On Wed, Sep 21, 2022 at 01:00:07PM +0200, Andreas Rheinhardt wrote:
> Michael Niedermayer:
> > On Mon, Sep 19, 2022 at 11:27:49PM +0200, Andreas Rheinhardt wrote:
> >> Only used there and by x86 snow asm code as fallback.
> >>
> >> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
> >> ---
> >>  libavcodec/snow.c     | 33 ---------------------------------
> >>  libavcodec/snow.h     |  3 ---
> >>  libavcodec/snow_dwt.c | 32 ++++++++++++++++++++++++++++++++
> >>  libavcodec/snow_dwt.h |  3 +++
> >>  4 files changed, 35 insertions(+), 36 deletions(-)
> >>
> >> diff --git a/libavcodec/snow.c b/libavcodec/snow.c
> >> index aa15fccc42..85ad6d10a1 100644
> >> --- a/libavcodec/snow.c
> >> +++ b/libavcodec/snow.c
> >> @@ -29,39 +29,6 @@
> >>  #include "snowdata.h"
> >>  
> >>  
> >> -void ff_snow_inner_add_yblock(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
> >> -                              int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
> >> -    int y, x;
> >> -    IDWTELEM * dst;
> >> -    for(y=0; y<b_h; y++){
> >> -        //FIXME ugly misuse of obmc_stride
> >> -        const uint8_t *obmc1= obmc + y*obmc_stride;
> >> -        const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
> >> -        const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
> >> -        const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
> >> -        dst = slice_buffer_get_line(sb, src_y + y);
> >> -        for(x=0; x<b_w; x++){
> >> -            int v=   obmc1[x] * block[3][x + y*src_stride]
> >> -                    +obmc2[x] * block[2][x + y*src_stride]
> >> -                    +obmc3[x] * block[1][x + y*src_stride]
> >> -                    +obmc4[x] * block[0][x + y*src_stride];
> >> -
> >> -            v <<= 8 - LOG2_OBMC_MAX;
> >> -            if(FRAC_BITS != 8){
> >> -                v >>= 8 - FRAC_BITS;
> >> -            }
> >> -            if(add){
> >> -                v += dst[x + src_x];
> >> -                v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
> >> -                if(v&(~255)) v= ~(v>>31);
> >> -                dst8[x + y*src_stride] = v;
> >> -            }else{
> >> -                dst[x + src_x] -= v;
> >> -            }
> >> -        }
> >> -    }
> >> -}
> >> -
> >>  int ff_snow_get_buffer(SnowContext *s, AVFrame *frame)
> >>  {
> >>      int ret, i;
> >> diff --git a/libavcodec/snow.h b/libavcodec/snow.h
> >> index ed0f9abb42..1c976b9ba7 100644
> >> --- a/libavcodec/snow.h
> >> +++ b/libavcodec/snow.h
> >> @@ -45,11 +45,8 @@
> >>  #define QSHIFT 5
> >>  #define QROOT (1<<QSHIFT)
> >>  #define LOSSLESS_QLOG -128
> >> -#define FRAC_BITS 4
> >>  #define MAX_REF_FRAMES 8
> >>  
> >> -#define LOG2_OBMC_MAX 8
> >> -#define OBMC_MAX (1<<(LOG2_OBMC_MAX))
> >>  typedef struct BlockNode{
> >>      int16_t mx;                 ///< Motion vector component X, see mv_scale
> >>      int16_t my;                 ///< Motion vector component Y, see mv_scale
> >> diff --git a/libavcodec/snow_dwt.c b/libavcodec/snow_dwt.c
> >> index 18b315ef66..9401d119d0 100644
> >> --- a/libavcodec/snow_dwt.c
> >> +++ b/libavcodec/snow_dwt.c
> >> @@ -25,6 +25,38 @@
> >>  #include "me_cmp.h"
> >>  #include "snow_dwt.h"
> >>  
> >> +void ff_snow_inner_add_yblock(const uint8_t *obmc, const int obmc_stride,
> >> +                              uint8_t **block, int b_w, int b_h,
> >> +                              int src_x, int src_y, int src_stride,
> >> +                              slice_buffer * sb, int add, uint8_t * dst8)
> >> +{
> >> +    for (int y = 0; y < b_h; y++) {
> >> +        //FIXME ugly misuse of obmc_stride
> >> +        const uint8_t *obmc1 = obmc  + y * obmc_stride;
> >> +        const uint8_t *obmc2 = obmc1 + (obmc_stride >> 1);
> >> +        const uint8_t *obmc3 = obmc1 + obmc_stride * (obmc_stride >> 1);
> >> +        const uint8_t *obmc4 = obmc3 + (obmc_stride >> 1);
> >> +        IDWTELEM *dst = slice_buffer_get_line(sb, src_y + y);
> >> +        for (int x = 0; x < b_w; x++) {
> >> +            int v = obmc1[x] * block[3][x + y*src_stride]
> >> +                  + obmc2[x] * block[2][x + y*src_stride]
> >> +                  + obmc3[x] * block[1][x + y*src_stride]
> >> +                  + obmc4[x] * block[0][x + y*src_stride];
> >> +
> >> +            v <<= 8 - LOG2_OBMC_MAX;
> >> +            if (FRAC_BITS != 8)
> >> +                v >>= 8 - FRAC_BITS;
> >> +            if (add) {
> >> +                v += dst[x + src_x];
> >> +                v = (v + (1 << (FRAC_BITS - 1))) >> FRAC_BITS;
> >> +                if (v & (~255)) v= ~(v>>31);
> >> +                dst8[x + y*src_stride] = v;
> >> +            } else
> >> +                dst[x + src_x] -= v;
> >> +        }
> >> +    }
> >> +}
> >> +
> > 
> > putting this in snow_dwt may be convenient but it is not part of the dwt so
> > this feels semantically wrong
> > 
> 
> If it is not part of the dwt, then why does SnowDWTContext have an
> inner_add_yblock function pointer whose C version is
> ff_snow_inner_add_yblock?

Iam not sure how this relates to what is part of a DWT
you can check the definition of a DWT and then look at this function
its not a DWT, its not part of some optimized DWT implementation either
what it is, is  part of snows OBMC+DWT related code.
The callback is a optimization of that, for the purpose of optimization this
is mixed together like that.

a
git grep inner_add_yblock origin
also shows mostly snowdsp matches not snowdwt

is it really a problem to leave dsp code in files with dsp in t the name ?
if so iam not against making an exception here

thx

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Nations do behave wisely once they have exhausted all other alternatives. 
-- Abba Eban