dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2024-12-27 22:28:22 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

Unroll Predictors 10, 11 and 12.

Browse Source 
We see the following speed-ups:
10 -> 13%
11 -> 13%
12 -> 13%

Change-Id: I4734fd388d0f4e508884d0b123976bf2cbe69d2f
This commit is contained in:
Vincent Rabaud 2017-06-08 19:05:44 +02:00
parent a1779a017b
commit 8f6df1d0b9
1 changed files with 79 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

128
src/dsp/lossless_sse2.c
View File

@ -272,9 +272,24 @@ GENERATE_PREDICTOR_2(9, upper[i + 1])
#undef GENERATE_PREDICTOR_2 #undef GENERATE_PREDICTOR_2
// Predictor10: average of (average of (L,TL), average of (T, TR)). // Predictor10: average of (average of (L,TL), average of (T, TR)).
#define DO_PRED10(OUT) do { \
__m128i avgLTL, avg; \
Average2_m128i(&L, &TL, &avgLTL); \
Average2_m128i(&avgTTR, &avgLTL, &avg); \
L = _mm_add_epi8(avg, src); \
out[i + (OUT)] = _mm_cvtsi128_si32(L); \
} while (0)
#define DO_PRED10_SHIFT do { \
/* Rotate the pre-computed values for the next iteration.*/ \
avgTTR = _mm_srli_si128(avgTTR, 4); \
TL = _mm_srli_si128(TL, 4); \
src = _mm_srli_si128(src, 4); \
} while (0)
static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper, static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) { int num_pixels, uint32_t* out) {
int i, j; int i;
__m128i L = _mm_cvtsi32_si128(out[-1]); __m128i L = _mm_cvtsi32_si128(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) { for (i = 0; i + 4 <= num_pixels; i += 4) {
__m128i src = _mm_loadu_si128((const __m128i*)&in[i]); __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
@ -283,77 +298,88 @@ static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]); const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
__m128i avgTTR; __m128i avgTTR;
Average2_m128i(&T, &TR, &avgTTR); Average2_m128i(&T, &TR, &avgTTR);
for (j = 0; j < 4; ++j) { DO_PRED10(0);
__m128i avgLTL, avg; DO_PRED10_SHIFT;
Average2_m128i(&L, &TL, &avgLTL); DO_PRED10(1);
Average2_m128i(&avgTTR, &avgLTL, &avg); DO_PRED10_SHIFT;
L = _mm_add_epi8(avg, src); DO_PRED10(2);
out[i + j] = _mm_cvtsi128_si32(L); DO_PRED10_SHIFT;
// Rotate the pre-computed values for the next iteration. DO_PRED10(3);
avgTTR = _mm_srli_si128(avgTTR, 4);
TL = _mm_srli_si128(TL, 4);
src = _mm_srli_si128(src, 4);
}
} }
if (i != num_pixels) { if (i != num_pixels) {
VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i); VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i);
} }
} }
#undef DO_PRED10
#undef DO_PRED10_SHIFT
// Predictor11: select. // Predictor11: select.
static void GetSumAbsDiff32(const __m128i* const A, const __m128i* const B, #define DO_PRED11(OUT) do { \
__m128i* const out) { const __m128i L_lo = _mm_unpacklo_epi32(L, T); \
// We can unpack with any value on the upper 32 bits, provided it's the same const __m128i TL_lo = _mm_unpacklo_epi32(TL, T); \
// on both operands (to that their sum of abs diff is zero). Here we use *A. const __m128i pb = _mm_sad_epu8(L_lo, TL_lo); /* pb = sum |L-TL|*/ \
const __m128i A_lo = _mm_unpacklo_epi32(*A, *A); const __m128i mask = _mm_cmpgt_epi32(pb, pa); \
const __m128i B_lo = _mm_unpacklo_epi32(*B, *A); const __m128i A = _mm_and_si128(mask, L); \
const __m128i A_hi = _mm_unpackhi_epi32(*A, *A); const __m128i B = _mm_andnot_si128(mask, T); \
const __m128i B_hi = _mm_unpackhi_epi32(*B, *A); const __m128i pred = _mm_or_si128(A, B); /* pred = (pa > b)? L : T*/ \
const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo); L = _mm_add_epi8(src, pred); \
const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi); out[i + (OUT)] = _mm_cvtsi128_si32(L); \
*out = _mm_packs_epi32(s_lo, s_hi); } while (0)
}
#define DO_PRED11_SHIFT do { \
/* Shift the pre-computed value for the next iteration.*/ \
T = _mm_srli_si128(T, 4); \
TL = _mm_srli_si128(TL, 4); \
src = _mm_srli_si128(src, 4); \
pa = _mm_srli_si128(pa, 4); \
} while (0)
static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper, static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) { int num_pixels, uint32_t* out) {
int i, j; int i;
__m128i pa;
__m128i L = _mm_cvtsi32_si128(out[-1]); __m128i L = _mm_cvtsi32_si128(out[-1]);
for (i = 0; i + 4 <= num_pixels; i += 4) { for (i = 0; i + 4 <= num_pixels; i += 4) {
__m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
__m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
__m128i src = _mm_loadu_si128((const __m128i*)&in[i]); __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
__m128i pa; {
GetSumAbsDiff32(&T, &TL, &pa); // pa = sum |T-TL| // We can unpack with any value on the upper 32 bits, provided it's the
for (j = 0; j < 4; ++j) { // same on both operands (so that their sum of abs diff is zero). Here we
const __m128i L_lo = _mm_unpacklo_epi32(L, L); // use T.
const __m128i TL_lo = _mm_unpacklo_epi32(TL, L); const __m128i T_lo = _mm_unpacklo_epi32(T, T);
const __m128i pb = _mm_sad_epu8(L_lo, TL_lo); // pb = sum |L-TL| const __m128i TL_lo = _mm_unpacklo_epi32(TL, T);
const __m128i mask = _mm_cmpgt_epi32(pb, pa); const __m128i T_hi = _mm_unpackhi_epi32(T, T);
const __m128i A = _mm_and_si128(mask, L); const __m128i TL_hi = _mm_unpackhi_epi32(TL, T);
const __m128i B = _mm_andnot_si128(mask, T); const __m128i s_lo = _mm_sad_epu8(T_lo, TL_lo);
const __m128i pred = _mm_or_si128(A, B); // pred = (L > T)? L : T const __m128i s_hi = _mm_sad_epu8(T_hi, TL_hi);
L = _mm_add_epi8(src, pred); pa = _mm_packs_epi32(s_lo, s_hi); // pa = sum |T-TL|
out[i + j] = _mm_cvtsi128_si32(L);
// Shift the pre-computed value for the next iteration.
T = _mm_srli_si128(T, 4);
TL = _mm_srli_si128(TL, 4);
src = _mm_srli_si128(src, 4);
pa = _mm_srli_si128(pa, 4);
} }
DO_PRED11(0);
DO_PRED11_SHIFT;
DO_PRED11(1);
DO_PRED11_SHIFT;
DO_PRED11(2);
DO_PRED11_SHIFT;
DO_PRED11(3);
} }
if (i != num_pixels) { if (i != num_pixels) {
VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i); VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i);
} }
} }
#undef DO_PRED11
#undef DO_PRED11_SHIFT
// Predictor12: ClampedAddSubtractFull. // Predictor12: ClampedAddSubtractFull.
#define DO_PRED12(DIFF, LANE, OUT) \ #define DO_PRED12(DIFF, LANE, OUT) do { \
do { \ const __m128i all = _mm_add_epi16(L, (DIFF)); \
const __m128i all = _mm_add_epi16(L, (DIFF)); \ const __m128i alls = _mm_packus_epi16(all, all); \
const __m128i alls = _mm_packus_epi16(all, all); \ const __m128i res = _mm_add_epi8(src, alls); \
const __m128i res = _mm_add_epi8(src, alls); \ out[i + (OUT)] = _mm_cvtsi128_si32(res); \
out[i + (OUT)] = _mm_cvtsi128_si32(res); \ L = _mm_unpacklo_epi8(res, zero); \
L = _mm_unpacklo_epi8(res, zero); \ } while (0)
#define DO_PRED12_SHIFT(DIFF, LANE) do { \
/* Shift the pre-computed value for the next iteration.*/ \ /* Shift the pre-computed value for the next iteration.*/ \
if (LANE == 0) (DIFF) = _mm_srli_si128((DIFF), 8); \ if (LANE == 0) (DIFF) = _mm_srli_si128((DIFF), 8); \
src = _mm_srli_si128(src, 4); \ src = _mm_srli_si128(src, 4); \
@ -377,8 +403,11 @@ static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper,
__m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo); __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo);
__m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi); __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi);
DO_PRED12(diff_lo, 0, 0); DO_PRED12(diff_lo, 0, 0);
DO_PRED12_SHIFT(diff_lo, 0);
DO_PRED12(diff_lo, 1, 1); DO_PRED12(diff_lo, 1, 1);
DO_PRED12_SHIFT(diff_lo, 1);
DO_PRED12(diff_hi, 0, 2); DO_PRED12(diff_hi, 0, 2);
DO_PRED12_SHIFT(diff_hi, 0);
DO_PRED12(diff_hi, 1, 3); DO_PRED12(diff_hi, 1, 3);
} }
if (i != num_pixels) { if (i != num_pixels) {
@ -386,6 +415,7 @@ static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper,
} }
} }
#undef DO_PRED12 #undef DO_PRED12
#undef DO_PRED12_SHIFT
// Due to averages with integers, values cannot be accumulated in parallel for // Due to averages with integers, values cannot be accumulated in parallel for
// predictors 13. // predictors 13.