shithub: libvpx

Download patch

ref: 4dca92345467fea3c4461152f1847352271ec883
parent: e7c453b613ffc2baab4b04cc514fd4a0a15c8927
author: Johann <johannkoenig@google.com>
date: Wed Dec 21 09:19:25 EST 2016 

postproc: vpx_mbpost_proc_across_ip_neon

The speedup is pretty poor. I would be concerned except the SSE2 is
worse:
Existing SSE2 improvement: 22%
New neon improvement: 35%

BUG=webm:1320

Change-Id: Ied598a261134aa6cbe69f96f58589d2bae17bf62


--- a/test/pp_filter_test.cc
+++ b/test/pp_filter_test.cc
@@ -595,6 +595,9 @@
 INSTANTIATE_TEST_CASE_P(
     NEON, VpxPostProcDownAndAcrossMbRowTest,
     ::testing::Values(vpx_post_proc_down_and_across_mb_row_neon));
+
+INSTANTIATE_TEST_CASE_P(NEON, VpxMbPostProcAcrossIpTest,
+                        ::testing::Values(vpx_mbpost_proc_across_ip_neon));
 #endif  // HAVE_NEON
 
 #if HAVE_MSA
--- a/vpx_dsp/arm/deblock_neon.c
+++ b/vpx_dsp/arm/deblock_neon.c
@@ -257,3 +257,130 @@
     dst_ptr += 8 * dst_stride;
   }
 }
+
+// sum += x;
+// sumsq += x * y;
+static void accumulate_sum_sumsq(const int16x4_t x, const int32x4_t xy,
+                                 int16x4_t *const sum, int32x4_t *const sumsq) {
+  const int16x4_t zero = vdup_n_s16(0);
+  const int32x4_t zeroq = vdupq_n_s32(0);
+
+  // Add in the first set because vext doesn't work with '0'.
+  *sum = vadd_s16(*sum, x);
+  *sumsq = vaddq_s32(*sumsq, xy);
+
+  // Shift x and xy to the right and sum. vext requires an immediate.
+  *sum = vadd_s16(*sum, vext_s16(zero, x, 1));
+  *sumsq = vaddq_s32(*sumsq, vextq_s32(zeroq, xy, 1));
+
+  *sum = vadd_s16(*sum, vext_s16(zero, x, 2));
+  *sumsq = vaddq_s32(*sumsq, vextq_s32(zeroq, xy, 2));
+
+  *sum = vadd_s16(*sum, vext_s16(zero, x, 3));
+  *sumsq = vaddq_s32(*sumsq, vextq_s32(zeroq, xy, 3));
+}
+
+// Generate mask based on (sumsq * 15 - sum * sum < flimit)
+static uint16x4_t calculate_mask(const int16x4_t sum, const int32x4_t sumsq,
+                                 const int32x4_t f, const int32x4_t fifteen) {
+  const int32x4_t a = vmulq_s32(sumsq, fifteen);
+  const int32x4_t b = vmlsl_s16(a, sum, sum);
+  const uint32x4_t mask32 = vcltq_s32(b, f);
+  return vmovn_u32(mask32);
+}
+
+static uint8x8_t combine_mask(const int16x4_t sum_low, const int16x4_t sum_high,
+                              const int32x4_t sumsq_low,
+                              const int32x4_t sumsq_high, const int32x4_t f) {
+  const int32x4_t fifteen = vdupq_n_s32(15);
+  const uint16x4_t mask16_low = calculate_mask(sum_low, sumsq_low, f, fifteen);
+  const uint16x4_t mask16_high =
+      calculate_mask(sum_high, sumsq_high, f, fifteen);
+  return vmovn_u16(vcombine_u16(mask16_low, mask16_high));
+}
+
+// Apply filter of (8 + sum + s[c]) >> 4.
+static uint8x8_t filter_pixels(const int16x8_t sum, const uint8x8_t s) {
+  const int16x8_t s16 = vreinterpretq_s16_u16(vmovl_u8(s));
+  const int16x8_t sum_s = vaddq_s16(sum, s16);
+
+  return vqrshrun_n_s16(sum_s, 4);
+}
+
+void vpx_mbpost_proc_across_ip_neon(uint8_t *src, int pitch, int rows, int cols,
+                                    int flimit) {
+  int row, col;
+  const int32x4_t f = vdupq_n_s32(flimit);
+
+  assert(cols % 8 == 0);
+
+  for (row = 0; row < rows; ++row) {
+    // Sum the first 8 elements, which are extended from s[0].
+    // sumsq gets primed with +16.
+    int sumsq = src[0] * src[0] * 9 + 16;
+    int sum = src[0] * 9;
+
+    uint8x8_t left_context, s, right_context;
+    int16x4_t sum_low, sum_high;
+    int32x4_t sumsq_low, sumsq_high;
+
+    // Sum (+square) the next 6 elements.
+    // Skip [0] because it's included above.
+    for (col = 1; col <= 6; ++col) {
+      sumsq += src[col] * src[col];
+      sum += src[col];
+    }
+
+    // Prime the sums. Later the loop uses the _high values to prime the new
+    // vectors.
+    sumsq_high = vdupq_n_s32(sumsq);
+    sum_high = vdup_n_s16(sum);
+
+    // Manually extend the left border.
+    left_context = vdup_n_u8(src[0]);
+
+    for (col = 0; col < cols; col += 8) {
+      uint8x8_t mask, output;
+      int16x8_t x, y;
+      int32x4_t xy_low, xy_high;
+
+      s = vld1_u8(src + col);
+
+      if (col + 8 == cols) {
+        // Last row. Extend border.
+        right_context = vdup_n_u8(src[col + 7]);
+      } else {
+        right_context = vld1_u8(src + col + 7);
+      }
+
+      x = vreinterpretq_s16_u16(vsubl_u8(right_context, left_context));
+      y = vreinterpretq_s16_u16(vaddl_u8(right_context, left_context));
+      xy_low = vmull_s16(vget_low_s16(x), vget_low_s16(y));
+      xy_high = vmull_s16(vget_high_s16(x), vget_high_s16(y));
+
+      // Catch up to the last sum'd value.
+      sum_low = vdup_lane_s16(sum_high, 3);
+      sumsq_low = vdupq_lane_s32(vget_high_s32(sumsq_high), 1);
+
+      accumulate_sum_sumsq(vget_low_s16(x), xy_low, &sum_low, &sumsq_low);
+
+      // Need to do this sequentially because we need the max value from
+      // sum_low.
+      sum_high = vdup_lane_s16(sum_low, 3);
+      sumsq_high = vdupq_lane_s32(vget_high_s32(sumsq_low), 1);
+
+      accumulate_sum_sumsq(vget_high_s16(x), xy_high, &sum_high, &sumsq_high);
+
+      mask = combine_mask(sum_low, sum_high, sumsq_low, sumsq_high, f);
+
+      output = filter_pixels(vcombine_s16(sum_low, sum_high), s);
+      output = vbsl_u8(mask, output, s);
+
+      vst1_u8(src + col, output);
+
+      left_context = s;
+    }
+
+    src += pitch;
+  }
+}
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1753,7 +1753,7 @@
     specialize qw/vpx_mbpost_proc_down sse2 msa/;
 
     add_proto qw/void vpx_mbpost_proc_across_ip/, "unsigned char *dst, int pitch, int rows, int cols,int flimit";
-    specialize qw/vpx_mbpost_proc_across_ip sse2 msa/;
+    specialize qw/vpx_mbpost_proc_across_ip sse2 neon msa/;
 
     add_proto qw/void vpx_post_proc_down_and_across_mb_row/, "unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size";
     specialize qw/vpx_post_proc_down_and_across_mb_row sse2 neon msa/;