Add SSE2 function for Inverse Cross-color Transform

Lossless decoding is now ~3% faster.

Change-Id: Idafb5c73e5cfb272cc3661d841f79971f9da0743
This commit is contained in:
Urvang Joshi
2014-04-01 15:52:25 -07:00
parent 26029568b7
commit d4813f0cb2
3 changed files with 144 additions and 66 deletions

View File

@ -13,12 +13,14 @@
#include "./dsp.h"
#include <assert.h>
#if defined(WEBP_USE_SSE2)
#include <emmintrin.h>
#include "./lossless.h"
//------------------------------------------------------------------------------
// Predictors
// Predictor Transform
static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1,
uint32_t c2) {
@ -118,7 +120,7 @@ static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
}
//------------------------------------------------------------------------------
// Colorspace conversion functions
// Subtract-Green Transform
static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
const __m128i mask = _mm_set1_epi32(0x0000ff00);
@ -152,6 +154,65 @@ static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
VP8LAddGreenToBlueAndRed_C(argb_data + i, num_pixels - i);
}
//------------------------------------------------------------------------------
// Color Transform
static WEBP_INLINE __m128i ColorTransformDelta(__m128i color_pred,
__m128i color) {
// We simulate signed 8-bit multiplication as:
// * Left shift the two (8-bit) numbers by 8 bits,
// * Perform a 16-bit signed multiplication and retain the higher 16-bits.
const __m128i color_pred_shifted = _mm_slli_epi32(color_pred, 8);
const __m128i color_shifted = _mm_slli_epi32(color, 8);
// Note: This performs multiplication on 8 packed 16-bit numbers, 4 of which
// happen to be zeroes.
const __m128i signed_mult =
_mm_mulhi_epi16(color_pred_shifted, color_shifted);
return _mm_srli_epi32(signed_mult, 5);
}
static WEBP_INLINE void TransformColorInverse(const VP8LMultipliers* const m,
uint32_t* argb_data,
int num_pixels) {
const __m128i g_to_r = _mm_set1_epi32(m->green_to_red_); // multipliers
const __m128i g_to_b = _mm_set1_epi32(m->green_to_blue_);
const __m128i r_to_b = _mm_set1_epi32(m->red_to_blue_);
int i;
for (i = 0; i + 4 <= num_pixels; i += 4) {
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
const __m128i alpha_green_mask = _mm_set1_epi32(0xff00ff00); // masks
const __m128i red_mask = _mm_set1_epi32(0x00ff0000);
const __m128i green_mask = _mm_set1_epi32(0x0000ff00);
const __m128i lower_8bit_mask = _mm_set1_epi32(0x000000ff);
const __m128i ag = _mm_and_si128(in, alpha_green_mask); // alpha, green
const __m128i r = _mm_srli_epi32(_mm_and_si128(in, red_mask), 16);
const __m128i g = _mm_srli_epi32(_mm_and_si128(in, green_mask), 8);
const __m128i b = in;
const __m128i r_delta = ColorTransformDelta(g_to_r, g); // red
const __m128i r_new =
_mm_and_si128(_mm_add_epi32(r, r_delta), lower_8bit_mask);
const __m128i r_new_shifted = _mm_slli_epi32(r_new, 16);
const __m128i b_delta_1 = ColorTransformDelta(g_to_b, g); // blue
const __m128i b_delta_2 = ColorTransformDelta(r_to_b, r_new);
const __m128i b_delta = _mm_add_epi32(b_delta_1, b_delta_2);
const __m128i b_new =
_mm_and_si128(_mm_add_epi32(b, b_delta), lower_8bit_mask);
const __m128i out = _mm_or_si128(_mm_or_si128(ag, r_new_shifted), b_new);
_mm_storeu_si128((__m128i*)&argb_data[i], out);
}
// Fall-back to C-version for left-overs.
VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i);
}
//------------------------------------------------------------------------------
// Color-space conversion functions
static void ConvertBGRAToRGBA(const uint32_t* src,
int num_pixels, uint8_t* dst) {
const __m128i* in = (const __m128i*)src;
@ -298,6 +359,8 @@ void VP8LDspInitSSE2(void) {
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
VP8LTransformColorInverse = TransformColorInverse;
VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;
VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444;
VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565;