[FFmpeg-devel] [PR] Reduce size of snow allocations (PR #21793)

[mkver via ffmpeg-devel <ffmpeg-devel@ffmpeg.org>]

PR #21793 opened by mkver
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21793
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21793.patch


>From 9a1d99a7506414045c6e2b0f5433caa33370bb21 Mon Sep 17 00:00:00 2001
From: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Date: Mon, 9 Feb 2026 14:33:01 +0100
Subject: [PATCH 1/2] avcodec/snow: Only allocate emu_edge_buffer for encoder

Also allocate it during init and move it to the encoder's context.

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
---
 libavcodec/snow.c    | 6 +-----
 libavcodec/snow.h    | 1 -
 libavcodec/snowenc.c | 9 ++++++++-
 3 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/libavcodec/snow.c b/libavcodec/snow.c
index 006d84d8ce..094b893c96 100644
--- a/libavcodec/snow.c
+++ b/libavcodec/snow.c
@@ -540,10 +540,7 @@ int ff_snow_common_init_after_header(AVCodecContext *avctx) {
     int plane_index, level, orientation;
 
     if(!s->scratchbuf) {
-        int emu_buf_size;
-        emu_buf_size = FFMAX(s->mconly_picture->linesize[0], 2*avctx->width+256) * (2 * MB_SIZE + HTAPS_MAX - 1);
-        if (!FF_ALLOCZ_TYPED_ARRAY(s->scratchbuf,      FFMAX(s->mconly_picture->linesize[0], 2*avctx->width+256) * 7 * MB_SIZE) ||
-            !FF_ALLOCZ_TYPED_ARRAY(s->emu_edge_buffer, emu_buf_size))
+        if (!FF_ALLOCZ_TYPED_ARRAY(s->scratchbuf, FFMAX(s->mconly_picture->linesize[0], 2*avctx->width+256) * 7 * MB_SIZE))
             return AVERROR(ENOMEM);
     }
 
@@ -642,7 +639,6 @@ av_cold void ff_snow_common_end(SnowContext *s)
 
     av_freep(&s->block);
     av_freep(&s->scratchbuf);
-    av_freep(&s->emu_edge_buffer);
 
     for(i=0; i<MAX_REF_FRAMES; i++){
         av_frame_free(&s->last_picture[i]);
diff --git a/libavcodec/snow.h b/libavcodec/snow.h
index 83dc6c1256..82417c3324 100644
--- a/libavcodec/snow.h
+++ b/libavcodec/snow.h
@@ -168,7 +168,6 @@ typedef struct SnowContext{
     slice_buffer sb;
 
     uint8_t *scratchbuf;
-    uint8_t *emu_edge_buffer;
 
     AVMotionVector *avmv;
     unsigned avmv_size;
diff --git a/libavcodec/snowenc.c b/libavcodec/snowenc.c
index 20a41f11a4..f34130f184 100644
--- a/libavcodec/snowenc.c
+++ b/libavcodec/snowenc.c
@@ -68,6 +68,8 @@ typedef struct SnowEncContext {
 
     uint64_t encoding_error[SNOW_MAX_PLANES];
 
+    uint8_t *emu_edge_buffer;
+
     IDWTELEM obmc_scratchpad[MB_SIZE * MB_SIZE * 12 * 2];
 } SnowEncContext;
 
@@ -286,6 +288,10 @@ static av_cold int encode_init(AVCodecContext *avctx)
     if ((ret = get_encode_buffer(s, s->input_picture)) < 0)
         return ret;
 
+    enc->emu_edge_buffer = av_calloc(avctx->width + 128, 2 * (2 * MB_SIZE + HTAPS_MAX - 1));
+    if (!enc->emu_edge_buffer)
+        return AVERROR(ENOMEM);
+
     if (enc->motion_est == FF_ME_ITER) {
         int size= s->b_width * s->b_height << 2*s->block_max_depth;
         for(i=0; i<s->max_ref_frames; i++){
@@ -770,7 +776,7 @@ static int get_block_rd(SnowEncContext *enc, int mb_x, int mb_y,
     const uint8_t *src = s->input_picture->data[plane_index];
     IDWTELEM *pred = enc->obmc_scratchpad + plane_index * block_size * block_size * 4;
     uint8_t *cur = s->scratchbuf;
-    uint8_t *tmp = s->emu_edge_buffer;
+    uint8_t *tmp = enc->emu_edge_buffer;
     const int b_stride = s->b_width << s->block_max_depth;
     const int b_height = s->b_height<< s->block_max_depth;
     const int w= p->width;
@@ -2088,6 +2094,7 @@ static av_cold int encode_end(AVCodecContext *avctx)
 
     enc->m.s.me.temp = NULL;
     av_freep(&enc->m.s.me.scratchpad);
+    av_freep(&enc->emu_edge_buffer);
 
     av_freep(&avctx->stats_out);
 
-- 
2.52.0


>From 8de00b910c237943c7956f2851ce2ccf3014091d Mon Sep 17 00:00:00 2001
From: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Date: Mon, 9 Feb 2026 17:57:30 +0100
Subject: [PATCH 2/2] avcodec/snow: Reduce sizeof(SnowContext)

Each SubBand currently contains an array of 519 uint8_t[32],
yet most of these are unused: For both the decoder and the
encoder, at most 34 contexts are actually used: The only
variable index is context+2, where context is the result
of av_log2() and therefore in the 0..31 range.

There are also several accesses using compile-time indices,
the highest of which is 30. FATE passes with 31 contexts
and maybe these are enough, but I don't know.

Reducing the number to 34 reduces sizeof(SnowContext)
from 2141664B to 155104B here (on x64).

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
---
 libavcodec/snow.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/libavcodec/snow.h b/libavcodec/snow.h
index 82417c3324..f8f45b8763 100644
--- a/libavcodec/snow.h
+++ b/libavcodec/snow.h
@@ -92,7 +92,7 @@ typedef struct SubBand{
     int stride_line; ///< Stride measured in lines, not pixels.
     x_and_coeff * x_coeff;
     struct SubBand *parent;
-    uint8_t state[/*7*2*/ 7 + 512][32];
+    uint8_t state[34][32];
 }SubBand;
 
 typedef struct Plane{
-- 
2.52.0

_______________________________________________
ffmpeg-devel mailing list -- ffmpeg-devel@ffmpeg.org
To unsubscribe send an email to ffmpeg-devel-leave@ffmpeg.org