ref: 128715b5fdd513f9824bef32119285efa44c1d1b
parent: 32c62b5f38e9c935454b95e3aa1730ebd00f87cf
author: Ronald S. Bultje <rsbultje@gmail.com>
date: Wed Oct 24 15:21:46 EDT 2018
Make access into loopfilter masks safe across sb64 tile threads Fixes #100.
--- a/src/lf_apply.c
+++ b/src/lf_apply.c
@@ -38,7 +38,7 @@
const int have_left,
const uint8_t (*lvl)[4],
const ptrdiff_t b4_stride,
- const uint32_t (*const mask)[3],
+ const uint16_t (*const mask)[3][2],
pixel *dst, const ptrdiff_t ls,
const int w,
const int starty4, const int endy4)
@@ -48,12 +48,23 @@
// filter edges between columns (e.g. block1 | block2)
for (int x = 0; x < w; x++) {if (!have_left && !x) continue;
- dsp->lf.loop_filter_sb[0][0](&dst[x * 4], ls,
- starty4 ? (const uint32_t[4]) {- mask[x][0] >> starty4,
- mask[x][1] >> starty4,
- mask[x][2] >> starty4,
- } : mask[x],
+ uint32_t hmask[4];
+ if (!starty4) {+ hmask[0] = mask[x][0][0];
+ hmask[1] = mask[x][1][0];
+ hmask[2] = mask[x][2][0];
+ if (endy4 > 16) {+ hmask[0] |= mask[x][0][1] << 16;
+ hmask[1] |= mask[x][1][1] << 16;
+ hmask[2] |= mask[x][2][1] << 16;
+ }
+ } else {+ hmask[0] = mask[x][0][1];
+ hmask[1] = mask[x][1][1];
+ hmask[2] = mask[x][2][1];
+ }
+ hmask[3] = 0;
+ dsp->lf.loop_filter_sb[0][0](&dst[x * 4], ls, hmask,
(const uint8_t(*)[4]) &lvl[x][0], b4_stride,
&f->lf.lim_lut, endy4 - starty4);
}
@@ -63,7 +74,7 @@
const int have_top,
const uint8_t (*lvl)[4],
const ptrdiff_t b4_stride,
- const uint32_t (*const mask)[3],
+ const uint16_t (*const mask)[3][2],
pixel *dst, const ptrdiff_t ls,
const int w,
const int starty4, const int endy4)
@@ -77,7 +88,13 @@
y++, dst += 4 * PXSTRIDE(ls), lvl += b4_stride)
{if (!have_top && !y) continue;
- dsp->lf.loop_filter_sb[0][1](dst, ls, mask[y],
+ const uint32_t vmask[4] = {+ mask[y][0][0] | (mask[y][0][1] << 16),
+ mask[y][1][0] | (mask[y][1][1] << 16),
+ mask[y][2][0] | (mask[y][2][1] << 16),
+ 0,
+ };
+ dsp->lf.loop_filter_sb[0][1](dst, ls, vmask,
(const uint8_t(*)[4]) &lvl[0][1], b4_stride,
&f->lf.lim_lut, w);
}
@@ -87,10 +104,11 @@
const int have_left,
const uint8_t (*lvl)[4],
const ptrdiff_t b4_stride,
- const uint32_t (*const mask)[2],
+ const uint16_t (*const mask)[2][2],
pixel *const u, pixel *const v,
const ptrdiff_t ls, const int w,
- const int starty4, const int endy4)
+ const int starty4, const int endy4,
+ const int ss_ver)
{const Dav1dDSPContext *const dsp = f->dsp;
@@ -97,18 +115,23 @@
// filter edges between columns (e.g. block1 | block2)
for (int x = 0; x < w; x++) {if (!have_left && !x) continue;
- dsp->lf.loop_filter_sb[1][0](&u[x * 4], ls,
- starty4 ? (const uint32_t[3]) {- mask[x][0] >> starty4,
- mask[x][1] >> starty4,
- } : mask[x],
+ uint32_t hmask[3];
+ if (!starty4) {+ hmask[0] = mask[x][0][0];
+ hmask[1] = mask[x][1][0];
+ if (endy4 > (16 >> ss_ver)) {+ hmask[0] |= mask[x][0][1] << (16 >> ss_ver);
+ hmask[1] |= mask[x][1][1] << (16 >> ss_ver);
+ }
+ } else {+ hmask[0] = mask[x][0][1];
+ hmask[1] = mask[x][1][1];
+ }
+ hmask[2] = 0;
+ dsp->lf.loop_filter_sb[1][0](&u[x * 4], ls, hmask,
(const uint8_t(*)[4]) &lvl[x][2], b4_stride,
&f->lf.lim_lut, endy4 - starty4);
- dsp->lf.loop_filter_sb[1][0](&v[x * 4], ls,
- starty4 ? (const uint32_t[3]) {- mask[x][0] >> starty4,
- mask[x][1] >> starty4,
- } : mask[x],
+ dsp->lf.loop_filter_sb[1][0](&v[x * 4], ls, hmask,
(const uint8_t(*)[4]) &lvl[x][3], b4_stride,
&f->lf.lim_lut, endy4 - starty4);
}
@@ -118,10 +141,11 @@
const int have_top,
const uint8_t (*lvl)[4],
const ptrdiff_t b4_stride,
- const uint32_t (*const mask)[2],
+ const uint16_t (*const mask)[2][2],
pixel *const u, pixel *const v,
const ptrdiff_t ls, const int w,
- const int starty4, const int endy4)
+ const int starty4, const int endy4,
+ const int ss_hor)
{const Dav1dDSPContext *const dsp = f->dsp;
ptrdiff_t off_l = 0;
@@ -133,10 +157,15 @@
y++, off_l += 4 * PXSTRIDE(ls), lvl += b4_stride)
{if (!have_top && !y) continue;
- dsp->lf.loop_filter_sb[1][1](&u[off_l], ls, mask[y],
+ const uint32_t vmask[3] = {+ mask[y][0][0] | (mask[y][0][1] << (16 >> ss_hor)),
+ mask[y][1][0] | (mask[y][1][1] << (16 >> ss_hor)),
+ 0,
+ };
+ dsp->lf.loop_filter_sb[1][1](&u[off_l], ls, vmask,
(const uint8_t(*)[4]) &lvl[0][2], b4_stride,
&f->lf.lim_lut, w);
- dsp->lf.loop_filter_sb[1][1](&v[off_l], ls, mask[y],
+ dsp->lf.loop_filter_sb[1][1](&v[off_l], ls, vmask,
(const uint8_t(*)[4]) &lvl[0][3], b4_stride,
&f->lf.lim_lut, w);
}
@@ -157,6 +186,8 @@
const int halign = (f->bh + 31) & ~31;
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 vmask = 16 >> ss_ver, hmask = 16 >> ss_hor;
+ const unsigned vmax = 1 << vmask, hmax = 1 << hmask;
const unsigned endy4 = starty4 + imin(hy4 - sby * sbsz, sbsz);
const unsigned uv_endy4 = (endy4 + ss_ver) >> ss_ver;
@@ -169,24 +200,29 @@
const int bx4 = x & is_sb64 ? 16 : 0, cbx4 = bx4 >> ss_hor;
x >>= is_sb64;
- uint32_t *const y_hmask = lflvl[x].filter_y[0][bx4];
+ uint16_t (*const y_hmask)[2] = lflvl[x].filter_y[0][bx4];
for (unsigned y = starty4, mask = 1 << y; y < endy4; y++, mask <<= 1) {- const int idx = 2 * !!(y_hmask[2] & mask) + !!(y_hmask[1] & mask);
- y_hmask[2] &= ~mask;
- y_hmask[1] &= ~mask;
- y_hmask[0] &= ~mask;
- y_hmask[imin(idx, lpf_y[y - starty4])] |= mask;
+ const int sidx = mask >= 0x10000;
+ const unsigned smask = mask >> (sidx << 4);
+ const int idx = 2 * !!(y_hmask[2][sidx] & smask) +
+ !!(y_hmask[1][sidx] & smask);
+ y_hmask[2][sidx] &= ~smask;
+ y_hmask[1][sidx] &= ~smask;
+ y_hmask[0][sidx] &= ~smask;
+ y_hmask[imin(idx, lpf_y[y - starty4])][sidx] |= smask;
}
if (f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I400) {- uint32_t *const uv_hmask = lflvl[x].filter_uv[0][cbx4];
+ uint16_t (*const uv_hmask)[2] = lflvl[x].filter_uv[0][cbx4];
for (unsigned y = starty4 >> ss_ver, uv_mask = 1 << y; y < uv_endy4;
y++, uv_mask <<= 1)
{- const int idx = !!(uv_hmask[1] & uv_mask);
- uv_hmask[1] &= ~uv_mask;
- uv_hmask[0] &= ~uv_mask;
- uv_hmask[imin(idx, lpf_uv[y - (starty4 >> ss_ver)])] |= uv_mask;
+ const int sidx = uv_mask >= vmax;
+ const unsigned smask = uv_mask >> (sidx << (4 - ss_ver));
+ const int idx = !!(uv_hmask[1][sidx] & smask);
+ uv_hmask[1][sidx] &= ~smask;
+ uv_hmask[0][sidx] &= ~smask;
+ uv_hmask[imin(idx, lpf_uv[y - (starty4 >> ss_ver)])][sidx] |= smask;
}
}
lpf_y += halign;
@@ -199,22 +235,27 @@
for (x = 0, a = &f->a[f->sb128w * (start_of_tile_row - 1)];
x < f->sb128w; x++, a++)
{- uint32_t *const y_vmask = lflvl[x].filter_y[1][starty4];
+ uint16_t (*const y_vmask)[2] = lflvl[x].filter_y[1][starty4];
for (unsigned mask = 1, i = 0; i < 32; mask <<= 1, i++) {- const int idx = 2 * !!(y_vmask[2] & mask) + !!(y_vmask[1] & mask);
- y_vmask[2] &= ~mask;
- y_vmask[1] &= ~mask;
- y_vmask[0] &= ~mask;
- y_vmask[imin(idx, a->tx_lpf_y[i])] |= mask;
+ const int sidx = mask >= 0x10000;
+ const unsigned smask = mask >> (sidx << 4);
+ const int idx = 2 * !!(y_vmask[2][sidx] & smask) +
+ !!(y_vmask[1][sidx] & smask);
+ y_vmask[2][sidx] &= ~smask;
+ y_vmask[1][sidx] &= ~smask;
+ y_vmask[0][sidx] &= ~smask;
+ y_vmask[imin(idx, a->tx_lpf_y[i])][sidx] |= smask;
}
if (f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I400) {- uint32_t *const uv_vmask = lflvl[x].filter_uv[1][starty4 >> ss_ver];
- for (unsigned mask = 1, i = 0; i < (32U >> ss_hor); mask <<= 1, i++) {- const int idx = !!(uv_vmask[1] & mask);
- uv_vmask[1] &= ~mask;
- uv_vmask[0] &= ~mask;
- uv_vmask[imin(idx, a->tx_lpf_uv[i])] |= mask;
+ uint16_t (*const uv_vmask)[2] = lflvl[x].filter_uv[1][starty4 >> ss_ver];
+ for (unsigned uv_mask = 1, i = 0; i < (32U >> ss_hor); uv_mask <<= 1, i++) {+ const int sidx = uv_mask >= hmax;
+ const unsigned smask = uv_mask >> (sidx << (4 - ss_hor));
+ const int idx = !!(uv_vmask[1][sidx] & smask);
+ uv_vmask[1][sidx] &= ~smask;
+ uv_vmask[0][sidx] &= ~smask;
+ uv_vmask[imin(idx, a->tx_lpf_uv[i])][sidx] |= smask;
}
}
}
@@ -249,7 +290,7 @@
lflvl[x].filter_uv[0],
&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);
+ starty4 >> ss_ver, uv_endy4, ss_ver);
}
level_ptr = f->lf.level + f->b4_stride * (sby * sbsz >> ss_ver);
@@ -260,6 +301,6 @@
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);
+ starty4 >> ss_ver, uv_endy4, ss_hor);
}
}
--- a/src/lf_mask.c
+++ b/src/lf_mask.c
@@ -73,7 +73,7 @@
}
}
-static inline void mask_edges_inter(uint32_t (*const masks)[32][3],
+static inline void mask_edges_inter(uint16_t (*const masks)[32][3][2],
const int by4, const int bx4,
const int w4, const int h4, const int skip,
const enum RectTxfmSize max_tx,
@@ -91,21 +91,29 @@
// left block edge
unsigned mask = 1U << by4;
- for (y = 0; y < h4; y++, mask <<= 1)
- masks[0][bx4][imin(txa[0][0][y][0], l[y])] |= mask;
+ for (y = 0; y < h4; y++, mask <<= 1) {+ const int sidx = mask >= 0x10000;
+ const unsigned smask = mask >> (sidx << 4);
+ masks[0][bx4][imin(txa[0][0][y][0], l[y])][sidx] |= smask;
+ }
// top block edge
- for (x = 0, mask = 1U << bx4; x < w4; x++, mask <<= 1)
- masks[1][by4][imin(txa[1][0][0][x], a[x])] |= mask;
+ for (x = 0, mask = 1U << bx4; x < w4; x++, mask <<= 1) {+ const int sidx = mask >= 0x10000;
+ const unsigned smask = mask >> (sidx << 4);
+ masks[1][by4][imin(txa[1][0][0][x], a[x])][sidx] |= smask;
+ }
if (!skip) {// inner (tx) left|right edges
for (y = 0, mask = 1U << by4; y < h4; y++, mask <<= 1) {+ const int sidx = mask >= 0x10000U;
+ const unsigned smask = mask >> (sidx << 4);
int ltx = txa[0][0][y][0];
int step = txa[0][1][y][0];
for (x = step; x < w4; x += step) {const int rtx = txa[0][0][y][x];
- masks[0][bx4 + x][imin(rtx, ltx)] |= mask;
+ masks[0][bx4 + x][imin(rtx, ltx)][sidx] |= smask;
ltx = rtx;
step = txa[0][1][y][x];
}
@@ -115,11 +123,13 @@
// inner (tx) --- edges
// bottom
for (x = 0, mask = 1U << bx4; x < w4; x++, mask <<= 1) {+ const int sidx = mask >= 0x10000U;
+ const unsigned smask = mask >> (sidx << 4);
int ttx = txa[1][0][0][x];
int step = txa[1][1][0][x];
for (y = step; y < h4; y += step) {const int btx = txa[1][0][y][x];
- masks[1][by4 + y][imin(ttx, btx)] |= mask;
+ masks[1][by4 + y][imin(ttx, btx)][sidx] |= smask;
ttx = btx;
step = txa[1][1][y][x];
}
@@ -131,7 +141,7 @@
memcpy(a, txa[1][0][h4 - 1], w4);
}
-static inline void mask_edges_intra(uint32_t (*const masks)[32][3],
+static inline void mask_edges_intra(uint16_t (*const masks)[32][3][2],
const int by4, const int bx4,
const int w4, const int h4,
const enum RectTxfmSize tx,
@@ -144,19 +154,28 @@
// left block edge
unsigned mask = 1U << by4;
- for (y = 0; y < h4; y++, mask <<= 1)
- masks[0][bx4][imin(twl4c, l[y])] |= mask;
+ for (y = 0; y < h4; y++, mask <<= 1) {+ const int sidx = mask >= 0x10000;
+ const unsigned smask = mask >> (sidx << 4);
+ masks[0][bx4][imin(twl4c, l[y])][sidx] |= smask;
+ }
// top block edge
- for (x = 0, mask = 1U << bx4; x < w4; x++, mask <<= 1)
- masks[1][by4][imin(thl4c, a[x])] |= mask;
+ for (x = 0, mask = 1U << bx4; x < w4; x++, mask <<= 1) {+ const int sidx = mask >= 0x10000;
+ const unsigned smask = mask >> (sidx << 4);
+ masks[1][by4][imin(thl4c, a[x])][sidx] |= smask;
+ }
// inner (tx) left|right edges
const int hstep = t_dim->w;
unsigned t = 1U << by4;
unsigned inner = (((uint64_t) t) << h4) - t;
- for (x = hstep; x < w4; x += hstep)
- masks[0][bx4 + x][twl4c] |= inner;
+ unsigned inner1 = inner & 0xffff, inner2 = inner >> 16;
+ for (x = hstep; x < w4; x += hstep) {+ if (inner1) masks[0][bx4 + x][twl4c][0] |= inner1;
+ if (inner2) masks[0][bx4 + x][twl4c][1] |= inner2;
+ }
// top
// inner (tx) --- edges
@@ -164,41 +183,58 @@
const int vstep = t_dim->h;
t = 1U << bx4;
inner = (((uint64_t) t) << w4) - t;
- for (y = vstep; y < h4; y += vstep)
- masks[1][by4 + y][thl4c] |= inner;
+ inner1 = inner & 0xffff;
+ inner2 = inner >> 16;
+ for (y = vstep; y < h4; y += vstep) {+ if (inner1) masks[1][by4 + y][thl4c][0] |= inner1;
+ if (inner2) masks[1][by4 + y][thl4c][1] |= inner2;
+ }
memset(a, thl4c, w4);
memset(l, twl4c, h4);
}
-static inline void mask_edges_chroma(uint32_t (*const masks)[32][2],
+static inline void mask_edges_chroma(uint16_t (*const masks)[32][2][2],
const int cby4, const int cbx4,
const int cw4, const int ch4,
const int skip_inter,
const enum RectTxfmSize tx,
- uint8_t *const a, uint8_t *const l)
+ uint8_t *const a, uint8_t *const l,
+ const int ss_hor, const int ss_ver)
{const TxfmInfo *const t_dim = &dav1d_txfm_dimensions[tx];
const int twl4 = t_dim->lw, thl4 = t_dim->lh;
const int twl4c = !!twl4, thl4c = !!thl4;
int y, x;
+ const int vbits = 4 - ss_ver, hbits = 4 - ss_hor;
+ const int vmask = 16 >> ss_ver, hmask = 16 >> ss_hor;
+ const unsigned vmax = 1 << vmask, hmax = 1 << hmask;
// left block edge
unsigned mask = 1U << cby4;
- for (y = 0; y < ch4; y++, mask <<= 1)
- masks[0][cbx4][imin(twl4c, l[y])] |= mask;
+ for (y = 0; y < ch4; y++, mask <<= 1) {+ const int sidx = mask >= vmax;
+ const unsigned smask = mask >> (sidx << vbits);
+ masks[0][cbx4][imin(twl4c, l[y])][sidx] |= smask;
+ }
// top block edge
- for (x = 0, mask = 1U << cbx4; x < cw4; x++, mask <<= 1)
- masks[1][cby4][imin(thl4c, a[x])] |= mask;
+ for (x = 0, mask = 1U << cbx4; x < cw4; x++, mask <<= 1) {+ const int sidx = mask >= hmax;
+ const unsigned smask = mask >> (sidx << hbits);
+ masks[1][cby4][imin(thl4c, a[x])][sidx] |= smask;
+ }
if (!skip_inter) {// inner (tx) left|right edges
const int hstep = t_dim->w;
- int t = 1U << cby4;
+ unsigned t = 1U << cby4;
unsigned inner = (((uint64_t) t) << ch4) - t;
- for (x = hstep; x < cw4; x += hstep)
- masks[0][cbx4 + x][twl4c] |= inner;
+ unsigned inner1 = inner & ((1 << vmask) - 1), inner2 = inner >> vmask;
+ for (x = hstep; x < cw4; x += hstep) {+ if (inner1) masks[0][cbx4 + x][twl4c][0] |= inner1;
+ if (inner2) masks[0][cbx4 + x][twl4c][1] |= inner2;
+ }
// top
// inner (tx) --- edges
@@ -206,8 +242,11 @@
const int vstep = t_dim->h;
t = 1U << cbx4;
inner = (((uint64_t) t) << cw4) - t;
- for (y = vstep; y < ch4; y += vstep)
- masks[1][cby4 + y][thl4c] |= inner;
+ inner1 = inner & ((1 << hmask) - 1), inner2 = inner >> hmask;
+ for (y = vstep; y < ch4; y += vstep) {+ if (inner1) masks[1][cby4 + y][thl4c][0] |= inner1;
+ if (inner2) masks[1][cby4 + y][thl4c][1] |= inner2;
+ }
}
memset(a, thl4c, cw4);
@@ -271,7 +310,8 @@
level_cache_ptr += b4_stride;
}
- mask_edges_chroma(lflvl->filter_uv, cby4, cbx4, cbw4, cbh4, 0, uvtx, auv, luv);
+ mask_edges_chroma(lflvl->filter_uv, cby4, cbx4, cbw4, cbh4, 0, uvtx,
+ auv, luv, ss_hor, ss_ver);
}
void dav1d_create_lf_mask_inter(Av1Filter *const lflvl,
@@ -332,7 +372,8 @@
level_cache_ptr += b4_stride;
}
- mask_edges_chroma(lflvl->filter_uv, cby4, cbx4, cbw4, cbh4, skip, uvtx, auv, luv);
+ mask_edges_chroma(lflvl->filter_uv, cby4, cbx4, cbw4, cbh4, skip, uvtx,
+ auv, luv, ss_hor, ss_ver);
}
void dav1d_calc_eih(Av1FilterLUT *const lim_lut, const int filter_sharpness) {--- a/src/lf_mask.h
+++ b/src/lf_mask.h
@@ -50,8 +50,8 @@
// each struct describes one 128x128 area (1 or 4 SBs)
typedef struct Av1Filter {// each bit is 1 col
- uint32_t filter_y[2 /* 0=col, 1=row */][32][3];
- uint32_t filter_uv[2 /* 0=col, 1=row */][32][2];
+ uint16_t filter_y[2 /* 0=col, 1=row */][32][3][2];
+ uint16_t filter_uv[2 /* 0=col, 1=row */][32][2][2];
int8_t cdef_idx[4]; // -1 means "unset"
uint16_t noskip_mask[32][2];
Av1RestorationUnit lr[3][4];