[media] vivi: Move computations out of vivi_fillbuf linecopy loop
authorKirill Smelkov <kirr@mns.spb.ru>
Fri, 2 Nov 2012 12:10:32 +0000 (09:10 -0300)
committerMauro Carvalho Chehab <mchehab@redhat.com>
Fri, 21 Dec 2012 20:25:43 +0000 (18:25 -0200)
The "dev->mvcount % wmax" thing was showing high in profiles (we do it
for each line which ~ 500 per frame)
           ?     000010c0 <vivi_fillbuff>:
                 ...
      0,39 ? 70:???mov    0x3ff4(%edi),%esi
      0,22 ? 76:?  mov    0x2a0(%edi),%eax
      0,30 ?    ?  mov    -0x84(%ebp),%ebx
      0,35 ?    ?  mov    %eax,%edx
      0,04 ?    ?  mov    -0x7c(%ebp),%ecx
      0,35 ?    ?  sar    $0x1f,%edx
      0,44 ?    ?  idivl  -0x7c(%ebp)
     21,68 ?    ?  imul   %esi,%ecx
      0,70 ?    ?  imul   %esi,%ebx
      0,52 ?    ?  add    -0x88(%ebp),%ebx
      1,65 ?    ?  mov    %ebx,%eax
      0,22 ?    ?  imul   %edx,%esi
      0,04 ?    ?  lea    0x3f4(%edi,%esi,1),%edx
      2,18 ?    ?? call   vivi_fillbuff+0xa6
      0,74 ?    ?  addl   $0x1,-0x80(%ebp)
     62,69 ?    ?  mov    -0x7c(%ebp),%edx
      1,18 ?    ?  mov    -0x80(%ebp),%ecx
      0,35 ?    ?  add    %edx,-0x84(%ebp)
      0,61 ?    ?  cmp    %ecx,-0x8c(%ebp)
      0,22 ?    ???jne    70
so since all variables stay the same for all iterations let's move
computations out of the loop: the abovementioned division and
"width*pixelsize" too
before:
    # cmdline : /home/kirr/local/perf/bin/perf record -g -a sleep 20
    #
    # Samples: 49K of event 'cycles'
    # Event count (approx.): 16475832370
    #
    # Overhead          Command           Shared Object
    # ........  ...............  ......................
    #
        29.07%             rawv  libc-2.13.so            [.] __memcpy_ssse3
        20.57%           vivi-*  [kernel.kallsyms]       [k] memcpy
        10.20%             Xorg  [unknown]               [.] 0xa7301494
         5.16%           vivi-*  [vivi]                  [k] gen_text.constprop.6
         4.43%             rawv  [vivi]                  [k] gen_twopix
         4.36%           vivi-*  [vivi]                  [k] vivi_fillbuff
         2.42%             rawv  [vivi]                  [k] precalculate_line
         1.33%          swapper  [kernel.kallsyms]       [k] read_hpet
after:
    # cmdline : /home/kirr/local/perf/bin/perf record -g -a sleep 20
    #
    # Samples: 46K of event 'cycles'
    # Event count (approx.): 15574200568
    #
    # Overhead          Command         Shared Object
    # ........  ...............  ....................
    #
        27.99%             rawv  libc-2.13.so          [.] __memcpy_ssse3
        23.29%           vivi-*  [kernel.kallsyms]     [k] memcpy
        10.30%             Xorg  [unknown]             [.] 0xa75c98f8
         5.34%           vivi-*  [vivi]                [k] gen_text.constprop.6
         4.61%             rawv  [vivi]                [k] gen_twopix
         2.64%             rawv  [vivi]                [k] precalculate_line
         1.37%          swapper  [kernel.kallsyms]     [k] read_hpet
         0.79%             Xorg  [kernel.kallsyms]     [k] read_hpet
         0.64%             Xorg  [kernel.kallsyms]     [k] unix_poll
         0.45%             Xorg  [kernel.kallsyms]     [k] fget_light
         0.43%             rawv  libxcb.so.1.1.0       [.] 0x0000aae9
         0.40%            runsv  [kernel.kallsyms]     [k] ext2_try_to_allocate
         0.36%             Xorg  [kernel.kallsyms]     [k] _raw_spin_lock_irqsave
         0.31%           vivi-*  [vivi]                [k] vivi_fillbuff
(i.e. vivi_fillbuff own overhead is almost gone)

Signed-off-by: Kirill Smelkov <kirr@mns.spb.ru>
Acked-by: Hans Verkuil <hans.verkuil@cisco.com>
Signed-off-by: Mauro Carvalho Chehab <mchehab@redhat.com>
drivers/media/platform/vivi.c

index a3090f0312c18b0409b6ab05647e1f81502d9745..ec832e9461a8c2075569e5dccd69a14cccf663e6 100644 (file)
@@ -574,21 +574,22 @@ static void gen_text(struct vivi_dev *dev, char *basep,
 
 static void vivi_fillbuff(struct vivi_dev *dev, struct vivi_buffer *buf)
 {
-       int wmax = dev->width;
+       int stride = dev->width * dev->pixelsize;
        int hmax = dev->height;
        void *vbuf = vb2_plane_vaddr(&buf->vb, 0);
        unsigned ms;
        char str[100];
        int h, line = 1;
+       u8 *linestart;
        s32 gain;
 
        if (!vbuf)
                return;
 
+       linestart = dev->line + (dev->mv_count % dev->width) * dev->pixelsize;
+
        for (h = 0; h < hmax; h++)
-               memcpy(vbuf + h * wmax * dev->pixelsize,
-                      dev->line + (dev->mv_count % wmax) * dev->pixelsize,
-                      wmax * dev->pixelsize);
+               memcpy(vbuf + h * stride, linestart, stride);
 
        /* Updates stream time */