mirror of
https://github.com/webmproject/libwebp.git
synced 2025-02-13 15:32:53 +01:00
SSE2: 53% faster TransformColor[Inverse]
Changed the code (again) to process 4 pixels at a time. Loop is more involved, but overall it's faster. Removed the SSE4.1 implementation which is now slower than SSE2. Change-Id: I7734e371033ad8929ace7f7e1373ba930d9bb5f1
This commit is contained in:
Pascal Massimino
parent
49073da6d6
commit
fc6c75a2a2
@ -41,30 +41,34 @@ static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
|
||||
// Color Transform
|
||||
|
||||
static WEBP_INLINE void TransformColor(const VP8LMultipliers* const m,
|
||||
uint32_t* argb_data,
|
||||
int num_pixels) {
|
||||
// Used to collect the two parts of the delta (horizontal add) with madd.
|
||||
const __m128i kCstAdd = _mm_set1_epi16(1);
|
||||
uint32_t* argb_data, int num_pixels) {
|
||||
// sign-extended multiplying constants, pre-shifted by 5.
|
||||
#define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend
|
||||
const __m128i mults = _mm_set_epi16(
|
||||
CST(green_to_red_), 0, CST(red_to_blue_), CST(green_to_blue_),
|
||||
CST(green_to_red_), 0, CST(red_to_blue_), CST(green_to_blue_));
|
||||
const __m128i mults_rb = _mm_set_epi16(
|
||||
CST(green_to_red_), CST(green_to_blue_),
|
||||
CST(green_to_red_), CST(green_to_blue_),
|
||||
CST(green_to_red_), CST(green_to_blue_),
|
||||
CST(green_to_red_), CST(green_to_blue_));
|
||||
const __m128i mults_b2 = _mm_set_epi16(
|
||||
CST(red_to_blue_), 0, CST(red_to_blue_), 0,
|
||||
CST(red_to_blue_), 0, CST(red_to_blue_), 0);
|
||||
#undef CST
|
||||
const __m128i zero = _mm_setzero_si128();
|
||||
const __m128i mask = _mm_set1_epi32(0xff);
|
||||
const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks
|
||||
const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff); // red-blue masks
|
||||
int i;
|
||||
for (i = 0; i + 2 <= num_pixels; i += 2) {
|
||||
const __m128i in = _mm_loadl_epi64((__m128i*)&argb_data[i]); // argb
|
||||
const __m128i A = _mm_unpacklo_epi8(zero, in);
|
||||
const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(1, 0, 2, 1)); // gxrg
|
||||
const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(1, 0, 2, 1));
|
||||
const __m128i D = _mm_mulhi_epi16(C, mults); // dr | 0 | db1 | db2
|
||||
const __m128i E = _mm_madd_epi16(D, kCstAdd); // 0 | dr | 0 | db
|
||||
const __m128i F = _mm_and_si128(E, mask);
|
||||
const __m128i G = _mm_packus_epi16(F, zero); // dr | 0 | db | 0
|
||||
const __m128i out = _mm_sub_epi8(in, G);
|
||||
_mm_storel_epi64((__m128i*)&argb_data[i], out);
|
||||
for (i = 0; i + 4 <= num_pixels; i += 4) {
|
||||
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
|
||||
const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0
|
||||
const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
|
||||
const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0
|
||||
const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1
|
||||
const __m128i E = _mm_slli_epi16(in, 8); // r 0 b 0
|
||||
const __m128i F = _mm_mulhi_epi16(E, mults_b2); // x db2 0 0
|
||||
const __m128i G = _mm_srli_epi32(F, 16); // 0 0 x db2
|
||||
const __m128i H = _mm_add_epi8(G, D); // x dr x db
|
||||
const __m128i I = _mm_and_si128(H, mask_rb); // 0 dr 0 db
|
||||
const __m128i out = _mm_sub_epi8(in, I);
|
||||
_mm_storeu_si128((__m128i*)&argb_data[i], out);
|
||||
}
|
||||
// fallthrough and finish off with plain-C
|
||||
VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
|
||||
|
||||
@ -35,40 +35,6 @@ static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
|
||||
VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
|
||||
}
|
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Color Transform
|
||||
|
||||
static WEBP_INLINE void TransformColor(const VP8LMultipliers* const m,
|
||||
uint32_t* argb_data, int num_pixels) {
|
||||
// Shuffle constant to spread green and red to some *upper* byte locations.
|
||||
const __m128i kCst_g0rg = _mm_set_epi8(5, -1, -1, -1, 6, -1, 5, -1,
|
||||
1, -1, -1, -1, 2, -1, 1, -1);
|
||||
// Shuffling constant to collect deltas from uint32 to uint8 locations.
|
||||
const __m128i kCstShuffle = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1,
|
||||
-1, 12, -1, 8, -1, 4, -1, 0);
|
||||
// Used to collect the two parts of the delta (horizontal add) with madd.
|
||||
const __m128i kCstAdd = _mm_set1_epi16(1);
|
||||
// sign-extended multiplying constants, pre-shifted by 5.
|
||||
#define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend
|
||||
const __m128i mults = _mm_set_epi16(
|
||||
CST(green_to_red_), 0, CST(red_to_blue_), CST(green_to_blue_),
|
||||
CST(green_to_red_), 0, CST(red_to_blue_), CST(green_to_blue_));
|
||||
#undef CST
|
||||
|
||||
int i;
|
||||
for (i = 0; i + 2 <= num_pixels; i += 2) {
|
||||
const __m128i in = _mm_loadl_epi64((__m128i*)&argb_data[i]); // argb
|
||||
const __m128i A = _mm_shuffle_epi8(in, kCst_g0rg); // g | 0 | r | g
|
||||
const __m128i B = _mm_mulhi_epi16(A, mults); // dr | 0 | db1 | db2
|
||||
const __m128i C = _mm_madd_epi16(B, kCstAdd); // dr | 0 | db | 0
|
||||
const __m128i D = _mm_shuffle_epi8(C, kCstShuffle); // 0 | dr | 0 | db
|
||||
const __m128i out = _mm_sub_epi8(in, D);
|
||||
_mm_storel_epi64((__m128i*)&argb_data[i], out);
|
||||
}
|
||||
// fallthrough and finish off with plain-C
|
||||
VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
|
||||
}
|
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Entry point
|
||||
|
||||
@ -76,7 +42,6 @@ extern void VP8LEncDspInitSSE41(void);
|
||||
|
||||
WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE41(void) {
|
||||
VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
|
||||
VP8LTransformColor = TransformColor;
|
||||
}
|
||||
|
||||
#else // !WEBP_USE_SSE41
|
||||
|
||||
@ -179,33 +179,30 @@ static WEBP_INLINE void TransformColorInverse(const VP8LMultipliers* const m,
|
||||
int num_pixels) {
|
||||
// sign-extended multiplying constants, pre-shifted by 5.
|
||||
#define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend
|
||||
const __m128i mults_r = _mm_set_epi16(
|
||||
0, CST(green_to_red_), 0, CST(green_to_red_),
|
||||
0, CST(green_to_red_), 0, CST(green_to_red_));
|
||||
const __m128i mults_b1 = _mm_set_epi16(
|
||||
0, CST(green_to_blue_), 0, CST(green_to_blue_),
|
||||
0, CST(green_to_blue_), 0, CST(green_to_blue_));
|
||||
const __m128i mults_rb = _mm_set_epi16(
|
||||
CST(green_to_red_), CST(green_to_blue_),
|
||||
CST(green_to_red_), CST(green_to_blue_),
|
||||
CST(green_to_red_), CST(green_to_blue_),
|
||||
CST(green_to_red_), CST(green_to_blue_));
|
||||
const __m128i mults_b2 = _mm_set_epi16(
|
||||
CST(red_to_blue_), 0, CST(red_to_blue_), 0,
|
||||
CST(red_to_blue_), 0, CST(red_to_blue_), 0);
|
||||
#undef CST
|
||||
const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks
|
||||
const __m128i mask_b = _mm_set1_epi32(0x000000ff); // blue mask
|
||||
const __m128i mask_r = _mm_set1_epi32(0x00ff0000); // red mask
|
||||
int i;
|
||||
for (i = 0; i + 4 <= num_pixels; i += 4) {
|
||||
const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
|
||||
const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0
|
||||
const __m128i C = _mm_mulhi_epi16(A, mults_r); // 0 0 x dr1
|
||||
const __m128i D = _mm_mulhi_epi16(A, mults_b1); // 0 0 x db1
|
||||
const __m128i E = _mm_and_si128(_mm_slli_epi32(C, 16), mask_r); // 0 dr 0 0
|
||||
const __m128i F = _mm_and_si128(D, mask_b); // 0 0 0 db1
|
||||
const __m128i G = _mm_add_epi8(in, E); // a r' g b
|
||||
const __m128i H = _mm_slli_epi16(G, 8); // r' 0 b 0
|
||||
const __m128i I = _mm_mulhi_epi16(H, mults_b2); // x db2 0 0
|
||||
const __m128i J = _mm_and_si128(_mm_srli_epi32(I, 16), mask_b); // db2
|
||||
const __m128i K = _mm_add_epi8(G, F);
|
||||
const __m128i out = _mm_add_epi8(K, J);
|
||||
const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
|
||||
const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0
|
||||
const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1
|
||||
const __m128i E = _mm_add_epi8(in, D); // x r' x b'
|
||||
const __m128i F = _mm_slli_epi16(E, 8); // r' 0 b' 0
|
||||
const __m128i G = _mm_mulhi_epi16(F, mults_b2); // x db2 0 0
|
||||
const __m128i H = _mm_srli_epi32(G, 8); // 0 x db2 0
|
||||
const __m128i I = _mm_add_epi8(H, F); // r' x b'' 0
|
||||
const __m128i J = _mm_srli_epi16(I, 8); // 0 r' 0 b''
|
||||
const __m128i out = _mm_or_si128(J, A);
|
||||
_mm_storeu_si128((__m128i*)&argb_data[i], out);
|
||||
}
|
||||
// Fall-back to C-version for left-overs.
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user