dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2024-12-26 13:48:21 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

SSE2 version of the fwd transform and the squared sum metric

Browse Source 
average 10% faster encoding.

Patch by Christian Duvivier (cduvivier at google dot com)

Change-Id: Iff1bba402b280522af323f73e0c817a3d665354a
This commit is contained in:
Pascal Massimino 2011-04-26 00:01:33 -07:00
parent 2ab4b72f53
commit fb29c26201
1 changed files with 199 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

199
src/enc/dsp_sse2.c
View File

@ -232,6 +232,203 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
}
}
static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
int16_t* out) {
const __m128i zero = _mm_setzero_si128();
const __m128i seven = _mm_set1_epi16(7);
const __m128i k7500 = _mm_set1_epi32(7500);
const __m128i k14500 = _mm_set1_epi32(14500);
const __m128i k51000 = _mm_set1_epi32(51000);
const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16));
const __m128i k5352_2217 = _mm_set_epi16(5352, 2217, 5352, 2217,
5352, 2217, 5352, 2217);
const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352,
2217, -5352, 2217, -5352);
__m128i v01, v32;
// Difference between src and ref and initial transpose.
{
// Load src and convert to 16b.
const __m128i src0 = _mm_loadl_epi64((__m128i *)&src[0 * BPS]);
const __m128i src1 = _mm_loadl_epi64((__m128i *)&src[1 * BPS]);
const __m128i src2 = _mm_loadl_epi64((__m128i *)&src[2 * BPS]);
const __m128i src3 = _mm_loadl_epi64((__m128i *)&src[3 * BPS]);
const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
// Load ref and convert to 16b.
const __m128i ref0 = _mm_loadl_epi64((__m128i *)&ref[0 * BPS]);
const __m128i ref1 = _mm_loadl_epi64((__m128i *)&ref[1 * BPS]);
const __m128i ref2 = _mm_loadl_epi64((__m128i *)&ref[2 * BPS]);
const __m128i ref3 = _mm_loadl_epi64((__m128i *)&ref[3 * BPS]);
const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
// Compute difference.
const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
// Transpose.
// 00 01 02 03 0 0 0 0
// 10 11 12 13 0 0 0 0
// 20 21 22 23 0 0 0 0
// 30 31 32 33 0 0 0 0
const __m128i transpose0_0 = _mm_unpacklo_epi16(diff0, diff1);
const __m128i transpose0_1 = _mm_unpacklo_epi16(diff2, diff3);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));
// a02 a12 a22 a32 a03 a13 a23 a33
// a00 a10 a20 a30 a01 a11 a21 a31
// a03 a13 a23 a33 a02 a12 a22 a32
}
// First pass and subsequent transpose.
{
// Same operations are done on the (0,3) and (1,2) pairs.
// b0 = (a0 + a3) << 3
// b1 = (a1 + a2) << 3
// b3 = (a0 - a3) << 3
// b2 = (a1 - a2) << 3
const __m128i a01 = _mm_add_epi16(v01, v32);
const __m128i a32 = _mm_sub_epi16(v01, v32);
const __m128i b01 = _mm_slli_epi16(a01, 3);
const __m128i b32 = _mm_slli_epi16(a32, 3);
const __m128i b11 = _mm_unpackhi_epi64(b01, b01);
const __m128i b22 = _mm_unpackhi_epi64(b32, b32);
// e0 = b0 + b1
// e2 = b0 - b1
const __m128i e0 = _mm_add_epi16(b01, b11);
const __m128i e2 = _mm_sub_epi16(b01, b11);
const __m128i e02 = _mm_unpacklo_epi64(e0, e2);
// e1 = (b3 * 5352 + b2 * 2217 + 14500) >> 12
// e3 = (b3 * 2217 - b2 * 5352 + 7500) >> 12
const __m128i b23 = _mm_unpacklo_epi16(b22, b32);
const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
const __m128i d1 = _mm_add_epi32(c1, k14500);
const __m128i d3 = _mm_add_epi32(c3, k7500);
const __m128i e1 = _mm_srai_epi32(d1, 12);
const __m128i e3 = _mm_srai_epi32(d3, 12);
const __m128i e13 = _mm_packs_epi32(e1, e3);
// Transpose.
// 00 01 02 03 20 21 22 23
// 10 11 12 13 30 31 32 33
const __m128i transpose0_0 = _mm_unpacklo_epi16(e02, e13);
const __m128i transpose0_1 = _mm_unpackhi_epi16(e02, e13);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));
// 02 12 22 32 03 13 23 33
// 00 10 20 30 01 11 21 31
// 03 13 23 33 02 12 22 32
}
// Second pass
{
// Same operations are done on the (0,3) and (1,2) pairs.
// a0 = v0 + v3
// a1 = v1 + v2
// a3 = v0 - v3
// a2 = v1 - v2
const __m128i a01 = _mm_add_epi16(v01, v32);
const __m128i a32 = _mm_sub_epi16(v01, v32);
const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
// d0 = (a0 + a1 + 7) >> 4;
// d2 = (a0 - a1 + 7) >> 4;
const __m128i b0 = _mm_add_epi16(a01, a11);
const __m128i b2 = _mm_sub_epi16(a01, a11);
const __m128i c0 = _mm_add_epi16(b0, seven);
const __m128i c2 = _mm_add_epi16(b2, seven);
const __m128i d0 = _mm_srai_epi16(c0, 4);
const __m128i d2 = _mm_srai_epi16(c2, 4);
// f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
// f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
const __m128i b23 = _mm_unpacklo_epi16(a22, a32);
const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one);
const __m128i d3 = _mm_add_epi32(c3, k51000);
const __m128i e1 = _mm_srai_epi32(d1, 16);
const __m128i e3 = _mm_srai_epi32(d3, 16);
const __m128i f1 = _mm_packs_epi32(e1, e1);
const __m128i f3 = _mm_packs_epi32(e3, e3);
// f1 = f1 + (a3 != 0);
// The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
// desired (0, 1), we add one earlier through k12000_plus_one.
const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
_mm_storel_epi64((__m128i *)&out[ 0], d0);
_mm_storel_epi64((__m128i *)&out[ 4], g1);
_mm_storel_epi64((__m128i *)&out[ 8], d2);
_mm_storel_epi64((__m128i *)&out[12], f3);
}
}
//-----------------------------------------------------------------------------
// Metric
static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) {
const __m128i zero = _mm_set1_epi16(0);
// Load values.
const __m128i a0 = _mm_loadl_epi64((__m128i *)&a[BPS * 0]);
const __m128i a1 = _mm_loadl_epi64((__m128i *)&a[BPS * 1]);
const __m128i a2 = _mm_loadl_epi64((__m128i *)&a[BPS * 2]);
const __m128i a3 = _mm_loadl_epi64((__m128i *)&a[BPS * 3]);
const __m128i b0 = _mm_loadl_epi64((__m128i *)&b[BPS * 0]);
const __m128i b1 = _mm_loadl_epi64((__m128i *)&b[BPS * 1]);
const __m128i b2 = _mm_loadl_epi64((__m128i *)&b[BPS * 2]);
const __m128i b3 = _mm_loadl_epi64((__m128i *)&b[BPS * 3]);
// Combine pair of lines and convert to 16b.
const __m128i a01 = _mm_unpacklo_epi32(a0, a1);
const __m128i a23 = _mm_unpacklo_epi32(a2, a3);
const __m128i b01 = _mm_unpacklo_epi32(b0, b1);
const __m128i b23 = _mm_unpacklo_epi32(b2, b3);
const __m128i a01s = _mm_unpacklo_epi8(a01, zero);
const __m128i a23s = _mm_unpacklo_epi8(a23, zero);
const __m128i b01s = _mm_unpacklo_epi8(b01, zero);
const __m128i b23s = _mm_unpacklo_epi8(b23, zero);
// Compute differences; (a-b)^2 = (abs(a-b))^2 = (sat8(a-b) + sat8(b-a))^2
// TODO(cduvivier): Dissassemble and figure out why this is fastest. We don't
// need absolute values, there is no need to do calculation
// in 8bit as we are already in 16bit, ... Yet this is what
// benchmarks the fastest!
const __m128i d0 = _mm_subs_epu8(a01s, b01s);
const __m128i d1 = _mm_subs_epu8(b01s, a01s);
const __m128i d2 = _mm_subs_epu8(a23s, b23s);
const __m128i d3 = _mm_subs_epu8(b23s, a23s);
// Square and add them all together.
const __m128i madd0 = _mm_madd_epi16(d0, d0);
const __m128i madd1 = _mm_madd_epi16(d1, d1);
const __m128i madd2 = _mm_madd_epi16(d2, d2);
const __m128i madd3 = _mm_madd_epi16(d3, d3);
const __m128i sum0 = _mm_add_epi32(madd0, madd1);
const __m128i sum1 = _mm_add_epi32(madd2, madd3);
const __m128i sum2 = _mm_add_epi32(sum0, sum1);
int32_t tmp[4];
_mm_storeu_si128((__m128i *)tmp, sum2);
return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
}
//-----------------------------------------------------------------------------
// Texture distortion
//
@ -567,6 +764,8 @@ extern void VP8EncDspInitSSE2(void);
void VP8EncDspInitSSE2(void) {
VP8EncQuantizeBlock = QuantizeBlockSSE2;
VP8ITransform = ITransformSSE2;
VP8FTransform = FTransformSSE2;
VP8SSE4x4 = SSE4x4SSE2;
VP8TDisto4x4 = Disto4x4SSE2;
VP8TDisto16x16 = Disto16x16SSE2;
}