VSX version of vpx_mbpost_proc_ip

Low bit depth version only. Passes the VpxMbPostProcAcrossIpTest.

VpxMbPostProcAcrossIpTest Speed Test (POWER8 Model 2.1)
C time = 188.5ms (±0.2ms), VSX time = 65.2ms (±0.1ms) [2.9x]

Change-Id: I1cf72365d94a9d7f1e9323925a87a30e3bd5cfe2

test/pp_filter_test.cc

@@ -164,11 +164,21 @@ TEST_P(VpxPostProcDownAndAcrossMbRowTest, CheckCvsAssembly) {
 }
 
 class VpxMbPostProcAcrossIpTest
-    : public ::testing::TestWithParam<VpxMbPostProcAcrossIpFunc> {
+    : public AbstractBench,
+      public ::testing::TestWithParam<VpxMbPostProcAcrossIpFunc> {
  public:
+  VpxMbPostProcAcrossIpTest()
+      : rows(16), cols(16), mbPostProcAcrossIp(GetParam()),
+        src(Buffer<uint8_t>(rows, cols, 8, 8, 17, 8)) {}
   virtual void TearDown() { libvpx_test::ClearSystemState(); }
 
  protected:
+  const int rows;
+  const int cols;
+  const VpxMbPostProcAcrossIpFunc mbPostProcAcrossIp;
+  Buffer<uint8_t> src;
+  void run();
+
   void SetCols(unsigned char *s, int rows, int cols, int src_width) {
     for (int r = 0; r < rows; r++) {
       for (int c = 0; c < cols; c++) {
@@ -197,11 +207,11 @@ class VpxMbPostProcAcrossIpTest
   }
 };
 
-TEST_P(VpxMbPostProcAcrossIpTest, CheckLowFilterOutput) {
+void VpxMbPostProcAcrossIpTest::run() {
-  const int rows = 16;
+  mbPostProcAcrossIp(src.TopLeftPixel(), src.stride(), rows, cols, q2mbl(0));
-  const int cols = 16;
+}
 
-  Buffer<uint8_t> src = Buffer<uint8_t>(cols, rows, 8, 8, 17, 8);
+TEST_P(VpxMbPostProcAcrossIpTest, CheckLowFilterOutput) {
   ASSERT_TRUE(src.Init());
   src.SetPadding(10);
   SetCols(src.TopLeftPixel(), rows, cols, src.stride());
@@ -215,15 +225,11 @@ TEST_P(VpxMbPostProcAcrossIpTest, CheckLowFilterOutput) {
 }
 
 TEST_P(VpxMbPostProcAcrossIpTest, CheckMediumFilterOutput) {
-  const int rows = 16;
-  const int cols = 16;
-
-  Buffer<uint8_t> src = Buffer<uint8_t>(cols, rows, 8, 8, 17, 8);
   ASSERT_TRUE(src.Init());
   src.SetPadding(10);
   SetCols(src.TopLeftPixel(), rows, cols, src.stride());
 
-  static const unsigned char kExpectedOutput[cols] = {
+  static const unsigned char kExpectedOutput[] = {
     2, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 13
   };
 
@@ -232,15 +238,11 @@ TEST_P(VpxMbPostProcAcrossIpTest, CheckMediumFilterOutput) {
 }
 
 TEST_P(VpxMbPostProcAcrossIpTest, CheckHighFilterOutput) {
-  const int rows = 16;
-  const int cols = 16;
-
-  Buffer<uint8_t> src = Buffer<uint8_t>(cols, rows, 8, 8, 17, 8);
   ASSERT_TRUE(src.Init());
   src.SetPadding(10);
   SetCols(src.TopLeftPixel(), rows, cols, src.stride());
 
-  static const unsigned char kExpectedOutput[cols] = {
+  static const unsigned char kExpectedOutput[] = {
     2, 2, 3, 4, 4, 5, 6, 7, 8, 9, 10, 11, 11, 12, 13, 13
   };
 
@@ -254,9 +256,6 @@ TEST_P(VpxMbPostProcAcrossIpTest, CheckHighFilterOutput) {
 }
 
 TEST_P(VpxMbPostProcAcrossIpTest, CheckCvsAssembly) {
-  const int rows = 16;
-  const int cols = 16;
-
   Buffer<uint8_t> c_mem = Buffer<uint8_t>(cols, rows, 8, 8, 17, 8);
   ASSERT_TRUE(c_mem.Init());
   Buffer<uint8_t> asm_mem = Buffer<uint8_t>(cols, rows, 8, 8, 17, 8);
@@ -279,6 +278,16 @@ TEST_P(VpxMbPostProcAcrossIpTest, CheckCvsAssembly) {
   }
 }
 
+TEST_P(VpxMbPostProcAcrossIpTest, DISABLED_Speed) {
+  ASSERT_TRUE(src.Init());
+  src.SetPadding(10);
+
+  SetCols(src.TopLeftPixel(), rows, cols, src.stride());
+
+  runNTimes(100000);
+  printMedian("16x16");
+}
+
 class VpxMbPostProcDownTest
     : public AbstractBench,
       public ::testing::TestWithParam<VpxMbPostProcDownFunc> {
@@ -491,6 +500,10 @@ INSTANTIATE_TEST_CASE_P(MSA, VpxMbPostProcDownTest,
 #endif  // HAVE_MSA
 
 #if HAVE_VSX
+
+INSTANTIATE_TEST_CASE_P(VSX, VpxMbPostProcAcrossIpTest,
+                        ::testing::Values(vpx_mbpost_proc_across_ip_vsx));
+
 INSTANTIATE_TEST_CASE_P(VSX, VpxMbPostProcDownTest,
                         ::testing::Values(vpx_mbpost_proc_down_vsx));
 #endif  // HAVE_VSX

vpx_dsp/ppc/deblock_vsx.c

@@ -19,9 +19,170 @@ 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,
+// C: s[c + 7]
-                                       0x14, 0x15, 0x08, 0x09, 0x18, 0x19,
+static INLINE int16x8_t next7l_s16(uint8x16_t c) {
-                                       0x0C, 0x0D, 0x1C, 0x1D };
+  static const uint8x16_t next7_perm = {
+    0x07, 0x10, 0x08, 0x11, 0x09, 0x12, 0x0A, 0x13,
+    0x0B, 0x14, 0x0C, 0x15, 0x0D, 0x16, 0x0E, 0x17,
+  };
+  return (int16x8_t)vec_perm(c, vec_zeros_u8, next7_perm);
+}
+
+// Slide across window and add.
+static INLINE int16x8_t slide_sum_s16(int16x8_t x) {
+  // x = A B C D E F G H
+  //
+  // 0 A B C D E F G
+  const int16x8_t sum1 = vec_add(x, vec_slo(x, vec_splats((int8_t)(2 << 3))));
+  // 0 0 A B C D E F
+  const int16x8_t sum2 = vec_add(vec_slo(x, vec_splats((int8_t)(4 << 3))),
+                                 // 0 0 0 A B C D E
+                                 vec_slo(x, vec_splats((int8_t)(6 << 3))));
+  // 0 0 0 0 A B C D
+  const int16x8_t sum3 = vec_add(vec_slo(x, vec_splats((int8_t)(8 << 3))),
+                                 // 0 0 0 0 0 A B C
+                                 vec_slo(x, vec_splats((int8_t)(10 << 3))));
+  // 0 0 0 0 0 0 A B
+  const int16x8_t sum4 = vec_add(vec_slo(x, vec_splats((int8_t)(12 << 3))),
+                                 // 0 0 0 0 0 0 0 A
+                                 vec_slo(x, vec_splats((int8_t)(14 << 3))));
+  return vec_add(vec_add(sum1, sum2), vec_add(sum3, sum4));
+}
+
+// Slide across window and add.
+static INLINE int32x4_t slide_sumsq_s32(int32x4_t xsq_even, int32x4_t xsq_odd) {
+  //   0 A C E
+  // + 0 B D F
+  int32x4_t sumsq_1 = vec_add(vec_slo(xsq_even, vec_splats((int8_t)(4 << 3))),
+                              vec_slo(xsq_odd, vec_splats((int8_t)(4 << 3))));
+  //   0 0 A C
+  // + 0 0 B D
+  int32x4_t sumsq_2 = vec_add(vec_slo(xsq_even, vec_splats((int8_t)(8 << 3))),
+                              vec_slo(xsq_odd, vec_splats((int8_t)(8 << 3))));
+  //   0 0 0 A
+  // + 0 0 0 B
+  int32x4_t sumsq_3 = vec_add(vec_slo(xsq_even, vec_splats((int8_t)(12 << 3))),
+                              vec_slo(xsq_odd, vec_splats((int8_t)(12 << 3))));
+  sumsq_1 = vec_add(sumsq_1, xsq_even);
+  sumsq_2 = vec_add(sumsq_2, sumsq_3);
+  return vec_add(sumsq_1, sumsq_2);
+}
+
+// C: (b + sum + val) >> 4
+static INLINE int16x8_t filter_s16(int16x8_t b, int16x8_t sum, int16x8_t val) {
+  return vec_sra(vec_add(vec_add(b, sum), val), vec_splats((uint16_t)4));
+}
+
+// C: sumsq * 15 - sum * sum
+static INLINE bool16x8_t mask_s16(int32x4_t sumsq_even, int32x4_t sumsq_odd,
+                                  int16x8_t sum, int32x4_t lim) {
+  static const uint8x16_t mask_merge = { 0x00, 0x01, 0x10, 0x11, 0x04, 0x05,
+                                         0x14, 0x15, 0x08, 0x09, 0x18, 0x19,
+                                         0x0C, 0x0D, 0x1C, 0x1D };
+  const int32x4_t sumsq_odd_scaled =
+      vec_mul(sumsq_odd, vec_splats((int32_t)15));
+  const int32x4_t sumsq_even_scaled =
+      vec_mul(sumsq_even, vec_splats((int32_t)15));
+  const int32x4_t thres_odd = vec_sub(sumsq_odd_scaled, vec_mulo(sum, sum));
+  const int32x4_t thres_even = vec_sub(sumsq_even_scaled, vec_mule(sum, sum));
+
+  const bool32x4_t mask_odd = vec_cmplt(thres_odd, lim);
+  const bool32x4_t mask_even = vec_cmplt(thres_even, lim);
+  return vec_perm((bool16x8_t)mask_even, (bool16x8_t)mask_odd, mask_merge);
+}
+
+void vpx_mbpost_proc_across_ip_vsx(unsigned char *src, int pitch, int rows,
+                                   int cols, int flimit) {
+  int row, col;
+  const int32x4_t lim = vec_splats(flimit);
+
+  // 8 columns are processed at a time.
+  assert(cols % 8 == 0);
+
+  for (row = 0; row < rows; row++) {
+    // The sum is signed and requires at most 13 bits.
+    // (8 bits + sign) * 15 (4 bits)
+    int16x8_t sum;
+    // The sum of squares requires at most 20 bits.
+    // (16 bits + sign) * 15 (4 bits)
+    int32x4_t sumsq_even, sumsq_odd;
+
+    // Fill left context with first col.
+    int16x8_t left_ctx = vec_splats((int16_t)src[0]);
+    int16_t s = src[0] * 9;
+    int32_t ssq = src[0] * src[0] * 9 + 16;
+
+    // Fill the next 6 columns of the sliding window with cols 2 to 7.
+    for (col = 1; col <= 6; ++col) {
+      s += src[col];
+      ssq += src[col] * src[col];
+    }
+    // Set this sum to every element in the window.
+    sum = vec_splats(s);
+    sumsq_even = vec_splats(ssq);
+    sumsq_odd = vec_splats(ssq);
+
+    for (col = 0; col < cols; col += 8) {
+      bool16x8_t mask;
+      int16x8_t filtered, masked;
+      uint8x16_t out;
+
+      const uint8x16_t val = vec_vsx_ld(0, src + col);
+      const int16x8_t val_high = unpack_to_s16_h(val);
+
+      // C: s[c + 7]
+      const int16x8_t right_ctx = (col + 8 == cols)
+                                      ? vec_splats((int16_t)src[col + 7])
+                                      : next7l_s16(val);
+
+      // C: x = s[c + 7] - s[c - 8];
+      const int16x8_t x = vec_sub(right_ctx, left_ctx);
+      const int32x4_t xsq_even =
+          vec_sub(vec_mule(right_ctx, right_ctx), vec_mule(left_ctx, left_ctx));
+      const int32x4_t xsq_odd =
+          vec_sub(vec_mulo(right_ctx, right_ctx), vec_mulo(left_ctx, left_ctx));
+
+      const int32x4_t sumsq_tmp = slide_sumsq_s32(xsq_even, xsq_odd);
+      // A C E G
+      // 0 B D F
+      // 0 A C E
+      // 0 0 B D
+      // 0 0 A C
+      // 0 0 0 B
+      // 0 0 0 A
+      sumsq_even = vec_add(sumsq_even, sumsq_tmp);
+      // B D F G
+      // A C E G
+      // 0 B D F
+      // 0 A C E
+      // 0 0 B D
+      // 0 0 A C
+      // 0 0 0 B
+      // 0 0 0 A
+      sumsq_odd = vec_add(sumsq_odd, vec_add(sumsq_tmp, xsq_odd));
+
+      sum = vec_add(sum, slide_sum_s16(x));
+
+      // C: (8 + sum + s[c]) >> 4
+      filtered = filter_s16(vec_splats((int16_t)8), sum, val_high);
+      // C: sumsq * 15 - sum * sum
+      mask = mask_s16(sumsq_even, sumsq_odd, sum, lim);
+      masked = vec_sel(val_high, filtered, mask);
+
+      out = vec_perm((uint8x16_t)masked, vec_vsx_ld(0, src + col), load_merge);
+      vec_vsx_st(out, 0, src + col);
+
+      // Update window sum and square sum
+      sum = vec_splat(sum, 7);
+      sumsq_even = vec_splat(sumsq_odd, 3);
+      sumsq_odd = vec_splat(sumsq_odd, 3);
+
+      // C: s[c - 8] (for next iteration)
+      left_ctx = val_high;
+    }
+    src += pitch;
+  }
+}
 
 void vpx_mbpost_proc_down_vsx(uint8_t *dst, int pitch, int rows, int cols,
                               int flimit) {
@@ -66,9 +227,6 @@ void vpx_mbpost_proc_down_vsx(uint8_t *dst, int pitch, int rows, int cols,
 
     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;
@@ -96,19 +254,11 @@ void vpx_mbpost_proc_down_vsx(uint8_t *dst, int pitch, int rows, int cols,
       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 = filter_s16(rv, sum, window[8]);
-      filtered = vec_sra(filtered, vec_splats((uint16_t)4));
 
-      mask_odd = vec_cmplt(thres_odd, lim);
+      // C: sumsq * 15 - sum * sum
-      mask_even = vec_cmplt(thres_even, lim);
+      mask = mask_s16(sumsq_even, sumsq_odd, sum, 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

vpx_dsp/ppc/types_vsx.h

@@ -67,10 +67,13 @@ static const uint8x16_t xxpermdi3_perm = { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D,
 #endif
 #endif
 
+static const uint8x16_t vec_zeros_u8 = { 0, 0, 0, 0, 0, 0, 0, 0,
+                                         0, 0, 0, 0, 0, 0, 0, 0 };
 static const int16x8_t vec_zeros_s16 = { 0, 0, 0, 0, 0, 0, 0, 0 };
 static const int16x8_t vec_ones_s16 = { 1, 1, 1, 1, 1, 1, 1, 1 };
 static const uint16x8_t vec_ones_u16 = { 1, 1, 1, 1, 1, 1, 1, 1 };
 static const uint32x4_t vec_ones_u32 = { 1, 1, 1, 1 };
+static const int32x4_t vec_zeros_s32 = { 0, 0, 0, 0 };
 static const uint16x8_t vec_shift_sign_s16 = { 15, 15, 15, 15, 15, 15, 15, 15 };
 static const uint32x4_t vec_shift_sign_s32 = { 31, 31, 31, 31 };
 static const uint8x16_t vec_perm64 = { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D,

vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -1601,7 +1601,7 @@ if (vpx_config("CONFIG_POSTPROC") eq "yes" || vpx_config("CONFIG_VP9_POSTPROC")
     specialize qw/vpx_mbpost_proc_down sse2 neon msa vsx/;
 
     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 neon msa/;
+    specialize qw/vpx_mbpost_proc_across_ip sse2 neon msa vsx/;
 
     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/;