eessppeelllloo/ffmpeg
1
0
Fork 0
mirror of https://git.ffmpeg.org/ffmpeg.git synced 2025-12-05 14:30:00 +01:00
Code Issues Packages Projects Releases Wiki Activity

vulkan/prores: normalize coefficients during IDCT

Browse Source 
This allows increased internal precision.
In addition, we can introduce an offset to the DC coefficient
during the second IDCT step, to remove a per-element addition
in the output codepath.
Finally, by processing columns first we can remove the barrier
after loading coefficients.

Signed-off-by: averne <averne381@gmail.com>
This commit is contained in:
averne
2025-11-29 17:25:17 +01:00
parent 1982add485
commit 1c5bb1b12d
1 changed files with 37 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
libavcodec/vulkan/prores_idct.comp
View File

@@ -37,19 +37,27 @@ void put_px(uint tex_idx, ivec2 pos, uint v)
#endif #endif
} }
const float idct_8x8_scales[] = { const float idct_scale[64] = {
0.353553390593274f, // cos(4 * pi/16) / 2 0.1250000000000000, 0.1733799806652684, 0.1633203706095471, 0.1469844503024199,
0.490392640201615f, // cos(1 * pi/16) / 2 0.1250000000000000, 0.0982118697983878, 0.0676495125182746, 0.0344874224103679,
0.461939766255643f, // cos(2 * pi/16) / 2 0.1733799806652684, 0.2404849415639108, 0.2265318615882219, 0.2038732892122293,
0.415734806151273f, // cos(3 * pi/16) / 2 0.1733799806652684, 0.1362237766939547, 0.0938325693794663, 0.0478354290456362,
0.353553390593274f, // cos(4 * pi/16) / 2 0.1633203706095471, 0.2265318615882219, 0.2133883476483184, 0.1920444391778541,
0.277785116509801f, // cos(5 * pi/16) / 2 0.1633203706095471, 0.1283199917898342, 0.0883883476483185, 0.0450599888754343,
0.191341716182545f, // cos(6 * pi/16) / 2 0.1469844503024199, 0.2038732892122293, 0.1920444391778541, 0.1728354290456362,
0.097545161008064f, // cos(7 * pi/16) / 2 0.1469844503024199, 0.1154849415639109, 0.0795474112858021, 0.0405529186026822,
0.1250000000000000, 0.1733799806652684, 0.1633203706095471, 0.1469844503024199,
0.1250000000000000, 0.0982118697983878, 0.0676495125182746, 0.0344874224103679,
0.0982118697983878, 0.1362237766939547, 0.1283199917898342, 0.1154849415639109,
0.0982118697983878, 0.0771645709543638, 0.0531518809229535, 0.0270965939155924,
0.0676495125182746, 0.0938325693794663, 0.0883883476483185, 0.0795474112858021,
0.0676495125182746, 0.0531518809229535, 0.0366116523516816, 0.0186644585125857,
0.0344874224103679, 0.0478354290456362, 0.0450599888754343, 0.0405529186026822,
0.0344874224103679, 0.0270965939155924, 0.0186644585125857, 0.0095150584360892,
}; };
/* 7.4 Inverse Transform */ /* 7.4 Inverse Transform */
void idct(uint block, uint offset, uint stride) void idct8(uint block, uint offset, uint stride)
{ {
float t0, t1, t2, t3, t4, t5, t6, t7, u8; float t0, t1, t2, t3, t4, t5, t6, t7, u8;
float u0, u1, u2, u3, u4, u5, u6, u7; float u0, u1, u2, u3, u4, u5, u6, u7;
@@ -117,6 +125,12 @@ void main(void)
uint chroma_shift = comp != 0 ? log2_chroma_w : 0; uint chroma_shift = comp != 0 ? log2_chroma_w : 0;
bool act = gid.x < mb_width << (4 - chroma_shift); bool act = gid.x < mb_width << (4 - chroma_shift);
/**
* Normalize coefficients to [-1, 1] for increased precision during the iDCT.
* DCT coeffs have the range of a 12-bit signed integer (7.4 Inverse Transform).
*/
const float norm = 1.0f / (1 << 11);
/* Coalesced load of DCT coeffs in shared memory, inverse quantization */ /* Coalesced load of DCT coeffs in shared memory, inverse quantization */
if (act) { if (act) {
/** /**
@@ -131,28 +145,31 @@ void main(void)
int qscale = qidx > 128 ? (qidx - 96) << 2 : qidx; int qscale = qidx > 128 ? (qidx - 96) << 2 : qidx;
[[unroll]] for (uint i = 0; i < 8; ++i) { [[unroll]] for (uint i = 0; i < 8; ++i) {
uint cidx = (i << 3) + idx;
int c = sign_extend(int(get_px(comp, ivec2(gid.x, (gid.y << 3) + i))), 16); int c = sign_extend(int(get_px(comp, ivec2(gid.x, (gid.y << 3) + i))), 16);
float v = float(c * qscale * int(qmat[(i << 3) + idx])); float v = float(c * qscale * int(qmat[cidx])) * norm;
blocks[block][i * 9 + idx] = v * idct_8x8_scales[idx] * idct_8x8_scales[i]; blocks[block][i * 9 + idx] = v * idct_scale[cidx];
} }
} }
/* Row-wise iDCT */
barrier();
idct(block, idx * 9, 1);
/* Column-wise iDCT */ /* Column-wise iDCT */
idct8(block, idx, 9);
barrier(); barrier();
idct(block, idx, 9);
float fact = 1.0f / (1 << (12 - depth)), off = 1 << (depth - 1); /* Remap [-1, 1] to [0, 2] to remove a per-element addition in the output loop */
blocks[block][idx * 9] += 1.0f;
/* Row-wise iDCT */
idct8(block, idx * 9, 1);
barrier();
float fact = 1 << (depth - 1);
int maxv = (1 << depth) - 1; int maxv = (1 << depth) - 1;
/* 7.5.1 Color Component Samples. Rescale, clamp and write back to global memory */ /* 7.5.1 Color Component Samples. Rescale, clamp and write back to global memory */
barrier();
if (act) { if (act) {
[[unroll]] for (uint i = 0; i < 8; ++i) { [[unroll]] for (uint i = 0; i < 8; ++i) {
float v = round(blocks[block][i * 9 + idx] * fact + off); float v = round(blocks[block][i * 9 + idx] * fact);
put_px(comp, ivec2(gid.x, (gid.y << 3) + i), clamp(int(v), 0, maxv)); put_px(comp, ivec2(gid.x, (gid.y << 3) + i), clamp(int(v), 0, maxv));
} }
} }