ref: d322d45170b16b2e9c89ff6063c3c0ea1cb7d9f9
dir: /src/ppc/cdef_init_tmpl.c/
/*
* Copyright © 2019, Luca Barbato
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
* ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <stdlib.h>
#include "common/bitdepth.h"
#include "common/intops.h"
#include "src/cdef.h"
#include "src/cpu.h"
#include "src/ppc/types.h"
#if BITDEPTH == 8
static inline i16x8 vconstrain(const i16x8 diff, const int16_t threshold,
const int damping)
{
const i16x8 zero = vec_splat_s16(0);
if (!threshold) return zero;
const uint16_t shift = imax(0, damping - ulog2(threshold));
const i16x8 abs_diff = vec_abs(diff);
const b16x8 mask = vec_cmplt(diff, zero);
const i16x8 thr = vec_splats(threshold);
const i16x8 sub = vec_sub(thr, vec_sra(abs_diff, vec_splats(shift)));
const i16x8 max = vec_max(zero, sub);
const i16x8 min = vec_min(abs_diff, max);
const i16x8 neg = vec_sub(zero, min);
return vec_sel(min, neg, mask);
}
static inline void copy4xN(uint16_t *tmp, const ptrdiff_t tmp_stride,
const uint8_t *src, const ptrdiff_t src_stride,
const uint8_t (*left)[2], uint8_t *const top[2],
const int w, const int h,
const enum CdefEdgeFlags edges)
{
const u16x8 fill = vec_splats((uint16_t)INT16_MAX);
u16x8 l0;
u16x8 l1;
int y_start = -2, y_end = h + 2;
// Copy top and bottom first
if (!(edges & CDEF_HAVE_TOP)) {
l0 = fill;
l1 = fill;
y_start = 0;
} else {
l0 = u8h_to_u16(vec_vsx_ld(0, top[0] - 2));
l1 = u8h_to_u16(vec_vsx_ld(0, top[1] - 2));
}
vec_st(l0, 0, tmp - 2 * 8);
vec_st(l1, 0, tmp - 1 * 8);
if (!(edges & CDEF_HAVE_BOTTOM)) {
l0 = fill;
l1 = fill;
y_end -= 2;
} else {
l0 = u8h_to_u16(vec_vsx_ld(0, src - 2 + (h + 0) * src_stride));
l1 = u8h_to_u16(vec_vsx_ld(0, src - 2 + (h + 1) * src_stride));
}
vec_st(l0, 0, tmp + (h + 0) * 8);
vec_st(l1, 0, tmp + (h + 1) * 8);
for (int y = 0; y < h; y++) {
u16x8 l = u8h_to_u16(vec_vsx_ld(0, src - 2 + y * src_stride));
vec_st(l, 0, tmp + y * 8);
}
if (!(edges & CDEF_HAVE_LEFT)) {
for (int y = y_start; y < y_end; y++) {
tmp[y * 8] = INT16_MAX;
tmp[1 + y * 8] = INT16_MAX;
}
} else {
for (int y = 0; y < h; y++) {
tmp[y * 8] = left[y][0];
tmp[1 + y * 8] = left[y][1];
}
}
if (!(edges & CDEF_HAVE_RIGHT)) {
for (int y = y_start; y < y_end; y++) {
tmp[- 2 + (y + 1) * 8] = INT16_MAX;
tmp[- 1 + (y + 1) * 8] = INT16_MAX;
}
}
}
static inline void copy8xN(uint16_t *tmp, const ptrdiff_t tmp_stride,
const uint8_t *src, const ptrdiff_t src_stride,
const uint8_t (*left)[2], uint8_t *const top[2],
const int w, const int h,
const enum CdefEdgeFlags edges)
{
const u16x8 fill = vec_splats((uint16_t)INT16_MAX);
u16x8 l0h, l0l;
u16x8 l1h, l1l;
int y_start = -2, y_end = h + 2;
// Copy top and bottom first
if (!(edges & CDEF_HAVE_TOP)) {
l0h = fill;
l0l = fill;
l1h = fill;
l1l = fill;
y_start = 0;
} else {
u8x16 l0 = vec_vsx_ld(0, top[0] - 2);
u8x16 l1 = vec_vsx_ld(0, top[1] - 2);
l0h = u8h_to_u16(l0);
l0l = u8l_to_u16(l0);
l1h = u8h_to_u16(l1);
l1l = u8l_to_u16(l1);
}
vec_st(l0h, 0, tmp - 4 * 8);
vec_st(l0l, 0, tmp - 3 * 8);
vec_st(l1h, 0, tmp - 2 * 8);
vec_st(l1l, 0, tmp - 1 * 8);
if (!(edges & CDEF_HAVE_BOTTOM)) {
l0h = fill;
l0l = fill;
l1h = fill;
l1l = fill;
y_end -= 2;
} else {
u8x16 l0 = vec_vsx_ld(0, src - 2 + (h + 0) * src_stride);
u8x16 l1 = vec_vsx_ld(0, src - 2 + (h + 1) * src_stride);
l0h = u8h_to_u16(l0);
l0l = u8l_to_u16(l0);
l1h = u8h_to_u16(l1);
l1l = u8l_to_u16(l1);
}
vec_st(l0h, 0, tmp + (h + 0) * 16);
vec_st(l0l, 0, tmp + (h + 0) * 16 + 8);
vec_st(l1h, 0, tmp + (h + 1) * 16);
vec_st(l1l, 0, tmp + (h + 1) * 16 + 8);
for (int y = 0; y < h; y++) {
u8x16 l = vec_vsx_ld(0, src - 2 + y * src_stride);
u16x8 lh = u8h_to_u16(l);
u16x8 ll = u8l_to_u16(l);
vec_st(lh, 0, tmp + y * 16);
vec_st(ll, 0, tmp + 8 + y * 16);
}
if (!(edges & CDEF_HAVE_LEFT)) {
for (int y = y_start; y < y_end; y++) {
tmp[y * 16] = INT16_MAX;
tmp[1 + y * 16] = INT16_MAX;
}
} else {
for (int y = 0; y < h; y++) {
tmp[y * 16] = left[y][0];
tmp[1 + y * 16] = left[y][1];
}
}
if (!(edges & CDEF_HAVE_RIGHT)) {
for (int y = y_start; y < y_end; y++) {
tmp[- 6 + (y + 1) * 16] = INT16_MAX;
tmp[- 5 + (y + 1) * 16] = INT16_MAX;
}
}
}
static inline i16x8 max_mask(i16x8 a, i16x8 b) {
const i16x8 I16X8_INT16_MAX = vec_splats((int16_t)INT16_MAX);
const b16x8 mask = vec_cmpeq(a, I16X8_INT16_MAX);
const i16x8 val = vec_sel(a, b, mask);
return vec_max(val, b);
}
#define LOAD_PIX(addr) \
const i16x8 px = (i16x8)vec_vsx_ld(0, addr); \
i16x8 max = px; \
i16x8 min = px; \
i16x8 sum = vec_splat_s16(0);
#define LOAD_PIX4(addr) \
const i16x8 a = (i16x8)vec_vsx_ld(0, addr); \
const i16x8 b = (i16x8)vec_vsx_ld(0, addr + tmp_stride); \
const i16x8 px = vec_xxpermdi(a, b, 0); \
i16x8 max = px; \
i16x8 min = px; \
i16x8 sum = vec_splat_s16(0);
#define LOAD_DIR(p, addr, o0, o1) \
const i16x8 p ## 0 = (i16x8)vec_vsx_ld(0, addr + o0); \
const i16x8 p ## 1 = (i16x8)vec_vsx_ld(0, addr - o0); \
const i16x8 p ## 2 = (i16x8)vec_vsx_ld(0, addr + o1); \
const i16x8 p ## 3 = (i16x8)vec_vsx_ld(0, addr - o1);
#define LOAD_DIR4(p, addr, o0, o1) \
LOAD_DIR(p ## a, addr, o0, o1) \
LOAD_DIR(p ## b, addr + tmp_stride, o0, o1) \
const i16x8 p ## 0 = vec_xxpermdi(p ## a ## 0, p ## b ## 0, 0); \
const i16x8 p ## 1 = vec_xxpermdi(p ## a ## 1, p ## b ## 1, 0); \
const i16x8 p ## 2 = vec_xxpermdi(p ## a ## 2, p ## b ## 2, 0); \
const i16x8 p ## 3 = vec_xxpermdi(p ## a ## 3, p ## b ## 3, 0);
#define CONSTRAIN(p, strength) \
const i16x8 p ## _d0 = vec_sub(p ## 0, px); \
const i16x8 p ## _d1 = vec_sub(p ## 1, px); \
const i16x8 p ## _d2 = vec_sub(p ## 2, px); \
const i16x8 p ## _d3 = vec_sub(p ## 3, px); \
\
i16x8 p ## _c0 = vconstrain(p ## _d0, strength, damping); \
i16x8 p ## _c1 = vconstrain(p ## _d1, strength, damping); \
i16x8 p ## _c2 = vconstrain(p ## _d2, strength, damping); \
i16x8 p ## _c3 = vconstrain(p ## _d3, strength, damping);
#define MIN_MAX(p) \
max = max_mask(p ## 0, max); \
min = vec_min(p ## 0, min); \
max = max_mask(p ## 1, max); \
min = vec_min(p ## 1, min); \
max = max_mask(p ## 2, max); \
min = vec_min(p ## 2, min); \
max = max_mask(p ## 3, max); \
min = vec_min(p ## 3, min);
#define PRI_0(p) \
p ## _c0 = vec_add(vec_sl(p ## _c0, vec_splat_u16(1)), vec_sl(p ## _c0, vec_splats(tap_even))); \
p ## _c1 = vec_add(vec_sl(p ## _c1, vec_splat_u16(1)), vec_sl(p ## _c1, vec_splats(tap_even)));
#define PRI_1(p) \
p ## _c2 = vec_sub(vec_sl(p ## _c2, vec_splat_u16(2)), vec_sl(p ## _c2, vec_splats(tap_even))); \
p ## _c3 = vec_sub(vec_sl(p ## _c3, vec_splat_u16(2)), vec_sl(p ## _c3, vec_splats(tap_even)));
#define SEC_0(p) \
p ## _c0 = vec_sl(p ## _c0, vec_splat_u16(1)); \
p ## _c1 = vec_sl(p ## _c1, vec_splat_u16(1)); \
p ## _c2 = vec_sl(p ## _c2, vec_splat_u16(1)); \
p ## _c3 = vec_sl(p ## _c3, vec_splat_u16(1));
#define UPDATE_SUM(p) \
const i16x8 p ## sum0 = vec_add(p ## _c0, p ## _c1); \
const i16x8 p ## sum1 = vec_add(p ## _c2, p ## _c3); \
sum = vec_add(sum, p ## sum0); \
sum = vec_add(sum, p ## sum1);
static inline void
filter_4xN(pixel *dst, const ptrdiff_t dst_stride,
const pixel (*left)[2], /*const*/ pixel *const top[2],
const int w, const int h, const int pri_strength,
const int sec_strength, const int dir,
const int damping, const enum CdefEdgeFlags edges,
const ptrdiff_t tmp_stride, uint16_t *tmp)
{
const int8_t cdef_directions[8 /* dir */][2 /* pass */] = {
{ -1 * tmp_stride + 1, -2 * tmp_stride + 2 },
{ 0 * tmp_stride + 1, -1 * tmp_stride + 2 },
{ 0 * tmp_stride + 1, 0 * tmp_stride + 2 },
{ 0 * tmp_stride + 1, 1 * tmp_stride + 2 },
{ 1 * tmp_stride + 1, 2 * tmp_stride + 2 },
{ 1 * tmp_stride + 0, 2 * tmp_stride + 1 },
{ 1 * tmp_stride + 0, 2 * tmp_stride + 0 },
{ 1 * tmp_stride + 0, 2 * tmp_stride - 1 }
};
const int bitdepth_min_8 = bitdepth_from_max(bitdepth_max) - 8;
const uint16_t tap_even = !((pri_strength >> bitdepth_min_8) & 1);
const int off1 = cdef_directions[dir][0];
const int off1_1 = cdef_directions[dir][1];
const int off2 = cdef_directions[(dir + 2) & 7][0];
const int off3 = cdef_directions[(dir + 6) & 7][0];
const int off2_1 = cdef_directions[(dir + 2) & 7][1];
const int off3_1 = cdef_directions[(dir + 6) & 7][1];
copy4xN(tmp - 2, tmp_stride, dst, dst_stride, left, top, w, h, edges);
for (int y = 0; y < h / 2; y++) {
LOAD_PIX4(tmp)
// Primary pass
LOAD_DIR4(p, tmp, off1, off1_1)
CONSTRAIN(p, pri_strength)
MIN_MAX(p)
PRI_0(p)
PRI_1(p)
UPDATE_SUM(p)
// Secondary pass 1
LOAD_DIR4(s, tmp, off2, off3)
CONSTRAIN(s, sec_strength)
MIN_MAX(s)
SEC_0(s)
UPDATE_SUM(s)
// Secondary pass 2
LOAD_DIR4(s2, tmp, off2_1, off3_1)
CONSTRAIN(s2, sec_strength)
MIN_MAX(s2)
UPDATE_SUM(s2)
// Store
i16x8 bias = vec_and((i16x8)vec_cmplt(sum, vec_splat_s16(0)), vec_splat_s16(1));
bias = vec_sub(vec_splat_s16(8), bias);
i16x8 unclamped = vec_add(px, vec_sra(vec_add(sum, bias), vec_splat_u16(4)));
i16x8 vdst = vec_max(vec_min(unclamped, max), min);
dst[0] = vdst[0];
dst[1] = vdst[1];
dst[2] = vdst[2];
dst[3] = vdst[3];
tmp += tmp_stride;
dst += PXSTRIDE(dst_stride);
dst[0] = vdst[4];
dst[1] = vdst[5];
dst[2] = vdst[6];
dst[3] = vdst[7];
tmp += tmp_stride;
dst += PXSTRIDE(dst_stride);
}
}
static inline void
filter_8xN(pixel *dst, const ptrdiff_t dst_stride,
const pixel (*left)[2], /*const*/ pixel *const top[2],
const int w, const int h, const int pri_strength,
const int sec_strength, const int dir,
const int damping, const enum CdefEdgeFlags edges,
const ptrdiff_t tmp_stride, uint16_t *tmp)
{
const int8_t cdef_directions[8 /* dir */][2 /* pass */] = {
{ -1 * tmp_stride + 1, -2 * tmp_stride + 2 },
{ 0 * tmp_stride + 1, -1 * tmp_stride + 2 },
{ 0 * tmp_stride + 1, 0 * tmp_stride + 2 },
{ 0 * tmp_stride + 1, 1 * tmp_stride + 2 },
{ 1 * tmp_stride + 1, 2 * tmp_stride + 2 },
{ 1 * tmp_stride + 0, 2 * tmp_stride + 1 },
{ 1 * tmp_stride + 0, 2 * tmp_stride + 0 },
{ 1 * tmp_stride + 0, 2 * tmp_stride - 1 }
};
const int bitdepth_min_8 = bitdepth_from_max(bitdepth_max) - 8;
const uint16_t tap_even = !((pri_strength >> bitdepth_min_8) & 1);
const int off1 = cdef_directions[dir][0];
const int off1_1 = cdef_directions[dir][1];
const int off2 = cdef_directions[(dir + 2) & 7][0];
const int off3 = cdef_directions[(dir + 6) & 7][0];
const int off2_1 = cdef_directions[(dir + 2) & 7][1];
const int off3_1 = cdef_directions[(dir + 6) & 7][1];
copy8xN(tmp - 2, tmp_stride, dst, dst_stride, left, top, w, h, edges);
for (int y = 0; y < h; y++) {
LOAD_PIX(tmp)
// Primary pass
LOAD_DIR(p, tmp, off1, off1_1)
CONSTRAIN(p, pri_strength)
MIN_MAX(p)
PRI_0(p)
PRI_1(p)
UPDATE_SUM(p)
// Secondary pass 1
LOAD_DIR(s, tmp, off2, off3)
CONSTRAIN(s, sec_strength)
MIN_MAX(s)
SEC_0(s)
UPDATE_SUM(s)
// Secondary pass 2
LOAD_DIR(s2, tmp, off2_1, off3_1)
CONSTRAIN(s2, sec_strength)
MIN_MAX(s2)
UPDATE_SUM(s2)
// Store
i16x8 bias = vec_and((i16x8)vec_cmplt(sum, vec_splat_s16(0)), vec_splat_s16(1));
bias = vec_sub(vec_splat_s16(8), bias);
i16x8 unclamped = vec_add(px, vec_sra(vec_add(sum, bias), vec_splat_u16(4)));
i16x8 vdst = vec_max(vec_min(unclamped, max), min);
dst[0] = vdst[0];
dst[1] = vdst[1];
dst[2] = vdst[2];
dst[3] = vdst[3];
dst[4] = vdst[4];
dst[5] = vdst[5];
dst[6] = vdst[6];
dst[7] = vdst[7];
tmp += tmp_stride;
dst += PXSTRIDE(dst_stride);
}
}
#define cdef_fn(w, h, tmp_stride) \
static void cdef_filter_##w##x##h##_vsx(pixel *const dst, \
const ptrdiff_t dst_stride, \
const pixel (*left)[2], \
/*const*/ pixel *const top[2], \
const int pri_strength, \
const int sec_strength, \
const int dir, \
const int damping, \
const enum CdefEdgeFlags edges) \
{ \
ALIGN_STK_16(uint16_t, tmp_buf, 12 * tmp_stride,); \
uint16_t *tmp = tmp_buf + 2 * tmp_stride + 2; \
filter_##w##xN(dst, dst_stride, left, top, w, h, pri_strength, sec_strength, \
dir, damping, edges, tmp_stride, tmp); \
}
cdef_fn(4, 4, 8);
cdef_fn(4, 8, 8);
cdef_fn(8, 8, 16);
#endif
COLD void bitfn(dav1d_cdef_dsp_init_ppc)(Dav1dCdefDSPContext *const c) {
const unsigned flags = dav1d_get_cpu_flags();
if (!(flags & DAV1D_PPC_CPU_FLAG_VSX)) return;
#if BITDEPTH == 8
// c->dir = dav1d_cdef_find_dir_vsx;
c->fb[0] = cdef_filter_8x8_vsx;
c->fb[1] = cdef_filter_4x8_vsx;
c->fb[2] = cdef_filter_4x4_vsx;
#endif
}