ref: 36647aaa1aec66983d2d470c3346e0102a5a240b
parent: 76646c7d94199f4e15dddd4c7e6151dc51103ed7
author: Ronald S. Bultje <rsbultje@gmail.com>
date: Fri Oct 5 13:44:18 EDT 2018
Vertical loopfilter simplifications - remove edge condition in maxifzero(), since it's always true; - make subsampled level array packed; - handle a SB worth of 4px blocks per DSP function call, so that SIMD will be more efficient.
--- a/src/lf_apply.c
+++ b/src/lf_apply.c
@@ -34,12 +34,11 @@
#include "src/lf_apply.h"
-static inline int maxifzero(const uint8_t (*const a)[4], const int have_b,
+static inline int maxifzero(const uint8_t (*const a)[4],
const uint8_t (*const b)[4], const int diridx)
{const int a_val = (*a)[diridx];
if (a_val) return a_val;
- if (!have_b) return a_val;
return (*b)[diridx];
}
@@ -64,7 +63,7 @@
for (unsigned x = 1; hm & ~(x - 1); l++, x <<= 1, ptr += 4) { if ((have_left || x > 1) && (hm & x)) {- const int L = maxifzero(l, have_left || x > 1, &l[-1], 0);
+ const int L = maxifzero(l, &l[-1], 0);
if (!L) continue;
const int H = L >> 4;
const int E = f->lf.lim_lut.e[L], I = f->lf.lim_lut.i[L];
@@ -82,6 +81,7 @@
const ptrdiff_t b4_stride,
const uint32_t (*const mask)[3],
pixel *dst, const ptrdiff_t ls,
+ const int w,
const int starty4, const int endy4)
{const Dav1dDSPContext *const dsp = f->dsp;
@@ -92,22 +92,10 @@
for (int y = starty4; y < endy4;
y++, dst += 4 * PXSTRIDE(ls), lvl += b4_stride)
{- pixel *ptr = dst;
- const uint8_t (*l)[4] = lvl;
- const uint32_t *const vmask = mask[y];
- const unsigned vm = vmask[0] | vmask[1] | vmask[2];
-
- for (unsigned x = 1; vm & ~(x - 1); x <<= 1, ptr += 4, l++) {- if ((have_top || y) && (vm & x)) {- const int L = maxifzero(l, have_top || y, &l[-b4_stride], 1);
- if (!L) continue;
- const int H = L >> 4;
- const int E = f->lf.lim_lut.e[L], I = f->lf.lim_lut.i[L];
- const int idx = (vmask[2] & x) ? 2 : !!(vmask[1] & x);
-
- dsp->lf.loop_filter[idx][1](ptr, ls, E, I, H);
- }
- }
+ if (!have_top && !y) continue;
+ dsp->lf.loop_filter_sb128y(dst, ls, mask[y],
+ (const uint8_t(*)[4]) &lvl[0][1], b4_stride,
+ &f->lf.lim_lut, w);
}
}
@@ -125,12 +113,10 @@
ptrdiff_t off_l;
const int ss_ver = f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
const int ss_hor = f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
- const int hstep = 1 << ss_hor;
// filter edges between columns (e.g. block1 | block2)
- lvl += ss_hor + ss_ver * b4_stride;
for (off_l = 0, y = starty4; y < endy4;
- y++, off_l += 4 * PXSTRIDE(ls), lvl += b4_stride << ss_ver)
+ y++, off_l += 4 * PXSTRIDE(ls), lvl += b4_stride)
{ptrdiff_t off = off_l;
const uint8_t (*l)[4] = lvl;
@@ -137,11 +123,11 @@
const uint32_t *const hmask = mask[y];
const unsigned hm = hmask[0] | hmask[1];
- for (unsigned x = 1; hm & ~(x - 1); l += hstep, x <<= 1, off += 4) {+ for (unsigned x = 1; hm & ~(x - 1); l++, x <<= 1, off += 4) { if ((have_left || x > 1) && (hm & x)) {const int idx = !!(hmask[1] & x);
- const int Lu = maxifzero(l, have_left || x > 1, &l[-hstep], 2);
+ const int Lu = maxifzero(l, &l[-1], 2);
if (Lu) {const int H = Lu >> 4;
const int E = f->lf.lim_lut.e[Lu], I = f->lf.lim_lut.i[Lu];
@@ -149,7 +135,7 @@
dsp->lf.loop_filter_uv[idx][0](&u[off], ls, E, I, H);
}
- const int Lv = maxifzero(l, have_left || x > 1, &l[-hstep], 3);
+ const int Lv = maxifzero(l, &l[-1], 3);
if (Lv) {const int H = Lv >> 4;
const int E = f->lf.lim_lut.e[Lv], I = f->lf.lim_lut.i[Lv];
@@ -167,51 +153,26 @@
const ptrdiff_t b4_stride,
const uint32_t (*const mask)[2],
pixel *const u, pixel *const v,
- const ptrdiff_t ls,
+ const ptrdiff_t ls, const int w,
const int starty4, const int endy4)
{const Dav1dDSPContext *const dsp = f->dsp;
int y;
ptrdiff_t off_l;
- const int ss_ver = f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
- const int ss_hor = f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
- const int hstep = 1 << ss_hor;
// block1
// filter edges between rows (e.g. ------)
// block2
- lvl += ss_ver * b4_stride + ss_hor;
for (off_l = 0, y = starty4; y < endy4;
- y++, off_l += 4 * PXSTRIDE(ls), lvl += b4_stride << ss_ver)
+ y++, off_l += 4 * PXSTRIDE(ls), lvl += b4_stride)
{- ptrdiff_t off = off_l;
- const uint8_t (*l)[4] = lvl;
- const uint32_t *const vmask = mask[y];
- const unsigned vm = vmask[0] | vmask[1];
-
- for (unsigned x = 1; vm & ~(x - 1); x <<= 1, off += 4, l += hstep) {- if ((have_top || y) && (vm & x)) {- const int idx = !!(vmask[1] & x);
-
- const int Lu = maxifzero(l, have_top || y,
- &l[-(b4_stride << ss_ver)], 2);
- if (Lu) {- const int H = Lu >> 4;
- const int E = f->lf.lim_lut.e[Lu], I = f->lf.lim_lut.i[Lu];
-
- dsp->lf.loop_filter_uv[idx][1](&u[off], ls, E, I, H);
- }
-
- const int Lv = maxifzero(l, have_top || y,
- &l[-(b4_stride << ss_ver)], 3);
- if (Lv) {- const int H = Lv >> 4;
- const int E = f->lf.lim_lut.e[Lv], I = f->lf.lim_lut.i[Lv];
-
- dsp->lf.loop_filter_uv[idx][1](&v[off], ls, E, I, H);
- }
- }
- }
+ if (!have_top && !y) continue;
+ dsp->lf.loop_filter_sb128uv(&u[off_l], ls, mask[y],
+ (const uint8_t(*)[4]) &lvl[0][2], b4_stride,
+ &f->lf.lim_lut, w);
+ dsp->lf.loop_filter_sb128uv(&v[off_l], ls, mask[y],
+ (const uint8_t(*)[4]) &lvl[0][3], b4_stride,
+ &f->lf.lim_lut, w);
}
}
@@ -303,8 +264,8 @@
level_ptr = f->lf.level + f->b4_stride * sby * sbsz;
for (ptr = p[0], x = 0; x < f->sb128w; x++, ptr += 128, level_ptr += 32) {filter_plane_rows_y(f, have_top, level_ptr, f->b4_stride,
- lflvl[x].filter_y[1],
- ptr, f->cur.p.stride[0], starty4, endy4);
+ lflvl[x].filter_y[1], ptr, f->cur.p.stride[0],
+ imin(32, f->bw - x * 32), starty4, endy4);
}
if (!f->frame_hdr.loopfilter.level_u && !f->frame_hdr.loopfilter.level_v)
@@ -311,9 +272,9 @@
return;
ptrdiff_t uv_off;
- level_ptr = f->lf.level + f->b4_stride * sby * sbsz;
+ level_ptr = f->lf.level + f->b4_stride * (sby * sbsz >> ss_ver);
for (uv_off = 0, have_left = 0, x = 0; x < f->sb128w;
- x++, have_left = 1, uv_off += 128 >> ss_hor, level_ptr += 32)
+ x++, have_left = 1, uv_off += 128 >> ss_hor, level_ptr += 32 >> ss_hor)
{filter_plane_cols_uv(f, have_left, level_ptr, f->b4_stride,
lflvl[x].filter_uv[0],
@@ -321,13 +282,14 @@
starty4 >> ss_ver, uv_endy4);
}
- level_ptr = f->lf.level + f->b4_stride * sby * sbsz;
+ level_ptr = f->lf.level + f->b4_stride * (sby * sbsz >> ss_ver);
for (uv_off = 0, x = 0; x < f->sb128w;
- x++, uv_off += 128 >> ss_hor, level_ptr += 32)
+ x++, uv_off += 128 >> ss_hor, level_ptr += 32 >> ss_hor)
{filter_plane_rows_uv(f, have_top, level_ptr, f->b4_stride,
lflvl[x].filter_uv[1],
&p[1][uv_off], &p[2][uv_off], f->cur.p.stride[1],
+ (imin(32, f->bw - x * 32) + ss_hor) >> ss_hor,
starty4 >> ss_ver, uv_endy4);
}
}
--- a/src/lf_mask.c
+++ b/src/lf_mask.c
@@ -233,7 +233,7 @@
}
void dav1d_create_lf_mask_intra(Av1Filter *const lflvl,
- uint8_t (*level_cache)[4],
+ uint8_t (*const level_cache)[4],
const ptrdiff_t b4_stride,
const Av1FrameHeader *const hdr,
const uint8_t (*filter_level)[8][2],
@@ -255,15 +255,13 @@
const int bx4 = bx & 31;
const int by4 = by & 31;
- level_cache += by * b4_stride + bx;
+ uint8_t (*level_cache_ptr)[4] = level_cache + by * b4_stride + bx;
for (int y = 0; y < bh4; y++) { for (int x = 0; x < bw4; x++) {- level_cache[x][0] = filter_level[0][0][0];
- level_cache[x][1] = filter_level[1][0][0];
- level_cache[x][2] = filter_level[2][0][0];
- level_cache[x][3] = filter_level[3][0][0];
+ level_cache_ptr[x][0] = filter_level[0][0][0];
+ level_cache_ptr[x][1] = filter_level[1][0][0];
}
- level_cache += b4_stride;
+ level_cache_ptr += b4_stride;
}
mask_edges_intra(lflvl->filter_y, by4, bx4, bw4, bh4, ytx, ay, ly);
@@ -277,11 +275,20 @@
const int cbx4 = bx4 >> ss_hor;
const int cby4 = by4 >> ss_ver;
+ level_cache_ptr = level_cache + (by >> ss_ver) * b4_stride + (bx >> ss_hor);
+ for (int y = 0; y < cbh4; y++) {+ for (int x = 0; x < cbw4; x++) {+ level_cache_ptr[x][2] = filter_level[2][0][0];
+ level_cache_ptr[x][3] = filter_level[3][0][0];
+ }
+ level_cache_ptr += b4_stride;
+ }
+
mask_edges_chroma(lflvl->filter_uv, cby4, cbx4, cbw4, cbh4, 0, uvtx, auv, luv);
}
void dav1d_create_lf_mask_inter(Av1Filter *const lflvl,
- uint8_t (*level_cache)[4],
+ uint8_t (*const level_cache)[4],
const ptrdiff_t b4_stride,
const Av1FrameHeader *const hdr,
const uint8_t (*filter_level)[8][2],
@@ -303,15 +310,13 @@
const int bx4 = bx & 31;
const int by4 = by & 31;
- level_cache += by * b4_stride + bx;
+ uint8_t (*level_cache_ptr)[4] = level_cache + by * b4_stride + bx;
for (int y = 0; y < bh4; y++) { for (int x = 0; x < bw4; x++) {- level_cache[x][0] = filter_level[0][0][0];
- level_cache[x][1] = filter_level[1][0][0];
- level_cache[x][2] = filter_level[2][0][0];
- level_cache[x][3] = filter_level[3][0][0];
+ level_cache_ptr[x][0] = filter_level[0][0][0];
+ level_cache_ptr[x][1] = filter_level[1][0][0];
}
- level_cache += b4_stride;
+ level_cache_ptr += b4_stride;
}
mask_edges_inter(lflvl->filter_y, by4, bx4, bw4, bh4, skip,
@@ -325,6 +330,15 @@
const int cbh4 = (bh4 + ss_ver) >> ss_ver;
const int cbx4 = bx4 >> ss_hor;
const int cby4 = by4 >> ss_ver;
+
+ level_cache_ptr = level_cache + (by >> ss_ver) * b4_stride + (bx >> ss_hor);
+ for (int y = 0; y < cbh4; y++) {+ for (int x = 0; x < cbw4; x++) {+ level_cache_ptr[x][2] = filter_level[2][0][0];
+ level_cache_ptr[x][3] = filter_level[3][0][0];
+ }
+ level_cache_ptr += b4_stride;
+ }
mask_edges_chroma(lflvl->filter_uv, cby4, cbx4, cbw4, cbh4, skip, uvtx, auv, luv);
}
--- a/src/loopfilter.c
+++ b/src/loopfilter.c
@@ -67,7 +67,6 @@
fm &= abs(p3 - p2) <= I && abs(q3 - q2) <= I;
}
}
-
if (!fm) continue;
if (wd >= 16) {@@ -181,6 +180,42 @@
#undef lf_4_fn
#undef lf_4_fns
+static void loop_filter_v_sb128y_c(pixel *dst, const ptrdiff_t stride,
+ const uint32_t *const vmask,
+ const uint8_t (*l)[4], ptrdiff_t b4_stride,
+ const Av1FilterLUT *lut, const int w)
+{+ const unsigned vm = vmask[0] | vmask[1] | vmask[2];
+ for (unsigned x = 1; vm & ~(x - 1); x <<= 1, dst += 4, l++) {+ if (vm & x) {+ const int L = l[0][0] ? l[0][0] : l[-b4_stride][0];
+ if (!L) continue;
+ const int H = L >> 4;
+ const int E = lut->e[L], I = lut->i[L];
+ const int idx = (vmask[2] & x) ? 2 : !!(vmask[1] & x);
+ loop_filter(dst, E, I, H, 1, PXSTRIDE(stride), 4 << idx);
+ }
+ }
+}
+
+static void loop_filter_v_sb128uv_c(pixel *dst, const ptrdiff_t stride,
+ const uint32_t *const vmask,
+ const uint8_t (*l)[4], ptrdiff_t b4_stride,
+ const Av1FilterLUT *lut, const int w)
+{+ const unsigned vm = vmask[0] | vmask[1];
+ for (unsigned x = 1; vm & ~(x - 1); x <<= 1, dst += 4, l++) {+ if (vm & x) {+ const int L = l[0][0] ? l[0][0] : l[-b4_stride][0];
+ if (!L) continue;
+ const int H = L >> 4;
+ const int E = lut->e[L], I = lut->i[L];
+ const int idx = !!(vmask[1] & x);
+ loop_filter(dst, E, I, H, 1, PXSTRIDE(stride), 4 + 2 * idx);
+ }
+ }
+}
+
void bitfn(dav1d_loop_filter_dsp_init)(Dav1dLoopFilterDSPContext *const c) {c->loop_filter[0][0] = loop_filter_h_4wd_4px_c;
c->loop_filter[0][1] = loop_filter_v_4wd_4px_c;
@@ -193,4 +228,7 @@
c->loop_filter_uv[0][1] = loop_filter_v_4wd_4px_c;
c->loop_filter_uv[1][0] = loop_filter_h_6wd_4px_c;
c->loop_filter_uv[1][1] = loop_filter_v_6wd_4px_c;
+
+ c->loop_filter_sb128y = loop_filter_v_sb128y_c;
+ c->loop_filter_sb128uv = loop_filter_v_sb128uv_c;
}
--- a/src/loopfilter.h
+++ b/src/loopfilter.h
@@ -34,11 +34,18 @@
#include "common/bitdepth.h"
#include "src/levels.h"
+#include "src/lf_mask.h"
#define decl_loopfilter_fn(name) \
void (name)(pixel *dst, ptrdiff_t stride, int mb_lim, int lim, int hev_thr)
typedef decl_loopfilter_fn(*loopfilter_fn);
+#define decl_loopfilter_sb_fn(name) \
+void (name)(pixel *dst, ptrdiff_t stride, const uint32_t *mask, \
+ const uint8_t (*lvl)[4], ptrdiff_t lvl_stride, \
+ const Av1FilterLUT *lut, int w)
+typedef decl_loopfilter_sb_fn(*loopfilter_sb_fn);
+
typedef struct Dav1dLoopFilterDSPContext {/*
* dimension 1: filter taps (0=4, 1=8, 2=16 for luma; 0=4, 1=6 for chroma)
@@ -48,6 +55,8 @@
*/
loopfilter_fn loop_filter[3][2];
loopfilter_fn loop_filter_uv[2][2];
+ loopfilter_sb_fn loop_filter_sb128y;
+ loopfilter_sb_fn loop_filter_sb128uv;
} Dav1dLoopFilterDSPContext;
void dav1d_loop_filter_dsp_init_8bpc(Dav1dLoopFilterDSPContext *c);