static inline void DotProductWithScaleNeon(int32_t* cross_correlation, const int16_t* vector1, const int16_t* vector2, size_t length, int scaling) { size_t i = 0; size_t len1 = length >> 3; size_t len2 = length & 7; int64x2_t sum0 = vdupq_n_s64(0); int64x2_t sum1 = vdupq_n_s64(0); for (i = len1; i > 0; i -= 1) { int16x8_t seq1_16x8 = vld1q_s16(vector1); int16x8_t seq2_16x8 = vld1q_s16(vector2); #if defined(WEBRTC_ARCH_ARM64) int32x4_t tmp0 = vmull_s16(vget_low_s16(seq1_16x8), vget_low_s16(seq2_16x8)); int32x4_t tmp1 = vmull_high_s16(seq1_16x8, seq2_16x8); #else int32x4_t tmp0 = vmull_s16(vget_low_s16(seq1_16x8), vget_low_s16(seq2_16x8)); int32x4_t tmp1 = vmull_s16(vget_high_s16(seq1_16x8), vget_high_s16(seq2_16x8)); #endif sum0 = vpadalq_s32(sum0, tmp0); sum1 = vpadalq_s32(sum1, tmp1); vector1 += 8; vector2 += 8; } // Calculate the rest of the samples. int64_t sum_res = 0; for (i = len2; i > 0; i -= 1) { sum_res += WEBRTC_SPL_MUL_16_16(*vector1, *vector2); vector1++; vector2++; } sum0 = vaddq_s64(sum0, sum1); #if defined(WEBRTC_ARCH_ARM64) int64_t sum2 = vaddvq_s64(sum0); *cross_correlation = (int32_t)((sum2 + sum_res) >> scaling); #else int64x1_t shift = vdup_n_s64(-scaling); int64x1_t sum2 = vadd_s64(vget_low_s64(sum0), vget_high_s64(sum0)); sum2 = vadd_s64(sum2, vdup_n_s64(sum_res)); sum2 = vshl_s64(sum2, shift); vst1_lane_s32(cross_correlation, vreinterpret_s32_s64(sum2), 0); #endif }
int64x1_t test_vdup_n_s64(int64_t v1) { // CHECK: test_vdup_n_s64 return vdup_n_s64(v1); // CHECK: fmov {{d[0-9]+}}, {{x[0-9]+}} }
void vpx_fdct4x4_neon(const int16_t *input, tran_low_t *final_output, int stride) { int i; // input[M * stride] * 16 int16x4_t input_0 = vshl_n_s16(vld1_s16(input + 0 * stride), 4); int16x4_t input_1 = vshl_n_s16(vld1_s16(input + 1 * stride), 4); int16x4_t input_2 = vshl_n_s16(vld1_s16(input + 2 * stride), 4); int16x4_t input_3 = vshl_n_s16(vld1_s16(input + 3 * stride), 4); // If the very first value != 0, then add 1. if (input[0] != 0) { const int16x4_t one = vreinterpret_s16_s64(vdup_n_s64(1)); input_0 = vadd_s16(input_0, one); } for (i = 0; i < 2; ++i) { const int16x8_t input_01 = vcombine_s16(input_0, input_1); const int16x8_t input_32 = vcombine_s16(input_3, input_2); // in_0 +/- in_3, in_1 +/- in_2 const int16x8_t s_01 = vaddq_s16(input_01, input_32); const int16x8_t s_32 = vsubq_s16(input_01, input_32); // step_0 +/- step_1, step_2 +/- step_3 const int16x4_t s_0 = vget_low_s16(s_01); const int16x4_t s_1 = vget_high_s16(s_01); const int16x4_t s_2 = vget_high_s16(s_32); const int16x4_t s_3 = vget_low_s16(s_32); // (s_0 +/- s_1) * cospi_16_64 // Must expand all elements to s32. See 'needs32' comment in fwd_txfm.c. const int32x4_t s_0_p_s_1 = vaddl_s16(s_0, s_1); const int32x4_t s_0_m_s_1 = vsubl_s16(s_0, s_1); const int32x4_t temp1 = vmulq_n_s32(s_0_p_s_1, cospi_16_64); const int32x4_t temp2 = vmulq_n_s32(s_0_m_s_1, cospi_16_64); // fdct_round_shift int16x4_t out_0 = vrshrn_n_s32(temp1, DCT_CONST_BITS); int16x4_t out_2 = vrshrn_n_s32(temp2, DCT_CONST_BITS); // s_3 * cospi_8_64 + s_2 * cospi_24_64 // s_3 * cospi_24_64 - s_2 * cospi_8_64 const int32x4_t s_3_cospi_8_64 = vmull_n_s16(s_3, cospi_8_64); const int32x4_t s_3_cospi_24_64 = vmull_n_s16(s_3, cospi_24_64); const int32x4_t temp3 = vmlal_n_s16(s_3_cospi_8_64, s_2, cospi_24_64); const int32x4_t temp4 = vmlsl_n_s16(s_3_cospi_24_64, s_2, cospi_8_64); // fdct_round_shift int16x4_t out_1 = vrshrn_n_s32(temp3, DCT_CONST_BITS); int16x4_t out_3 = vrshrn_n_s32(temp4, DCT_CONST_BITS); transpose_s16_4x4d(&out_0, &out_1, &out_2, &out_3); input_0 = out_0; input_1 = out_1; input_2 = out_2; input_3 = out_3; } { // Not quite a rounding shift. Only add 1 despite shifting by 2. const int16x8_t one = vdupq_n_s16(1); int16x8_t out_01 = vcombine_s16(input_0, input_1); int16x8_t out_23 = vcombine_s16(input_2, input_3); out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2); out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2); store_s16q_to_tran_low(final_output + 0 * 8, out_01); store_s16q_to_tran_low(final_output + 1 * 8, out_23); } }
inline int64x1_t vdup_n(const s64 & val) { return vdup_n_s64(val); }
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); }