Optimize VP8SetResidualCoeffs.

Brings down WebP lossy encoding timings by 5%

Change-Id: Ia4a2fab0a887aaaf7841ce6d9ee16270d3e15489
This commit is contained in:
skal
2014-06-03 06:44:04 +02:00
parent ac591cf22e
commit 6679f8996f
6 changed files with 107 additions and 44 deletions

View File

@ -13,6 +13,12 @@
#include "./cost.h"
#if defined(WEBP_USE_SSE2)
#include <emmintrin.h>
#endif // WEBP_USE_SSE2
#include "../utils/utils.h"
//------------------------------------------------------------------------------
// Boolean-cost cost table
@ -536,15 +542,13 @@ extern int VP8GetResidualCostMIPS32(int ctx0, const VP8Residual* const res);
VP8GetResidualCostFunc VP8GetResidualCost;
void VP8GetResidualCostInit(void) {
if (VP8GetResidualCost == NULL) {
VP8GetResidualCost = GetResidualCost;
if (VP8GetCPUInfo != NULL) {
VP8GetResidualCost = GetResidualCost;
if (VP8GetCPUInfo != NULL) {
#if defined(WEBP_USE_MIPS32)
if (VP8GetCPUInfo(kMIPS32)) {
VP8GetResidualCost = VP8GetResidualCostMIPS32;
}
#endif
if (VP8GetCPUInfo(kMIPS32)) {
VP8GetResidualCost = VP8GetResidualCostMIPS32;
}
#endif
}
}
@ -560,7 +564,8 @@ void VP8InitResidual(int first, int coeff_type,
res->first = first;
}
void VP8SetResidualCoeffs(const int16_t* const coeffs, VP8Residual* const res) {
static void SetResidualCoeffs(const int16_t* const coeffs,
VP8Residual* const res) {
int n;
res->last = -1;
for (n = 15; n >= res->first; --n) {
@ -572,6 +577,27 @@ void VP8SetResidualCoeffs(const int16_t* const coeffs, VP8Residual* const res) {
res->coeffs = coeffs;
}
//------------------------------------------------------------------------------
// init function
#if defined(WEBP_USE_SSE2)
extern void VP8SetResidualCoeffsSSE2(const int16_t* const coeffs,
VP8Residual* const res);
#endif // WEBP_USE_SSE2
VP8SetResidualCoeffsFunc VP8SetResidualCoeffs;
void VP8SetResidualCoeffsInit(void) {
VP8SetResidualCoeffs = SetResidualCoeffs;
if (VP8GetCPUInfo != NULL) {
#if defined(WEBP_USE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8SetResidualCoeffs = VP8SetResidualCoeffsSSE2;
}
#endif
}
}
//------------------------------------------------------------------------------
// Mode costs

View File

@ -37,7 +37,13 @@ typedef struct {
void VP8InitResidual(int first, int coeff_type,
VP8Encoder* const enc, VP8Residual* const res);
void VP8SetResidualCoeffs(const int16_t* const coeffs, VP8Residual* const res);
typedef void (*VP8SetResidualCoeffsFunc)(const int16_t* const coeffs,
VP8Residual* const res);
extern VP8SetResidualCoeffsFunc VP8SetResidualCoeffs;
extern void VP8SetResidualCoeffsInit(void); // must be called first
int VP8RecordCoeffs(int ctx, const VP8Residual* const res);
// approximate cost per level:

View File

@ -253,6 +253,7 @@ static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
ResetFilterHeader(enc);
ResetBoundaryPredictions(enc);
VP8GetResidualCostInit();
VP8SetResidualCoeffsInit();
VP8EncInitAlpha(enc);
// lower quality means smaller output -> we modulate a little the page