optimize ppusim with threads

7 files changed, 67 insertions, 25 deletions

diff --git a/src/demo.c b/src/demo.c
index 6248989..019750a 100644
--- a/src/demo.c
+++ b/src/demo.c
@@ -12,7 +12,7 @@ void hh_demo_setup() {
 	hh_ppu_update_sprite(1, HH_DBG_SPRITE_CHECKERBOARD);
 
 	// background pattern
-	hh_ppu_update_color(0, 1, (hh_ppu_rgb_color_t) {0x8, 0x8, 0x8});
+	hh_ppu_update_color(0, 1, (hh_ppu_rgb_color_t) {0x4, 0x4, 0x4});
 	for (unsigned i = 0; i < HH_PPU_BG_CANVAS_TILES_H * HH_PPU_BG_CANVAS_TILES_V; i++) {
 		hh_ppu_update_background(i, (hh_s_ppu_loc_bam_entry) {
 			.horizontal_flip = false,
@@ -48,8 +48,8 @@ void hh_demo_loop(unsigned long frame) {
 	});
 
 	for (unsigned i = 0; i < HH_DEMO_BALL_COUNT; i++) {
-		g_hh_demo_balls[i].position_x = HH_PPU_SCREEN_WIDTH/2  - HH_PPU_SPRITE_WIDTH/2  + (int)(60 * (double)sin((1*(double)frame + 6*(double)i) / 30));
-		g_hh_demo_balls[i].position_y = HH_PPU_SCREEN_HEIGHT/2 - HH_PPU_SPRITE_HEIGHT/2 + (int)(30 * (double)sin((2*(double)frame + 6*(double)i) / 30));
+		g_hh_demo_balls[i].position_x = HH_PPU_SCREEN_WIDTH/2  - HH_PPU_SPRITE_WIDTH/2  + (int)(60 * (double)sin((1*(double)frame / 10) + (double)i * 12));
+		g_hh_demo_balls[i].position_y = HH_PPU_SCREEN_HEIGHT/2 - HH_PPU_SPRITE_HEIGHT/2 + (int)(30 * (double)sin((2*(double)frame / 10) + (double)i * 12));
 		hh_ppu_update_foreground(i, g_hh_demo_balls[i]);
 	}
 }
diff --git a/src/ds.mk b/src/ds.mk
index 18107ad..f70c1c9 100644
--- a/src/ds.mk
+++ b/src/ds.mk
@@ -5,5 +5,6 @@ LFLAGS += -lSDL2
 
 DESKTOP_SRCS += ppusim/sim.c \
 								ppusim/mem.c \
-								ppusim/pixel.c
+								ppusim/pixel.c \
+								ppusim/work.c
 
diff --git a/src/ppusim/pixel.c b/src/ppusim/pixel.c
index 2163c10..6e915e6 100644
--- a/src/ppusim/pixel.c
+++ b/src/ppusim/pixel.c
@@ -1,21 +1,11 @@
 #include <SDL2/SDL.h>
 #include <stdbool.h>
 
+#include "ppusim/work.h"
 #include "ppu/internals.h"
 #include "ppusim/mem.h"
-#include "ppusim/sim.h"
 #include "ppu/consts.h"
 
-static void hh_ppusim_draw(SDL_Renderer* r, unsigned px, unsigned py, unsigned cr, unsigned cg, unsigned cb) {
-	SDL_SetRenderDrawColor(r, cr, cg, cb, 255);
-	SDL_RenderFillRect(r, &(SDL_Rect) {
-		.x = px * HH_PPUSIM_UPSCALE_FACTOR,
-		.y = py * HH_PPUSIM_UPSCALE_FACTOR,
-		.w = HH_PPUSIM_UPSCALE_FACTOR,
-		.h = HH_PPUSIM_UPSCALE_FACTOR
-	});
-}
-
 /* transform xy if tile is flipped */
 static uint16_t hh_ppusim_apply_transform(unsigned x, unsigned y, bool fliph, bool flipv) {
 	unsigned tx = fliph ? HH_PPU_SPRITE_WIDTH - x - 1 : x;
@@ -77,7 +67,7 @@ static uint8_t hh_ppusim_fg_pixel(unsigned x, unsigned y) {
 	return cidx;
 }
 
-void hh_ppusim_pixel(SDL_Renderer* r, unsigned x, unsigned y) {
+void hh_ppusim_pixel(uint8_t* s, unsigned x, unsigned y) {
 	uint8_t bg_cidx = hh_ppusim_bg_pixel(x, y);
 	uint8_t fg_cidx = hh_ppusim_fg_pixel(x, y);
 	uint8_t cidx = (fg_cidx & HH_MASK(3)) == 0 ? bg_cidx : fg_cidx;
@@ -87,6 +77,9 @@ void hh_ppusim_pixel(SDL_Renderer* r, unsigned x, unsigned y) {
 		HH_RESIZE(pal_rgb, 7, 4),
 		HH_RESIZE(pal_rgb, 3, 0)
 	};
-	hh_ppusim_draw(r, x, y, rgb[0] << 4, rgb[1] << 4, rgb[2] << 4);
+
+	s[0] = rgb[0] << 4;
+	s[1] = rgb[1] << 4;
+	s[2] = rgb[2] << 4;
 }
 
diff --git a/src/ppusim/pixel.h b/src/ppusim/pixel.h
index 8044560..475fe11 100644
--- a/src/ppusim/pixel.h
+++ b/src/ppusim/pixel.h
@@ -1,6 +1,6 @@
 #pragma once
 
-#include <SDL2/SDL.h>
+#include "ppusim/work.h"
 
-void hh_ppusim_pixel(SDL_Renderer*, unsigned x, unsigned y);
+void hh_ppusim_pixel(uint8_t* s, unsigned x, unsigned y);
 
diff --git a/src/ppusim/sim.c b/src/ppusim/sim.c
index 8f01359..4226a9b 100644
--- a/src/ppusim/sim.c
+++ b/src/ppusim/sim.c
@@ -7,7 +7,7 @@
 #include "ppu/ppu.h"
 #include "ppusim/mem.h"
 #include "ppusim/sim.h"
-#include "ppusim/pixel.h"
+#include "ppusim/work.h"
 
 SDL_Window *g_hh_window = NULL;
 SDL_Renderer *g_hh_renderer = NULL;
@@ -38,11 +38,7 @@ void hh_loop() {
 		hh_ppu_vblank_interrupt();
 
 		SDL_RenderClear(g_hh_renderer);
-
-		for (unsigned x = 0; x < HH_PPU_SCREEN_WIDTH; x++)
-			for (unsigned y = 0; y < HH_PPU_SCREEN_HEIGHT; y++)
-				hh_ppusim_pixel(g_hh_renderer, x, y);
-
+		hh_ppusim_draw_frame(g_hh_renderer);
 		SDL_SetRenderDrawColor(g_hh_renderer, 0, 0, 0, 255);
 		SDL_RenderPresent(g_hh_renderer);
 
diff --git a/src/ppusim/work.c b/src/ppusim/work.c
new file mode 100644
index 0000000..2073907
--- /dev/null
+++ b/src/ppusim/work.c
@@ -0,0 +1,38 @@
+#include <SDL2/SDL.h>
+#include <pthread.h>
+#include <stdio.h>
+
+#include "ppusim/work.h"
+#include "ppusim/sim.h"
+#include "ppu/consts.h"
+#include "ppusim/pixel.h"
+
+pthread_t g_hh_ppusim_scanline_threads[HH_PPU_SCREEN_HEIGHT];
+hh_s_ppusim_screen g_hh_ppusim_screen;
+
+void* hh_ppusim_draw_scanline(void* scanline_context) {
+	unsigned scanline = (unsigned long) scanline_context;
+	for (unsigned x = 0; x < HH_PPU_SCREEN_WIDTH; x++)
+		hh_ppusim_pixel(g_hh_ppusim_screen[scanline][x], x, scanline);
+	return NULL;
+}
+
+void hh_ppusim_draw_frame(SDL_Renderer* renderer) {
+	for (unsigned y = 0; y < HH_PPU_SCREEN_HEIGHT; y++)
+		pthread_create(&g_hh_ppusim_scanline_threads[y], NULL, hh_ppusim_draw_scanline, (void*)(unsigned long)y);
+	for (unsigned y = 0; y < HH_PPU_SCREEN_HEIGHT; y++)
+		pthread_join(g_hh_ppusim_scanline_threads[y], NULL);
+	
+	for (unsigned x = 0; x < HH_PPU_SCREEN_WIDTH; x++) {
+		for (unsigned y = 0; y < HH_PPU_SCREEN_HEIGHT; y++) {
+			SDL_SetRenderDrawColor(renderer, g_hh_ppusim_screen[y][x][0], g_hh_ppusim_screen[y][x][1], g_hh_ppusim_screen[y][x][2], 255);
+			SDL_RenderFillRect(renderer, &(SDL_Rect) {
+				.x = x * HH_PPUSIM_UPSCALE_FACTOR,
+				.y = y * HH_PPUSIM_UPSCALE_FACTOR,
+				.w = HH_PPUSIM_UPSCALE_FACTOR,
+				.h = HH_PPUSIM_UPSCALE_FACTOR,
+			});
+		}
+	}
+}
+
diff --git a/src/ppusim/work.h b/src/ppusim/work.h
new file mode 100644
index 0000000..1171a22
--- /dev/null
+++ b/src/ppusim/work.h
@@ -0,0 +1,14 @@
+#pragma once
+
+#include <SDL2/SDL.h>
+
+#include "ppu/consts.h"
+
+typedef uint8_t hh_s_ppusim_color[3];
+typedef hh_s_ppusim_color hh_s_ppusim_scanline[HH_PPU_SCREEN_WIDTH];
+typedef hh_s_ppusim_scanline hh_s_ppusim_screen[HH_PPU_SCREEN_HEIGHT];
+
+void* hh_ppusim_draw_scanline(void*);
+
+void hh_ppusim_draw_frame(SDL_Renderer*);
+