Commit 7a85c05dd8 for ffmpeg
commit 7a85c05dd877abd8ed502f8d1370ec70acd4065b
Author: Lynne <dev@lynne.ee>
Date: Sat Nov 7 15:25:53 2026 +0800
avcodec/vulkan_apv: decode DC codes longer than 32 bits
The entropy decoder reads each code from a 32-bit window, but a DC
coefficient difference beyond 32768 with a small Rice parameter takes up
to 34 bits with its sign. Such codes were cut short, which threw off the
rest of the tile. They occur at 12 bits and above at very low QPs, for
instance between flat blocks at opposite ends of the range.
Read DC differences in parts instead: the prefix, the bits after it,
then the sign. There is one per block, and AC levels, which take at most
32 bits with their sign, keep reading from a single window.
diff --git a/libavcodec/vulkan/apv_decode.comp.glsl b/libavcodec/vulkan/apv_decode.comp.glsl
index 914c130b0c..8d05a40f23 100644
--- a/libavcodec/vulkan/apv_decode.comp.glsl
+++ b/libavcodec/vulkan/apv_decode.comp.glsl
@@ -97,11 +97,10 @@ int apv_read_vlc(int k)
}
/*
- * As above, with the trailing sign bit folded into the same window. The caller
- * says whether a sign bit is present (DC: when val != 0; AC levels: always);
- * sign is only valid when it is.
+ * As above, for AC levels, with their trailing sign bit folded into the same
+ * window.
*/
-int apv_read_vlc_sign(int k, bool sign_always, out bool sign)
+int apv_read_vlc_sign(int k, out bool sign)
{
/* Top 32 bits; the longest legal code is 31 bits, +1 for the sign */
uint bits = show_bits(gb, 32);
@@ -126,16 +125,36 @@ int apv_read_vlc_sign(int k, bool sign_always, out bool sign)
int val = caseA ? valA : (caseB ? valB : valC);
int len = caseA ? lenA : (caseB ? lenB : lenC);
- bool has_sign = sign_always || val != 0;
- sign = has_sign && bool((bits >> (31 - len)) & 1u);
- len += has_sign ? 1 : 0;
-
- skip_bits(gb, len);
+ sign = bool((bits >> (31 - len)) & 1u);
+ skip_bits(gb, len + 1);
if (!caseA && !caseB && badC)
val = APV_MAX_TRANS_COEFF + 1;
return val;
}
+/*
+ * DC differences go up to 65535, whose codes take up to 34 bits with the
+ * sign, more than one window holds. There is one per block, so read the
+ * prefix, the k or n + k bits after it and the sign one after the other.
+ */
+int apv_read_dc_diff(int k, out bool sign)
+{
+ uint bits = show_bits(gb, 32);
+ int n = clamp(31 - findMSB(bits << 2), 0, 15);
+
+ bool caseA = bits >= 0x80000000u;
+ bool caseB = bits < 0x40000000u;
+
+ int prefix = caseA ? 1 : (caseB ? 2 : 3 + n);
+ int base = caseA ? 0 : (caseB ? 1 << k : ((1 << n) + 1) << k);
+ int len = caseA || caseB ? k : n + k;
+
+ skip_bits(gb, prefix);
+ int val = base + int(get_bits(gb, len));
+ sign = val != 0 && get_bits(gb, 1) != 0u;
+ return val;
+}
+
/* ff_zigzag_direct, packed: each byte is the raster index (y*8 + x). */
const uint8_t zigzag[64] = {
uint8_t( 0), uint8_t( 1), uint8_t( 8), uint8_t(16),
@@ -165,7 +184,7 @@ void decode_block(uint cbase, int cstride, ivec2 pos, uint comp)
int dc_coeff;
bool dc_sign;
- int abs_diff = apv_read_vlc_sign(prev_k_dc, false, dc_sign);
+ int abs_diff = apv_read_dc_diff(prev_k_dc, dc_sign);
dc_coeff = prev_dc + (dc_sign ? -abs_diff : abs_diff);
@@ -202,8 +221,7 @@ void decode_block(uint cbase, int cstride, ivec2 pos, uint comp)
bool sign_ac_coeff;
k_param = clamp(prev_level >> 2, 0, 4);
- abs_ac_coeff_minus1 = apv_read_vlc_sign(k_param, true,
- sign_ac_coeff);
+ abs_ac_coeff_minus1 = apv_read_vlc_sign(k_param, sign_ac_coeff);
if (sign_ac_coeff)
level = -abs_ac_coeff_minus1 - 1;