ref: e94000aa35c5beb8901952abb9c0d037d3c977f1
dir: /vpx_dsp/ppc/deblock_vsx.c/
/*
* Copyright (c) 2018 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* 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 "./vpx_dsp_rtcd.h"
#include "vpx_dsp/ppc/types_vsx.h"
extern const int16_t vpx_rv[];
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 st8_perm = { 0x00, 0x01, 0x02, 0x03, 0x04, 0x05,
0x06, 0x07, 0x18, 0x19, 0x1A, 0x1B,
0x1C, 0x1D, 0x1E, 0x1F };
static INLINE uint8x16_t apply_filter(uint8x16_t ctx[4], uint8x16_t v,
uint8x16_t filter) {
const uint8x16_t k1 = vec_avg(ctx[0], ctx[1]);
const uint8x16_t k2 = vec_avg(ctx[3], ctx[2]);
const uint8x16_t k3 = vec_avg(k1, k2);
const uint8x16_t f_a = vec_max(vec_absd(v, ctx[0]), vec_absd(v, ctx[1]));
const uint8x16_t f_b = vec_max(vec_absd(v, ctx[2]), vec_absd(v, ctx[3]));
const bool8x16_t mask = vec_cmplt(vec_max(f_a, f_b), filter);
return vec_sel(v, vec_avg(k3, v), mask);
}
static INLINE void vert_ctx(uint8x16_t ctx[4], int col, uint8_t *src,
int stride) {
ctx[0] = vec_vsx_ld(col - 2 * stride, src);
ctx[1] = vec_vsx_ld(col - stride, src);
ctx[2] = vec_vsx_ld(col + stride, src);
ctx[3] = vec_vsx_ld(col + 2 * stride, src);
}
static INLINE void horz_ctx(uint8x16_t ctx[4], uint8x16_t left_ctx,
uint8x16_t v, uint8x16_t right_ctx) {
static const uint8x16_t l2_perm = { 0x0E, 0x0F, 0x10, 0x11, 0x12, 0x13,
0x14, 0x15, 0x16, 0x17, 0x18, 0x19,
0x1A, 0x1B, 0x1C, 0x1D };
static const uint8x16_t l1_perm = { 0x0F, 0x10, 0x11, 0x12, 0x13, 0x14,
0x15, 0x16, 0x17, 0x18, 0x19, 0x1A,
0x1B, 0x1C, 0x1D, 0x1E };
static const uint8x16_t r1_perm = { 0x01, 0x02, 0x03, 0x04, 0x05, 0x06,
0x07, 0x08, 0x09, 0x0A, 0x0B, 0x0C,
0x0D, 0x0E, 0x0F, 0x10 };
static const uint8x16_t r2_perm = { 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D,
0x0E, 0x0F, 0x10, 0x11 };
ctx[0] = vec_perm(left_ctx, v, l2_perm);
ctx[1] = vec_perm(left_ctx, v, l1_perm);
ctx[2] = vec_perm(v, right_ctx, r1_perm);
ctx[3] = vec_perm(v, right_ctx, r2_perm);
}
void vpx_post_proc_down_and_across_mb_row_vsx(unsigned char *src_ptr,
unsigned char *dst_ptr,
int src_pixels_per_line,
int dst_pixels_per_line, int cols,
unsigned char *f, int size) {
int row, col;
uint8x16_t ctx[4], out, v, left_ctx;
for (row = 0; row < size; row++) {
for (col = 0; col < cols - 8; col += 16) {
const uint8x16_t filter = vec_vsx_ld(col, f);
v = vec_vsx_ld(col, src_ptr);
vert_ctx(ctx, col, src_ptr, src_pixels_per_line);
vec_vsx_st(apply_filter(ctx, v, filter), col, dst_ptr);
}
if (col != cols) {
const uint8x16_t filter = vec_vsx_ld(col, f);
v = vec_vsx_ld(col, src_ptr);
vert_ctx(ctx, col, src_ptr, src_pixels_per_line);
out = apply_filter(ctx, v, filter);
vec_vsx_st(vec_perm(out, v, st8_perm), col, dst_ptr);
}
/* now post_proc_across */
left_ctx = vec_splats(dst_ptr[0]);
v = vec_vsx_ld(0, dst_ptr);
for (col = 0; col < cols - 8; col += 16) {
const uint8x16_t filter = vec_vsx_ld(col, f);
const uint8x16_t right_ctx = (col + 16 == cols)
? vec_splats(dst_ptr[cols - 1])
: vec_vsx_ld(col, dst_ptr + 16);
horz_ctx(ctx, left_ctx, v, right_ctx);
vec_vsx_st(apply_filter(ctx, v, filter), col, dst_ptr);
left_ctx = v;
v = right_ctx;
}
if (col != cols) {
const uint8x16_t filter = vec_vsx_ld(col, f);
const uint8x16_t right_ctx = vec_splats(dst_ptr[cols - 1]);
horz_ctx(ctx, left_ctx, v, right_ctx);
out = apply_filter(ctx, v, filter);
vec_vsx_st(vec_perm(out, v, st8_perm), col, dst_ptr);
}
src_ptr += src_pixels_per_line;
dst_ptr += dst_pixels_per_line;
}
}
// C: s[c + 7]
static INLINE int16x8_t next7l_s16(uint8x16_t c) {
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) {
int col, row, i;
int16x8_t window[16];
const int32x4_t lim = vec_splats(flimit);
// 8 columns are processed at a time.
assert(cols % 8 == 0);
// If rows is less than 8 the bottom border extension fails.
assert(rows >= 8);
for (col = 0; col < cols; col += 8) {
// The sum is signed and requires at most 13 bits.
// (8 bits + sign) * 15 (4 bits)
int16x8_t r1, sum;
// The sum of squares requires at most 20 bits.
// (16 bits + sign) * 15 (4 bits)
int32x4_t sumsq_even, sumsq_odd;
r1 = unpack_to_s16_h(vec_vsx_ld(0, dst));
// Fill sliding window with first row.
for (i = 0; i <= 8; i++) {
window[i] = r1;
}
// First 9 rows of the sliding window are the same.
// sum = r1 * 9
sum = vec_mladd(r1, vec_splats((int16_t)9), vec_zeros_s16);
// sumsq = r1 * r1 * 9
sumsq_even = vec_mule(sum, r1);
sumsq_odd = vec_mulo(sum, r1);
// Fill the next 6 rows of the sliding window with rows 2 to 7.
for (i = 1; i <= 6; ++i) {
const int16x8_t next_row = unpack_to_s16_h(vec_vsx_ld(i * pitch, dst));
window[i + 8] = next_row;
sum = vec_add(sum, next_row);
sumsq_odd = vec_add(sumsq_odd, vec_mulo(next_row, next_row));
sumsq_even = vec_add(sumsq_even, vec_mule(next_row, next_row));
}
for (row = 0; row < rows; row++) {
int32x4_t d15_even, d15_odd, d0_even, d0_odd;
bool16x8_t mask;
int16x8_t filtered, masked;
uint8x16_t out;
const int16x8_t rv = vec_vsx_ld(0, vpx_rv + (row & 127));
// Move the sliding window
if (row + 7 < rows) {
window[15] = unpack_to_s16_h(vec_vsx_ld((row + 7) * pitch, dst));
} else {
window[15] = window[14];
}
// C: sum += s[7 * pitch] - s[-8 * pitch];
sum = vec_add(sum, vec_sub(window[15], window[0]));
// C: sumsq += s[7 * pitch] * s[7 * pitch] - s[-8 * pitch] * s[-8 *
// pitch];
// Optimization Note: Caching a squared-window for odd and even is
// slower than just repeating the multiplies.
d15_odd = vec_mulo(window[15], window[15]);
d15_even = vec_mule(window[15], window[15]);
d0_odd = vec_mulo(window[0], window[0]);
d0_even = vec_mule(window[0], window[0]);
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: (vpx_rv[(r & 127) + (c & 7)] + sum + s[0]) >> 4
filtered = filter_s16(rv, sum, window[8]);
// C: sumsq * 15 - sum * sum
mask = mask_s16(sumsq_even, sumsq_odd, sum, lim);
masked = vec_sel(window[8], filtered, mask);
// TODO(ltrudeau) If cols % 16 == 0, we could just process 16 per
// iteration
out = vec_perm((uint8x16_t)masked, vec_vsx_ld(0, dst + row * pitch),
load_merge);
vec_vsx_st(out, 0, dst + row * pitch);
// Optimization Note: Turns out that the following loop is faster than
// using pointers to manage the sliding window.
for (i = 1; i < 16; i++) {
window[i - 1] = window[i];
}
}
dst += 8;
}
}