Improved vp9_quantize_fp_neon()

Eliminated instructions by using better neon instructions
and rearranging the loop.

On a Nexus 7, vpxenc (in realtime mode, speed -12)
reported a performance improvement of ~1.0%.

Change-Id: I6b1700e79318f647ea67ef25e954c308932950ec

diff --git a/vp9/encoder/arm/neon/vp9_quantize_neon.c b/vp9/encoder/arm/neon/vp9_quantize_neon.c
index 2d5ec79b369f898691ab5ac2f85496a0c541da1f..8c13d0da672c69725d5f1366a793158cdc915bae 100644 (file)
--- a/vp9/encoder/arm/neon/vp9_quantize_neon.c
+++ b/vp9/encoder/arm/neon/vp9_quantize_neon.c
@@ -28,7 +28,6 @@ void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
                           int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
                           int zbin_oq_value, uint16_t *eob_ptr,
                           const int16_t *scan, const int16_t *iscan) {
-  int i;
   // TODO(jingning) Decide the need of these arguments after the
   // quantization process is completed.
   (void)zbin_ptr;
@@ -39,7 +38,7 @@ void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
   if (!skip_block) {
     // Quantization pass: All coefficients with index >= zero_flag are
     // skippable. Note: zero_flag can be zero.
-
+    int i;
     const int16x8_t v_zero = vdupq_n_s16(0);
     const int16x8_t v_one = vdupq_n_s16(1);
     int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
@@ -50,13 +49,37 @@ void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
     v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
     v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
     v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
-
-    for (i = 0; i < count; i += 8) {
+    // process dc and the first seven ac coeffs
+    {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[0]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
+                                           vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
+                                           vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
+                                            vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[0], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[0], v_dqcoeff);
+      v_round = vmovq_n_s16(round_ptr[1]);
+      v_quant = vmovq_n_s16(quant_ptr[1]);
+      v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    }
+    // now process the rest of the ac coeffs
+    for (i = 8; i < count; i += 8) {
       const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
       const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
       const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
-      const int16x8_t v_abs_coeff = vabsq_s16(v_coeff);
-      const int16x8_t v_tmp = vqaddq_s16(v_abs_coeff, v_round);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
       const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
                                            vget_low_s16(v_quant));
       const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
@@ -65,19 +88,13 @@ void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
                                             vshrn_n_s32(v_tmp_hi, 16));
       const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
       const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
-      const int16x8_t v_nz_iscan =
-          vandq_s16(vmvnq_s16(vreinterpretq_s16_u16(v_nz_mask)), v_iscan_plus1);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
       const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
       const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
       const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
-
       v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
-
       vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
       vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
-      v_round = vmovq_n_s16(round_ptr[1]);
-      v_quant = vmovq_n_s16(quant_ptr[1]);
-      v_dequant = vmovq_n_s16(dequant_ptr[1]);
     }
     {
       const int16x4_t v_eobmax_3210 =