merge the (rest of) texmem branch
diff --git a/progs/demos/texdown.c b/progs/demos/texdown.c
index 79525a0..fc98fdd 100644
--- a/progs/demos/texdown.c
+++ b/progs/demos/texdown.c
@@ -38,8 +38,8 @@
#include <GL/glut.h>
-static GLsizei MaxSize = 1024;
-static GLsizei TexWidth = 256, TexHeight = 256, TexBorder = 0;
+static GLsizei MaxSize = 2048;
+static GLsizei TexWidth = 1024, TexHeight = 1024, TexBorder = 0;
static GLboolean ScaleAndBias = GL_FALSE;
static GLboolean SubImage = GL_FALSE;
static GLdouble DownloadRate = 0.0; /* texels/sec */
@@ -47,6 +47,32 @@
static GLuint Mode = 0;
+/* Try and avoid L2 cache effects by cycling through a small number of
+ * textures.
+ *
+ * At the initial size of 1024x1024x4 == 4mbyte, say 8 textures will
+ * keep us out of most caches at 32mb total.
+ *
+ * This turns into a fairly interesting question of what exactly you
+ * expect to be in cache in normal usage, and what you think should be
+ * outside. There's no rules for this, no reason to favour one usage
+ * over another except what the application you care about happens to
+ * resemble most closely.
+ *
+ * - Should the client texture image be in L2 cache? Has it just been
+ * generated or read from disk?
+ * - Does the application really use >1 texture, or is it constantly
+ * updating one image in-place?
+ *
+ * Different answers will favour different texture upload mechanisms.
+ * To upload an image that is purely outside of cache, a DMA-based
+ * upload will probably win, whereas for small, in-cache textures,
+ * copying looks good.
+ */
+#define NR_TEXOBJ 4
+static GLuint TexObj[NR_TEXOBJ];
+
+
struct FormatRec {
GLenum Format;
GLenum Type;
@@ -116,25 +142,57 @@
}
}
+/* On x86, there is a performance cliff for memcpy to texture memory
+ * for sources below 64 byte alignment. We do our best with this in
+ * the driver, but it is better if the images are correctly aligned to
+ * start with:
+ */
+#define ALIGN (1<<12)
+
+static unsigned align(unsigned value, unsigned a)
+{
+ return (value + a - 1) & ~(a-1);
+}
+
+static int MIN2(int a, int b)
+{
+ return a < b ? a : b;
+}
static void
MeasureDownloadRate(void)
{
const int w = TexWidth + 2 * TexBorder;
const int h = TexHeight + 2 * TexBorder;
- const int bytes = w * h * BytesPerTexel(Format);
+ const int image_bytes = align(w * h * BytesPerTexel(Format), ALIGN);
+ const int bytes = image_bytes * NR_TEXOBJ;
+ GLubyte *orig_texImage, *orig_getImage;
GLubyte *texImage, *getImage;
GLdouble t0, t1, time;
int count;
int i;
+ int offset = 0;
+ GLdouble total = 0; /* ints will tend to overflow */
- texImage = (GLubyte *) malloc(bytes);
- getImage = (GLubyte *) malloc(bytes);
- if (!texImage || !getImage) {
+ printf("allocating %d bytes for %d %dx%d images\n",
+ bytes, NR_TEXOBJ, w, h);
+
+ orig_texImage = (GLubyte *) malloc(bytes + ALIGN);
+ orig_getImage = (GLubyte *) malloc(image_bytes + ALIGN);
+ if (!orig_texImage || !orig_getImage) {
DownloadRate = 0.0;
return;
}
+ printf("alloc %p %p\n", orig_texImage, orig_getImage);
+
+ texImage = (GLubyte *)align((unsigned)orig_texImage, ALIGN);
+ getImage = (GLubyte *)align((unsigned)orig_getImage, ALIGN);
+
+ for (i = 1; !(((unsigned)texImage) & i); i<<=1)
+ ;
+ printf("texture image alignment: %d bytes (%p)\n", i, texImage);
+
for (i = 0; i < bytes; i++) {
texImage[i] = i & 0xff;
}
@@ -166,16 +224,50 @@
count = 0;
t0 = glutGet(GLUT_ELAPSED_TIME) * 0.001;
do {
+ int img = count%NR_TEXOBJ;
+ GLubyte *img_ptr = texImage + img * image_bytes;
+
+ glBindTexture(GL_TEXTURE_2D, TexObj[img]);
+
if (SubImage && count > 0) {
- glTexSubImage2D(GL_TEXTURE_2D, 0, -TexBorder, -TexBorder, w, h,
+ /* Only update a portion of the image each iteration. This
+ * is presumably why you'd want to use texsubimage, otherwise
+ * you may as well just call teximage again.
+ *
+ * A bigger question is whether to use a pointer that moves
+ * with each call, ie does the incoming data come from L2
+ * cache under normal circumstances, or is it pulled from
+ * uncached memory?
+ *
+ * There's a good argument to say L2 cache, ie you'd expect
+ * the data to have been recently generated. It's possible
+ * that it could have come from a file read, which may or may
+ * not have gone through the cpu.
+ */
+ glTexSubImage2D(GL_TEXTURE_2D, 0,
+ -TexBorder,
+ -TexBorder + offset * h/8,
+ w,
+ h/8,
FormatTable[Format].Format,
- FormatTable[Format].Type, texImage);
+ FormatTable[Format].Type,
+#if 1
+ texImage /* likely in L2$ */
+#else
+ img_ptr + offset * bytes/8 /* unlikely in L2$ */
+#endif
+ );
+ offset += 1;
+ offset %= 8;
+ total += w * h / 8;
}
else {
glTexImage2D(GL_TEXTURE_2D, 0,
FormatTable[Format].IntFormat, w, h, TexBorder,
FormatTable[Format].Format,
- FormatTable[Format].Type, texImage);
+ FormatTable[Format].Type,
+ img_ptr);
+ total += w*h;
}
/* draw a tiny polygon to force texture into texram */
@@ -192,25 +284,12 @@
glDisable(GL_TEXTURE_2D);
- printf("w*h=%d count=%d time=%f\n", w*h, count, time);
- DownloadRate = w * h * count / time;
+ printf("total texels=%f time=%f\n", total, time);
+ DownloadRate = total / time;
-#if 0
- if (!ScaleAndBias) {
- /* verify texture readback */
- glGetTexImage(GL_TEXTURE_2D, 0,
- FormatTable[Format].Format,
- FormatTable[Format].Type, getImage);
- for (i = 0; i < w * h; i++) {
- if (texImage[i] != getImage[i]) {
- printf("[%d] %d != %d\n", i, texImage[i], getImage[i]);
- }
- }
- }
-#endif
- free(texImage);
- free(getImage);
+ free(orig_texImage);
+ free(orig_getImage);
{
GLint err = glGetError();