Source code
Revision control
Copy as Markdown
Other Tools
/*
* Copyright (c) 2024, Alliance for Open Media. All rights reserved.
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#include <immintrin.h>
#include "config/av1_rtcd.h"
#include "av1/common/resize.h"
#include "aom_dsp/x86/synonyms.h"
#define ROW_OFFSET 5
#define PROCESS_RESIZE_Y_WD8 \
/* ah0 ah1 ... ah7 */ \
const __m128i AH = _mm_add_epi16(l0, l7); \
/* bg0 bg1 ... bh7 */ \
const __m128i BG = _mm_add_epi16(l1, l6); \
/* cf0 cf1 ... cf7 */ \
const __m128i CF = _mm_add_epi16(l2, l5); \
/* de0 de1 ... de7 */ \
const __m128i DE = _mm_add_epi16(l3, l4); \
\
/* ah0 bg0 ... ah3 bg3 */ \
const __m128i AHBG_low = _mm_unpacklo_epi16(AH, BG); \
/*cf0 de0 ... cf2 de2 */ \
const __m128i CFDE_low = _mm_unpacklo_epi16(CF, DE); \
\
/* ah4 bg4... ah7 bg7 */ \
const __m128i AHBG_hi = _mm_unpackhi_epi16(AH, BG); \
/* cf4 de4... cf7 de7 */ \
const __m128i CFDE_hi = _mm_unpackhi_epi16(CF, DE); \
\
/* r00 r01 r02 r03 */ \
const __m128i r00 = _mm_madd_epi16(AHBG_low, coeffs_y[0]); \
const __m128i r01 = _mm_madd_epi16(CFDE_low, coeffs_y[1]); \
__m128i r0 = _mm_add_epi32(r00, r01); \
/* r04 r05 r06 r07 */ \
const __m128i r10 = _mm_madd_epi16(AHBG_hi, coeffs_y[0]); \
const __m128i r11 = _mm_madd_epi16(CFDE_hi, coeffs_y[1]); \
__m128i r1 = _mm_add_epi32(r10, r11); \
\
r0 = _mm_add_epi32(r0, round_const_bits); \
r1 = _mm_add_epi32(r1, round_const_bits); \
r0 = _mm_sra_epi32(r0, round_shift_bits); \
r1 = _mm_sra_epi32(r1, round_shift_bits); \
\
/* r00 ... r07 (8 values of each 16bit) */ \
const __m128i res_16b = _mm_packs_epi32(r0, r1); \
/* r00 ... r07 | r00 ... r07 (16 values of each 8bit) */ \
const __m128i res_8b0 = _mm_packus_epi16(res_16b, res_16b); \
\
__m128i res = _mm_min_epu8(res_8b0, clip_pixel); \
res = _mm_max_epu8(res, zero); \
_mm_storel_epi64((__m128i *)&output[(i / 2) * out_stride + j], res); \
\
l0 = l2; \
l1 = l3; \
l2 = l4; \
l3 = l5; \
l4 = l6; \
l5 = l7; \
data += 2 * stride;
static inline void prepare_filter_coeffs(const int16_t *filter,
__m128i *const coeffs /* [2] */) {
// f0 f1 f2 f3 x x x x
const __m128i sym_even_filter = _mm_loadl_epi64((__m128i *)filter);
// f1 f0 f3 f2 x x x x
const __m128i tmp1 = _mm_shufflelo_epi16(sym_even_filter, 0xb1);
// f3 f2 f3 f2 ...
coeffs[0] = _mm_shuffle_epi32(tmp1, 0x55);
// f1 f0 f1 f0 ...
coeffs[1] = _mm_shuffle_epi32(tmp1, 0x00);
}
bool av1_resize_vert_dir_sse2(uint8_t *intbuf, uint8_t *output, int out_stride,
int height, int height2, int stride,
int start_col) {
// For the GM tool, the input layer height or width is assured to be an even
// number. Hence the function 'down2_symodd()' is not invoked and SIMD
// optimization of the same is not implemented.
// When the input height is less than 8 and even, the potential input
// heights are limited to 2, 4, or 6. These scenarios require seperate
// handling due to padding requirements. Invoking the C function here will
// eliminate the need for conditional statements within the subsequent SIMD
// code to manage these cases.
if (height & 1 || height < 8) {
return av1_resize_vert_dir_c(intbuf, output, out_stride, height, height2,
stride, start_col);
}
__m128i coeffs_y[2];
const int bits = FILTER_BITS;
const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1);
const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
const uint8_t max_pixel = 255;
const __m128i clip_pixel = _mm_set1_epi8((char)max_pixel);
const __m128i zero = _mm_setzero_si128();
prepare_filter_coeffs(av1_down2_symeven_half_filter, coeffs_y);
const int remain_col = stride % 8;
for (int j = start_col; j < stride - remain_col; j += 8) {
uint8_t *data = &intbuf[j];
// d0 ... d7
const __m128i l8_3 = _mm_loadl_epi64((__m128i *)(data + 0 * stride));
// Padding top 3 rows with the last available row at the top.
// a0 ... a7
const __m128i l8_0 = l8_3;
// b0 ... b7
const __m128i l8_1 = l8_3;
// c0 ... c7
const __m128i l8_2 = l8_3;
// e0 ... e7
const __m128i l8_4 = _mm_loadl_epi64((__m128i *)(data + 1 * stride));
// f0 ... f7
const __m128i l8_5 = _mm_loadl_epi64((__m128i *)(data + 2 * stride));
// Convert to 16bit as addition of 2 source pixel crosses 8 bit.
__m128i l0 = _mm_unpacklo_epi8(l8_0, zero); // A(128bit) = a0 - a7(16 bit)
__m128i l1 = _mm_unpacklo_epi8(l8_1, zero); // B(128bit) = b0 - b7(16 bit)
__m128i l2 = _mm_unpacklo_epi8(l8_2, zero); // C(128bit) = c0 - c7(16 bit)
__m128i l3 = _mm_unpacklo_epi8(l8_3, zero); // D(128bit) = d0 - d7(16 bit)
__m128i l4 = _mm_unpacklo_epi8(l8_4, zero); // E(128bit) = e0 - e7(16 bit)
__m128i l5 = _mm_unpacklo_epi8(l8_5, zero); // F(128bit) = f0 - f7(16 bit)
// Increment the pointer such that the loading starts from row G.
data = data + 3 * stride;
// The core vertical SIMD processes 2 input rows simultaneously to generate
// output corresponding to 1 row. To streamline the core loop and eliminate
// the need for conditional checks, the remaining rows 4 are processed
// separately.
for (int i = 0; i < height - 4; i += 2) {
// g0 ... g7
__m128i l8_6 = _mm_loadl_epi64((__m128i *)(data));
// h0 ... h7
__m128i l8_7 = _mm_loadl_epi64((__m128i *)(data + stride));
__m128i l6 = _mm_unpacklo_epi8(l8_6, zero); // G(128bit):g0-g7(16b)
__m128i l7 = _mm_unpacklo_epi8(l8_7, zero); // H(128bit):h0-h7(16b)
PROCESS_RESIZE_Y_WD8
}
__m128i l8_6 = _mm_loadl_epi64((__m128i *)(data));
__m128i l6 = _mm_unpacklo_epi8(l8_6, zero);
// Process the last 4 input rows here.
for (int i = height - 4; i < height; i += 2) {
__m128i l7 = l6;
PROCESS_RESIZE_Y_WD8
}
}
if (remain_col)
return av1_resize_vert_dir_c(intbuf, output, out_stride, height, height2,
stride, stride - remain_col);
return true;
}
// Blends a and b using mask and returns the result.
static inline __m128i blend(__m128i a, __m128i b, __m128i mask) {
const __m128i masked_b = _mm_and_si128(mask, b);
const __m128i masked_a = _mm_andnot_si128(mask, a);
return (_mm_or_si128(masked_a, masked_b));
}
// Masks used for width 16 pixels, with left and right padding
// requirements.
static const uint8_t left_padding_mask[16] = {
255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
};
static const uint8_t right_padding_mask[16] = { 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 255, 255,
255, 255, 255, 255 };
static const uint8_t mask_16[16] = {
255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0,
};
void av1_resize_horz_dir_sse2(const uint8_t *const input, int in_stride,
uint8_t *intbuf, int height, int filtered_length,
int width2) {
assert(height % 2 == 0);
// Invoke C for width less than 16.
if (filtered_length < 16) {
av1_resize_horz_dir_c(input, in_stride, intbuf, height, filtered_length,
width2);
return;
}
__m128i coeffs_x[2];
const int bits = FILTER_BITS;
const int dst_stride = width2;
const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1);
const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
const uint8_t max_pixel = 255;
const __m128i clip_pixel = _mm_set1_epi8((char)max_pixel);
const __m128i zero = _mm_setzero_si128();
const __m128i start_pad_mask = _mm_loadu_si128((__m128i *)left_padding_mask);
const __m128i end_pad_mask = _mm_loadu_si128((__m128i *)right_padding_mask);
const __m128i mask_even = _mm_loadu_si128((__m128i *)mask_16);
prepare_filter_coeffs(av1_down2_symeven_half_filter, coeffs_x);
for (int i = 0; i < height; ++i) {
int filter_offset = 0;
int row01_offset = ROW_OFFSET;
int remain_col = filtered_length;
// To avoid pixel over-read at frame boundary, processing of 16 pixels
// is done using the core loop only if sufficient number of pixels required
// for the load are present.The remaining pixels are processed separately.
for (int j = 0; j <= filtered_length - 16; j += 16) {
if (remain_col == 18 || remain_col == 20) {
break;
}
const int is_last_cols16 = (j == filtered_length - 16);
// While processing the last 16 pixels of the row, ensure that only valid
// pixels are loaded.
if (is_last_cols16) row01_offset = 0;
const int in_idx = i * in_stride + j - filter_offset;
const int out_idx = i * dst_stride + j / 2;
remain_col -= 16;
// a0 a1 a2 a3 .... a15
__m128i row00 = _mm_loadu_si128((__m128i *)&input[in_idx]);
// a8 a9 a10 a11 .... a23
__m128i row01 = _mm_loadu_si128(
(__m128i *)&input[in_idx + row01_offset + filter_offset]);
filter_offset = 3;
// Pad start pixels to the left, while processing the first pixels in the
// row.
if (j == 0) {
const __m128i start_pixel_row0 =
_mm_set1_epi8((char)input[i * in_stride]);
row00 =
blend(_mm_slli_si128(row00, 3), start_pixel_row0, start_pad_mask);
}
// Pad end pixels to the right, while processing the last pixels in the
// row.
if (is_last_cols16) {
const __m128i end_pixel_row0 =
_mm_set1_epi8((char)input[i * in_stride + filtered_length - 1]);
row01 = blend(_mm_srli_si128(row01, ROW_OFFSET), end_pixel_row0,
end_pad_mask);
}
// a2 a3 a4 a5 a6 a7 a8 a9 .... a17
const __m128i row0_1 = _mm_unpacklo_epi64(_mm_srli_si128(row00, 2),
_mm_srli_si128(row01, 2));
// a4 a5 a6 a7 a9 10 a11 a12 .... a19
const __m128i row0_2 = _mm_unpacklo_epi64(_mm_srli_si128(row00, 4),
_mm_srli_si128(row01, 4));
// a6 a7 a8 a9 a10 a11 a12 a13 .... a21
const __m128i row0_3 = _mm_unpacklo_epi64(_mm_srli_si128(row00, 6),
_mm_srli_si128(row01, 6));
// a0 a2 a4 a6 a8 a10 a12 a14 (each 16 bit)
const __m128i s0 = _mm_and_si128(row00, mask_even);
// a1 a3 a5 a7 a9 a11 a13 a15
const __m128i s1 = _mm_and_si128(_mm_srli_epi16(row00, 8), mask_even);
// a2 a4 a6 a8 a10 a12 a14 a16
const __m128i s2 = _mm_and_si128(row0_1, mask_even);
// a3 a5 a7 a9 a11 a13 a15 a17
const __m128i s3 = _mm_and_si128(_mm_srli_epi16(row0_1, 8), mask_even);
// a4 a6 a8 a10 a12 a14 a16 a18
const __m128i s4 = _mm_and_si128(row0_2, mask_even);
// a5 a7 a9 a11 a13 a15 a17 a19
const __m128i s5 = _mm_and_si128(_mm_srli_epi16(row0_2, 8), mask_even);
// a6 a8 a10 a12 a14 a16 a18 a20
const __m128i s6 = _mm_and_si128(row0_3, mask_even);
// a7 a9 a11 a13 a15 a17 a19 a21
const __m128i s7 = _mm_and_si128(_mm_srli_epi16(row0_3, 8), mask_even);
// a0a7 a2a9 a4a11 .... a12a19 a14a21
const __m128i s07 = _mm_add_epi16(s0, s7);
// a1a6 a3a8 a5a10 .... a13a18 a15a20
const __m128i s16 = _mm_add_epi16(s1, s6);
// a2a5 a4a7 a6a9 .... a14a17 a16a19
const __m128i s25 = _mm_add_epi16(s2, s5);
// a3a4 a5a6 a7a8 .... a15a16 a17a18
const __m128i s34 = _mm_add_epi16(s3, s4);
// a0a7 a1a6 a2a9 a3a8 a4a11 a5a10 a6a13 a7a12
const __m128i s1607_low = _mm_unpacklo_epi16(s07, s16);
// a2a5 a3a4 a4a7 a5a6 a6a9 a7a8 a8a11 a9a10
const __m128i s3425_low = _mm_unpacklo_epi16(s25, s34);
// a8a15 a9a14 a10a17 a11a16 a12a19 a13a18 a14a21 a15a20
const __m128i s1607_high = _mm_unpackhi_epi16(s07, s16);
// a10a13 a11a12 a12a15 a13a14 a14a17 a15a16 a16a19 a17a18
const __m128i s3425_high = _mm_unpackhi_epi16(s25, s34);
const __m128i r01_0 = _mm_madd_epi16(s3425_low, coeffs_x[1]);
const __m128i r01_1 = _mm_madd_epi16(s1607_low, coeffs_x[0]);
const __m128i r01_2 = _mm_madd_epi16(s3425_high, coeffs_x[1]);
const __m128i r01_3 = _mm_madd_epi16(s1607_high, coeffs_x[0]);
// Result of first 8 pixels of row0 (a0 to a7).
// r0_0 r0_1 r0_2 r0_3
__m128i r00 = _mm_add_epi32(r01_0, r01_1);
r00 = _mm_add_epi32(r00, round_const_bits);
r00 = _mm_sra_epi32(r00, round_shift_bits);
// Result of next 8 pixels of row0 (a8 to 15).
// r0_4 r0_5 r0_6 r0_7
__m128i r01 = _mm_add_epi32(r01_2, r01_3);
r01 = _mm_add_epi32(r01, round_const_bits);
r01 = _mm_sra_epi32(r01, round_shift_bits);
// r0_0 r0_1 r1_2 r0_3 r0_4 r0_5 r0_6 r0_7
const __m128i res_16 = _mm_packs_epi32(r00, r01);
const __m128i res_8 = _mm_packus_epi16(res_16, res_16);
__m128i res = _mm_min_epu8(res_8, clip_pixel);
res = _mm_max_epu8(res, zero);
// r0_0 r0_1 r1_2 r0_3 r0_4 r0_5 r0_6 r0_7
_mm_storel_epi64((__m128i *)&intbuf[out_idx], res);
}
int wd_processed = filtered_length - remain_col;
if (remain_col) {
const int in_idx = (in_stride * i);
const int out_idx = (wd_processed / 2) + width2 * i;
down2_symeven(input + in_idx, filtered_length, intbuf + out_idx,
wd_processed);
}
}
}