VSX version of vpx_mbpost_proc_down

Low bit depth version only. Passes the VpxMbPostProcDownTest.

VpxMbPostProcDownTest Speed Test (POWER8 Model 2.1)
Full calculations:
C time = 195.4 ms, VSX time = 33.7 ms (5.8x)

Change-Id: If1aca7c135de036a1ab7923c0d1e6733bfe27ef7

1 files changed, 128 insertions, 0 deletions

diff --git a/vpx_dsp/ppc/deblock_vsx.c b/vpx_dsp/ppc/deblock_vsx.c
new file mode 100644
index 000000000..4a031832c
--- /dev/null
+++ b/vpx_dsp/ppc/deblock_vsx.c
@@ -0,0 +1,128 @@
+/*
+ *  Copyright (c) 2018 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/ppc/types_vsx.h"
+
+extern const int16_t vpx_rv[];
+
+static const uint8x16_t load_merge = { 0x00, 0x02, 0x04, 0x06, 0x08, 0x0A,
+                                       0x0C, 0x0E, 0x18, 0x19, 0x1A, 0x1B,
+                                       0x1C, 0x1D, 0x1E, 0x1F };
+
+static const uint8x16_t mask_merge = { 0x00, 0x01, 0x10, 0x11, 0x04, 0x05,
+                                       0x14, 0x15, 0x08, 0x09, 0x18, 0x19,
+                                       0x0C, 0x0D, 0x1C, 0x1D };
+
+void vpx_mbpost_proc_down_vsx(uint8_t *dst, int pitch, int rows, int cols,
+                              int flimit) {
+  int col, row, i;
+  int16x8_t window[16];
+  const int32x4_t lim = vec_splats(flimit);
+
+  // 8 columns are processed at a time.
+  assert(cols % 8 == 0);
+  // If rows is less than 8 the bottom border extension fails.
+  assert(rows >= 8);
+
+  for (col = 0; col < cols; col += 8) {
+    // The sum is signed and requires at most 13 bits.
+    // (8 bits + sign) * 15 (4 bits)
+    int16x8_t r1, sum;
+    // The sum of squares requires at most 20 bits.
+    // (16 bits + sign) * 15 (4 bits)
+    int32x4_t sumsq_even, sumsq_odd;
+
+    r1 = unpack_to_s16_h(vec_vsx_ld(0, dst));
+    // Fill sliding window with first row.
+    for (i = 0; i <= 8; i++) {
+      window[i] = r1;
+    }
+    // First 9 rows of the sliding window are the same.
+    // sum = r1 * 9
+    sum = vec_mladd(r1, vec_splats((int16_t)9), vec_zeros_s16);
+
+    // sumsq = r1 * r1 * 9
+    sumsq_even = vec_mule(sum, r1);
+    sumsq_odd = vec_mulo(sum, r1);
+
+    // Fill the next 6 rows of the sliding window with rows 2 to 7.
+    for (i = 1; i <= 6; ++i) {
+      const int16x8_t next_row = unpack_to_s16_h(vec_vsx_ld(i * pitch, dst));
+      window[i + 8] = next_row;
+      sum = vec_add(sum, next_row);
+      sumsq_odd = vec_add(sumsq_odd, vec_mulo(next_row, next_row));
+      sumsq_even = vec_add(sumsq_even, vec_mule(next_row, next_row));
+    }
+
+    for (row = 0; row < rows; row++) {
+      int32x4_t d15_even, d15_odd, d0_even, d0_odd;
+      int32x4_t sumsq_odd_scaled, sumsq_even_scaled;
+      int32x4_t thres_odd, thres_even;
+      bool32x4_t mask_odd, mask_even;
+      bool16x8_t mask;
+      int16x8_t filtered, masked;
+      uint8x16_t out;
+
+      const int16x8_t rv = vec_vsx_ld(0, vpx_rv + (row & 127));
+
+      // Move the sliding window
+      if (row + 7 < rows) {
+        window[15] = unpack_to_s16_h(vec_vsx_ld((row + 7) * pitch, dst));
+      } else {
+        window[15] = window[14];
+      }
+
+      // C: sum += s[7 * pitch] - s[-8 * pitch];
+      sum = vec_add(sum, vec_sub(window[15], window[0]));
+
+      // C: sumsq += s[7 * pitch] * s[7 * pitch] - s[-8 * pitch] * s[-8 *
+      // pitch];
+      // Optimization Note: Caching a squared-window for odd and even is
+      // slower than just repeating the multiplies.
+      d15_odd = vec_mulo(window[15], window[15]);
+      d15_even = vec_mule(window[15], window[15]);
+      d0_odd = vec_mulo(window[0], window[0]);
+      d0_even = vec_mule(window[0], window[0]);
+      sumsq_odd = vec_add(sumsq_odd, vec_sub(d15_odd, d0_odd));
+      sumsq_even = vec_add(sumsq_even, vec_sub(d15_even, d0_even));
+
+      // C: sumsq * 15 - sum * sum
+      sumsq_odd_scaled = vec_mul(sumsq_odd, vec_splats((int32_t)15));
+      sumsq_even_scaled = vec_mul(sumsq_even, vec_splats((int32_t)15));
+      thres_odd = vec_sub(sumsq_odd_scaled, vec_mulo(sum, sum));
+      thres_even = vec_sub(sumsq_even_scaled, vec_mule(sum, sum));
+
+      // C: (vpx_rv[(r & 127) + (c & 7)] + sum + s[0]) >> 4
+      filtered = vec_add(vec_add(rv, sum), window[8]);
+      filtered = vec_sra(filtered, vec_splats((uint16_t)4));
+
+      mask_odd = vec_cmplt(thres_odd, lim);
+      mask_even = vec_cmplt(thres_even, lim);
+      mask = vec_perm((bool16x8_t)mask_even, (bool16x8_t)mask_odd, mask_merge);
+      masked = vec_sel(window[8], filtered, mask);
+
+      // TODO(ltrudeau) If cols % 16 == 0, we could just process 16 per
+      // iteration
+      out = vec_perm((uint8x16_t)masked, vec_vsx_ld(0, dst + row * pitch),
+                     load_merge);
+      vec_vsx_st(out, 0, dst + row * pitch);
+
+      // Optimization Note: Turns out that the following loop is faster than
+      // using pointers to manage the sliding window.
+      for (i = 1; i < 16; i++) {
+        window[i - 1] = window[i];
+      }
+    }
+    dst += 8;
+  }
+}