inline uint32x2_t vsub(const uint32x2_t & v0, const uint32x2_t & v1) { return vsub_u32(v0, v1); }
unsigned int vp8_variance_halfpixvar16x16_hv_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {
int i;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
int16x4_t d0s16, d1s16, d2s16, d3s16, d10s16, d11s16, d12s16, d13s16;
int16x4_t d18s16, d19s16, d20s16, d21s16, d22s16, d23s16, d24s16, d25s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64, d2s64, d3s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8;
uint16x8_t q0u16, q1u16, q5u16, q6u16, q9u16, q10u16, q11u16, q12u16;
int32x4_t q13s32, q14s32, q15s32;
int64x2_t q0s64, q1s64, q5s64;
q13s32 = vdupq_n_s32(0);
q14s32 = vdupq_n_s32(0);
q15s32 = vdupq_n_s32(0);
q0u8 = vld1q_u8(src_ptr);
q1u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q1u8 = vextq_u8(q0u8, q1u8, 1);
q0u8 = vrhaddq_u8(q0u8, q1u8);
for (i = 0; i < 4; i++) { // vp8_filt_fpo16x16s_4_0_loop_neon
q2u8 = vld1q_u8(src_ptr);
q3u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q4u8 = vld1q_u8(src_ptr);
q5u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q6u8 = vld1q_u8(src_ptr);
q7u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q8u8 = vld1q_u8(src_ptr);
q9u8 = vld1q_u8(src_ptr + 16);
src_ptr += source_stride;
q3u8 = vextq_u8(q2u8, q3u8, 1);
q5u8 = vextq_u8(q4u8, q5u8, 1);
q7u8 = vextq_u8(q6u8, q7u8, 1);
q9u8 = vextq_u8(q8u8, q9u8, 1);
q1u8 = vrhaddq_u8(q2u8, q3u8);
q2u8 = vrhaddq_u8(q4u8, q5u8);
q3u8 = vrhaddq_u8(q6u8, q7u8);
q4u8 = vrhaddq_u8(q8u8, q9u8);
q0u8 = vrhaddq_u8(q0u8, q1u8);
q1u8 = vrhaddq_u8(q1u8, q2u8);
q2u8 = vrhaddq_u8(q2u8, q3u8);
q3u8 = vrhaddq_u8(q3u8, q4u8);
q5u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q6u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q7u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q8u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
d0u8 = vget_low_u8(q0u8);
d1u8 = vget_high_u8(q0u8);
d2u8 = vget_low_u8(q1u8);
d3u8 = vget_high_u8(q1u8);
d4u8 = vget_low_u8(q2u8);
d5u8 = vget_high_u8(q2u8);
d6u8 = vget_low_u8(q3u8);
d7u8 = vget_high_u8(q3u8);
q9u16 = vsubl_u8(d0u8, vget_low_u8(q5u8));
q10u16 = vsubl_u8(d1u8, vget_high_u8(q5u8));
q11u16 = vsubl_u8(d2u8, vget_low_u8(q6u8));
q12u16 = vsubl_u8(d3u8, vget_high_u8(q6u8));
q0u16 = vsubl_u8(d4u8, vget_low_u8(q7u8));
q1u16 = vsubl_u8(d5u8, vget_high_u8(q7u8));
q5u16 = vsubl_u8(d6u8, vget_low_u8(q8u8));
q6u16 = vsubl_u8(d7u8, vget_high_u8(q8u8));
d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q9u16));
q14s32 = vmlal_s16(q14s32, d18s16, d18s16);
q15s32 = vmlal_s16(q15s32, d19s16, d19s16);
d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q10u16));
q14s32 = vmlal_s16(q14s32, d20s16, d20s16);
q15s32 = vmlal_s16(q15s32, d21s16, d21s16);
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q11u16));
q14s32 = vmlal_s16(q14s32, d22s16, d22s16);
q15s32 = vmlal_s16(q15s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q12u16));
q14s32 = vmlal_s16(q14s32, d24s16, d24s16);
q15s32 = vmlal_s16(q15s32, d25s16, d25s16);
d0s16 = vreinterpret_s16_u16(vget_low_u16(q0u16));
d1s16 = vreinterpret_s16_u16(vget_high_u16(q0u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q0u16));
q14s32 = vmlal_s16(q14s32, d0s16, d0s16);
q15s32 = vmlal_s16(q15s32, d1s16, d1s16);
d2s16 = vreinterpret_s16_u16(vget_low_u16(q1u16));
d3s16 = vreinterpret_s16_u16(vget_high_u16(q1u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q1u16));
q14s32 = vmlal_s16(q14s32, d2s16, d2s16);
q15s32 = vmlal_s16(q15s32, d3s16, d3s16);
d10s16 = vreinterpret_s16_u16(vget_low_u16(q5u16));
d11s16 = vreinterpret_s16_u16(vget_high_u16(q5u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q5u16));
q14s32 = vmlal_s16(q14s32, d10s16, d10s16);
q15s32 = vmlal_s16(q15s32, d11s16, d11s16);
d12s16 = vreinterpret_s16_u16(vget_low_u16(q6u16));
d13s16 = vreinterpret_s16_u16(vget_high_u16(q6u16));
q13s32 = vpadalq_s16(q13s32, vreinterpretq_s16_u16(q6u16));
q14s32 = vmlal_s16(q14s32, d12s16, d12s16);
q15s32 = vmlal_s16(q15s32, d13s16, d13s16);
q0u8 = q4u8;
}
q15s32 = vaddq_s32(q14s32, q15s32);
q0s64 = vpaddlq_s32(q13s32);
q1s64 = vpaddlq_s32(q15s32);
d0s64 = vget_low_s64(q0s64);
d1s64 = vget_high_s64(q0s64);
d2s64 = vget_low_s64(q1s64);
d3s64 = vget_high_s64(q1s64);
d0s64 = vadd_s64(d0s64, d1s64);
d1s64 = vadd_s64(d2s64, d3s64);
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
void mdrc5b_apply_limiter(MDRC5B_LOCAL_STRUCT_T *HeapPtr)
{
unsigned int LaIdx;
unsigned int NumMainCh;
unsigned int Samples;
unsigned int ch, k, n;
MMlong *Ptr;
MMlong *Ptr2;
MMlong *MemOutPtr;
MMshort PeakdB;
MMlong PeakMax;
int RmsMeasure;
MMshort LimiterAtCoef;
MMshort LimiterReCoef;
MMshort LimiterGainMant[MDRC5B_BLOCK_SIZE + 1];
MMshort LimiterGainExp;
MMshort LimiterTargetGaindB;
unsigned int LimiterHoldRem;
unsigned int LimiterHtSamp;
MMshort Exp, TargetGain;
MMshort MaxShiftBits;
unsigned int lookahead_len = (unsigned int) HeapPtr->LimiterLALen;
unsigned int cpt1, cpt2;
uint32x2x2_t Temp_u32x2x2;
uint32x2_t Ldbits_u32x2, Ldbits2_u32x2;
uint32x2_t bsl_u32x2;
int32x2_t LimGainMant_32x2;
int64x2_t TempX_64x2, MemOut_64x2;
int64x2_t Tmp_64x2;
int64x2_t LimiterGainExp_64x2, sample_64x2;
int64x1_t TempX_64x1, sample_64x1;
int32_t *LimiterGainMant_ptr;
int32x2_t Tmp_32x2, Ldbits_32x2, n_32x2;
int32x2_t TempX_low_32x2, TempX_high_32x2;
int32x2x2_t Tmp_32x2x2;
int64x1_t Peak_64x1, PeakMax_64x1, Tmp_64x1, diffX_64x1;
int64x1_t Peak_scale_pow_64x1, Peak_scale_64x1, Zero_s64x1;
int64x1_t MaxShiftBits_neg_64x1, MaxShiftBits_hd_64x1;
int64x2_t diffX_64x2;
uint64x1_t bsl_u64x1;
int32x2_t LimiterPeakCoef_32x2, diffX_low_32x2, diffX_high_32x2;
int32x2_t TargetGain_32x2;
uint32x2x2_t Peak_u32x2x2;
uint32x2_t Peak_exp_u32x2, Peak_exp2_u32x2, Peak_mant_u32x2;
int32x2_t x_32x2, xn_32x2, PeakdB_32x2, Peak_exp_32x2;
int32x2_t LimiterTargetGaindB_32x2, Exp_32x2, LimiterCoef_32x2;
int32x4_t Tmp_32x4;
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_APPLY_LIMITER)
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT)
Samples = (unsigned int) HeapPtr->BlockSize;
NumMainCh = (unsigned int) HeapPtr->NumMainCh;
TempX_64x2 = vdupq_n_s64(0);
for(ch = 0; ch < NumMainCh; ch++)
{
Ptr = HeapPtr->MainInBuf[ch];
// compute the number of bits needs to be shifted to avoid overflow
for(k = (Samples >> 1); k > 0; k--)
{
sample_64x2 = vld1q_s64(Ptr);
Ptr +=2;
sample_64x2 = veorq_s64(sample_64x2, vshrq_n_s64(sample_64x2, 63));
TempX_64x2 = vorrq_s64(TempX_64x2, sample_64x2);
}
if(Samples & 1)
{
sample_64x1 = vld1_s64(Ptr);
sample_64x1 = veor_s64(sample_64x1, vshr_n_s64(sample_64x1, 63));
TempX_64x2 = vorrq_s64(TempX_64x2, vcombine_s64(sample_64x1, sample_64x1));
}
}
TempX_64x1 = vorr_s64(vget_low_s64(TempX_64x2), vget_high_s64(TempX_64x2));
Temp_u32x2x2 = vuzp_u32(vreinterpret_u32_s64(TempX_64x1), vreinterpret_u32_s64(TempX_64x1));
bsl_u32x2 = vceq_u32(Temp_u32x2x2.val[1], vdup_n_u32(0)); // MSB == 0 ?
// use clz instead of cls because we are sure that input value is positive
// and because cls(LSB) could be wrong (if MSB is equal to 0 and bit 31 of LSL is 1)
// thus clz result will be 1 more than cls result (that's why you may see (Ldbits - 1)
// instead of Ldbits below)
Ldbits_u32x2 = vadd_u32(vclz_u32(Temp_u32x2x2.val[0]), vdup_n_u32(32)); // clz(LSB)+32
Ldbits2_u32x2 = vclz_u32(Temp_u32x2x2.val[1]); // clz(MSB)
Ldbits_u32x2 = vbsl_u32(bsl_u32x2, Ldbits_u32x2, Ldbits2_u32x2); // MSB == 0 ? clz(LSB)+32 : clz(MSB)
bsl_u32x2 = vceq_u32(Ldbits_u32x2, vdup_n_u32(64)); // Ldbits == 64 ? (i.e. TempX == 0 ?)
// the aim of MaxShiftBits is that sample will be shifted so that it occupies
// 24 significant bits for 24 bits samples or 32 significant bits for 32 bits samples
// but we are in 64 bits architecture on CA9/NEON
// so we must right shift of ((64 - 24) - (Ldbits - 1)) bits for 24 bits samples
// or of ((64 - 32) - (Ldbits - 1)) bits for 32 bits samples
// and we add 1 because it was done this way on MMDSP (I don't know why !)
#ifdef SAMPLES_24_BITS
// MaxShiftBits = ((64 - 24) - (Ldbits - 1)) + 1
// = 42 - Ldbits
Ldbits_32x2 = vsub_s32(vdup_n_s32(42), vreinterpret_s32_u32(Ldbits_u32x2));
#else // SAMPLES_24_BITS
// MaxShiftBits = ((64 - 32) - (Ldbits - 1)) + 1
// = 34 - Ldbits
Ldbits_32x2 = vsub_s32(vdup_n_s32(34), vreinterpret_s32_u32(Ldbits_u32x2));
#endif // SAMPLES_24_BITS
Ldbits_32x2 = vmax_s32(vdup_n_s32(1), Ldbits_32x2);
Ldbits_32x2 = vbsl_s32(bsl_u32x2, vdup_n_s32(1), Ldbits_32x2); // if(TempX == 0) Ldbits = 1
MaxShiftBits = vget_lane_s32(Ldbits_32x2, 0);
STOP_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT)
#ifdef DEBUG_LIMITER_OUTPUT
if((debug_cpt_samples >= DEBUG_CPT_MIN) && (debug_cpt_samples <= DEBUG_CPT_MAX))
{
char string[100];
debug_write_string("MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT\n");
sprintf(string, "MaxShiftBits=%d\n", MaxShiftBits);
debug_write_string(string);
}
#endif // DEBUG_LIMITER_OUTPUT
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_INSERT_NEW_SUBBAND)
// insert the new subband samples into the lookahead buffers
RmsMeasure = HeapPtr->Limiter.RmsMeasure;
LaIdx = (unsigned int) HeapPtr->LimiterLaIdx;
if(LaIdx + Samples >= lookahead_len)
{
cpt1 = lookahead_len - LaIdx;
cpt2 = Samples - cpt1;
// update index
HeapPtr->LimiterLaIdx = (int) cpt2;
}
else
{
cpt1 = Samples;
cpt2 = 0;
// update index
HeapPtr->LimiterLaIdx = (int) (LaIdx + Samples);
}
LimiterPeakCoef_32x2 = vdup_n_s32(HeapPtr->LimiterPeakAtCoef); // LimiterPeakAtCoef, LimiterPeakAtCoef
LimiterPeakCoef_32x2 = vset_lane_s32(HeapPtr->LimiterPeakReCoef, LimiterPeakCoef_32x2, 1); // LimiterPeakAtCoef, LimiterPeakReCoef
Peak_scale_64x1 = vdup_n_s64(HeapPtr->PrevShiftBits - MaxShiftBits);
Peak_scale_pow_64x1 = vshl_n_s64(Peak_scale_64x1, 1);
MaxShiftBits_neg_64x1 = vdup_n_s64(-MaxShiftBits);
#ifdef SAMPLES_24_BITS
MaxShiftBits_hd_64x1 = vdup_n_s64(24 - MaxShiftBits);
#else // SAMPLES_24_BITS
MaxShiftBits_hd_64x1 = vdup_n_s64(32 - MaxShiftBits);
#endif // SAMPLES_24_BITS
PeakMax_64x1 = vdup_n_s64(0);
for(ch = 0; ch < NumMainCh; ch++)
{
Ptr = HeapPtr->MainInBuf[ch];
Ptr2 = HeapPtr->LimiterLABuf[ch] + LaIdx; // go to the first valid sample
Peak_64x1 = vdup_n_s64(HeapPtr->LimiterPeak[ch]);
if(RmsMeasure)
{
// compensate Peak according to the previous shift bits
Peak_64x1 = vqrshl_s64(Peak_64x1, Peak_scale_pow_64x1); // neg value => shift right rounding
// rms measure
for(k = cpt1; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
Tmp_64x1 = vqrshl_s64(Tmp_64x1, MaxShiftBits_neg_64x1);
Tmp_64x2 = vcombine_s64(Tmp_64x1, Tmp_64x1);
Tmp_32x2x2 = vuzp_s32(vget_low_s32(vreinterpretq_s32_s64(Tmp_64x2)), vget_high_s32(vreinterpretq_s32_s64(Tmp_64x2)));
Tmp_32x2 = Tmp_32x2x2.val[0]; // LSB of Tmp_64x2 (MSB is dummy)
TempX_64x2 = vqdmull_s32(Tmp_32x2, Tmp_32x2);
TempX_64x1 = vget_low_s64(TempX_64x2);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
Ptr2 = HeapPtr->LimiterLABuf[ch];
for(k = cpt2; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
Tmp_64x1 = vqrshl_s64(Tmp_64x1, MaxShiftBits_neg_64x1);
Tmp_64x2 = vcombine_s64(Tmp_64x1, Tmp_64x1);
Tmp_32x2x2 = vuzp_s32(vget_low_s32(vreinterpretq_s32_s64(Tmp_64x2)), vget_high_s32(vreinterpretq_s32_s64(Tmp_64x2)));
Tmp_32x2 = Tmp_32x2x2.val[0]; // LSB of Tmp_64x2 (MSB is dummy)
TempX_64x2 = vqdmull_s32(Tmp_32x2, Tmp_32x2);
TempX_64x1 = vget_low_s64(TempX_64x2);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
}
else
{
// compensate Peak according to the previous shift bits
Peak_64x1 = vqrshl_s64(Peak_64x1, Peak_scale_64x1);
// amplitude measure
Zero_s64x1 = vdup_n_s64(0);
for(k = cpt1; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Tmp_64x1)
TempX_64x1 = vqsub_s64(Zero_s64x1, Tmp_64x1); // -Tmp_64x1
TempX_64x1 = vbsl_s64(bsl_u64x1, TempX_64x1, Tmp_64x1);
TempX_64x1 = vqrshl_s64(TempX_64x1, MaxShiftBits_hd_64x1);
TempX_64x2 = vcombine_s64(TempX_64x1, TempX_64x1);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
Ptr2 = HeapPtr->LimiterLABuf[ch];
for(k = cpt2; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Tmp_64x1)
TempX_64x1 = vqsub_s64(Zero_s64x1, Tmp_64x1); // -Tmp_64x1
TempX_64x1 = vbsl_s64(bsl_u64x1, TempX_64x1, Tmp_64x1);
TempX_64x1 = vqrshl_s64(TempX_64x1, MaxShiftBits_hd_64x1);
TempX_64x2 = vcombine_s64(TempX_64x1, TempX_64x1);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
}
HeapPtr->LimiterPeak[ch] = vget_lane_s64(Peak_64x1, 0); // save history
} // for(ch = 0...)
PeakMax = vget_lane_s64(PeakMax_64x1, 0);
HeapPtr->PrevShiftBits = MaxShiftBits;
STOP_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_INSERT_NEW_SUBBAND)
if(PeakMax < MDRC5B_ALMOST_ZERO_THRESH)
{
PeakdB = (MDRC5B_POWER_DB_MINUS_INF << 16); // 8.16, [-128.0, 127.0] dB
}
else
{
Peak_u32x2x2 = vuzp_u32(vreinterpret_u32_s64(PeakMax_64x1), vreinterpret_u32_s64(PeakMax_64x1));
bsl_u32x2 = vceq_u32(Peak_u32x2x2.val[1], vdup_n_u32(0));
Peak_exp_u32x2 = vadd_u32(vclz_u32(Peak_u32x2x2.val[0]), vdup_n_u32(32));
Peak_exp2_u32x2 = vclz_u32(Peak_u32x2x2.val[1]);
Peak_exp_u32x2 = vbsl_u32(bsl_u32x2, Peak_exp_u32x2, Peak_exp2_u32x2);
Peak_mant_u32x2 = vrshrn_n_u64(vshlq_u64(vreinterpretq_u64_s64(vcombine_s64(PeakMax_64x1, PeakMax_64x1)), vreinterpretq_s64_u64(vmovl_u32(Peak_exp_u32x2))), 32);
// if(Peak_mant >= sqrt(0.5))
// {
// Peak_exp--;
// Peak_mant >>= 1;
// }
bsl_u32x2 = vcge_u32(Peak_mant_u32x2, vdup_n_u32(0xB504F334));
Peak_exp_u32x2 = vbsl_u32(bsl_u32x2, vsub_u32(Peak_exp_u32x2, vdup_n_u32(1)), Peak_exp_u32x2);
Peak_mant_u32x2 = vbsl_u32(bsl_u32x2, vrshr_n_u32(Peak_mant_u32x2, 1), Peak_mant_u32x2);
Peak_exp_32x2 = vreinterpret_s32_u32(Peak_exp_u32x2);
#ifdef SAMPLES_24_BITS
// correction of 16 bits if input samples are 24 bits
Peak_exp_32x2 = vsub_s32(Peak_exp_32x2, vdup_n_s32(16));
#endif // SAMPLES_24_BITS
// at this point : sqrt(0.5)/2 <= Peak_mant < sqrt(0.5)
//
// ln(1+x) = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9 - x^10/10 ... accuracy OK if |x| < 0.5
// sqrt(0.5)/2 <= Peak_mant < sqrt(0.5) => sqrt(0.5)-1 <= 2*Peak_mant-1 < 2*sqrt(0.5)-1
// => ln(Peak_mant) = ln(1+x)-ln(2) with x=2*Peak_mant-1, i.e. |x| < 0.414214...
// x=2*PeakMax_mant-1 in Q31
// => sqrt(0.5)-1 <= x < 2*sqrt(0.5)-1
x_32x2 = vreinterpret_s32_u32(vsub_u32(Peak_mant_u32x2, vdup_n_u32(0x80000000)));
PeakdB_32x2 = x_32x2; // PeakdB = x
xn_32x2 = vqrdmulh_s32(x_32x2, x_32x2); // xn = x^2
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 1)); // PeakdB = x - x^2/2
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^3
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x2AAAAAAB))); // PeakdB = x - x^2/2 + x^3/3
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^4
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 2)); // PeakdB = x - x^2/2 + x^3/3 - x^4/4
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^5
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x1999999A))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^6
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x15555555))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^7
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x12492492))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^8
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 3)); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^9
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x0E38E38E))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^10
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x0CCCCCCD))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9 - x^10/10
// at this point : PeakMaxdB contains ln(1+x) in Q31
if(RmsMeasure)
{
// dB(power) = 10*log10(power)
// PeakMaxdB = 10*log10(PeakMax)+20*log10(2)*(HEADROOM+MaxShiftBits)
// = 10*ln(PeakMax)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant*2^(-PeakMax_exp))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*(ln(PeakMax_mant)-PeakMax_exp*ln(2))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant)-PeakMax_exp*10*ln(2)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant)+10*ln(2)/ln(10)*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
//
// => RmsdB = 10/ln(10)*ln(1+x)+10*ln(2)/ln(10)*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
// => RmsdB (Q16) = 0x457CB*ln(1+x)+0x302A3*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
// fractional mutiply 0x457CB*ln(1+x) in Q16
PeakdB_32x2 = vqrdmulh_s32(PeakdB_32x2, vdup_n_s32(0x457CB));
// PeakdB_exp = 2*(HEADROOM+MaxShiftBits)-PeakdB_exp
Peak_exp_32x2 = vsub_s32(vdup_n_s32(2 * (HEADROOM + MaxShiftBits)), Peak_exp_32x2);
// PeakMaxdB final value (integer mac 0x302A3*PeakdB_exp)
PeakdB_32x2 = vmla_s32(PeakdB_32x2, Peak_exp_32x2, vdup_n_s32(0x302A3));
}
else
{
// dB(power) = 20*log10(abs)
// PeakMaxdB = 20*log10(PeakMax)+20*log10(2)*(HEADROOM+MaxShiftBits)
// = 20*ln(PeakMax)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant*2^(-PeakMax_exp))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*(ln(PeakMax_mant)-PeakMax_exp*ln(2))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant)-PeakMax_exp*20*ln(2)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits-PeakMax_exp)
//
// => RmsdB = 20/ln(10)*ln(1+x)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits-PeakMax_exp)
// => RmsdB (Q16) = 0x8AF96*ln(1+x)+0x60546*(HEADROOM+MaxShiftBits-PeakMax_exp)
// fractional mutiply 0x8AF96*ln(1+x) in Q16
PeakdB_32x2 = vqrdmulh_s32(PeakdB_32x2, vdup_n_s32(0x8AF96));
// PeakdB_exp = HEADROOM+MaxShiftBits-PeakdB_exp
Peak_exp_32x2 = vsub_s32(vdup_n_s32(HEADROOM + MaxShiftBits), Peak_exp_32x2);
// PeakMaxdB final value (integer mac 0x60546*PeakdB_exp)
PeakdB_32x2 = vmla_s32(PeakdB_32x2, Peak_exp_32x2, vdup_n_s32(0x60546));
}
PeakdB = vget_lane_s32(PeakdB_32x2, 0);
}
#ifdef DEBUG_LIMITER_OUTPUT
if((debug_cpt_samples >= DEBUG_CPT_MIN) && (debug_cpt_samples <= DEBUG_CPT_MAX))
{
char string[100];
debug_write_string("MRDC5B_LIMITER_PEAKMAX_PEAKDB\n");
sprintf(string, "PeakMax=0x%012llX, HEADROOM+MaxShiftBits=%d => PeakdB=0x%06X\n",
#ifdef SAMPLES_24_BITS
PeakMax & 0xFFFFFFFFFFFFLL,
#else // SAMPLES_24_BITS
(PeakMax >> 16) & 0xFFFFFFFFFFFFLL,
#endif // SAMPLES_24_BITS
HEADROOM + MaxShiftBits,
PeakdB & 0xFFFFFF);
debug_write_string(string);
}
unsigned int vp8_sub_pixel_variance16x16_neon_func(
const unsigned char *src_ptr,
int src_pixels_per_line,
int xoffset,
int yoffset,
const unsigned char *dst_ptr,
int dst_pixels_per_line,
unsigned int *sse) {
int i;
DECLARE_ALIGNED_ARRAY(16, unsigned char, tmp, 528);
unsigned char *tmpp;
unsigned char *tmpp2;
uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
uint8x8_t d10u8, d11u8, d12u8, d13u8, d14u8, d15u8, d16u8, d17u8, d18u8;
uint8x8_t d19u8, d20u8, d21u8;
int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64, d2s64, d3s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8;
uint8x16_t q10u8, q11u8, q12u8, q13u8, q14u8, q15u8;
uint16x8_t q1u16, q2u16, q3u16, q4u16, q5u16, q6u16, q7u16, q8u16;
uint16x8_t q9u16, q10u16, q11u16, q12u16, q13u16, q14u16;
int32x4_t q8s32, q9s32, q10s32;
int64x2_t q0s64, q1s64, q5s64;
tmpp2 = tmp + 272;
tmpp = tmp;
if (xoffset == 0) { // secondpass_bfilter16x16_only
d0u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][1]);
q11u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
for (i = 4; i > 0; i--) {
q12u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q13u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q14u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
q15u8 = vld1q_u8(src_ptr);
src_ptr += src_pixels_per_line;
__builtin_prefetch(src_ptr);
__builtin_prefetch(src_ptr + src_pixels_per_line);
__builtin_prefetch(src_ptr + src_pixels_per_line * 2);
q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8);
q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8);
q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8);
q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8);
q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8);
q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8);
q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8);
q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8);
q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8);
q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8);
q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8);
q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8);
q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8);
q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8);
q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8);
q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8);
d2u8 = vqrshrn_n_u16(q1u16, 7);
d3u8 = vqrshrn_n_u16(q2u16, 7);
d4u8 = vqrshrn_n_u16(q3u16, 7);
d5u8 = vqrshrn_n_u16(q4u16, 7);
d6u8 = vqrshrn_n_u16(q5u16, 7);
d7u8 = vqrshrn_n_u16(q6u16, 7);
d8u8 = vqrshrn_n_u16(q7u16, 7);
d9u8 = vqrshrn_n_u16(q8u16, 7);
q1u8 = vcombine_u8(d2u8, d3u8);
q2u8 = vcombine_u8(d4u8, d5u8);
q3u8 = vcombine_u8(d6u8, d7u8);
q4u8 = vcombine_u8(d8u8, d9u8);
q11u8 = q15u8;
vst1q_u8((uint8_t *)tmpp2, q1u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q2u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q3u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q4u8);
tmpp2 += 16;
}
} else if (yoffset == 0) { // firstpass_bfilter16x16_only
d0u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][1]);
for (i = 4; i > 0 ; i--) {
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
__builtin_prefetch(src_ptr);
__builtin_prefetch(src_ptr + src_pixels_per_line);
__builtin_prefetch(src_ptr + src_pixels_per_line * 2);
q7u16 = vmull_u8(d2u8, d0u8);
q8u16 = vmull_u8(d3u8, d0u8);
q9u16 = vmull_u8(d5u8, d0u8);
q10u16 = vmull_u8(d6u8, d0u8);
q11u16 = vmull_u8(d8u8, d0u8);
q12u16 = vmull_u8(d9u8, d0u8);
q13u16 = vmull_u8(d11u8, d0u8);
q14u16 = vmull_u8(d12u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
d11u8 = vext_u8(d11u8, d12u8, 1);
q7u16 = vmlal_u8(q7u16, d2u8, d1u8);
q9u16 = vmlal_u8(q9u16, d5u8, d1u8);
q11u16 = vmlal_u8(q11u16, d8u8, d1u8);
q13u16 = vmlal_u8(q13u16, d11u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
d12u8 = vext_u8(d12u8, d13u8, 1);
q8u16 = vmlal_u8(q8u16, d3u8, d1u8);
q10u16 = vmlal_u8(q10u16, d6u8, d1u8);
q12u16 = vmlal_u8(q12u16, d9u8, d1u8);
q14u16 = vmlal_u8(q14u16, d12u8, d1u8);
d14u8 = vqrshrn_n_u16(q7u16, 7);
d15u8 = vqrshrn_n_u16(q8u16, 7);
d16u8 = vqrshrn_n_u16(q9u16, 7);
d17u8 = vqrshrn_n_u16(q10u16, 7);
d18u8 = vqrshrn_n_u16(q11u16, 7);
d19u8 = vqrshrn_n_u16(q12u16, 7);
d20u8 = vqrshrn_n_u16(q13u16, 7);
d21u8 = vqrshrn_n_u16(q14u16, 7);
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
q10u8 = vcombine_u8(d20u8, d21u8);
vst1q_u8((uint8_t *)tmpp2, q7u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q8u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q9u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q10u8);
tmpp2 += 16;
}
} else {
d0u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[xoffset][1]);
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
// First Pass: output_height lines x output_width columns (17x16)
for (i = 3; i > 0; i--) {
q7u16 = vmull_u8(d2u8, d0u8);
q8u16 = vmull_u8(d3u8, d0u8);
q9u16 = vmull_u8(d5u8, d0u8);
q10u16 = vmull_u8(d6u8, d0u8);
q11u16 = vmull_u8(d8u8, d0u8);
q12u16 = vmull_u8(d9u8, d0u8);
q13u16 = vmull_u8(d11u8, d0u8);
q14u16 = vmull_u8(d12u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
d11u8 = vext_u8(d11u8, d12u8, 1);
q7u16 = vmlal_u8(q7u16, d2u8, d1u8);
q9u16 = vmlal_u8(q9u16, d5u8, d1u8);
q11u16 = vmlal_u8(q11u16, d8u8, d1u8);
q13u16 = vmlal_u8(q13u16, d11u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
d12u8 = vext_u8(d12u8, d13u8, 1);
q8u16 = vmlal_u8(q8u16, d3u8, d1u8);
q10u16 = vmlal_u8(q10u16, d6u8, d1u8);
q12u16 = vmlal_u8(q12u16, d9u8, d1u8);
q14u16 = vmlal_u8(q14u16, d12u8, d1u8);
d14u8 = vqrshrn_n_u16(q7u16, 7);
d15u8 = vqrshrn_n_u16(q8u16, 7);
d16u8 = vqrshrn_n_u16(q9u16, 7);
d17u8 = vqrshrn_n_u16(q10u16, 7);
d18u8 = vqrshrn_n_u16(q11u16, 7);
d19u8 = vqrshrn_n_u16(q12u16, 7);
d20u8 = vqrshrn_n_u16(q13u16, 7);
d21u8 = vqrshrn_n_u16(q14u16, 7);
d2u8 = vld1_u8(src_ptr);
d3u8 = vld1_u8(src_ptr + 8);
d4u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d5u8 = vld1_u8(src_ptr);
d6u8 = vld1_u8(src_ptr + 8);
d7u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d8u8 = vld1_u8(src_ptr);
d9u8 = vld1_u8(src_ptr + 8);
d10u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
d11u8 = vld1_u8(src_ptr);
d12u8 = vld1_u8(src_ptr + 8);
d13u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
q10u8 = vcombine_u8(d20u8, d21u8);
vst1q_u8((uint8_t *)tmpp, q7u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q8u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q9u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q10u8);
tmpp += 16;
}
// First-pass filtering for rest 5 lines
d14u8 = vld1_u8(src_ptr);
d15u8 = vld1_u8(src_ptr + 8);
d16u8 = vld1_u8(src_ptr + 16);
src_ptr += src_pixels_per_line;
q9u16 = vmull_u8(d2u8, d0u8);
q10u16 = vmull_u8(d3u8, d0u8);
q11u16 = vmull_u8(d5u8, d0u8);
q12u16 = vmull_u8(d6u8, d0u8);
q13u16 = vmull_u8(d8u8, d0u8);
q14u16 = vmull_u8(d9u8, d0u8);
d2u8 = vext_u8(d2u8, d3u8, 1);
d5u8 = vext_u8(d5u8, d6u8, 1);
d8u8 = vext_u8(d8u8, d9u8, 1);
q9u16 = vmlal_u8(q9u16, d2u8, d1u8);
q11u16 = vmlal_u8(q11u16, d5u8, d1u8);
q13u16 = vmlal_u8(q13u16, d8u8, d1u8);
d3u8 = vext_u8(d3u8, d4u8, 1);
d6u8 = vext_u8(d6u8, d7u8, 1);
d9u8 = vext_u8(d9u8, d10u8, 1);
q10u16 = vmlal_u8(q10u16, d3u8, d1u8);
q12u16 = vmlal_u8(q12u16, d6u8, d1u8);
q14u16 = vmlal_u8(q14u16, d9u8, d1u8);
q1u16 = vmull_u8(d11u8, d0u8);
q2u16 = vmull_u8(d12u8, d0u8);
q3u16 = vmull_u8(d14u8, d0u8);
q4u16 = vmull_u8(d15u8, d0u8);
d11u8 = vext_u8(d11u8, d12u8, 1);
d14u8 = vext_u8(d14u8, d15u8, 1);
q1u16 = vmlal_u8(q1u16, d11u8, d1u8);
q3u16 = vmlal_u8(q3u16, d14u8, d1u8);
d12u8 = vext_u8(d12u8, d13u8, 1);
d15u8 = vext_u8(d15u8, d16u8, 1);
q2u16 = vmlal_u8(q2u16, d12u8, d1u8);
q4u16 = vmlal_u8(q4u16, d15u8, d1u8);
d10u8 = vqrshrn_n_u16(q9u16, 7);
d11u8 = vqrshrn_n_u16(q10u16, 7);
d12u8 = vqrshrn_n_u16(q11u16, 7);
d13u8 = vqrshrn_n_u16(q12u16, 7);
d14u8 = vqrshrn_n_u16(q13u16, 7);
d15u8 = vqrshrn_n_u16(q14u16, 7);
d16u8 = vqrshrn_n_u16(q1u16, 7);
d17u8 = vqrshrn_n_u16(q2u16, 7);
d18u8 = vqrshrn_n_u16(q3u16, 7);
d19u8 = vqrshrn_n_u16(q4u16, 7);
q5u8 = vcombine_u8(d10u8, d11u8);
q6u8 = vcombine_u8(d12u8, d13u8);
q7u8 = vcombine_u8(d14u8, d15u8);
q8u8 = vcombine_u8(d16u8, d17u8);
q9u8 = vcombine_u8(d18u8, d19u8);
vst1q_u8((uint8_t *)tmpp, q5u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q6u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q7u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q8u8);
tmpp += 16;
vst1q_u8((uint8_t *)tmpp, q9u8);
// secondpass_filter
d0u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][0]);
d1u8 = vdup_n_u8(bilinear_taps_coeff[yoffset][1]);
tmpp = tmp;
tmpp2 = tmpp + 272;
q11u8 = vld1q_u8(tmpp);
tmpp += 16;
for (i = 4; i > 0; i--) {
q12u8 = vld1q_u8(tmpp);
tmpp += 16;
q13u8 = vld1q_u8(tmpp);
tmpp += 16;
q14u8 = vld1q_u8(tmpp);
tmpp += 16;
q15u8 = vld1q_u8(tmpp);
tmpp += 16;
q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8);
q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8);
q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8);
q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8);
q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8);
q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8);
q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8);
q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8);
q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8);
q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8);
q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8);
q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8);
q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8);
q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8);
q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8);
q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8);
d2u8 = vqrshrn_n_u16(q1u16, 7);
d3u8 = vqrshrn_n_u16(q2u16, 7);
d4u8 = vqrshrn_n_u16(q3u16, 7);
d5u8 = vqrshrn_n_u16(q4u16, 7);
d6u8 = vqrshrn_n_u16(q5u16, 7);
d7u8 = vqrshrn_n_u16(q6u16, 7);
d8u8 = vqrshrn_n_u16(q7u16, 7);
d9u8 = vqrshrn_n_u16(q8u16, 7);
q1u8 = vcombine_u8(d2u8, d3u8);
q2u8 = vcombine_u8(d4u8, d5u8);
q3u8 = vcombine_u8(d6u8, d7u8);
q4u8 = vcombine_u8(d8u8, d9u8);
q11u8 = q15u8;
vst1q_u8((uint8_t *)tmpp2, q1u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q2u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q3u8);
tmpp2 += 16;
vst1q_u8((uint8_t *)tmpp2, q4u8);
tmpp2 += 16;
}
}
// sub_pixel_variance16x16_neon
q8s32 = vdupq_n_s32(0);
q9s32 = vdupq_n_s32(0);
q10s32 = vdupq_n_s32(0);
tmpp = tmp + 272;
for (i = 0; i < 8; i++) { // sub_pixel_variance16x16_neon_loop
q0u8 = vld1q_u8(tmpp);
tmpp += 16;
q1u8 = vld1q_u8(tmpp);
tmpp += 16;
q2u8 = vld1q_u8(dst_ptr);
dst_ptr += dst_pixels_per_line;
q3u8 = vld1q_u8(dst_ptr);
dst_ptr += dst_pixels_per_line;
d0u8 = vget_low_u8(q0u8);
d1u8 = vget_high_u8(q0u8);
d2u8 = vget_low_u8(q1u8);
d3u8 = vget_high_u8(q1u8);
q11u16 = vsubl_u8(d0u8, vget_low_u8(q2u8));
q12u16 = vsubl_u8(d1u8, vget_high_u8(q2u8));
q13u16 = vsubl_u8(d2u8, vget_low_u8(q3u8));
q14u16 = vsubl_u8(d3u8, vget_high_u8(q3u8));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
}
q10s32 = vaddq_s32(q10s32, q9s32);
q0s64 = vpaddlq_s32(q8s32);
q1s64 = vpaddlq_s32(q10s32);
d0s64 = vget_low_s64(q0s64);
d1s64 = vget_high_s64(q0s64);
d2s64 = vget_low_s64(q1s64);
d3s64 = vget_high_s64(q1s64);
d0s64 = vadd_s64(d0s64, d1s64);
d1s64 = vadd_s64(d2s64, d3s64);
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 8);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
unsigned int vp8_variance16x8_neon(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {
int i;
int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
uint32x2_t d0u32, d10u32;
int64x1_t d0s64, d1s64;
uint8x16_t q0u8, q1u8, q2u8, q3u8;
uint16x8_t q11u16, q12u16, q13u16, q14u16;
int32x4_t q8s32, q9s32, q10s32;
int64x2_t q0s64, q1s64, q5s64;
q8s32 = vdupq_n_s32(0);
q9s32 = vdupq_n_s32(0);
q10s32 = vdupq_n_s32(0);
for (i = 0; i < 4; i++) { // variance16x8_neon_loop
q0u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
q1u8 = vld1q_u8(src_ptr);
src_ptr += source_stride;
__builtin_prefetch(src_ptr);
q2u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
q3u8 = vld1q_u8(ref_ptr);
ref_ptr += recon_stride;
__builtin_prefetch(ref_ptr);
q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
}
q10s32 = vaddq_s32(q10s32, q9s32);
q0s64 = vpaddlq_s32(q8s32);
q1s64 = vpaddlq_s32(q10s32);
d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
vreinterpret_s32_s64(d0s64));
vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
return vget_lane_u32(d0u32, 0);
}
