[FFmpeg-devel] [PATCH 1/2] avcodec/ffv1enc: replace 3 double precision log2() by 1 single precision log2()

[FFmpeg-devel] [PATCH 1/2] avcodec/ffv1enc: replace 3 double precision log2() by 1 single precision log2()

[Michael Niedermayer]

This makes the remap finding heuristic much faster

Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
---
 libavcodec/ffv1enc.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/libavcodec/ffv1enc.c b/libavcodec/ffv1enc.c
index a09b31218db..2a188ade776 100644
--- a/libavcodec/ffv1enc.c
+++ b/libavcodec/ffv1enc.c
@@ -1390,7 +1390,7 @@ static void encode_float32_remap(FFV1Context *f, FFV1SliceContext *sc,
             int last_mul_index = -1;
             int mul_count = 1 << log2_mul_count;
 
-            score_sum[log2_mul_count] += log2_mul_count;
+            score_sum[log2_mul_count] += log2_mul_count * log2_mul_count;
             for (int i= 0; i<pixel_num; i++) {
                 int64_t val = sc->unit[p][i].val;
                 int mul_index = (val + 1LL)*mul_count >> 32;
@@ -1414,11 +1414,17 @@ static void encode_float32_remap(FFV1Context *f, FFV1SliceContext *sc,
                         }
 
                         cost = FFMAX((delta + mul/2)  / mul, 1);
-                        score_tab[si] += log2(cost);
-                        if (mul > 1)
-                            score_tab[si] += log2(fabs(delta - cost*mul)+1) * (1 + (mul_count > 1));
+                        float score = 1;
+                        if (mul > 1) {
+                            score *= (fabs(delta - cost*mul)+1);
+                            if (mul_count > 1)
+                                score *= score;
+                        }
+                        score *= cost;
+                        score *= score;
                         if (mul_index != last_mul_index)
-                            score_tab[si] += 0.5*log2(mul);
+                            score *= mul;
+                        score_tab[si] += log2f(score);
                     }
                 }
                 last_val = val;
-- 
2.49.0

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

[FFmpeg-devel] [PATCH 2/2] avcodec/ffv1enc: replace the remaining log2() by magic

[Michael Niedermayer]

big secret, reading a float as the corresponding integer is a good
approximation of log2() for numbers not too close to 1.0. At the same
time it maintains strict monotonicity

this reduces run time from 19sec to 17sec for the slowest of my testcases
(with default remap_optimizer 3)

Compression is about the same:

-rw-r----- 1 michael michael  497603370 Mar 30 15:23 float-303503-fixed-40-optim-3-1log2F.nut
-rw-r----- 1 michael michael  497603374 Mar 28 11:27 float-303503-fixed-40-optim-3b.nut
-rw-r----- 1 michael michael  549938852 Mar 28 11:27 float-303503-float16-40-optim-3b.nut
-rw-r----- 1 michael michael  549938857 Mar 30 15:23 float-303503-float16-40-optim-3-1log2F.nut
-rw-r----- 1 michael michael 1150827841 Mar 28 11:28 float-303503-float-40-optim-3b.nut
-rw-r----- 1 michael michael 1150832913 Mar 30 15:22 float-303503-float-40-optim-3ref-log2F2.nut

Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
---
 libavcodec/ffv1enc.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/libavcodec/ffv1enc.c b/libavcodec/ffv1enc.c
index 2a188ade776..a09ab300389 100644
--- a/libavcodec/ffv1enc.c
+++ b/libavcodec/ffv1enc.c
@@ -1375,6 +1375,7 @@ static void encode_float32_remap(FFV1Context *f, FFV1SliceContext *sc,
     const int log2_mul_step       = ((int[]){  1,  8,  1,  1,  1,   1})[f->remap_optimizer];
     const int bruteforce_count    = ((int[]){  0,  0,  0,  1,  1,   1})[f->remap_optimizer];
     const int stair_mode          = ((int[]){  0,  0,  0,  1,  0,   0})[f->remap_optimizer];
+    const int magic_log2          = ((int[]){  1,  1,  1,  1,  0,   0})[f->remap_optimizer];
 
     av_assert0 (pixel_num <= 65536);
 
@@ -1390,7 +1391,9 @@ static void encode_float32_remap(FFV1Context *f, FFV1SliceContext *sc,
             int last_mul_index = -1;
             int mul_count = 1 << log2_mul_count;
 
-            score_sum[log2_mul_count] += log2_mul_count * log2_mul_count;
+            score_sum[log2_mul_count] = 2 * log2_mul_count;
+            if (magic_log2)
+                score_sum[log2_mul_count] = av_float2int((float)mul_count * mul_count);
             for (int i= 0; i<pixel_num; i++) {
                 int64_t val = sc->unit[p][i].val;
                 int mul_index = (val + 1LL)*mul_count >> 32;
@@ -1424,7 +1427,10 @@ static void encode_float32_remap(FFV1Context *f, FFV1SliceContext *sc,
                         score *= score;
                         if (mul_index != last_mul_index)
                             score *= mul;
-                        score_tab[si] += log2f(score);
+                        if (magic_log2) {
+                            score_tab[si] += av_float2int(score);
+                        } else
+                            score_tab[si] += log2f(score);
                     }
                 }
                 last_val = val;
-- 
2.49.0

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

Re: [FFmpeg-devel] [PATCH 1/2] avcodec/ffv1enc: replace 3 double precision log2() by 1 single precision log2()

[Michael Niedermayer]

[-- Attachment #1.1: Type: text/plain, Size: 506 bytes --]

On Sun, Mar 30, 2025 at 05:00:22PM +0200, Michael Niedermayer wrote:
> This makes the remap finding heuristic much faster
> 
> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
> ---
>  libavcodec/ffv1enc.c | 16 +++++++++++-----
>  1 file changed, 11 insertions(+), 5 deletions(-)

will apply patchset

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

The greatest way to live with honor in this world is to be what we pretend
to be. -- Socrates

[-- Attachment #1.2: signature.asc --]
[-- Type: application/pgp-signature, Size: 195 bytes --]

[-- Attachment #2: Type: text/plain, Size: 251 bytes --]

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".