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

NEON: faster fancy upsampling

Browse Source 
2-3% faster decoding overall

Change-Id: I2c53e50dc7e0ade5245cff8cc5d7b96a14062955
This commit is contained in:
Pascal Massimino 2016-12-21 15:23:54 +01:00
parent 58fc507842
commit b016cb91c5
1 changed files with 24 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

61
src/dsp/upsampling_neon.c
View File

@ -28,47 +28,34 @@
// U/V upsampling
// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels.
#define UPSAMPLE_16PIXELS(r1, r2, out) { \
uint8x8_t a = vld1_u8(r1); \
uint8x8_t b = vld1_u8(r1 + 1); \
uint8x8_t c = vld1_u8(r2); \
uint8x8_t d = vld1_u8(r2 + 1); \
\
uint16x8_t al = vshll_n_u8(a, 1); \
uint16x8_t bl = vshll_n_u8(b, 1); \
uint16x8_t cl = vshll_n_u8(c, 1); \
uint16x8_t dl = vshll_n_u8(d, 1); \
\
uint8x8_t diag1, diag2; \
uint16x8_t sl; \
\
#define UPSAMPLE_16PIXELS(r1, r2, out) do { \
const uint8x8_t a = vld1_u8(r1 + 0); \
const uint8x8_t b = vld1_u8(r1 + 1); \
const uint8x8_t c = vld1_u8(r2 + 0); \
const uint8x8_t d = vld1_u8(r2 + 1); \
/* a + b + c + d */ \
sl = vaddl_u8(a, b); \
sl = vaddw_u8(sl, c); \
sl = vaddw_u8(sl, d); \
const uint16x8_t ad = vaddl_u8(a, d); \
const uint16x8_t bc = vaddl_u8(b, c); \
const uint16x8_t abcd = vaddq_u16(ad, bc); \
/* 3a + b + c + 3d */ \
const uint16x8_t al = vaddq_u16(abcd, vshlq_n_u16(ad, 1)); \
/* a + 3b + 3c + d */ \
const uint16x8_t bl = vaddq_u16(abcd, vshlq_n_u16(bc, 1)); \
\
al = vaddq_u16(sl, al); /* 3a + b + c + d */ \
bl = vaddq_u16(sl, bl); /* a + 3b + c + d */ \
const uint8x8_t diag2 = vshrn_n_u16(al, 3); \
const uint8x8_t diag1 = vshrn_n_u16(bl, 3); \
\
al = vaddq_u16(al, dl); /* 3a + b + c + 3d */ \
bl = vaddq_u16(bl, cl); /* a + 3b + 3c + d */ \
const uint8x8_t A = vrhadd_u8(a, diag1); \
const uint8x8_t B = vrhadd_u8(b, diag2); \
const uint8x8_t C = vrhadd_u8(c, diag2); \
const uint8x8_t D = vrhadd_u8(d, diag1); \
\
diag2 = vshrn_n_u16(al, 3); \
diag1 = vshrn_n_u16(bl, 3); \
\
a = vrhadd_u8(a, diag1); \
b = vrhadd_u8(b, diag2); \
c = vrhadd_u8(c, diag2); \
d = vrhadd_u8(d, diag1); \
\
{ \
uint8x8x2_t a_b, c_d; \
INIT_VECTOR2(a_b, a, b); \
INIT_VECTOR2(c_d, c, d); \
vst2_u8(out, a_b); \
vst2_u8(out + 32, c_d); \
} \
}
uint8x8x2_t A_B, C_D; \
INIT_VECTOR2(A_B, A, B); \
INIT_VECTOR2(C_D, C, D); \
vst2_u8(out + 0, A_B); \
vst2_u8(out + 32, C_D); \
} while (0)
// Turn the macro into a function for reducing code-size when non-critical
static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,