Hi,
Last week I worked on adding the ring buffer logic into SwsSlice and fixing
some seg faults. As I'm having some problems with git send-email through
gmail I've attached the patch.
From cb66ff7f4cfe1879fe9303a84efe49fca40c6d46 Mon Sep 17 00:00:00 2001
From: Pedro Arthur <bygran...@gmail.com>
Date: Fri, 3 Jul 2015 10:55:45 -0300
Subject: [PATCH 1/4] swscale: fix seg fault when accessing src slice
---
libswscale/slice.c | 79 ++++++++++++++++++++++++++-----------------
libswscale/swscale.c | 5 ++-
libswscale/swscale_internal.h | 3 +-
3 files changed, 54 insertions(+), 33 deletions(-)
diff --git a/libswscale/slice.c b/libswscale/slice.c
index 770b974..8cc071e 100644
--- a/libswscale/slice.c
+++ b/libswscale/slice.c
@@ -1,7 +1,7 @@
#include "swscale_internal.h"
-static int alloc_slice(SwsSlice * s, enum AVPixelFormat fmt, int lumLines, int chrLines, int h_sub_sample, int v_sub_sample)
+static int alloc_slice(SwsSlice * s, enum AVPixelFormat fmt, int lumLines, int chrLines, int h_sub_sample, int v_sub_sample, int ring)
{
int i;
int err = 0;
@@ -15,10 +15,11 @@ static int alloc_slice(SwsSlice * s, enum AVPixelFormat fmt, int lumLines, int c
s->h_chr_sub_sample = h_sub_sample;
s->v_chr_sub_sample = v_sub_sample;
s->fmt = fmt;
+ s->is_ring = ring;
for (i = 0; i < 4; ++i)
{
- s->plane[i].line = av_malloc_array(sizeof(uint8_t*), size[i]);
+ s->plane[i].line = av_malloc_array(sizeof(uint8_t*), size[i] * ( ring == 0 ? 1 : 2));
if (!s->plane[i].line)
{
err = AVERROR(ENOMEM);
@@ -45,7 +46,7 @@ static void free_slice(SwsSlice *s)
av_freep(&s->plane[i].line);
}
-int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int srcW, int lumY, int lumH, int chrY, int chrH, int skip)
+int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int srcW, int lumY, int lumH, int chrY, int chrH)
{
int i = 0;
@@ -53,30 +54,39 @@ int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int src
chrY,
chrY,
lumY};
-
- const int stride1[4] = {stride[0],
- stride[1] << skip,
- stride[2] << skip,
- stride[3]};
-
- const int height[4] = {lumH,
- chrH,
- chrH,
- lumH};
+
+ const int end[4] = {lumY +lumH,
+ chrY + chrH,
+ chrY + chrH,
+ lumY + lumH};
s->width = srcW;
for (i = 0; i < 4; ++i)
{
int j;
- int lines = height[i];
+ int lines = end[i];
lines = s->plane[i].available_lines < lines ? s->plane[i].available_lines : lines;
- s->plane[i].sliceY = start[i];
- s->plane[i].sliceH = lines;
+ if (end[i] > s->plane[i].sliceY+s->plane[i].sliceH)
+ {
+ if (start[i] <= s->plane[i].sliceY+1)
+ s->plane[i].sliceY = FFMIN(start[i], s->plane[i].sliceY);
+ else
+ s->plane[i].sliceY = start[i];
+ s->plane[i].sliceH = end[i] - s->plane[i].sliceY;
+ }
+ else
+ {
+ if (end[i] >= s->plane[i].sliceY)
+ s->plane[i].sliceH = s->plane[i].sliceY + s->plane[i].sliceH - start[i];
+ else
+ s->plane[i].sliceH = end[i] - start[i];
+ s->plane[i].sliceY = start[i];
+ }
- for (j = 0; j < lines; j+= 1 << skip)
- s->plane[i].line[j] = src[i] + (start[i] + j) * stride1[i];
+ for (j = start[i]; j < lines; j+= 1)
+ s->plane[i].line[j] = src[i] + (start[i] + j) * stride[i];
}
@@ -191,13 +201,14 @@ static int lum_convert(SwsContext *c, SwsFilterDescriptor *desc, int sliceY, int
ConvertInstance * instance = desc->instance;
uint32_t * pal = instance->pal;
- int sp = sliceY - desc->src->plane[0].sliceY;
+ int sp0 = sliceY - desc->src->plane[0].sliceY;
+ int sp1 = (sliceY >> desc->src->v_chr_sub_sample) - desc->src->plane[1].sliceY;
int dp = sliceY - desc->dst->plane[0].sliceY;
- const uint8_t * src[4] = { desc->src->plane[0].line[sp],
- desc->src->plane[1].line[sp],
- desc->src->plane[2].line[sp],
- desc->src->plane[3].line[sp]};
+ const uint8_t * src[4] = { desc->src->plane[0].line[sp0],
+ desc->src->plane[1].line[sp1],
+ desc->src->plane[2].line[sp1],
+ desc->src->plane[3].line[sp0]};
uint8_t * dst = desc->dst->plane[0].line[0/*dp*/];
desc->dst->plane[0].sliceY = sliceY;
@@ -304,13 +315,15 @@ static int chr_convert(SwsContext *c, SwsFilterDescriptor *desc, int sliceY, int
ConvertInstance * instance = desc->instance;
uint32_t * pal = instance->pal;
- int sp = sliceY - desc->src->plane[1].sliceY;
+ int sp0 = (sliceY - (desc->src->plane[0].sliceY >> desc->src->v_chr_sub_sample)) << desc->src->v_chr_sub_sample;
+ int sp1 = sliceY - desc->src->plane[1].sliceY;
int dp = sliceY - desc->dst->plane[1].sliceY;
- const uint8_t * src[4] = { desc->src->plane[0].line[sp],
- desc->src->plane[1].line[sp],
- desc->src->plane[2].line[sp],
- desc->src->plane[3].line[sp]};
+ const uint8_t * src[4] = { desc->src->plane[0].line[sp0],
+ desc->src->plane[1].line[sp1],
+ desc->src->plane[2].line[sp1],
+ desc->src->plane[3].line[sp0]};
+
uint8_t * dst1 = desc->dst->plane[1].line[0/*dp*/];
uint8_t * dst2 = desc->dst->plane[2].line[0/*dp*/];
@@ -374,6 +387,7 @@ int ff_init_filters(SwsContext * c)
int need_lum_conv = c->lumToYV12 || c->readLumPlanar || c->alpToYV12 || c->readAlpPlanar;
int need_chr_conv = c->chrToYV12 || c->readChrPlanar;
int srcIdx, dstIdx;
+ int maxSize;
uint32_t * pal = usePal(c->srcFormat) ? c->pal_yuv : (uint32_t*)c->input_rgb2yuv_table;
@@ -390,9 +404,11 @@ int ff_init_filters(SwsContext * c)
c->desc = av_malloc_array(sizeof(SwsFilterDescriptor), c->numDesc);
c->slice = av_malloc_array(sizeof(SwsSlice), c->numSlice);
- for (i = 0; i < c->numSlice-1; ++i)
- alloc_slice(&c->slice[i], c->srcFormat, c->vLumFilterSize, c->vChrFilterSize, c->chrSrcHSubSample, c->chrSrcVSubSample);
- alloc_slice(&c->slice[i], c->srcFormat, c->vLumFilterSize, c->vChrFilterSize, c->chrDstHSubSample, c->chrDstVSubSample);
+ maxSize = FFMAX(c->vLumFilterSize, c->vChrFilterSize << c->chrSrcVSubSample);
+ alloc_slice(&c->slice[0], c->srcFormat, c->srcH, c->chrSrcH, c->chrSrcHSubSample, c->chrSrcVSubSample, 0);
+ for (i = 1; i < c->numSlice-1; ++i)
+ alloc_slice(&c->slice[i], c->srcFormat, c->vLumFilterSize, c->vChrFilterSize, c->chrSrcHSubSample, c->chrSrcVSubSample, 0);
+ alloc_slice(&c->slice[i], c->srcFormat, c->vLumFilterSize, c->vChrFilterSize, c->chrDstHSubSample, c->chrDstVSubSample, 0);
index = 0;
srcIdx = 0;
@@ -450,3 +466,4 @@ int ff_free_filters(SwsContext *c)
}
return 1;
}
+
diff --git a/libswscale/swscale.c b/libswscale/swscale.c
index 4770f13..457a29a 100644
--- a/libswscale/swscale.c
+++ b/libswscale/swscale.c
@@ -451,6 +451,10 @@ static int swscale(SwsContext *c, const uint8_t *src[],
}
lastDstY = dstY;
+ ff_init_slice_from_src(src_slice, (uint8_t**)src, srcStride, c->srcW,
+ srcSliceY, srcSliceH,
+ chrSrcSliceY, chrSrcSliceH);
+
for (; dstY < dstH; dstY++) {
const int chrDstY = dstY >> c->chrDstVSubSample;
@@ -508,7 +512,6 @@ static int swscale(SwsContext *c, const uint8_t *src[],
line_pool[2] = &chrVPixBuf[chrBufIndex + 1];
line_pool[3] = alpPixBuf ? &alpPixBuf[lumBufIndex + 1] : NULL;
- ff_init_slice_from_src(src_slice, (uint8_t**)src, srcStride, c->srcW, lastInLumBuf + 1, lastLumSrcY - lastInLumBuf, lastInChrBuf + 1, lastChrSrcY - lastInChrBuf, 0);
ff_init_slice_from_lp(dst_slice, (uint8_t ***)line_pool, dstW, lastInLumBuf + 1, lastLumSrcY - lastInLumBuf, lastInChrBuf + 1, lastChrSrcY - lastInChrBuf);
#endif
diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h
index 0ba37bd..1055a73 100644
--- a/libswscale/swscale_internal.h
+++ b/libswscale/swscale_internal.h
@@ -932,6 +932,7 @@ typedef struct SwsSlice
int width;
int h_chr_sub_sample;
int v_chr_sub_sample;
+ int is_ring;
enum AVPixelFormat fmt;
SwsPlane plane[MAX_SLICE_PLANES];
} SwsSlice;
@@ -960,7 +961,7 @@ typedef struct ScaleInstance
int xInc;
} ScaleInstance;
-int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int srcW, int lumY, int lumH, int chrY, int chrH, int skip);
+int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int srcW, int lumY, int lumH, int chrY, int chrH);
int ff_init_slice_from_lp(SwsSlice *s, uint8_t ***linesPool, int dstW, int lumY, int lumH, int chrY, int chrH);
int ff_init_filters(SwsContext *c);
int ff_free_filters(SwsContext *c);
--
1.9.1
From f29b7d39493ba329bdf4583f671b7a42a36b6491 Mon Sep 17 00:00:00 2001
From: Pedro Arthur <bygran...@gmail.com>
Date: Sun, 5 Jul 2015 21:42:57 -0300
Subject: [PATCH 2/4] swscale: WIP SwsSlice ring buffer
---
libswscale/slice.c | 78 +++++++++++++++++++++++++++++++++++++++----
libswscale/swscale.c | 39 ++++++++++++++++++++--
libswscale/swscale_internal.h | 1 +
3 files changed, 109 insertions(+), 9 deletions(-)
diff --git a/libswscale/slice.c b/libswscale/slice.c
index 8cc071e..ca31df3 100644
--- a/libswscale/slice.c
+++ b/libswscale/slice.c
@@ -46,6 +46,57 @@ static void free_slice(SwsSlice *s)
av_freep(&s->plane[i].line);
}
+static int alloc_lines(SwsSlice *s, int width)
+{
+ int i;
+ for (i = 0; i < 4; ++i)
+ {
+ int n = s->plane[i].available_lines;
+ int j;
+ for (j = 0; j < n; ++j)
+ {
+ s->plane[i].line[j] = av_mallocz(width);
+ if (s->is_ring)
+ s->plane[i].line[j+n] = s->plane[i].line[j];
+ }
+ }
+ return 1;
+}
+
+int ff_rotate_slice(SwsSlice *s, int lum, int chr)
+{
+ int i;
+ if (lum)
+ {
+ for (i = 0; i < 4; i+=3)
+ {
+ int n = s->plane[i].available_lines;
+ int l = s->plane[i].sliceH;
+
+ if (l+lum >= n * 2)
+ {
+ s->plane[i].sliceY += n;
+ s->plane[i].sliceH -= n;
+ }
+ }
+ }
+ if (chr)
+ {
+ for (i = 1; i < 3; ++i)
+ {
+ int n = s->plane[i].available_lines;
+ int l = s->plane[i].sliceH;
+
+ if (l+chr >= n * 2)
+ {
+ s->plane[i].sliceY += n;
+ s->plane[i].sliceH -= n;
+ }
+ }
+ }
+ return 1;
+}
+
int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int srcW, int lumY, int lumH, int chrY, int chrH)
{
int i = 0;
@@ -171,6 +222,7 @@ static int lum_h_scale(SwsContext *c, SwsFilterDescriptor *desc, int sliceY, int
if (c->lumConvertRange)
c->lumConvertRange((int16_t*)dst[dst_pos], dstW);
+ desc->dst->plane[0].sliceH += 1;
if (desc->alpha)
{
@@ -180,7 +232,7 @@ static int lum_h_scale(SwsContext *c, SwsFilterDescriptor *desc, int sliceY, int
src_pos = sliceY - desc->src->plane[3].sliceY;
dst_pos = sliceY - desc->dst->plane[3].sliceY;
-
+ desc->dst->plane[3].sliceH += 1;
if (!c->hyscale_fast) {
c->hyScale(c, (int16_t*)dst[dst_pos], dstW, (const uint8_t *)src[src_pos], instance->filter,
@@ -306,6 +358,8 @@ static int chr_h_scale(SwsContext *c, SwsFilterDescriptor *desc, int sliceY, int
if (c->chrConvertRange)
c->chrConvertRange((uint16_t*)dst1[dst_pos1], (uint16_t*)dst2[dst_pos2], dstW);
+ desc->dst->plane[1].sliceH += 1;
+ desc->dst->plane[2].sliceH += 1;
return 1;
}
@@ -387,10 +441,13 @@ int ff_init_filters(SwsContext * c)
int need_lum_conv = c->lumToYV12 || c->readLumPlanar || c->alpToYV12 || c->readAlpPlanar;
int need_chr_conv = c->chrToYV12 || c->readChrPlanar;
int srcIdx, dstIdx;
- int maxSize;
+ int dst_stride = FFALIGN(c->dstW * sizeof(int16_t) + 66, 16);
uint32_t * pal = usePal(c->srcFormat) ? c->pal_yuv : (uint32_t*)c->input_rgb2yuv_table;
+ if (c->dstBpc == 16)
+ dst_stride <<= 1;
+
num_ydesc = need_lum_conv ? 2 : 1;
num_cdesc = c->needs_hcscale ? (need_chr_conv ? 2 : 1) : 0;
@@ -404,19 +461,23 @@ int ff_init_filters(SwsContext * c)
c->desc = av_malloc_array(sizeof(SwsFilterDescriptor), c->numDesc);
c->slice = av_malloc_array(sizeof(SwsSlice), c->numSlice);
- maxSize = FFMAX(c->vLumFilterSize, c->vChrFilterSize << c->chrSrcVSubSample);
+
alloc_slice(&c->slice[0], c->srcFormat, c->srcH, c->chrSrcH, c->chrSrcHSubSample, c->chrSrcVSubSample, 0);
for (i = 1; i < c->numSlice-1; ++i)
+ {
alloc_slice(&c->slice[i], c->srcFormat, c->vLumFilterSize, c->vChrFilterSize, c->chrSrcHSubSample, c->chrSrcVSubSample, 0);
- alloc_slice(&c->slice[i], c->srcFormat, c->vLumFilterSize, c->vChrFilterSize, c->chrDstHSubSample, c->chrDstVSubSample, 0);
+ alloc_lines(&c->slice[i], FFALIGN(c->srcW*2+78, 16));
+ }
+ alloc_slice(&c->slice[i], c->srcFormat, c->vLumFilterSize, c->vChrFilterSize, c->chrDstHSubSample, c->chrDstVSubSample, 1);
+ alloc_lines(&c->slice[i], dst_stride);
index = 0;
srcIdx = 0;
dstIdx = 1;
// temp slice for color space conversion
- if (need_lum_conv || need_chr_conv)
- init_slice_1(&c->slice[dstIdx], c->formatConvBuffer, (c->formatConvBuffer + FFALIGN(c->srcW*2+78, 16)), c->srcW, 0, c->vLumFilterSize);
+ //if (need_lum_conv || need_chr_conv)
+ // init_slice_1(&c->slice[dstIdx], c->formatConvBuffer, (c->formatConvBuffer + FFALIGN(c->srcW*2+78, 16)), c->srcW, 0, c->vLumFilterSize);
if (need_lum_conv)
{
@@ -467,3 +528,8 @@ int ff_free_filters(SwsContext *c)
return 1;
}
+
+
+
+
+
diff --git a/libswscale/swscale.c b/libswscale/swscale.c
index 457a29a..7d6c3dd 100644
--- a/libswscale/swscale.c
+++ b/libswscale/swscale.c
@@ -455,6 +455,16 @@ static int swscale(SwsContext *c, const uint8_t *src[],
srcSliceY, srcSliceH,
chrSrcSliceY, chrSrcSliceH);
+ dst_slice->plane[0].sliceY = lastInLumBuf + 1;
+ dst_slice->plane[1].sliceY = lastInChrBuf + 1;
+ dst_slice->plane[2].sliceY = lastInChrBuf + 1;
+ dst_slice->plane[3].sliceY = lastInLumBuf + 1;
+
+ dst_slice->plane[0].sliceH =
+ dst_slice->plane[1].sliceH =
+ dst_slice->plane[2].sliceH =
+ dst_slice->plane[3].sliceH = 0;
+ dst_slice->width = dstW;
for (; dstY < dstH; dstY++) {
const int chrDstY = dstY >> c->chrDstVSubSample;
@@ -479,10 +489,20 @@ static int swscale(SwsContext *c, const uint8_t *src[],
int enough_lines;
// handle holes (FAST_BILINEAR & weird filters)
- if (firstLumSrcY > lastInLumBuf)
+ if (firstLumSrcY > lastInLumBuf) {
lastInLumBuf = firstLumSrcY - 1;
- if (firstChrSrcY > lastInChrBuf)
+ dst_slice->plane[0].sliceY = lastInLumBuf + 1;
+ dst_slice->plane[3].sliceY = lastInLumBuf + 1;
+ dst_slice->plane[0].sliceH =
+ dst_slice->plane[3].sliceH = 0;
+ }
+ if (firstChrSrcY > lastInChrBuf) {
lastInChrBuf = firstChrSrcY - 1;
+ dst_slice->plane[1].sliceY = lastInChrBuf + 1;
+ dst_slice->plane[2].sliceY = lastInChrBuf + 1;
+ dst_slice->plane[1].sliceH =
+ dst_slice->plane[2].sliceH = 0;
+ }
av_assert0(firstLumSrcY >= lastInLumBuf - vLumBufSize + 1);
av_assert0(firstChrSrcY >= lastInChrBuf - vChrBufSize + 1);
@@ -506,7 +526,7 @@ static int swscale(SwsContext *c, const uint8_t *src[],
#define NEW_FILTER 1
-#if NEW_FILTER
+#if 0 //NEW_FILTER
line_pool[0] = &lumPixBuf[lumBufIndex + 1];
line_pool[1] = &chrUPixBuf[chrBufIndex + 1];
line_pool[2] = &chrVPixBuf[chrBufIndex + 1];
@@ -515,6 +535,7 @@ static int swscale(SwsContext *c, const uint8_t *src[],
ff_init_slice_from_lp(dst_slice, (uint8_t ***)line_pool, dstW, lastInLumBuf + 1, lastLumSrcY - lastInLumBuf, lastInChrBuf + 1, lastChrSrcY - lastInChrBuf);
#endif
+ ff_rotate_slice(dst_slice, lastLumSrcY - lastInLumBuf, lastChrSrcY - lastInChrBuf);
// Do horizontal scaling
while (lastInLumBuf < lastLumSrcY) {
const uint8_t *src1[4] = {
@@ -599,11 +620,23 @@ static int swscale(SwsContext *c, const uint8_t *src[],
}
{
+
+#if NEW_FILTER
+ const int16_t **lumSrcPtr = (const int16_t **)(void*) dst_slice->plane[0].line + dst_slice->plane[0].sliceH - vLumFilterSize;
+ const int16_t **chrUSrcPtr = (const int16_t **)(void*) dst_slice->plane[1].line + dst_slice->plane[1].sliceH - vChrFilterSize;
+ const int16_t **chrVSrcPtr = (const int16_t **)(void*) dst_slice->plane[2].line + dst_slice->plane[2].sliceH - vChrFilterSize;
+ const int16_t **alpSrcPtr = (CONFIG_SWSCALE_ALPHA && alpPixBuf) ?
+ (const int16_t **)(void*) dst_slice->plane[3].line + dst_slice->plane[3].sliceH - vLumFilterSize : NULL;
+#else
const int16_t **lumSrcPtr = (const int16_t **)(void*) lumPixBuf + lumBufIndex + firstLumSrcY - lastInLumBuf + vLumBufSize;
const int16_t **chrUSrcPtr = (const int16_t **)(void*) chrUPixBuf + chrBufIndex + firstChrSrcY - lastInChrBuf + vChrBufSize;
const int16_t **chrVSrcPtr = (const int16_t **)(void*) chrVPixBuf + chrBufIndex + firstChrSrcY - lastInChrBuf + vChrBufSize;
const int16_t **alpSrcPtr = (CONFIG_SWSCALE_ALPHA && alpPixBuf) ?
(const int16_t **)(void*) alpPixBuf + lumBufIndex + firstLumSrcY - lastInLumBuf + vLumBufSize : NULL;
+#endif
+
+
+
int16_t *vLumFilter = c->vLumFilter;
int16_t *vChrFilter = c->vChrFilter;
diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h
index 1055a73..d832343 100644
--- a/libswscale/swscale_internal.h
+++ b/libswscale/swscale_internal.h
@@ -965,5 +965,6 @@ int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int src
int ff_init_slice_from_lp(SwsSlice *s, uint8_t ***linesPool, int dstW, int lumY, int lumH, int chrY, int chrH);
int ff_init_filters(SwsContext *c);
int ff_free_filters(SwsContext *c);
+int ff_rotate_slice(SwsSlice *s, int lum, int chr);
#endif /* SWSCALE_SWSCALE_INTERNAL_H */
--
1.9.1
From d1405d52dc93217e490e15405ed986d3fb0d181a Mon Sep 17 00:00:00 2001
From: Pedro Arthur <bygran...@gmail.com>
Date: Wed, 8 Jul 2015 17:04:34 -0300
Subject: [PATCH 3/4] swscale: fix seg fault when dst format is grayscale
---
libswscale/slice.c | 50 +++++++++++++++++++++++++++++++++++++++++++++++---
1 file changed, 47 insertions(+), 3 deletions(-)
diff --git a/libswscale/slice.c b/libswscale/slice.c
index ca31df3..9286ac2 100644
--- a/libswscale/slice.c
+++ b/libswscale/slice.c
@@ -432,6 +432,47 @@ static int init_desc_chscale(SwsFilterDescriptor *desc, SwsSlice *src, SwsSlice
return 1;
}
+static void fill_ones(SwsSlice *s, int n, int is16bit)
+{
+ int i;
+ for (i = 0; i < 4; ++i)
+ {
+ int j;
+ int size = s->plane[i].available_lines;
+ for (int j = 0; j < size; ++j)
+ {
+ int k;
+ int end = is16bit ? n>>1: n;
+
+ if (is16bit)
+ for (k = 0; k < end; ++k)
+ ((int32_t*)(s->plane[i].line[j]))[k] = 1<<18;
+ else
+ for (k = 0; k < end; ++k)
+ ((int16_t*)(s->plane[i].line[j]))[k] = 1<<14;
+ }
+ }
+}
+
+static int no_chr_scale(SwsContext *c, SwsFilterDescriptor *desc, int sliceY, int sliceH)
+{
+ desc->dst->plane[1].sliceY = sliceY + sliceH - desc->dst->plane[1].available_lines;
+ desc->dst->plane[1].sliceH = desc->dst->plane[1].available_lines;
+ desc->dst->plane[2].sliceY = sliceY + sliceH - desc->dst->plane[2].available_lines;
+ desc->dst->plane[2].sliceH = desc->dst->plane[2].available_lines;
+ return 0;
+}
+
+static int init_desc_no_chr(SwsFilterDescriptor *desc, SwsSlice * src, SwsSlice *dst)
+{
+ desc->src = src;
+ desc->dst = dst;
+ desc->alpha = 0;
+ desc->instance = NULL;
+ desc->process = &no_chr_scale;
+ return 0;
+}
+
int ff_init_filters(SwsContext * c)
{
int i;
@@ -449,7 +490,7 @@ int ff_init_filters(SwsContext * c)
dst_stride <<= 1;
num_ydesc = need_lum_conv ? 2 : 1;
- num_cdesc = c->needs_hcscale ? (need_chr_conv ? 2 : 1) : 0;
+ num_cdesc = /*c->needs_hcscale ? */(need_chr_conv ? 2 : 1)/* : 0*/;
c->numSlice = FFMAX(num_ydesc, num_cdesc) + 1;
c->numDesc = num_ydesc + num_cdesc;
@@ -470,6 +511,7 @@ int ff_init_filters(SwsContext * c)
}
alloc_slice(&c->slice[i], c->srcFormat, c->vLumFilterSize, c->vChrFilterSize, c->chrDstHSubSample, c->chrDstVSubSample, 1);
alloc_lines(&c->slice[i], dst_stride);
+ fill_ones(&c->slice[i], dst_stride>>1, c->dstBpc == 16);
index = 0;
srcIdx = 0;
@@ -494,7 +536,6 @@ int ff_init_filters(SwsContext * c)
++index;
- if (c->needs_hcscale)
{
srcIdx = 0;
dstIdx = 1;
@@ -506,7 +547,10 @@ int ff_init_filters(SwsContext * c)
}
dstIdx = FFMAX(num_ydesc, num_cdesc);
- init_desc_chscale(&c->desc[index], &c->slice[srcIdx], &c->slice[dstIdx], c->hChrFilter, c->hChrFilterPos, c->hChrFilterSize, c->chrXInc);
+ if (c->needs_hcscale)
+ init_desc_chscale(&c->desc[index], &c->slice[srcIdx], &c->slice[dstIdx], c->hChrFilter, c->hChrFilterPos, c->hChrFilterSize, c->chrXInc);
+ else
+ init_desc_no_chr(&c->desc[index], &c->slice[srcIdx], &c->slice[dstIdx]);
}
return 1;
--
1.9.1
From 02480244b886e82ff421c4dd0889a86171babb6e Mon Sep 17 00:00:00 2001
From: Pedro Arthur <bygran...@gmail.com>
Date: Wed, 8 Jul 2015 22:05:21 -0300
Subject: [PATCH 4/4] swscale: added slice lines freeing functions
+removed unused functions
---
libswscale/slice.c | 122 +++++++++++++++++-------------------------
libswscale/swscale.c | 10 ----
libswscale/swscale_internal.h | 1 +
3 files changed, 51 insertions(+), 82 deletions(-)
diff --git a/libswscale/slice.c b/libswscale/slice.c
index 9286ac2..04633e3 100644
--- a/libswscale/slice.c
+++ b/libswscale/slice.c
@@ -1,7 +1,47 @@
#include "swscale_internal.h"
+static void free_lines(SwsSlice *s)
+{
+ int i;
+ for (i = 0; i < 4; ++i)
+ {
+ int n = s->plane[i].available_lines;
+ int j;
+ for (j = 0; j < n; ++j)
+ {
+ av_freep(&s->plane[i].line[j]);
+ if (s->is_ring)
+ s->plane[i].line[j+n] = NULL;
+ }
+ }
+ s->should_free_lines = 0;
+}
+
+static int alloc_lines(SwsSlice *s, int width)
+{
+ int i;
+ s->should_free_lines = 1;
+
+ for (i = 0; i < 4; ++i)
+ {
+ int n = s->plane[i].available_lines;
+ int j;
+ for (j = 0; j < n; ++j)
+ {
+ s->plane[i].line[j] = av_mallocz(width);
+ if (!s->plane[i].line[j])
+ {
+ free_lines(s);
+ return AVERROR(ENOMEM);
+ }
+ if (s->is_ring)
+ s->plane[i].line[j+n] = s->plane[i].line[j];
+ }
+ }
+ return 1;
+}
-static int alloc_slice(SwsSlice * s, enum AVPixelFormat fmt, int lumLines, int chrLines, int h_sub_sample, int v_sub_sample, int ring)
+static int alloc_slice(SwsSlice *s, enum AVPixelFormat fmt, int lumLines, int chrLines, int h_sub_sample, int v_sub_sample, int ring)
{
int i;
int err = 0;
@@ -16,15 +56,21 @@ static int alloc_slice(SwsSlice * s, enum AVPixelFormat fmt, int lumLines, int c
s->v_chr_sub_sample = v_sub_sample;
s->fmt = fmt;
s->is_ring = ring;
+ s->should_free_lines = 0;
for (i = 0; i < 4; ++i)
{
- s->plane[i].line = av_malloc_array(sizeof(uint8_t*), size[i] * ( ring == 0 ? 1 : 2));
+ int j;
+ int n = size[i] * ( ring == 0 ? 1 : 2);
+ s->plane[i].line = av_malloc_array(sizeof(uint8_t*), n);
if (!s->plane[i].line)
{
err = AVERROR(ENOMEM);
break;
}
+ for (int j = 0; j < n; ++j)
+ s->plane[i].line[j] = NULL;
+
s->plane[i].available_lines = size[i];
s->plane[i].sliceY = 0;
s->plane[i].sliceH = 0;
@@ -42,27 +88,12 @@ static int alloc_slice(SwsSlice * s, enum AVPixelFormat fmt, int lumLines, int c
static void free_slice(SwsSlice *s)
{
int i;
+ if (s->should_free_lines)
+ free_lines(s);
for (i = 0; i < 4; ++i)
av_freep(&s->plane[i].line);
}
-static int alloc_lines(SwsSlice *s, int width)
-{
- int i;
- for (i = 0; i < 4; ++i)
- {
- int n = s->plane[i].available_lines;
- int j;
- for (j = 0; j < n; ++j)
- {
- s->plane[i].line[j] = av_mallocz(width);
- if (s->is_ring)
- s->plane[i].line[j+n] = s->plane[i].line[j];
- }
- }
- return 1;
-}
-
int ff_rotate_slice(SwsSlice *s, int lum, int chr)
{
int i;
@@ -144,59 +175,6 @@ int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int src
return 1;
}
-int ff_init_slice_from_lp(SwsSlice *s, uint8_t ***linesPool, int dstW, int lumY, int lumH, int chrY, int chrH)
-{
- int i;
- const int start[4] = {lumY,
- chrY,
- chrY,
- lumY};
-
- const int height[4] = {lumH,
- chrH,
- chrH,
- lumH};
- s->width = dstW;
- for (i = 0; i < 4; ++i)
- {
- int j;
- int lines = height[i];
- lines = s->plane[i].available_lines < lines ? s->plane[i].available_lines : lines;
-
- s->plane[i].sliceY = start[i];
- s->plane[i].sliceH = lines;
-
- for (j = 0; j < lines; ++j)
- {
- uint8_t * v = linesPool[i] ? linesPool[i][j] : NULL;
- s->plane[i].line[j] = v;
- }
-
- }
- return 1;
-}
-
-static int init_slice_1(SwsSlice *s, uint8_t *v, uint8_t *v2, int dstW, int sliceY, int sliceH)
-{
- int i;
- uint8_t *ptr[4] = {v, v, v2, v2};
- s->width = dstW;
- for (i = 0; i < 4; ++i)
- {
- int j;
- int lines = s->plane[i].available_lines;
-
- s->plane[i].sliceY = sliceY;
- s->plane[i].sliceH = lines;
-
- for (j = 0; j < lines; ++j)
- s->plane[i].line[j] = ptr[i];
-
- }
- return 1;
-}
-
-
static int lum_h_scale(SwsContext *c, SwsFilterDescriptor *desc, int sliceY, int sliceH)
{
ScaleInstance *instance = desc->instance;
diff --git a/libswscale/swscale.c b/libswscale/swscale.c
index 7d6c3dd..5facb99 100644
--- a/libswscale/swscale.c
+++ b/libswscale/swscale.c
@@ -525,16 +525,6 @@ static int swscale(SwsContext *c, const uint8_t *src[],
#define NEW_FILTER 1
-
-#if 0 //NEW_FILTER
- line_pool[0] = &lumPixBuf[lumBufIndex + 1];
- line_pool[1] = &chrUPixBuf[chrBufIndex + 1];
- line_pool[2] = &chrVPixBuf[chrBufIndex + 1];
- line_pool[3] = alpPixBuf ? &alpPixBuf[lumBufIndex + 1] : NULL;
-
- ff_init_slice_from_lp(dst_slice, (uint8_t ***)line_pool, dstW, lastInLumBuf + 1, lastLumSrcY - lastInLumBuf, lastInChrBuf + 1, lastChrSrcY - lastInChrBuf);
-
-#endif
ff_rotate_slice(dst_slice, lastLumSrcY - lastInLumBuf, lastChrSrcY - lastInChrBuf);
// Do horizontal scaling
while (lastInLumBuf < lastLumSrcY) {
diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h
index d832343..82713fd 100644
--- a/libswscale/swscale_internal.h
+++ b/libswscale/swscale_internal.h
@@ -933,6 +933,7 @@ typedef struct SwsSlice
int h_chr_sub_sample;
int v_chr_sub_sample;
int is_ring;
+ int should_free_lines;
enum AVPixelFormat fmt;
SwsPlane plane[MAX_SLICE_PLANES];
} SwsSlice;
--
1.9.1
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
