dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2024-12-28 14:38:21 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

dec_neon: add TM8uv

Browse Source 
~68% faster

reuses TM4() adding support for the additional rows, the columns were
already being done.

Change-Id: I6eac17e58cd1c636082bf7281f70f884ec399a6b
This commit is contained in:
James Zern 2014-11-24 21:04:50 -08:00
parent e18571393d
commit 9de9074c92
1 changed files with 44 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

66
src/dsp/dec_neon.c
View File

@ -1281,33 +1281,48 @@ static void DC4(uint8_t* dst) { // DC
} }
} }
static void TM4(uint8_t* dst) { // TrueMotion // TrueMotion (4x4 + 8x8)
static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) {
const uint8x8_t TL = vdup_n_u8(dst[-BPS - 1]); // top-left pixel 'A[-1]' const uint8x8_t TL = vdup_n_u8(dst[-BPS - 1]); // top-left pixel 'A[-1]'
const uint8x8_t T = vld1_u8(dst - BPS); // top row 'A[0..3]' const uint8x8_t T = vld1_u8(dst - BPS); // top row 'A[0..3]'
const int16x8_t d = vreinterpretq_s16_u16(vsubl_u8(T, TL)); // A[c] - A[-1] const int16x8_t d = vreinterpretq_s16_u16(vsubl_u8(T, TL)); // A[c] - A[-1]
const int16x8_t l0 = ConvertU8ToS16(vld1_u8(dst + 0 * BPS - 1)); // left edge int y;
const int16x8_t l1 = ConvertU8ToS16(vld1_u8(dst + 1 * BPS - 1)); for (y = 0; y < size; y += 4) {
const int16x8_t l2 = ConvertU8ToS16(vld1_u8(dst + 2 * BPS - 1)); // left edge
const int16x8_t l3 = ConvertU8ToS16(vld1_u8(dst + 3 * BPS - 1)); const int16x8_t l0 = ConvertU8ToS16(vld1_u8(dst + 0 * BPS - 1));
const int16x8_t L0 = vdupq_lane_s16(vget_low_s16(l0), 0); const int16x8_t l1 = ConvertU8ToS16(vld1_u8(dst + 1 * BPS - 1));
const int16x8_t L1 = vdupq_lane_s16(vget_low_s16(l1), 0); const int16x8_t l2 = ConvertU8ToS16(vld1_u8(dst + 2 * BPS - 1));
const int16x8_t L2 = vdupq_lane_s16(vget_low_s16(l2), 0); const int16x8_t l3 = ConvertU8ToS16(vld1_u8(dst + 3 * BPS - 1));
const int16x8_t L3 = vdupq_lane_s16(vget_low_s16(l3), 0); const int16x8_t L0 = vdupq_lane_s16(vget_low_s16(l0), 0);
const int16x8_t r0 = vaddq_s16(L0, d); // L[r] + A[c] - A[-1] const int16x8_t L1 = vdupq_lane_s16(vget_low_s16(l1), 0);
const int16x8_t r1 = vaddq_s16(L1, d); const int16x8_t L2 = vdupq_lane_s16(vget_low_s16(l2), 0);
const int16x8_t r2 = vaddq_s16(L2, d); const int16x8_t L3 = vdupq_lane_s16(vget_low_s16(l3), 0);
const int16x8_t r3 = vaddq_s16(L3, d); const int16x8_t r0 = vaddq_s16(L0, d); // L[r] + A[c] - A[-1]
// Saturate and store the result. const int16x8_t r1 = vaddq_s16(L1, d);
const uint32x2_t r0_u32 = vreinterpret_u32_u8(vqmovun_s16(r0)); const int16x8_t r2 = vaddq_s16(L2, d);
const uint32x2_t r1_u32 = vreinterpret_u32_u8(vqmovun_s16(r1)); const int16x8_t r3 = vaddq_s16(L3, d);
const uint32x2_t r2_u32 = vreinterpret_u32_u8(vqmovun_s16(r2)); // Saturate and store the result.
const uint32x2_t r3_u32 = vreinterpret_u32_u8(vqmovun_s16(r3)); const uint32x2_t r0_u32 = vreinterpret_u32_u8(vqmovun_s16(r0));
vst1_lane_u32((uint32_t*)(dst + 0 * BPS), r0_u32, 0); const uint32x2_t r1_u32 = vreinterpret_u32_u8(vqmovun_s16(r1));
vst1_lane_u32((uint32_t*)(dst + 1 * BPS), r1_u32, 0); const uint32x2_t r2_u32 = vreinterpret_u32_u8(vqmovun_s16(r2));
vst1_lane_u32((uint32_t*)(dst + 2 * BPS), r2_u32, 0); const uint32x2_t r3_u32 = vreinterpret_u32_u8(vqmovun_s16(r3));
vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3_u32, 0); if (size == 4) {
vst1_lane_u32((uint32_t*)(dst + 0 * BPS), r0_u32, 0);
vst1_lane_u32((uint32_t*)(dst + 1 * BPS), r1_u32, 0);
vst1_lane_u32((uint32_t*)(dst + 2 * BPS), r2_u32, 0);
vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3_u32, 0);
} else {
vst1_u32((uint32_t*)(dst + 0 * BPS), r0_u32);
vst1_u32((uint32_t*)(dst + 1 * BPS), r1_u32);
vst1_u32((uint32_t*)(dst + 2 * BPS), r2_u32);
vst1_u32((uint32_t*)(dst + 3 * BPS), r3_u32);
}
dst += 4 * BPS;
}
} }
static void TM4(uint8_t* dst) { return TrueMotion(dst, 4); }
static void VE4(uint8_t* dst) { // vertical static void VE4(uint8_t* dst) { // vertical
// NB: avoid vld1_u64 here as an alignment hint may be added -> SIGBUS. // NB: avoid vld1_u64 here as an alignment hint may be added -> SIGBUS.
const uint64x1_t A0 = vreinterpret_u64_u8(vld1_u8(dst - BPS - 1)); // top row const uint64x1_t A0 = vreinterpret_u64_u8(vld1_u8(dst - BPS - 1)); // top row
@ -1371,6 +1386,11 @@ static void LD4(uint8_t* dst) { // Down-left
vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3, 0); vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3, 0);
} }
//------------------------------------------------------------------------------
// Chroma
static void TM8uv(uint8_t* dst) { return TrueMotion(dst, 8); }
#endif // WEBP_USE_NEON #endif // WEBP_USE_NEON
//------------------------------------------------------------------------------ //------------------------------------------------------------------------------
@ -1407,5 +1427,7 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitNEON(void) {
VP8PredLuma4[2] = VE4; VP8PredLuma4[2] = VE4;
VP8PredLuma4[4] = RD4; VP8PredLuma4[4] = RD4;
VP8PredLuma4[6] = LD4; VP8PredLuma4[6] = LD4;
VP8PredChroma8[1] = TM8uv;
#endif // WEBP_USE_NEON #endif // WEBP_USE_NEON
} }