generic-library/webp
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

lossless_neon: remove predictors 5-13

Browse Source 
operating on single uint32's isn't helped by NEON.
this improves aarch64 performance by ~4%

Change-Id: I9fb25a8962de7b80e893e756ee7c76393cfd40c7
This commit is contained in:
James Zern 2015-07-28 19:44:58 -07:00
parent ca221bbc48
commit 2a010f992a
1 changed files with 0 additions and 131 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

131
src/dsp/lossless_neon.c
View File

@ -139,125 +139,6 @@ static void ConvertBGRAToRGB(const uint32_t* src,
#endif // !WORK_AROUND_GCC
//------------------------------------------------------------------------------
#ifdef WEBP_USE_INTRINSICS
static WEBP_INLINE uint32_t Average2(const uint32_t* const a,
const uint32_t* const b) {
const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
const uint8x8_t avg = vhadd_u8(a0, b0);
return vget_lane_u32(vreinterpret_u32_u8(avg), 0);
}
static WEBP_INLINE uint32_t Average3(const uint32_t* const a,
const uint32_t* const b,
const uint32_t* const c) {
const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
const uint8x8_t c0 = vreinterpret_u8_u64(vcreate_u64(*c));
const uint8x8_t avg1 = vhadd_u8(a0, c0);
const uint8x8_t avg2 = vhadd_u8(avg1, b0);
return vget_lane_u32(vreinterpret_u32_u8(avg2), 0);
}
static WEBP_INLINE uint32_t Average4(const uint32_t* const a,
const uint32_t* const b,
const uint32_t* const c,
const uint32_t* const d) {
const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
const uint8x8_t c0 = vreinterpret_u8_u64(vcreate_u64(*c));
const uint8x8_t d0 = vreinterpret_u8_u64(vcreate_u64(*d));
const uint8x8_t avg1 = vhadd_u8(a0, b0);
const uint8x8_t avg2 = vhadd_u8(c0, d0);
const uint8x8_t avg3 = vhadd_u8(avg1, avg2);
return vget_lane_u32(vreinterpret_u32_u8(avg3), 0);
}
static uint32_t Predictor5(uint32_t left, const uint32_t* const top) {
return Average3(&left, top + 0, top + 1);
}
static uint32_t Predictor6(uint32_t left, const uint32_t* const top) {
return Average2(&left, top - 1);
}
static uint32_t Predictor7(uint32_t left, const uint32_t* const top) {
return Average2(&left, top + 0);
}
static uint32_t Predictor8(uint32_t left, const uint32_t* const top) {
(void)left;
return Average2(top - 1, top + 0);
}
static uint32_t Predictor9(uint32_t left, const uint32_t* const top) {
(void)left;
return Average2(top + 0, top + 1);
}
static uint32_t Predictor10(uint32_t left, const uint32_t* const top) {
return Average4(&left, top - 1, top + 0, top + 1);
}
//------------------------------------------------------------------------------
static WEBP_INLINE uint32_t Select(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
const uint8x8_t bc = vabd_u8(p1, p2); // |b-c|
const uint8x8_t ac = vabd_u8(p0, p2); // |a-c|
const int16x4_t sum_bc = vreinterpret_s16_u16(vpaddl_u8(bc));
const int16x4_t sum_ac = vreinterpret_s16_u16(vpaddl_u8(ac));
const int32x2_t diff = vpaddl_s16(vsub_s16(sum_bc, sum_ac));
const int32_t pa_minus_pb = vget_lane_s32(diff, 0);
return (pa_minus_pb <= 0) ? *c0 : *c1;
}
static uint32_t Predictor11(uint32_t left, const uint32_t* const top) {
return Select(top + 0, &left, top - 1);
}
static WEBP_INLINE uint32_t ClampedAddSubtractFull(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
const uint16x8_t sum0 = vaddl_u8(p0, p1); // add and widen
const uint16x8_t sum1 = vqsubq_u16(sum0, vmovl_u8(p2)); // widen and subtract
const uint8x8_t out = vqmovn_u16(sum1); // narrow and clamp
return vget_lane_u32(vreinterpret_u32_u8(out), 0);
}
static uint32_t Predictor12(uint32_t left, const uint32_t* const top) {
return ClampedAddSubtractFull(&left, top + 0, top - 1);
}
static WEBP_INLINE uint32_t ClampedAddSubtractHalf(const uint32_t* const c0,
const uint32_t* const c1,
const uint32_t* const c2) {
const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
const uint8x8_t avg = vhadd_u8(p0, p1); // Average(c0,c1)
const uint8x8_t ab = vshr_n_u8(vqsub_u8(avg, p2), 1); // (a-b)>>1 saturated
const uint8x8_t ba = vshr_n_u8(vqsub_u8(p2, avg), 1); // (b-a)>>1 saturated
const uint8x8_t out = vqsub_u8(vqadd_u8(avg, ab), ba);
return vget_lane_u32(vreinterpret_u32_u8(out), 0);
}
static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
return ClampedAddSubtractHalf(&left, top + 0, top - 1);
}
#endif // WEBP_USE_INTRINSICS
//------------------------------------------------------------------------------
// Subtract-Green Transform
@ -318,18 +199,6 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitNEON(void) {
VP8LConvertBGRAToBGR = ConvertBGRAToBGR;
VP8LConvertBGRAToRGB = ConvertBGRAToRGB;
#ifdef WEBP_USE_INTRINSICS
VP8LPredictors[5] = Predictor5;
VP8LPredictors[6] = Predictor6;
VP8LPredictors[7] = Predictor7;
VP8LPredictors[8] = Predictor8;
VP8LPredictors[9] = Predictor9;
VP8LPredictors[10] = Predictor10;
VP8LPredictors[11] = Predictor11;
VP8LPredictors[12] = Predictor12;
VP8LPredictors[13] = Predictor13;
#endif
VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
}