dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2025-02-12 15:02:53 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

* fix off-by-one zthresh calculation

Browse Source 
* remove the sharpening for non luma-AC coeffs
* adjust the bias a little bit to compensate for this

Using the multiply-by-reciprocal doesn't always give the same result
as the exact divide, given the QFIX fixed-point precision we use.
-> removed few now-unneeded SSE2 instructions (and checked for
bit-exactness using -noasm)

Change-Id: Ib68057cbdd69c4e589af56a01a8e7085db762c24
This commit is contained in:
skal 2013-12-09 13:56:04 +01:00
parent ad9dec0c24
commit d513bb62bc
2 changed files with 13 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/dsp/enc_sse2.c
View File

@ -830,8 +830,6 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]); const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]);
const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]); const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]);
const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]); const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]);
const __m128i zthresh0 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[0]);
const __m128i zthresh8 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[8]);
// sign(in) = in >> 15 (0x0000 if positive, 0xffff if negative) // sign(in) = in >> 15 (0x0000 if positive, 0xffff if negative)
const __m128i sign0 = _mm_srai_epi16(in0, 15); const __m128i sign0 = _mm_srai_epi16(in0, 15);
@ -894,17 +892,8 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
in0 = _mm_mullo_epi16(out0, q0); in0 = _mm_mullo_epi16(out0, q0);
in8 = _mm_mullo_epi16(out8, q8); in8 = _mm_mullo_epi16(out8, q8);
// if (coeff <= mtx->zthresh_) {in=0; out=0;} _mm_storeu_si128((__m128i*)&in[0], in0);
{ _mm_storeu_si128((__m128i*)&in[8], in8);
__m128i cmp0 = _mm_cmpgt_epi16(coeff0, zthresh0);
__m128i cmp8 = _mm_cmpgt_epi16(coeff8, zthresh8);
in0 = _mm_and_si128(in0, cmp0);
in8 = _mm_and_si128(in8, cmp8);
_mm_storeu_si128((__m128i*)&in[0], in0);
_mm_storeu_si128((__m128i*)&in[8], in8);
out0 = _mm_and_si128(out0, cmp0);
out8 = _mm_and_si128(out8, cmp8);
}
// zigzag the output before storing it. // zigzag the output before storing it.
// //

22
src/enc/quant.c
View File

@ -169,19 +169,13 @@ static const uint16_t kAcTable2[128] = {
385, 393, 401, 409, 416, 424, 432, 440 385, 393, 401, 409, 416, 424, 432, 440
}; };
static const uint16_t kCoeffThresh[16] = {
0, 10, 20, 30,
10, 20, 30, 30,
20, 30, 30, 30,
30, 30, 30, 30
};
static const uint8_t kBiasMatrices[3][2] = { // [luma-ac,luma-dc,chroma][dc,ac] static const uint8_t kBiasMatrices[3][2] = { // [luma-ac,luma-dc,chroma][dc,ac]
{ 96, 110 }, { 96, 112 }, { 112, 120 } { 96, 110 }, { 96, 108 }, { 110, 115 }
}; };
// Sharpening by (slightly) raising the hi-frequency coeffs. // Sharpening by (slightly) raising the hi-frequency coeffs.
// Hack-ish but helpful for mid-bitrate range. Use with care. // Hack-ish but helpful for mid-bitrate range. Use with care.
#define SHARPEN_BITS 11 // number of descaling bits for sharpening bias
static const uint8_t kFreqSharpening[16] = { static const uint8_t kFreqSharpening[16] = {
0, 30, 60, 90, 0, 30, 60, 90,
30, 60, 90, 90, 30, 60, 90, 90,
@ -204,9 +198,15 @@ static int ExpandMatrix(VP8Matrix* const m, int type) {
const int bias = kBiasMatrices[type][is_ac_coeff]; const int bias = kBiasMatrices[type][is_ac_coeff];
m->iq_[i] = (1 << QFIX) / m->q_[i]; m->iq_[i] = (1 << QFIX) / m->q_[i];
m->bias_[i] = BIAS(bias); m->bias_[i] = BIAS(bias);
// TODO(skal): tune kCoeffThresh[] // zthresh_ is the exact value such that QUANTDIV(coeff, iQ, B) is:
m->zthresh_[i] = ((256 /*+ kCoeffThresh[i]*/ - bias) * m->q_[i] + 127) >> 8; // * zero if coeff <= zthresh
m->sharpen_[i] = (kFreqSharpening[i] * m->q_[i]) >> 11; // * non-zero if coeff > zthresh
m->zthresh_[i] = ((1 << QFIX) - 1 - m->bias_[i]) / m->iq_[i];
if (type == 0) { // we only use sharpening for AC luma coeffs
m->sharpen_[i] = (kFreqSharpening[i] * m->q_[i]) >> SHARPEN_BITS;
} else {
m->sharpen_[i] = 0;
}
sum += m->q_[i]; sum += m->q_[i];
} }
return (sum + 8) >> 4; return (sum + 8) >> 4;