Revision 4bdc44c7

View differences:

libswscale/internal_bfin.S
24 24

  
25 25

  
26 26
/*
27
    YUV420 to RGB565 conversion.  This routine takes a YUV 420 planar macroblock
28
    and converts it to RGB565.  R:5 bits, G:6 bits, B:5 bits.. packed into shorts
27
YUV420 to RGB565 conversion.  This routine takes a YUV 420 planar macroblock
28
and converts it to RGB565.  R:5 bits, G:6 bits, B:5 bits.. packed into shorts
29 29

  
30 30

  
31
    The following calculation is used for the conversion:
31
The following calculation is used for the conversion:
32 32

  
33
      r = clipz((y-oy)*cy  + crv*(v-128))
34
      g = clipz((y-oy)*cy  + cgv*(v-128) + cgu*(u-128))
35
      b = clipz((y-oy)*cy  + cbu*(u-128))
33
  r = clipz((y-oy)*cy  + crv*(v-128))
34
  g = clipz((y-oy)*cy  + cgv*(v-128) + cgu*(u-128))
35
  b = clipz((y-oy)*cy  + cbu*(u-128))
36 36

  
37
    y,u,v are pre scaled by a factor of 4 i.e. left shifted to gain precision.
37
y,u,v are pre scaled by a factor of 4 i.e. left shifted to gain precision.
38 38

  
39 39

  
40
    New factorization to eliminate the truncation error which was
41
    occuring due to the byteop3p.
40
New factorization to eliminate the truncation error which was
41
occuring due to the byteop3p.
42 42

  
43 43

  
44
  1) use the bytop16m to subtract quad bytes we use this in U8 this
45
   then so the offsets need to be renormalized to 8bits.
44
1) use the bytop16m to subtract quad bytes we use this in U8 this
45
 then so the offsets need to be renormalized to 8bits.
46 46

  
47
  2) scale operands up by a factor of 4 not 8 because Blackfin
48
     multiplies include a shift.
47
2) scale operands up by a factor of 4 not 8 because Blackfin
48
   multiplies include a shift.
49 49

  
50
  3) compute into the accumulators cy*yx0, cy*yx1
50
3) compute into the accumulators cy*yx0, cy*yx1
51 51

  
52
  4) compute each of the linear equations
53
      r = clipz((y-oy)*cy  + crv*(v-128))
52
4) compute each of the linear equations
53
     r = clipz((y - oy) * cy  + crv * (v - 128))
54 54

  
55
      g = clipz((y-oy)*cy  + cgv*(v-128) + cgu*(u-128))
55
     g = clipz((y - oy) * cy  + cgv * (v - 128) + cgu * (u - 128))
56 56

  
57
      b = clipz((y-oy)*cy  + cbu*(u-128))
57
     b = clipz((y - oy) * cy  + cbu * (u - 128))
58 58

  
59
     reuse of the accumulators requires that we actually multiply
60
     twice once with addition and the second time with a subtaction.
59
   reuse of the accumulators requires that we actually multiply
60
   twice once with addition and the second time with a subtaction.
61 61

  
62
     because of this we need to compute the equations in the order R B
63
     then G saving the writes for B in the case of 24/32 bit color
64
     formats.
62
   because of this we need to compute the equations in the order R B
63
   then G saving the writes for B in the case of 24/32 bit color
64
   formats.
65 65

  
66
    api: yuv2rgb_kind (uint8_t *Y, uint8_t *U, uint8_t *V, int *out,
67
                       int dW, uint32_t *coeffs);
66
   api: yuv2rgb_kind (uint8_t *Y, uint8_t *U, uint8_t *V, int *out,
67
                      int dW, uint32_t *coeffs);
68 68

  
69
        A          B
70
        ---        ---
71
        i2 = cb    i3 = cr
72
        i1 = coeff i0 = y
69
       A          B
70
       ---        ---
71
       i2 = cb    i3 = cr
72
       i1 = coeff i0 = y
73 73

  
74
  Where coeffs have the following layout in memory.
74
Where coeffs have the following layout in memory.
75 75

  
76
  uint32_t oy,oc,zero,cy,crv,rmask,cbu,bmask,cgu,cgv;
76
uint32_t oy,oc,zero,cy,crv,rmask,cbu,bmask,cgu,cgv;
77 77

  
78
  coeffs is a pointer to oy.
78
coeffs is a pointer to oy.
79 79

  
80
  the {rgb} masks are only utilized by the 565 packing algorithm. Note the data
81
  replication is used to simplify the internal algorithms for the dual mac architecture
82
  of BlackFin.
80
the {rgb} masks are only utilized by the 565 packing algorithm. Note the data
81
replication is used to simplify the internal algorithms for the dual mac architecture
82
of BlackFin.
83 83

  
84
  All routines are exported with _ff_bfin_ as a symbol prefix
84
All routines are exported with _ff_bfin_ as a symbol prefix
85 85

  
86
  rough performance gain compared against -O3:
86
rough performance gain compared against -O3:
87 87

  
88
  2779809/1484290 187.28%
89

  
90
  which translates to ~33c/pel to ~57c/pel for the reference vs 17.5
91
  c/pel for the optimized implementations. Not sure why there is such a
92
  huge variation on the reference codes on Blackfin I guess it must have
93
  to do with the memory system.
88
2779809/1484290 187.28%
94 89

  
90
which translates to ~33c/pel to ~57c/pel for the reference vs 17.5
91
c/pel for the optimized implementations. Not sure why there is such a
92
huge variation on the reference codes on Blackfin I guess it must have
93
to do with the memory system.
95 94
*/
96 95

  
97 96
#define mL3 .text
libswscale/yuv2rgb_altivec.c
21 21
 */
22 22

  
23 23
/*
24
  convert I420 YV12 to RGB in various formats,
25
    it rejects images that are not in 420 formats
26
    it rejects images that don't have widths of multiples of 16
27
    it rejects images that don't have heights of multiples of 2
28
  reject defers to C simulation codes.
24
convert I420 YV12 to RGB in various formats,
25
  it rejects images that are not in 420 formats
26
  it rejects images that don't have widths of multiples of 16
27
  it rejects images that don't have heights of multiples of 2
28
reject defers to C simulation codes.
29 29

  
30
  lots of optimizations to be done here
30
lots of optimizations to be done here
31 31

  
32
  1. need to fix saturation code, I just couldn't get it to fly with packs and adds.
33
     so we currently use max min to clip
32
1. need to fix saturation code, I just couldn't get it to fly with packs and adds.
33
   so we currently use max min to clip
34 34

  
35
  2. the inefficient use of chroma loading needs a bit of brushing up
35
2. the inefficient use of chroma loading needs a bit of brushing up
36 36

  
37
  3. analysis of pipeline stalls needs to be done, use shark to identify pipeline stalls
37
3. analysis of pipeline stalls needs to be done, use shark to identify pipeline stalls
38 38

  
39 39

  
40
  MODIFIED to calculate coeffs from currently selected color space.
41
  MODIFIED core to be a macro which you spec the output format.
42
  ADDED UYVY conversion which is never called due to some thing in SWSCALE.
43
  CORRECTED algorithim selection to be strict on input formats.
44
  ADDED runtime detection of altivec.
40
MODIFIED to calculate coeffs from currently selected color space.
41
MODIFIED core to be a macro which you spec the output format.
42
ADDED UYVY conversion which is never called due to some thing in SWSCALE.
43
CORRECTED algorithim selection to be strict on input formats.
44
ADDED runtime detection of altivec.
45 45

  
46
  ADDED altivec_yuv2packedX vertical scl + RGB converter
46
ADDED altivec_yuv2packedX vertical scl + RGB converter
47 47

  
48
  March 27,2004
49
  PERFORMANCE ANALYSIS
48
March 27,2004
49
PERFORMANCE ANALYSIS
50 50

  
51
  The C version use 25% of the processor or ~250Mips for D1 video rawvideo used as test
52
  The ALTIVEC version uses 10% of the processor or ~100Mips for D1 video same sequence
51
The C version use 25% of the processor or ~250Mips for D1 video rawvideo used as test
52
The ALTIVEC version uses 10% of the processor or ~100Mips for D1 video same sequence
53 53

  
54
  720*480*30  ~10MPS
54
720*480*30  ~10MPS
55 55

  
56
  so we have roughly 10clocks per pixel this is too high something has to be wrong.
56
so we have roughly 10clocks per pixel this is too high something has to be wrong.
57 57

  
58
  OPTIMIZED clip codes to utilize vec_max and vec_packs removing the need for vec_min.
58
OPTIMIZED clip codes to utilize vec_max and vec_packs removing the need for vec_min.
59 59

  
60
  OPTIMIZED DST OUTPUT cache/dma controls. we are pretty much
61
  guaranteed to have the input video frame it was just decompressed so
62
  it probably resides in L1 caches.  However we are creating the
63
  output video stream this needs to use the DSTST instruction to
64
  optimize for the cache.  We couple this with the fact that we are
65
  not going to be visiting the input buffer again so we mark it Least
66
  Recently Used.  This shaves 25% of the processor cycles off.
60
OPTIMIZED DST OUTPUT cache/dma controls. we are pretty much
61
guaranteed to have the input video frame it was just decompressed so
62
it probably resides in L1 caches.  However we are creating the
63
output video stream this needs to use the DSTST instruction to
64
optimize for the cache.  We couple this with the fact that we are
65
not going to be visiting the input buffer again so we mark it Least
66
Recently Used.  This shaves 25% of the processor cycles off.
67 67

  
68
  Now MEMCPY is the largest mips consumer in the system, probably due
69
  to the inefficient X11 stuff.
68
Now MEMCPY is the largest mips consumer in the system, probably due
69
to the inefficient X11 stuff.
70 70

  
71
  GL libraries seem to be very slow on this machine 1.33Ghz PB running
72
  Jaguar, this is not the case for my 1Ghz PB.  I thought it might be
73
  a versioning issues, however I have libGL.1.2.dylib for both
74
  machines. ((We need to figure this out now))
71
GL libraries seem to be very slow on this machine 1.33Ghz PB running
72
Jaguar, this is not the case for my 1Ghz PB.  I thought it might be
73
a versioning issues, however I have libGL.1.2.dylib for both
74
machines. ((We need to figure this out now))
75 75

  
76
  GL2 libraries work now with patch for RGB32
76
GL2 libraries work now with patch for RGB32
77 77

  
78
  NOTE quartz vo driver ARGB32_to_RGB24 consumes 30% of the processor
78
NOTE quartz vo driver ARGB32_to_RGB24 consumes 30% of the processor
79 79

  
80
  Integrated luma prescaling adjustment for saturation/contrast/brightness adjustment.
80
Integrated luma prescaling adjustment for saturation/contrast/brightness adjustment.
81 81
*/
82 82

  
83 83
#include <stdio.h>

Also available in: Unified diff