vpx/vpx_dsp/arm/deblock_neon.c
Johann c23970ec25 postproc: vpx_mbpost_proc_down_neon
This was much more amenable to optimization than the across filter.
Speedup of almost 2.5x

BUG=webm:1320

Change-Id: I49acc0f9cb2e7642303df90132cbc938acade4c4
2017-01-09 10:21:56 -08:00

486 lines
15 KiB
C

/*
* Copyright (c) 2016 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 <arm_neon.h>
#include <assert.h>
#include "./vpx_dsp_rtcd.h"
#include "vpx/vpx_integer.h"
#include "vpx_dsp/arm/transpose_neon.h"
extern const int16_t vpx_rv[];
static uint8x8_t average_k_out(const uint8x8_t a2, const uint8x8_t a1,
const uint8x8_t v0, const uint8x8_t b1,
const uint8x8_t b2) {
const uint8x8_t k1 = vrhadd_u8(a2, a1);
const uint8x8_t k2 = vrhadd_u8(b2, b1);
const uint8x8_t k3 = vrhadd_u8(k1, k2);
return vrhadd_u8(k3, v0);
}
static uint8x8_t generate_mask(const uint8x8_t a2, const uint8x8_t a1,
const uint8x8_t v0, const uint8x8_t b1,
const uint8x8_t b2, const uint8x8_t filter) {
const uint8x8_t a2_v0 = vabd_u8(a2, v0);
const uint8x8_t a1_v0 = vabd_u8(a1, v0);
const uint8x8_t b1_v0 = vabd_u8(b1, v0);
const uint8x8_t b2_v0 = vabd_u8(b2, v0);
uint8x8_t max = vmax_u8(a2_v0, a1_v0);
max = vmax_u8(b1_v0, max);
max = vmax_u8(b2_v0, max);
return vclt_u8(max, filter);
}
static uint8x8_t generate_output(const uint8x8_t a2, const uint8x8_t a1,
const uint8x8_t v0, const uint8x8_t b1,
const uint8x8_t b2, const uint8x8_t filter) {
const uint8x8_t k_out = average_k_out(a2, a1, v0, b1, b2);
const uint8x8_t mask = generate_mask(a2, a1, v0, b1, b2, filter);
return vbsl_u8(mask, k_out, v0);
}
// Same functions but for uint8x16_t.
static uint8x16_t average_k_outq(const uint8x16_t a2, const uint8x16_t a1,
const uint8x16_t v0, const uint8x16_t b1,
const uint8x16_t b2) {
const uint8x16_t k1 = vrhaddq_u8(a2, a1);
const uint8x16_t k2 = vrhaddq_u8(b2, b1);
const uint8x16_t k3 = vrhaddq_u8(k1, k2);
return vrhaddq_u8(k3, v0);
}
static uint8x16_t generate_maskq(const uint8x16_t a2, const uint8x16_t a1,
const uint8x16_t v0, const uint8x16_t b1,
const uint8x16_t b2, const uint8x16_t filter) {
const uint8x16_t a2_v0 = vabdq_u8(a2, v0);
const uint8x16_t a1_v0 = vabdq_u8(a1, v0);
const uint8x16_t b1_v0 = vabdq_u8(b1, v0);
const uint8x16_t b2_v0 = vabdq_u8(b2, v0);
uint8x16_t max = vmaxq_u8(a2_v0, a1_v0);
max = vmaxq_u8(b1_v0, max);
max = vmaxq_u8(b2_v0, max);
return vcltq_u8(max, filter);
}
static uint8x16_t generate_outputq(const uint8x16_t a2, const uint8x16_t a1,
const uint8x16_t v0, const uint8x16_t b1,
const uint8x16_t b2,
const uint8x16_t filter) {
const uint8x16_t k_out = average_k_outq(a2, a1, v0, b1, b2);
const uint8x16_t mask = generate_maskq(a2, a1, v0, b1, b2, filter);
return vbslq_u8(mask, k_out, v0);
}
void vpx_post_proc_down_and_across_mb_row_neon(uint8_t *src_ptr,
uint8_t *dst_ptr, int src_stride,
int dst_stride, int cols,
uint8_t *f, int size) {
uint8_t *src, *dst;
int row;
int col;
// Process a stripe of macroblocks. The stripe will be a multiple of 16 (for
// Y) or 8 (for U/V) wide (cols) and the height (size) will be 16 (for Y) or 8
// (for U/V).
assert((size == 8 || size == 16) && cols % 8 == 0);
// While columns of length 16 can be processed, load them.
for (col = 0; col < cols - 8; col += 16) {
uint8x16_t a0, a1, a2, a3, a4, a5, a6, a7;
src = src_ptr - 2 * src_stride;
dst = dst_ptr;
a0 = vld1q_u8(src);
src += src_stride;
a1 = vld1q_u8(src);
src += src_stride;
a2 = vld1q_u8(src);
src += src_stride;
a3 = vld1q_u8(src);
src += src_stride;
for (row = 0; row < size; row += 4) {
uint8x16_t v_out_0, v_out_1, v_out_2, v_out_3;
const uint8x16_t filterq = vld1q_u8(f + col);
a4 = vld1q_u8(src);
src += src_stride;
a5 = vld1q_u8(src);
src += src_stride;
a6 = vld1q_u8(src);
src += src_stride;
a7 = vld1q_u8(src);
src += src_stride;
v_out_0 = generate_outputq(a0, a1, a2, a3, a4, filterq);
v_out_1 = generate_outputq(a1, a2, a3, a4, a5, filterq);
v_out_2 = generate_outputq(a2, a3, a4, a5, a6, filterq);
v_out_3 = generate_outputq(a3, a4, a5, a6, a7, filterq);
vst1q_u8(dst, v_out_0);
dst += dst_stride;
vst1q_u8(dst, v_out_1);
dst += dst_stride;
vst1q_u8(dst, v_out_2);
dst += dst_stride;
vst1q_u8(dst, v_out_3);
dst += dst_stride;
// Rotate over to the next slot.
a0 = a4;
a1 = a5;
a2 = a6;
a3 = a7;
}
src_ptr += 16;
dst_ptr += 16;
}
// Clean up any left over column of length 8.
if (col != cols) {
uint8x8_t a0, a1, a2, a3, a4, a5, a6, a7;
src = src_ptr - 2 * src_stride;
dst = dst_ptr;
a0 = vld1_u8(src);
src += src_stride;
a1 = vld1_u8(src);
src += src_stride;
a2 = vld1_u8(src);
src += src_stride;
a3 = vld1_u8(src);
src += src_stride;
for (row = 0; row < size; row += 4) {
uint8x8_t v_out_0, v_out_1, v_out_2, v_out_3;
const uint8x8_t filter = vld1_u8(f + col);
a4 = vld1_u8(src);
src += src_stride;
a5 = vld1_u8(src);
src += src_stride;
a6 = vld1_u8(src);
src += src_stride;
a7 = vld1_u8(src);
src += src_stride;
v_out_0 = generate_output(a0, a1, a2, a3, a4, filter);
v_out_1 = generate_output(a1, a2, a3, a4, a5, filter);
v_out_2 = generate_output(a2, a3, a4, a5, a6, filter);
v_out_3 = generate_output(a3, a4, a5, a6, a7, filter);
vst1_u8(dst, v_out_0);
dst += dst_stride;
vst1_u8(dst, v_out_1);
dst += dst_stride;
vst1_u8(dst, v_out_2);
dst += dst_stride;
vst1_u8(dst, v_out_3);
dst += dst_stride;
// Rotate over to the next slot.
a0 = a4;
a1 = a5;
a2 = a6;
a3 = a7;
}
// Not strictly necessary but makes resetting dst_ptr easier.
dst_ptr += 8;
}
dst_ptr -= cols;
for (row = 0; row < size; row += 8) {
uint8x8_t a0, a1, a2, a3;
uint8x8_t b0, b1, b2, b3, b4, b5, b6, b7;
src = dst_ptr;
dst = dst_ptr;
// Load 8 values, transpose 4 of them, and discard 2 because they will be
// reloaded later.
load_and_transpose_u8_4x8(src, dst_stride, &a0, &a1, &a2, &a3);
a3 = a1;
a2 = a1 = a0; // Extend left border.
src += 2;
for (col = 0; col < cols; col += 8) {
uint8x8_t v_out_0, v_out_1, v_out_2, v_out_3, v_out_4, v_out_5, v_out_6,
v_out_7;
// Although the filter is meant to be applied vertically and is instead
// being applied horizontally here it's OK because it's set in blocks of 8
// (or 16).
const uint8x8_t filter = vld1_u8(f + col);
load_and_transpose_u8_8x8(src, dst_stride, &b0, &b1, &b2, &b3, &b4, &b5,
&b6, &b7);
if (col + 8 == cols) {
// Last row. Extend border (b5).
b6 = b7 = b5;
}
v_out_0 = generate_output(a0, a1, a2, a3, b0, filter);
v_out_1 = generate_output(a1, a2, a3, b0, b1, filter);
v_out_2 = generate_output(a2, a3, b0, b1, b2, filter);
v_out_3 = generate_output(a3, b0, b1, b2, b3, filter);
v_out_4 = generate_output(b0, b1, b2, b3, b4, filter);
v_out_5 = generate_output(b1, b2, b3, b4, b5, filter);
v_out_6 = generate_output(b2, b3, b4, b5, b6, filter);
v_out_7 = generate_output(b3, b4, b5, b6, b7, filter);
transpose_and_store_u8_8x8(dst, dst_stride, v_out_0, v_out_1, v_out_2,
v_out_3, v_out_4, v_out_5, v_out_6, v_out_7);
a0 = b4;
a1 = b5;
a2 = b6;
a3 = b7;
src += 8;
dst += 8;
}
dst_ptr += 8 * dst_stride;
}
}
// sum += x;
// sumsq += x * y;
static void accumulate_sum_sumsq(const int16x4_t x, const int32x4_t xy,
int16x4_t *const sum, int32x4_t *const sumsq) {
const int16x4_t zero = vdup_n_s16(0);
const int32x4_t zeroq = vdupq_n_s32(0);
// Add in the first set because vext doesn't work with '0'.
*sum = vadd_s16(*sum, x);
*sumsq = vaddq_s32(*sumsq, xy);
// Shift x and xy to the right and sum. vext requires an immediate.
*sum = vadd_s16(*sum, vext_s16(zero, x, 1));
*sumsq = vaddq_s32(*sumsq, vextq_s32(zeroq, xy, 1));
*sum = vadd_s16(*sum, vext_s16(zero, x, 2));
*sumsq = vaddq_s32(*sumsq, vextq_s32(zeroq, xy, 2));
*sum = vadd_s16(*sum, vext_s16(zero, x, 3));
*sumsq = vaddq_s32(*sumsq, vextq_s32(zeroq, xy, 3));
}
// Generate mask based on (sumsq * 15 - sum * sum < flimit)
static uint16x4_t calculate_mask(const int16x4_t sum, const int32x4_t sumsq,
const int32x4_t f, const int32x4_t fifteen) {
const int32x4_t a = vmulq_s32(sumsq, fifteen);
const int32x4_t b = vmlsl_s16(a, sum, sum);
const uint32x4_t mask32 = vcltq_s32(b, f);
return vmovn_u32(mask32);
}
static uint8x8_t combine_mask(const int16x4_t sum_low, const int16x4_t sum_high,
const int32x4_t sumsq_low,
const int32x4_t sumsq_high, const int32x4_t f) {
const int32x4_t fifteen = vdupq_n_s32(15);
const uint16x4_t mask16_low = calculate_mask(sum_low, sumsq_low, f, fifteen);
const uint16x4_t mask16_high =
calculate_mask(sum_high, sumsq_high, f, fifteen);
return vmovn_u16(vcombine_u16(mask16_low, mask16_high));
}
// Apply filter of (8 + sum + s[c]) >> 4.
static uint8x8_t filter_pixels(const int16x8_t sum, const uint8x8_t s) {
const int16x8_t s16 = vreinterpretq_s16_u16(vmovl_u8(s));
const int16x8_t sum_s = vaddq_s16(sum, s16);
return vqrshrun_n_s16(sum_s, 4);
}
void vpx_mbpost_proc_across_ip_neon(uint8_t *src, int pitch, int rows, int cols,
int flimit) {
int row, col;
const int32x4_t f = vdupq_n_s32(flimit);
assert(cols % 8 == 0);
for (row = 0; row < rows; ++row) {
// Sum the first 8 elements, which are extended from s[0].
// sumsq gets primed with +16.
int sumsq = src[0] * src[0] * 9 + 16;
int sum = src[0] * 9;
uint8x8_t left_context, s, right_context;
int16x4_t sum_low, sum_high;
int32x4_t sumsq_low, sumsq_high;
// Sum (+square) the next 6 elements.
// Skip [0] because it's included above.
for (col = 1; col <= 6; ++col) {
sumsq += src[col] * src[col];
sum += src[col];
}
// Prime the sums. Later the loop uses the _high values to prime the new
// vectors.
sumsq_high = vdupq_n_s32(sumsq);
sum_high = vdup_n_s16(sum);
// Manually extend the left border.
left_context = vdup_n_u8(src[0]);
for (col = 0; col < cols; col += 8) {
uint8x8_t mask, output;
int16x8_t x, y;
int32x4_t xy_low, xy_high;
s = vld1_u8(src + col);
if (col + 8 == cols) {
// Last row. Extend border.
right_context = vdup_n_u8(src[col + 7]);
} else {
right_context = vld1_u8(src + col + 7);
}
x = vreinterpretq_s16_u16(vsubl_u8(right_context, left_context));
y = vreinterpretq_s16_u16(vaddl_u8(right_context, left_context));
xy_low = vmull_s16(vget_low_s16(x), vget_low_s16(y));
xy_high = vmull_s16(vget_high_s16(x), vget_high_s16(y));
// Catch up to the last sum'd value.
sum_low = vdup_lane_s16(sum_high, 3);
sumsq_low = vdupq_lane_s32(vget_high_s32(sumsq_high), 1);
accumulate_sum_sumsq(vget_low_s16(x), xy_low, &sum_low, &sumsq_low);
// Need to do this sequentially because we need the max value from
// sum_low.
sum_high = vdup_lane_s16(sum_low, 3);
sumsq_high = vdupq_lane_s32(vget_high_s32(sumsq_low), 1);
accumulate_sum_sumsq(vget_high_s16(x), xy_high, &sum_high, &sumsq_high);
mask = combine_mask(sum_low, sum_high, sumsq_low, sumsq_high, f);
output = filter_pixels(vcombine_s16(sum_low, sum_high), s);
output = vbsl_u8(mask, output, s);
vst1_u8(src + col, output);
left_context = s;
}
src += pitch;
}
}
// Apply filter of (vpx_rv + sum + s[c]) >> 4.
static uint8x8_t filter_pixels_rv(const int16x8_t sum, const uint8x8_t s,
const int16x8_t rv) {
const int16x8_t s16 = vreinterpretq_s16_u16(vmovl_u8(s));
const int16x8_t sum_s = vaddq_s16(sum, s16);
const int16x8_t rounded = vaddq_s16(sum_s, rv);
return vqshrun_n_s16(rounded, 4);
}
void vpx_mbpost_proc_down_neon(uint8_t *dst, int pitch, int rows, int cols,
int flimit) {
int row, col, i;
const int32x4_t f = vdupq_n_s32(flimit);
uint8x8_t below_context = vdup_n_u8(0);
// 8 columns are processed at a time.
// If rows is less than 8 the bottom border extension fails.
assert(cols % 8 == 0);
assert(rows >= 8);
// Load and keep the first 8 values in memory. Process a vertical stripe that
// is 8 wide.
for (col = 0; col < cols; col += 8) {
uint8x8_t s, above_context[8];
int16x8_t sum, sum_tmp;
int32x4_t sumsq_low, sumsq_high;
// Load and extend the top border.
s = vld1_u8(dst);
for (i = 0; i < 8; i++) {
above_context[i] = s;
}
sum_tmp = vreinterpretq_s16_u16(vmovl_u8(s));
// sum * 9
sum = vmulq_n_s16(sum_tmp, 9);
// (sum * 9) * sum == sum * sum * 9
sumsq_low = vmull_s16(vget_low_s16(sum), vget_low_s16(sum_tmp));
sumsq_high = vmull_s16(vget_high_s16(sum), vget_high_s16(sum_tmp));
// Load and discard the next 6 values to prime sum and sumsq.
for (i = 1; i <= 6; ++i) {
const uint8x8_t a = vld1_u8(dst + i * pitch);
const int16x8_t b = vreinterpretq_s16_u16(vmovl_u8(a));
sum = vaddq_s16(sum, b);
sumsq_low = vmlal_s16(sumsq_low, vget_low_s16(b), vget_low_s16(b));
sumsq_high = vmlal_s16(sumsq_high, vget_high_s16(b), vget_high_s16(b));
}
for (row = 0; row < rows; ++row) {
uint8x8_t mask, output;
int16x8_t x, y;
int32x4_t xy_low, xy_high;
s = vld1_u8(dst + row * pitch);
// Extend the bottom border.
if (row + 7 < rows) {
below_context = vld1_u8(dst + (row + 7) * pitch);
}
x = vreinterpretq_s16_u16(vsubl_u8(below_context, above_context[0]));
y = vreinterpretq_s16_u16(vaddl_u8(below_context, above_context[0]));
xy_low = vmull_s16(vget_low_s16(x), vget_low_s16(y));
xy_high = vmull_s16(vget_high_s16(x), vget_high_s16(y));
sum = vaddq_s16(sum, x);
sumsq_low = vaddq_s32(sumsq_low, xy_low);
sumsq_high = vaddq_s32(sumsq_high, xy_high);
mask = combine_mask(vget_low_s16(sum), vget_high_s16(sum), sumsq_low,
sumsq_high, f);
output = filter_pixels_rv(sum, s, vld1q_s16(vpx_rv + (row & 127)));
output = vbsl_u8(mask, output, s);
vst1_u8(dst + row * pitch, output);
above_context[0] = above_context[1];
above_context[1] = above_context[2];
above_context[2] = above_context[3];
above_context[3] = above_context[4];
above_context[4] = above_context[5];
above_context[5] = above_context[6];
above_context[6] = above_context[7];
above_context[7] = s;
}
dst += 8;
}
}