FFmpeg/libavcodec/vulkan/ffv1_dec.comp

378 lines
12 KiB
Text

/*
* FFv1 codec
*
* Copyright (c) 2024 Lynne <dev@lynne.ee>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#ifndef RGB
#define LADDR(p) (p)
#else
#define RGB_LINECACHE 2
#define RGB_LBUF (RGB_LINECACHE - 1)
#define LADDR(p) (ivec2((p).x, ((p).y & RGB_LBUF)))
#endif
#ifdef RGB
ivec2 get_pred(readonly uimage2D pred, ivec2 sp, ivec2 off, int comp, int sw, uint8_t quant_table_idx)
{
const ivec2 yoff_border1 = expectEXT(off.x == 0, false) ? ivec2(1, -1) : ivec2(0, 0);
/* Thanks to the same coincidence as below, we can skip checking if off == 0, 1 */
VTYPE3 top = VTYPE3(TYPE(imageLoad(pred, sp + LADDR(off + ivec2(-1, -1) + yoff_border1))[comp]),
TYPE(imageLoad(pred, sp + LADDR(off + ivec2(0, -1)))[comp]),
TYPE(imageLoad(pred, sp + LADDR(off + ivec2(min(1, sw - off.x - 1), -1)))[comp]));
/* Normally, we'd need to check if off != ivec2(0, 0) here, since otherwise, we must
* return zero. However, ivec2(-1, 0) + ivec2(1, -1) == ivec2(0, -1), e.g. previous
* row, 0 offset, same slice, which is zero since we zero out the buffer for RGB */
TYPE cur = TYPE(imageLoad(pred, sp + LADDR(off + ivec2(-1, 0) + yoff_border1))[comp]);
int base = quant_table[quant_table_idx][0][(cur - top[0]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][1][(top[0] - top[1]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][2][(top[1] - top[2]) & MAX_QUANT_TABLE_MASK];
if (expectEXT(extend_lookup[quant_table_idx] > 0, false)) {
TYPE cur2 = TYPE(0);
if (expectEXT(off.x > 0, true)) {
const ivec2 yoff_border2 = expectEXT(off.x == 1, false) ? ivec2(-1, -1) : ivec2(-2, 0);
cur2 = TYPE(imageLoad(pred, sp + LADDR(off + yoff_border2))[comp]);
}
base += quant_table[quant_table_idx][3][(cur2 - cur) & MAX_QUANT_TABLE_MASK];
/* top-2 became current upon swap */
TYPE top2 = TYPE(imageLoad(pred, sp + LADDR(off))[comp]);
base += quant_table[quant_table_idx][4][(top2 - top[1]) & MAX_QUANT_TABLE_MASK];
}
/* context, prediction */
return ivec2(base, predict(cur, VTYPE2(top)));
}
#else
ivec2 get_pred(readonly uimage2D pred, ivec2 sp, ivec2 off, int comp, int sw, uint8_t quant_table_idx)
{
const ivec2 yoff_border1 = off.x == 0 ? ivec2(1, -1) : ivec2(0, 0);
sp += off;
VTYPE3 top = VTYPE3(TYPE(0),
TYPE(0),
TYPE(0));
if (off.y > 0 && off != ivec2(0, 1))
top[0] = TYPE(imageLoad(pred, sp + ivec2(-1, -1) + yoff_border1)[comp]);
if (off.y > 0) {
top[1] = TYPE(imageLoad(pred, sp + ivec2(0, -1))[comp]);
top[2] = TYPE(imageLoad(pred, sp + ivec2(min(1, sw - off.x - 1), -1))[comp]);
}
TYPE cur = TYPE(0);
if (off != ivec2(0, 0))
cur = TYPE(imageLoad(pred, sp + ivec2(-1, 0) + yoff_border1)[comp]);
int base = quant_table[quant_table_idx][0][(cur - top[0]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][1][(top[0] - top[1]) & MAX_QUANT_TABLE_MASK] +
quant_table[quant_table_idx][2][(top[1] - top[2]) & MAX_QUANT_TABLE_MASK];
if ((quant_table[quant_table_idx][3][127] != 0) ||
(quant_table[quant_table_idx][4][127] != 0)) {
TYPE cur2 = TYPE(0);
if (off.x > 0 && off != ivec2(1, 0)) {
const ivec2 yoff_border2 = off.x == 1 ? ivec2(1, -1) : ivec2(0, 0);
cur2 = TYPE(imageLoad(pred, sp + ivec2(-2, 0) + yoff_border2)[comp]);
}
base += quant_table[quant_table_idx][3][(cur2 - cur) & MAX_QUANT_TABLE_MASK];
TYPE top2 = TYPE(0);
if (off.y > 1)
top2 = TYPE(imageLoad(pred, sp + ivec2(0, -2))[comp]);
base += quant_table[quant_table_idx][4][(top2 - top[1]) & MAX_QUANT_TABLE_MASK];
}
/* context, prediction */
return ivec2(base, predict(cur, VTYPE2(top)));
}
#endif
#ifndef GOLOMB
#ifdef CACHED_SYMBOL_READER
shared uint8_t state[CONTEXT_SIZE];
#define READ(c, off) get_rac_direct(c, state[off])
#else
#define READ(c, off) get_rac(c, uint64_t(slice_state) + (state_off + off))
#endif
int get_isymbol(inout RangeCoder c, uint state_off)
{
if (READ(c, 0))
return 0;
uint e = 1;
for (; e < 33; e++)
if (!READ(c, min(e, 10)))
break;
if (expectEXT(e == 1, false)) {
return READ(c, 11) ? -1 : 1;
} else if (expectEXT(e == 33, false)) {
corrupt = true;
return 0;
}
int a = 1;
for (uint i = e + 20; i >= 22; i--) {
a <<= 1;
a |= int(READ(c, min(i, 31)));
}
return READ(c, min(e + 10, 21)) ? -a : a;
}
void decode_line_pcm(inout SliceContext sc, ivec2 sp, int w, int y, int p, int bits)
{
#ifndef RGB
if (p > 0 && p < 3) {
w >>= chroma_shift.x;
sp >>= chroma_shift;
}
#endif
for (int x = 0; x < w; x++) {
uint v = 0;
for (int i = (bits - 1); i >= 0; i--)
v |= uint(get_rac_equi(sc.c)) << i;
imageStore(dec[p], sp + LADDR(ivec2(x, y)), uvec4(v));
}
}
void decode_line(inout SliceContext sc, ivec2 sp, int w,
int y, int p, int bits, uint state_off,
uint8_t quant_table_idx, const int run_index)
{
#ifndef RGB
if (p > 0 && p < 3) {
w >>= chroma_shift.x;
sp >>= chroma_shift;
}
#endif
for (int x = 0; x < w; x++) {
ivec2 pr = get_pred(dec[p], sp, ivec2(x, y), 0, w,
quant_table_idx);
uint context_off = state_off + CONTEXT_SIZE*abs(pr[0]);
#ifdef CACHED_SYMBOL_READER
u8buf sb = u8buf(uint64_t(slice_state) + context_off + gl_LocalInvocationID.x);
state[gl_LocalInvocationID.x] = sb.v;
barrier();
if (gl_LocalInvocationID.x == 0) {
#endif
int diff = get_isymbol(sc.c, context_off);
if (pr[0] < 0)
diff = -diff;
uint v = zero_extend(pr[1] + diff, bits);
imageStore(dec[p], sp + LADDR(ivec2(x, y)), uvec4(v));
#ifdef CACHED_SYMBOL_READER
}
sb.v = state[gl_LocalInvocationID.x];
#endif
}
}
#else /* GOLOMB */
void decode_line(inout SliceContext sc, ivec2 sp, int w,
int y, int p, int bits, uint state_off,
uint8_t quant_table_idx, inout int run_index)
{
#ifndef RGB
if (p > 0 && p < 3) {
w >>= chroma_shift.x;
sp >>= chroma_shift;
}
#endif
int run_count = 0;
int run_mode = 0;
for (int x = 0; x < w; x++) {
ivec2 pos = sp + ivec2(x, y);
int diff;
ivec2 pr = get_pred(dec[p], sp, ivec2(x, y), 0, w,
quant_table_idx);
VlcState sb = VlcState(uint64_t(slice_state) + state_off + VLC_STATE_SIZE*abs(pr[0]));
if (pr[0] == 0 && run_mode == 0)
run_mode = 1;
if (run_mode != 0) {
if (run_count == 0 && run_mode == 1) {
int tmp_idx = int(log2_run[run_index]);
if (get_bit(sc.gb)) {
run_count = 1 << tmp_idx;
if (x + run_count <= w)
run_index++;
} else {
if (tmp_idx != 0) {
run_count = int(get_bits(sc.gb, tmp_idx));
} else
run_count = 0;
if (run_index != 0)
run_index--;
run_mode = 2;
}
}
run_count--;
if (run_count < 0) {
run_mode = 0;
run_count = 0;
diff = read_vlc_symbol(sc.gb, sb, bits);
if (diff >= 0)
diff++;
} else {
diff = 0;
}
} else {
diff = read_vlc_symbol(sc.gb, sb, bits);
}
if (pr[0] < 0)
diff = -diff;
uint v = zero_extend(pr[1] + diff, bits);
imageStore(dec[p], sp + LADDR(ivec2(x, y)), uvec4(v));
}
}
#endif
#ifdef RGB
ivec4 transform_sample(ivec4 pix, ivec2 rct_coef)
{
pix.b -= rct_offset;
pix.r -= rct_offset;
pix.g -= (pix.b*rct_coef.y + pix.r*rct_coef.x) >> 2;
pix.b += pix.g;
pix.r += pix.g;
return ivec4(pix[fmt_lut[0]], pix[fmt_lut[1]],
pix[fmt_lut[2]], pix[fmt_lut[3]]);
}
void writeout_rgb(in SliceContext sc, ivec2 sp, int w, int y, bool apply_rct)
{
for (uint x = gl_LocalInvocationID.x; x < w; x += gl_WorkGroupSize.x) {
ivec2 lpos = sp + LADDR(ivec2(x, y));
ivec2 pos = sc.slice_pos + ivec2(x, y);
ivec4 pix;
pix.r = int(imageLoad(dec[2], lpos)[0]);
pix.g = int(imageLoad(dec[0], lpos)[0]);
pix.b = int(imageLoad(dec[1], lpos)[0]);
if (transparency != 0)
pix.a = int(imageLoad(dec[3], lpos)[0]);
if (expectEXT(apply_rct, true))
pix = transform_sample(pix, sc.slice_rct_coef);
imageStore(dst[0], pos, pix);
if (planar_rgb != 0) {
for (int i = 1; i < color_planes; i++)
imageStore(dst[i], pos, ivec4(pix[i]));
}
}
}
#endif
void decode_slice(inout SliceContext sc, const uint slice_idx)
{
int run_index = 0;
int w = sc.slice_dim.x;
ivec2 sp = sc.slice_pos;
#ifndef RGB
int bits = bits_per_raw_sample;
#else
int bits = 9;
if (bits != 8 || sc.slice_coding_mode != 0)
bits = bits_per_raw_sample + int(sc.slice_coding_mode != 1);
sp.y = int(gl_WorkGroupID.y)*RGB_LINECACHE;
#endif
/* PCM coding */
#ifndef GOLOMB
if (sc.slice_coding_mode == 1) {
if (gl_LocalInvocationID.x > 0)
return;
#ifndef RGB
for (int p = 0; p < planes; p++) {
int h = sc.slice_dim.y;
if (p > 0 && p < 3)
h >>= chroma_shift.y;
for (int y = 0; y < h; y++)
decode_line_pcm(sc, sp, w, y, p, bits);
}
#else
for (int y = 0; y < sc.slice_dim.y; y++) {
for (int p = 0; p < color_planes; p++)
decode_line_pcm(sc, sp, w, y, p, bits);
writeout_rgb(sc, sp, w, y, false);
}
#endif
} else
/* Arithmetic coding */
#endif
{
u8vec4 quant_table_idx = sc.quant_table_idx.xyyz;
u32vec4 slice_state_off = (slice_idx*codec_planes + uvec4(0, 1, 1, 2))*plane_state_size;
#ifndef RGB
for (int p = 0; p < planes; p++) {
int h = sc.slice_dim.y;
if (p > 0 && p < 3)
h >>= chroma_shift.y;
for (int y = 0; y < h; y++)
decode_line(sc, sp, w, y, p, bits,
slice_state_off[p], quant_table_idx[p], run_index);
}
#else
for (int y = 0; y < sc.slice_dim.y; y++) {
for (int p = 0; p < color_planes; p++)
decode_line(sc, sp, w, y, p, bits,
slice_state_off[p], quant_table_idx[p], run_index);
writeout_rgb(sc, sp, w, y, true);
}
#endif
}
}
void main(void)
{
const uint slice_idx = gl_WorkGroupID.y*gl_NumWorkGroups.x + gl_WorkGroupID.x;
decode_slice(slice_ctx[slice_idx], slice_idx);
}