Optimize Neon implementation of vpx_int_pro_row

Double the number of accumulator registers to remove the bottleneck.
Also peel the first loop iteration.

Change-Id: I6a90680369f9c33cdfe14ea547ac1569ec3f50de

1 files changed, 44 insertions, 45 deletions

diff --git a/vpx_dsp/arm/avg_neon.c b/vpx_dsp/arm/avg_neon.c
index 2fe65d112..22164242c 100644
--- a/vpx_dsp/arm/avg_neon.c
+++ b/vpx_dsp/arm/avg_neon.c
@@ -70,54 +70,53 @@ int vpx_satd_neon(const tran_low_t *coeff, int length) {
 void vpx_int_pro_row_neon(int16_t hbuf[16], uint8_t const *ref,
                           const int ref_stride, const int height) {
   int i;
-  uint16x8_t vec_sum_lo = vdupq_n_u16(0);
-  uint16x8_t vec_sum_hi = vdupq_n_u16(0);
-  const int shift_factor = ((height >> 5) + 3) * -1;
-  const int16x8_t vec_shift = vdupq_n_s16(shift_factor);
-
-  for (i = 0; i < height; i += 8) {
-    const uint8x16_t vec_row1 = vld1q_u8(ref);
-    const uint8x16_t vec_row2 = vld1q_u8(ref + ref_stride);
-    const uint8x16_t vec_row3 = vld1q_u8(ref + ref_stride * 2);
-    const uint8x16_t vec_row4 = vld1q_u8(ref + ref_stride * 3);
-    const uint8x16_t vec_row5 = vld1q_u8(ref + ref_stride * 4);
-    const uint8x16_t vec_row6 = vld1q_u8(ref + ref_stride * 5);
-    const uint8x16_t vec_row7 = vld1q_u8(ref + ref_stride * 6);
-    const uint8x16_t vec_row8 = vld1q_u8(ref + ref_stride * 7);
-
-    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row1));
-    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row1));
-
-    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row2));
-    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row2));
-
-    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row3));
-    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row3));
-
-    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row4));
-    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row4));
-
-    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row5));
-    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row5));
-
-    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row6));
-    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row6));
-
-    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row7));
-    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row7));
-
-    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row8));
-    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row8));
-
-    ref += ref_stride * 8;
+  uint8x16_t r0, r1, r2, r3;
+  uint16x8_t sum_lo[2], sum_hi[2];
+  uint16x8_t tmp_lo[2], tmp_hi[2];
+  int16x8_t avg_lo, avg_hi;
+
+  const int norm_factor = (height >> 5) + 3;
+  const int16x8_t neg_norm_factor = vdupq_n_s16(-norm_factor);
+
+  assert(height >= 4 && height % 4 == 0);
+
+  r0 = vld1q_u8(ref + 0 * ref_stride);
+  r1 = vld1q_u8(ref + 1 * ref_stride);
+  r2 = vld1q_u8(ref + 2 * ref_stride);
+  r3 = vld1q_u8(ref + 3 * ref_stride);
+
+  sum_lo[0] = vaddl_u8(vget_low_u8(r0), vget_low_u8(r1));
+  sum_hi[0] = vaddl_u8(vget_high_u8(r0), vget_high_u8(r1));
+  sum_lo[1] = vaddl_u8(vget_low_u8(r2), vget_low_u8(r3));
+  sum_hi[1] = vaddl_u8(vget_high_u8(r2), vget_high_u8(r3));
+
+  for (i = 4; i < height; i += 4) {
+    r0 = vld1q_u8(ref + 0 * ref_stride);
+    r1 = vld1q_u8(ref + 1 * ref_stride);
+    r2 = vld1q_u8(ref + 2 * ref_stride);
+    r3 = vld1q_u8(ref + 3 * ref_stride);
+
+    tmp_lo[0] = vaddl_u8(vget_low_u8(r0), vget_low_u8(r1));
+    tmp_hi[0] = vaddl_u8(vget_high_u8(r0), vget_high_u8(r1));
+    tmp_lo[1] = vaddl_u8(vget_low_u8(r2), vget_low_u8(r3));
+    tmp_hi[1] = vaddl_u8(vget_high_u8(r2), vget_high_u8(r3));
+
+    sum_lo[0] = vaddq_u16(sum_lo[0], tmp_lo[0]);
+    sum_hi[0] = vaddq_u16(sum_hi[0], tmp_hi[0]);
+    sum_lo[1] = vaddq_u16(sum_lo[1], tmp_lo[1]);
+    sum_hi[1] = vaddq_u16(sum_hi[1], tmp_hi[1]);
+
+    ref += 4 * ref_stride;
   }
 
-  vec_sum_lo = vshlq_u16(vec_sum_lo, vec_shift);
-  vec_sum_hi = vshlq_u16(vec_sum_hi, vec_shift);
+  sum_lo[0] = vaddq_u16(sum_lo[0], sum_lo[1]);
+  sum_hi[0] = vaddq_u16(sum_hi[0], sum_hi[1]);
+
+  avg_lo = vshlq_s16(vreinterpretq_s16_u16(sum_lo[0]), neg_norm_factor);
+  avg_hi = vshlq_s16(vreinterpretq_s16_u16(sum_hi[0]), neg_norm_factor);
 
-  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_lo));
-  hbuf += 8;
-  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_hi));
+  vst1q_s16(hbuf, avg_lo);
+  vst1q_s16(hbuf + 8, avg_hi);
 }
 
 int16_t vpx_int_pro_col_neon(uint8_t const *ref, const int width) {