vpx/vpx_dsp/arm/fdct_neon.c
Johann 105503b839 neon fdct: 4x4 implementation
Approximately twice as fast as C implementation.

BUG=webm:1424

Change-Id: I3c0307fb08ddc23df42545cd089a78e2ed5c9d3f
2017-05-17 07:38:18 -07:00

93 lines
3.4 KiB
C

/*
* Copyright (c) 2017 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 "./vpx_config.h"
#include "vpx_dsp/txfm_common.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_dsp/arm/idct_neon.h"
#include "vpx_dsp/arm/mem_neon.h"
#include "vpx_dsp/arm/transpose_neon.h"
void vpx_fdct4x4_neon(const int16_t *input, tran_low_t *final_output,
int stride) {
int i;
// input[M * stride] * 16
int16x4_t input_0 = vshl_n_s16(vld1_s16(input + 0 * stride), 4);
int16x4_t input_1 = vshl_n_s16(vld1_s16(input + 1 * stride), 4);
int16x4_t input_2 = vshl_n_s16(vld1_s16(input + 2 * stride), 4);
int16x4_t input_3 = vshl_n_s16(vld1_s16(input + 3 * stride), 4);
// If the very first value != 0, then add 1.
if (input[0] != 0) {
const int16x4_t one = vreinterpret_s16_s64(vdup_n_s64(1));
input_0 = vadd_s16(input_0, one);
}
for (i = 0; i < 2; ++i) {
const int16x8_t input_01 = vcombine_s16(input_0, input_1);
const int16x8_t input_32 = vcombine_s16(input_3, input_2);
// in_0 +/- in_3, in_1 +/- in_2
const int16x8_t s_01 = vaddq_s16(input_01, input_32);
const int16x8_t s_32 = vsubq_s16(input_01, input_32);
// step_0 +/- step_1, step_2 +/- step_3
const int16x4_t s_0 = vget_low_s16(s_01);
const int16x4_t s_1 = vget_high_s16(s_01);
const int16x4_t s_2 = vget_high_s16(s_32);
const int16x4_t s_3 = vget_low_s16(s_32);
// (s_0 +/- s_1) * cospi_16_64
// Must expand all elements to s32. See 'needs32' comment in fwd_txfm.c.
const int32x4_t s_0_p_s_1 = vaddl_s16(s_0, s_1);
const int32x4_t s_0_m_s_1 = vsubl_s16(s_0, s_1);
const int32x4_t temp1 = vmulq_n_s32(s_0_p_s_1, (int16_t)cospi_16_64);
const int32x4_t temp2 = vmulq_n_s32(s_0_m_s_1, (int16_t)cospi_16_64);
// fdct_round_shift
int16x4_t out_0 = vrshrn_n_s32(temp1, DCT_CONST_BITS);
int16x4_t out_2 = vrshrn_n_s32(temp2, DCT_CONST_BITS);
// s_3 * cospi_8_64 + s_2 * cospi_24_64
// s_3 * cospi_24_64 - s_2 * cospi_8_64
const int32x4_t s_3_cospi_8_64 = vmull_n_s16(s_3, (int16_t)cospi_8_64);
const int32x4_t s_3_cospi_24_64 = vmull_n_s16(s_3, (int16_t)cospi_24_64);
const int32x4_t temp3 =
vmlal_n_s16(s_3_cospi_8_64, s_2, (int16_t)cospi_24_64);
const int32x4_t temp4 =
vmlsl_n_s16(s_3_cospi_24_64, s_2, (int16_t)cospi_8_64);
// fdct_round_shift
int16x4_t out_1 = vrshrn_n_s32(temp3, DCT_CONST_BITS);
int16x4_t out_3 = vrshrn_n_s32(temp4, DCT_CONST_BITS);
transpose_s16_4x4d(&out_0, &out_1, &out_2, &out_3);
input_0 = out_0;
input_1 = out_1;
input_2 = out_2;
input_3 = out_3;
}
{
// Not quite a rounding shift. Only add 1 despite shifting by 2.
const int16x8_t one = vdupq_n_s16(1);
int16x8_t out_01 = vcombine_s16(input_0, input_1);
int16x8_t out_23 = vcombine_s16(input_2, input_3);
out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2);
out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2);
store_s16q_to_tran_low(final_output + 0 * 8, out_01);
store_s16q_to_tran_low(final_output + 1 * 8, out_23);
}
}