On 22 Apr, this message from Benjamin Herrenschmidt echoed through cyberspace:
>>Other point: performance. I am getting around 70 dropped frames out of
>>each block of 200... Which makes DVD useable, but impossible to really
>>_watch_ a movie. What are others getting? This is on a G4/400 TiBook.
>>Has anybody made any atttempt at using Altivec to speed things up?
>
> I'm pretty sure most of the performance problem is not due to decoding
> but to PCI throughput in Xv. This could probably be improved either by
> using bust mastering within Xv (using DRI) but that would be quite
> complicated, or by figuring out a way for Xv to map the YUV overlay
> buffer write-through cacheable.
How about the attached patch then? It adds an mmap() operation to
aty128fb, thus X maps the framebuffer writethru:
aty128fb: Mapping uncached: start=a0000000, len=8192.
aty128fb: Mapping uncached: start=a0000000, len=8192.
aty128fb: Mapping uncached: start=a0000000, len=8192.
aty128fb: Mapping WriteThru: start=a4000000, len=8388608.
VFS: Disk change detected on device ide1(22,0)
^^^^^^^^^^^
This is the DVD drive ;-)
I.e. X has one mmap() on the framebuffer, and that one is set writethru.
But performance in xine doesn't improve :( still have ~70 out of 200
frames dropped.
So I guess we need to look at other solutions... Note that while playing
a DVD, top shows one xine process consuming around 70% CPU (video
decoder?), X has 15-20%, and two other xine processes consume around 5%
each.
Patch for mmap is attached (against recent benh 2.4 kernels).
Cheers
Michel
-----------------------------------------------------------------------
Michel Lanners | " Read Philosophy. Study Art.
23, Rue Paul Henkes | Ask Questions. Make Mistakes.
L-1710 Luxembourg |
email [EMAIL PROTECTED] |
http://www.cpu.lu/~mlan | Learn Always. "
--- linux-2.4.benh-244pre1/drivers/video/aty128fb.c Thu Apr 12 20:27:03 2001
+++ linux-2.4.benh/drivers/video/aty128fb.c Mon Apr 23 13:17:19 2001
@@ -351,6 +351,8 @@
static int aty128fb_ioctl(struct inode *inode, struct file *file, u_int cmd,
u_long arg, int con, struct fb_info *info);
static int aty128fb_rasterimg(struct fb_info *info, int start);
+static int aty128fb_mmap(struct fb_info *info, struct file *file,
+ struct vm_area_struct *vma);
/*
@@ -443,6 +445,7 @@
fb_pan_display: aty128fb_pan_display,
fb_ioctl: aty128fb_ioctl,
fb_rasterimg: aty128fb_rasterimg,
+ fb_mmap: aty128fb_mmap,
};
#ifdef CONFIG_PMAC_BACKLIGHT
@@ -1693,6 +1696,53 @@
}
+/* Private mmap since we want to have a different caching on the framebuffer
+ * for aty128fb.
+ * Note there's no locking in here; it's done in fb_mmap() in fbmem.c.
+ */
+static int aty128fb_mmap(struct fb_info *info, struct file *file,
+ struct vm_area_struct *vma)
+{
+ struct fb_ops *fb = info->fbops;
+ struct fb_fix_screeninfo fix;
+ struct fb_var_screeninfo var;
+ unsigned long off, start;
+ u32 len;
+
+ fb->fb_get_fix(&fix, PROC_CONSOLE(info), info);
+ off = vma->vm_pgoff << PAGE_SHIFT;
+
+ /* frame buffer memory */
+ start = fix.smem_start;
+ len = PAGE_ALIGN((start & ~PAGE_MASK)+fix.smem_len);
+ if (off >= len) {
+ /* memory mapped io */
+ off -= len;
+ fb->fb_get_var(&var, PROC_CONSOLE(info), info);
+ if (var.accel_flags)
+ return -EINVAL;
+ start = fix.mmio_start;
+ len = PAGE_ALIGN((start & ~PAGE_MASK)+fix.mmio_len);
+ pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE|_PAGE_GUARDED;
+ printk("aty128fb: Mapping uncached: start=%lp, len=%d.\n",
start, len);
+ } else {
+ /* framebuffer */
+ pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
+ printk("aty128fb: Mapping WriteThru: start=%lp, len=%d.\n",
start, len);
+ }
+ start &= PAGE_MASK;
+ if ((vma->vm_end - vma->vm_start + off) > len)
+ return -EINVAL;
+ off += start;
+ vma->vm_pgoff = off >> PAGE_SHIFT;
+ if (io_remap_page_range(vma->vm_start, off,
+ vma->vm_end - vma->vm_start, vma->vm_page_prot))
+ return -EAGAIN;
+
+ return 0;
+}
+
+
#ifndef MODULE
int __init
aty128fb_setup(char *options)
