ref: 394020383d3da02e602050f9d5e8e671a6e3cb8b
dir: /vpx_dsp/arm/idct_neon.h/
/*
* 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.
*/
#ifndef VPX_DSP_ARM_IDCT_NEON_H_
#define VPX_DSP_ARM_IDCT_NEON_H_
#include <arm_neon.h>
#include "./vpx_config.h"
#include "vpx_dsp/arm/transpose_neon.h"
#include "vpx_dsp/vpx_dsp_common.h"
DECLARE_ALIGNED(16, static const int16_t, kCospi[8]) = {
16384 /* cospi_0_64 */, 15137 /* cospi_8_64 */,
11585 /* cospi_16_64 */, 6270 /* cospi_24_64 */,
16069 /* cospi_4_64 */, 13623 /* cospi_12_64 */,
-9102 /* -cospi_20_64 */, 3196 /* cospi_28_64 */
};
//------------------------------------------------------------------------------
// Helper functions used to load tran_low_t into int16, narrowing if necessary.
static INLINE int16x8x2_t load_tran_low_to_s16x2q(const tran_low_t *buf) {
#if CONFIG_VP9_HIGHBITDEPTH
const int32x4x2_t v0 = vld2q_s32(buf);
const int32x4x2_t v1 = vld2q_s32(buf + 8);
const int16x4_t s0 = vmovn_s32(v0.val[0]);
const int16x4_t s1 = vmovn_s32(v0.val[1]);
const int16x4_t s2 = vmovn_s32(v1.val[0]);
const int16x4_t s3 = vmovn_s32(v1.val[1]);
int16x8x2_t res;
res.val[0] = vcombine_s16(s0, s2);
res.val[1] = vcombine_s16(s1, s3);
return res;
#else
return vld2q_s16(buf);
#endif
}
static INLINE int16x8_t load_tran_low_to_s16q(const tran_low_t *buf) {
#if CONFIG_VP9_HIGHBITDEPTH
const int32x4_t v0 = vld1q_s32(buf);
const int32x4_t v1 = vld1q_s32(buf + 4);
const int16x4_t s0 = vmovn_s32(v0);
const int16x4_t s1 = vmovn_s32(v1);
return vcombine_s16(s0, s1);
#else
return vld1q_s16(buf);
#endif
}
static INLINE int16x4_t load_tran_low_to_s16d(const tran_low_t *buf) {
#if CONFIG_VP9_HIGHBITDEPTH
const int32x4_t v0 = vld1q_s32(buf);
return vmovn_s32(v0);
#else
return vld1_s16(buf);
#endif
}
//------------------------------------------------------------------------------
// Multiply a by a_const. Saturate, shift and narrow by 14.
static INLINE int16x8_t multiply_shift_and_narrow_s16(const int16x8_t a,
const int16_t a_const) {
// Shift by 14 + rounding will be within 16 bits for well formed streams.
// See WRAPLOW and dct_const_round_shift for details.
// This instruction doubles the result and returns the high half, essentially
// resulting in a right shift by 15. By multiplying the constant first that
// becomes a right shift by 14.
// The largest possible value used here is
// vpx_dsp/txfm_common.h:cospi_1_64 = 16364 (* 2 = 32728) a which falls *just*
// within the range of int16_t (+32767 / -32768) even when negated.
return vqrdmulhq_n_s16(a, a_const * 2);
}
// Add a and b, then multiply by ab_const. Shift and narrow by 14.
static INLINE int16x8_t add_multiply_shift_and_narrow_s16(
const int16x8_t a, const int16x8_t b, const int16_t ab_const) {
// In both add_ and it's pair, sub_, the input for well-formed streams will be
// well within 16 bits (input to the idct is the difference between two frames
// and will be within -255 to 255, or 9 bits)
// However, for inputs over about 25,000 (valid for int16_t, but not for idct
// input) this function can not use vaddq_s16.
// In order to match existing behavior and intentionally out of range tests,
// expand the addition up to 32 bits to prevent truncation.
int32x4_t temp_low = vaddl_s16(vget_low_s16(a), vget_low_s16(b));
int32x4_t temp_high = vaddl_s16(vget_high_s16(a), vget_high_s16(b));
temp_low = vmulq_n_s32(temp_low, ab_const);
temp_high = vmulq_n_s32(temp_high, ab_const);
return vcombine_s16(vrshrn_n_s32(temp_low, 14), vrshrn_n_s32(temp_high, 14));
}
// Subtract b from a, then multiply by ab_const. Shift and narrow by 14.
static INLINE int16x8_t sub_multiply_shift_and_narrow_s16(
const int16x8_t a, const int16x8_t b, const int16_t ab_const) {
int32x4_t temp_low = vsubl_s16(vget_low_s16(a), vget_low_s16(b));
int32x4_t temp_high = vsubl_s16(vget_high_s16(a), vget_high_s16(b));
temp_low = vmulq_n_s32(temp_low, ab_const);
temp_high = vmulq_n_s32(temp_high, ab_const);
return vcombine_s16(vrshrn_n_s32(temp_low, 14), vrshrn_n_s32(temp_high, 14));
}
// Multiply a by a_const and b by b_const, then accumulate. Shift and narrow by
// 14.
static INLINE int16x8_t multiply_accumulate_shift_and_narrow_s16(
const int16x8_t a, const int16_t a_const, const int16x8_t b,
const int16_t b_const) {
int32x4_t temp_low = vmull_n_s16(vget_low_s16(a), a_const);
int32x4_t temp_high = vmull_n_s16(vget_high_s16(a), a_const);
temp_low = vmlal_n_s16(temp_low, vget_low_s16(b), b_const);
temp_high = vmlal_n_s16(temp_high, vget_high_s16(b), b_const);
return vcombine_s16(vrshrn_n_s32(temp_low, 14), vrshrn_n_s32(temp_high, 14));
}
static INLINE void load_and_transpose_s16_8x8(const int16_t *a, int a_stride,
int16x8_t *a0, int16x8_t *a1,
int16x8_t *a2, int16x8_t *a3,
int16x8_t *a4, int16x8_t *a5,
int16x8_t *a6, int16x8_t *a7) {
*a0 = vld1q_s16(a);
a += a_stride;
*a1 = vld1q_s16(a);
a += a_stride;
*a2 = vld1q_s16(a);
a += a_stride;
*a3 = vld1q_s16(a);
a += a_stride;
*a4 = vld1q_s16(a);
a += a_stride;
*a5 = vld1q_s16(a);
a += a_stride;
*a6 = vld1q_s16(a);
a += a_stride;
*a7 = vld1q_s16(a);
transpose_s16_8x8(a0, a1, a2, a3, a4, a5, a6, a7);
}
// Shift the output down by 6 and add it to the destination buffer.
static INLINE void add_and_store_u8_s16(const int16x8_t a0, const int16x8_t a1,
const int16x8_t a2, const int16x8_t a3,
const int16x8_t a4, const int16x8_t a5,
const int16x8_t a6, const int16x8_t a7,
uint8_t *b, const int b_stride) {
uint8x8_t b0, b1, b2, b3, b4, b5, b6, b7;
int16x8_t c0, c1, c2, c3, c4, c5, c6, c7;
b0 = vld1_u8(b);
b += b_stride;
b1 = vld1_u8(b);
b += b_stride;
b2 = vld1_u8(b);
b += b_stride;
b3 = vld1_u8(b);
b += b_stride;
b4 = vld1_u8(b);
b += b_stride;
b5 = vld1_u8(b);
b += b_stride;
b6 = vld1_u8(b);
b += b_stride;
b7 = vld1_u8(b);
b -= (7 * b_stride);
// c = b + (a >> 6)
c0 = vrsraq_n_s16(vreinterpretq_s16_u16(vmovl_u8(b0)), a0, 6);
c1 = vrsraq_n_s16(vreinterpretq_s16_u16(vmovl_u8(b1)), a1, 6);
c2 = vrsraq_n_s16(vreinterpretq_s16_u16(vmovl_u8(b2)), a2, 6);
c3 = vrsraq_n_s16(vreinterpretq_s16_u16(vmovl_u8(b3)), a3, 6);
c4 = vrsraq_n_s16(vreinterpretq_s16_u16(vmovl_u8(b4)), a4, 6);
c5 = vrsraq_n_s16(vreinterpretq_s16_u16(vmovl_u8(b5)), a5, 6);
c6 = vrsraq_n_s16(vreinterpretq_s16_u16(vmovl_u8(b6)), a6, 6);
c7 = vrsraq_n_s16(vreinterpretq_s16_u16(vmovl_u8(b7)), a7, 6);
b0 = vqmovun_s16(c0);
b1 = vqmovun_s16(c1);
b2 = vqmovun_s16(c2);
b3 = vqmovun_s16(c3);
b4 = vqmovun_s16(c4);
b5 = vqmovun_s16(c5);
b6 = vqmovun_s16(c6);
b7 = vqmovun_s16(c7);
vst1_u8(b, b0);
b += b_stride;
vst1_u8(b, b1);
b += b_stride;
vst1_u8(b, b2);
b += b_stride;
vst1_u8(b, b3);
b += b_stride;
vst1_u8(b, b4);
b += b_stride;
vst1_u8(b, b5);
b += b_stride;
vst1_u8(b, b6);
b += b_stride;
vst1_u8(b, b7);
}
static INLINE void idct4x4_16_kernel_bd8(const int16x4_t cospis,
int16x8_t *const a0,
int16x8_t *const a1) {
int16x4_t b0, b1, b2, b3;
int32x4_t c0, c1, c2, c3;
int16x8_t d0, d1;
transpose_s16_4x4q(a0, a1);
b0 = vget_low_s16(*a0);
b1 = vget_high_s16(*a0);
b2 = vget_low_s16(*a1);
b3 = vget_high_s16(*a1);
c0 = vmull_lane_s16(b0, cospis, 2);
c2 = vmull_lane_s16(b1, cospis, 2);
c1 = vsubq_s32(c0, c2);
c0 = vaddq_s32(c0, c2);
c2 = vmull_lane_s16(b2, cospis, 3);
c3 = vmull_lane_s16(b2, cospis, 1);
c2 = vmlsl_lane_s16(c2, b3, cospis, 1);
c3 = vmlal_lane_s16(c3, b3, cospis, 3);
b0 = vrshrn_n_s32(c0, 14);
b1 = vrshrn_n_s32(c1, 14);
b2 = vrshrn_n_s32(c2, 14);
b3 = vrshrn_n_s32(c3, 14);
d0 = vcombine_s16(b0, b1);
d1 = vcombine_s16(b3, b2);
*a0 = vaddq_s16(d0, d1);
*a1 = vsubq_s16(d0, d1);
}
#endif // VPX_DSP_ARM_IDCT_NEON_H_