static void interpolate5LineNeon(uint16 *dst, const uint16 *srcA, const uint16 *srcB, int width, int k1, int k2) { uint16x4_t kRedBlueMask_4 = vdup_n_u16(ColorMask::kRedBlueMask); uint16x4_t kGreenMask_4 = vdup_n_u16(ColorMask::kGreenMask); uint16x4_t k1_4 = vdup_n_u16(k1); uint16x4_t k2_4 = vdup_n_u16(k2); while (width >= 4) { uint16x4_t srcA_4 = vld1_u16(srcA); uint16x4_t srcB_4 = vld1_u16(srcB); uint16x4_t p1_4 = srcB_4; uint16x4_t p2_4 = srcA_4; uint16x4_t p1_rb_4 = vand_u16(p1_4, kRedBlueMask_4); uint16x4_t p1_g_4 = vand_u16(p1_4, kGreenMask_4); uint16x4_t p2_rb_4 = vand_u16(p2_4, kRedBlueMask_4); uint16x4_t p2_g_4 = vand_u16(p2_4, kGreenMask_4); uint32x4_t tmp_rb_4 = vshrq_n_u32(vmlal_u16(vmull_u16(p2_rb_4, k2_4), p1_rb_4, k1_4), 3); uint32x4_t tmp_g_4 = vshrq_n_u32(vmlal_u16(vmull_u16(p2_g_4, k2_4), p1_g_4, k1_4), 3); uint16x4_t p_rb_4 = vmovn_u32(tmp_rb_4); p_rb_4 = vand_u16(p_rb_4, kRedBlueMask_4); uint16x4_t p_g_4 = vmovn_u32(tmp_g_4); p_g_4 = vand_u16(p_g_4, kGreenMask_4); uint16x4_t result_4 = p_rb_4 | p_g_4; vst1_u16(dst, result_4); dst += 4; srcA += 4; srcB += 4; width -= 4; } }
void WebRtcAecm_CalcLinearEnergiesNeon(AecmCore* aecm, const uint16_t* far_spectrum, int32_t* echo_est, uint32_t* far_energy, uint32_t* echo_energy_adapt, uint32_t* echo_energy_stored) { int16_t* start_stored_p = aecm->channelStored; int16_t* start_adapt_p = aecm->channelAdapt16; int32_t* echo_est_p = echo_est; const int16_t* end_stored_p = aecm->channelStored + PART_LEN; const uint16_t* far_spectrum_p = far_spectrum; int16x8_t store_v, adapt_v; uint16x8_t spectrum_v; uint32x4_t echo_est_v_low, echo_est_v_high; uint32x4_t far_energy_v, echo_stored_v, echo_adapt_v; far_energy_v = vdupq_n_u32(0); echo_adapt_v = vdupq_n_u32(0); echo_stored_v = vdupq_n_u32(0); // Get energy for the delayed far end signal and estimated // echo using both stored and adapted channels. // The C code: // for (i = 0; i < PART_LEN1; i++) { // echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], // far_spectrum[i]); // (*far_energy) += (uint32_t)(far_spectrum[i]); // *echo_energy_adapt += aecm->channelAdapt16[i] * far_spectrum[i]; // (*echo_energy_stored) += (uint32_t)echo_est[i]; // } while (start_stored_p < end_stored_p) { spectrum_v = vld1q_u16(far_spectrum_p); adapt_v = vld1q_s16(start_adapt_p); store_v = vld1q_s16(start_stored_p); far_energy_v = vaddw_u16(far_energy_v, vget_low_u16(spectrum_v)); far_energy_v = vaddw_u16(far_energy_v, vget_high_u16(spectrum_v)); echo_est_v_low = vmull_u16(vreinterpret_u16_s16(vget_low_s16(store_v)), vget_low_u16(spectrum_v)); echo_est_v_high = vmull_u16(vreinterpret_u16_s16(vget_high_s16(store_v)), vget_high_u16(spectrum_v)); vst1q_s32(echo_est_p, vreinterpretq_s32_u32(echo_est_v_low)); vst1q_s32(echo_est_p + 4, vreinterpretq_s32_u32(echo_est_v_high)); echo_stored_v = vaddq_u32(echo_est_v_low, echo_stored_v); echo_stored_v = vaddq_u32(echo_est_v_high, echo_stored_v); echo_adapt_v = vmlal_u16(echo_adapt_v, vreinterpret_u16_s16(vget_low_s16(adapt_v)), vget_low_u16(spectrum_v)); echo_adapt_v = vmlal_u16(echo_adapt_v, vreinterpret_u16_s16(vget_high_s16(adapt_v)), vget_high_u16(spectrum_v)); start_stored_p += 8; start_adapt_p += 8; far_spectrum_p += 8; echo_est_p += 8; } AddLanes(far_energy, far_energy_v); AddLanes(echo_energy_stored, echo_stored_v); AddLanes(echo_energy_adapt, echo_adapt_v); echo_est[PART_LEN] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[PART_LEN], far_spectrum[PART_LEN]); *echo_energy_stored += (uint32_t)echo_est[PART_LEN]; *far_energy += (uint32_t)far_spectrum[PART_LEN]; *echo_energy_adapt += aecm->channelAdapt16[PART_LEN] * far_spectrum[PART_LEN]; }