1-2% faster quantization in SSE2

C-version is a bit faster too (sub-1% faster on ARM)

Change-Id: I077262042f1d0937aba1ecf15174f2c51bf6cd97
This commit is contained in:
skal
2014-02-13 15:55:30 -08:00
parent b2fbc36c26
commit 0235d5e44b
4 changed files with 64 additions and 62 deletions

View File

@ -592,13 +592,13 @@ static int TrellisQuantizeBlock(const VP8EncIterator* const it,
// traverse trellis.
for (n = first; n <= last; ++n) {
const int j = kZigzag[n];
const int Q = mtx->q_[j];
const int iQ = mtx->iq_[j];
const int B = BIAS(0x00); // neutral bias
const uint32_t Q = mtx->q_[j];
const uint32_t iQ = mtx->iq_[j];
const uint32_t B = BIAS(0x00); // neutral bias
// note: it's important to take sign of the _original_ coeff,
// so we don't have to consider level < 0 afterward.
const int sign = (in[j] < 0);
const int coeff0 = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
const uint32_t coeff0 = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
int level0 = QUANTDIV(coeff0, iQ, B);
if (level0 > MAX_LEVEL) level0 = MAX_LEVEL;

View File

@ -166,8 +166,8 @@ typedef int64_t score_t; // type used for scores, rate, distortion
#define BIAS(b) ((b) << (QFIX - 8))
// Fun fact: this is the _only_ line where we're actually being lossy and
// discarding bits.
static WEBP_INLINE int QUANTDIV(int n, int iQ, int B) {
return (n * iQ + B) >> QFIX;
static WEBP_INLINE int QUANTDIV(uint32_t n, uint32_t iQ, uint32_t B) {
return (int)((n * iQ + B) >> QFIX);
}
// size of histogram used by CollectHistogram.
@ -236,8 +236,8 @@ typedef struct {
typedef struct VP8Matrix {
uint16_t q_[16]; // quantizer steps
uint16_t iq_[16]; // reciprocals, fixed point.
uint16_t bias_[16]; // rounding bias
uint16_t zthresh_[16]; // value under which a coefficient is zeroed
uint32_t bias_[16]; // rounding bias
uint32_t zthresh_[16]; // value below which a coefficient is zeroed
uint16_t sharpen_[16]; // frequency boosters for slight sharpening
} VP8Matrix;