Merged OpenGL branch

7 files changed, 347 insertions, 133 deletions

diff --git a/Makefile b/Makefile
index 40c070a..b403988 100644
--- a/Makefile
+++ b/Makefile
@@ -1,4 +1,4 @@
-LIBS=sdl
+LIBS=sdl glew gl
 LDFLAGS=-lm `pkg-config --libs $(LIBS)`
 ifdef DEBUG
 CFLAGS=-ggdb -std=gnu99 `pkg-config --cflags-only-I $(LIBS)` -Wreturn-type -Werror=return-type
diff --git a/blastem.c b/blastem.c
index b57952e..85f92d3 100644
--- a/blastem.c
+++ b/blastem.c
@@ -278,9 +278,9 @@ m68k_context * sync_components(m68k_context * context, uint32_t address)
 	adjust_int_cycle(context, v_context);
 	if (address) {
 		if (break_on_sync) {
-			break_on_sync = 0;
-			debugger(context, address);
-		}
+		break_on_sync = 0;
+		debugger(context, address);
+	}
 		if (save_state) {
 			save_state = 0;
 			while (!z_context->pc)
@@ -1507,7 +1507,7 @@ void set_speed_percent(genesis_context * context, uint32_t percent)
 	context->master_clock = ((uint64_t)context->normal_clock * (uint64_t)percent) / 100;
 	while (context->ym->current_cycle != context->psg->cycles) {
 		sync_sound(context, context->psg->cycles + MCLKS_PER_PSG);
-	}
+}
 	ym_adjust_master_clock(context->ym, context->master_clock);
 	psg_adjust_master_clock(context->psg, context->master_clock);
 }
@@ -1773,7 +1773,7 @@ int main(int argc, char ** argv)
 	char * romfname = NULL;
 	FILE *address_log = NULL;
 	char * statefile = NULL;
-	uint8_t fullscreen = 0;
+	uint8_t fullscreen = 0, use_gl = 0;
 	for (int i = 1; i < argc; i++) {
 		if (argv[i][0] == '-') {
 			switch(argv[i][1]) {
@@ -1783,6 +1783,9 @@ int main(int argc, char ** argv)
 			case 'f':
 				fullscreen = 1;
 				break;
+			case 'g':
+				use_gl = 1;
+				break;
 			case 'l':
 				address_log = fopen("address.log", "w");
 				break;
@@ -1885,7 +1888,7 @@ int main(int argc, char ** argv)
 		fps = 50;
 	}
 	if (!headless) {
-		render_init(width, height, title, fps, fullscreen);
+		render_init(width, height, title, fps, fullscreen, use_gl);
 	}
 	vdp_context v_context;
 	genesis_context gen;
diff --git a/default.f.glsl b/default.f.glsl
new file mode 100644
index 0000000..f7bb89a
--- /dev/null
+++ b/default.f.glsl
@@ -0,0 +1,14 @@
+#version 110
+
+uniform sampler2D textures[2];
+
+varying vec2 texcoord;
+
+void main()
+{
+	gl_FragColor = mix(
+		texture2D(textures[0], texcoord),
+		texture2D(textures[1], vec2(texcoord.x, texcoord.y - 1.0/512.0)),
+		sin((texcoord.y * 512.0 - 0.75) * 3.14159265359) / 2.0 + 0.5
+	);
+}
diff --git a/default.v.glsl b/default.v.glsl
new file mode 100644
index 0000000..367c128
--- /dev/null
+++ b/default.v.glsl
@@ -0,0 +1,10 @@
+#version 110
+
+attribute vec2 pos;
+varying vec2 texcoord;
+
+void main()
+{
+	gl_Position = vec4(pos, 0.0, 1.0);
+	texcoord = pos * vec2(320.0/1024.0, 240.0/-512.0) + vec2(320.0/1024.0, 240.0/512.0);
+}
diff --git a/render.h b/render.h
index acca888..b5dbeda 100644
--- a/render.h
+++ b/render.h
@@ -1,6 +1,6 @@
 /*
  Copyright 2013 Michael Pavone
- This file is part of BlastEm. 
+ This file is part of BlastEm.
  BlastEm is free software distributed under the terms of the GNU General Public License version 3 or greater. See COPYING for full license text.
 */
 #ifndef RENDER_H_
@@ -9,9 +9,17 @@
 #include "vdp.h"
 #include "psg.h"
 #include "ym2612.h"
+
+typedef struct {
+	void *oddbuf;
+	void *evenbuf;
+	int  stride;
+} surface_info;
+
 uint32_t render_map_color(uint8_t r, uint8_t g, uint8_t b);
+void render_alloc_surfaces(vdp_context * context);
 uint8_t render_depth();
-void render_init(int width, int height, char * title, uint32_t fps, uint8_t fullscreen);
+void render_init(int width, int height, char * title, uint32_t fps, uint8_t fullscreen, uint8_t use_gl);
 void render_context(vdp_context * context);
 void render_wait_quit(vdp_context * context);
 void render_wait_psg(psg_context * context);
diff --git a/render_sdl.c b/render_sdl.c
index 3cbc604..3c01eeb 100644
--- a/render_sdl.c
+++ b/render_sdl.c
@@ -1,6 +1,6 @@
 /*
  Copyright 2013 Michael Pavone
- This file is part of BlastEm. 
+ This file is part of BlastEm.
  BlastEm is free software distributed under the terms of the GNU General Public License version 3 or greater. See COPYING for full license text.
 */
 #include <stdlib.h>
@@ -9,9 +9,14 @@
 #include "blastem.h"
 #include "io.h"
 
+#ifndef DISABLE_OPENGL
+#include <GL/glew.h>
+#endif
+
 SDL_Surface *screen;
 uint8_t render_dbg = 0;
 uint8_t debug_pal = 0;
+uint8_t render_gl;
 
 uint32_t last_frame = 0;
 
@@ -85,7 +90,104 @@ int render_num_joysticks()
 
 uint32_t render_map_color(uint8_t r, uint8_t g, uint8_t b)
 {
-	return SDL_MapRGB(screen->format, r, g, b);
+	if (render_gl) {
+		return 255 << 24 | r << 16 | g << 8 | b;
+	} else {
+		return SDL_MapRGB(screen->format, r, g, b);
+	}
+}
+
+GLuint textures[3], buffers[2], vshader, fshader, program, un_textures[2], at_pos;
+
+const GLfloat vertex_data[] = {
+	-1.0f, -1.0f,
+	 1.0f, -1.0f,
+	-1.0f,  1.0f,
+	 1.0f,  1.0f
+};
+
+const GLushort element_data[] = {0, 1, 2, 3};
+
+GLuint load_shader(char * fname, GLenum shader_type)
+{
+	FILE * f = fopen(fname, "r");
+	if (!f) {
+		fprintf(stderr, "Failed to open shader file %s for reading\n", fname);
+		return 0;
+	}
+	fseek(f, 0, SEEK_END);
+	long fsize = ftell(f);
+	fseek(f, 0, SEEK_SET);
+	GLchar * text = malloc(fsize);
+	if (fread(text, 1, fsize, f) != fsize) {
+		fprintf(stderr, "Error reading from shader file %s\n", fname);
+		free(text);
+		return 0;
+	}
+	GLuint ret = glCreateShader(shader_type);
+	glShaderSource(ret, 1, (const GLchar **)&text, (const GLint *)&fsize);
+	free(text);
+	glCompileShader(ret);
+	GLint compile_status, loglen;
+	glGetShaderiv(ret, GL_COMPILE_STATUS, &compile_status);
+	if (!compile_status) {
+		fprintf(stderr, "Shader %s failed to compile\n", fname);
+		glGetShaderiv(ret, GL_INFO_LOG_LENGTH, &loglen);
+		text = malloc(loglen);
+		glGetShaderInfoLog(ret, loglen, NULL, text);
+		fputs(text, stderr);
+		free(text);
+		glDeleteShader(ret);
+		return 0;
+	}
+	return ret;
+}
+
+void render_alloc_surfaces(vdp_context * context)
+{
+	if (render_gl) {
+		context->oddbuf = context->framebuf = malloc(320 * 240 * 4 * 2);
+		memset(context->oddbuf, 0, 320 * 240 * 4 * 2);
+		context->evenbuf = ((char *)context->oddbuf) + 320 * 240 * 4;
+		glGenTextures(3, textures);
+		for (int i = 0; i < 3; i++)
+		{
+			glBindTexture(GL_TEXTURE_2D, textures[i]);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+			if (i < 2) {
+				glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 512, 256, 0, GL_BGRA, GL_UNSIGNED_BYTE, i ? context->evenbuf : context->oddbuf);
+			} else {
+				uint32_t blank = 255 << 24;
+				glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 1, 1, 0, GL_BGRA, GL_UNSIGNED_BYTE, &blank);
+			}
+		}
+		glGenBuffers(2, buffers);
+		glBindBuffer(GL_ARRAY_BUFFER, buffers[0]);
+		glBufferData(GL_ARRAY_BUFFER, sizeof(vertex_data), vertex_data, GL_STATIC_DRAW);
+		glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, buffers[1]);
+		glBufferData(GL_ELEMENT_ARRAY_BUFFER, sizeof(element_data), element_data, GL_STATIC_DRAW);
+		vshader = load_shader("default.v.glsl", GL_VERTEX_SHADER);
+		fshader = load_shader("default.f.glsl", GL_FRAGMENT_SHADER);
+		program = glCreateProgram();
+		glAttachShader(program, vshader);
+		glAttachShader(program, fshader);
+		glLinkProgram(program);
+		GLint link_status;
+		glGetProgramiv(program, GL_LINK_STATUS, &link_status);
+		if (!link_status) {
+			fputs("Failed to link shader program\n", stderr);
+			exit(1);
+		}
+		un_textures[0] = glGetUniformLocation(program, "textures[0]");
+		un_textures[1] = glGetUniformLocation(program, "textures[1]");
+		at_pos = glGetAttribLocation(program, "pos");
+	} else {
+		context->oddbuf = context->framebuf = malloc(320 * 240 * screen->format->BytesPerPixel * 2);
+		context->evenbuf = ((char *)context->oddbuf) + 320 * 240 * screen->format->BytesPerPixel;
+	}
 }
 
 uint8_t render_depth()
@@ -95,62 +197,98 @@ uint8_t render_depth()
 
 char * caption = NULL;
 
-void render_init(int width, int height, char * title, uint32_t fps, uint8_t fullscreen)
+void render_init(int width, int height, char * title, uint32_t fps, uint8_t fullscreen, uint8_t use_gl)
 {
 	if (SDL_Init(SDL_INIT_VIDEO | SDL_INIT_AUDIO | SDL_INIT_JOYSTICK) < 0) {
-        fprintf(stderr, "Unable to init SDL: %s\n", SDL_GetError());
-        exit(1);
-    }
-    atexit(SDL_Quit);
-    atexit(render_close_audio);
-    printf("width: %d, height: %d\n", width, height);
-    uint32_t flags = SDL_ANYFORMAT;
-    if (fullscreen) {
-    	flags |= SDL_FULLSCREEN | SDL_HWSURFACE | SDL_DOUBLEBUF;
-    } else {
-    	flags |= SDL_SWSURFACE;
-    }
-    screen = SDL_SetVideoMode(width, height, 32, flags);
-    if (!screen) {
-    	fprintf(stderr, "Unable to get SDL surface: %s\n", SDL_GetError());
-        exit(1);
-    }
-    if (screen->format->BytesPerPixel != 2 && screen->format->BytesPerPixel != 4) {
-    	fprintf(stderr, "BlastEm requires a 16-bit or 32-bit surface, SDL returned a %d-bit surface\n", screen->format->BytesPerPixel * 8);
-    	exit(1);
-    }
-    SDL_WM_SetCaption(title, title);
+		fprintf(stderr, "Unable to init SDL: %s\n", SDL_GetError());
+		exit(1);
+	}
+	atexit(SDL_Quit);
+	atexit(render_close_audio);
+	printf("width: %d, height: %d\n", width, height);
+	uint32_t flags = SDL_ANYFORMAT;
+
+#ifndef DISABLE_OPENGL
+	if (use_gl)
+	{
+		SDL_GL_SetAttribute(SDL_GL_RED_SIZE, 5);
+		SDL_GL_SetAttribute(SDL_GL_GREEN_SIZE, 5);
+		SDL_GL_SetAttribute(SDL_GL_BLUE_SIZE, 5);
+		SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
+		SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
+		flags = SDL_OPENGL;
+		if (fullscreen) {
+			flags |= SDL_FULLSCREEN;
+		}
+	} else {
+#else
+	{
+#endif
+		if (fullscreen) {
+			flags |= SDL_FULLSCREEN | SDL_HWSURFACE | SDL_DOUBLEBUF;
+		} else {
+			flags |= SDL_SWSURFACE;
+		}
+	}
+	screen = SDL_SetVideoMode(width, height, 32, flags);
+	if (!screen) {
+		fprintf(stderr, "Unable to get SDL surface: %s\n", SDL_GetError());
+		exit(1);
+	}
+	if (!use_gl && screen->format->BytesPerPixel != 2 && screen->format->BytesPerPixel != 4) {
+		fprintf(stderr, "BlastEm requires a 16-bit or 32-bit surface, SDL returned a %d-bit surface\n", screen->format->BytesPerPixel * 8);
+		exit(1);
+	}
+#ifndef DISABLE_OPENGL
+	//TODO: fallback on standard rendering if OpenGL 2.0 is unavailable or if init fails
+	if (use_gl)
+	{
+		GLenum res = glewInit();
+		if (res != GLEW_OK)
+		{
+			fprintf(stderr, "Initialization of GLEW failed with code %d\n", res);
+			exit(1);
+		}
+		if (!GLEW_VERSION_2_0)
+		{
+			fputs("OpenGL 2.0 is unable, falling back to standard SDL rendering\n", stderr);
+			exit(1);
+		}
+	}
+	render_gl = use_gl;
+#endif
+	SDL_WM_SetCaption(title, title);
 	caption = title;
-    min_delay = 0;
-    for (int i = 0; i < 100; i++) {
-    	uint32_t start = SDL_GetTicks();
-    	SDL_Delay(1);
-    	uint32_t delay = SDL_GetTicks()-start;
-    	if (delay > min_delay) {
-    		min_delay = delay;
-    	}
-    }
-    if (!min_delay) {
-    	min_delay = 1;
-    }
-    printf("minimum delay: %d\n", min_delay);
-    
-    frame_delay = 1000/fps;
-    
-    audio_mutex = SDL_CreateMutex();
-    psg_cond = SDL_CreateCond();
-    ym_cond = SDL_CreateCond();
-    audio_ready = SDL_CreateCond();
-    
-    SDL_AudioSpec desired, actual;
+	min_delay = 0;
+	for (int i = 0; i < 100; i++) {
+		uint32_t start = SDL_GetTicks();
+		SDL_Delay(1);
+		uint32_t delay = SDL_GetTicks()-start;
+		if (delay > min_delay) {
+			min_delay = delay;
+		}
+	}
+	if (!min_delay) {
+		min_delay = 1;
+	}
+	printf("minimum delay: %d\n", min_delay);
+
+	frame_delay = 1000/fps;
+
+	audio_mutex = SDL_CreateMutex();
+	psg_cond = SDL_CreateCond();
+	ym_cond = SDL_CreateCond();
+	audio_ready = SDL_CreateCond();
+
+	SDL_AudioSpec desired, actual;
     char * rate_str = tern_find_ptr(config, "audiorate");
    	int rate = rate_str ? atoi(rate_str) : 0;
    	if (!rate) {
    		rate = 48000;
    	}
     desired.freq = rate;
-    desired.format = AUDIO_S16SYS;
-    desired.channels = 2;
+	desired.format = AUDIO_S16SYS;
+	desired.channels = 2;
     char * samples_str = tern_find_ptr(config, "audiobuffer");
    	int samples = samples_str ? atoi(samples_str) : 0;
    	if (!samples) {
@@ -158,29 +296,62 @@ void render_init(int width, int height, char * title, uint32_t fps, uint8_t full
    	}
     printf("config says: %d\n", samples);
     desired.samples = samples*2;
-    desired.callback = audio_callback;
-    desired.userdata = NULL;
-    
-    if (SDL_OpenAudio(&desired, &actual) < 0) {
-    	fprintf(stderr, "Unable to open SDL audio: %s\n", SDL_GetError());
-    	exit(1);
-    }
-    buffer_samples = actual.samples;
-    sample_rate = actual.freq;
-    printf("Initialized audio at frequency %d with a %d sample buffer\n", actual.freq, actual.samples);
-    SDL_PauseAudio(0);
-    num_joysticks = SDL_NumJoysticks();
-    if (num_joysticks > MAX_JOYSTICKS) {
-    	num_joysticks = MAX_JOYSTICKS;
-    }
-    for (int i = 0; i < num_joysticks; i++) {
-    	printf("Joystick %d: %s\n", i, SDL_JoystickName(i));
-    	SDL_Joystick * joy = joysticks[i] = SDL_JoystickOpen(i);
-    	if (joy) {
-    		printf("\tNum Axes: %d\n\tNum Buttons: %d\n\tNum Hats: %d\n", SDL_JoystickNumAxes(joy), SDL_JoystickNumButtons(joy), SDL_JoystickNumHats(joy));
-    	}
-    }
-    SDL_JoystickEventState(SDL_ENABLE);
+	desired.callback = audio_callback;
+	desired.userdata = NULL;
+
+	if (SDL_OpenAudio(&desired, &actual) < 0) {
+		fprintf(stderr, "Unable to open SDL audio: %s\n", SDL_GetError());
+		exit(1);
+	}
+	buffer_samples = actual.samples;
+	sample_rate = actual.freq;
+	printf("Initialized audio at frequency %d with a %d sample buffer\n", actual.freq, actual.samples);
+	SDL_PauseAudio(0);
+	num_joysticks = SDL_NumJoysticks();
+	if (num_joysticks > MAX_JOYSTICKS) {
+		num_joysticks = MAX_JOYSTICKS;
+	}
+	for (int i = 0; i < num_joysticks; i++) {
+		printf("Joystick %d: %s\n", i, SDL_JoystickName(i));
+		SDL_Joystick * joy = joysticks[i] = SDL_JoystickOpen(i);
+		if (joy) {
+			printf("\tNum Axes: %d\n\tNum Buttons: %d\n\tNum Hats: %d\n", SDL_JoystickNumAxes(joy), SDL_JoystickNumButtons(joy), SDL_JoystickNumHats(joy));
+		}
+	}
+	SDL_JoystickEventState(SDL_ENABLE);
+}
+
+void render_context_gl(vdp_context * context)
+{
+	glBindTexture(GL_TEXTURE_2D, textures[context->framebuf == context->oddbuf ? 0 : 1]);
+	glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 320, 240, GL_BGRA, GL_UNSIGNED_BYTE, context->framebuf);;
+
+	glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
+	glClear(GL_COLOR_BUFFER_BIT);
+
+	glUseProgram(program);
+	glActiveTexture(GL_TEXTURE0);
+	glBindTexture(GL_TEXTURE_2D, textures[0]);
+	glUniform1i(un_textures[0], 0);
+
+	glActiveTexture(GL_TEXTURE1);
+	glBindTexture(GL_TEXTURE_2D, (context->regs[REG_MODE_4] & BIT_INTERLACE) ? textures[1] : textures[2]);
+	glUniform1i(un_textures[1], 1);
+
+	glBindBuffer(GL_ARRAY_BUFFER, buffers[0]);
+	glVertexAttribPointer(at_pos, 2, GL_FLOAT, GL_FALSE, sizeof(GLfloat[2]), (void *)0);
+	glEnableVertexAttribArray(at_pos);
+
+	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, buffers[1]);
+	glDrawElements(GL_TRIANGLE_STRIP, 4, GL_UNSIGNED_SHORT, (void *)0);
+
+	glDisableVertexAttribArray(at_pos);
+
+	SDL_GL_SwapBuffers();
+	if (context->regs[REG_MODE_4] & BIT_INTERLACE)
+	{
+		context->framebuf = context->framebuf == context->oddbuf ? context->evenbuf : context->oddbuf;
+	}
 }
 
 uint32_t blankbuf[320*240];
@@ -188,65 +359,70 @@ uint32_t blankbuf[320*240];
 void render_context(vdp_context * context)
 {
 	uint16_t *buf_16;
-	uint32_t *buf_32; 
+	uint32_t *buf_32;
 	uint8_t b,g,r;
 	last_frame = SDL_GetTicks();
+	if (render_gl)
+	{
+		render_context_gl(context);
+		return;
+	}
 	if (SDL_MUSTLOCK(screen)) {
 		if (SDL_LockSurface(screen) < 0) {
 			return;
 		}
-    }
-    uint16_t repeat_x = screen->clip_rect.w / 320;
-    uint16_t repeat_y = screen->clip_rect.h / 240;
-    if (repeat_x > repeat_y) {
-    	repeat_x = repeat_y;
-    } else {
-    	repeat_y = repeat_x;
-    }
-    int othermask = repeat_y >> 1;
-    
-    if (screen->format->BytesPerPixel == 2) {
-    	uint16_t *otherbuf = (context->regs[REG_MODE_4] & BIT_INTERLACE) ? context->evenbuf : (uint16_t *)blankbuf;
-    	uint16_t * oddbuf = context->oddbuf;
-    	buf_16 = (uint16_t *)screen->pixels;
-    	for (int y = 0; y < 240; y++) {
-    		for (int i = 0; i < repeat_y; i++,buf_16 += screen->pitch/2) {
-        		uint16_t *line = buf_16;
-        		uint16_t *src_line = (i & othermask ? otherbuf : oddbuf) + y * 320;
-		    	for (int x = 0; x < 320; x++) {
-		    		uint16_t color = *(src_line++);
-		    		for (int j = 0; j < repeat_x; j++) {
-		    			*(line++) = color;
-		    		}
-		    	}
-		    }
-    	}
-    } else {
-    	uint32_t *otherbuf = (context->regs[REG_MODE_4] & BIT_INTERLACE) ? context->evenbuf : (uint32_t *)blankbuf;
-    	uint32_t * oddbuf = context->oddbuf;
-    	buf_32 = (uint32_t *)screen->pixels;
-    	for (int y = 0; y < 240; y++) {
-    		for (int i = 0; i < repeat_y; i++,buf_32 += screen->pitch/4) {
-        		uint32_t *line = buf_32;
-        		uint32_t *src_line = (i & othermask ? otherbuf : oddbuf) + y * 320;
-		    	for (int x = 0; x < 320; x++) {
-		    		uint32_t color = *(src_line++);
-		    		for (int j = 0; j < repeat_x; j++) {
-		    			*(line++) = color;
-		    		}
-		    	}
-		    }
-    	}
-    }
-    if ( SDL_MUSTLOCK(screen) ) {
-        SDL_UnlockSurface(screen);
-    }
+	}
+	uint16_t repeat_x = screen->clip_rect.w / 320;
+	uint16_t repeat_y = screen->clip_rect.h / 240;
+	if (repeat_x > repeat_y) {
+		repeat_x = repeat_y;
+	} else {
+		repeat_y = repeat_x;
+	}
+	int othermask = repeat_y >> 1;
+
+	if (screen->format->BytesPerPixel == 2) {
+		uint16_t *otherbuf = (context->regs[REG_MODE_4] & BIT_INTERLACE) ? context->evenbuf : (uint16_t *)blankbuf;
+		uint16_t * oddbuf = context->oddbuf;
+		buf_16 = (uint16_t *)screen->pixels;
+		for (int y = 0; y < 240; y++) {
+			for (int i = 0; i < repeat_y; i++,buf_16 += screen->pitch/2) {
+				uint16_t *line = buf_16;
+				uint16_t *src_line = (i & othermask ? otherbuf : oddbuf) + y * 320;
+				for (int x = 0; x < 320; x++) {
+					uint16_t color = *(src_line++);
+					for (int j = 0; j < repeat_x; j++) {
+						*(line++) = color;
+					}
+				}
+			}
+		}
+	} else {
+		uint32_t *otherbuf = (context->regs[REG_MODE_4] & BIT_INTERLACE) ? context->evenbuf : (uint32_t *)blankbuf;
+		uint32_t * oddbuf = context->oddbuf;
+		buf_32 = (uint32_t *)screen->pixels;
+		for (int y = 0; y < 240; y++) {
+			for (int i = 0; i < repeat_y; i++,buf_32 += screen->pitch/4) {
+				uint32_t *line = buf_32;
+				uint32_t *src_line = (i & othermask ? otherbuf : oddbuf) + y * 320;
+				for (int x = 0; x < 320; x++) {
+					uint32_t color = *(src_line++);
+					for (int j = 0; j < repeat_x; j++) {
+						*(line++) = color;
+					}
+				}
+			}
+		}
+	}
+	if ( SDL_MUSTLOCK(screen) ) {
+		SDL_UnlockSurface(screen);
+	}
     //SDL_UpdateRect(screen, 0, 0, screen->clip_rect.w, screen->clip_rect.h);
     SDL_Flip(screen);
-    if (context->regs[REG_MODE_4] & BIT_INTERLACE)
-    {
-    	context->framebuf = context->framebuf == context->oddbuf ? context->evenbuf : context->oddbuf;
-    }
+	if (context->regs[REG_MODE_4] & BIT_INTERLACE)
+	{
+		context->framebuf = context->framebuf == context->oddbuf ? context->evenbuf : context->oddbuf;
+	}
 }
 
 int render_joystick_num_buttons(int joystick)
diff --git a/vdp.c b/vdp.c
index c1e04f0..e962a67 100644
--- a/vdp.c
+++ b/vdp.c
@@ -50,10 +50,13 @@ void init_vdp_context(vdp_context * context)
 	memset(context, 0, sizeof(*context));
 	context->vdpmem = malloc(VRAM_SIZE);
 	memset(context->vdpmem, 0, VRAM_SIZE);
-	context->oddbuf = context->framebuf = malloc(FRAMEBUF_ENTRIES * (render_depth() / 8));
+	/*context->oddbuf = context->framebuf = malloc(FRAMEBUF_ENTRIES * (render_depth() / 8));
 	memset(context->framebuf, 0, FRAMEBUF_ENTRIES * (render_depth() / 8));
 	context->evenbuf = malloc(FRAMEBUF_ENTRIES * (render_depth() / 8));
 	memset(context->evenbuf, 0, FRAMEBUF_ENTRIES * (render_depth() / 8));
+	*/
+	render_alloc_surfaces(context);
+	context->framebuf = context->oddbuf;
 	context->linebuf = malloc(LINEBUF_SIZE + SCROLL_BUFFER_SIZE*2);
 	memset(context->linebuf, 0, LINEBUF_SIZE + SCROLL_BUFFER_SIZE*2);
 	context->tmp_buf_a = context->linebuf + LINEBUF_SIZE;