wimmel_gl: Use a second texture unit for search region

Speeds up calulation from ~11s to ~10s.

2 files changed, 43 insertions, 8 deletions

diff --git a/wimmel.frag b/wimmel.frag
index b148852..42b0139 100644
--- a/wimmel.frag
+++ b/wimmel.frag
@@ -1,5 +1,6 @@
 varying vec2 v_texcoord;
 uniform sampler2D tex;
+uniform sampler2D tex2;
 uniform vec2 pixelsize;
 /* The rect to match against */
 uniform vec4 srect;
@@ -13,8 +14,8 @@ main()
 
 	for (i = 0.0; i < srect[2] && diff < barrier; i += pixelsize.x) {
 		for (j = 0.0; j < srect[3] && diff < barrier; j += pixelsize.y) {
-			vec4 match = texture2D(tex, srect.xy + vec2(i,j));
 			vec4 orig = texture2D(tex, v_texcoord + vec2(i,j));
+			vec4 match = texture2D(tex2, vec2(i, j) / srect.zw);
 
 			diff += int(any(greaterThan(abs(orig - match), vec4(0.2))));
 		}
diff --git a/wimmel_gl.c b/wimmel_gl.c
index 0432823..2d33edb 100644
--- a/wimmel_gl.c
+++ b/wimmel_gl.c
@@ -16,9 +16,10 @@
 
 EGLDisplay dpy;
 GLuint vbo;
-GLint proj_uniform, tex_uniform, pixelsize_uniform,
+GLint proj_uniform, tex_uniform, tex2_uniform, pixelsize_uniform,
       search_rect_uniform, barrier_uniform;
 GLuint texture;
+GLuint texture2;
 int width, height;
 GLuint program;
 GdkPixbuf *pixbuf;
@@ -78,6 +79,7 @@ init_shaders()
 	glUseProgram(program);
 	proj_uniform = glGetUniformLocation(program, "proj");
 	tex_uniform = glGetUniformLocation(program, "tex");
+	tex2_uniform = glGetUniformLocation(program, "tex2");
 	pixelsize_uniform = glGetUniformLocation(program, "pixelsize");
 	search_rect_uniform = glGetUniformLocation(program, "srect");
 	barrier_uniform = glGetUniformLocation(program, "barrier");
@@ -113,14 +115,22 @@ init_shaders()
 	return 0;
 }
 
+static inline int
+align(int value, int alignment)
+{
+   return (value + alignment - 1) & ~(alignment - 1);
+}
+
 static void
-create_texture(GdkPixbuf *pixbuf)
+create_texture(GdkPixbuf *pixbuf, GLuint *texture, int i)
 {
-	void *data;
+	guchar *data, *buffer = NULL;
+	int width, height, stride, nch, new_stride;
 	GLenum format;
 
-	glGenTextures(1, &texture);
-	glBindTexture(GL_TEXTURE_2D, texture);
+	glActiveTexture(GL_TEXTURE0 + i);
+	glGenTextures(1, texture);
+	glBindTexture(GL_TEXTURE_2D, *texture);
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
@@ -132,10 +142,28 @@ create_texture(GdkPixbuf *pixbuf)
 		format = GL_RGBA;
 	else
 		format = GL_RGB;
+	
+	width = gdk_pixbuf_get_width(pixbuf);
+	height = gdk_pixbuf_get_height(pixbuf);
+	stride = gdk_pixbuf_get_rowstride(pixbuf);
+	nch = gdk_pixbuf_get_n_channels(pixbuf);
+
+	new_stride = align(nch * width, 4);
+
+	if (stride != new_stride) {
+		buffer = malloc(new_stride * height);
+		for (int row = 0; row < height; ++row)
+			memcpy(buffer + row * new_stride,
+			       data + row * stride,
+			       width * nch);
+		data = buffer;
+	}
 
 	glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB,
 		     width, height, 0,
 		     format, GL_UNSIGNED_BYTE, data);
+	if (buffer != NULL)
+		free(buffer);
 }
 
 static void
@@ -155,6 +183,7 @@ draw(void)
 
 	glUniformMatrix4fv(proj_uniform, 1, GL_FALSE, proj);
 	glUniform1i(tex_uniform, 0);
+	glUniform1i(tex2_uniform, 1);
 	glUniform2fv(pixelsize_uniform, 1, pixelsize);
 
 	search[0] = 1.0 * x1 / width;
@@ -167,7 +196,6 @@ draw(void)
 
 	glDisable(GL_BLEND);
 
-	glBindTexture(GL_TEXTURE_2D, texture);
 	glBindBuffer(GL_ARRAY_BUFFER, vbo);
 	glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, 5 * sizeof(GL_FLOAT), NULL);
 	glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, 5 * sizeof(GL_FLOAT), (GLfloat *)0 + 3);
@@ -233,6 +261,7 @@ main(int argc, char **argv)
 {
 	gint stride;
 	guchar *buffer;
+	GdkPixbuf *match;
 
 	g_type_init();
 
@@ -254,7 +283,12 @@ main(int argc, char **argv)
 	mheight = atoi(argv[5]);
 
 	init_shaders();
-	create_texture(pixbuf);
+	create_texture(pixbuf, &texture, 0);
+
+	match = gdk_pixbuf_new_subpixbuf(pixbuf, x1, y1, mwidth, mheight);
+	create_texture(match, &texture2, 1);
+	g_object_unref(match);
+
 	glViewport(0,0,width,height);
 	glClearColor(0.0f, 0.0f, 0.0f, 1.0f);