blob: 273712a9d49337ea0e1d9c07fca9653c34ee4837 [file] [log] [blame]
/*
* Copyright (c) 2020, 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 <arm_neon.h>
#include <assert.h>
#include "av1/common/av1_txfm.h"
#include "av1/encoder/av1_fwd_txfm1d_cfg.h"
#include "aom_dsp/txfm_common.h"
#include "aom_ports/mem.h"
#include "config/av1_rtcd.h"
#include "config/aom_config.h"
static INLINE int32x4_t half_btf_neon(const int32_t *w0, const int32x4_t *n0,
const int32_t *w1, const int32x4_t *n1,
const int32x4_t v_bit) {
int32x4_t x;
x = vmulq_n_s32(*n0, *w0);
x = vmlaq_n_s32(x, *n1, *w1);
x = vrshlq_s32(x, v_bit);
return x;
}
static INLINE int32x4_t half_btf_neon_m(const int32_t *w0, const int32x4_t *n0,
const int32_t *w1, const int32x4_t *n1,
const int32x4_t v_bit) {
int32x4_t x;
x = vmulq_n_s32(*n0, *w0);
x = vmlsq_n_s32(x, *n1, *w1);
x = vrshlq_s32(x, v_bit);
return x;
}
#if defined(__aarch64__)
#define TRANSPOSE_4X4(x0, x1, x2, x3, y0, y1, y2, y3) \
do { \
int32x4x2_t swap_low = vtrnq_s32(x0, x1); \
int32x4x2_t swap_high = vtrnq_s32(x2, x3); \
y0 = vreinterpretq_s32_s64( \
vzip1q_s64(vreinterpretq_s64_s32(swap_low.val[0]), \
vreinterpretq_s64_s32(swap_high.val[0]))); \
y1 = vreinterpretq_s32_s64( \
vzip1q_s64(vreinterpretq_s64_s32(swap_low.val[1]), \
vreinterpretq_s64_s32(swap_high.val[1]))); \
y2 = vreinterpretq_s32_s64( \
vzip2q_s64(vreinterpretq_s64_s32(swap_low.val[0]), \
vreinterpretq_s64_s32(swap_high.val[0]))); \
y3 = vreinterpretq_s32_s64( \
vzip2q_s64(vreinterpretq_s64_s32(swap_low.val[1]), \
vreinterpretq_s64_s32(swap_high.val[1]))); \
} while (0)
#else
#define TRANSPOSE_4X4(x0, x1, x2, x3, y0, y1, y2, y3) \
do { \
int32x4x2_t swap_low = vtrnq_s32(x0, x1); \
int32x4x2_t swap_high = vtrnq_s32(x2, x3); \
y0 = vextq_s32(vextq_s32(swap_low.val[0], swap_low.val[0], 2), \
swap_high.val[0], 2); \
y1 = vextq_s32(vextq_s32(swap_low.val[1], swap_low.val[1], 2), \
swap_high.val[1], 2); \
y2 = vextq_s32(swap_low.val[0], \
vextq_s32(swap_high.val[0], swap_high.val[0], 2), 2); \
y3 = vextq_s32(swap_low.val[1], \
vextq_s32(swap_high.val[1], swap_high.val[1], 2), 2); \
} while (0)
#endif // (__aarch64__)
static INLINE void transpose_8x8(const int32x4_t *in, int32x4_t *out) {
TRANSPOSE_4X4(in[0], in[2], in[4], in[6], out[0], out[2], out[4], out[6]);
TRANSPOSE_4X4(in[1], in[3], in[5], in[7], out[8], out[10], out[12], out[14]);
TRANSPOSE_4X4(in[8], in[10], in[12], in[14], out[1], out[3], out[5], out[7]);
TRANSPOSE_4X4(in[9], in[11], in[13], in[15], out[9], out[11], out[13],
out[15]);
}
static INLINE void transpose_16x16(const int32x4_t *in, int32x4_t *out) {
// Upper left 8x8
TRANSPOSE_4X4(in[0], in[4], in[8], in[12], out[0], out[4], out[8], out[12]);
TRANSPOSE_4X4(in[1], in[5], in[9], in[13], out[16], out[20], out[24],
out[28]);
TRANSPOSE_4X4(in[16], in[20], in[24], in[28], out[1], out[5], out[9],
out[13]);
TRANSPOSE_4X4(in[17], in[21], in[25], in[29], out[17], out[21], out[25],
out[29]);
// Upper right 8x8
TRANSPOSE_4X4(in[2], in[6], in[10], in[14], out[32], out[36], out[40],
out[44]);
TRANSPOSE_4X4(in[3], in[7], in[11], in[15], out[48], out[52], out[56],
out[60]);
TRANSPOSE_4X4(in[18], in[22], in[26], in[30], out[33], out[37], out[41],
out[45]);
TRANSPOSE_4X4(in[19], in[23], in[27], in[31], out[49], out[53], out[57],
out[61]);
// Lower left 8x8
TRANSPOSE_4X4(in[32], in[36], in[40], in[44], out[2], out[6], out[10],
out[14]);
TRANSPOSE_4X4(in[33], in[37], in[41], in[45], out[18], out[22], out[26],
out[30]);
TRANSPOSE_4X4(in[48], in[52], in[56], in[60], out[3], out[7], out[11],
out[15]);
TRANSPOSE_4X4(in[49], in[53], in[57], in[61], out[19], out[23], out[27],
out[31]);
// Lower right 8x8
TRANSPOSE_4X4(in[34], in[38], in[42], in[46], out[34], out[38], out[42],
out[46]);
TRANSPOSE_4X4(in[35], in[39], in[43], in[47], out[50], out[54], out[58],
out[62]);
TRANSPOSE_4X4(in[50], in[54], in[58], in[62], out[35], out[39], out[43],
out[47]);
TRANSPOSE_4X4(in[51], in[55], in[59], in[63], out[51], out[55], out[59],
out[63]);
}
static INLINE void av1_round_shift_rect_array_32_neon(int32x4_t *input,
int32x4_t *output,
const int size,
const int bit,
const int val) {
const int32x4_t sqrt2 = vdupq_n_s32(val);
const int32x4_t v_bit = vdupq_n_s32(-bit);
int i;
for (i = 0; i < size; i++) {
const int32x4_t r0 = vrshlq_s32(input[i], v_bit);
const int32x4_t r1 = vmulq_s32(sqrt2, r0);
output[i] = vrshrq_n_s32(r1, NewSqrt2Bits);
}
}
#define btf_32_neon_type0(w0, w1, in0, in1, out0, out1, v_cos_bit) \
do { \
out0 = vmulq_n_s32(in0, w0); \
out0 = vmlaq_n_s32(out0, in1, w1); \
out0 = vrshlq_s32(out0, v_cos_bit); \
out1 = vmulq_n_s32(in0, w1); \
out1 = vmlsq_n_s32(out1, in1, w0); \
out1 = vrshlq_s32(out1, v_cos_bit); \
} while (0)
#define btf_32_neon_type1(w0, w1, in0, in1, out0, out1, bit) \
do { \
btf_32_neon_type0(w1, w0, in1, in0, out0, out1, bit); \
} while (0)
static INLINE void load_buffer_4x4(const int16_t *input, int32x4_t *in,
int stride, int flipud, int fliplr,
const int32x4_t *v_shift) {
int16x4_t v0, v1, v2, v3;
if (!flipud) {
v0 = vld1_s16(input + 0 * stride);
v1 = vld1_s16(input + 1 * stride);
v2 = vld1_s16(input + 2 * stride);
v3 = vld1_s16(input + 3 * stride);
} else {
v0 = vld1_s16(input + 3 * stride);
v1 = vld1_s16(input + 2 * stride);
v2 = vld1_s16(input + 1 * stride);
v3 = vld1_s16(input + 0 * stride);
}
if (fliplr) {
v0 = vrev64_s16(v0);
v1 = vrev64_s16(v1);
v2 = vrev64_s16(v2);
v3 = vrev64_s16(v3);
}
in[0] = vshlq_s32(vmovl_s16(v0), *v_shift);
in[1] = vshlq_s32(vmovl_s16(v1), *v_shift);
in[2] = vshlq_s32(vmovl_s16(v2), *v_shift);
in[3] = vshlq_s32(vmovl_s16(v3), *v_shift);
}
static void fdct4x4_neon(int32x4_t *in, int32x4_t *out, int bit,
const int num_col) {
const int32_t *cospi = cospi_arr(bit);
const int32x4_t cospi32 = vdupq_n_s32(cospi[32]);
const int32x4_t cospi48 = vdupq_n_s32(cospi[48]);
const int32x4_t cospi16 = vdupq_n_s32(cospi[16]);
int32x4_t s0, s1, s2, s3;
int32x4_t u0, u1, u2, u3;
int32x4_t v0, v2;
int endidx = 3 * num_col;
s0 = vaddq_s32(in[0], in[endidx]);
s3 = vsubq_s32(in[0], in[endidx]);
endidx -= num_col;
s1 = vaddq_s32(in[num_col], in[endidx]);
s2 = vsubq_s32(in[num_col], in[endidx]);
u0 = vmulq_s32(s0, cospi32);
u1 = vmulq_s32(s1, cospi32);
u2 = vaddq_s32(u0, u1);
v0 = vsubq_s32(u0, u1);
const int32x4_t v_bit = vdupq_n_s32(-bit);
u0 = vrshlq_s32(u2, v_bit);
u2 = vrshlq_s32(v0, v_bit);
v0 = vmulq_s32(s2, cospi48);
v2 = vmlaq_s32(v0, s3, cospi16);
u1 = vrshlq_s32(v2, v_bit);
v0 = vmulq_s32(s3, cospi48);
v2 = vmlsq_s32(v0, s2, cospi16);
u3 = vrshlq_s32(v2, v_bit);
TRANSPOSE_4X4(u0, u1, u2, u3, out[0], out[1], out[2], out[3]);
}
static INLINE void write_buffer_4x4(int32x4_t *res, int32_t *output) {
vst1q_s32((output + 0 * 4), res[0]);
vst1q_s32((output + 1 * 4), res[1]);
vst1q_s32((output + 2 * 4), res[2]);
vst1q_s32((output + 3 * 4), res[3]);
}
static void fadst4x4_neon(int32x4_t *in, int32x4_t *out, int bit,
const int num_col) {
const int32_t *sinpi = sinpi_arr(bit);
const int32x4_t sinpi4x = vld1q_s32(&sinpi[1]);
const int32x4_t sinpi1 = vdupq_lane_s32(vget_low_s32(sinpi4x), 0);
const int32x4_t sinpi2 = vdupq_lane_s32(vget_low_s32(sinpi4x), 1);
const int32x4_t sinpi3 = vdupq_lane_s32(vget_high_s32(sinpi4x), 0);
const int32x4_t sinpi4 = vdupq_lane_s32(vget_high_s32(sinpi4x), 1);
int32x4_t t;
int32x4_t s0, s1, s2, s3, s7;
int32x4_t x0, x1, x2, x3;
int32x4_t u0, u1, u2, u3;
int idx = 0 * num_col;
s0 = vmulq_s32(in[idx], sinpi1);
s1 = vmulq_s32(in[idx], sinpi4);
t = vaddq_s32(in[idx], in[idx + num_col]);
idx += 2 * num_col;
x3 = vmulq_s32(in[idx], sinpi3);
idx += num_col;
s7 = vsubq_s32(t, in[idx]);
t = vmlaq_s32(s0, in[idx - 2 * num_col], sinpi2);
x0 = vmlaq_s32(t, in[idx], sinpi4);
x1 = vmulq_s32(s7, sinpi3);
t = vmlsq_s32(s1, in[idx - 2 * num_col], sinpi1);
x2 = vmlaq_s32(t, in[idx], sinpi2);
s0 = vaddq_s32(x0, x3);
s1 = x1;
s2 = vsubq_s32(x2, x3);
t = vsubq_s32(x2, x0);
s3 = vaddq_s32(t, x3);
const int32x4_t v_bit = vdupq_n_s32(-bit);
u0 = vrshlq_s32(s0, v_bit);
u1 = vrshlq_s32(s1, v_bit);
u2 = vrshlq_s32(s2, v_bit);
u3 = vrshlq_s32(s3, v_bit);
TRANSPOSE_4X4(u0, u1, u2, u3, out[0], out[1], out[2], out[3]);
}
static void idtx4x4_neon(int32x4_t *in, int32x4_t *out, int bit, int col_num) {
(void)bit;
int32x4_t fact = vdupq_n_s32(NewSqrt2);
int32x4_t a_low;
int i;
for (i = 0; i < 4; i++) {
a_low = vmulq_s32(in[i * col_num], fact);
out[i] = vrshrq_n_s32(a_low, NewSqrt2Bits);
}
TRANSPOSE_4X4(out[0], out[1], out[2], out[3], out[0], out[1], out[2], out[3]);
}
void av1_fwd_txfm2d_4x4_neon(const int16_t *input, int32_t *coeff,
int input_stride, TX_TYPE tx_type, int bd) {
int32x4_t in[4];
const int8_t *shift = av1_fwd_txfm_shift_ls[TX_4X4];
const int txw_idx = get_txw_idx(TX_4X4);
const int txh_idx = get_txh_idx(TX_4X4);
int32x4_t v_shift0 = vdupq_n_s32(shift[0]);
switch (tx_type) {
case DCT_DCT:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
fdct4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fdct4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case ADST_DCT:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
fadst4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fdct4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case DCT_ADST:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
fdct4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fadst4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case ADST_ADST:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
fadst4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fadst4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case FLIPADST_DCT:
load_buffer_4x4(input, in, input_stride, 1, 0, &v_shift0);
fadst4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fdct4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case DCT_FLIPADST:
load_buffer_4x4(input, in, input_stride, 0, 1, &v_shift0);
fdct4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fadst4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case FLIPADST_FLIPADST:
load_buffer_4x4(input, in, input_stride, 1, 1, &v_shift0);
fadst4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fadst4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case ADST_FLIPADST:
load_buffer_4x4(input, in, input_stride, 0, 1, &v_shift0);
fadst4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fadst4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case FLIPADST_ADST:
load_buffer_4x4(input, in, input_stride, 1, 0, &v_shift0);
fadst4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
fadst4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case IDTX:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
idtx4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
idtx4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case V_DCT:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
fdct4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
idtx4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case H_DCT:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
idtx4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
fdct4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case V_ADST:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
fadst4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
idtx4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case H_ADST:
load_buffer_4x4(input, in, input_stride, 0, 0, &v_shift0);
idtx4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
fadst4x4_neon(in, in, av1_fwd_cos_bit_col[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case V_FLIPADST:
load_buffer_4x4(input, in, input_stride, 1, 0, &v_shift0);
fadst4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
idtx4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
case H_FLIPADST:
load_buffer_4x4(input, in, input_stride, 0, 1, &v_shift0);
idtx4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
fadst4x4_neon(in, in, av1_fwd_cos_bit_row[txw_idx][txh_idx], 1);
write_buffer_4x4(in, coeff);
break;
default: assert(0);
}
(void)bd;
}
static INLINE void load_buffer_8x8(const int16_t *input, int32x4_t *in,
int stride, int flipud, int fliplr,
const int shift) {
if (!flipud) {
in[0] = vreinterpretq_s32_s16(vld1q_s16((input + 0 * stride)));
in[1] = vreinterpretq_s32_s16(vld1q_s16((input + 1 * stride)));
in[2] = vreinterpretq_s32_s16(vld1q_s16((input + 2 * stride)));
in[3] = vreinterpretq_s32_s16(vld1q_s16((input + 3 * stride)));
in[4] = vreinterpretq_s32_s16(vld1q_s16((input + 4 * stride)));
in[5] = vreinterpretq_s32_s16(vld1q_s16((input + 5 * stride)));
in[6] = vreinterpretq_s32_s16(vld1q_s16((input + 6 * stride)));
in[7] = vreinterpretq_s32_s16(vld1q_s16((input + 7 * stride)));
} else {
in[0] = vreinterpretq_s32_s16(vld1q_s16((input + 7 * stride)));
in[1] = vreinterpretq_s32_s16(vld1q_s16((input + 6 * stride)));
in[2] = vreinterpretq_s32_s16(vld1q_s16((input + 5 * stride)));
in[3] = vreinterpretq_s32_s16(vld1q_s16((input + 4 * stride)));
in[4] = vreinterpretq_s32_s16(vld1q_s16((input + 3 * stride)));
in[5] = vreinterpretq_s32_s16(vld1q_s16((input + 2 * stride)));
in[6] = vreinterpretq_s32_s16(vld1q_s16((input + 1 * stride)));
in[7] = vreinterpretq_s32_s16(vld1q_s16((input + 0 * stride)));
}
if (fliplr) {
in[0] = vreinterpretq_s32_s16(vrev64q_s16(vreinterpretq_s16_s32(in[0])));
in[0] = vextq_s32(in[0], in[0], 2);
in[1] = vreinterpretq_s32_s16(vrev64q_s16(vreinterpretq_s16_s32(in[1])));
in[1] = vextq_s32(in[1], in[1], 2);
in[2] = vreinterpretq_s32_s16(vrev64q_s16(vreinterpretq_s16_s32(in[2])));
in[2] = vextq_s32(in[2], in[2], 2);
in[3] = vreinterpretq_s32_s16(vrev64q_s16(vreinterpretq_s16_s32(in[3])));
in[3] = vextq_s32(in[3], in[3], 2);
in[4] = vreinterpretq_s32_s16(vrev64q_s16(vreinterpretq_s16_s32(in[4])));
in[4] = vextq_s32(in[4], in[4], 2);
in[5] = vreinterpretq_s32_s16(vrev64q_s16(vreinterpretq_s16_s32(in[5])));
in[5] = vextq_s32(in[5], in[5], 2);
in[6] = vreinterpretq_s32_s16(vrev64q_s16(vreinterpretq_s16_s32(in[6])));
in[6] = vextq_s32(in[6], in[6], 2);
in[7] = vreinterpretq_s32_s16(vrev64q_s16(vreinterpretq_s16_s32(in[7])));
in[7] = vextq_s32(in[7], in[7], 2);
}
int16x4_t u = vget_high_s16(vreinterpretq_s16_s32(in[4]));
in[8] = vmovl_s16(vget_low_s16(vreinterpretq_s16_s32(in[4])));
in[9] = vmovl_s16(u);
u = vget_high_s16(vreinterpretq_s16_s32(in[5]));
in[10] = vmovl_s16(vget_low_s16(vreinterpretq_s16_s32(in[5])));
in[11] = vmovl_s16(u);
u = vget_high_s16(vreinterpretq_s16_s32(in[6]));
in[12] = vmovl_s16(vget_low_s16(vreinterpretq_s16_s32(in[6])));
in[13] = vmovl_s16(u);
u = vget_high_s16(vreinterpretq_s16_s32(in[7]));
in[14] = vmovl_s16(vget_low_s16(vreinterpretq_s16_s32(in[7])));
in[15] = vmovl_s16(u);
u = vget_high_s16(vreinterpretq_s16_s32(in[3]));
in[6] = vmovl_s16(vget_low_s16(vreinterpretq_s16_s32(in[3])));
in[7] = vmovl_s16(u);
u = vget_high_s16(vreinterpretq_s16_s32(in[2]));
in[4] = vmovl_s16(vget_low_s16(vreinterpretq_s16_s32(in[2])));
in[5] = vmovl_s16(u);
u = vget_high_s16(vreinterpretq_s16_s32(in[1]));
in[2] = vmovl_s16(vget_low_s16(vreinterpretq_s16_s32(in[1])));
in[3] = vmovl_s16(u);
u = vget_high_s16(vreinterpretq_s16_s32(in[0]));
in[0] = vmovl_s16(vget_low_s16(vreinterpretq_s16_s32(in[0])));
in[1] = vmovl_s16(u);
const int32x4_t v_shift = vdupq_n_s32(shift);
in[0] = vshlq_s32(in[0], v_shift);
in[1] = vshlq_s32(in[1], v_shift);
in[2] = vshlq_s32(in[2], v_shift);
in[3] = vshlq_s32(in[3], v_shift);
in[4] = vshlq_s32(in[4], v_shift);
in[5] = vshlq_s32(in[5], v_shift);
in[6] = vshlq_s32(in[6], v_shift);
in[7] = vshlq_s32(in[7], v_shift);
in[8] = vshlq_s32(in[8], v_shift);
in[9] = vshlq_s32(in[9], v_shift);
in[10] = vshlq_s32(in[10], v_shift);
in[11] = vshlq_s32(in[11], v_shift);
in[12] = vshlq_s32(in[12], v_shift);
in[13] = vshlq_s32(in[13], v_shift);
in[14] = vshlq_s32(in[14], v_shift);
in[15] = vshlq_s32(in[15], v_shift);
}
static INLINE void col_txfm_8x8_rounding(int32x4_t *in,
const int32x4_t *v_shift) {
in[0] = vrshlq_s32(in[0], *v_shift);
in[1] = vrshlq_s32(in[1], *v_shift);
in[2] = vrshlq_s32(in[2], *v_shift);
in[3] = vrshlq_s32(in[3], *v_shift);
in[4] = vrshlq_s32(in[4], *v_shift);
in[5] = vrshlq_s32(in[5], *v_shift);
in[6] = vrshlq_s32(in[6], *v_shift);
in[7] = vrshlq_s32(in[7], *v_shift);
in[8] = vrshlq_s32(in[8], *v_shift);
in[9] = vrshlq_s32(in[9], *v_shift);
in[10] = vrshlq_s32(in[10], *v_shift);
in[11] = vrshlq_s32(in[11], *v_shift);
in[12] = vrshlq_s32(in[12], *v_shift);
in[13] = vrshlq_s32(in[13], *v_shift);
in[14] = vrshlq_s32(in[14], *v_shift);
in[15] = vrshlq_s32(in[15], *v_shift);
}
static INLINE void col_txfm_4x8_rounding(int32x4_t *in,
const int32x4_t *v_shift) {
in[0] = vrshlq_s32(in[0], *v_shift);
in[1] = vrshlq_s32(in[1], *v_shift);
in[2] = vrshlq_s32(in[2], *v_shift);
in[3] = vrshlq_s32(in[3], *v_shift);
in[4] = vrshlq_s32(in[4], *v_shift);
in[5] = vrshlq_s32(in[5], *v_shift);
in[6] = vrshlq_s32(in[6], *v_shift);
in[7] = vrshlq_s32(in[7], *v_shift);
}
static INLINE void write_buffer_8x8(const int32x4_t *res, int32_t *output) {
vst1q_s32(output + 0 * 4, res[0]);
vst1q_s32(output + 1 * 4, res[1]);
vst1q_s32(output + 2 * 4, res[2]);
vst1q_s32(output + 3 * 4, res[3]);
vst1q_s32(output + 4 * 4, res[4]);
vst1q_s32(output + 5 * 4, res[5]);
vst1q_s32(output + 6 * 4, res[6]);
vst1q_s32(output + 7 * 4, res[7]);
vst1q_s32(output + 8 * 4, res[8]);
vst1q_s32(output + 9 * 4, res[9]);
vst1q_s32(output + 10 * 4, res[10]);
vst1q_s32(output + 11 * 4, res[11]);
vst1q_s32(output + 12 * 4, res[12]);
vst1q_s32(output + 13 * 4, res[13]);
vst1q_s32(output + 14 * 4, res[14]);
vst1q_s32(output + 15 * 4, res[15]);
}
static INLINE void write_buffer_16x8(const int32x4_t *res, int32_t *output,
const int stride) {
vst1q_s32(output, res[0]);
vst1q_s32(output + 4, res[1]);
vst1q_s32(output + stride, res[2]);
vst1q_s32(output + stride + 4, res[3]);
vst1q_s32(output + (stride * 2), res[4]);
vst1q_s32(output + (stride * 2) + 4, res[5]);
vst1q_s32(output + (stride * 3), res[6]);
vst1q_s32(output + (stride * 3) + 4, res[7]);
vst1q_s32(output + (stride * 4), res[8]);
vst1q_s32(output + (stride * 4) + 4, res[9]);
vst1q_s32(output + (stride * 5), res[10]);
vst1q_s32(output + (stride * 5) + 4, res[11]);
vst1q_s32(output + (stride * 6), res[12]);
vst1q_s32(output + (stride * 6) + 4, res[13]);
vst1q_s32(output + (stride * 7), res[14]);
vst1q_s32(output + (stride * 7) + 4, res[15]);
}
static void fdct4x8_neon(int32x4_t *in, int32x4_t *out, int bit,
const int col_num) {
const int32_t *cospi = cospi_arr(bit);
const int32x4_t v_bit = vdupq_n_s32(-bit);
int32x4_t u[8], v[8];
int startidx = 0 * col_num;
int endidx = 7 * col_num;
// stage 0-1
u[0] = vaddq_s32(in[startidx], in[endidx]);
v[7] = vsubq_s32(in[startidx], in[endidx]);
startidx += col_num;
endidx -= col_num;
u[1] = vaddq_s32(in[startidx], in[endidx]);
u[6] = vsubq_s32(in[startidx], in[endidx]);
startidx += col_num;
endidx -= col_num;
u[2] = vaddq_s32(in[startidx], in[endidx]);
u[5] = vsubq_s32(in[startidx], in[endidx]);
startidx += col_num;
endidx -= col_num;
u[3] = vaddq_s32(in[startidx], in[endidx]);
v[4] = vsubq_s32(in[startidx], in[endidx]);
// stage 2
v[0] = vaddq_s32(u[0], u[3]);
v[3] = vsubq_s32(u[0], u[3]);
v[1] = vaddq_s32(u[1], u[2]);
v[2] = vsubq_s32(u[1], u[2]);
v[5] = vmulq_n_s32(u[6], cospi[32]);
v[5] = vmlsq_n_s32(v[5], u[5], cospi[32]);
v[5] = vrshlq_s32(v[5], v_bit);
u[0] = vmulq_n_s32(u[5], cospi[32]);
v[6] = vmlaq_n_s32(u[0], u[6], cospi[32]);
v[6] = vrshlq_s32(v[6], v_bit);
// stage 3
// type 0
v[0] = vmulq_n_s32(v[0], cospi[32]);
v[1] = vmulq_n_s32(v[1], cospi[32]);
u[0] = vaddq_s32(v[0], v[1]);
u[0] = vrshlq_s32(u[0], v_bit);
u[1] = vsubq_s32(v[0], v[1]);
u[1] = vrshlq_s32(u[1], v_bit);
// type 1
v[0] = vmulq_n_s32(v[2], cospi[48]);
u[2] = vmlaq_n_s32(v[0], v[3], cospi[16]);
u[2] = vrshlq_s32(u[2], v_bit);
v[1] = vmulq_n_s32(v[3], cospi[48]);
u[3] = vmlsq_n_s32(v[1], v[2], cospi[16]);
u[3] = vrshlq_s32(u[3], v_bit);
u[4] = vaddq_s32(v[4], v[5]);
u[5] = vsubq_s32(v[4], v[5]);
u[6] = vsubq_s32(v[7], v[6]);
u[7] = vaddq_s32(v[7], v[6]);
// stage 4-5
v[0] = vmulq_n_s32(u[4], cospi[56]);
v[0] = vmlaq_n_s32(v[0], u[7], cospi[8]);
out[1 * col_num] = vrshlq_s32(v[0], v_bit);
v[1] = vmulq_n_s32(u[7], cospi[56]);
v[0] = vmlsq_n_s32(v[1], u[4], cospi[8]);
out[7 * col_num] = vrshlq_s32(v[0], v_bit);
v[0] = vmulq_n_s32(u[5], cospi[24]);
v[0] = vmlaq_n_s32(v[0], u[6], cospi[40]);
out[5 * col_num] = vrshlq_s32(v[0], v_bit);
v[1] = vmulq_n_s32(u[6], cospi[24]);
v[0] = vmlsq_n_s32(v[1], u[5], cospi[40]);
out[3 * col_num] = vrshlq_s32(v[0], v_bit);
out[0 * col_num] = u[0];
out[4 * col_num] = u[1];
out[2 * col_num] = u[2];
out[6 * col_num] = u[3];
}
static void fdct8x8_neon(int32x4_t *in, int32x4_t *out, int bit,
const int col_num) {
fdct4x8_neon(in, out, bit, col_num);
fdct4x8_neon(in + 1, out + 1, bit, col_num);
}
static void fadst8x8_neon(int32x4_t *in, int32x4_t *out, int bit,
const int col_num) {
const int32_t *cospi = cospi_arr(bit);
const int32x4_t v_bit = vdupq_n_s32(-bit);
int32x4_t u0, u1, u2, u3, u4, u5, u6, u7;
int32x4_t v0, v1, v2, v3, v4, v5, v6, v7;
int32x4_t x, y;
int col;
for (col = 0; col < col_num; ++col) {
// stage 0-1
u0 = in[col_num * 0 + col];
u1 = vnegq_s32(in[col_num * 7 + col]);
u2 = vnegq_s32(in[col_num * 3 + col]);
u3 = in[col_num * 4 + col];
u4 = vnegq_s32(in[col_num * 1 + col]);
u5 = in[col_num * 6 + col];
u6 = in[col_num * 2 + col];
u7 = vnegq_s32(in[col_num * 5 + col]);
// stage 2
v0 = u0;
v1 = u1;
x = vmulq_n_s32(u2, cospi[32]);
y = vmulq_n_s32(u3, cospi[32]);
v2 = vaddq_s32(x, y);
v2 = vrshlq_s32(v2, v_bit);
v3 = vsubq_s32(x, y);
v3 = vrshlq_s32(v3, v_bit);
v4 = u4;
v5 = u5;
x = vmulq_n_s32(u6, cospi[32]);
y = vmulq_n_s32(u7, cospi[32]);
v6 = vaddq_s32(x, y);
v6 = vrshlq_s32(v6, v_bit);
v7 = vsubq_s32(x, y);
v7 = vrshlq_s32(v7, v_bit);
// stage 3
u0 = vaddq_s32(v0, v2);
u1 = vaddq_s32(v1, v3);
u2 = vsubq_s32(v0, v2);
u3 = vsubq_s32(v1, v3);
u4 = vaddq_s32(v4, v6);
u5 = vaddq_s32(v5, v7);
u6 = vsubq_s32(v4, v6);
u7 = vsubq_s32(v5, v7);
// stage 4
v0 = u0;
v1 = u1;
v2 = u2;
v3 = u3;
v4 = vmulq_n_s32(u4, cospi[16]);
v4 = vmlaq_n_s32(v4, u5, cospi[48]);
v4 = vrshlq_s32(v4, v_bit);
v5 = vmulq_n_s32(u4, cospi[48]);
v5 = vmlsq_n_s32(v5, u5, cospi[16]);
v5 = vrshlq_s32(v5, v_bit);
v6 = vmulq_n_s32(u7, cospi[16]);
v6 = vmlsq_n_s32(v6, u6, cospi[48]);
v6 = vrshlq_s32(v6, v_bit);
v7 = vmulq_n_s32(u6, cospi[16]);
v7 = vmlaq_n_s32(v7, u7, cospi[48]);
v7 = vrshlq_s32(v7, v_bit);
// stage 5
u0 = vaddq_s32(v0, v4);
u1 = vaddq_s32(v1, v5);
u2 = vaddq_s32(v2, v6);
u3 = vaddq_s32(v3, v7);
u4 = vsubq_s32(v0, v4);
u5 = vsubq_s32(v1, v5);
u6 = vsubq_s32(v2, v6);
u7 = vsubq_s32(v3, v7);
// stage 6
v0 = vmulq_n_s32(u0, cospi[4]);
v0 = vmlaq_n_s32(v0, u1, cospi[60]);
v0 = vrshlq_s32(v0, v_bit);
v1 = vmulq_n_s32(u0, cospi[60]);
v1 = vmlsq_n_s32(v1, u1, cospi[4]);
v1 = vrshlq_s32(v1, v_bit);
v2 = vmulq_n_s32(u2, cospi[20]);
v2 = vmlaq_n_s32(v2, u3, cospi[44]);
v2 = vrshlq_s32(v2, v_bit);
v3 = vmulq_n_s32(u2, cospi[44]);
v3 = vmlsq_n_s32(v3, u3, cospi[20]);
v3 = vrshlq_s32(v3, v_bit);
v4 = vmulq_n_s32(u4, cospi[36]);
v4 = vmlaq_n_s32(v4, u5, cospi[28]);
v4 = vrshlq_s32(v4, v_bit);
v5 = vmulq_n_s32(u4, cospi[28]);
v5 = vmlsq_n_s32(v5, u5, cospi[36]);
v5 = vrshlq_s32(v5, v_bit);
x = vmulq_n_s32(u6, cospi[52]);
v6 = vmlaq_n_s32(x, u7, cospi[12]);
v6 = vrshlq_s32(v6, v_bit);
v7 = vmulq_n_s32(u6, cospi[12]);
v7 = vmlsq_n_s32(v7, u7, cospi[52]);
v7 = vrshlq_s32(v7, v_bit);
// stage 7
out[col_num * 0 + col] = v1;
out[col_num * 1 + col] = v6;
out[col_num * 2 + col] = v3;
out[col_num * 3 + col] = v4;
out[col_num * 4 + col] = v5;
out[col_num * 5 + col] = v2;
out[col_num * 6 + col] = v7;
out[col_num * 7 + col] = v0;
}
}
static void idtx8x8_neon(int32x4_t *in, int32x4_t *out, int bit, int col_num) {
(void)bit;
for (int i = 0; i < col_num; i += 1) {
out[0 + 8 * i] = vshlq_n_s32(in[0 + 8 * i], 1);
out[1 + 8 * i] = vshlq_n_s32(in[1 + 8 * i], 1);
out[2 + 8 * i] = vshlq_n_s32(in[2 + 8 * i], 1);
out[3 + 8 * i] = vshlq_n_s32(in[3 + 8 * i], 1);
out[4 + 8 * i] = vshlq_n_s32(in[4 + 8 * i], 1);
out[5 + 8 * i] = vshlq_n_s32(in[5 + 8 * i], 1);
out[6 + 8 * i] = vshlq_n_s32(in[6 + 8 * i], 1);
out[7 + 8 * i] = vshlq_n_s32(in[7 + 8 * i], 1);
}
}
#if !CONFIG_REALTIME_ONLY
static void idtx32x8_neon(int32x4_t *in, int32x4_t *out, int bit, int col_num) {
(void)bit;
(void)col_num;
for (int j = 0; j < 2; j++) {
out[j + 8 * 0] = vshlq_n_s32(in[j + 8 * 0], 1);
out[j + 8 * 1] = vshlq_n_s32(in[j + 8 * 1], 1);
out[j + 8 * 2] = vshlq_n_s32(in[j + 8 * 2], 1);
out[j + 8 * 3] = vshlq_n_s32(in[j + 8 * 3], 1);
out[j + 8 * 4] = vshlq_n_s32(in[j + 8 * 4], 1);
out[j + 8 * 5] = vshlq_n_s32(in[j + 8 * 5], 1);
out[j + 8 * 6] = vshlq_n_s32(in[j + 8 * 6], 1);
out[j + 8 * 7] = vshlq_n_s32(in[j + 8 * 7], 1);
}
}
#endif
void av1_fwd_txfm2d_8x8_neon(const int16_t *input, int32_t *coeff, int stride,
TX_TYPE tx_type, int bd) {
int32x4_t in[16], out[16];
const int8_t *shift = av1_fwd_txfm_shift_ls[TX_8X8];
const int txw_idx = get_txw_idx(TX_8X8);
const int txh_idx = get_txh_idx(TX_8X8);
const int32x4_t v_shift1 = vdupq_n_s32(shift[1]);
switch (tx_type) {
case DCT_DCT:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
fdct8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fdct8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case ADST_DCT:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fdct8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case DCT_ADST:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
fdct8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fadst8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case ADST_ADST:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fadst8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case FLIPADST_DCT:
load_buffer_8x8(input, in, stride, 1, 0, shift[0]);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fdct8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case DCT_FLIPADST:
load_buffer_8x8(input, in, stride, 0, 1, shift[0]);
fdct8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fadst8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case FLIPADST_FLIPADST:
load_buffer_8x8(input, in, stride, 1, 1, shift[0]);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fadst8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case ADST_FLIPADST:
load_buffer_8x8(input, in, stride, 0, 1, shift[0]);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fadst8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case FLIPADST_ADST:
load_buffer_8x8(input, in, stride, 1, 0, shift[0]);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fadst8x8_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case IDTX:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
idtx8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
idtx8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case V_DCT:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
fdct8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
idtx8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case H_DCT:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
idtx8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fdct8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case V_ADST:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
idtx8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case H_ADST:
load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
idtx8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case V_FLIPADST:
load_buffer_8x8(input, in, stride, 1, 0, shift[0]);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
idtx8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
case H_FLIPADST:
load_buffer_8x8(input, in, stride, 0, 1, shift[0]);
idtx8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
col_txfm_8x8_rounding(out, &v_shift1);
transpose_8x8(out, in);
fadst8x8_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], 2);
transpose_8x8(out, in);
write_buffer_8x8(in, coeff);
break;
default: assert(0);
}
(void)bd;
}
// Hybrid Transform 16x16
static INLINE void convert_8x8_to_16x16(const int32x4_t *in, int32x4_t *out) {
int row_index = 0;
int dst_index = 0;
int src_index = 0;
// row 0, 1, .., 7
do {
out[dst_index] = in[src_index];
out[dst_index + 1] = in[src_index + 1];
out[dst_index + 2] = in[src_index + 16];
out[dst_index + 3] = in[src_index + 17];
dst_index += 4;
src_index += 2;
row_index += 1;
} while (row_index < 8);
// row 8, 9, ..., 15
src_index += 16;
do {
out[dst_index] = in[src_index];
out[dst_index + 1] = in[src_index + 1];
out[dst_index + 2] = in[src_index + 16];
out[dst_index + 3] = in[src_index + 17];
dst_index += 4;
src_index += 2;
row_index += 1;
} while (row_index < 16);
}
static INLINE void load_buffer_16x16(const int16_t *input, int32x4_t *out,
int stride, int flipud, int fliplr,
int shift) {
int32x4_t in[64];
// Load 4 8x8 blocks
const int16_t *topL = input;
const int16_t *topR = input + 8;
const int16_t *botL = input + 8 * stride;
const int16_t *botR = input + 8 * stride + 8;
const int16_t *tmp;
if (flipud) {
// Swap left columns
tmp = topL;
topL = botL;
botL = tmp;
// Swap right columns
tmp = topR;
topR = botR;
botR = tmp;
}
if (fliplr) {
// Swap top rows
tmp = topL;
topL = topR;
topR = tmp;
// Swap bottom rows
tmp = botL;
botL = botR;
botR = tmp;
}
// load first 8 columns
load_buffer_8x8(topL, &in[0], stride, flipud, fliplr, shift);
load_buffer_8x8(botL, &in[32], stride, flipud, fliplr, shift);
// load second 8 columns
load_buffer_8x8(topR, &in[16], stride, flipud, fliplr, shift);
load_buffer_8x8(botR, &in[48], stride, flipud, fliplr, shift);
convert_8x8_to_16x16(in, out);
}
static INLINE void load_buffer_8x16(const int16_t *input, int32x4_t *out,
int stride, int flipud, int fliplr,
int shift) {
const int16_t *topL = input;
const int16_t *botL = input + 8 * stride;
const int16_t *tmp;
if (flipud) {
tmp = topL;
topL = botL;
botL = tmp;
}
load_buffer_8x8(topL, out, stride, flipud, fliplr, shift);
load_buffer_8x8(botL, out + 16, stride, flipud, fliplr, shift);
}
static INLINE void load_buffer_8x4(const int16_t *input, int32x4_t *out,
int stride, int flipud, int fliplr,
const int32x4_t *v_shift) {
const int16_t *topL = input;
const int16_t *topR = input + 4;
const int16_t *tmp;
if (fliplr) {
tmp = topL;
topL = topR;
topR = tmp;
}
load_buffer_4x4(topL, out, stride, flipud, fliplr, v_shift);
load_buffer_4x4(topR, out + 4, stride, flipud, fliplr, v_shift);
}
static INLINE void load_buffer_16x4(const int16_t *input, int32x4_t *out,
int stride, int flipud, int fliplr,
const int32x4_t *v_shift) {
const int16_t *topL = input;
const int16_t *topR = input + 8;
const int16_t *tmp;
if (fliplr) {
tmp = topL;
topL = topR;
topR = tmp;
}
load_buffer_8x4(topL, out, stride, flipud, fliplr, v_shift);
load_buffer_8x4(topR, out + 8, stride, flipud, fliplr, v_shift);
}
static INLINE void load_buffer_4x8(const int16_t *input, int32x4_t *out,
int stride, int flipud, int fliplr,
const int32x4_t *v_shift) {
const int16_t *topL = input;
const int16_t *botL = input + 4 * stride;
const int16_t *tmp;
if (flipud) {
tmp = topL;
topL = botL;
botL = tmp;
}
load_buffer_4x4(topL, out, stride, flipud, fliplr, v_shift);
load_buffer_4x4(botL, out + 4, stride, flipud, fliplr, v_shift);
}
#if !CONFIG_REALTIME_ONLY
static INLINE void load_buffer_4x16(const int16_t *input, int32x4_t *out,
const int stride, const int flipud,
const int fliplr,
const int32x4_t *v_shift) {
const int16_t *topL = input;
const int16_t *botL = input + 8 * stride;
const int16_t *tmp;
if (flipud) {
tmp = topL;
topL = botL;
botL = tmp;
}
load_buffer_4x8(topL, out, stride, flipud, fliplr, v_shift);
load_buffer_4x8(botL, out + 8, stride, flipud, fliplr, v_shift);
}
#endif
static INLINE void load_buffer_32x8n(const int16_t *input, int32x4_t *out,
int stride, int flipud, int fliplr,
int shift, const int height) {
const int16_t *in = input;
int32x4_t *output = out;
for (int col = 0; col < height; col++) {
in = input + col * stride;
output = out + col * 8;
int32x4_t v_shift = vdupq_n_s32(shift);
load_buffer_4x4(in, output, 4, flipud, fliplr, &v_shift);
load_buffer_4x4((in + 16), (output + 4), 4, flipud, fliplr, &v_shift);
}
}
static void fdct16x16_neon(int32x4_t *in, int32x4_t *out, int bit,
const int col_num) {
const int32_t *cospi = cospi_arr(bit);
const int32x4_t v_bit = vdupq_n_s32(-bit);
int32x4_t u[16], v[16];
int col;
// Calculate the column 0, 1, 2, 3
for (col = 0; col < col_num; ++col) {
// stage 0
// stage 1
u[0] = vaddq_s32(in[0 * col_num + col], in[15 * col_num + col]);
u[15] = vsubq_s32(in[0 * col_num + col], in[15 * col_num + col]);
u[1] = vaddq_s32(in[1 * col_num + col], in[14 * col_num + col]);
u[14] = vsubq_s32(in[1 * col_num + col], in[14 * col_num + col]);
u[2] = vaddq_s32(in[2 * col_num + col], in[13 * col_num + col]);
u[13] = vsubq_s32(in[2 * col_num + col], in[13 * col_num + col]);
u[3] = vaddq_s32(in[3 * col_num + col], in[12 * col_num + col]);
u[12] = vsubq_s32(in[3 * col_num + col], in[12 * col_num + col]);
u[4] = vaddq_s32(in[4 * col_num + col], in[11 * col_num + col]);
u[11] = vsubq_s32(in[4 * col_num + col], in[11 * col_num + col]);
u[5] = vaddq_s32(in[5 * col_num + col], in[10 * col_num + col]);
u[10] = vsubq_s32(in[5 * col_num + col], in[10 * col_num + col]);
u[6] = vaddq_s32(in[6 * col_num + col], in[9 * col_num + col]);
u[9] = vsubq_s32(in[6 * col_num + col], in[9 * col_num + col]);
u[7] = vaddq_s32(in[7 * col_num + col], in[8 * col_num + col]);
u[8] = vsubq_s32(in[7 * col_num + col], in[8 * col_num + col]);
// stage 2
v[0] = vaddq_s32(u[0], u[7]);
v[7] = vsubq_s32(u[0], u[7]);
v[1] = vaddq_s32(u[1], u[6]);
v[6] = vsubq_s32(u[1], u[6]);
v[2] = vaddq_s32(u[2], u[5]);
v[5] = vsubq_s32(u[2], u[5]);
v[3] = vaddq_s32(u[3], u[4]);
v[4] = vsubq_s32(u[3], u[4]);
v[8] = u[8];
v[9] = u[9];
v[10] = vmulq_n_s32(u[13], cospi[32]);
v[10] = vmlsq_n_s32(v[10], u[10], cospi[32]);
v[10] = vrshlq_s32(v[10], v_bit);
v[13] = vmulq_n_s32(u[10], cospi[32]);
v[13] = vmlaq_n_s32(v[13], u[13], cospi[32]);
v[13] = vrshlq_s32(v[13], v_bit);
v[11] = vmulq_n_s32(u[12], cospi[32]);
v[11] = vmlsq_n_s32(v[11], u[11], cospi[32]);
v[11] = vrshlq_s32(v[11], v_bit);
v[12] = vmulq_n_s32(u[11], cospi[32]);
v[12] = vmlaq_n_s32(v[12], u[12], cospi[32]);
v[12] = vrshlq_s32(v[12], v_bit);
v[14] = u[14];
v[15] = u[15];
// stage 3
u[0] = vaddq_s32(v[0], v[3]);
u[3] = vsubq_s32(v[0], v[3]);
u[1] = vaddq_s32(v[1], v[2]);
u[2] = vsubq_s32(v[1], v[2]);
u[4] = v[4];
u[5] = vmulq_n_s32(v[6], cospi[32]);
u[5] = vmlsq_n_s32(u[5], v[5], cospi[32]);
u[5] = vrshlq_s32(u[5], v_bit);
u[6] = vmulq_n_s32(v[5], cospi[32]);
u[6] = vmlaq_n_s32(u[6], v[6], cospi[32]);
u[6] = vrshlq_s32(u[6], v_bit);
u[7] = v[7];
u[8] = vaddq_s32(v[8], v[11]);
u[11] = vsubq_s32(v[8], v[11]);
u[9] = vaddq_s32(v[9], v[10]);
u[10] = vsubq_s32(v[9], v[10]);
u[12] = vsubq_s32(v[15], v[12]);
u[15] = vaddq_s32(v[15], v[12]);
u[13] = vsubq_s32(v[14], v[13]);
u[14] = vaddq_s32(v[14], v[13]);
// stage 4
u[0] = vmulq_n_s32(u[0], cospi[32]);
u[1] = vmulq_n_s32(u[1], cospi[32]);
v[0] = vaddq_s32(u[0], u[1]);
v[0] = vrshlq_s32(v[0], v_bit);
v[1] = vsubq_s32(u[0], u[1]);
v[1] = vrshlq_s32(v[1], v_bit);
v[2] = vmulq_n_s32(u[2], cospi[48]);
v[2] = vmlaq_n_s32(v[2], u[3], cospi[16]);
v[2] = vrshlq_s32(v[2], v_bit);
v[3] = vmulq_n_s32(u[3], cospi[48]);
v[3] = vmlsq_n_s32(v[3], u[2], cospi[16]);
v[3] = vrshlq_s32(v[3], v_bit);
v[4] = vaddq_s32(u[4], u[5]);
v[5] = vsubq_s32(u[4], u[5]);
v[6] = vsubq_s32(u[7], u[6]);
v[7] = vaddq_s32(u[7], u[6]);
v[8] = u[8];
v[9] = vmulq_n_s32(u[14], cospi[48]);
v[9] = vmlsq_n_s32(v[9], u[9], cospi[16]);
v[9] = vrshlq_s32(v[9], v_bit);
v[14] = vmulq_n_s32(u[9], cospi[48]);
v[14] = vmlaq_n_s32(v[14], u[14], cospi[16]);
v[14] = vrshlq_s32(v[14], v_bit);
v[10] = vmulq_n_s32(u[13], -cospi[16]);
v[10] = vmlsq_n_s32(v[10], u[10], cospi[48]);
v[10] = vrshlq_s32(v[10], v_bit);
v[13] = vmulq_n_s32(u[10], -cospi[16]);
v[13] = vmlaq_n_s32(v[13], u[13], cospi[48]);
v[13] = vrshlq_s32(v[13], v_bit);
v[11] = u[11];
v[12] = u[12];
v[15] = u[15];
// stage 5
u[0] = v[0];
u[1] = v[1];
u[2] = v[2];
u[3] = v[3];
u[4] = vmulq_n_s32(v[4], cospi[56]);
u[4] = vmlaq_n_s32(u[4], v[7], cospi[8]);
u[4] = vrshlq_s32(u[4], v_bit);
u[7] = vmulq_n_s32(v[7], cospi[56]);
u[7] = vmlsq_n_s32(u[7], v[4], cospi[8]);
u[7] = vrshlq_s32(u[7], v_bit);
u[5] = vmulq_n_s32(v[5], cospi[24]);
u[5] = vmlaq_n_s32(u[5], v[6], cospi[40]);
u[5] = vrshlq_s32(u[5], v_bit);
u[6] = vmulq_n_s32(v[6], cospi[24]);
u[6] = vmlsq_n_s32(u[6], v[5], cospi[40]);
u[6] = vrshlq_s32(u[6], v_bit);
u[8] = vaddq_s32(v[8], v[9]);
u[9] = vsubq_s32(v[8], v[9]);
u[10] = vsubq_s32(v[11], v[10]);
u[11] = vaddq_s32(v[11], v[10]);
u[12] = vaddq_s32(v[12], v[13]);
u[13] = vsubq_s32(v[12], v[13]);
u[14] = vsubq_s32(v[15], v[14]);
u[15] = vaddq_s32(v[15], v[14]);
// stage 6
v[0] = u[0];
v[1] = u[1];
v[2] = u[2];
v[3] = u[3];
v[4] = u[4];
v[5] = u[5];
v[6] = u[6];
v[7] = u[7];
v[8] = vmulq_n_s32(u[8], cospi[60]);
v[8] = vmlaq_n_s32(v[8], u[15], cospi[4]);
v[8] = vrshlq_s32(v[8], v_bit);
v[15] = vmulq_n_s32(u[15], cospi[60]);
v[15] = vmlsq_n_s32(v[15], u[8], cospi[4]);
v[15] = vrshlq_s32(v[15], v_bit);
v[9] = vmulq_n_s32(u[9], cospi[28]);
v[9] = vmlaq_n_s32(v[9], u[14], cospi[36]);
v[9] = vrshlq_s32(v[9], v_bit);
v[14] = vmulq_n_s32(u[14], cospi[28]);
v[14] = vmlsq_n_s32(v[14], u[9], cospi[36]);
v[14] = vrshlq_s32(v[14], v_bit);
v[10] = vmulq_n_s32(u[10], cospi[44]);
v[10] = vmlaq_n_s32(v[10], u[13], cospi[20]);
v[10] = vrshlq_s32(v[10], v_bit);
v[13] = vmulq_n_s32(u[13], cospi[44]);
v[13] = vmlsq_n_s32(v[13], u[10], cospi[20]);
v[13] = vrshlq_s32(v[13], v_bit);
v[11] = vmulq_n_s32(u[11], cospi[12]);
v[11] = vmlaq_n_s32(v[11], u[12], cospi[52]);
v[11] = vrshlq_s32(v[11], v_bit);
v[12] = vmulq_n_s32(u[12], cospi[12]);
v[12] = vmlsq_n_s32(v[12], u[11], cospi[52]);
v[12] = vrshlq_s32(v[12], v_bit);
out[0 * col_num + col] = v[0];
out[1 * col_num + col] = v[8];
out[2 * col_num + col] = v[4];
out[3 * col_num + col] = v[12];
out[4 * col_num + col] = v[2];
out[5 * col_num + col] = v[10];
out[6 * col_num + col] = v[6];
out[7 * col_num + col] = v[14];
out[8 * col_num + col] = v[1];
out[9 * col_num + col] = v[9];
out[10 * col_num + col] = v[5];
out[11 * col_num + col] = v[13];
out[12 * col_num + col] = v[3];
out[13 * col_num + col] = v[11];
out[14 * col_num + col] = v[7];
out[15 * col_num + col] = v[15];
}
}
static void fadst16x16_neon(int32x4_t *in, int32x4_t *out, int bit,
const int num_cols) {
const int32_t *cospi = cospi_arr(bit);
const int32x4_t v_bit = vdupq_n_s32(-bit);
int32x4_t u[16], v[16], x, y;
int col;
for (col = 0; col < num_cols; ++col) {
// stage 0-1
u[0] = in[0 * num_cols + col];
u[1] = vnegq_s32(in[15 * num_cols + col]);
u[2] = vnegq_s32(in[7 * num_cols + col]);
u[3] = in[8 * num_cols + col];
u[4] = vnegq_s32(in[3 * num_cols + col]);
u[5] = in[12 * num_cols + col];
u[6] = in[4 * num_cols + col];
u[7] = vnegq_s32(in[11 * num_cols + col]);
u[8] = vnegq_s32(in[1 * num_cols + col]);
u[9] = in[14 * num_cols + col];
u[10] = in[6 * num_cols + col];
u[11] = vnegq_s32(in[9 * num_cols + col]);
u[12] = in[2 * num_cols + col];
u[13] = vnegq_s32(in[13 * num_cols + col]);
u[14] = vnegq_s32(in[5 * num_cols + col]);
u[15] = in[10 * num_cols + col];
// stage 2
v[0] = u[0];
v[1] = u[1];
x = vmulq_n_s32(u[2], cospi[32]);
y = vmulq_n_s32(u[3], cospi[32]);
v[2] = vaddq_s32(x, y);
v[2] = vrshlq_s32(v[2], v_bit);
v[3] = vsubq_s32(x, y);
v[3] = vrshlq_s32(v[3], v_bit);
v[4] = u[4];
v[5] = u[5];
x = vmulq_n_s32(u[6], cospi[32]);
y = vmulq_n_s32(u[7], cospi[32]);
v[6] = vaddq_s32(x, y);
v[6] = vrshlq_s32(v[6], v_bit);
v[7] = vsubq_s32(x, y);
v[7] = vrshlq_s32(v[7], v_bit);
v[8] = u[8];
v[9] = u[9];
x = vmulq_n_s32(u[10], cospi[32]);
y = vmulq_n_s32(u[11], cospi[32]);
v[10] = vaddq_s32(x, y);
v[10] = vrshlq_s32(v[10], v_bit);
v[11] = vsubq_s32(x, y);
v[11] = vrshlq_s32(v[11], v_bit);
v[12] = u[12];
v[13] = u[13];
x = vmulq_n_s32(u[14], cospi[32]);
y = vmulq_n_s32(u[15], cospi[32]);
v[14] = vaddq_s32(x, y);
v[14] = vrshlq_s32(v[14], v_bit);
v[15] = vsubq_s32(x, y);
v[15] = vrshlq_s32(v[15], v_bit);
// stage 3
u[0] = vaddq_s32(v[0], v[2]);
u[1] = vaddq_s32(v[1], v[3]);
u[2] = vsubq_s32(v[0], v[2]);
u[3] = vsubq_s32(v[1], v[3]);
u[4] = vaddq_s32(v[4], v[6]);
u[5] = vaddq_s32(v[5], v[7]);
u[6] = vsubq_s32(v[4], v[6]);
u[7] = vsubq_s32(v[5], v[7]);
u[8] = vaddq_s32(v[8], v[10]);
u[9] = vaddq_s32(v[9], v[11]);
u[10] = vsubq_s32(v[8], v[10]);
u[11] = vsubq_s32(v[9], v[11]);
u[12] = vaddq_s32(v[12], v[14]);
u[13] = vaddq_s32(v[13], v[15]);
u[14] = vsubq_s32(v[12], v[14]);
u[15] = vsubq_s32(v[13], v[15]);
// stage 4
v[0] = u[0];
v[1] = u[1];
v[2] = u[2];
v[3] = u[3];
v[4] = half_btf_neon(&cospi[16], &u[4], &cospi[48], &u[5], v_bit);
v[7] = half_btf_neon(&cospi[16], &u[6], &cospi[48], &u[7], v_bit);
v[5] = half_btf_neon_m(&cospi[48], &u[4], &cospi[16], &u[5], v_bit);
v[6] = half_btf_neon_m(&cospi[16], &u[7], &cospi[48], &u[6], v_bit);
v[8] = u[8];
v[9] = u[9];
v[10] = u[10];
v[11] = u[11];
v[12] = half_btf_neon(&cospi[16], &u[12], &cospi[48], &u[13], v_bit);
v[15] = half_btf_neon(&cospi[16], &u[14], &cospi[48], &u[15], v_bit);
v[13] = half_btf_neon_m(&cospi[48], &u[12], &cospi[16], &u[13], v_bit);
v[14] = half_btf_neon_m(&cospi[16], &u[15], &cospi[48], &u[14], v_bit);
// stage 5
u[0] = vaddq_s32(v[0], v[4]);
u[1] = vaddq_s32(v[1], v[5]);
u[2] = vaddq_s32(v[2], v[6]);
u[3] = vaddq_s32(v[3], v[7]);
u[4] = vsubq_s32(v[0], v[4]);
u[5] = vsubq_s32(v[1], v[5]);
u[6] = vsubq_s32(v[2], v[6]);
u[7] = vsubq_s32(v[3], v[7]);
u[8] = vaddq_s32(v[8], v[12]);
u[9] = vaddq_s32(v[9], v[13]);
u[10] = vaddq_s32(v[10], v[14]);
u[11] = vaddq_s32(v[11], v[15]);
u[12] = vsubq_s32(v[8], v[12]);
u[13] = vsubq_s32(v[9], v[13]);
u[14] = vsubq_s32(v[10], v[14]);
u[15] = vsubq_s32(v[11], v[15]);
// stage 6
v[0] = u[0];
v[1] = u[1];
v[2] = u[2];
v[3] = u[3];
v[4] = u[4];
v[5] = u[5];
v[6] = u[6];
v[7] = u[7];
v[8] = half_btf_neon(&cospi[8], &u[8], &cospi[56], &u[9], v_bit);
v[13] = half_btf_neon(&cospi[8], &u[12], &cospi[56], &u[13], v_bit);
v[9] = half_btf_neon_m(&cospi[56], &u[8], &cospi[8], &u[9], v_bit);
v[12] = half_btf_neon_m(&cospi[8], &u[13], &cospi[56], &u[12], v_bit);
v[10] = half_btf_neon(&cospi[40], &u[10], &cospi[24], &u[11], v_bit);
v[15] = half_btf_neon(&cospi[40], &u[14], &cospi[24], &u[15], v_bit);
v[11] = half_btf_neon_m(&cospi[24], &u[10], &cospi[40], &u[11], v_bit);
v[14] = half_btf_neon_m(&cospi[40], &u[15], &cospi[24], &u[14], v_bit);
// stage 7
u[0] = vaddq_s32(v[0], v[8]);
u[1] = vaddq_s32(v[1], v[9]);
u[2] = vaddq_s32(v[2], v[10]);
u[3] = vaddq_s32(v[3], v[11]);
u[4] = vaddq_s32(v[4], v[12]);
u[5] = vaddq_s32(v[5], v[13]);
u[6] = vaddq_s32(v[6], v[14]);
u[7] = vaddq_s32(v[7], v[15]);
u[8] = vsubq_s32(v[0], v[8]);
u[9] = vsubq_s32(v[1], v[9]);
u[10] = vsubq_s32(v[2], v[10]);
u[11] = vsubq_s32(v[3], v[11]);
u[12] = vsubq_s32(v[4], v[12]);
u[13] = vsubq_s32(v[5], v[13]);
u[14] = vsubq_s32(v[6], v[14]);
u[15] = vsubq_s32(v[7], v[15]);
// stage 8
v[0] = half_btf_neon(&cospi[2], &u[0], &cospi[62], &u[1], v_bit);
v[1] = half_btf_neon_m(&cospi[62], &u[0], &cospi[2], &u[1], v_bit);
v[2] = half_btf_neon(&cospi[10], &u[2], &cospi[54], &u[3], v_bit);
v[3] = half_btf_neon_m(&cospi[54], &u[2], &cospi[10], &u[3], v_bit);
v[4] = half_btf_neon(&cospi[18], &u[4], &cospi[46], &u[5], v_bit);
v[5] = half_btf_neon_m(&cospi[46], &u[4], &cospi[18], &u[5], v_bit);
v[6] = half_btf_neon(&cospi[26], &u[6], &cospi[38], &u[7], v_bit);
v[7] = half_btf_neon_m(&cospi[38], &u[6], &cospi[26], &u[7], v_bit);
v[8] = half_btf_neon(&cospi[34], &u[8], &cospi[30], &u[9], v_bit);
v[9] = half_btf_neon_m(&cospi[30], &u[8], &cospi[34], &u[9], v_bit);
v[10] = half_btf_neon(&cospi[42], &u[10], &cospi[22], &u[11], v_bit);
v[11] = half_btf_neon_m(&cospi[22], &u[10], &cospi[42], &u[11], v_bit);
v[12] = half_btf_neon(&cospi[50], &u[12], &cospi[14], &u[13], v_bit);
v[13] = half_btf_neon_m(&cospi[14], &u[12], &cospi[50], &u[13], v_bit);
v[14] = half_btf_neon(&cospi[58], &u[14], &cospi[6], &u[15], v_bit);
v[15] = half_btf_neon_m(&cospi[6], &u[14], &cospi[58], &u[15], v_bit);
// stage 9
out[0 * num_cols + col] = v[1];
out[1 * num_cols + col] = v[14];
out[2 * num_cols + col] = v[3];
out[3 * num_cols + col] = v[12];
out[4 * num_cols + col] = v[5];
out[5 * num_cols + col] = v[10];
out[6 * num_cols + col] = v[7];
out[7 * num_cols + col] = v[8];
out[8 * num_cols + col] = v[9];
out[9 * num_cols + col] = v[6];
out[10 * num_cols + col] = v[11];
out[11 * num_cols + col] = v[4];
out[12 * num_cols + col] = v[13];
out[13 * num_cols + col] = v[2];
out[14 * num_cols + col] = v[15];
out[15 * num_cols + col] = v[0];
}
}
static void col_txfm_16x16_rounding(int32x4_t *in, const int32x4_t *v_shift) {
// Note:
// We split 16x16 rounding into 4 sections of 8x8 rounding,
// instead of 4 columns
col_txfm_8x8_rounding(&in[0], v_shift);
col_txfm_8x8_rounding(&in[16], v_shift);
col_txfm_8x8_rounding(&in[32], v_shift);
col_txfm_8x8_rounding(&in[48], v_shift);
}
static void col_txfm_8x16_rounding(int32x4_t *in, const int32x4_t *v_shift) {
col_txfm_8x8_rounding(&in[0], v_shift);
col_txfm_8x8_rounding(&in[16], v_shift);
}
static void write_buffer_16x16(const int32x4_t *in, int32_t *output) {
const int size_8x8 = 16 * 4;
write_buffer_8x8(&in[0], output);
output += size_8x8;
write_buffer_8x8(&in[16], output);
output += size_8x8;
write_buffer_8x8(&in[32], output);
output += size_8x8;
write_buffer_8x8(&in[48], output);
}
static void idtx16x16_neon(int32x4_t *in, int32x4_t *out, int bit,
int col_num) {
(void)bit;
int32x4_t fact = vdupq_n_s32(2 * NewSqrt2);
int32x4_t offset = vdupq_n_s32(1 << (NewSqrt2Bits - 1));
int32x4_t a_low;
int num_iters = 16 * col_num;
for (int i = 0; i < num_iters; i++) {
a_low = vmulq_s32(in[i], fact);
a_low = vaddq_s32(a_low, offset);
out[i] = vshrq_n_s32(a_low, NewSqrt2Bits);
}
}
void av1_fwd_txfm2d_16x16_neon(const int16_t *input, int32_t *coeff, int stride,
TX_TYPE tx_type, int bd) {
int32x4_t in[64], out[64];
const int8_t *shift = av1_fwd_txfm_shift_ls[TX_16X16];
const int txw_idx = get_txw_idx(TX_16X16);
const int txh_idx = get_txh_idx(TX_16X16);
const int col_num = 4;
const int32x4_t v_shift = vdupq_n_s32(shift[1]);
switch (tx_type) {
case DCT_DCT:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
fdct16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fdct16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case ADST_DCT:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
fadst16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fdct16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case DCT_ADST:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
fdct16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fadst16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case ADST_ADST:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
fadst16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fadst16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case FLIPADST_DCT:
load_buffer_16x16(input, in, stride, 1, 0, shift[0]);
fadst16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fdct16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case DCT_FLIPADST:
load_buffer_16x16(input, in, stride, 0, 1, shift[0]);
fdct16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fadst16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case FLIPADST_FLIPADST:
load_buffer_16x16(input, in, stride, 1, 1, shift[0]);
fadst16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fadst16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case ADST_FLIPADST:
load_buffer_16x16(input, in, stride, 0, 1, shift[0]);
fadst16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fadst16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case FLIPADST_ADST:
load_buffer_16x16(input, in, stride, 1, 0, shift[0]);
fadst16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fadst16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case IDTX:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
idtx16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
idtx16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case V_DCT:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
fdct16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
idtx16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case H_DCT:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
idtx16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fdct16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case V_ADST:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
fadst16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
idtx16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case H_ADST:
load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
idtx16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fadst16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case V_FLIPADST:
load_buffer_16x16(input, in, stride, 1, 0, shift[0]);
fadst16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
idtx16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
case H_FLIPADST:
load_buffer_16x16(input, in, stride, 0, 1, shift[0]);
idtx16x16_neon(in, out, av1_fwd_cos_bit_col[txw_idx][txh_idx], col_num);
col_txfm_16x16_rounding(out, &v_shift);
transpose_16x16(out, in);
fadst16x16_neon(in, out, av1_fwd_cos_bit_row[txw_idx][txh_idx], col_num);
transpose_16x16(out, in);
write_buffer_16x16(in, coeff);
break;
default: assert(0);
}
(void)bd;
}
static INLINE void flip_buf_neon(int32x4_t *in, int32x4_t *out, int size) {
for (int i = 0; i < size; i += 2) in[30 - i] = out[i];
for (int i = 1; i < size; i += 2) in[size - i] = out[i];
}
typedef void (*fwd_transform_1d_neon)(int32x4_t *in, int32x4_t *out, int bit,
const int num_cols);
static const fwd_transform_1d_neon col_highbd_txfm8x8_arr[TX_TYPES] = {
fdct8x8_neon, // DCT_DCT
fadst8x8_neon, // ADST_DCT
fdct8x8_neon, // DCT_ADST
fadst8x8_neon, // ADST_ADST
fadst8x8_neon, // FLIPADST_DCT
fdct8x8_neon, // DCT_FLIPADST
fadst8x8_neon, // FLIPADST_FLIPADST
fadst8x8_neon, // ADST_FLIPADST
fadst8x8_neon, // FLIPADST_ADST
idtx8x8_neon, // IDTX
fdct8x8_neon, // V_DCT
idtx8x8_neon, // H_DCT
fadst8x8_neon, // V_ADST
idtx8x8_neon, // H_ADST
fadst8x8_neon, // V_FLIPADST
idtx8x8_neon // H_FLIPADST
};
#if !CONFIG_REALTIME_ONLY
static const fwd_transform_1d_neon row_highbd_txfm32x8_arr[TX_TYPES] = {
fdct8x8_neon, // DCT_DCT
NULL, // ADST_DCT
NULL, // DCT_ADST
NULL, // ADST_ADST
NULL, // FLIPADST_DCT
NULL, // DCT_FLIPADST
NULL, // FLIPADST_FLIPADST
NULL, // ADST_FLIPADST
NULL, // FLIPADST-ADST
idtx32x8_neon, // IDTX
NULL, // V_DCT
NULL, // H_DCT
NULL, // V_ADST
NULL, // H_ADST
NULL, // V_FLIPADST
NULL, // H_FLIPADST
};
#endif
static const fwd_transform_1d_neon col_highbd_txfm4x8_arr[TX_TYPES] = {
fdct4x8_neon, // DCT_DCT
fadst8x8_neon, // ADST_DCT
fdct4x8_neon, // DCT_ADST
fadst8x8_neon, // ADST_ADST
fadst8x8_neon, // FLIPADST_DCT
fdct4x8_neon, // DCT_FLIPADST
fadst8x8_neon, // FLIPADST_FLIPADST
fadst8x8_neon, // ADST_FLIPADST
fadst8x8_neon, // FLIPADST_ADST
idtx8x8_neon, // IDTX
fdct4x8_neon, // V_DCT
idtx8x8_neon, // H_DCT
fadst8x8_neon, // V_ADST
idtx8x8_neon, // H_ADST
fadst8x8_neon, // V_FLIPADST
idtx8x8_neon // H_FLIPADST
};
static const fwd_transform_1d_neon row_highbd_txfm8x16_arr[TX_TYPES] = {
fdct16x16_neon, // DCT_DCT
fdct16x16_neon, // ADST_DCT
fadst16x16_neon, // DCT_ADST
fadst16x16_neon, // ADST_ADST
fdct16x16_neon, // FLIPADST_DCT
fadst16x16_neon, // DCT_FLIPADST
fadst16x16_neon, // FLIPADST_FLIPADST
fadst16x16_neon, // ADST_FLIPADST
fadst16x16_neon, // FLIPADST_ADST
idtx16x16_neon, // IDTX
idtx16x16_neon, // V_DCT
fdct16x16_neon, // H_DCT
idtx16x16_neon, // V_ADST
fadst16x16_neon, // H_ADST
idtx16x16_neon, // V_FLIPADST
fadst16x16_neon // H_FLIPADST
};
static const fwd_transform_1d_neon col_highbd_txfm8x16_arr[TX_TYPES] = {
fdct16x16_neon, // DCT_DCT
fadst16x16_neon, // ADST_DCT
fdct16x16_neon, // DCT_ADST
fadst16x16_neon, // ADST_ADST
fadst16x16_neon, // FLIPADST_DCT
fdct16x16_neon, // DCT_FLIPADST
fadst16x16_neon, // FLIPADST_FLIPADST
fadst16x16_neon, // ADST_FLIPADST
fadst16x16_neon, // FLIPADST_ADST
idtx16x16_neon, // IDTX
fdct16x16_neon, // V_DCT
idtx16x16_neon, // H_DCT
fadst16x16_neon, // V_ADST
idtx16x16_neon, // H_ADST
fadst16x16_neon, // V_FLIPADST
idtx16x16_neon // H_FLIPADST
};
static const fwd_transform_1d_neon row_highbd_txfm8x8_arr[TX_TYPES] = {
fdct8x8_neon, // DCT_DCT
fdct8x8_neon, // ADST_DCT
fadst8x8_neon, // DCT_ADST
fadst8x8_neon, // ADST_ADST
fdct8x8_neon, // FLIPADST_DCT
fadst8x8_neon, // DCT_FLIPADST
fadst8x8_neon, // FLIPADST_FLIPADST
fadst8x8_neon, // ADST_FLIPADST
fadst8x8_neon, // FLIPADST_ADST
idtx8x8_neon, // IDTX
idtx8x8_neon, // V_DCT
fdct8x8_neon, // H_DCT
idtx8x8_neon, // V_ADST
fadst8x8_neon, // H_ADST
idtx8x8_neon, // V_FLIPADST
fadst8x8_neon // H_FLIPADST
};
static const fwd_transform_1d_neon row_highbd_txfm4x8_arr[TX_TYPES] = {
fdct4x8_neon, // DCT_DCT
fdct4x8_neon, // ADST_DCT
fadst8x8_neon, // DCT_ADST
fadst8x8_neon, // ADST_ADST
fdct4x8_neon, // FLIPADST_DCT
fadst8x8_neon, // DCT_FLIPADST
fadst8x8_neon, // FLIPADST_FLIPADST
fadst8x8_neon, // ADST_FLIPADST
fadst8x8_neon, // FLIPADST_ADST
idtx8x8_neon, // IDTX
idtx8x8_neon, // V_DCT
fdct4x8_neon, // H_DCT
idtx8x8_neon, // V_ADST
fadst8x8_neon, // H_ADST
idtx8x8_neon, // V_FLIPADST
fadst8x8_neon // H_FLIPADST
};
static const fwd_transform_1d_neon row_highbd_txfm4x4_arr[TX_TYPES] = {
fdct4x4_neon, // DCT_DCT
fdct4x4_neon, // ADST_DCT
fadst4x4_neon, // DCT_ADST
fadst4x4_neon, // ADST_ADST
fdct4x4_neon, // FLIPADST_DCT
fadst4x4_neon, // DCT_FLIPADST
fadst4x4_neon, // FLIPADST_FLIPADST
fadst4x4_neon, // ADST_FLIPADST
fadst4x4_neon, // FLIPADST_ADST
idtx4x4_neon, // IDTX
idtx4x4_neon, // V_DCT
fdct4x4_neon, // H_DCT
idtx4x4_neon, // V_ADST
fadst4x4_neon, // H_ADST
idtx4x4_neon, // V_FLIPADST
fadst4x4_neon // H_FLIPADST
};
static const fwd_transform_1d_neon col_highbd_txfm4x4_arr[TX_TYPES] = {
fdct4x4_neon, // DCT_DCT
fadst4x4_neon, // ADST_DCT
fdct4x4_neon, // DCT_ADST
fadst4x4_neon, // ADST_ADST
fadst4x4_neon, // FLIPADST_DCT
fdct4x4_neon, // DCT_FLIPADST
fadst4x4_neon, // FLIPADST_FLIPADST
fadst4x4_neon, // ADST_FLIPADST
fadst4x4_neon, // FLIPADST_ADST
idtx4x4_neon, // IDTX
fdct4x4_neon, // V_DCT
idtx4x4_neon, // H_DCT
fadst4x4_neon, // V_ADST
idtx4x4_neon, // H_ADST
fadst4x4_neon, // V_FLIPADST
idtx4x4_neon // H_FLIPADST
};
void av1_fdct32_new_neon(int32x4_t *input, int32x4_t *output, int cos_bit,
const int stride) {
int32x4_t buf0[32];
int32x4_t buf1[32];
const int32_t *cospi;
const int32x4_t v_cos_bit = vdupq_n_s32(-cos_bit);
int startidx = 0 * stride;
int endidx = 31 * stride;
// stage 0
// stage 1
buf1[0] = vaddq_s32(input[startidx], input[endidx]);
buf1[31] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[1] = vaddq_s32(input[startidx], input[endidx]);
buf1[30] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[2] = vaddq_s32(input[startidx], input[endidx]);
buf1[29] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[3] = vaddq_s32(input[startidx], input[endidx]);
buf1[28] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[4] = vaddq_s32(input[startidx], input[endidx]);
buf1[27] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[5] = vaddq_s32(input[startidx], input[endidx]);
buf1[26] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[6] = vaddq_s32(input[startidx], input[endidx]);
buf1[25] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[7] = vaddq_s32(input[startidx], input[endidx]);
buf1[24] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[8] = vaddq_s32(input[startidx], input[endidx]);
buf1[23] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[9] = vaddq_s32(input[startidx], input[endidx]);
buf1[22] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[10] = vaddq_s32(input[startidx], input[endidx]);
buf1[21] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[11] = vaddq_s32(input[startidx], input[endidx]);
buf1[20] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[12] = vaddq_s32(input[startidx], input[endidx]);
buf1[19] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[13] = vaddq_s32(input[startidx], input[endidx]);
buf1[18] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[14] = vaddq_s32(input[startidx], input[endidx]);
buf1[17] = vsubq_s32(input[startidx], input[endidx]);
startidx += stride;
endidx -= stride;
buf1[15] = vaddq_s32(input[startidx], input[endidx]);
buf1[16] = vsubq_s32(input[startidx], input[endidx]);
// stage 2
cospi = cospi_arr(cos_bit);
buf0[0] = vaddq_s32(buf1[0], buf1[15]);
buf0[15] = vsubq_s32(buf1[0], buf1[15]);
buf0[1] = vaddq_s32(buf1[1], buf1[14]);
buf0[14] = vsubq_s32(buf1[1], buf1[14]);
buf0[2] = vaddq_s32(buf1[2], buf1[13]);
buf0[13] = vsubq_s32(buf1[2], buf1[13]);
buf0[3] = vaddq_s32(buf1[3], buf1[12]);
buf0[12] = vsubq_s32(buf1[3], buf1[12]);
buf0[4] = vaddq_s32(buf1[4], buf1[11]);
buf0[11] = vsubq_s32(buf1[4], buf1[11]);
buf0[5] = vaddq_s32(buf1[5], buf1[10]);
buf0[10] = vsubq_s32(buf1[5], buf1[10]);
buf0[6] = vaddq_s32(buf1[6], buf1[9]);
buf0[9] = vsubq_s32(buf1[6], buf1[9]);
buf0[7] = vaddq_s32(buf1[7], buf1[8]);
buf0[8] = vsubq_s32(buf1[7], buf1[8]);
buf0[16] = buf1[16];
buf0[17] = buf1[17];
buf0[18] = buf1[18];
buf0[19] = buf1[19];
btf_32_neon_type0(-cospi[32], cospi[32], buf1[20], buf1[27], buf0[20],
buf0[27], v_cos_bit);
btf_32_neon_type0(-cospi[32], cospi[32], buf1[21], buf1[26], buf0[21],
buf0[26], v_cos_bit);
btf_32_neon_type0(-cospi[32], cospi[32], buf1[22], buf1[25], buf0[22],