}
}
+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);
+ 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);
+}
+
// TODO(lu_zero): Implement 8x8 and bigger block special cases
-static 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) {
+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(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) {
- const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
- const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- const int16x8_t s = unpack_to_s16_h(vec_vsx_ld(0, src_x));
- const int16x8_t f = vec_vsx_ld(0, x_filter);
- 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 + x);
+ uint8_t v;
+ convolve_line(&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;
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);
+}