Merge changes Ia5293d94,I90d481d3,Ia509d622,I54549b03,I89b635d6

* changes:
  ppc: Add convolve8_vsx and convolve8_avg_vsx
  ppc: Add convolve8_avg_vert_vsx
  ppc: Add convolve8_vert
  ppc: Add convolve8_horiz_avg
  ppc: Add convolve8_horiz

test/convolve_test.cc

@@ -1224,11 +1224,11 @@ INSTANTIATE_TEST_CASE_P(MSA, ConvolveTest,
 
 #if HAVE_VSX
 const ConvolveFunctions convolve8_vsx(
-    vpx_convolve_copy_vsx, vpx_convolve_avg_vsx, vpx_convolve8_horiz_c,
-    vpx_convolve8_avg_horiz_c, vpx_convolve8_vert_c, vpx_convolve8_avg_vert_c,
-    vpx_convolve8_c, vpx_convolve8_avg_c, vpx_scaled_horiz_c,
-    vpx_scaled_avg_horiz_c, vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
-    vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+    vpx_convolve_copy_vsx, vpx_convolve_avg_vsx, vpx_convolve8_horiz_vsx,
+    vpx_convolve8_avg_horiz_vsx, vpx_convolve8_vert_vsx,
+    vpx_convolve8_avg_vert_vsx, vpx_convolve8_vsx, vpx_convolve8_avg_vsx,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c, vpx_scaled_vert_c,
+    vpx_scaled_avg_vert_c, vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
 const ConvolveParam kArrayConvolve_vsx[] = { ALL_SIZES(convolve8_vsx) };
 INSTANTIATE_TEST_CASE_P(VSX, ConvolveTest,
                         ::testing::ValuesIn(kArrayConvolve_vsx));

vpx_dsp/ppc/vpx_convolve_vsx.c

@@ -7,8 +7,10 @@
  *  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 <string.h>
 #include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_filter.h"
 #include "vpx_dsp/ppc/types_vsx.h"
 
 // TODO(lu_zero): unroll
@@ -158,3 +160,259 @@ void vpx_convolve_avg_vsx(const uint8_t *src, ptrdiff_t src_stride,
     }
   }
 }
+
+static inline void convolve_line(uint8_t *dst, const int16x8_t s,
+                                 const int16x8_t f) {
+  const int32x4_t sum = vec_msum(s, f, vec_splat_s32(0));
+  const int32x4_t bias =
+      vec_sl(vec_splat_s32(1), vec_splat_u32(FILTER_BITS - 1));
+  const int32x4_t avg = vec_sr(vec_sums(sum, bias), vec_splat_u32(FILTER_BITS));
+  const uint8x16_t v = vec_splat(
+      vec_packsu(vec_pack(avg, vec_splat_s32(0)), vec_splat_s16(0)), 3);
+  vec_ste(v, 0, dst);
+}
+
+static inline void convolve_line_h(uint8_t *dst, const uint8_t *const src_x,
+                                   const int16_t *const x_filter) {
+  const int16x8_t s = unpack_to_s16_h(vec_vsx_ld(0, src_x));
+  const int16x8_t f = vec_vsx_ld(0, x_filter);
+
+  convolve_line(dst, s, f);
+}
+
+// TODO(lu_zero): Implement 8x8 and bigger block special cases
+static inline void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const InterpKernel *x_filters, int x0_q4,
+                                  int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      convolve_line_h(dst + x, &src[x_q4 >> SUBPEL_BITS],
+                      x_filters[x_q4 & SUBPEL_MASK]);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static inline void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                                      uint8_t *dst, ptrdiff_t dst_stride,
+                                      const InterpKernel *x_filters, int x0_q4,
+                                      int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      uint8_t v;
+      convolve_line_h(&v, &src[x_q4 >> SUBPEL_BITS],
+                      x_filters[x_q4 & SUBPEL_MASK]);
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + v, 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static uint8x16_t transpose_line_u8_8x8(uint8x16_t a, uint8x16_t b,
+                                        uint8x16_t c, uint8x16_t d,
+                                        uint8x16_t e, uint8x16_t f,
+                                        uint8x16_t g, uint8x16_t h) {
+  uint16x8_t ab = (uint16x8_t)vec_mergeh(a, b);
+  uint16x8_t cd = (uint16x8_t)vec_mergeh(c, d);
+  uint16x8_t ef = (uint16x8_t)vec_mergeh(e, f);
+  uint16x8_t gh = (uint16x8_t)vec_mergeh(g, h);
+
+  uint32x4_t abcd = (uint32x4_t)vec_mergeh(ab, cd);
+  uint32x4_t efgh = (uint32x4_t)vec_mergeh(ef, gh);
+
+  return (uint8x16_t)vec_mergeh(abcd, efgh);
+}
+
+static inline void convolve_line_v(uint8_t *dst, const uint8_t *const src_y,
+                                   ptrdiff_t src_stride,
+                                   const int16_t *const y_filter) {
+  uint8x16_t s0 = vec_vsx_ld(0, src_y + 0 * src_stride);
+  uint8x16_t s1 = vec_vsx_ld(0, src_y + 1 * src_stride);
+  uint8x16_t s2 = vec_vsx_ld(0, src_y + 2 * src_stride);
+  uint8x16_t s3 = vec_vsx_ld(0, src_y + 3 * src_stride);
+  uint8x16_t s4 = vec_vsx_ld(0, src_y + 4 * src_stride);
+  uint8x16_t s5 = vec_vsx_ld(0, src_y + 5 * src_stride);
+  uint8x16_t s6 = vec_vsx_ld(0, src_y + 6 * src_stride);
+  uint8x16_t s7 = vec_vsx_ld(0, src_y + 7 * src_stride);
+  const int16x8_t f = vec_vsx_ld(0, y_filter);
+  uint8_t buf[16];
+  const uint8x16_t s = transpose_line_u8_8x8(s0, s1, s2, s3, s4, s5, s6, s7);
+
+  vec_vsx_st(s, 0, buf);
+
+  convolve_line(dst, unpack_to_s16_h(s), f);
+}
+
+static inline void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const InterpKernel *y_filters, int y0_q4,
+                                 int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      convolve_line_v(dst + y * dst_stride,
+                      &src[(y_q4 >> SUBPEL_BITS) * src_stride], src_stride,
+                      y_filters[y_q4 & SUBPEL_MASK]);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static inline void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const InterpKernel *y_filters, int y0_q4,
+                                     int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      uint8_t v;
+      convolve_line_v(&v, &src[(y_q4 >> SUBPEL_BITS) * src_stride], src_stride,
+                      y_filters[y_q4 & SUBPEL_MASK]);
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] + v, 1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static inline void convolve(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const InterpKernel *const x_filters, int x0_q4,
+                            int x_step_q4, const InterpKernel *const y_filters,
+                            int y0_q4, int y_step_q4, int w, int h) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  DECLARE_ALIGNED(16, uint8_t, temp[64 * 135]);
+  const int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64,
+                 x_filters, x0_q4, x_step_q4, w, intermediate_height);
+  convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride,
+                y_filters, y0_q4, y_step_q4, w, h);
+}
+
+void vpx_convolve8_horiz_vsx(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4, int w,
+                             int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+                 w, h);
+}
+
+void vpx_convolve8_avg_horiz_vsx(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x, int x_step_q4,
+                                 const int16_t *filter_y, int y_step_q4, int w,
+                                 int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                     x_step_q4, w, h);
+}
+
+void vpx_convolve8_vert_vsx(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4, int w,
+                            int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, y_step_q4,
+                w, h);
+}
+
+void vpx_convolve8_avg_vert_vsx(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int x_step_q4,
+                                const int16_t *filter_y, int y_step_q4, int w,
+                                int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                    y_step_q4, w, h);
+}
+
+void vpx_convolve8_vsx(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                       ptrdiff_t dst_stride, const int16_t *filter_x,
+                       int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                       int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+           filters_y, y0_q4, y_step_q4, w, h);
+}
+
+void vpx_convolve8_avg_vsx(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4, int w,
+                           int h) {
+  // Fixed size intermediate buffer places limits on parameters.
+  DECLARE_ALIGNED(16, uint8_t, temp[64 * 64]);
+  assert(w <= 64);
+  assert(h <= 64);
+
+  vpx_convolve8_vsx(src, src_stride, temp, 64, filter_x, x_step_q4, filter_y,
+                    y_step_q4, w, h);
+  vpx_convolve_avg_vsx(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
+}

vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -338,22 +338,22 @@ add_proto qw/void vpx_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride,
 specialize qw/vpx_convolve_avg neon dspr2 msa sse2 vsx/;
 
 add_proto qw/void vpx_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8 sse2 ssse3 avx2 neon dspr2 msa/;
+specialize qw/vpx_convolve8 sse2 ssse3 avx2 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8_horiz sse2 ssse3 avx2 neon dspr2 msa/;
+specialize qw/vpx_convolve8_horiz sse2 ssse3 avx2 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8_vert sse2 ssse3 avx2 neon dspr2 msa/;
+specialize qw/vpx_convolve8_vert sse2 ssse3 avx2 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa/;
+specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa/;
+specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
-specialize qw/vpx_convolve8_avg_vert sse2 ssse3 neon dspr2 msa/;
+specialize qw/vpx_convolve8_avg_vert sse2 ssse3 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 specialize qw/vpx_scaled_2d ssse3/;