Rearrangement and refactoring and optimizations and more accuracy

1 files changed, 158 insertions, 69 deletions

diff --git a/mknes.h b/mknes.h
index d3e3840..c41d7af 100644
--- a/mknes.h
+++ b/mknes.h
@@ -19,55 +19,69 @@
 #define MIRROR_FOURSCREEN 2
 
 struct ppu_state {
-	uint16_t bg_shift_pattern_low;
-	uint16_t bg_shift_pattern_high;
-	uint16_t bg_shift_attrib_low;
-	uint16_t bg_shift_attrib_high;
-
-	uint16_t scanline;
-	uint16_t dot;
-	uint16_t vram_addr;
-	uint16_t temp_addr;
-	uint8_t fine_x;
-	uint8_t bg_next_tile_id;
-	uint8_t bg_next_tile_attrib;
-	uint8_t bg_next_tile_lsb;
-	uint8_t bg_next_tile_msb;
-	uint8_t oam_addr;
-	uint8_t oam_data;
-	uint8_t even_frame;
-
-	uint8_t reg_ctrl;
-	uint8_t reg_mask;
-	uint8_t reg_status;
-
-	uint8_t write_latch;
-	uint8_t vram_read_buffer;
-	uint8_t open_bus;
-
-	uint8_t sprite_count;
-	uint8_t palette[32];
-
-	uint8_t frame_ready;
-	// NOTE(peter): CACHELINE 2 start here!
-
-	uint8_t sprite_indexes[8];
-	uint8_t sprite_positions[8];
-	uint8_t sprite_priorities[8];
-	uint8_t sprite_shift_lo[8];
-	uint8_t sprite_shift_hi[8];
-	uint8_t secondary_oam[32];
-	uint8_t oam[256];
-
-	uint8_t input[2];				// Controller 1 & 2
-	uint8_t input_latch[2];		// Latched inputs after strobe
-	uint8_t input_bit[2];		// Current bit position being shifted out
-	uint8_t input_strobe;		// Control bit (0 or 1)
-} __attribute__((packed, aligned(64)));
+	uint16_t bg_shift_pattern_low;	// 0
+	uint16_t bg_shift_pattern_high;	// 2
+	uint16_t bg_shift_attrib_low;		// 4
+	uint16_t bg_shift_attrib_high;	// 6
+
+	uint16_t scanline;					// 8
+	uint16_t dot;							// 10
+	uint16_t vram_addr;					// 12
+	uint16_t temp_addr;					// 14
+	uint8_t fine_x;						// 16
+	uint8_t bg_next_tile_id;			// 17
+	uint8_t bg_next_tile_attrib;		// 18
+	uint8_t bg_next_tile_lsb;			// 19
+	uint8_t bg_next_tile_msb;			// 20
+	uint8_t oam_addr;						// 21
+	uint8_t oam_data;						// 22
+	uint8_t even_frame;					// 23
+
+	uint8_t reg_ctrl;						// 24
+	uint8_t reg_mask;						// 25
+	uint8_t reg_status;					// 26
+
+	uint8_t write_latch;					// 27
+	uint8_t vram_read_buffer;			// 28
+	uint8_t open_bus;						// 29
+
+	uint8_t sprite_count;				// 30
+	uint8_t overflow_scheduled_dot;	// 31
+	uint8_t palette[32];					// 32
+
+	// NOTE(peter): CACHELINE 2
+	uint8_t secondary_oam[32] __attribute__((aligned(64)));
+	uint8_t temp_secondary_oam[32];
+
+	// NOTE(peter): CACHELINE 3
+	struct sprite_data {
+		uint8_t shift_lo;
+		uint8_t shift_hi;
+		uint8_t position;
+		uint8_t priority;
+		uint8_t palette;
+	} __attribute__((packed, aligned(64))) sprites[8];
+
+	uint8_t input[2];						// 40 - Controller 1 & 2
+	uint8_t input_latch[2];				// 42 - Latched inputs after strobe
+	uint8_t input_bit[2];				// 44 - Current bit position being shifted out
+	uint8_t input_strobe;				// 46 - Control bit (0 or 1)
+	uint8_t frame_ready;					// 47
+	uint8_t sprite_zero_in_range;		// 48 - Boolean: is sprite 0 in range (will always be slot 0 if true)
+	uint8_t sprite_count_next;			// 49 - Sprite count for next scanline
+	// 15 bytes left.
+
+	// NOTE(peter): CACHELINE 4
+	uint8_t oam[256] __attribute__((aligned(64)));
+} __attribute__((packed));
 
 struct apu_state {
 	uint32_t frame_cycle;
 
+	uint16_t dmc_current_addr;
+	uint16_t dmc_bytes_remaining;
+	uint16_t dmc_sample_timer;
+
 	uint8_t mode;
 	uint8_t irq_inhibit;
 	uint8_t irq_pending;
@@ -79,11 +93,7 @@ struct apu_state {
 	uint8_t dmc_freq_index;
 	uint8_t dmc_sample_addr;
 	uint8_t dmc_sample_len;
-
-	uint16_t dmc_current_addr;
-	uint16_t dmc_bytes_remaining;
-	uint16_t dmc_sample_timer;
-} __attribute__((packed, aligned(64)));
+} __attribute__((packed));
 
 struct cpu_state {
 	size_t cycles;
@@ -96,15 +106,14 @@ struct cpu_state {
 	uint8_t n;		// Negative Flag
 	uint8_t v;		// Overflow Flag
 	// uint8_t b;		// Break Flag (Set by BRK instruction) -- does not exist outside the stack
-	uint8_t d;		// Decimal Flag
+	uint8_t d;		// Decimal Flag (exists but ignored on NES)
 	uint8_t i;		// Interrupt Disable Flag
 	uint8_t z;		// Zero Flag
 	uint8_t c;		// Carry Flag
 // --
 	uint8_t irq_pending;
 	uint8_t nmi_pending;
-} __attribute__((packed, aligned(64)));
-
+};
 
 struct ines_state {
 	uint8_t mapper;
@@ -112,24 +121,24 @@ struct ines_state {
 	uint8_t mirroring;	// 0 = H, 1 = V, 2 = 4-screen
 	uint32_t prg_size;
 	uint32_t chr_size;
-} __attribute__((packed, aligned(64)));
+};
 
 struct nes_state {
-	struct ppu_state ppu;
-	struct mapper_functions mapper_function;
-	union mapper_data mapper_data;
-	struct cpu_state cpu;
-	struct ines_state ines;
-	struct apu_state apu;
-
-	uint8_t ram[RAM_SIZE]			__attribute__((aligned(4096)));
-	uint8_t ciram[CIRAM_SIZE]		__attribute__((aligned(4096)));
-	uint8_t prg_rom[PRG_ROM_SIZE]	__attribute__((aligned(4096)));
-	uint8_t chr_rom[CHR_ROM_SIZE]	__attribute__((aligned(4096)));
-	uint8_t chr_ram[CHR_RAM_SIZE]	__attribute__((aligned(4096)));
-	uint8_t pixels[PIXELS_SIZE]	__attribute__((aligned(4096)));
-	uint8_t sram[SRAM_SIZE]			__attribute__((aligned(4096)));
-} __attribute__((packed, aligned(4096)));
+	struct ppu_state ppu								__attribute__((aligned(4096)));
+	struct mapper_functions mapper_function	__attribute__((aligned(64)));
+	union mapper_data mapper_data					__attribute__((aligned(64)));
+	struct cpu_state cpu								__attribute__((aligned(64)));
+	struct ines_state ines							__attribute__((aligned(64)));
+	struct apu_state apu								__attribute__((aligned(64)));
+
+	uint8_t ram[RAM_SIZE]							__attribute__((aligned(4096)));
+	uint8_t ciram[CIRAM_SIZE]						__attribute__((aligned(4096)));
+	uint8_t prg_rom[PRG_ROM_SIZE]					__attribute__((aligned(4096)));
+	uint8_t chr_rom[CHR_ROM_SIZE]					__attribute__((aligned(4096)));
+	uint8_t chr_ram[CHR_RAM_SIZE]					__attribute__((aligned(4096)));
+	uint8_t pixels[PIXELS_SIZE]					__attribute__((aligned(4096)));
+	uint8_t sram[SRAM_SIZE]							__attribute__((aligned(4096)));
+};
 
 __attribute__((aligned(4096))) static uint32_t nes_palette[65] = {
 	0x585858ff, 0x00237cff, 0x0d1099ff, 0x300092ff, 0x4f006cff, 0x600035ff, 0x5c0500ff, 0x461800ff,
@@ -142,3 +151,83 @@ __attribute__((aligned(4096))) static uint32_t nes_palette[65] = {
 	0xdee086ff, 0xc6ec87ff, 0xb2f29dff, 0xa7f0c3ff, 0xa8e7f0ff, 0xacacacff, 0x000000ff, 0x000000ff,
 	0xffffffff	// one extra for debug-coloring...
 };
+
+struct remake_state {
+	struct { int32_t x, y, w, h; } viewport;
+	int32_t mouse_dx;
+	int32_t mouse_dy;
+
+	float contrast;
+	float saturation;
+	float brightness;
+	float tone_data[4];
+
+	// OpenGL Objects
+	GLuint shader_program;
+	GLuint persistence_program;
+	// GLuint upscale_program;
+	GLuint upscale_warp_program;
+	GLuint bloom_extract_program;
+	GLuint bloom_blur_program;
+	GLuint bloom_warp_program;
+	GLuint bloom_composite_program;
+
+	GLuint texture;
+	GLuint persistence_texture;
+	GLuint persistence_output_texture;
+	GLuint crt_output_texture;
+	GLuint bloom_texture;
+	GLuint bloom_temp_texture;
+	GLuint bloom_warped_texture;
+	GLuint upscaled_source_texture;
+
+	GLuint persistence_fbo;
+	GLuint upscaled_source_fbo;
+	GLuint crt_fbo;
+	GLuint bloom_fbo;
+	GLuint bloom_temp_fbo;
+	GLuint bloom_warp_fbo;
+
+	GLuint vao;
+	GLuint vbo;
+	GLuint ebo;
+
+	// CRT Shader Uniforms
+	GLuint uniform_resolution;
+	GLuint uniform_src_image_size;
+	GLuint uniform_brightness;
+	GLuint uniform_tone;
+	GLuint uniform_crt_emulation;
+	GLuint uniform_apply_mask;
+	GLuint uniform_sampler_location;
+
+	// Bloom Shader Uniforms
+	GLuint bloom_uniform_threshold;
+	GLuint bloom_uniform_sampler;
+	GLuint blur_uniform_horizontal;
+	GLuint blur_uniform_sampler;
+	GLuint composite_uniform_bloom_strength;
+	GLuint composite_uniform_crt_sampler;
+	GLuint composite_uniform_bloom_sampler;
+
+	// Bloom settings
+	float bloom_threshold;
+	float bloom_strength;
+	uint32_t bloom_width;
+	uint32_t bloom_height;
+
+	// Phosphor persistence
+	float persistence_decay;
+
+	// Rendering & Dynamic Resolution
+	uint32_t render_width;	// The actual remake resolution (e.g., 360)
+	uint32_t render_height;	// The actual remake resolution (e.g., 270)
+	uint32_t frame_number;
+	uint8_t running;
+	uint8_t toggle_crt_emulation;
+	uint8_t toggle_bloom;
+	uint8_t fullscreen;
+	uint8_t viewport_changed;  // Flag to signal render thread to recreate FBOs
+};
+
+// static struct remake_state state;