vshrn_n_s32(dqcoeff_1, 1)));
#endif
- eob_max = vandq_u16(vtstq_s16(qcoeff, neg_one), v_iscan);
+ store_s16q_to_tran_low(qcoeff_ptr, v_qcoeff);
- store_s16q_to_tran_low(qcoeff_ptr, qcoeff);
+ *v_eobmax = get_max_lane_eob(iscan_ptr, *v_eobmax, v_nz_mask);
+}
- iscan += 8;
- coeff_ptr += 8;
- qcoeff_ptr += 8;
- dqcoeff_ptr += 8;
+void vp9_quantize_fp_32x32_neon(const tran_low_t *coeff_ptr, intptr_t count,
+ const int16_t *round_ptr,
+ const int16_t *quant_ptr,
+ tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+ const int16_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan) {
+ int16x8_t eob_max = vdupq_n_s16(-1);
+ // ROUND_POWER_OF_TWO(round_ptr[], 1)
+ int16x8_t round = vrshrq_n_s16(vld1q_s16(round_ptr), 1);
+ int16x8_t quant = vld1q_s16(quant_ptr);
+ int16x8_t dequant = vld1q_s16(dequant_ptr);
+ // dequant >> 2 is used similar to zbin as a threshold.
+ int16x8_t dequant_thresh = vshrq_n_s16(vld1q_s16(dequant_ptr), 2);
+ int i;
- {
- int i;
- const int16x8_t round = vrshrq_n_s16(vmovq_n_s16(round_ptr[1]), 1);
- const int16x8_t quant = vmovq_n_s16(quant_ptr[1]);
- const int16x8_t dequant_thresh =
- vshrq_n_s16(vmovq_n_s16(dequant_ptr[1]), 2);
-
- // Process the rest of the ac coeffs.
- for (i = 8; i < 32 * 32; i += 8) {
- const uint16x8_t v_iscan =
- vreinterpretq_u16_s16(vaddq_s16(vld1q_s16(iscan), one));
- const int16x8_t coeff = load_tran_low_to_s16q(coeff_ptr);
- const int16x8_t coeff_sign = vshrq_n_s16(coeff, 15);
- const int16x8_t coeff_abs = vabsq_s16(coeff);
- const int16x8_t dequant_mask =
- vreinterpretq_s16_u16(vcgeq_s16(coeff_abs, dequant_thresh));
-
- int16x8_t qcoeff = vqaddq_s16(coeff_abs, round);
- int32x4_t dqcoeff_0, dqcoeff_1;
-
- qcoeff = vqdmulhq_s16(qcoeff, quant);
- qcoeff = veorq_s16(qcoeff, coeff_sign);
- qcoeff = vsubq_s16(qcoeff, coeff_sign);
- qcoeff = vandq_s16(qcoeff, dequant_mask);
-
- dqcoeff_0 = vmull_n_s16(vget_low_s16(qcoeff), dequant_ptr[1]);
- dqcoeff_1 = vmull_n_s16(vget_high_s16(qcoeff), dequant_ptr[1]);
-
- dqcoeff_0 = vaddq_s32(dqcoeff_0, extract_sign_bit(dqcoeff_0));
- dqcoeff_1 = vaddq_s32(dqcoeff_1, extract_sign_bit(dqcoeff_1));
+ (void)scan;
+ (void)count;
-#if CONFIG_VP9_HIGHBITDEPTH
- vst1q_s32(dqcoeff_ptr, vshrq_n_s32(dqcoeff_0, 1));
- vst1q_s32(dqcoeff_ptr + 4, vshrq_n_s32(dqcoeff_1, 1));
-#else
- store_s16q_to_tran_low(
- dqcoeff_ptr,
- vcombine_s16(vshrn_n_s32(dqcoeff_0, 1), vshrn_n_s32(dqcoeff_1, 1)));
-#endif
+ // Process dc and the first seven ac coeffs.
+ quantize_fp_32x32_8(&round, &quant, &dequant, &dequant_thresh, coeff_ptr,
+ iscan, qcoeff_ptr, dqcoeff_ptr, &eob_max);
- eob_max =
- vmaxq_u16(eob_max, vandq_u16(vtstq_s16(qcoeff, neg_one), v_iscan));
+ update_fp_values(&round, &quant, &dequant);
+ dequant_thresh = vdupq_lane_s16(vget_low_s16(dequant_thresh), 1);
- store_s16q_to_tran_low(qcoeff_ptr, qcoeff);
+ iscan += 8;
+ coeff_ptr += 8;
+ qcoeff_ptr += 8;
+ dqcoeff_ptr += 8;
- iscan += 8;
- coeff_ptr += 8;
- qcoeff_ptr += 8;
- dqcoeff_ptr += 8;
- }
+ // Process the rest of the ac coeffs.
+ for (i = 8; i < 32 * 32; i += 8) {
+ quantize_fp_32x32_8(&round, &quant, &dequant, &dequant_thresh, coeff_ptr,
+ iscan, qcoeff_ptr, dqcoeff_ptr, &eob_max);
-#ifdef __aarch64__
- *eob_ptr = vmaxvq_u16(eob_max);
-#else
- {
- const uint16x4_t eob_max_0 =
- vmax_u16(vget_low_u16(eob_max), vget_high_u16(eob_max));
- const uint16x4_t eob_max_1 = vpmax_u16(eob_max_0, eob_max_0);
- const uint16x4_t eob_max_2 = vpmax_u16(eob_max_1, eob_max_1);
- vst1_lane_u16(eob_ptr, eob_max_2, 0);
- }
-#endif // __aarch64__
+ iscan += 8;
+ coeff_ptr += 8;
+ qcoeff_ptr += 8;
+ dqcoeff_ptr += 8;
}
+
+ *eob_ptr = get_max_eob(eob_max);
}
+
+ #if CONFIG_VP9_HIGHBITDEPTH
+ static VPX_FORCE_INLINE uint16x4_t
+ highbd_quantize_fp_4(const tran_low_t *coeff_ptr, tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr, int32x4_t v_quant_s32,
+ int32x4_t v_dequant_s32, int32x4_t v_round_s32) {
+ const int32x4_t v_coeff = vld1q_s32(coeff_ptr);
+ const int32x4_t v_coeff_sign =
+ vreinterpretq_s32_u32(vcltq_s32(v_coeff, vdupq_n_s32(0)));
+ const int32x4_t v_abs_coeff = vabsq_s32(v_coeff);
+ const int32x4_t v_tmp = vaddq_s32(v_abs_coeff, v_round_s32);
+ // const int abs_qcoeff = (int)((tmp * quant) >> 16);
+ const int32x4_t v_abs_qcoeff = vqdmulhq_s32(v_tmp, v_quant_s32);
+ // qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+ const int32x4_t v_qcoeff =
+ vsubq_s32(veorq_s32(v_abs_qcoeff, v_coeff_sign), v_coeff_sign);
+ const int32x4_t v_abs_dqcoeff = vmulq_s32(v_abs_qcoeff, v_dequant_s32);
+ // dqcoeff_ptr[rc] = (tran_low_t)((abs_dqcoeff ^ coeff_sign) - coeff_sign);
+ const int32x4_t v_dqcoeff =
+ vsubq_s32(veorq_s32(v_abs_dqcoeff, v_coeff_sign), v_coeff_sign);
+
+ vst1q_s32(qcoeff_ptr, v_qcoeff);
+ vst1q_s32(dqcoeff_ptr, v_dqcoeff);
+
+ // Packed nz_qcoeff_mask. Used to find eob.
+ return vmovn_u32(vceqq_s32(v_abs_qcoeff, vdupq_n_s32(0)));
+ }
+
+ void vp9_highbd_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t count,
+ const int16_t *round_ptr,
+ const int16_t *quant_ptr,
+ tran_low_t *qcoeff_ptr,
+ tran_low_t *dqcoeff_ptr,
+ const int16_t *dequant_ptr, uint16_t *eob_ptr,
+ const int16_t *scan, const int16_t *iscan) {
+ const int16x4_t v_zero = vdup_n_s16(0);
+ const int16x4_t v_quant = vld1_s16(quant_ptr);
+ const int16x4_t v_dequant = vld1_s16(dequant_ptr);
+ const int16x4_t v_round = vld1_s16(round_ptr);
+ int32x4_t v_round_s32 = vaddl_s16(v_round, v_zero);
+ int32x4_t v_quant_s32 = vshlq_n_s32(vaddl_s16(v_quant, v_zero), 15);
+ int32x4_t v_dequant_s32 = vaddl_s16(v_dequant, v_zero);
+ uint16x4_t v_mask_lo, v_mask_hi;
+ int16x8_t v_eobmax = vdupq_n_s16(-1);
+
+ (void)scan;
+
+ // DC and first 3 AC
+ v_mask_lo = highbd_quantize_fp_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr,
+ v_quant_s32, v_dequant_s32, v_round_s32);
+
+ // overwrite the DC constants with AC constants
+ v_round_s32 = vdupq_lane_s32(vget_low_s32(v_round_s32), 1);
+ v_quant_s32 = vdupq_lane_s32(vget_low_s32(v_quant_s32), 1);
+ v_dequant_s32 = vdupq_lane_s32(vget_low_s32(v_dequant_s32), 1);
+
+ // 4 more AC
+ v_mask_hi =
+ highbd_quantize_fp_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
+ v_quant_s32, v_dequant_s32, v_round_s32);
+
+ // Find the max lane eob for the first 8 coeffs.
+ v_eobmax =
+ get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
+
+ count -= 8;
+ do {
+ coeff_ptr += 8;
+ qcoeff_ptr += 8;
+ dqcoeff_ptr += 8;
+ iscan += 8;
+ v_mask_lo = highbd_quantize_fp_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr,
+ v_quant_s32, v_dequant_s32, v_round_s32);
+ v_mask_hi =
+ highbd_quantize_fp_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
+ v_quant_s32, v_dequant_s32, v_round_s32);
+ // Find the max lane eob for 8 coeffs.
+ v_eobmax =
+ get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
+ count -= 8;
+ } while (count);
+
+ *eob_ptr = get_max_eob(v_eobmax);
+ }
+ #endif // CONFIG_VP9_HIGHBITDEPTH