dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2024-12-25 21:28:22 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

perform two idct transforms at a time when possible

Browse Source 
patch by Christian Duvivier (cduvivier at google dot com)
This commit is contained in:
Pascal Massimino 2011-06-20 00:22:37 -07:00
parent 01af7b69cd
commit 96ed9ce0fb
4 changed files with 62 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
src/dec/dsp.c
View File

@ -62,7 +62,7 @@ static const int kC1 = 20091 + (1 << 16);
static const int kC2 = 35468;
#define MUL(a, b) (((a) * (b)) >> 16)
static void Transform(const int16_t* in, uint8_t* dst) {
static void TransformOne(const int16_t* in, uint8_t* dst) {
int C[4 * 4], *tmp;
int i;
tmp = C;
@ -102,11 +102,16 @@ static void Transform(const int16_t* in, uint8_t* dst) {
}
#undef MUL
static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
TransformOne(in, dst);
if (do_two) {
TransformOne(in + 16, dst + 4);
}
}
static void TransformUV(const int16_t* in, uint8_t* dst) {
VP8Transform(in + 0 * 16, dst);
VP8Transform(in + 1 * 16, dst + 4);
VP8Transform(in + 2 * 16, dst + 4 * BPS);
VP8Transform(in + 3 * 16, dst + 4 * BPS + 4);
VP8Transform(in + 0 * 16, dst, 1);
VP8Transform(in + 2 * 16, dst + 4 * BPS, 1);
}
static void TransformDC(const int16_t *in, uint8_t* dst) {
@ -129,7 +134,7 @@ static void TransformDCUV(const int16_t* in, uint8_t* dst) {
#undef STORE
// default C implementations:
VP8Idct VP8Transform = Transform;
VP8Idct2 VP8Transform = TransformTwo;
VP8Idct VP8TransformUV = TransformUV;
VP8Idct VP8TransformDC = TransformDC;
VP8Idct VP8TransformDCUV = TransformDCUV;

58
src/dec/dsp_sse2.c
View File

@ -22,7 +22,7 @@ extern "C" {
//-----------------------------------------------------------------------------
// Transforms (Paragraph 14.4)
static void TransformSSE2(const int16_t* in, uint8_t* dst) {
static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
// This implementation makes use of 16-bit fixed point versions of two
// multiply constants:
// K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@ -43,8 +43,9 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst) {
const __m128i k2 = _mm_set1_epi16(-30068);
__m128i T0, T1, T2, T3;
// Load the transform coefficients. The second half of the vectors will just
// contain random value we'll never use nor store.
// Load and concatenate the transform coefficients (we'll do two transforms
// in parallel). In the case of only one transform, the second half of the
// vectors will just contain random value we'll never use nor store.
__m128i in0, in1, in2, in3;
{
in0 = _mm_loadl_epi64((__m128i*)&in[0]);
@ -55,6 +56,20 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst) {
// a01 a11 a21 a31 x x x x
// a02 a12 a22 a32 x x x x
// a03 a13 a23 a33 x x x x
if (do_two) {
const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]);
const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]);
const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]);
const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]);
in0 = _mm_unpacklo_epi64(in0, inB0);
in1 = _mm_unpacklo_epi64(in1, inB1);
in2 = _mm_unpacklo_epi64(in2, inB2);
in3 = _mm_unpacklo_epi64(in3, inB3);
// a00 a10 a20 a30 b00 b10 b20 b30
// a01 a11 a21 a31 b01 b11 b21 b31
// a02 a12 a22 a32 b02 b12 b22 b32
// a03 a13 a23 a33 b03 b13 b23 b33
}
}
// Vertical pass and subsequent transpose.
@ -179,10 +194,20 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst) {
{
const __m128i zero = _mm_set1_epi16(0);
// Load the reference(s).
__m128i dst0 = _mm_cvtsi32_si128(*(int*)&dst[0 * BPS]);
__m128i dst1 = _mm_cvtsi32_si128(*(int*)&dst[1 * BPS]);
__m128i dst2 = _mm_cvtsi32_si128(*(int*)&dst[2 * BPS]);
__m128i dst3 = _mm_cvtsi32_si128(*(int*)&dst[3 * BPS]);
__m128i dst0, dst1, dst2, dst3;
if (do_two) {
// Load eight bytes/pixels per line.
dst0 = _mm_loadl_epi64((__m128i*)&dst[0 * BPS]);
dst1 = _mm_loadl_epi64((__m128i*)&dst[1 * BPS]);
dst2 = _mm_loadl_epi64((__m128i*)&dst[2 * BPS]);
dst3 = _mm_loadl_epi64((__m128i*)&dst[3 * BPS]);
} else {
// Load four bytes/pixels per line.
dst0 = _mm_cvtsi32_si128(*(int*)&dst[0 * BPS]);
dst1 = _mm_cvtsi32_si128(*(int*)&dst[1 * BPS]);
dst2 = _mm_cvtsi32_si128(*(int*)&dst[2 * BPS]);
dst3 = _mm_cvtsi32_si128(*(int*)&dst[3 * BPS]);
}
// Convert to 16b.
dst0 = _mm_unpacklo_epi8(dst0, zero);
dst1 = _mm_unpacklo_epi8(dst1, zero);
@ -198,11 +223,20 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst) {
dst1 = _mm_packus_epi16(dst1, dst1);
dst2 = _mm_packus_epi16(dst2, dst2);
dst3 = _mm_packus_epi16(dst3, dst3);
// Store the results, four bytes/pixels per line.
*((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(dst0);
*((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(dst1);
*((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(dst2);
*((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(dst3);
// Store the results.
if (do_two) {
// Store eight bytes/pixels per line.
_mm_storel_epi64((__m128i*)&dst[0 * BPS], dst0);
_mm_storel_epi64((__m128i*)&dst[1 * BPS], dst1);
_mm_storel_epi64((__m128i*)&dst[2 * BPS], dst2);
_mm_storel_epi64((__m128i*)&dst[3 * BPS], dst3);
} else {
// Store four bytes/pixels per line.
*((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(dst0);
*((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(dst1);
*((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(dst2);
*((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(dst3);
}
}
}

4
src/dec/frame.c
View File

@ -379,7 +379,7 @@ void VP8ReconstructBlock(VP8Decoder* const dec) {
uint8_t* const dst = y_dst + kScan[n];
VP8PredLuma4[dec->imodes_[n]](dst);
if (dec->non_zero_ac_ & (1 << n)) {
VP8Transform(coeffs + n * 16, dst);
VP8Transform(coeffs + n * 16, dst, 0);
} else if (dec->non_zero_ & (1 << n)) { // only DC is present
VP8TransformDC(coeffs + n * 16, dst);
}
@ -391,7 +391,7 @@ void VP8ReconstructBlock(VP8Decoder* const dec) {
for (n = 0; n < 16; n++) {
uint8_t* const dst = y_dst + kScan[n];
if (dec->non_zero_ac_ & (1 << n)) {
VP8Transform(coeffs + n * 16, dst);
VP8Transform(coeffs + n * 16, dst, 0);
} else if (dec->non_zero_ & (1 << n)) { // only DC is present
VP8TransformDC(coeffs + n * 16, dst);
}

4
src/dec/vp8i.h
View File

@ -297,7 +297,9 @@ int VP8DecodeLayer(VP8Decoder* const dec);
// in dsp.c
typedef void (*VP8Idct)(const int16_t* coeffs, uint8_t* dst);
extern VP8Idct VP8Transform;
// when doing two transforms, coeffs is actually int16_t[2][16].
typedef void (*VP8Idct2)(const int16_t* coeffs, uint8_t* dst, int do_two);
extern VP8Idct2 VP8Transform;
extern VP8Idct VP8TransformUV;
extern VP8Idct VP8TransformDC;
extern VP8Idct VP8TransformDCUV;