 __attribute__((noinline, hot, optimize("no-jump-tables", "unroll-loops")))
 static void ppu_tick(struct nes_state *state) {