ppc: Add convolve8_vert

Only the generic one again, speedups for 8x8 and larger blocks
to come later.

Change-Id: Ia509d6225984b4930ec03928c9bcbf51486da99f

1 files changed, 84 insertions, 8 deletions

diff --git a/vpx_dsp/ppc/vpx_convolve_vsx.c b/vpx_dsp/ppc/vpx_convolve_vsx.c
index af7032d9f..5a789c768 100644
--- a/vpx_dsp/ppc/vpx_convolve_vsx.c
+++ b/vpx_dsp/ppc/vpx_convolve_vsx.c
@@ -160,10 +160,8 @@ void vpx_convolve_avg_vsx(const uint8_t *src, ptrdiff_t src_stride,
   }
 }
 
-static inline void convolve_line(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);
+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));
@@ -173,6 +171,14 @@ static inline void convolve_line(uint8_t *dst, const uint8_t *const src_x,
   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,
@@ -184,8 +190,8 @@ static inline void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
   for (y = 0; y < h; ++y) {
     int x_q4 = x0_q4;
     for (x = 0; x < w; ++x) {
-      convolve_line(dst + x, &src[x_q4 >> SUBPEL_BITS],
-                    x_filters[x_q4 & SUBPEL_MASK]);
+      convolve_line_h(dst + x, &src[x_q4 >> SUBPEL_BITS],
+                      x_filters[x_q4 & SUBPEL_MASK]);
       x_q4 += x_step_q4;
     }
     src += src_stride;
@@ -204,8 +210,8 @@ static inline void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
     int x_q4 = x0_q4;
     for (x = 0; x < w; ++x) {
       uint8_t v;
-      convolve_line(&v, &src[x_q4 >> SUBPEL_BITS],
-                    x_filters[x_q4 & SUBPEL_MASK]);
+      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;
     }
@@ -214,6 +220,61 @@ static inline void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_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;
+  }
+}
+
 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,
@@ -243,3 +304,18 @@ void vpx_convolve8_avg_horiz_vsx(const uint8_t *src, ptrdiff_t src_stride,
   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);
+}