From 3c0df27067d0e0758fc72bbe31c5cab7eb2c8ed2 Mon Sep 17 00:00:00 2001
From: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Date: Tue, 25 Mar 2025 18:00:25 +0100
Subject: [PATCH 04/11] avcodec/mpegvideo_enc: Don't use unnecessarily much
stack
encode_thread() puts two MPVEncContexts (2*6516B here)
on the stack and zeroes one of them in order to
temporarily store the variables that get changed
during encoding a macroblock (when there is more than
one candidate type for a macroblock). This is wasteful
and therefore this commit adds a small (328B here) structure
to store exactly the fields that actually need to be backed
up. Then one can extend MPVEncContext without fearing
too use up to much stack.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
---
libavcodec/mpegvideo_enc.c | 158 +++++++++++++++++++++----------------
1 file changed, 89 insertions(+), 69 deletions(-)
diff --git a/libavcodec/mpegvideo_enc.c b/libavcodec/mpegvideo_enc.c
index 16086bf067..fb3ad2e25c 100644
--- a/libavcodec/mpegvideo_enc.c
+++ b/libavcodec/mpegvideo_enc.c
@@ -2631,80 +2631,100 @@ static void encode_mb(MPVEncContext *const s, int motion_x, int motion_y)
encode_mb_internal(s, motion_x, motion_y, 16, 16, 12, 0, 0, CHROMA_444);
}
-static inline void copy_context_before_encode(MPVEncContext *const d,
- const MPVEncContext *const s)
-{
- int i;
-
- memcpy(d->c.last_mv, s->c.last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster than a loop?
-
- /* MPEG-1 */
- d->c.mb_skip_run = s->c.mb_skip_run;
- for(i=0; i<3; i++)
- d->c.last_dc[i] = s->c.last_dc[i];
-
- /* statistics */
- d->mv_bits= s->mv_bits;
- d->i_tex_bits= s->i_tex_bits;
- d->p_tex_bits= s->p_tex_bits;
- d->i_count= s->i_count;
- d->misc_bits= s->misc_bits;
- d->last_bits= 0;
-
- d->c.mb_skipped = 0;
- d->c.qscale = s->c.qscale;
- d->dquant= s->dquant;
-
- d->esc3_level_length= s->esc3_level_length;
+typedef struct MBBackup {
+ struct {
+ int mv[2][4][2];
+ int last_mv[2][2][2];
+ int mv_type, mv_dir;
+ int last_dc[3];
+ int mb_intra, mb_skipped, mb_skip_run;
+ int qscale;
+ int block_last_index[8];
+ int interlaced_dct;
+ int16_t (*block)[64];
+ } c;
+ int mv_bits, i_tex_bits, p_tex_bits, i_count, misc_bits, last_bits;
+ int dquant;
+ int esc3_level_length;
+ PutBitContext pb, pb2, tex_pb;
+} MBBackup;
+
+#define COPY_CONTEXT(BEFORE, AFTER, DST_TYPE, SRC_TYPE) \
+static inline void BEFORE ##_context_before_encode(DST_TYPE *const d, \
+ const SRC_TYPE *const s) \
+{ \
+ /* FIXME is memcpy faster than a loop? */ \
+ memcpy(d->c.last_mv, s->c.last_mv, 2*2*2*sizeof(int)); \
+ \
+ /* MPEG-1 */ \
+ d->c.mb_skip_run = s->c.mb_skip_run; \
+ for (int i = 0; i < 3; i++) \
+ d->c.last_dc[i] = s->c.last_dc[i]; \
+ \
+ /* statistics */ \
+ d->mv_bits = s->mv_bits; \
+ d->i_tex_bits = s->i_tex_bits; \
+ d->p_tex_bits = s->p_tex_bits; \
+ d->i_count = s->i_count; \
+ d->misc_bits = s->misc_bits; \
+ d->last_bits = 0; \
+ \
+ d->c.mb_skipped = 0; \
+ d->c.qscale = s->c.qscale; \
+ d->dquant = s->dquant; \
+ \
+ d->esc3_level_length = s->esc3_level_length; \
+} \
+ \
+static inline void AFTER ## _context_after_encode(DST_TYPE *const d, \
+ const SRC_TYPE *const s, \
+ int data_partitioning) \
+{ \
+ /* FIXME is memcpy faster than a loop? */ \
+ memcpy(d->c.mv, s->c.mv, 2*4*2*sizeof(int)); \
+ memcpy(d->c.last_mv, s->c.last_mv, 2*2*2*sizeof(int)); \
+ \
+ /* MPEG-1 */ \
+ d->c.mb_skip_run = s->c.mb_skip_run; \
+ for (int i = 0; i < 3; i++) \
+ d->c.last_dc[i] = s->c.last_dc[i]; \
+ \
+ /* statistics */ \
+ d->mv_bits = s->mv_bits; \
+ d->i_tex_bits = s->i_tex_bits; \
+ d->p_tex_bits = s->p_tex_bits; \
+ d->i_count = s->i_count; \
+ d->misc_bits = s->misc_bits; \
+ \
+ d->c.mb_intra = s->c.mb_intra; \
+ d->c.mb_skipped = s->c.mb_skipped; \
+ d->c.mv_type = s->c.mv_type; \
+ d->c.mv_dir = s->c.mv_dir; \
+ d->pb = s->pb; \
+ if (data_partitioning) { \
+ d->pb2 = s->pb2; \
+ d->tex_pb = s->tex_pb; \
+ } \
+ d->c.block = s->c.block; \
+ for (int i = 0; i < 8; i++) \
+ d->c.block_last_index[i] = s->c.block_last_index[i]; \
+ d->c.interlaced_dct = s->c.interlaced_dct; \
+ d->c.qscale = s->c.qscale; \
+ \
+ d->esc3_level_length = s->esc3_level_length; \
}
-static inline void copy_context_after_encode(MPVEncContext *const d,
- const MPVEncContext *const s,
- int data_partitioning)
-{
- int i;
-
- memcpy(d->c.mv, s->c.mv, 2*4*2*sizeof(int));
- memcpy(d->c.last_mv, s->c.last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster than a loop?
-
- /* MPEG-1 */
- d->c.mb_skip_run = s->c.mb_skip_run;
- for(i=0; i<3; i++)
- d->c.last_dc[i] = s->c.last_dc[i];
-
- /* statistics */
- d->mv_bits= s->mv_bits;
- d->i_tex_bits= s->i_tex_bits;
- d->p_tex_bits= s->p_tex_bits;
- d->i_count= s->i_count;
- d->misc_bits= s->misc_bits;
-
- d->c.mb_intra = s->c.mb_intra;
- d->c.mb_skipped = s->c.mb_skipped;
- d->c.mv_type = s->c.mv_type;
- d->c.mv_dir = s->c.mv_dir;
- d->pb= s->pb;
- if (data_partitioning) {
- d->pb2= s->pb2;
- d->tex_pb= s->tex_pb;
- }
- d->c.block = s->c.block;
- for(i=0; i<8; i++)
- d->c.block_last_index[i] = s->c.block_last_index[i];
- d->c.interlaced_dct = s->c.interlaced_dct;
- d->c.qscale = s->c.qscale;
-
- d->esc3_level_length= s->esc3_level_length;
-}
+COPY_CONTEXT(backup, save, MBBackup, MPVEncContext)
+COPY_CONTEXT(reset, store, MPVEncContext, MBBackup)
-static void encode_mb_hq(MPVEncContext *const s, MPVEncContext *const backup, MPVEncContext *const best,
+static void encode_mb_hq(MPVEncContext *const s, MBBackup *const backup, MBBackup *const best,
PutBitContext pb[2], PutBitContext pb2[2], PutBitContext tex_pb[2],
int *dmin, int *next_block, int motion_x, int motion_y)
{
int score;
uint8_t *dest_backup[3];
- copy_context_before_encode(s, backup);
+ reset_context_before_encode(s, backup);
s->c.block = s->c.blocks[*next_block];
s->pb = pb[*next_block];
@@ -2744,7 +2764,7 @@ static void encode_mb_hq(MPVEncContext *const s, MPVEncContext *const backup, MP
*dmin= score;
*next_block^=1;
- copy_context_after_encode(best, s, s->c.data_partitioning);
+ save_context_after_encode(best, s, s->c.data_partitioning);
}
}
@@ -2962,7 +2982,7 @@ static int encode_thread(AVCodecContext *c, void *arg){
MPVEncContext *const s = *(void**)arg;
int chr_h = 16 >> s->c.chroma_y_shift;
int i;
- MPVEncContext best_s = { 0 }, backup_s;
+ MBBackup best_s = { 0 }, backup_s;
uint8_t bit_buf[2][MAX_MB_BYTES];
uint8_t bit_buf2[2][MAX_MB_BYTES];
uint8_t bit_buf_tex[2][MAX_MB_BYTES];
@@ -3163,7 +3183,7 @@ static int encode_thread(AVCodecContext *c, void *arg){
int next_block=0;
int pb_bits_count, pb2_bits_count, tex_pb_bits_count;
- copy_context_before_encode(&backup_s, s);
+ backup_context_before_encode(&backup_s, s);
backup_s.pb= s->pb;
if (s->c.data_partitioning) {
backup_s.pb2= s->pb2;
@@ -3388,7 +3408,7 @@ static int encode_thread(AVCodecContext *c, void *arg){
}
}
- copy_context_after_encode(s, &best_s, s->c.data_partitioning);
+ store_context_after_encode(s, &best_s, s->c.data_partitioning);
pb_bits_count= put_bits_count(&s->pb);
flush_put_bits(&s->pb);
--
2.45.2