diff options
author | Urvang Joshi <urvang@google.com> | 2018-05-09 15:04:31 -0400 |
---|---|---|
committer | Urvang Joshi <urvang@google.com> | 2018-05-10 18:53:52 +0000 |
commit | 698720b9fc22747788dcaba9a08c79a6dc873b04 (patch) | |
tree | 31d486266af66963e42e3a2526ffd9da14c83411 /aom_dsp/arm | |
parent | 6eb623047128b05e8cdd0b4768392d376429130b (diff) |
Further remove old transform code
*iht*, *fht*, aom_{highbd_}fdct, aom_{highbd_}fadst, aom_{highbd_}idct,
aom_{highbd_}iadst etc.
Reduces libaom.so size:
Before: 7322288 bytes
After: 7121552 bytes
Change-Id: I803a7700b330577f921daad2dc230c7c3b6a5da7
Diffstat (limited to 'aom_dsp/arm')
-rw-r--r-- | aom_dsp/arm/idct16x16_1_add_neon.asm | 201 | ||||
-rw-r--r-- | aom_dsp/arm/idct16x16_1_add_neon.c | 59 | ||||
-rw-r--r-- | aom_dsp/arm/idct16x16_add_neon.asm | 1182 | ||||
-rw-r--r-- | aom_dsp/arm/idct16x16_add_neon.c | 1295 | ||||
-rw-r--r-- | aom_dsp/arm/idct16x16_neon.c | 152 | ||||
-rw-r--r-- | aom_dsp/arm/idct32x32_1_add_neon.asm | 147 | ||||
-rw-r--r-- | aom_dsp/arm/idct32x32_1_add_neon.c | 141 | ||||
-rw-r--r-- | aom_dsp/arm/idct32x32_add_neon.asm | 1302 | ||||
-rw-r--r-- | aom_dsp/arm/idct32x32_add_neon.c | 686 | ||||
-rw-r--r-- | aom_dsp/arm/idct4x4_1_add_neon.asm | 71 | ||||
-rw-r--r-- | aom_dsp/arm/idct4x4_1_add_neon.c | 47 | ||||
-rw-r--r-- | aom_dsp/arm/idct4x4_add_neon.asm | 193 | ||||
-rw-r--r-- | aom_dsp/arm/idct4x4_add_neon.c | 146 | ||||
-rw-r--r-- | aom_dsp/arm/idct8x8_1_add_neon.asm | 91 | ||||
-rw-r--r-- | aom_dsp/arm/idct8x8_1_add_neon.c | 62 | ||||
-rw-r--r-- | aom_dsp/arm/idct8x8_add_neon.asm | 522 | ||||
-rw-r--r-- | aom_dsp/arm/idct8x8_add_neon.c | 509 |
17 files changed, 0 insertions, 6806 deletions
diff --git a/aom_dsp/arm/idct16x16_1_add_neon.asm b/aom_dsp/arm/idct16x16_1_add_neon.asm deleted file mode 100644 index d01c4bc03..000000000 --- a/aom_dsp/arm/idct16x16_1_add_neon.asm +++ /dev/null @@ -1,201 +0,0 @@ -; -; Copyright (c) 2016, 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. -; - - - - EXPORT |aom_idct16x16_1_add_neon| - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - -;void aom_idct16x16_1_add_neon(int16_t *input, uint8_t *dest, -; int dest_stride) -; -; r0 int16_t input -; r1 uint8_t *dest -; r2 int dest_stride) - -|aom_idct16x16_1_add_neon| PROC - ldrsh r0, [r0] - - ; generate cospi_16_64 = 11585 - mov r12, #0x2d00 - add r12, #0x41 - - ; out = dct_const_round_shift(input[0] * cospi_16_64) - mul r0, r0, r12 ; input[0] * cospi_16_64 - add r0, r0, #0x2000 ; +(1 << ((DCT_CONST_BITS) - 1)) - asr r0, r0, #14 ; >> DCT_CONST_BITS - - ; out = dct_const_round_shift(out * cospi_16_64) - mul r0, r0, r12 ; out * cospi_16_64 - mov r12, r1 ; save dest - add r0, r0, #0x2000 ; +(1 << ((DCT_CONST_BITS) - 1)) - asr r0, r0, #14 ; >> DCT_CONST_BITS - - ; a1 = ROUND_POWER_OF_TWO(out, 6) - add r0, r0, #32 ; + (1 <<((6) - 1)) - asr r0, r0, #6 ; >> 6 - - vdup.s16 q0, r0 ; duplicate a1 - mov r0, #8 - sub r2, #8 - - ; load destination data row0 - row3 - vld1.64 {d2}, [r1], r0 - vld1.64 {d3}, [r1], r2 - vld1.64 {d4}, [r1], r0 - vld1.64 {d5}, [r1], r2 - vld1.64 {d6}, [r1], r0 - vld1.64 {d7}, [r1], r2 - vld1.64 {d16}, [r1], r0 - vld1.64 {d17}, [r1], r2 - - vaddw.u8 q9, q0, d2 ; dest[x] + a1 - vaddw.u8 q10, q0, d3 ; dest[x] + a1 - vaddw.u8 q11, q0, d4 ; dest[x] + a1 - vaddw.u8 q12, q0, d5 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r0 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r0 - vst1.64 {d31}, [r12], r2 - - vaddw.u8 q9, q0, d6 ; dest[x] + a1 - vaddw.u8 q10, q0, d7 ; dest[x] + a1 - vaddw.u8 q11, q0, d16 ; dest[x] + a1 - vaddw.u8 q12, q0, d17 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r0 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r0 - vst1.64 {d31}, [r12], r2 - - ; load destination data row4 - row7 - vld1.64 {d2}, [r1], r0 - vld1.64 {d3}, [r1], r2 - vld1.64 {d4}, [r1], r0 - vld1.64 {d5}, [r1], r2 - vld1.64 {d6}, [r1], r0 - vld1.64 {d7}, [r1], r2 - vld1.64 {d16}, [r1], r0 - vld1.64 {d17}, [r1], r2 - - vaddw.u8 q9, q0, d2 ; dest[x] + a1 - vaddw.u8 q10, q0, d3 ; dest[x] + a1 - vaddw.u8 q11, q0, d4 ; dest[x] + a1 - vaddw.u8 q12, q0, d5 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r0 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r0 - vst1.64 {d31}, [r12], r2 - - vaddw.u8 q9, q0, d6 ; dest[x] + a1 - vaddw.u8 q10, q0, d7 ; dest[x] + a1 - vaddw.u8 q11, q0, d16 ; dest[x] + a1 - vaddw.u8 q12, q0, d17 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r0 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r0 - vst1.64 {d31}, [r12], r2 - - ; load destination data row8 - row11 - vld1.64 {d2}, [r1], r0 - vld1.64 {d3}, [r1], r2 - vld1.64 {d4}, [r1], r0 - vld1.64 {d5}, [r1], r2 - vld1.64 {d6}, [r1], r0 - vld1.64 {d7}, [r1], r2 - vld1.64 {d16}, [r1], r0 - vld1.64 {d17}, [r1], r2 - - vaddw.u8 q9, q0, d2 ; dest[x] + a1 - vaddw.u8 q10, q0, d3 ; dest[x] + a1 - vaddw.u8 q11, q0, d4 ; dest[x] + a1 - vaddw.u8 q12, q0, d5 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r0 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r0 - vst1.64 {d31}, [r12], r2 - - vaddw.u8 q9, q0, d6 ; dest[x] + a1 - vaddw.u8 q10, q0, d7 ; dest[x] + a1 - vaddw.u8 q11, q0, d16 ; dest[x] + a1 - vaddw.u8 q12, q0, d17 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r0 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r0 - vst1.64 {d31}, [r12], r2 - - ; load destination data row12 - row15 - vld1.64 {d2}, [r1], r0 - vld1.64 {d3}, [r1], r2 - vld1.64 {d4}, [r1], r0 - vld1.64 {d5}, [r1], r2 - vld1.64 {d6}, [r1], r0 - vld1.64 {d7}, [r1], r2 - vld1.64 {d16}, [r1], r0 - vld1.64 {d17}, [r1], r2 - - vaddw.u8 q9, q0, d2 ; dest[x] + a1 - vaddw.u8 q10, q0, d3 ; dest[x] + a1 - vaddw.u8 q11, q0, d4 ; dest[x] + a1 - vaddw.u8 q12, q0, d5 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r0 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r0 - vst1.64 {d31}, [r12], r2 - - vaddw.u8 q9, q0, d6 ; dest[x] + a1 - vaddw.u8 q10, q0, d7 ; dest[x] + a1 - vaddw.u8 q11, q0, d16 ; dest[x] + a1 - vaddw.u8 q12, q0, d17 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r0 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r0 - vst1.64 {d31}, [r12], r2 - - bx lr - ENDP ; |aom_idct16x16_1_add_neon| - - END diff --git a/aom_dsp/arm/idct16x16_1_add_neon.c b/aom_dsp/arm/idct16x16_1_add_neon.c deleted file mode 100644 index 196b2a890..000000000 --- a/aom_dsp/arm/idct16x16_1_add_neon.c +++ /dev/null @@ -1,59 +0,0 @@ -/* - * Copyright (c) 2016, 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 "aom_dsp/inv_txfm.h" -#include "aom_ports/mem.h" - -void aom_idct16x16_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) { - uint8x8_t d2u8, d3u8, d30u8, d31u8; - uint64x1_t d2u64, d3u64, d4u64, d5u64; - uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16; - int16x8_t q0s16; - uint8_t *d1, *d2; - int16_t i, j, a1; - int16_t out = dct_const_round_shift(input[0] * cospi_16_64); - out = dct_const_round_shift(out * cospi_16_64); - a1 = ROUND_POWER_OF_TWO(out, 6); - - q0s16 = vdupq_n_s16(a1); - q0u16 = vreinterpretq_u16_s16(q0s16); - - for (d1 = d2 = dest, i = 0; i < 4; i++) { - for (j = 0; j < 2; j++) { - d2u64 = vld1_u64((const uint64_t *)d1); - d3u64 = vld1_u64((const uint64_t *)(d1 + 8)); - d1 += dest_stride; - d4u64 = vld1_u64((const uint64_t *)d1); - d5u64 = vld1_u64((const uint64_t *)(d1 + 8)); - d1 += dest_stride; - - q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64)); - q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64)); - q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64)); - q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64)); - - d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16)); - d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16)); - d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16)); - d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16)); - - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8)); - vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d3u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8)); - vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d31u8)); - d2 += dest_stride; - } - } - return; -} diff --git a/aom_dsp/arm/idct16x16_add_neon.asm b/aom_dsp/arm/idct16x16_add_neon.asm deleted file mode 100644 index 4a8f8f183..000000000 --- a/aom_dsp/arm/idct16x16_add_neon.asm +++ /dev/null @@ -1,1182 +0,0 @@ -; -; Copyright (c) 2016, 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. -; - -; - - EXPORT |aom_idct16x16_256_add_neon_pass1| - EXPORT |aom_idct16x16_256_add_neon_pass2| - EXPORT |aom_idct16x16_10_add_neon_pass1| - EXPORT |aom_idct16x16_10_add_neon_pass2| - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - - ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15. - MACRO - TRANSPOSE8X8 - vswp d17, d24 - vswp d23, d30 - vswp d21, d28 - vswp d19, d26 - vtrn.32 q8, q10 - vtrn.32 q9, q11 - vtrn.32 q12, q14 - vtrn.32 q13, q15 - vtrn.16 q8, q9 - vtrn.16 q10, q11 - vtrn.16 q12, q13 - vtrn.16 q14, q15 - MEND - - AREA Block, CODE, READONLY ; name this block of code -;void |aom_idct16x16_256_add_neon_pass1|(int16_t *input, -; int16_t *output, int output_stride) -; -; r0 int16_t input -; r1 int16_t *output -; r2 int output_stride) - -; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output -; will be stored back into q8-q15 registers. This function will touch q0-q7 -; registers and use them as buffer during calculation. -|aom_idct16x16_256_add_neon_pass1| PROC - - ; TODO(hkuang): Find a better way to load the elements. - ; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15 - vld2.s16 {q8,q9}, [r0]! - vld2.s16 {q9,q10}, [r0]! - vld2.s16 {q10,q11}, [r0]! - vld2.s16 {q11,q12}, [r0]! - vld2.s16 {q12,q13}, [r0]! - vld2.s16 {q13,q14}, [r0]! - vld2.s16 {q14,q15}, [r0]! - vld2.s16 {q1,q2}, [r0]! - vmov.s16 q15, q1 - - ; generate cospi_28_64 = 3196 - mov r3, #0xc00 - add r3, #0x7c - - ; generate cospi_4_64 = 16069 - mov r12, #0x3e00 - add r12, #0xc5 - - ; transpose the input data - TRANSPOSE8X8 - - ; stage 3 - vdup.16 d0, r3 ; duplicate cospi_28_64 - vdup.16 d1, r12 ; duplicate cospi_4_64 - - ; preloading to avoid stall - ; generate cospi_12_64 = 13623 - mov r3, #0x3500 - add r3, #0x37 - - ; generate cospi_20_64 = 9102 - mov r12, #0x2300 - add r12, #0x8e - - ; step2[4] * cospi_28_64 - vmull.s16 q2, d18, d0 - vmull.s16 q3, d19, d0 - - ; step2[4] * cospi_4_64 - vmull.s16 q5, d18, d1 - vmull.s16 q6, d19, d1 - - ; temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64 - vmlsl.s16 q2, d30, d1 - vmlsl.s16 q3, d31, d1 - - ; temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64 - vmlal.s16 q5, d30, d0 - vmlal.s16 q6, d31, d0 - - vdup.16 d2, r3 ; duplicate cospi_12_64 - vdup.16 d3, r12 ; duplicate cospi_20_64 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d8, q2, #14 ; >> 14 - vqrshrn.s32 d9, q3, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d14, q5, #14 ; >> 14 - vqrshrn.s32 d15, q6, #14 ; >> 14 - - ; preloading to avoid stall - ; generate cospi_16_64 = 11585 - mov r3, #0x2d00 - add r3, #0x41 - - ; generate cospi_24_64 = 6270 - mov r12, #0x1800 - add r12, #0x7e - - ; step2[5] * cospi_12_64 - vmull.s16 q2, d26, d2 - vmull.s16 q3, d27, d2 - - ; step2[5] * cospi_20_64 - vmull.s16 q9, d26, d3 - vmull.s16 q15, d27, d3 - - ; temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64 - vmlsl.s16 q2, d22, d3 - vmlsl.s16 q3, d23, d3 - - ; temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64 - vmlal.s16 q9, d22, d2 - vmlal.s16 q15, d23, d2 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d10, q2, #14 ; >> 14 - vqrshrn.s32 d11, q3, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d12, q9, #14 ; >> 14 - vqrshrn.s32 d13, q15, #14 ; >> 14 - - ; stage 4 - vdup.16 d30, r3 ; cospi_16_64 - - ; step1[0] * cospi_16_64 - vmull.s16 q2, d16, d30 - vmull.s16 q11, d17, d30 - - ; step1[1] * cospi_16_64 - vmull.s16 q0, d24, d30 - vmull.s16 q1, d25, d30 - - ; generate cospi_8_64 = 15137 - mov r3, #0x3b00 - add r3, #0x21 - - vdup.16 d30, r12 ; duplicate cospi_24_64 - vdup.16 d31, r3 ; duplicate cospi_8_64 - - ; temp1 = (step1[0] + step1[1]) * cospi_16_64 - vadd.s32 q3, q2, q0 - vadd.s32 q12, q11, q1 - - ; temp2 = (step1[0] - step1[1]) * cospi_16_64 - vsub.s32 q13, q2, q0 - vsub.s32 q1, q11, q1 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d16, q3, #14 ; >> 14 - vqrshrn.s32 d17, q12, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d18, q13, #14 ; >> 14 - vqrshrn.s32 d19, q1, #14 ; >> 14 - - ; step1[2] * cospi_24_64 - step1[3] * cospi_8_64; - ; step1[2] * cospi_8_64 - vmull.s16 q0, d20, d31 - vmull.s16 q1, d21, d31 - - ; step1[2] * cospi_24_64 - vmull.s16 q12, d20, d30 - vmull.s16 q13, d21, d30 - - ; temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64 - vmlal.s16 q0, d28, d30 - vmlal.s16 q1, d29, d30 - - ; temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64 - vmlsl.s16 q12, d28, d31 - vmlsl.s16 q13, d29, d31 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d22, q0, #14 ; >> 14 - vqrshrn.s32 d23, q1, #14 ; >> 14 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d20, q12, #14 ; >> 14 - vqrshrn.s32 d21, q13, #14 ; >> 14 - - vsub.s16 q13, q4, q5 ; step2[5] = step1[4] - step1[5]; - vadd.s16 q4, q4, q5 ; step2[4] = step1[4] + step1[5]; - vsub.s16 q14, q7, q6 ; step2[6] = -step1[6] + step1[7]; - vadd.s16 q15, q6, q7 ; step2[7] = step1[6] + step1[7]; - - ; generate cospi_16_64 = 11585 - mov r3, #0x2d00 - add r3, #0x41 - - ; stage 5 - vadd.s16 q0, q8, q11 ; step1[0] = step2[0] + step2[3]; - vadd.s16 q1, q9, q10 ; step1[1] = step2[1] + step2[2]; - vsub.s16 q2, q9, q10 ; step1[2] = step2[1] - step2[2]; - vsub.s16 q3, q8, q11 ; step1[3] = step2[0] - step2[3]; - - vdup.16 d16, r3; ; duplicate cospi_16_64 - - ; step2[5] * cospi_16_64 - vmull.s16 q11, d26, d16 - vmull.s16 q12, d27, d16 - - ; step2[6] * cospi_16_64 - vmull.s16 q9, d28, d16 - vmull.s16 q10, d29, d16 - - ; temp1 = (step2[6] - step2[5]) * cospi_16_64 - vsub.s32 q6, q9, q11 - vsub.s32 q13, q10, q12 - - ; temp2 = (step2[5] + step2[6]) * cospi_16_64 - vadd.s32 q9, q9, q11 - vadd.s32 q10, q10, q12 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d10, q6, #14 ; >> 14 - vqrshrn.s32 d11, q13, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d12, q9, #14 ; >> 14 - vqrshrn.s32 d13, q10, #14 ; >> 14 - - ; stage 6 - vadd.s16 q8, q0, q15 ; step2[0] = step1[0] + step1[7]; - vadd.s16 q9, q1, q6 ; step2[1] = step1[1] + step1[6]; - vadd.s16 q10, q2, q5 ; step2[2] = step1[2] + step1[5]; - vadd.s16 q11, q3, q4 ; step2[3] = step1[3] + step1[4]; - vsub.s16 q12, q3, q4 ; step2[4] = step1[3] - step1[4]; - vsub.s16 q13, q2, q5 ; step2[5] = step1[2] - step1[5]; - vsub.s16 q14, q1, q6 ; step2[6] = step1[1] - step1[6]; - vsub.s16 q15, q0, q15 ; step2[7] = step1[0] - step1[7]; - - ; store the data - vst1.64 {d16}, [r1], r2 - vst1.64 {d17}, [r1], r2 - vst1.64 {d18}, [r1], r2 - vst1.64 {d19}, [r1], r2 - vst1.64 {d20}, [r1], r2 - vst1.64 {d21}, [r1], r2 - vst1.64 {d22}, [r1], r2 - vst1.64 {d23}, [r1], r2 - vst1.64 {d24}, [r1], r2 - vst1.64 {d25}, [r1], r2 - vst1.64 {d26}, [r1], r2 - vst1.64 {d27}, [r1], r2 - vst1.64 {d28}, [r1], r2 - vst1.64 {d29}, [r1], r2 - vst1.64 {d30}, [r1], r2 - vst1.64 {d31}, [r1], r2 - - bx lr - ENDP ; |aom_idct16x16_256_add_neon_pass1| - -;void aom_idct16x16_256_add_neon_pass2(int16_t *src, -; int16_t *output, -; int16_t *pass1Output, -; int16_t skip_adding, -; uint8_t *dest, -; int dest_stride) -; -; r0 int16_t *src -; r1 int16_t *output, -; r2 int16_t *pass1Output, -; r3 int16_t skip_adding, -; r4 uint8_t *dest, -; r5 int dest_stride) - -; idct16 stage1 - stage7 on all the elements loaded in q8-q15. The output -; will be stored back into q8-q15 registers. This function will touch q0-q7 -; registers and use them as buffer during calculation. -|aom_idct16x16_256_add_neon_pass2| PROC - push {r3-r9} - - ; TODO(hkuang): Find a better way to load the elements. - ; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15 - vld2.s16 {q8,q9}, [r0]! - vld2.s16 {q9,q10}, [r0]! - vld2.s16 {q10,q11}, [r0]! - vld2.s16 {q11,q12}, [r0]! - vld2.s16 {q12,q13}, [r0]! - vld2.s16 {q13,q14}, [r0]! - vld2.s16 {q14,q15}, [r0]! - vld2.s16 {q0,q1}, [r0]! - vmov.s16 q15, q0; - - ; generate cospi_30_64 = 1606 - mov r3, #0x0600 - add r3, #0x46 - - ; generate cospi_2_64 = 16305 - mov r12, #0x3f00 - add r12, #0xb1 - - ; transpose the input data - TRANSPOSE8X8 - - ; stage 3 - vdup.16 d12, r3 ; duplicate cospi_30_64 - vdup.16 d13, r12 ; duplicate cospi_2_64 - - ; preloading to avoid stall - ; generate cospi_14_64 = 12665 - mov r3, #0x3100 - add r3, #0x79 - - ; generate cospi_18_64 = 10394 - mov r12, #0x2800 - add r12, #0x9a - - ; step1[8] * cospi_30_64 - vmull.s16 q2, d16, d12 - vmull.s16 q3, d17, d12 - - ; step1[8] * cospi_2_64 - vmull.s16 q1, d16, d13 - vmull.s16 q4, d17, d13 - - ; temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64 - vmlsl.s16 q2, d30, d13 - vmlsl.s16 q3, d31, d13 - - ; temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64 - vmlal.s16 q1, d30, d12 - vmlal.s16 q4, d31, d12 - - vdup.16 d30, r3 ; duplicate cospi_14_64 - vdup.16 d31, r12 ; duplicate cospi_18_64 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d0, q2, #14 ; >> 14 - vqrshrn.s32 d1, q3, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d14, q1, #14 ; >> 14 - vqrshrn.s32 d15, q4, #14 ; >> 14 - - ; preloading to avoid stall - ; generate cospi_22_64 = 7723 - mov r3, #0x1e00 - add r3, #0x2b - - ; generate cospi_10_64 = 14449 - mov r12, #0x3800 - add r12, #0x71 - - ; step1[9] * cospi_14_64 - vmull.s16 q2, d24, d30 - vmull.s16 q3, d25, d30 - - ; step1[9] * cospi_18_64 - vmull.s16 q4, d24, d31 - vmull.s16 q5, d25, d31 - - ; temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64 - vmlsl.s16 q2, d22, d31 - vmlsl.s16 q3, d23, d31 - - ; temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64 - vmlal.s16 q4, d22, d30 - vmlal.s16 q5, d23, d30 - - vdup.16 d30, r3 ; duplicate cospi_22_64 - vdup.16 d31, r12 ; duplicate cospi_10_64 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d2, q2, #14 ; >> 14 - vqrshrn.s32 d3, q3, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d12, q4, #14 ; >> 14 - vqrshrn.s32 d13, q5, #14 ; >> 14 - - ; step1[10] * cospi_22_64 - vmull.s16 q11, d20, d30 - vmull.s16 q12, d21, d30 - - ; step1[10] * cospi_10_64 - vmull.s16 q4, d20, d31 - vmull.s16 q5, d21, d31 - - ; temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64 - vmlsl.s16 q11, d26, d31 - vmlsl.s16 q12, d27, d31 - - ; temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64 - vmlal.s16 q4, d26, d30 - vmlal.s16 q5, d27, d30 - - ; preloading to avoid stall - ; generate cospi_6_64 = 15679 - mov r3, #0x3d00 - add r3, #0x3f - - ; generate cospi_26_64 = 4756 - mov r12, #0x1200 - add r12, #0x94 - - vdup.16 d30, r3 ; duplicate cospi_6_64 - vdup.16 d31, r12 ; duplicate cospi_26_64 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d4, q11, #14 ; >> 14 - vqrshrn.s32 d5, q12, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d11, q5, #14 ; >> 14 - vqrshrn.s32 d10, q4, #14 ; >> 14 - - ; step1[11] * cospi_6_64 - vmull.s16 q10, d28, d30 - vmull.s16 q11, d29, d30 - - ; step1[11] * cospi_26_64 - vmull.s16 q12, d28, d31 - vmull.s16 q13, d29, d31 - - ; temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64 - vmlsl.s16 q10, d18, d31 - vmlsl.s16 q11, d19, d31 - - ; temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64 - vmlal.s16 q12, d18, d30 - vmlal.s16 q13, d19, d30 - - vsub.s16 q9, q0, q1 ; step1[9]=step2[8]-step2[9] - vadd.s16 q0, q0, q1 ; step1[8]=step2[8]+step2[9] - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d6, q10, #14 ; >> 14 - vqrshrn.s32 d7, q11, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d8, q12, #14 ; >> 14 - vqrshrn.s32 d9, q13, #14 ; >> 14 - - ; stage 3 - vsub.s16 q10, q3, q2 ; step1[10]=-step2[10]+step2[11] - vadd.s16 q11, q2, q3 ; step1[11]=step2[10]+step2[11] - vadd.s16 q12, q4, q5 ; step1[12]=step2[12]+step2[13] - vsub.s16 q13, q4, q5 ; step1[13]=step2[12]-step2[13] - vsub.s16 q14, q7, q6 ; step1[14]=-step2[14]+tep2[15] - vadd.s16 q7, q6, q7 ; step1[15]=step2[14]+step2[15] - - ; stage 4 - ; generate cospi_24_64 = 6270 - mov r3, #0x1800 - add r3, #0x7e - - ; generate cospi_8_64 = 15137 - mov r12, #0x3b00 - add r12, #0x21 - - ; -step1[9] * cospi_8_64 + step1[14] * cospi_24_64 - vdup.16 d30, r12 ; duplicate cospi_8_64 - vdup.16 d31, r3 ; duplicate cospi_24_64 - - ; step1[9] * cospi_24_64 - vmull.s16 q2, d18, d31 - vmull.s16 q3, d19, d31 - - ; step1[14] * cospi_24_64 - vmull.s16 q4, d28, d31 - vmull.s16 q5, d29, d31 - - ; temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64 - vmlal.s16 q2, d28, d30 - vmlal.s16 q3, d29, d30 - - ; temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64 - vmlsl.s16 q4, d18, d30 - vmlsl.s16 q5, d19, d30 - - rsb r12, #0 - vdup.16 d30, r12 ; duplicate -cospi_8_64 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d12, q2, #14 ; >> 14 - vqrshrn.s32 d13, q3, #14 ; >> 14 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d2, q4, #14 ; >> 14 - vqrshrn.s32 d3, q5, #14 ; >> 14 - - vmov.s16 q3, q11 - vmov.s16 q4, q12 - - ; - step1[13] * cospi_8_64 - vmull.s16 q11, d26, d30 - vmull.s16 q12, d27, d30 - - ; -step1[10] * cospi_8_64 - vmull.s16 q8, d20, d30 - vmull.s16 q9, d21, d30 - - ; temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64 - vmlsl.s16 q11, d20, d31 - vmlsl.s16 q12, d21, d31 - - ; temp1 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64 - vmlal.s16 q8, d26, d31 - vmlal.s16 q9, d27, d31 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d4, q11, #14 ; >> 14 - vqrshrn.s32 d5, q12, #14 ; >> 14 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d10, q8, #14 ; >> 14 - vqrshrn.s32 d11, q9, #14 ; >> 14 - - ; stage 5 - vadd.s16 q8, q0, q3 ; step1[8] = step2[8]+step2[11]; - vadd.s16 q9, q1, q2 ; step1[9] = step2[9]+step2[10]; - vsub.s16 q10, q1, q2 ; step1[10] = step2[9]-step2[10]; - vsub.s16 q11, q0, q3 ; step1[11] = step2[8]-step2[11]; - vsub.s16 q12, q7, q4 ; step1[12] =-step2[12]+step2[15]; - vsub.s16 q13, q6, q5 ; step1[13] =-step2[13]+step2[14]; - vadd.s16 q14, q6, q5 ; step1[14] =step2[13]+step2[14]; - vadd.s16 q15, q7, q4 ; step1[15] =step2[12]+step2[15]; - - ; stage 6. - ; generate cospi_16_64 = 11585 - mov r12, #0x2d00 - add r12, #0x41 - - vdup.16 d14, r12 ; duplicate cospi_16_64 - - ; step1[13] * cospi_16_64 - vmull.s16 q3, d26, d14 - vmull.s16 q4, d27, d14 - - ; step1[10] * cospi_16_64 - vmull.s16 q0, d20, d14 - vmull.s16 q1, d21, d14 - - ; temp1 = (-step1[10] + step1[13]) * cospi_16_64 - vsub.s32 q5, q3, q0 - vsub.s32 q6, q4, q1 - - ; temp2 = (step1[10] + step1[13]) * cospi_16_64 - vadd.s32 q10, q3, q0 - vadd.s32 q4, q4, q1 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d4, q5, #14 ; >> 14 - vqrshrn.s32 d5, q6, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d10, q10, #14 ; >> 14 - vqrshrn.s32 d11, q4, #14 ; >> 14 - - ; step1[11] * cospi_16_64 - vmull.s16 q0, d22, d14 - vmull.s16 q1, d23, d14 - - ; step1[12] * cospi_16_64 - vmull.s16 q13, d24, d14 - vmull.s16 q6, d25, d14 - - ; temp1 = (-step1[11] + step1[12]) * cospi_16_64 - vsub.s32 q10, q13, q0 - vsub.s32 q4, q6, q1 - - ; temp2 = (step1[11] + step1[12]) * cospi_16_64 - vadd.s32 q13, q13, q0 - vadd.s32 q6, q6, q1 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d6, q10, #14 ; >> 14 - vqrshrn.s32 d7, q4, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d8, q13, #14 ; >> 14 - vqrshrn.s32 d9, q6, #14 ; >> 14 - - mov r4, #16 ; pass1Output stride - ldr r3, [sp] ; load skip_adding - cmp r3, #0 ; check if need adding dest data - beq skip_adding_dest - - ldr r7, [sp, #28] ; dest used to save element 0-7 - mov r9, r7 ; save dest pointer for later use - ldr r8, [sp, #32] ; load dest_stride - - ; stage 7 - ; load the data in pass1 - vld1.s16 {q0}, [r2], r4 ; load data step2[0] - vld1.s16 {q1}, [r2], r4 ; load data step2[1] - vld1.s16 {q10}, [r2], r4 ; load data step2[2] - vld1.s16 {q11}, [r2], r4 ; load data step2[3] - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vld1.64 {d13}, [r7], r8 ; load destinatoin data - vadd.s16 q12, q0, q15 ; step2[0] + step2[15] - vadd.s16 q13, q1, q14 ; step2[1] + step2[14] - vrshr.s16 q12, q12, #6 ; ROUND_POWER_OF_TWO - vrshr.s16 q13, q13, #6 ; ROUND_POWER_OF_TWO - vaddw.u8 q12, q12, d12 ; + dest[j * dest_stride + i] - vaddw.u8 q13, q13, d13 ; + dest[j * dest_stride + i] - vqmovun.s16 d12, q12 ; clip pixel - vqmovun.s16 d13, q13 ; clip pixel - vst1.64 {d12}, [r9], r8 ; store the data - vst1.64 {d13}, [r9], r8 ; store the data - vsub.s16 q14, q1, q14 ; step2[1] - step2[14] - vsub.s16 q15, q0, q15 ; step2[0] - step2[15] - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vld1.64 {d13}, [r7], r8 ; load destinatoin data - vadd.s16 q12, q10, q5 ; step2[2] + step2[13] - vadd.s16 q13, q11, q4 ; step2[3] + step2[12] - vrshr.s16 q12, q12, #6 ; ROUND_POWER_OF_TWO - vrshr.s16 q13, q13, #6 ; ROUND_POWER_OF_TWO - vaddw.u8 q12, q12, d12 ; + dest[j * dest_stride + i] - vaddw.u8 q13, q13, d13 ; + dest[j * dest_stride + i] - vqmovun.s16 d12, q12 ; clip pixel - vqmovun.s16 d13, q13 ; clip pixel - vst1.64 {d12}, [r9], r8 ; store the data - vst1.64 {d13}, [r9], r8 ; store the data - vsub.s16 q4, q11, q4 ; step2[3] - step2[12] - vsub.s16 q5, q10, q5 ; step2[2] - step2[13] - vld1.s16 {q0}, [r2], r4 ; load data step2[4] - vld1.s16 {q1}, [r2], r4 ; load data step2[5] - vld1.s16 {q10}, [r2], r4 ; load data step2[6] - vld1.s16 {q11}, [r2], r4 ; load data step2[7] - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vld1.64 {d13}, [r7], r8 ; load destinatoin data - vadd.s16 q12, q0, q3 ; step2[4] + step2[11] - vadd.s16 q13, q1, q2 ; step2[5] + step2[10] - vrshr.s16 q12, q12, #6 ; ROUND_POWER_OF_TWO - vrshr.s16 q13, q13, #6 ; ROUND_POWER_OF_TWO - vaddw.u8 q12, q12, d12 ; + dest[j * dest_stride + i] - vaddw.u8 q13, q13, d13 ; + dest[j * dest_stride + i] - vqmovun.s16 d12, q12 ; clip pixel - vqmovun.s16 d13, q13 ; clip pixel - vst1.64 {d12}, [r9], r8 ; store the data - vst1.64 {d13}, [r9], r8 ; store the data - vsub.s16 q2, q1, q2 ; step2[5] - step2[10] - vsub.s16 q3, q0, q3 ; step2[4] - step2[11] - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vld1.64 {d13}, [r7], r8 ; load destinatoin data - vadd.s16 q12, q10, q9 ; step2[6] + step2[9] - vadd.s16 q13, q11, q8 ; step2[7] + step2[8] - vrshr.s16 q12, q12, #6 ; ROUND_POWER_OF_TWO - vrshr.s16 q13, q13, #6 ; ROUND_POWER_OF_TWO - vaddw.u8 q12, q12, d12 ; + dest[j * dest_stride + i] - vaddw.u8 q13, q13, d13 ; + dest[j * dest_stride + i] - vqmovun.s16 d12, q12 ; clip pixel - vqmovun.s16 d13, q13 ; clip pixel - vst1.64 {d12}, [r9], r8 ; store the data - vst1.64 {d13}, [r9], r8 ; store the data - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vld1.64 {d13}, [r7], r8 ; load destinatoin data - vsub.s16 q8, q11, q8 ; step2[7] - step2[8] - vsub.s16 q9, q10, q9 ; step2[6] - step2[9] - - ; store the data output 8,9,10,11,12,13,14,15 - vrshr.s16 q8, q8, #6 ; ROUND_POWER_OF_TWO - vaddw.u8 q8, q8, d12 ; + dest[j * dest_stride + i] - vqmovun.s16 d12, q8 ; clip pixel - vst1.64 {d12}, [r9], r8 ; store the data - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vrshr.s16 q9, q9, #6 - vaddw.u8 q9, q9, d13 ; + dest[j * dest_stride + i] - vqmovun.s16 d13, q9 ; clip pixel - vst1.64 {d13}, [r9], r8 ; store the data - vld1.64 {d13}, [r7], r8 ; load destinatoin data - vrshr.s16 q2, q2, #6 - vaddw.u8 q2, q2, d12 ; + dest[j * dest_stride + i] - vqmovun.s16 d12, q2 ; clip pixel - vst1.64 {d12}, [r9], r8 ; store the data - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vrshr.s16 q3, q3, #6 - vaddw.u8 q3, q3, d13 ; + dest[j * dest_stride + i] - vqmovun.s16 d13, q3 ; clip pixel - vst1.64 {d13}, [r9], r8 ; store the data - vld1.64 {d13}, [r7], r8 ; load destinatoin data - vrshr.s16 q4, q4, #6 - vaddw.u8 q4, q4, d12 ; + dest[j * dest_stride + i] - vqmovun.s16 d12, q4 ; clip pixel - vst1.64 {d12}, [r9], r8 ; store the data - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vrshr.s16 q5, q5, #6 - vaddw.u8 q5, q5, d13 ; + dest[j * dest_stride + i] - vqmovun.s16 d13, q5 ; clip pixel - vst1.64 {d13}, [r9], r8 ; store the data - vld1.64 {d13}, [r7], r8 ; load destinatoin data - vrshr.s16 q14, q14, #6 - vaddw.u8 q14, q14, d12 ; + dest[j * dest_stride + i] - vqmovun.s16 d12, q14 ; clip pixel - vst1.64 {d12}, [r9], r8 ; store the data - vld1.64 {d12}, [r7], r8 ; load destinatoin data - vrshr.s16 q15, q15, #6 - vaddw.u8 q15, q15, d13 ; + dest[j * dest_stride + i] - vqmovun.s16 d13, q15 ; clip pixel - vst1.64 {d13}, [r9], r8 ; store the data - b end_idct16x16_pass2 - -skip_adding_dest - ; stage 7 - ; load the data in pass1 - mov r5, #24 - mov r3, #8 - - vld1.s16 {q0}, [r2], r4 ; load data step2[0] - vld1.s16 {q1}, [r2], r4 ; load data step2[1] - vadd.s16 q12, q0, q15 ; step2[0] + step2[15] - vadd.s16 q13, q1, q14 ; step2[1] + step2[14] - vld1.s16 {q10}, [r2], r4 ; load data step2[2] - vld1.s16 {q11}, [r2], r4 ; load data step2[3] - vst1.64 {d24}, [r1], r3 ; store output[0] - vst1.64 {d25}, [r1], r5 - vst1.64 {d26}, [r1], r3 ; store output[1] - vst1.64 {d27}, [r1], r5 - vadd.s16 q12, q10, q5 ; step2[2] + step2[13] - vadd.s16 q13, q11, q4 ; step2[3] + step2[12] - vsub.s16 q14, q1, q14 ; step2[1] - step2[14] - vsub.s16 q15, q0, q15 ; step2[0] - step2[15] - vst1.64 {d24}, [r1], r3 ; store output[2] - vst1.64 {d25}, [r1], r5 - vst1.64 {d26}, [r1], r3 ; store output[3] - vst1.64 {d27}, [r1], r5 - vsub.s16 q4, q11, q4 ; step2[3] - step2[12] - vsub.s16 q5, q10, q5 ; step2[2] - step2[13] - vld1.s16 {q0}, [r2], r4 ; load data step2[4] - vld1.s16 {q1}, [r2], r4 ; load data step2[5] - vadd.s16 q12, q0, q3 ; step2[4] + step2[11] - vadd.s16 q13, q1, q2 ; step2[5] + step2[10] - vld1.s16 {q10}, [r2], r4 ; load data step2[6] - vld1.s16 {q11}, [r2], r4 ; load data step2[7] - vst1.64 {d24}, [r1], r3 ; store output[4] - vst1.64 {d25}, [r1], r5 - vst1.64 {d26}, [r1], r3 ; store output[5] - vst1.64 {d27}, [r1], r5 - vadd.s16 q12, q10, q9 ; step2[6] + step2[9] - vadd.s16 q13, q11, q8 ; step2[7] + step2[8] - vsub.s16 q2, q1, q2 ; step2[5] - step2[10] - vsub.s16 q3, q0, q3 ; step2[4] - step2[11] - vsub.s16 q8, q11, q8 ; step2[7] - step2[8] - vsub.s16 q9, q10, q9 ; step2[6] - step2[9] - vst1.64 {d24}, [r1], r3 ; store output[6] - vst1.64 {d25}, [r1], r5 - vst1.64 {d26}, [r1], r3 ; store output[7] - vst1.64 {d27}, [r1], r5 - - ; store the data output 8,9,10,11,12,13,14,15 - vst1.64 {d16}, [r1], r3 - vst1.64 {d17}, [r1], r5 - vst1.64 {d18}, [r1], r3 - vst1.64 {d19}, [r1], r5 - vst1.64 {d4}, [r1], r3 - vst1.64 {d5}, [r1], r5 - vst1.64 {d6}, [r1], r3 - vst1.64 {d7}, [r1], r5 - vst1.64 {d8}, [r1], r3 - vst1.64 {d9}, [r1], r5 - vst1.64 {d10}, [r1], r3 - vst1.64 {d11}, [r1], r5 - vst1.64 {d28}, [r1], r3 - vst1.64 {d29}, [r1], r5 - vst1.64 {d30}, [r1], r3 - vst1.64 {d31}, [r1], r5 -end_idct16x16_pass2 - pop {r3-r9} - bx lr - ENDP ; |aom_idct16x16_256_add_neon_pass2| - -;void |aom_idct16x16_10_add_neon_pass1|(int16_t *input, -; int16_t *output, int output_stride) -; -; r0 int16_t input -; r1 int16_t *output -; r2 int output_stride) - -; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output -; will be stored back into q8-q15 registers. This function will touch q0-q7 -; registers and use them as buffer during calculation. -|aom_idct16x16_10_add_neon_pass1| PROC - - ; TODO(hkuang): Find a better way to load the elements. - ; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15 - vld2.s16 {q8,q9}, [r0]! - vld2.s16 {q9,q10}, [r0]! - vld2.s16 {q10,q11}, [r0]! - vld2.s16 {q11,q12}, [r0]! - vld2.s16 {q12,q13}, [r0]! - vld2.s16 {q13,q14}, [r0]! - vld2.s16 {q14,q15}, [r0]! - vld2.s16 {q1,q2}, [r0]! - vmov.s16 q15, q1 - - ; generate cospi_28_64*2 = 6392 - mov r3, #0x1800 - add r3, #0xf8 - - ; generate cospi_4_64*2 = 32138 - mov r12, #0x7d00 - add r12, #0x8a - - ; transpose the input data - TRANSPOSE8X8 - - ; stage 3 - vdup.16 q0, r3 ; duplicate cospi_28_64*2 - vdup.16 q1, r12 ; duplicate cospi_4_64*2 - - ; The following instructions use vqrdmulh to do the - ; dct_const_round_shift(step2[4] * cospi_28_64). vvqrdmulh will multiply, - ; double, and return the high 16 bits, effectively giving >> 15. Doubling - ; the constant will change this to >> 14. - ; dct_const_round_shift(step2[4] * cospi_28_64); - vqrdmulh.s16 q4, q9, q0 - - ; preloading to avoid stall - ; generate cospi_16_64*2 = 23170 - mov r3, #0x5a00 - add r3, #0x82 - - ; dct_const_round_shift(step2[4] * cospi_4_64); - vqrdmulh.s16 q7, q9, q1 - - ; stage 4 - vdup.16 q1, r3 ; cospi_16_64*2 - - ; generate cospi_16_64 = 11585 - mov r3, #0x2d00 - add r3, #0x41 - - vdup.16 d4, r3; ; duplicate cospi_16_64 - - ; dct_const_round_shift(step1[0] * cospi_16_64) - vqrdmulh.s16 q8, q8, q1 - - ; step2[6] * cospi_16_64 - vmull.s16 q9, d14, d4 - vmull.s16 q10, d15, d4 - - ; step2[5] * cospi_16_64 - vmull.s16 q12, d9, d4 - vmull.s16 q11, d8, d4 - - ; temp1 = (step2[6] - step2[5]) * cospi_16_64 - vsub.s32 q15, q10, q12 - vsub.s32 q6, q9, q11 - - ; temp2 = (step2[5] + step2[6]) * cospi_16_64 - vadd.s32 q9, q9, q11 - vadd.s32 q10, q10, q12 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d11, q15, #14 ; >> 14 - vqrshrn.s32 d10, q6, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d12, q9, #14 ; >> 14 - vqrshrn.s32 d13, q10, #14 ; >> 14 - - ; stage 6 - vadd.s16 q2, q8, q7 ; step2[0] = step1[0] + step1[7]; - vadd.s16 q10, q8, q5 ; step2[2] = step1[2] + step1[5]; - vadd.s16 q11, q8, q4 ; step2[3] = step1[3] + step1[4]; - vadd.s16 q9, q8, q6 ; step2[1] = step1[1] + step1[6]; - vsub.s16 q12, q8, q4 ; step2[4] = step1[3] - step1[4]; - vsub.s16 q13, q8, q5 ; step2[5] = step1[2] - step1[5]; - vsub.s16 q14, q8, q6 ; step2[6] = step1[1] - step1[6]; - vsub.s16 q15, q8, q7 ; step2[7] = step1[0] - step1[7]; - - ; store the data - vst1.64 {d4}, [r1], r2 - vst1.64 {d5}, [r1], r2 - vst1.64 {d18}, [r1], r2 - vst1.64 {d19}, [r1], r2 - vst1.64 {d20}, [r1], r2 - vst1.64 {d21}, [r1], r2 - vst1.64 {d22}, [r1], r2 - vst1.64 {d23}, [r1], r2 - vst1.64 {d24}, [r1], r2 - vst1.64 {d25}, [r1], r2 - vst1.64 {d26}, [r1], r2 - vst1.64 {d27}, [r1], r2 - vst1.64 {d28}, [r1], r2 - vst1.64 {d29}, [r1], r2 - vst1.64 {d30}, [r1], r2 - vst1.64 {d31}, [r1], r2 - - bx lr - ENDP ; |aom_idct16x16_10_add_neon_pass1| - -;void aom_idct16x16_10_add_neon_pass2(int16_t *src, -; int16_t *output, -; int16_t *pass1Output, -; int16_t skip_adding, -; uint8_t *dest, -; int dest_stride) -; -; r0 int16_t *src -; r1 int16_t *output, -; r2 int16_t *pass1Output, -; r3 int16_t skip_adding, -; r4 uint8_t *dest, -; r5 int dest_stride) - -; idct16 stage1 - stage7 on all the elements loaded in q8-q15. The output -; will be stored back into q8-q15 registers. This function will touch q0-q7 -; registers and use them as buffer during calculation. -|aom_idct16x16_10_add_neon_pass2| PROC - push {r3-r9} - - ; TODO(hkuang): Find a better way to load the elements. - ; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15 - vld2.s16 {q8,q9}, [r0]! - vld2.s16 {q9,q10}, [r0]! - vld2.s16 {q10,q11}, [r0]! - vld2.s16 {q11,q12}, [r0]! - vld2.s16 {q12,q13}, [r0]! - vld2.s16 {q13,q14}, [r0]! - vld2.s16 {q14,q15}, [r0]! - vld2.s16 {q0,q1}, [r0]! - vmov.s16 q15, q0; - - ; generate 2*cospi_30_64 = 3212 - mov r3, #0xc00 - add r3, #0x8c - - ; generate 2*cospi_2_64 = 32610 - mov r12, #0x7f00 - add r12, #0x62 - - ; transpose the input data - TRANSPOSE8X8 - - ; stage 3 - vdup.16 q6, r3 ; duplicate 2*cospi_30_64 - - ; dct_const_round_shift(step1[8] * cospi_30_64) - vqrdmulh.s16 q0, q8, q6 - - vdup.16 q6, r12 ; duplicate 2*cospi_2_64 - - ; dct_const_round_shift(step1[8] * cospi_2_64) - vqrdmulh.s16 q7, q8, q6 - - ; preloading to avoid stall - ; generate 2*cospi_26_64 = 9512 - mov r12, #0x2500 - add r12, #0x28 - rsb r12, #0 - vdup.16 q15, r12 ; duplicate -2*cospi_26_64 - - ; generate 2*cospi_6_64 = 31358 - mov r3, #0x7a00 - add r3, #0x7e - vdup.16 q14, r3 ; duplicate 2*cospi_6_64 - - ; dct_const_round_shift(- step1[12] * cospi_26_64) - vqrdmulh.s16 q3, q9, q15 - - ; dct_const_round_shift(step1[12] * cospi_6_64) - vqrdmulh.s16 q4, q9, q14 - - ; stage 4 - ; generate cospi_24_64 = 6270 - mov r3, #0x1800 - add r3, #0x7e - vdup.16 d31, r3 ; duplicate cospi_24_64 - - ; generate cospi_8_64 = 15137 - mov r12, #0x3b00 - add r12, #0x21 - vdup.16 d30, r12 ; duplicate cospi_8_64 - - ; step1[14] * cospi_24_64 - vmull.s16 q12, d14, d31 - vmull.s16 q5, d15, d31 - - ; step1[9] * cospi_24_64 - vmull.s16 q2, d0, d31 - vmull.s16 q11, d1, d31 - - ; temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64 - vmlsl.s16 q12, d0, d30 - vmlsl.s16 q5, d1, d30 - - ; temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64 - vmlal.s16 q2, d14, d30 - vmlal.s16 q11, d15, d30 - - rsb r12, #0 - vdup.16 d30, r12 ; duplicate -cospi_8_64 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d2, q12, #14 ; >> 14 - vqrshrn.s32 d3, q5, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d12, q2, #14 ; >> 14 - vqrshrn.s32 d13, q11, #14 ; >> 14 - - ; - step1[13] * cospi_8_64 - vmull.s16 q10, d8, d30 - vmull.s16 q13, d9, d30 - - ; -step1[10] * cospi_8_64 - vmull.s16 q8, d6, d30 - vmull.s16 q9, d7, d30 - - ; temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64 - vmlsl.s16 q10, d6, d31 - vmlsl.s16 q13, d7, d31 - - ; temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64 - vmlal.s16 q8, d8, d31 - vmlal.s16 q9, d9, d31 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d4, q10, #14 ; >> 14 - vqrshrn.s32 d5, q13, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d10, q8, #14 ; >> 14 - vqrshrn.s32 d11, q9, #14 ; >> 14 - - ; stage 5 - vadd.s16 q8, q0, q3 ; step1[8] = step2[8]+step2[11]; - vadd.s16 q9, q1, q2 ; step1[9] = step2[9]+step2[10]; - vsub.s16 q10, q1, q2 ; step1[10] = step2[9]-step2[10]; - vsub.s16 q11, q0, q3 ; step1[11] = step2[8]-step2[11]; - vsub.s16 q12, q7, q4 ; step1[12] =-step2[12]+step2[15]; - vsub.s16 q13, q6, q5 ; step1[13] =-step2[13]+step2[14]; - vadd.s16 q14, q6, q5 ; step1[14] =step2[13]+step2[14]; - vadd.s16 q15, q7, q4 ; step1[15] =step2[12]+step2[15]; - - ; stage 6. - ; generate cospi_16_64 = 11585 - mov r12, #0x2d00 - add r12, #0x41 - - vdup.16 d14, r12 ; duplicate cospi_16_64 - - ; step1[13] * cospi_16_64 - vmull.s16 q3, d26, d14 - vmull.s16 q4, d27, d14 - - ; step1[10] * cospi_16_64 - vmull.s16 q0, d20, d14 - vmull.s16 q1, d21, d14 - - ; temp1 = (-step1[10] + step1[13]) * cospi_16_64 - vsub.s32 q5, q3, q0 - vsub.s32 q6, q4, q1 - - ; temp2 = (step1[10] + step1[13]) * cospi_16_64 - vadd.s32 q0, q3, q0 - vadd.s32 q1, q4, q1 - - ; dct_const_round_shift(temp1) - vqrshrn.s32 d4, q5, #14 ; >> 14 - vqrshrn.s32 d5, q6, #14 ; >> 14 - - ; dct_const_round_shift(temp2) - vqrshrn.s32 d10, q0, #14 ; >> 14 - vqrshrn.s32 d11, q1, #14 ; >> 14 - - ; step1[11] * cospi_16_64 - vmull.s16 q0, d22, d14 - vmull.s16 q1, d23, d14 - - ; step1[12] * cospi_16_64 - vmull.s16 q13, d24, d14 - vmull.s16 q6, d25, d14 - - ; temp1 = (-step1[11] + step1[12]) * cospi_16_64 - vsub.s32 q10, q13, q0 - vsub.s32 q4, q6, q1 - - ; temp2 = (step1[11] + step1[12]) * cospi_16_64 - vadd.s32 q13, q13, q0 - vadd.s32 q6, q6, q1 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d6, q10, #14 ; >> 14 - vqrshrn.s32 d7, q4, #14 ; >> 14 - - ; dct_const_round_shift((step1[11] + step1[12]) * cospi_16_64); - vqrshrn.s32 d8, q13, #14 ; >> 14 - vqrshrn.s32 d9, q6, #14 ; >> 14 - - mov r4, #16 ; pass1Output stride - ldr r3, [sp] ; load skip_adding - - ; stage 7 - ; load the data in pass1 - mov r5, #24 - mov r3, #8 - - vld1.s16 {q0}, [r2], r4 ; load data step2[0] - vld1.s16 {q1}, [r2], r4 ; load data step2[1] - vadd.s16 q12, q0, q15 ; step2[0] + step2[15] - vadd.s16 q13, q1, q14 ; step2[1] + step2[14] - vld1.s16 {q10}, [r2], r4 ; load data step2[2] - vld1.s16 {q11}, [r2], r4 ; load data step2[3] - vst1.64 {d24}, [r1], r3 ; store output[0] - vst1.64 {d25}, [r1], r5 - vst1.64 {d26}, [r1], r3 ; store output[1] - vst1.64 {d27}, [r1], r5 - vadd.s16 q12, q10, q5 ; step2[2] + step2[13] - vadd.s16 q13, q11, q4 ; step2[3] + step2[12] - vsub.s16 q14, q1, q14 ; step2[1] - step2[14] - vsub.s16 q15, q0, q15 ; step2[0] - step2[15] - vst1.64 {d24}, [r1], r3 ; store output[2] - vst1.64 {d25}, [r1], r5 - vst1.64 {d26}, [r1], r3 ; store output[3] - vst1.64 {d27}, [r1], r5 - vsub.s16 q4, q11, q4 ; step2[3] - step2[12] - vsub.s16 q5, q10, q5 ; step2[2] - step2[13] - vld1.s16 {q0}, [r2], r4 ; load data step2[4] - vld1.s16 {q1}, [r2], r4 ; load data step2[5] - vadd.s16 q12, q0, q3 ; step2[4] + step2[11] - vadd.s16 q13, q1, q2 ; step2[5] + step2[10] - vld1.s16 {q10}, [r2], r4 ; load data step2[6] - vld1.s16 {q11}, [r2], r4 ; load data step2[7] - vst1.64 {d24}, [r1], r3 ; store output[4] - vst1.64 {d25}, [r1], r5 - vst1.64 {d26}, [r1], r3 ; store output[5] - vst1.64 {d27}, [r1], r5 - vadd.s16 q12, q10, q9 ; step2[6] + step2[9] - vadd.s16 q13, q11, q8 ; step2[7] + step2[8] - vsub.s16 q2, q1, q2 ; step2[5] - step2[10] - vsub.s16 q3, q0, q3 ; step2[4] - step2[11] - vsub.s16 q8, q11, q8 ; step2[7] - step2[8] - vsub.s16 q9, q10, q9 ; step2[6] - step2[9] - vst1.64 {d24}, [r1], r3 ; store output[6] - vst1.64 {d25}, [r1], r5 - vst1.64 {d26}, [r1], r3 ; store output[7] - vst1.64 {d27}, [r1], r5 - - ; store the data output 8,9,10,11,12,13,14,15 - vst1.64 {d16}, [r1], r3 - vst1.64 {d17}, [r1], r5 - vst1.64 {d18}, [r1], r3 - vst1.64 {d19}, [r1], r5 - vst1.64 {d4}, [r1], r3 - vst1.64 {d5}, [r1], r5 - vst1.64 {d6}, [r1], r3 - vst1.64 {d7}, [r1], r5 - vst1.64 {d8}, [r1], r3 - vst1.64 {d9}, [r1], r5 - vst1.64 {d10}, [r1], r3 - vst1.64 {d11}, [r1], r5 - vst1.64 {d28}, [r1], r3 - vst1.64 {d29}, [r1], r5 - vst1.64 {d30}, [r1], r3 - vst1.64 {d31}, [r1], r5 -end_idct10_16x16_pass2 - pop {r3-r9} - bx lr - ENDP ; |aom_idct16x16_10_add_neon_pass2| - END diff --git a/aom_dsp/arm/idct16x16_add_neon.c b/aom_dsp/arm/idct16x16_add_neon.c deleted file mode 100644 index b4cb7a0cd..000000000 --- a/aom_dsp/arm/idct16x16_add_neon.c +++ /dev/null @@ -1,1295 +0,0 @@ -/* - * Copyright (c) 2016, 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 "./aom_config.h" -#include "aom_dsp/txfm_common.h" - -static INLINE void TRANSPOSE8X8(int16x8_t *q8s16, int16x8_t *q9s16, - int16x8_t *q10s16, int16x8_t *q11s16, - int16x8_t *q12s16, int16x8_t *q13s16, - int16x8_t *q14s16, int16x8_t *q15s16) { - int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16; - int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16; - int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32; - int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16; - - d16s16 = vget_low_s16(*q8s16); - d17s16 = vget_high_s16(*q8s16); - d18s16 = vget_low_s16(*q9s16); - d19s16 = vget_high_s16(*q9s16); - d20s16 = vget_low_s16(*q10s16); - d21s16 = vget_high_s16(*q10s16); - d22s16 = vget_low_s16(*q11s16); - d23s16 = vget_high_s16(*q11s16); - d24s16 = vget_low_s16(*q12s16); - d25s16 = vget_high_s16(*q12s16); - d26s16 = vget_low_s16(*q13s16); - d27s16 = vget_high_s16(*q13s16); - d28s16 = vget_low_s16(*q14s16); - d29s16 = vget_high_s16(*q14s16); - d30s16 = vget_low_s16(*q15s16); - d31s16 = vget_high_s16(*q15s16); - - *q8s16 = vcombine_s16(d16s16, d24s16); // vswp d17, d24 - *q9s16 = vcombine_s16(d18s16, d26s16); // vswp d19, d26 - *q10s16 = vcombine_s16(d20s16, d28s16); // vswp d21, d28 - *q11s16 = vcombine_s16(d22s16, d30s16); // vswp d23, d30 - *q12s16 = vcombine_s16(d17s16, d25s16); - *q13s16 = vcombine_s16(d19s16, d27s16); - *q14s16 = vcombine_s16(d21s16, d29s16); - *q15s16 = vcombine_s16(d23s16, d31s16); - - q0x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(*q8s16), vreinterpretq_s32_s16(*q10s16)); - q1x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(*q9s16), vreinterpretq_s32_s16(*q11s16)); - q2x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(*q12s16), vreinterpretq_s32_s16(*q14s16)); - q3x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(*q13s16), vreinterpretq_s32_s16(*q15s16)); - - q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]), // q8 - vreinterpretq_s16_s32(q1x2s32.val[0])); // q9 - q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]), // q10 - vreinterpretq_s16_s32(q1x2s32.val[1])); // q11 - q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]), // q12 - vreinterpretq_s16_s32(q3x2s32.val[0])); // q13 - q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]), // q14 - vreinterpretq_s16_s32(q3x2s32.val[1])); // q15 - - *q8s16 = q0x2s16.val[0]; - *q9s16 = q0x2s16.val[1]; - *q10s16 = q1x2s16.val[0]; - *q11s16 = q1x2s16.val[1]; - *q12s16 = q2x2s16.val[0]; - *q13s16 = q2x2s16.val[1]; - *q14s16 = q3x2s16.val[0]; - *q15s16 = q3x2s16.val[1]; - return; -} - -void aom_idct16x16_256_add_neon_pass1(int16_t *in, int16_t *out, - int output_stride) { - int16x4_t d0s16, d1s16, d2s16, d3s16; - int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16; - int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16; - int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16; - uint64x1_t d16u64, d17u64, d18u64, d19u64, d20u64, d21u64, d22u64, d23u64; - uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64; - int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16; - int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16; - int32x4_t q0s32, q1s32, q2s32, q3s32, q5s32, q6s32, q9s32; - int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32; - int16x8x2_t q0x2s16; - - q0x2s16 = vld2q_s16(in); - q8s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q9s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q10s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q11s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q12s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q13s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q14s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q15s16 = q0x2s16.val[0]; - - TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - d16s16 = vget_low_s16(q8s16); - d17s16 = vget_high_s16(q8s16); - d18s16 = vget_low_s16(q9s16); - d19s16 = vget_high_s16(q9s16); - d20s16 = vget_low_s16(q10s16); - d21s16 = vget_high_s16(q10s16); - d22s16 = vget_low_s16(q11s16); - d23s16 = vget_high_s16(q11s16); - d24s16 = vget_low_s16(q12s16); - d25s16 = vget_high_s16(q12s16); - d26s16 = vget_low_s16(q13s16); - d27s16 = vget_high_s16(q13s16); - d28s16 = vget_low_s16(q14s16); - d29s16 = vget_high_s16(q14s16); - d30s16 = vget_low_s16(q15s16); - d31s16 = vget_high_s16(q15s16); - - // stage 3 - d0s16 = vdup_n_s16((int16_t)cospi_28_64); - d1s16 = vdup_n_s16((int16_t)cospi_4_64); - - q2s32 = vmull_s16(d18s16, d0s16); - q3s32 = vmull_s16(d19s16, d0s16); - q5s32 = vmull_s16(d18s16, d1s16); - q6s32 = vmull_s16(d19s16, d1s16); - - q2s32 = vmlsl_s16(q2s32, d30s16, d1s16); - q3s32 = vmlsl_s16(q3s32, d31s16, d1s16); - q5s32 = vmlal_s16(q5s32, d30s16, d0s16); - q6s32 = vmlal_s16(q6s32, d31s16, d0s16); - - d2s16 = vdup_n_s16((int16_t)cospi_12_64); - d3s16 = vdup_n_s16((int16_t)cospi_20_64); - - d8s16 = vqrshrn_n_s32(q2s32, 14); - d9s16 = vqrshrn_n_s32(q3s32, 14); - d14s16 = vqrshrn_n_s32(q5s32, 14); - d15s16 = vqrshrn_n_s32(q6s32, 14); - q4s16 = vcombine_s16(d8s16, d9s16); - q7s16 = vcombine_s16(d14s16, d15s16); - - q2s32 = vmull_s16(d26s16, d2s16); - q3s32 = vmull_s16(d27s16, d2s16); - q9s32 = vmull_s16(d26s16, d3s16); - q15s32 = vmull_s16(d27s16, d3s16); - - q2s32 = vmlsl_s16(q2s32, d22s16, d3s16); - q3s32 = vmlsl_s16(q3s32, d23s16, d3s16); - q9s32 = vmlal_s16(q9s32, d22s16, d2s16); - q15s32 = vmlal_s16(q15s32, d23s16, d2s16); - - d10s16 = vqrshrn_n_s32(q2s32, 14); - d11s16 = vqrshrn_n_s32(q3s32, 14); - d12s16 = vqrshrn_n_s32(q9s32, 14); - d13s16 = vqrshrn_n_s32(q15s32, 14); - q5s16 = vcombine_s16(d10s16, d11s16); - q6s16 = vcombine_s16(d12s16, d13s16); - - // stage 4 - d30s16 = vdup_n_s16((int16_t)cospi_16_64); - - q2s32 = vmull_s16(d16s16, d30s16); - q11s32 = vmull_s16(d17s16, d30s16); - q0s32 = vmull_s16(d24s16, d30s16); - q1s32 = vmull_s16(d25s16, d30s16); - - d30s16 = vdup_n_s16((int16_t)cospi_24_64); - d31s16 = vdup_n_s16((int16_t)cospi_8_64); - - q3s32 = vaddq_s32(q2s32, q0s32); - q12s32 = vaddq_s32(q11s32, q1s32); - q13s32 = vsubq_s32(q2s32, q0s32); - q1s32 = vsubq_s32(q11s32, q1s32); - - d16s16 = vqrshrn_n_s32(q3s32, 14); - d17s16 = vqrshrn_n_s32(q12s32, 14); - d18s16 = vqrshrn_n_s32(q13s32, 14); - d19s16 = vqrshrn_n_s32(q1s32, 14); - q8s16 = vcombine_s16(d16s16, d17s16); - q9s16 = vcombine_s16(d18s16, d19s16); - - q0s32 = vmull_s16(d20s16, d31s16); - q1s32 = vmull_s16(d21s16, d31s16); - q12s32 = vmull_s16(d20s16, d30s16); - q13s32 = vmull_s16(d21s16, d30s16); - - q0s32 = vmlal_s16(q0s32, d28s16, d30s16); - q1s32 = vmlal_s16(q1s32, d29s16, d30s16); - q12s32 = vmlsl_s16(q12s32, d28s16, d31s16); - q13s32 = vmlsl_s16(q13s32, d29s16, d31s16); - - d22s16 = vqrshrn_n_s32(q0s32, 14); - d23s16 = vqrshrn_n_s32(q1s32, 14); - d20s16 = vqrshrn_n_s32(q12s32, 14); - d21s16 = vqrshrn_n_s32(q13s32, 14); - q10s16 = vcombine_s16(d20s16, d21s16); - q11s16 = vcombine_s16(d22s16, d23s16); - - q13s16 = vsubq_s16(q4s16, q5s16); - q4s16 = vaddq_s16(q4s16, q5s16); - q14s16 = vsubq_s16(q7s16, q6s16); - q15s16 = vaddq_s16(q6s16, q7s16); - d26s16 = vget_low_s16(q13s16); - d27s16 = vget_high_s16(q13s16); - d28s16 = vget_low_s16(q14s16); - d29s16 = vget_high_s16(q14s16); - - // stage 5 - q0s16 = vaddq_s16(q8s16, q11s16); - q1s16 = vaddq_s16(q9s16, q10s16); - q2s16 = vsubq_s16(q9s16, q10s16); - q3s16 = vsubq_s16(q8s16, q11s16); - - d16s16 = vdup_n_s16((int16_t)cospi_16_64); - - q11s32 = vmull_s16(d26s16, d16s16); - q12s32 = vmull_s16(d27s16, d16s16); - q9s32 = vmull_s16(d28s16, d16s16); - q10s32 = vmull_s16(d29s16, d16s16); - - q6s32 = vsubq_s32(q9s32, q11s32); - q13s32 = vsubq_s32(q10s32, q12s32); - q9s32 = vaddq_s32(q9s32, q11s32); - q10s32 = vaddq_s32(q10s32, q12s32); - - d10s16 = vqrshrn_n_s32(q6s32, 14); - d11s16 = vqrshrn_n_s32(q13s32, 14); - d12s16 = vqrshrn_n_s32(q9s32, 14); - d13s16 = vqrshrn_n_s32(q10s32, 14); - q5s16 = vcombine_s16(d10s16, d11s16); - q6s16 = vcombine_s16(d12s16, d13s16); - - // stage 6 - q8s16 = vaddq_s16(q0s16, q15s16); - q9s16 = vaddq_s16(q1s16, q6s16); - q10s16 = vaddq_s16(q2s16, q5s16); - q11s16 = vaddq_s16(q3s16, q4s16); - q12s16 = vsubq_s16(q3s16, q4s16); - q13s16 = vsubq_s16(q2s16, q5s16); - q14s16 = vsubq_s16(q1s16, q6s16); - q15s16 = vsubq_s16(q0s16, q15s16); - - d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16)); - d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16)); - d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16)); - d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16)); - d20u64 = vreinterpret_u64_s16(vget_low_s16(q10s16)); - d21u64 = vreinterpret_u64_s16(vget_high_s16(q10s16)); - d22u64 = vreinterpret_u64_s16(vget_low_s16(q11s16)); - d23u64 = vreinterpret_u64_s16(vget_high_s16(q11s16)); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16)); - d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16)); - d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16)); - d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16)); - - // store the data - output_stride >>= 1; // output_stride / 2, out is int16_t - vst1_u64((uint64_t *)out, d16u64); - out += output_stride; - vst1_u64((uint64_t *)out, d17u64); - out += output_stride; - vst1_u64((uint64_t *)out, d18u64); - out += output_stride; - vst1_u64((uint64_t *)out, d19u64); - out += output_stride; - vst1_u64((uint64_t *)out, d20u64); - out += output_stride; - vst1_u64((uint64_t *)out, d21u64); - out += output_stride; - vst1_u64((uint64_t *)out, d22u64); - out += output_stride; - vst1_u64((uint64_t *)out, d23u64); - out += output_stride; - vst1_u64((uint64_t *)out, d24u64); - out += output_stride; - vst1_u64((uint64_t *)out, d25u64); - out += output_stride; - vst1_u64((uint64_t *)out, d26u64); - out += output_stride; - vst1_u64((uint64_t *)out, d27u64); - out += output_stride; - vst1_u64((uint64_t *)out, d28u64); - out += output_stride; - vst1_u64((uint64_t *)out, d29u64); - out += output_stride; - vst1_u64((uint64_t *)out, d30u64); - out += output_stride; - vst1_u64((uint64_t *)out, d31u64); - return; -} - -void aom_idct16x16_256_add_neon_pass2(int16_t *src, int16_t *out, - int16_t *pass1Output, int16_t skip_adding, - uint8_t *dest, int dest_stride) { - uint8_t *d; - uint8x8_t d12u8, d13u8; - int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16; - int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16; - int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16; - int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16; - uint64x1_t d24u64, d25u64, d26u64, d27u64; - int64x1_t d12s64, d13s64; - uint16x8_t q2u16, q3u16, q4u16, q5u16, q8u16; - uint16x8_t q9u16, q12u16, q13u16, q14u16, q15u16; - int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16; - int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16; - int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32; - int32x4_t q10s32, q11s32, q12s32, q13s32; - int16x8x2_t q0x2s16; - - q0x2s16 = vld2q_s16(src); - q8s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q9s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q10s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q11s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q12s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q13s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q14s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q15s16 = q0x2s16.val[0]; - - TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - d16s16 = vget_low_s16(q8s16); - d17s16 = vget_high_s16(q8s16); - d18s16 = vget_low_s16(q9s16); - d19s16 = vget_high_s16(q9s16); - d20s16 = vget_low_s16(q10s16); - d21s16 = vget_high_s16(q10s16); - d22s16 = vget_low_s16(q11s16); - d23s16 = vget_high_s16(q11s16); - d24s16 = vget_low_s16(q12s16); - d25s16 = vget_high_s16(q12s16); - d26s16 = vget_low_s16(q13s16); - d27s16 = vget_high_s16(q13s16); - d28s16 = vget_low_s16(q14s16); - d29s16 = vget_high_s16(q14s16); - d30s16 = vget_low_s16(q15s16); - d31s16 = vget_high_s16(q15s16); - - // stage 3 - d12s16 = vdup_n_s16((int16_t)cospi_30_64); - d13s16 = vdup_n_s16((int16_t)cospi_2_64); - - q2s32 = vmull_s16(d16s16, d12s16); - q3s32 = vmull_s16(d17s16, d12s16); - q1s32 = vmull_s16(d16s16, d13s16); - q4s32 = vmull_s16(d17s16, d13s16); - - q2s32 = vmlsl_s16(q2s32, d30s16, d13s16); - q3s32 = vmlsl_s16(q3s32, d31s16, d13s16); - q1s32 = vmlal_s16(q1s32, d30s16, d12s16); - q4s32 = vmlal_s16(q4s32, d31s16, d12s16); - - d0s16 = vqrshrn_n_s32(q2s32, 14); - d1s16 = vqrshrn_n_s32(q3s32, 14); - d14s16 = vqrshrn_n_s32(q1s32, 14); - d15s16 = vqrshrn_n_s32(q4s32, 14); - q0s16 = vcombine_s16(d0s16, d1s16); - q7s16 = vcombine_s16(d14s16, d15s16); - - d30s16 = vdup_n_s16((int16_t)cospi_14_64); - d31s16 = vdup_n_s16((int16_t)cospi_18_64); - - q2s32 = vmull_s16(d24s16, d30s16); - q3s32 = vmull_s16(d25s16, d30s16); - q4s32 = vmull_s16(d24s16, d31s16); - q5s32 = vmull_s16(d25s16, d31s16); - - q2s32 = vmlsl_s16(q2s32, d22s16, d31s16); - q3s32 = vmlsl_s16(q3s32, d23s16, d31s16); - q4s32 = vmlal_s16(q4s32, d22s16, d30s16); - q5s32 = vmlal_s16(q5s32, d23s16, d30s16); - - d2s16 = vqrshrn_n_s32(q2s32, 14); - d3s16 = vqrshrn_n_s32(q3s32, 14); - d12s16 = vqrshrn_n_s32(q4s32, 14); - d13s16 = vqrshrn_n_s32(q5s32, 14); - q1s16 = vcombine_s16(d2s16, d3s16); - q6s16 = vcombine_s16(d12s16, d13s16); - - d30s16 = vdup_n_s16((int16_t)cospi_22_64); - d31s16 = vdup_n_s16((int16_t)cospi_10_64); - - q11s32 = vmull_s16(d20s16, d30s16); - q12s32 = vmull_s16(d21s16, d30s16); - q4s32 = vmull_s16(d20s16, d31s16); - q5s32 = vmull_s16(d21s16, d31s16); - - q11s32 = vmlsl_s16(q11s32, d26s16, d31s16); - q12s32 = vmlsl_s16(q12s32, d27s16, d31s16); - q4s32 = vmlal_s16(q4s32, d26s16, d30s16); - q5s32 = vmlal_s16(q5s32, d27s16, d30s16); - - d4s16 = vqrshrn_n_s32(q11s32, 14); - d5s16 = vqrshrn_n_s32(q12s32, 14); - d11s16 = vqrshrn_n_s32(q5s32, 14); - d10s16 = vqrshrn_n_s32(q4s32, 14); - q2s16 = vcombine_s16(d4s16, d5s16); - q5s16 = vcombine_s16(d10s16, d11s16); - - d30s16 = vdup_n_s16((int16_t)cospi_6_64); - d31s16 = vdup_n_s16((int16_t)cospi_26_64); - - q10s32 = vmull_s16(d28s16, d30s16); - q11s32 = vmull_s16(d29s16, d30s16); - q12s32 = vmull_s16(d28s16, d31s16); - q13s32 = vmull_s16(d29s16, d31s16); - - q10s32 = vmlsl_s16(q10s32, d18s16, d31s16); - q11s32 = vmlsl_s16(q11s32, d19s16, d31s16); - q12s32 = vmlal_s16(q12s32, d18s16, d30s16); - q13s32 = vmlal_s16(q13s32, d19s16, d30s16); - - d6s16 = vqrshrn_n_s32(q10s32, 14); - d7s16 = vqrshrn_n_s32(q11s32, 14); - d8s16 = vqrshrn_n_s32(q12s32, 14); - d9s16 = vqrshrn_n_s32(q13s32, 14); - q3s16 = vcombine_s16(d6s16, d7s16); - q4s16 = vcombine_s16(d8s16, d9s16); - - // stage 3 - q9s16 = vsubq_s16(q0s16, q1s16); - q0s16 = vaddq_s16(q0s16, q1s16); - q10s16 = vsubq_s16(q3s16, q2s16); - q11s16 = vaddq_s16(q2s16, q3s16); - q12s16 = vaddq_s16(q4s16, q5s16); - q13s16 = vsubq_s16(q4s16, q5s16); - q14s16 = vsubq_s16(q7s16, q6s16); - q7s16 = vaddq_s16(q6s16, q7s16); - - // stage 4 - d18s16 = vget_low_s16(q9s16); - d19s16 = vget_high_s16(q9s16); - d20s16 = vget_low_s16(q10s16); - d21s16 = vget_high_s16(q10s16); - d26s16 = vget_low_s16(q13s16); - d27s16 = vget_high_s16(q13s16); - d28s16 = vget_low_s16(q14s16); - d29s16 = vget_high_s16(q14s16); - - d30s16 = vdup_n_s16((int16_t)cospi_8_64); - d31s16 = vdup_n_s16((int16_t)cospi_24_64); - - q2s32 = vmull_s16(d18s16, d31s16); - q3s32 = vmull_s16(d19s16, d31s16); - q4s32 = vmull_s16(d28s16, d31s16); - q5s32 = vmull_s16(d29s16, d31s16); - - q2s32 = vmlal_s16(q2s32, d28s16, d30s16); - q3s32 = vmlal_s16(q3s32, d29s16, d30s16); - q4s32 = vmlsl_s16(q4s32, d18s16, d30s16); - q5s32 = vmlsl_s16(q5s32, d19s16, d30s16); - - d12s16 = vqrshrn_n_s32(q2s32, 14); - d13s16 = vqrshrn_n_s32(q3s32, 14); - d2s16 = vqrshrn_n_s32(q4s32, 14); - d3s16 = vqrshrn_n_s32(q5s32, 14); - q1s16 = vcombine_s16(d2s16, d3s16); - q6s16 = vcombine_s16(d12s16, d13s16); - - q3s16 = q11s16; - q4s16 = q12s16; - - d30s16 = vdup_n_s16(-cospi_8_64); - q11s32 = vmull_s16(d26s16, d30s16); - q12s32 = vmull_s16(d27s16, d30s16); - q8s32 = vmull_s16(d20s16, d30s16); - q9s32 = vmull_s16(d21s16, d30s16); - - q11s32 = vmlsl_s16(q11s32, d20s16, d31s16); - q12s32 = vmlsl_s16(q12s32, d21s16, d31s16); - q8s32 = vmlal_s16(q8s32, d26s16, d31s16); - q9s32 = vmlal_s16(q9s32, d27s16, d31s16); - - d4s16 = vqrshrn_n_s32(q11s32, 14); - d5s16 = vqrshrn_n_s32(q12s32, 14); - d10s16 = vqrshrn_n_s32(q8s32, 14); - d11s16 = vqrshrn_n_s32(q9s32, 14); - q2s16 = vcombine_s16(d4s16, d5s16); - q5s16 = vcombine_s16(d10s16, d11s16); - - // stage 5 - q8s16 = vaddq_s16(q0s16, q3s16); - q9s16 = vaddq_s16(q1s16, q2s16); - q10s16 = vsubq_s16(q1s16, q2s16); - q11s16 = vsubq_s16(q0s16, q3s16); - q12s16 = vsubq_s16(q7s16, q4s16); - q13s16 = vsubq_s16(q6s16, q5s16); - q14s16 = vaddq_s16(q6s16, q5s16); - q15s16 = vaddq_s16(q7s16, q4s16); - - // stage 6 - d20s16 = vget_low_s16(q10s16); - d21s16 = vget_high_s16(q10s16); - d22s16 = vget_low_s16(q11s16); - d23s16 = vget_high_s16(q11s16); - d24s16 = vget_low_s16(q12s16); - d25s16 = vget_high_s16(q12s16); - d26s16 = vget_low_s16(q13s16); - d27s16 = vget_high_s16(q13s16); - - d14s16 = vdup_n_s16((int16_t)cospi_16_64); - - q3s32 = vmull_s16(d26s16, d14s16); - q4s32 = vmull_s16(d27s16, d14s16); - q0s32 = vmull_s16(d20s16, d14s16); - q1s32 = vmull_s16(d21s16, d14s16); - - q5s32 = vsubq_s32(q3s32, q0s32); - q6s32 = vsubq_s32(q4s32, q1s32); - q10s32 = vaddq_s32(q3s32, q0s32); - q4s32 = vaddq_s32(q4s32, q1s32); - - d4s16 = vqrshrn_n_s32(q5s32, 14); - d5s16 = vqrshrn_n_s32(q6s32, 14); - d10s16 = vqrshrn_n_s32(q10s32, 14); - d11s16 = vqrshrn_n_s32(q4s32, 14); - q2s16 = vcombine_s16(d4s16, d5s16); - q5s16 = vcombine_s16(d10s16, d11s16); - - q0s32 = vmull_s16(d22s16, d14s16); - q1s32 = vmull_s16(d23s16, d14s16); - q13s32 = vmull_s16(d24s16, d14s16); - q6s32 = vmull_s16(d25s16, d14s16); - - q10s32 = vsubq_s32(q13s32, q0s32); - q4s32 = vsubq_s32(q6s32, q1s32); - q13s32 = vaddq_s32(q13s32, q0s32); - q6s32 = vaddq_s32(q6s32, q1s32); - - d6s16 = vqrshrn_n_s32(q10s32, 14); - d7s16 = vqrshrn_n_s32(q4s32, 14); - d8s16 = vqrshrn_n_s32(q13s32, 14); - d9s16 = vqrshrn_n_s32(q6s32, 14); - q3s16 = vcombine_s16(d6s16, d7s16); - q4s16 = vcombine_s16(d8s16, d9s16); - - // stage 7 - if (skip_adding != 0) { - d = dest; - // load the data in pass1 - q0s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q1s16 = vld1q_s16(pass1Output); - pass1Output += 8; - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - d13s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - - q12s16 = vaddq_s16(q0s16, q15s16); - q13s16 = vaddq_s16(q1s16, q14s16); - q12s16 = vrshrq_n_s16(q12s16, 6); - q13s16 = vrshrq_n_s16(q13s16, 6); - q12u16 = - vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64)); - q13u16 = - vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16)); - d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8)); - d += dest_stride; - q14s16 = vsubq_s16(q1s16, q14s16); - q15s16 = vsubq_s16(q0s16, q15s16); - - q10s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q11s16 = vld1q_s16(pass1Output); - pass1Output += 8; - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - d13s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q12s16 = vaddq_s16(q10s16, q5s16); - q13s16 = vaddq_s16(q11s16, q4s16); - q12s16 = vrshrq_n_s16(q12s16, 6); - q13s16 = vrshrq_n_s16(q13s16, 6); - q12u16 = - vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64)); - q13u16 = - vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16)); - d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8)); - d += dest_stride; - q4s16 = vsubq_s16(q11s16, q4s16); - q5s16 = vsubq_s16(q10s16, q5s16); - - q0s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q1s16 = vld1q_s16(pass1Output); - pass1Output += 8; - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - d13s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q12s16 = vaddq_s16(q0s16, q3s16); - q13s16 = vaddq_s16(q1s16, q2s16); - q12s16 = vrshrq_n_s16(q12s16, 6); - q13s16 = vrshrq_n_s16(q13s16, 6); - q12u16 = - vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64)); - q13u16 = - vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16)); - d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8)); - d += dest_stride; - q2s16 = vsubq_s16(q1s16, q2s16); - q3s16 = vsubq_s16(q0s16, q3s16); - - q10s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q11s16 = vld1q_s16(pass1Output); - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - d13s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q12s16 = vaddq_s16(q10s16, q9s16); - q13s16 = vaddq_s16(q11s16, q8s16); - q12s16 = vrshrq_n_s16(q12s16, 6); - q13s16 = vrshrq_n_s16(q13s16, 6); - q12u16 = - vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64)); - q13u16 = - vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16)); - d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8)); - d += dest_stride; - q8s16 = vsubq_s16(q11s16, q8s16); - q9s16 = vsubq_s16(q10s16, q9s16); - - // store the data out 8,9,10,11,12,13,14,15 - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q8s16 = vrshrq_n_s16(q8s16, 6); - q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_s64(d12s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q9s16 = vrshrq_n_s16(q9s16, 6); - q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_s64(d12s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q2s16 = vrshrq_n_s16(q2s16, 6); - q2u16 = vaddw_u8(vreinterpretq_u16_s16(q2s16), vreinterpret_u8_s64(d12s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q2u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q3s16 = vrshrq_n_s16(q3s16, 6); - q3u16 = vaddw_u8(vreinterpretq_u16_s16(q3s16), vreinterpret_u8_s64(d12s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q3u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q4s16 = vrshrq_n_s16(q4s16, 6); - q4u16 = vaddw_u8(vreinterpretq_u16_s16(q4s16), vreinterpret_u8_s64(d12s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q4u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q5s16 = vrshrq_n_s16(q5s16, 6); - q5u16 = vaddw_u8(vreinterpretq_u16_s16(q5s16), vreinterpret_u8_s64(d12s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q5u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - - d12s64 = vld1_s64((int64_t *)dest); - dest += dest_stride; - q14s16 = vrshrq_n_s16(q14s16, 6); - q14u16 = - vaddw_u8(vreinterpretq_u16_s16(q14s16), vreinterpret_u8_s64(d12s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q14u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - d += dest_stride; - - d12s64 = vld1_s64((int64_t *)dest); - q15s16 = vrshrq_n_s16(q15s16, 6); - q15u16 = - vaddw_u8(vreinterpretq_u16_s16(q15s16), vreinterpret_u8_s64(d12s64)); - d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q15u16)); - vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8)); - } else { // skip_adding_dest - q0s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q1s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q12s16 = vaddq_s16(q0s16, q15s16); - q13s16 = vaddq_s16(q1s16, q14s16); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - vst1_u64((uint64_t *)out, d24u64); - out += 4; - vst1_u64((uint64_t *)out, d25u64); - out += 12; - vst1_u64((uint64_t *)out, d26u64); - out += 4; - vst1_u64((uint64_t *)out, d27u64); - out += 12; - q14s16 = vsubq_s16(q1s16, q14s16); - q15s16 = vsubq_s16(q0s16, q15s16); - - q10s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q11s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q12s16 = vaddq_s16(q10s16, q5s16); - q13s16 = vaddq_s16(q11s16, q4s16); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - vst1_u64((uint64_t *)out, d24u64); - out += 4; - vst1_u64((uint64_t *)out, d25u64); - out += 12; - vst1_u64((uint64_t *)out, d26u64); - out += 4; - vst1_u64((uint64_t *)out, d27u64); - out += 12; - q4s16 = vsubq_s16(q11s16, q4s16); - q5s16 = vsubq_s16(q10s16, q5s16); - - q0s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q1s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q12s16 = vaddq_s16(q0s16, q3s16); - q13s16 = vaddq_s16(q1s16, q2s16); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - vst1_u64((uint64_t *)out, d24u64); - out += 4; - vst1_u64((uint64_t *)out, d25u64); - out += 12; - vst1_u64((uint64_t *)out, d26u64); - out += 4; - vst1_u64((uint64_t *)out, d27u64); - out += 12; - q2s16 = vsubq_s16(q1s16, q2s16); - q3s16 = vsubq_s16(q0s16, q3s16); - - q10s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q11s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q12s16 = vaddq_s16(q10s16, q9s16); - q13s16 = vaddq_s16(q11s16, q8s16); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - vst1_u64((uint64_t *)out, d24u64); - out += 4; - vst1_u64((uint64_t *)out, d25u64); - out += 12; - vst1_u64((uint64_t *)out, d26u64); - out += 4; - vst1_u64((uint64_t *)out, d27u64); - out += 12; - q8s16 = vsubq_s16(q11s16, q8s16); - q9s16 = vsubq_s16(q10s16, q9s16); - - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q8s16))); - out += 4; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q8s16))); - out += 12; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q9s16))); - out += 4; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q9s16))); - out += 12; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q2s16))); - out += 4; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q2s16))); - out += 12; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q3s16))); - out += 4; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q3s16))); - out += 12; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q4s16))); - out += 4; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q4s16))); - out += 12; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q5s16))); - out += 4; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q5s16))); - out += 12; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q14s16))); - out += 4; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q14s16))); - out += 12; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q15s16))); - out += 4; - vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q15s16))); - } - return; -} - -void aom_idct16x16_10_add_neon_pass1(int16_t *in, int16_t *out, - int output_stride) { - int16x4_t d4s16; - int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16; - uint64x1_t d4u64, d5u64, d18u64, d19u64, d20u64, d21u64, d22u64, d23u64; - uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64; - int16x8_t q0s16, q1s16, q2s16, q4s16, q5s16, q6s16, q7s16; - int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16; - int32x4_t q6s32, q9s32; - int32x4_t q10s32, q11s32, q12s32, q15s32; - int16x8x2_t q0x2s16; - - q0x2s16 = vld2q_s16(in); - q8s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q9s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q10s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q11s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q12s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q13s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q14s16 = q0x2s16.val[0]; - in += 16; - q0x2s16 = vld2q_s16(in); - q15s16 = q0x2s16.val[0]; - - TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - // stage 3 - q0s16 = vdupq_n_s16((int16_t)(cospi_28_64 * 2)); - q1s16 = vdupq_n_s16((int16_t)(cospi_4_64 * 2)); - - q4s16 = vqrdmulhq_s16(q9s16, q0s16); - q7s16 = vqrdmulhq_s16(q9s16, q1s16); - - // stage 4 - q1s16 = vdupq_n_s16((int16_t)(cospi_16_64 * 2)); - d4s16 = vdup_n_s16((int16_t)cospi_16_64); - - q8s16 = vqrdmulhq_s16(q8s16, q1s16); - - d8s16 = vget_low_s16(q4s16); - d9s16 = vget_high_s16(q4s16); - d14s16 = vget_low_s16(q7s16); - d15s16 = vget_high_s16(q7s16); - q9s32 = vmull_s16(d14s16, d4s16); - q10s32 = vmull_s16(d15s16, d4s16); - q12s32 = vmull_s16(d9s16, d4s16); - q11s32 = vmull_s16(d8s16, d4s16); - - q15s32 = vsubq_s32(q10s32, q12s32); - q6s32 = vsubq_s32(q9s32, q11s32); - q9s32 = vaddq_s32(q9s32, q11s32); - q10s32 = vaddq_s32(q10s32, q12s32); - - d11s16 = vqrshrn_n_s32(q15s32, 14); - d10s16 = vqrshrn_n_s32(q6s32, 14); - d12s16 = vqrshrn_n_s32(q9s32, 14); - d13s16 = vqrshrn_n_s32(q10s32, 14); - q5s16 = vcombine_s16(d10s16, d11s16); - q6s16 = vcombine_s16(d12s16, d13s16); - - // stage 6 - q2s16 = vaddq_s16(q8s16, q7s16); - q9s16 = vaddq_s16(q8s16, q6s16); - q10s16 = vaddq_s16(q8s16, q5s16); - q11s16 = vaddq_s16(q8s16, q4s16); - q12s16 = vsubq_s16(q8s16, q4s16); - q13s16 = vsubq_s16(q8s16, q5s16); - q14s16 = vsubq_s16(q8s16, q6s16); - q15s16 = vsubq_s16(q8s16, q7s16); - - d4u64 = vreinterpret_u64_s16(vget_low_s16(q2s16)); - d5u64 = vreinterpret_u64_s16(vget_high_s16(q2s16)); - d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16)); - d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16)); - d20u64 = vreinterpret_u64_s16(vget_low_s16(q10s16)); - d21u64 = vreinterpret_u64_s16(vget_high_s16(q10s16)); - d22u64 = vreinterpret_u64_s16(vget_low_s16(q11s16)); - d23u64 = vreinterpret_u64_s16(vget_high_s16(q11s16)); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16)); - d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16)); - d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16)); - d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16)); - - // store the data - output_stride >>= 1; // output_stride / 2, out is int16_t - vst1_u64((uint64_t *)out, d4u64); - out += output_stride; - vst1_u64((uint64_t *)out, d5u64); - out += output_stride; - vst1_u64((uint64_t *)out, d18u64); - out += output_stride; - vst1_u64((uint64_t *)out, d19u64); - out += output_stride; - vst1_u64((uint64_t *)out, d20u64); - out += output_stride; - vst1_u64((uint64_t *)out, d21u64); - out += output_stride; - vst1_u64((uint64_t *)out, d22u64); - out += output_stride; - vst1_u64((uint64_t *)out, d23u64); - out += output_stride; - vst1_u64((uint64_t *)out, d24u64); - out += output_stride; - vst1_u64((uint64_t *)out, d25u64); - out += output_stride; - vst1_u64((uint64_t *)out, d26u64); - out += output_stride; - vst1_u64((uint64_t *)out, d27u64); - out += output_stride; - vst1_u64((uint64_t *)out, d28u64); - out += output_stride; - vst1_u64((uint64_t *)out, d29u64); - out += output_stride; - vst1_u64((uint64_t *)out, d30u64); - out += output_stride; - vst1_u64((uint64_t *)out, d31u64); - return; -} - -void aom_idct16x16_10_add_neon_pass2(int16_t *src, int16_t *out, - int16_t *pass1Output, int16_t skip_adding, - uint8_t *dest, int dest_stride) { - int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16; - int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16; - int16x4_t d20s16, d21s16, d22s16, d23s16; - int16x4_t d24s16, d25s16, d26s16, d27s16, d30s16, d31s16; - uint64x1_t d4u64, d5u64, d6u64, d7u64, d8u64, d9u64, d10u64, d11u64; - uint64x1_t d16u64, d17u64, d18u64, d19u64; - uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64; - int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16; - int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16; - int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32; - int32x4_t q10s32, q11s32, q12s32, q13s32; - int16x8x2_t q0x2s16; - (void)skip_adding; - (void)dest; - (void)dest_stride; - - q0x2s16 = vld2q_s16(src); - q8s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q9s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q10s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q11s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q12s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q13s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q14s16 = q0x2s16.val[0]; - src += 16; - q0x2s16 = vld2q_s16(src); - q15s16 = q0x2s16.val[0]; - - TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - // stage 3 - q6s16 = vdupq_n_s16((int16_t)(cospi_30_64 * 2)); - q0s16 = vqrdmulhq_s16(q8s16, q6s16); - q6s16 = vdupq_n_s16((int16_t)(cospi_2_64 * 2)); - q7s16 = vqrdmulhq_s16(q8s16, q6s16); - - q15s16 = vdupq_n_s16(-cospi_26_64 * 2); - q14s16 = vdupq_n_s16((int16_t)(cospi_6_64 * 2)); - q3s16 = vqrdmulhq_s16(q9s16, q15s16); - q4s16 = vqrdmulhq_s16(q9s16, q14s16); - - // stage 4 - d0s16 = vget_low_s16(q0s16); - d1s16 = vget_high_s16(q0s16); - d6s16 = vget_low_s16(q3s16); - d7s16 = vget_high_s16(q3s16); - d8s16 = vget_low_s16(q4s16); - d9s16 = vget_high_s16(q4s16); - d14s16 = vget_low_s16(q7s16); - d15s16 = vget_high_s16(q7s16); - - d30s16 = vdup_n_s16((int16_t)cospi_8_64); - d31s16 = vdup_n_s16((int16_t)cospi_24_64); - - q12s32 = vmull_s16(d14s16, d31s16); - q5s32 = vmull_s16(d15s16, d31s16); - q2s32 = vmull_s16(d0s16, d31s16); - q11s32 = vmull_s16(d1s16, d31s16); - - q12s32 = vmlsl_s16(q12s32, d0s16, d30s16); - q5s32 = vmlsl_s16(q5s32, d1s16, d30s16); - q2s32 = vmlal_s16(q2s32, d14s16, d30s16); - q11s32 = vmlal_s16(q11s32, d15s16, d30s16); - - d2s16 = vqrshrn_n_s32(q12s32, 14); - d3s16 = vqrshrn_n_s32(q5s32, 14); - d12s16 = vqrshrn_n_s32(q2s32, 14); - d13s16 = vqrshrn_n_s32(q11s32, 14); - q1s16 = vcombine_s16(d2s16, d3s16); - q6s16 = vcombine_s16(d12s16, d13s16); - - d30s16 = vdup_n_s16(-cospi_8_64); - q10s32 = vmull_s16(d8s16, d30s16); - q13s32 = vmull_s16(d9s16, d30s16); - q8s32 = vmull_s16(d6s16, d30s16); - q9s32 = vmull_s16(d7s16, d30s16); - - q10s32 = vmlsl_s16(q10s32, d6s16, d31s16); - q13s32 = vmlsl_s16(q13s32, d7s16, d31s16); - q8s32 = vmlal_s16(q8s32, d8s16, d31s16); - q9s32 = vmlal_s16(q9s32, d9s16, d31s16); - - d4s16 = vqrshrn_n_s32(q10s32, 14); - d5s16 = vqrshrn_n_s32(q13s32, 14); - d10s16 = vqrshrn_n_s32(q8s32, 14); - d11s16 = vqrshrn_n_s32(q9s32, 14); - q2s16 = vcombine_s16(d4s16, d5s16); - q5s16 = vcombine_s16(d10s16, d11s16); - - // stage 5 - q8s16 = vaddq_s16(q0s16, q3s16); - q9s16 = vaddq_s16(q1s16, q2s16); - q10s16 = vsubq_s16(q1s16, q2s16); - q11s16 = vsubq_s16(q0s16, q3s16); - q12s16 = vsubq_s16(q7s16, q4s16); - q13s16 = vsubq_s16(q6s16, q5s16); - q14s16 = vaddq_s16(q6s16, q5s16); - q15s16 = vaddq_s16(q7s16, q4s16); - - // stage 6 - d20s16 = vget_low_s16(q10s16); - d21s16 = vget_high_s16(q10s16); - d22s16 = vget_low_s16(q11s16); - d23s16 = vget_high_s16(q11s16); - d24s16 = vget_low_s16(q12s16); - d25s16 = vget_high_s16(q12s16); - d26s16 = vget_low_s16(q13s16); - d27s16 = vget_high_s16(q13s16); - - d14s16 = vdup_n_s16((int16_t)cospi_16_64); - q3s32 = vmull_s16(d26s16, d14s16); - q4s32 = vmull_s16(d27s16, d14s16); - q0s32 = vmull_s16(d20s16, d14s16); - q1s32 = vmull_s16(d21s16, d14s16); - - q5s32 = vsubq_s32(q3s32, q0s32); - q6s32 = vsubq_s32(q4s32, q1s32); - q0s32 = vaddq_s32(q3s32, q0s32); - q4s32 = vaddq_s32(q4s32, q1s32); - - d4s16 = vqrshrn_n_s32(q5s32, 14); - d5s16 = vqrshrn_n_s32(q6s32, 14); - d10s16 = vqrshrn_n_s32(q0s32, 14); - d11s16 = vqrshrn_n_s32(q4s32, 14); - q2s16 = vcombine_s16(d4s16, d5s16); - q5s16 = vcombine_s16(d10s16, d11s16); - - q0s32 = vmull_s16(d22s16, d14s16); - q1s32 = vmull_s16(d23s16, d14s16); - q13s32 = vmull_s16(d24s16, d14s16); - q6s32 = vmull_s16(d25s16, d14s16); - - q10s32 = vsubq_s32(q13s32, q0s32); - q4s32 = vsubq_s32(q6s32, q1s32); - q13s32 = vaddq_s32(q13s32, q0s32); - q6s32 = vaddq_s32(q6s32, q1s32); - - d6s16 = vqrshrn_n_s32(q10s32, 14); - d7s16 = vqrshrn_n_s32(q4s32, 14); - d8s16 = vqrshrn_n_s32(q13s32, 14); - d9s16 = vqrshrn_n_s32(q6s32, 14); - q3s16 = vcombine_s16(d6s16, d7s16); - q4s16 = vcombine_s16(d8s16, d9s16); - - // stage 7 - q0s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q1s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q12s16 = vaddq_s16(q0s16, q15s16); - q13s16 = vaddq_s16(q1s16, q14s16); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - vst1_u64((uint64_t *)out, d24u64); - out += 4; - vst1_u64((uint64_t *)out, d25u64); - out += 12; - vst1_u64((uint64_t *)out, d26u64); - out += 4; - vst1_u64((uint64_t *)out, d27u64); - out += 12; - q14s16 = vsubq_s16(q1s16, q14s16); - q15s16 = vsubq_s16(q0s16, q15s16); - - q10s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q11s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q12s16 = vaddq_s16(q10s16, q5s16); - q13s16 = vaddq_s16(q11s16, q4s16); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - vst1_u64((uint64_t *)out, d24u64); - out += 4; - vst1_u64((uint64_t *)out, d25u64); - out += 12; - vst1_u64((uint64_t *)out, d26u64); - out += 4; - vst1_u64((uint64_t *)out, d27u64); - out += 12; - q4s16 = vsubq_s16(q11s16, q4s16); - q5s16 = vsubq_s16(q10s16, q5s16); - - q0s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q1s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q12s16 = vaddq_s16(q0s16, q3s16); - q13s16 = vaddq_s16(q1s16, q2s16); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - vst1_u64((uint64_t *)out, d24u64); - out += 4; - vst1_u64((uint64_t *)out, d25u64); - out += 12; - vst1_u64((uint64_t *)out, d26u64); - out += 4; - vst1_u64((uint64_t *)out, d27u64); - out += 12; - q2s16 = vsubq_s16(q1s16, q2s16); - q3s16 = vsubq_s16(q0s16, q3s16); - - q10s16 = vld1q_s16(pass1Output); - pass1Output += 8; - q11s16 = vld1q_s16(pass1Output); - q12s16 = vaddq_s16(q10s16, q9s16); - q13s16 = vaddq_s16(q11s16, q8s16); - d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16)); - d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16)); - d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16)); - d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16)); - vst1_u64((uint64_t *)out, d24u64); - out += 4; - vst1_u64((uint64_t *)out, d25u64); - out += 12; - vst1_u64((uint64_t *)out, d26u64); - out += 4; - vst1_u64((uint64_t *)out, d27u64); - out += 12; - q8s16 = vsubq_s16(q11s16, q8s16); - q9s16 = vsubq_s16(q10s16, q9s16); - - d4u64 = vreinterpret_u64_s16(vget_low_s16(q2s16)); - d5u64 = vreinterpret_u64_s16(vget_high_s16(q2s16)); - d6u64 = vreinterpret_u64_s16(vget_low_s16(q3s16)); - d7u64 = vreinterpret_u64_s16(vget_high_s16(q3s16)); - d8u64 = vreinterpret_u64_s16(vget_low_s16(q4s16)); - d9u64 = vreinterpret_u64_s16(vget_high_s16(q4s16)); - d10u64 = vreinterpret_u64_s16(vget_low_s16(q5s16)); - d11u64 = vreinterpret_u64_s16(vget_high_s16(q5s16)); - d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16)); - d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16)); - d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16)); - d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16)); - d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16)); - d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16)); - d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16)); - d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16)); - - vst1_u64((uint64_t *)out, d16u64); - out += 4; - vst1_u64((uint64_t *)out, d17u64); - out += 12; - vst1_u64((uint64_t *)out, d18u64); - out += 4; - vst1_u64((uint64_t *)out, d19u64); - out += 12; - vst1_u64((uint64_t *)out, d4u64); - out += 4; - vst1_u64((uint64_t *)out, d5u64); - out += 12; - vst1_u64((uint64_t *)out, d6u64); - out += 4; - vst1_u64((uint64_t *)out, d7u64); - out += 12; - vst1_u64((uint64_t *)out, d8u64); - out += 4; - vst1_u64((uint64_t *)out, d9u64); - out += 12; - vst1_u64((uint64_t *)out, d10u64); - out += 4; - vst1_u64((uint64_t *)out, d11u64); - out += 12; - vst1_u64((uint64_t *)out, d28u64); - out += 4; - vst1_u64((uint64_t *)out, d29u64); - out += 12; - vst1_u64((uint64_t *)out, d30u64); - out += 4; - vst1_u64((uint64_t *)out, d31u64); - return; -} diff --git a/aom_dsp/arm/idct16x16_neon.c b/aom_dsp/arm/idct16x16_neon.c deleted file mode 100644 index db0d4905b..000000000 --- a/aom_dsp/arm/idct16x16_neon.c +++ /dev/null @@ -1,152 +0,0 @@ -/* - * Copyright (c) 2016, 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 "aom_dsp/aom_dsp_common.h" - -void aom_idct16x16_256_add_neon_pass1(const int16_t *input, int16_t *output, - int output_stride); -void aom_idct16x16_256_add_neon_pass2(const int16_t *src, int16_t *output, - int16_t *pass1Output, int16_t skip_adding, - uint8_t *dest, int dest_stride); -void aom_idct16x16_10_add_neon_pass1(const int16_t *input, int16_t *output, - int output_stride); -void aom_idct16x16_10_add_neon_pass2(const int16_t *src, int16_t *output, - int16_t *pass1Output, int16_t skip_adding, - uint8_t *dest, int dest_stride); - -#if HAVE_NEON_ASM -/* For ARM NEON, d8-d15 are callee-saved registers, and need to be saved. */ -extern void aom_push_neon(int64_t *store); -extern void aom_pop_neon(int64_t *store); -#endif // HAVE_NEON_ASM - -void aom_idct16x16_256_add_neon(const int16_t *input, uint8_t *dest, - int dest_stride) { -#if HAVE_NEON_ASM - int64_t store_reg[8]; -#endif - int16_t pass1_output[16 * 16] = { 0 }; - int16_t row_idct_output[16 * 16] = { 0 }; - -#if HAVE_NEON_ASM - // save d8-d15 register values. - aom_push_neon(store_reg); -#endif - - /* Parallel idct on the upper 8 rows */ - // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the - // stage 6 result in pass1_output. - aom_idct16x16_256_add_neon_pass1(input, pass1_output, 8); - - // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines - // with result in pass1(pass1_output) to calculate final result in stage 7 - // which will be saved into row_idct_output. - aom_idct16x16_256_add_neon_pass2(input + 1, row_idct_output, pass1_output, 0, - dest, dest_stride); - - /* Parallel idct on the lower 8 rows */ - // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the - // stage 6 result in pass1_output. - aom_idct16x16_256_add_neon_pass1(input + 8 * 16, pass1_output, 8); - - // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines - // with result in pass1(pass1_output) to calculate final result in stage 7 - // which will be saved into row_idct_output. - aom_idct16x16_256_add_neon_pass2(input + 8 * 16 + 1, row_idct_output + 8, - pass1_output, 0, dest, dest_stride); - - /* Parallel idct on the left 8 columns */ - // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the - // stage 6 result in pass1_output. - aom_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8); - - // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines - // with result in pass1(pass1_output) to calculate final result in stage 7. - // Then add the result to the destination data. - aom_idct16x16_256_add_neon_pass2(row_idct_output + 1, row_idct_output, - pass1_output, 1, dest, dest_stride); - - /* Parallel idct on the right 8 columns */ - // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the - // stage 6 result in pass1_output. - aom_idct16x16_256_add_neon_pass1(row_idct_output + 8 * 16, pass1_output, 8); - - // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines - // with result in pass1(pass1_output) to calculate final result in stage 7. - // Then add the result to the destination data. - aom_idct16x16_256_add_neon_pass2(row_idct_output + 8 * 16 + 1, - row_idct_output + 8, pass1_output, 1, - dest + 8, dest_stride); - -#if HAVE_NEON_ASM - // restore d8-d15 register values. - aom_pop_neon(store_reg); -#endif - - return; -} - -void aom_idct16x16_10_add_neon(const int16_t *input, uint8_t *dest, - int dest_stride) { -#if HAVE_NEON_ASM - int64_t store_reg[8]; -#endif - int16_t pass1_output[16 * 16] = { 0 }; - int16_t row_idct_output[16 * 16] = { 0 }; - -#if HAVE_NEON_ASM - // save d8-d15 register values. - aom_push_neon(store_reg); -#endif - - /* Parallel idct on the upper 8 rows */ - // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the - // stage 6 result in pass1_output. - aom_idct16x16_10_add_neon_pass1(input, pass1_output, 8); - - // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines - // with result in pass1(pass1_output) to calculate final result in stage 7 - // which will be saved into row_idct_output. - aom_idct16x16_10_add_neon_pass2(input + 1, row_idct_output, pass1_output, 0, - dest, dest_stride); - - /* Skip Parallel idct on the lower 8 rows as they are all 0s */ - - /* Parallel idct on the left 8 columns */ - // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the - // stage 6 result in pass1_output. - aom_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8); - - // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines - // with result in pass1(pass1_output) to calculate final result in stage 7. - // Then add the result to the destination data. - aom_idct16x16_256_add_neon_pass2(row_idct_output + 1, row_idct_output, - pass1_output, 1, dest, dest_stride); - - /* Parallel idct on the right 8 columns */ - // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the - // stage 6 result in pass1_output. - aom_idct16x16_256_add_neon_pass1(row_idct_output + 8 * 16, pass1_output, 8); - - // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines - // with result in pass1(pass1_output) to calculate final result in stage 7. - // Then add the result to the destination data. - aom_idct16x16_256_add_neon_pass2(row_idct_output + 8 * 16 + 1, - row_idct_output + 8, pass1_output, 1, - dest + 8, dest_stride); - -#if HAVE_NEON_ASM - // restore d8-d15 register values. - aom_pop_neon(store_reg); -#endif - - return; -} diff --git a/aom_dsp/arm/idct32x32_1_add_neon.asm b/aom_dsp/arm/idct32x32_1_add_neon.asm deleted file mode 100644 index b04df2d0b..000000000 --- a/aom_dsp/arm/idct32x32_1_add_neon.asm +++ /dev/null @@ -1,147 +0,0 @@ -; -; Copyright (c) 2016, 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. -; - - - EXPORT |aom_idct32x32_1_add_neon| - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - - ;TODO(hkuang): put the following macros in a seperate - ;file so other idct function could also use them. - MACRO - LD_16x8 $src, $stride - vld1.8 {q8}, [$src], $stride - vld1.8 {q9}, [$src], $stride - vld1.8 {q10}, [$src], $stride - vld1.8 {q11}, [$src], $stride - vld1.8 {q12}, [$src], $stride - vld1.8 {q13}, [$src], $stride - vld1.8 {q14}, [$src], $stride - vld1.8 {q15}, [$src], $stride - MEND - - MACRO - ADD_DIFF_16x8 $diff - vqadd.u8 q8, q8, $diff - vqadd.u8 q9, q9, $diff - vqadd.u8 q10, q10, $diff - vqadd.u8 q11, q11, $diff - vqadd.u8 q12, q12, $diff - vqadd.u8 q13, q13, $diff - vqadd.u8 q14, q14, $diff - vqadd.u8 q15, q15, $diff - MEND - - MACRO - SUB_DIFF_16x8 $diff - vqsub.u8 q8, q8, $diff - vqsub.u8 q9, q9, $diff - vqsub.u8 q10, q10, $diff - vqsub.u8 q11, q11, $diff - vqsub.u8 q12, q12, $diff - vqsub.u8 q13, q13, $diff - vqsub.u8 q14, q14, $diff - vqsub.u8 q15, q15, $diff - MEND - - MACRO - ST_16x8 $dst, $stride - vst1.8 {q8}, [$dst], $stride - vst1.8 {q9}, [$dst], $stride - vst1.8 {q10},[$dst], $stride - vst1.8 {q11},[$dst], $stride - vst1.8 {q12},[$dst], $stride - vst1.8 {q13},[$dst], $stride - vst1.8 {q14},[$dst], $stride - vst1.8 {q15},[$dst], $stride - MEND - -;void aom_idct32x32_1_add_neon(int16_t *input, uint8_t *dest, -; int dest_stride) -; -; r0 int16_t input -; r1 uint8_t *dest -; r2 int dest_stride - -|aom_idct32x32_1_add_neon| PROC - push {lr} - pld [r1] - add r3, r1, #16 ; r3 dest + 16 for second loop - ldrsh r0, [r0] - - ; generate cospi_16_64 = 11585 - mov r12, #0x2d00 - add r12, #0x41 - - ; out = dct_const_round_shift(input[0] * cospi_16_64) - mul r0, r0, r12 ; input[0] * cospi_16_64 - add r0, r0, #0x2000 ; +(1 << ((DCT_CONST_BITS) - 1)) - asr r0, r0, #14 ; >> DCT_CONST_BITS - - ; out = dct_const_round_shift(out * cospi_16_64) - mul r0, r0, r12 ; out * cospi_16_64 - mov r12, r1 ; save dest - add r0, r0, #0x2000 ; +(1 << ((DCT_CONST_BITS) - 1)) - asr r0, r0, #14 ; >> DCT_CONST_BITS - - ; a1 = ROUND_POWER_OF_TWO(out, 6) - add r0, r0, #32 ; + (1 <<((6) - 1)) - asrs r0, r0, #6 ; >> 6 - bge diff_positive_32_32 - -diff_negative_32_32 - neg r0, r0 - usat r0, #8, r0 - vdup.u8 q0, r0 - mov r0, #4 - -diff_negative_32_32_loop - sub r0, #1 - LD_16x8 r1, r2 - SUB_DIFF_16x8 q0 - ST_16x8 r12, r2 - - LD_16x8 r1, r2 - SUB_DIFF_16x8 q0 - ST_16x8 r12, r2 - cmp r0, #2 - moveq r1, r3 - moveq r12, r3 - cmp r0, #0 - bne diff_negative_32_32_loop - pop {pc} - -diff_positive_32_32 - usat r0, #8, r0 - vdup.u8 q0, r0 - mov r0, #4 - -diff_positive_32_32_loop - sub r0, #1 - LD_16x8 r1, r2 - ADD_DIFF_16x8 q0 - ST_16x8 r12, r2 - - LD_16x8 r1, r2 - ADD_DIFF_16x8 q0 - ST_16x8 r12, r2 - cmp r0, #2 - moveq r1, r3 - moveq r12, r3 - cmp r0, #0 - bne diff_positive_32_32_loop - pop {pc} - - ENDP ; |aom_idct32x32_1_add_neon| - END diff --git a/aom_dsp/arm/idct32x32_1_add_neon.c b/aom_dsp/arm/idct32x32_1_add_neon.c deleted file mode 100644 index 547567c5b..000000000 --- a/aom_dsp/arm/idct32x32_1_add_neon.c +++ /dev/null @@ -1,141 +0,0 @@ -/* - * Copyright (c) 2016, 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 "./aom_config.h" - -#include "aom_dsp/inv_txfm.h" -#include "aom_ports/mem.h" - -static INLINE void LD_16x8(uint8_t *d, int d_stride, uint8x16_t *q8u8, - uint8x16_t *q9u8, uint8x16_t *q10u8, - uint8x16_t *q11u8, uint8x16_t *q12u8, - uint8x16_t *q13u8, uint8x16_t *q14u8, - uint8x16_t *q15u8) { - *q8u8 = vld1q_u8(d); - d += d_stride; - *q9u8 = vld1q_u8(d); - d += d_stride; - *q10u8 = vld1q_u8(d); - d += d_stride; - *q11u8 = vld1q_u8(d); - d += d_stride; - *q12u8 = vld1q_u8(d); - d += d_stride; - *q13u8 = vld1q_u8(d); - d += d_stride; - *q14u8 = vld1q_u8(d); - d += d_stride; - *q15u8 = vld1q_u8(d); - return; -} - -static INLINE void ADD_DIFF_16x8(uint8x16_t qdiffu8, uint8x16_t *q8u8, - uint8x16_t *q9u8, uint8x16_t *q10u8, - uint8x16_t *q11u8, uint8x16_t *q12u8, - uint8x16_t *q13u8, uint8x16_t *q14u8, - uint8x16_t *q15u8) { - *q8u8 = vqaddq_u8(*q8u8, qdiffu8); - *q9u8 = vqaddq_u8(*q9u8, qdiffu8); - *q10u8 = vqaddq_u8(*q10u8, qdiffu8); - *q11u8 = vqaddq_u8(*q11u8, qdiffu8); - *q12u8 = vqaddq_u8(*q12u8, qdiffu8); - *q13u8 = vqaddq_u8(*q13u8, qdiffu8); - *q14u8 = vqaddq_u8(*q14u8, qdiffu8); - *q15u8 = vqaddq_u8(*q15u8, qdiffu8); - return; -} - -static INLINE void SUB_DIFF_16x8(uint8x16_t qdiffu8, uint8x16_t *q8u8, - uint8x16_t *q9u8, uint8x16_t *q10u8, - uint8x16_t *q11u8, uint8x16_t *q12u8, - uint8x16_t *q13u8, uint8x16_t *q14u8, - uint8x16_t *q15u8) { - *q8u8 = vqsubq_u8(*q8u8, qdiffu8); - *q9u8 = vqsubq_u8(*q9u8, qdiffu8); - *q10u8 = vqsubq_u8(*q10u8, qdiffu8); - *q11u8 = vqsubq_u8(*q11u8, qdiffu8); - *q12u8 = vqsubq_u8(*q12u8, qdiffu8); - *q13u8 = vqsubq_u8(*q13u8, qdiffu8); - *q14u8 = vqsubq_u8(*q14u8, qdiffu8); - *q15u8 = vqsubq_u8(*q15u8, qdiffu8); - return; -} - -static INLINE void ST_16x8(uint8_t *d, int d_stride, uint8x16_t *q8u8, - uint8x16_t *q9u8, uint8x16_t *q10u8, - uint8x16_t *q11u8, uint8x16_t *q12u8, - uint8x16_t *q13u8, uint8x16_t *q14u8, - uint8x16_t *q15u8) { - vst1q_u8(d, *q8u8); - d += d_stride; - vst1q_u8(d, *q9u8); - d += d_stride; - vst1q_u8(d, *q10u8); - d += d_stride; - vst1q_u8(d, *q11u8); - d += d_stride; - vst1q_u8(d, *q12u8); - d += d_stride; - vst1q_u8(d, *q13u8); - d += d_stride; - vst1q_u8(d, *q14u8); - d += d_stride; - vst1q_u8(d, *q15u8); - return; -} - -void aom_idct32x32_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) { - uint8x16_t q0u8, q8u8, q9u8, q10u8, q11u8, q12u8, q13u8, q14u8, q15u8; - int i, j, dest_stride8; - uint8_t *d; - int16_t a1; - int16_t out = dct_const_round_shift(input[0] * cospi_16_64); - - out = dct_const_round_shift(out * cospi_16_64); - a1 = ROUND_POWER_OF_TWO(out, 6); - - dest_stride8 = dest_stride * 8; - if (a1 >= 0) { // diff_positive_32_32 - a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1; - q0u8 = vdupq_n_u8(a1); - for (i = 0; i < 2; i++, dest += 16) { // diff_positive_32_32_loop - d = dest; - for (j = 0; j < 4; j++) { - LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8, - &q14u8, &q15u8); - ADD_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8, - &q14u8, &q15u8); - ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8, - &q14u8, &q15u8); - d += dest_stride8; - } - } - } else { // diff_negative_32_32 - a1 = -a1; - a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1; - q0u8 = vdupq_n_u8(a1); - for (i = 0; i < 2; i++, dest += 16) { // diff_negative_32_32_loop - d = dest; - for (j = 0; j < 4; j++) { - LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8, - &q14u8, &q15u8); - SUB_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8, - &q14u8, &q15u8); - ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8, - &q14u8, &q15u8); - d += dest_stride8; - } - } - } - return; -} diff --git a/aom_dsp/arm/idct32x32_add_neon.asm b/aom_dsp/arm/idct32x32_add_neon.asm deleted file mode 100644 index e7793fb16..000000000 --- a/aom_dsp/arm/idct32x32_add_neon.asm +++ /dev/null @@ -1,1302 +0,0 @@ -; -; Copyright (c) 2016, 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. -; - -; - -;TODO(cd): adjust these constant to be able to use vqdmulh for faster -; dct_const_round_shift(a * b) within butterfly calculations. -cospi_1_64 EQU 16364 -cospi_2_64 EQU 16305 -cospi_3_64 EQU 16207 -cospi_4_64 EQU 16069 -cospi_5_64 EQU 15893 -cospi_6_64 EQU 15679 -cospi_7_64 EQU 15426 -cospi_8_64 EQU 15137 -cospi_9_64 EQU 14811 -cospi_10_64 EQU 14449 -cospi_11_64 EQU 14053 -cospi_12_64 EQU 13623 -cospi_13_64 EQU 13160 -cospi_14_64 EQU 12665 -cospi_15_64 EQU 12140 -cospi_16_64 EQU 11585 -cospi_17_64 EQU 11003 -cospi_18_64 EQU 10394 -cospi_19_64 EQU 9760 -cospi_20_64 EQU 9102 -cospi_21_64 EQU 8423 -cospi_22_64 EQU 7723 -cospi_23_64 EQU 7005 -cospi_24_64 EQU 6270 -cospi_25_64 EQU 5520 -cospi_26_64 EQU 4756 -cospi_27_64 EQU 3981 -cospi_28_64 EQU 3196 -cospi_29_64 EQU 2404 -cospi_30_64 EQU 1606 -cospi_31_64 EQU 804 - - - EXPORT |aom_idct32x32_1024_add_neon| - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - - AREA Block, CODE, READONLY - - ; -------------------------------------------------------------------------- - ; Load from transposed_buffer - ; q13 = transposed_buffer[first_offset] - ; q14 = transposed_buffer[second_offset] - ; for proper address calculation, the last offset used when manipulating - ; transposed_buffer must be passed in. use 0 for first use. - MACRO - LOAD_FROM_TRANSPOSED $prev_offset, $first_offset, $second_offset - ; address calculation with proper stride and loading - add r0, #($first_offset - $prev_offset )*8*2 - vld1.s16 {q14}, [r0] - add r0, #($second_offset - $first_offset)*8*2 - vld1.s16 {q13}, [r0] - ; (used) two registers (q14, q13) - MEND - ; -------------------------------------------------------------------------- - ; Load from output (used as temporary storage) - ; reg1 = output[first_offset] - ; reg2 = output[second_offset] - ; for proper address calculation, the last offset used when manipulating - ; output, whether reading or storing) must be passed in. use 0 for first - ; use. - MACRO - LOAD_FROM_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2 - ; address calculation with proper stride and loading - add r1, #($first_offset - $prev_offset )*32*2 - vld1.s16 {$reg1}, [r1] - add r1, #($second_offset - $first_offset)*32*2 - vld1.s16 {$reg2}, [r1] - ; (used) two registers ($reg1, $reg2) - MEND - ; -------------------------------------------------------------------------- - ; Store into output (sometimes as as temporary storage) - ; output[first_offset] = reg1 - ; output[second_offset] = reg2 - ; for proper address calculation, the last offset used when manipulating - ; output, whether reading or storing) must be passed in. use 0 for first - ; use. - MACRO - STORE_IN_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2 - ; address calculation with proper stride and storing - add r1, #($first_offset - $prev_offset )*32*2 - vst1.16 {$reg1}, [r1] - add r1, #($second_offset - $first_offset)*32*2 - vst1.16 {$reg2}, [r1] - MEND - ; -------------------------------------------------------------------------- - ; Combine-add results with current destination content - ; q6-q9 contain the results (out[j * 32 + 0-31]) - MACRO - STORE_COMBINE_CENTER_RESULTS - ; load dest[j * dest_stride + 0-31] - vld1.s16 {d8}, [r10], r2 - vld1.s16 {d11}, [r9], r11 - vld1.s16 {d9}, [r10] - vld1.s16 {d10}, [r9] - ; ROUND_POWER_OF_TWO - vrshr.s16 q7, q7, #6 - vrshr.s16 q8, q8, #6 - vrshr.s16 q9, q9, #6 - vrshr.s16 q6, q6, #6 - ; add to dest[j * dest_stride + 0-31] - vaddw.u8 q7, q7, d9 - vaddw.u8 q8, q8, d10 - vaddw.u8 q9, q9, d11 - vaddw.u8 q6, q6, d8 - ; clip pixel - vqmovun.s16 d9, q7 - vqmovun.s16 d10, q8 - vqmovun.s16 d11, q9 - vqmovun.s16 d8, q6 - ; store back into dest[j * dest_stride + 0-31] - vst1.16 {d9}, [r10], r11 - vst1.16 {d10}, [r9], r2 - vst1.16 {d8}, [r10] - vst1.16 {d11}, [r9] - ; update pointers (by dest_stride * 2) - sub r9, r9, r2, lsl #1 - add r10, r10, r2, lsl #1 - MEND - ; -------------------------------------------------------------------------- - ; Combine-add results with current destination content - ; q6-q9 contain the results (out[j * 32 + 0-31]) - MACRO - STORE_COMBINE_CENTER_RESULTS_LAST - ; load dest[j * dest_stride + 0-31] - vld1.s16 {d8}, [r10], r2 - vld1.s16 {d11}, [r9], r11 - vld1.s16 {d9}, [r10] - vld1.s16 {d10}, [r9] - ; ROUND_POWER_OF_TWO - vrshr.s16 q7, q7, #6 - vrshr.s16 q8, q8, #6 - vrshr.s16 q9, q9, #6 - vrshr.s16 q6, q6, #6 - ; add to dest[j * dest_stride + 0-31] - vaddw.u8 q7, q7, d9 - vaddw.u8 q8, q8, d10 - vaddw.u8 q9, q9, d11 - vaddw.u8 q6, q6, d8 - ; clip pixel - vqmovun.s16 d9, q7 - vqmovun.s16 d10, q8 - vqmovun.s16 d11, q9 - vqmovun.s16 d8, q6 - ; store back into dest[j * dest_stride + 0-31] - vst1.16 {d9}, [r10], r11 - vst1.16 {d10}, [r9], r2 - vst1.16 {d8}, [r10]! - vst1.16 {d11}, [r9]! - ; update pointers (by dest_stride * 2) - sub r9, r9, r2, lsl #1 - add r10, r10, r2, lsl #1 - MEND - ; -------------------------------------------------------------------------- - ; Combine-add results with current destination content - ; q4-q7 contain the results (out[j * 32 + 0-31]) - MACRO - STORE_COMBINE_EXTREME_RESULTS - ; load dest[j * dest_stride + 0-31] - vld1.s16 {d4}, [r7], r2 - vld1.s16 {d7}, [r6], r11 - vld1.s16 {d5}, [r7] - vld1.s16 {d6}, [r6] - ; ROUND_POWER_OF_TWO - vrshr.s16 q5, q5, #6 - vrshr.s16 q6, q6, #6 - vrshr.s16 q7, q7, #6 - vrshr.s16 q4, q4, #6 - ; add to dest[j * dest_stride + 0-31] - vaddw.u8 q5, q5, d5 - vaddw.u8 q6, q6, d6 - vaddw.u8 q7, q7, d7 - vaddw.u8 q4, q4, d4 - ; clip pixel - vqmovun.s16 d5, q5 - vqmovun.s16 d6, q6 - vqmovun.s16 d7, q7 - vqmovun.s16 d4, q4 - ; store back into dest[j * dest_stride + 0-31] - vst1.16 {d5}, [r7], r11 - vst1.16 {d6}, [r6], r2 - vst1.16 {d7}, [r6] - vst1.16 {d4}, [r7] - ; update pointers (by dest_stride * 2) - sub r6, r6, r2, lsl #1 - add r7, r7, r2, lsl #1 - MEND - ; -------------------------------------------------------------------------- - ; Combine-add results with current destination content - ; q4-q7 contain the results (out[j * 32 + 0-31]) - MACRO - STORE_COMBINE_EXTREME_RESULTS_LAST - ; load dest[j * dest_stride + 0-31] - vld1.s16 {d4}, [r7], r2 - vld1.s16 {d7}, [r6], r11 - vld1.s16 {d5}, [r7] - vld1.s16 {d6}, [r6] - ; ROUND_POWER_OF_TWO - vrshr.s16 q5, q5, #6 - vrshr.s16 q6, q6, #6 - vrshr.s16 q7, q7, #6 - vrshr.s16 q4, q4, #6 - ; add to dest[j * dest_stride + 0-31] - vaddw.u8 q5, q5, d5 - vaddw.u8 q6, q6, d6 - vaddw.u8 q7, q7, d7 - vaddw.u8 q4, q4, d4 - ; clip pixel - vqmovun.s16 d5, q5 - vqmovun.s16 d6, q6 - vqmovun.s16 d7, q7 - vqmovun.s16 d4, q4 - ; store back into dest[j * dest_stride + 0-31] - vst1.16 {d5}, [r7], r11 - vst1.16 {d6}, [r6], r2 - vst1.16 {d7}, [r6]! - vst1.16 {d4}, [r7]! - ; update pointers (by dest_stride * 2) - sub r6, r6, r2, lsl #1 - add r7, r7, r2, lsl #1 - MEND - ; -------------------------------------------------------------------------- - ; Touches q8-q12, q15 (q13-q14 are preserved) - ; valid output registers are anything but q8-q11 - MACRO - DO_BUTTERFLY $regC, $regD, $regA, $regB, $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4 - ; TODO(cd): have special case to re-use constants when they are similar for - ; consecutive butterflies - ; TODO(cd): have special case when both constants are the same, do the - ; additions/subtractions before the multiplies. - ; generate the constants - ; generate scalar constants - mov r8, #$first_constant & 0xFF00 - mov r12, #$second_constant & 0xFF00 - add r8, #$first_constant & 0x00FF - add r12, #$second_constant & 0x00FF - ; generate vector constants - vdup.16 d30, r8 - vdup.16 d31, r12 - ; (used) two for inputs (regA-regD), one for constants (q15) - ; do some multiplications (ordered for maximum latency hiding) - vmull.s16 q8, $regC, d30 - vmull.s16 q10, $regA, d31 - vmull.s16 q9, $regD, d30 - vmull.s16 q11, $regB, d31 - vmull.s16 q12, $regC, d31 - ; (used) five for intermediate (q8-q12), one for constants (q15) - ; do some addition/subtractions (to get back two register) - vsub.s32 q8, q8, q10 - vsub.s32 q9, q9, q11 - ; do more multiplications (ordered for maximum latency hiding) - vmull.s16 q10, $regD, d31 - vmull.s16 q11, $regA, d30 - vmull.s16 q15, $regB, d30 - ; (used) six for intermediate (q8-q12, q15) - ; do more addition/subtractions - vadd.s32 q11, q12, q11 - vadd.s32 q10, q10, q15 - ; (used) four for intermediate (q8-q11) - ; dct_const_round_shift - vqrshrn.s32 $reg1, q8, #14 - vqrshrn.s32 $reg2, q9, #14 - vqrshrn.s32 $reg3, q11, #14 - vqrshrn.s32 $reg4, q10, #14 - ; (used) two for results, well four d registers - MEND - ; -------------------------------------------------------------------------- - ; Touches q8-q12, q15 (q13-q14 are preserved) - ; valid output registers are anything but q8-q11 - MACRO - DO_BUTTERFLY_STD $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4 - DO_BUTTERFLY d28, d29, d26, d27, $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4 - MEND - ; -------------------------------------------------------------------------- - -;void aom_idct32x32_1024_add_neon(int16_t *input, uint8_t *dest, int dest_stride); -; -; r0 int16_t *input, -; r1 uint8_t *dest, -; r2 int dest_stride) -; loop counters -; r4 bands loop counter -; r5 pass loop counter -; r8 transpose loop counter -; combine-add pointers -; r6 dest + 31 * dest_stride, descending (30, 29, 28, ...) -; r7 dest + 0 * dest_stride, ascending (1, 2, 3, ...) -; r9 dest + 15 * dest_stride, descending (14, 13, 12, ...) -; r10 dest + 16 * dest_stride, ascending (17, 18, 19, ...) - -|aom_idct32x32_1024_add_neon| PROC - ; This function does one pass of idct32x32 transform. - ; - ; This is done by transposing the input and then doing a 1d transform on - ; columns. In the first pass, the transposed columns are the original - ; rows. In the second pass, after the transposition, the colums are the - ; original columns. - ; The 1d transform is done by looping over bands of eight columns (the - ; idct32_bands loop). For each band, the transform input transposition - ; is done on demand, one band of four 8x8 matrices at a time. The four - ; matrices are transposed by pairs (the idct32_transpose_pair loop). - push {r4-r11} - vpush {d8-d15} - ; stack operation - ; internal buffer used to transpose 8 lines into before transforming them - ; int16_t transpose_buffer[32 * 8]; - ; at sp + [4096, 4607] - ; results of the first pass (transpose and transform rows) - ; int16_t pass1[32 * 32]; - ; at sp + [0, 2047] - ; results of the second pass (transpose and transform columns) - ; int16_t pass2[32 * 32]; - ; at sp + [2048, 4095] - sub sp, sp, #512+2048+2048 - - ; r6 = dest + 31 * dest_stride - ; r7 = dest + 0 * dest_stride - ; r9 = dest + 15 * dest_stride - ; r10 = dest + 16 * dest_stride - rsb r6, r2, r2, lsl #5 - rsb r9, r2, r2, lsl #4 - add r10, r1, r2, lsl #4 - mov r7, r1 - add r6, r6, r1 - add r9, r9, r1 - ; r11 = -dest_stride - neg r11, r2 - ; r3 = input - mov r3, r0 - ; parameters for first pass - ; r0 = transpose_buffer[32 * 8] - add r0, sp, #4096 - ; r1 = pass1[32 * 32] - mov r1, sp - - mov r5, #0 ; initialize pass loop counter -idct32_pass_loop - mov r4, #4 ; initialize bands loop counter -idct32_bands_loop - mov r8, #2 ; initialize transpose loop counter -idct32_transpose_pair_loop - ; Load two horizontally consecutive 8x8 16bit data matrices. The first one - ; into q0-q7 and the second one into q8-q15. There is a stride of 64, - ; adjusted to 32 because of the two post-increments. - vld1.s16 {q8}, [r3]! - vld1.s16 {q0}, [r3]! - add r3, #32 - vld1.s16 {q9}, [r3]! - vld1.s16 {q1}, [r3]! - add r3, #32 - vld1.s16 {q10}, [r3]! - vld1.s16 {q2}, [r3]! - add r3, #32 - vld1.s16 {q11}, [r3]! - vld1.s16 {q3}, [r3]! - add r3, #32 - vld1.s16 {q12}, [r3]! - vld1.s16 {q4}, [r3]! - add r3, #32 - vld1.s16 {q13}, [r3]! - vld1.s16 {q5}, [r3]! - add r3, #32 - vld1.s16 {q14}, [r3]! - vld1.s16 {q6}, [r3]! - add r3, #32 - vld1.s16 {q15}, [r3]! - vld1.s16 {q7}, [r3]! - - ; Transpose the two 8x8 16bit data matrices. - vswp d17, d24 - vswp d23, d30 - vswp d21, d28 - vswp d19, d26 - vswp d1, d8 - vswp d7, d14 - vswp d5, d12 - vswp d3, d10 - vtrn.32 q8, q10 - vtrn.32 q9, q11 - vtrn.32 q12, q14 - vtrn.32 q13, q15 - vtrn.32 q0, q2 - vtrn.32 q1, q3 - vtrn.32 q4, q6 - vtrn.32 q5, q7 - vtrn.16 q8, q9 - vtrn.16 q10, q11 - vtrn.16 q12, q13 - vtrn.16 q14, q15 - vtrn.16 q0, q1 - vtrn.16 q2, q3 - vtrn.16 q4, q5 - vtrn.16 q6, q7 - - ; Store both matrices after each other. There is a stride of 32, which - ; adjusts to nothing because of the post-increments. - vst1.16 {q8}, [r0]! - vst1.16 {q9}, [r0]! - vst1.16 {q10}, [r0]! - vst1.16 {q11}, [r0]! - vst1.16 {q12}, [r0]! - vst1.16 {q13}, [r0]! - vst1.16 {q14}, [r0]! - vst1.16 {q15}, [r0]! - vst1.16 {q0}, [r0]! - vst1.16 {q1}, [r0]! - vst1.16 {q2}, [r0]! - vst1.16 {q3}, [r0]! - vst1.16 {q4}, [r0]! - vst1.16 {q5}, [r0]! - vst1.16 {q6}, [r0]! - vst1.16 {q7}, [r0]! - - ; increment pointers by adjusted stride (not necessary for r0/out) - ; go back by 7*32 for the seven lines moved fully by read and add - ; go back by 32 for the eigth line only read - ; advance by 16*2 to go the next pair - sub r3, r3, #7*32*2 + 32 - 16*2 - ; transpose pair loop processing - subs r8, r8, #1 - bne idct32_transpose_pair_loop - - ; restore r0/input to its original value - sub r0, r0, #32*8*2 - - ; Instead of doing the transforms stage by stage, it is done by loading - ; some input values and doing as many stages as possible to minimize the - ; storing/loading of intermediate results. To fit within registers, the - ; final coefficients are cut into four blocks: - ; BLOCK A: 16-19,28-31 - ; BLOCK B: 20-23,24-27 - ; BLOCK C: 8-10,11-15 - ; BLOCK D: 0-3,4-7 - ; Blocks A and C are straight calculation through the various stages. In - ; block B, further calculations are performed using the results from - ; block A. In block D, further calculations are performed using the results - ; from block C and then the final calculations are done using results from - ; block A and B which have been combined at the end of block B. - - ; -------------------------------------------------------------------------- - ; BLOCK A: 16-19,28-31 - ; -------------------------------------------------------------------------- - ; generate 16,17,30,31 - ; -------------------------------------------------------------------------- - ; part of stage 1 - ;temp1 = input[1 * 32] * cospi_31_64 - input[31 * 32] * cospi_1_64; - ;temp2 = input[1 * 32] * cospi_1_64 + input[31 * 32] * cospi_31_64; - ;step1b[16][i] = dct_const_round_shift(temp1); - ;step1b[31][i] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 0, 1, 31 - DO_BUTTERFLY_STD cospi_31_64, cospi_1_64, d0, d1, d4, d5 - ; -------------------------------------------------------------------------- - ; part of stage 1 - ;temp1 = input[17 * 32] * cospi_15_64 - input[15 * 32] * cospi_17_64; - ;temp2 = input[17 * 32] * cospi_17_64 + input[15 * 32] * cospi_15_64; - ;step1b[17][i] = dct_const_round_shift(temp1); - ;step1b[30][i] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 31, 17, 15 - DO_BUTTERFLY_STD cospi_15_64, cospi_17_64, d2, d3, d6, d7 - ; -------------------------------------------------------------------------- - ; part of stage 2 - ;step2[16] = step1b[16][i] + step1b[17][i]; - ;step2[17] = step1b[16][i] - step1b[17][i]; - ;step2[30] = -step1b[30][i] + step1b[31][i]; - ;step2[31] = step1b[30][i] + step1b[31][i]; - vadd.s16 q4, q0, q1 - vsub.s16 q13, q0, q1 - vadd.s16 q6, q2, q3 - vsub.s16 q14, q2, q3 - ; -------------------------------------------------------------------------- - ; part of stage 3 - ;temp1 = step1b[30][i] * cospi_28_64 - step1b[17][i] * cospi_4_64; - ;temp2 = step1b[30][i] * cospi_4_64 - step1b[17][i] * cospi_28_64; - ;step3[17] = dct_const_round_shift(temp1); - ;step3[30] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_28_64, cospi_4_64, d10, d11, d14, d15 - ; -------------------------------------------------------------------------- - ; generate 18,19,28,29 - ; -------------------------------------------------------------------------- - ; part of stage 1 - ;temp1 = input[9 * 32] * cospi_23_64 - input[23 * 32] * cospi_9_64; - ;temp2 = input[9 * 32] * cospi_9_64 + input[23 * 32] * cospi_23_64; - ;step1b[18][i] = dct_const_round_shift(temp1); - ;step1b[29][i] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 15, 9, 23 - DO_BUTTERFLY_STD cospi_23_64, cospi_9_64, d0, d1, d4, d5 - ; -------------------------------------------------------------------------- - ; part of stage 1 - ;temp1 = input[25 * 32] * cospi_7_64 - input[7 * 32] * cospi_25_64; - ;temp2 = input[25 * 32] * cospi_25_64 + input[7 * 32] * cospi_7_64; - ;step1b[19][i] = dct_const_round_shift(temp1); - ;step1b[28][i] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 23, 25, 7 - DO_BUTTERFLY_STD cospi_7_64, cospi_25_64, d2, d3, d6, d7 - ; -------------------------------------------------------------------------- - ; part of stage 2 - ;step2[18] = -step1b[18][i] + step1b[19][i]; - ;step2[19] = step1b[18][i] + step1b[19][i]; - ;step2[28] = step1b[28][i] + step1b[29][i]; - ;step2[29] = step1b[28][i] - step1b[29][i]; - vsub.s16 q13, q3, q2 - vadd.s16 q3, q3, q2 - vsub.s16 q14, q1, q0 - vadd.s16 q2, q1, q0 - ; -------------------------------------------------------------------------- - ; part of stage 3 - ;temp1 = step1b[18][i] * (-cospi_4_64) - step1b[29][i] * (-cospi_28_64); - ;temp2 = step1b[18][i] * (-cospi_28_64) + step1b[29][i] * (-cospi_4_64); - ;step3[29] = dct_const_round_shift(temp1); - ;step3[18] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD (-cospi_4_64), (-cospi_28_64), d2, d3, d0, d1 - ; -------------------------------------------------------------------------- - ; combine 16-19,28-31 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;step1[16] = step1b[16][i] + step1b[19][i]; - ;step1[17] = step1b[17][i] + step1b[18][i]; - ;step1[18] = step1b[17][i] - step1b[18][i]; - ;step1[29] = step1b[30][i] - step1b[29][i]; - ;step1[30] = step1b[30][i] + step1b[29][i]; - ;step1[31] = step1b[31][i] + step1b[28][i]; - vadd.s16 q8, q4, q2 - vadd.s16 q9, q5, q0 - vadd.s16 q10, q7, q1 - vadd.s16 q15, q6, q3 - vsub.s16 q13, q5, q0 - vsub.s16 q14, q7, q1 - STORE_IN_OUTPUT 0, 16, 31, q8, q15 - STORE_IN_OUTPUT 31, 17, 30, q9, q10 - ; -------------------------------------------------------------------------- - ; part of stage 5 - ;temp1 = step1b[29][i] * cospi_24_64 - step1b[18][i] * cospi_8_64; - ;temp2 = step1b[29][i] * cospi_8_64 + step1b[18][i] * cospi_24_64; - ;step2[18] = dct_const_round_shift(temp1); - ;step2[29] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d0, d1, d2, d3 - STORE_IN_OUTPUT 30, 29, 18, q1, q0 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;step1[19] = step1b[16][i] - step1b[19][i]; - ;step1[28] = step1b[31][i] - step1b[28][i]; - vsub.s16 q13, q4, q2 - vsub.s16 q14, q6, q3 - ; -------------------------------------------------------------------------- - ; part of stage 5 - ;temp1 = step1b[28][i] * cospi_24_64 - step1b[19][i] * cospi_8_64; - ;temp2 = step1b[28][i] * cospi_8_64 + step1b[19][i] * cospi_24_64; - ;step2[19] = dct_const_round_shift(temp1); - ;step2[28] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d8, d9, d12, d13 - STORE_IN_OUTPUT 18, 19, 28, q4, q6 - ; -------------------------------------------------------------------------- - - - ; -------------------------------------------------------------------------- - ; BLOCK B: 20-23,24-27 - ; -------------------------------------------------------------------------- - ; generate 20,21,26,27 - ; -------------------------------------------------------------------------- - ; part of stage 1 - ;temp1 = input[5 * 32] * cospi_27_64 - input[27 * 32] * cospi_5_64; - ;temp2 = input[5 * 32] * cospi_5_64 + input[27 * 32] * cospi_27_64; - ;step1b[20][i] = dct_const_round_shift(temp1); - ;step1b[27][i] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 7, 5, 27 - DO_BUTTERFLY_STD cospi_27_64, cospi_5_64, d0, d1, d4, d5 - ; -------------------------------------------------------------------------- - ; part of stage 1 - ;temp1 = input[21 * 32] * cospi_11_64 - input[11 * 32] * cospi_21_64; - ;temp2 = input[21 * 32] * cospi_21_64 + input[11 * 32] * cospi_11_64; - ;step1b[21][i] = dct_const_round_shift(temp1); - ;step1b[26][i] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 27, 21, 11 - DO_BUTTERFLY_STD cospi_11_64, cospi_21_64, d2, d3, d6, d7 - ; -------------------------------------------------------------------------- - ; part of stage 2 - ;step2[20] = step1b[20][i] + step1b[21][i]; - ;step2[21] = step1b[20][i] - step1b[21][i]; - ;step2[26] = -step1b[26][i] + step1b[27][i]; - ;step2[27] = step1b[26][i] + step1b[27][i]; - vsub.s16 q13, q0, q1 - vadd.s16 q0, q0, q1 - vsub.s16 q14, q2, q3 - vadd.s16 q2, q2, q3 - ; -------------------------------------------------------------------------- - ; part of stage 3 - ;temp1 = step1b[26][i] * cospi_12_64 - step1b[21][i] * cospi_20_64; - ;temp2 = step1b[26][i] * cospi_20_64 + step1b[21][i] * cospi_12_64; - ;step3[21] = dct_const_round_shift(temp1); - ;step3[26] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_12_64, cospi_20_64, d2, d3, d6, d7 - ; -------------------------------------------------------------------------- - ; generate 22,23,24,25 - ; -------------------------------------------------------------------------- - ; part of stage 1 - ;temp1 = input[13 * 32] * cospi_19_64 - input[19 * 32] * cospi_13_64; - ;temp2 = input[13 * 32] * cospi_13_64 + input[19 * 32] * cospi_19_64; - ;step1b[22][i] = dct_const_round_shift(temp1); - ;step1b[25][i] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 11, 13, 19 - DO_BUTTERFLY_STD cospi_19_64, cospi_13_64, d10, d11, d14, d15 - ; -------------------------------------------------------------------------- - ; part of stage 1 - ;temp1 = input[29 * 32] * cospi_3_64 - input[3 * 32] * cospi_29_64; - ;temp2 = input[29 * 32] * cospi_29_64 + input[3 * 32] * cospi_3_64; - ;step1b[23][i] = dct_const_round_shift(temp1); - ;step1b[24][i] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 19, 29, 3 - DO_BUTTERFLY_STD cospi_3_64, cospi_29_64, d8, d9, d12, d13 - ; -------------------------------------------------------------------------- - ; part of stage 2 - ;step2[22] = -step1b[22][i] + step1b[23][i]; - ;step2[23] = step1b[22][i] + step1b[23][i]; - ;step2[24] = step1b[24][i] + step1b[25][i]; - ;step2[25] = step1b[24][i] - step1b[25][i]; - vsub.s16 q14, q4, q5 - vadd.s16 q5, q4, q5 - vsub.s16 q13, q6, q7 - vadd.s16 q6, q6, q7 - ; -------------------------------------------------------------------------- - ; part of stage 3 - ;temp1 = step1b[22][i] * (-cospi_20_64) - step1b[25][i] * (-cospi_12_64); - ;temp2 = step1b[22][i] * (-cospi_12_64) + step1b[25][i] * (-cospi_20_64); - ;step3[25] = dct_const_round_shift(temp1); - ;step3[22] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD (-cospi_20_64), (-cospi_12_64), d8, d9, d14, d15 - ; -------------------------------------------------------------------------- - ; combine 20-23,24-27 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;step1[22] = step1b[22][i] + step1b[21][i]; - ;step1[23] = step1b[23][i] + step1b[20][i]; - vadd.s16 q10, q7, q1 - vadd.s16 q11, q5, q0 - ;step1[24] = step1b[24][i] + step1b[27][i]; - ;step1[25] = step1b[25][i] + step1b[26][i]; - vadd.s16 q12, q6, q2 - vadd.s16 q15, q4, q3 - ; -------------------------------------------------------------------------- - ; part of stage 6 - ;step3[16] = step1b[16][i] + step1b[23][i]; - ;step3[17] = step1b[17][i] + step1b[22][i]; - ;step3[22] = step1b[17][i] - step1b[22][i]; - ;step3[23] = step1b[16][i] - step1b[23][i]; - LOAD_FROM_OUTPUT 28, 16, 17, q14, q13 - vadd.s16 q8, q14, q11 - vadd.s16 q9, q13, q10 - vsub.s16 q13, q13, q10 - vsub.s16 q11, q14, q11 - STORE_IN_OUTPUT 17, 17, 16, q9, q8 - ; -------------------------------------------------------------------------- - ; part of stage 6 - ;step3[24] = step1b[31][i] - step1b[24][i]; - ;step3[25] = step1b[30][i] - step1b[25][i]; - ;step3[30] = step1b[30][i] + step1b[25][i]; - ;step3[31] = step1b[31][i] + step1b[24][i]; - LOAD_FROM_OUTPUT 16, 30, 31, q14, q9 - vsub.s16 q8, q9, q12 - vadd.s16 q10, q14, q15 - vsub.s16 q14, q14, q15 - vadd.s16 q12, q9, q12 - STORE_IN_OUTPUT 31, 30, 31, q10, q12 - ; -------------------------------------------------------------------------- - ; TODO(cd) do some register allocation change to remove these push/pop - vpush {q8} ; [24] - vpush {q11} ; [23] - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;temp1 = (step1b[25][i] - step1b[22][i]) * cospi_16_64; - ;temp2 = (step1b[25][i] + step1b[22][i]) * cospi_16_64; - ;step1[22] = dct_const_round_shift(temp1); - ;step1[25] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29 - STORE_IN_OUTPUT 31, 25, 22, q14, q13 - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;temp1 = (step1b[24][i] - step1b[23][i]) * cospi_16_64; - ;temp2 = (step1b[24][i] + step1b[23][i]) * cospi_16_64; - ;step1[23] = dct_const_round_shift(temp1); - ;step1[24] = dct_const_round_shift(temp2); - ; TODO(cd) do some register allocation change to remove these push/pop - vpop {q13} ; [23] - vpop {q14} ; [24] - DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29 - STORE_IN_OUTPUT 22, 24, 23, q14, q13 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;step1[20] = step1b[23][i] - step1b[20][i]; - ;step1[27] = step1b[24][i] - step1b[27][i]; - vsub.s16 q14, q5, q0 - vsub.s16 q13, q6, q2 - ; -------------------------------------------------------------------------- - ; part of stage 5 - ;temp1 = step1b[20][i] * (-cospi_8_64) - step1b[27][i] * (-cospi_24_64); - ;temp2 = step1b[20][i] * (-cospi_24_64) + step1b[27][i] * (-cospi_8_64); - ;step2[27] = dct_const_round_shift(temp1); - ;step2[20] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d10, d11, d12, d13 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;step1[21] = step1b[22][i] - step1b[21][i]; - ;step1[26] = step1b[25][i] - step1b[26][i]; - vsub.s16 q14, q7, q1 - vsub.s16 q13, q4, q3 - ; -------------------------------------------------------------------------- - ; part of stage 5 - ;temp1 = step1b[21][i] * (-cospi_8_64) - step1b[26][i] * (-cospi_24_64); - ;temp2 = step1b[21][i] * (-cospi_24_64) + step1b[26][i] * (-cospi_8_64); - ;step2[26] = dct_const_round_shift(temp1); - ;step2[21] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d0, d1, d2, d3 - ; -------------------------------------------------------------------------- - ; part of stage 6 - ;step3[18] = step1b[18][i] + step1b[21][i]; - ;step3[19] = step1b[19][i] + step1b[20][i]; - ;step3[20] = step1b[19][i] - step1b[20][i]; - ;step3[21] = step1b[18][i] - step1b[21][i]; - LOAD_FROM_OUTPUT 23, 18, 19, q14, q13 - vadd.s16 q8, q14, q1 - vadd.s16 q9, q13, q6 - vsub.s16 q13, q13, q6 - vsub.s16 q1, q14, q1 - STORE_IN_OUTPUT 19, 18, 19, q8, q9 - ; -------------------------------------------------------------------------- - ; part of stage 6 - ;step3[27] = step1b[28][i] - step1b[27][i]; - ;step3[28] = step1b[28][i] + step1b[27][i]; - ;step3[29] = step1b[29][i] + step1b[26][i]; - ;step3[26] = step1b[29][i] - step1b[26][i]; - LOAD_FROM_OUTPUT 19, 28, 29, q8, q9 - vsub.s16 q14, q8, q5 - vadd.s16 q10, q8, q5 - vadd.s16 q11, q9, q0 - vsub.s16 q0, q9, q0 - STORE_IN_OUTPUT 29, 28, 29, q10, q11 - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;temp1 = (step1b[27][i] - step1b[20][i]) * cospi_16_64; - ;temp2 = (step1b[27][i] + step1b[20][i]) * cospi_16_64; - ;step1[20] = dct_const_round_shift(temp1); - ;step1[27] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29 - STORE_IN_OUTPUT 29, 20, 27, q13, q14 - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;temp1 = (step1b[26][i] - step1b[21][i]) * cospi_16_64; - ;temp2 = (step1b[26][i] + step1b[21][i]) * cospi_16_64; - ;step1[21] = dct_const_round_shift(temp1); - ;step1[26] = dct_const_round_shift(temp2); - DO_BUTTERFLY d0, d1, d2, d3, cospi_16_64, cospi_16_64, d2, d3, d0, d1 - STORE_IN_OUTPUT 27, 21, 26, q1, q0 - ; -------------------------------------------------------------------------- - - - ; -------------------------------------------------------------------------- - ; BLOCK C: 8-10,11-15 - ; -------------------------------------------------------------------------- - ; generate 8,9,14,15 - ; -------------------------------------------------------------------------- - ; part of stage 2 - ;temp1 = input[2 * 32] * cospi_30_64 - input[30 * 32] * cospi_2_64; - ;temp2 = input[2 * 32] * cospi_2_64 + input[30 * 32] * cospi_30_64; - ;step2[8] = dct_const_round_shift(temp1); - ;step2[15] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 3, 2, 30 - DO_BUTTERFLY_STD cospi_30_64, cospi_2_64, d0, d1, d4, d5 - ; -------------------------------------------------------------------------- - ; part of stage 2 - ;temp1 = input[18 * 32] * cospi_14_64 - input[14 * 32] * cospi_18_64; - ;temp2 = input[18 * 32] * cospi_18_64 + input[14 * 32] * cospi_14_64; - ;step2[9] = dct_const_round_shift(temp1); - ;step2[14] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 30, 18, 14 - DO_BUTTERFLY_STD cospi_14_64, cospi_18_64, d2, d3, d6, d7 - ; -------------------------------------------------------------------------- - ; part of stage 3 - ;step3[8] = step1b[8][i] + step1b[9][i]; - ;step3[9] = step1b[8][i] - step1b[9][i]; - ;step3[14] = step1b[15][i] - step1b[14][i]; - ;step3[15] = step1b[15][i] + step1b[14][i]; - vsub.s16 q13, q0, q1 - vadd.s16 q0, q0, q1 - vsub.s16 q14, q2, q3 - vadd.s16 q2, q2, q3 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;temp1 = step1b[14][i] * cospi_24_64 - step1b[9][i] * cospi_8_64; - ;temp2 = step1b[14][i] * cospi_8_64 + step1b[9][i] * cospi_24_64; - ;step1[9] = dct_const_round_shift(temp1); - ;step1[14] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d2, d3, d6, d7 - ; -------------------------------------------------------------------------- - ; generate 10,11,12,13 - ; -------------------------------------------------------------------------- - ; part of stage 2 - ;temp1 = input[10 * 32] * cospi_22_64 - input[22 * 32] * cospi_10_64; - ;temp2 = input[10 * 32] * cospi_10_64 + input[22 * 32] * cospi_22_64; - ;step2[10] = dct_const_round_shift(temp1); - ;step2[13] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 14, 10, 22 - DO_BUTTERFLY_STD cospi_22_64, cospi_10_64, d10, d11, d14, d15 - ; -------------------------------------------------------------------------- - ; part of stage 2 - ;temp1 = input[26 * 32] * cospi_6_64 - input[6 * 32] * cospi_26_64; - ;temp2 = input[26 * 32] * cospi_26_64 + input[6 * 32] * cospi_6_64; - ;step2[11] = dct_const_round_shift(temp1); - ;step2[12] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 22, 26, 6 - DO_BUTTERFLY_STD cospi_6_64, cospi_26_64, d8, d9, d12, d13 - ; -------------------------------------------------------------------------- - ; part of stage 3 - ;step3[10] = step1b[11][i] - step1b[10][i]; - ;step3[11] = step1b[11][i] + step1b[10][i]; - ;step3[12] = step1b[12][i] + step1b[13][i]; - ;step3[13] = step1b[12][i] - step1b[13][i]; - vsub.s16 q14, q4, q5 - vadd.s16 q5, q4, q5 - vsub.s16 q13, q6, q7 - vadd.s16 q6, q6, q7 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;temp1 = step1b[10][i] * (-cospi_8_64) - step1b[13][i] * (-cospi_24_64); - ;temp2 = step1b[10][i] * (-cospi_24_64) + step1b[13][i] * (-cospi_8_64); - ;step1[13] = dct_const_round_shift(temp1); - ;step1[10] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d8, d9, d14, d15 - ; -------------------------------------------------------------------------- - ; combine 8-10,11-15 - ; -------------------------------------------------------------------------- - ; part of stage 5 - ;step2[8] = step1b[8][i] + step1b[11][i]; - ;step2[9] = step1b[9][i] + step1b[10][i]; - ;step2[10] = step1b[9][i] - step1b[10][i]; - vadd.s16 q8, q0, q5 - vadd.s16 q9, q1, q7 - vsub.s16 q13, q1, q7 - ;step2[13] = step1b[14][i] - step1b[13][i]; - ;step2[14] = step1b[14][i] + step1b[13][i]; - ;step2[15] = step1b[15][i] + step1b[12][i]; - vsub.s16 q14, q3, q4 - vadd.s16 q10, q3, q4 - vadd.s16 q15, q2, q6 - STORE_IN_OUTPUT 26, 8, 15, q8, q15 - STORE_IN_OUTPUT 15, 9, 14, q9, q10 - ; -------------------------------------------------------------------------- - ; part of stage 6 - ;temp1 = (step1b[13][i] - step1b[10][i]) * cospi_16_64; - ;temp2 = (step1b[13][i] + step1b[10][i]) * cospi_16_64; - ;step3[10] = dct_const_round_shift(temp1); - ;step3[13] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7 - STORE_IN_OUTPUT 14, 13, 10, q3, q1 - ; -------------------------------------------------------------------------- - ; part of stage 5 - ;step2[11] = step1b[8][i] - step1b[11][i]; - ;step2[12] = step1b[15][i] - step1b[12][i]; - vsub.s16 q13, q0, q5 - vsub.s16 q14, q2, q6 - ; -------------------------------------------------------------------------- - ; part of stage 6 - ;temp1 = (step1b[12][i] - step1b[11][i]) * cospi_16_64; - ;temp2 = (step1b[12][i] + step1b[11][i]) * cospi_16_64; - ;step3[11] = dct_const_round_shift(temp1); - ;step3[12] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7 - STORE_IN_OUTPUT 10, 11, 12, q1, q3 - ; -------------------------------------------------------------------------- - - - ; -------------------------------------------------------------------------- - ; BLOCK D: 0-3,4-7 - ; -------------------------------------------------------------------------- - ; generate 4,5,6,7 - ; -------------------------------------------------------------------------- - ; part of stage 3 - ;temp1 = input[4 * 32] * cospi_28_64 - input[28 * 32] * cospi_4_64; - ;temp2 = input[4 * 32] * cospi_4_64 + input[28 * 32] * cospi_28_64; - ;step3[4] = dct_const_round_shift(temp1); - ;step3[7] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 6, 4, 28 - DO_BUTTERFLY_STD cospi_28_64, cospi_4_64, d0, d1, d4, d5 - ; -------------------------------------------------------------------------- - ; part of stage 3 - ;temp1 = input[20 * 32] * cospi_12_64 - input[12 * 32] * cospi_20_64; - ;temp2 = input[20 * 32] * cospi_20_64 + input[12 * 32] * cospi_12_64; - ;step3[5] = dct_const_round_shift(temp1); - ;step3[6] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 28, 20, 12 - DO_BUTTERFLY_STD cospi_12_64, cospi_20_64, d2, d3, d6, d7 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;step1[4] = step1b[4][i] + step1b[5][i]; - ;step1[5] = step1b[4][i] - step1b[5][i]; - ;step1[6] = step1b[7][i] - step1b[6][i]; - ;step1[7] = step1b[7][i] + step1b[6][i]; - vsub.s16 q13, q0, q1 - vadd.s16 q0, q0, q1 - vsub.s16 q14, q2, q3 - vadd.s16 q2, q2, q3 - ; -------------------------------------------------------------------------- - ; part of stage 5 - ;temp1 = (step1b[6][i] - step1b[5][i]) * cospi_16_64; - ;temp2 = (step1b[5][i] + step1b[6][i]) * cospi_16_64; - ;step2[5] = dct_const_round_shift(temp1); - ;step2[6] = dct_const_round_shift(temp2); - DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7 - ; -------------------------------------------------------------------------- - ; generate 0,1,2,3 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;temp1 = (input[0 * 32] - input[16 * 32]) * cospi_16_64; - ;temp2 = (input[0 * 32] + input[16 * 32]) * cospi_16_64; - ;step1[1] = dct_const_round_shift(temp1); - ;step1[0] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 12, 0, 16 - DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d10, d11, d14, d15 - ; -------------------------------------------------------------------------- - ; part of stage 4 - ;temp1 = input[8 * 32] * cospi_24_64 - input[24 * 32] * cospi_8_64; - ;temp2 = input[8 * 32] * cospi_8_64 + input[24 * 32] * cospi_24_64; - ;step1[2] = dct_const_round_shift(temp1); - ;step1[3] = dct_const_round_shift(temp2); - LOAD_FROM_TRANSPOSED 16, 8, 24 - DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d28, d29, d12, d13 - ; -------------------------------------------------------------------------- - ; part of stage 5 - ;step2[0] = step1b[0][i] + step1b[3][i]; - ;step2[1] = step1b[1][i] + step1b[2][i]; - ;step2[2] = step1b[1][i] - step1b[2][i]; - ;step2[3] = step1b[0][i] - step1b[3][i]; - vadd.s16 q4, q7, q6 - vsub.s16 q7, q7, q6 - vsub.s16 q6, q5, q14 - vadd.s16 q5, q5, q14 - ; -------------------------------------------------------------------------- - ; combine 0-3,4-7 - ; -------------------------------------------------------------------------- - ; part of stage 6 - ;step3[0] = step1b[0][i] + step1b[7][i]; - ;step3[1] = step1b[1][i] + step1b[6][i]; - ;step3[2] = step1b[2][i] + step1b[5][i]; - ;step3[3] = step1b[3][i] + step1b[4][i]; - vadd.s16 q8, q4, q2 - vadd.s16 q9, q5, q3 - vadd.s16 q10, q6, q1 - vadd.s16 q11, q7, q0 - ;step3[4] = step1b[3][i] - step1b[4][i]; - ;step3[5] = step1b[2][i] - step1b[5][i]; - ;step3[6] = step1b[1][i] - step1b[6][i]; - ;step3[7] = step1b[0][i] - step1b[7][i]; - vsub.s16 q12, q7, q0 - vsub.s16 q13, q6, q1 - vsub.s16 q14, q5, q3 - vsub.s16 q15, q4, q2 - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;step1[0] = step1b[0][i] + step1b[15][i]; - ;step1[1] = step1b[1][i] + step1b[14][i]; - ;step1[14] = step1b[1][i] - step1b[14][i]; - ;step1[15] = step1b[0][i] - step1b[15][i]; - LOAD_FROM_OUTPUT 12, 14, 15, q0, q1 - vadd.s16 q2, q8, q1 - vadd.s16 q3, q9, q0 - vsub.s16 q4, q9, q0 - vsub.s16 q5, q8, q1 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[14 * 32] = step1b[14][i] + step1b[17][i]; - ;output[15 * 32] = step1b[15][i] + step1b[16][i]; - ;output[16 * 32] = step1b[15][i] - step1b[16][i]; - ;output[17 * 32] = step1b[14][i] - step1b[17][i]; - LOAD_FROM_OUTPUT 15, 16, 17, q0, q1 - vadd.s16 q8, q4, q1 - vadd.s16 q9, q5, q0 - vsub.s16 q6, q5, q0 - vsub.s16 q7, q4, q1 - - cmp r5, #0 - bgt idct32_bands_end_2nd_pass - -idct32_bands_end_1st_pass - STORE_IN_OUTPUT 17, 16, 17, q6, q7 - STORE_IN_OUTPUT 17, 14, 15, q8, q9 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 0 * 32] = step1b[0][i] + step1b[31][i]; - ;output[ 1 * 32] = step1b[1][i] + step1b[30][i]; - ;output[30 * 32] = step1b[1][i] - step1b[30][i]; - ;output[31 * 32] = step1b[0][i] - step1b[31][i]; - LOAD_FROM_OUTPUT 15, 30, 31, q0, q1 - vadd.s16 q4, q2, q1 - vadd.s16 q5, q3, q0 - vsub.s16 q6, q3, q0 - vsub.s16 q7, q2, q1 - STORE_IN_OUTPUT 31, 30, 31, q6, q7 - STORE_IN_OUTPUT 31, 0, 1, q4, q5 - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;step1[2] = step1b[2][i] + step1b[13][i]; - ;step1[3] = step1b[3][i] + step1b[12][i]; - ;step1[12] = step1b[3][i] - step1b[12][i]; - ;step1[13] = step1b[2][i] - step1b[13][i]; - LOAD_FROM_OUTPUT 1, 12, 13, q0, q1 - vadd.s16 q2, q10, q1 - vadd.s16 q3, q11, q0 - vsub.s16 q4, q11, q0 - vsub.s16 q5, q10, q1 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[12 * 32] = step1b[12][i] + step1b[19][i]; - ;output[13 * 32] = step1b[13][i] + step1b[18][i]; - ;output[18 * 32] = step1b[13][i] - step1b[18][i]; - ;output[19 * 32] = step1b[12][i] - step1b[19][i]; - LOAD_FROM_OUTPUT 13, 18, 19, q0, q1 - vadd.s16 q8, q4, q1 - vadd.s16 q9, q5, q0 - vsub.s16 q6, q5, q0 - vsub.s16 q7, q4, q1 - STORE_IN_OUTPUT 19, 18, 19, q6, q7 - STORE_IN_OUTPUT 19, 12, 13, q8, q9 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 2 * 32] = step1b[2][i] + step1b[29][i]; - ;output[ 3 * 32] = step1b[3][i] + step1b[28][i]; - ;output[28 * 32] = step1b[3][i] - step1b[28][i]; - ;output[29 * 32] = step1b[2][i] - step1b[29][i]; - LOAD_FROM_OUTPUT 13, 28, 29, q0, q1 - vadd.s16 q4, q2, q1 - vadd.s16 q5, q3, q0 - vsub.s16 q6, q3, q0 - vsub.s16 q7, q2, q1 - STORE_IN_OUTPUT 29, 28, 29, q6, q7 - STORE_IN_OUTPUT 29, 2, 3, q4, q5 - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;step1[4] = step1b[4][i] + step1b[11][i]; - ;step1[5] = step1b[5][i] + step1b[10][i]; - ;step1[10] = step1b[5][i] - step1b[10][i]; - ;step1[11] = step1b[4][i] - step1b[11][i]; - LOAD_FROM_OUTPUT 3, 10, 11, q0, q1 - vadd.s16 q2, q12, q1 - vadd.s16 q3, q13, q0 - vsub.s16 q4, q13, q0 - vsub.s16 q5, q12, q1 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[10 * 32] = step1b[10][i] + step1b[21][i]; - ;output[11 * 32] = step1b[11][i] + step1b[20][i]; - ;output[20 * 32] = step1b[11][i] - step1b[20][i]; - ;output[21 * 32] = step1b[10][i] - step1b[21][i]; - LOAD_FROM_OUTPUT 11, 20, 21, q0, q1 - vadd.s16 q8, q4, q1 - vadd.s16 q9, q5, q0 - vsub.s16 q6, q5, q0 - vsub.s16 q7, q4, q1 - STORE_IN_OUTPUT 21, 20, 21, q6, q7 - STORE_IN_OUTPUT 21, 10, 11, q8, q9 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 4 * 32] = step1b[4][i] + step1b[27][i]; - ;output[ 5 * 32] = step1b[5][i] + step1b[26][i]; - ;output[26 * 32] = step1b[5][i] - step1b[26][i]; - ;output[27 * 32] = step1b[4][i] - step1b[27][i]; - LOAD_FROM_OUTPUT 11, 26, 27, q0, q1 - vadd.s16 q4, q2, q1 - vadd.s16 q5, q3, q0 - vsub.s16 q6, q3, q0 - vsub.s16 q7, q2, q1 - STORE_IN_OUTPUT 27, 26, 27, q6, q7 - STORE_IN_OUTPUT 27, 4, 5, q4, q5 - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;step1[6] = step1b[6][i] + step1b[9][i]; - ;step1[7] = step1b[7][i] + step1b[8][i]; - ;step1[8] = step1b[7][i] - step1b[8][i]; - ;step1[9] = step1b[6][i] - step1b[9][i]; - LOAD_FROM_OUTPUT 5, 8, 9, q0, q1 - vadd.s16 q2, q14, q1 - vadd.s16 q3, q15, q0 - vsub.s16 q4, q15, q0 - vsub.s16 q5, q14, q1 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 8 * 32] = step1b[8][i] + step1b[23][i]; - ;output[ 9 * 32] = step1b[9][i] + step1b[22][i]; - ;output[22 * 32] = step1b[9][i] - step1b[22][i]; - ;output[23 * 32] = step1b[8][i] - step1b[23][i]; - LOAD_FROM_OUTPUT 9, 22, 23, q0, q1 - vadd.s16 q8, q4, q1 - vadd.s16 q9, q5, q0 - vsub.s16 q6, q5, q0 - vsub.s16 q7, q4, q1 - STORE_IN_OUTPUT 23, 22, 23, q6, q7 - STORE_IN_OUTPUT 23, 8, 9, q8, q9 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 6 * 32] = step1b[6][i] + step1b[25][i]; - ;output[ 7 * 32] = step1b[7][i] + step1b[24][i]; - ;output[24 * 32] = step1b[7][i] - step1b[24][i]; - ;output[25 * 32] = step1b[6][i] - step1b[25][i]; - LOAD_FROM_OUTPUT 9, 24, 25, q0, q1 - vadd.s16 q4, q2, q1 - vadd.s16 q5, q3, q0 - vsub.s16 q6, q3, q0 - vsub.s16 q7, q2, q1 - STORE_IN_OUTPUT 25, 24, 25, q6, q7 - STORE_IN_OUTPUT 25, 6, 7, q4, q5 - - ; restore r0 by removing the last offset from the last - ; operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2 - sub r0, r0, #24*8*2 - ; restore r1 by removing the last offset from the last - ; operation (STORE_IN_OUTPUT 24, 6, 7) => 7*32*2 - ; advance by 8 columns => 8*2 - sub r1, r1, #7*32*2 - 8*2 - ; advance by 8 lines (8*32*2) - ; go back by the two pairs from the loop (32*2) - add r3, r3, #8*32*2 - 32*2 - - ; bands loop processing - subs r4, r4, #1 - bne idct32_bands_loop - - ; parameters for second pass - ; the input of pass2 is the result of pass1. we have to remove the offset - ; of 32 columns induced by the above idct32_bands_loop - sub r3, r1, #32*2 - ; r1 = pass2[32 * 32] - add r1, sp, #2048 - - ; pass loop processing - add r5, r5, #1 - b idct32_pass_loop - -idct32_bands_end_2nd_pass - STORE_COMBINE_CENTER_RESULTS - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 0 * 32] = step1b[0][i] + step1b[31][i]; - ;output[ 1 * 32] = step1b[1][i] + step1b[30][i]; - ;output[30 * 32] = step1b[1][i] - step1b[30][i]; - ;output[31 * 32] = step1b[0][i] - step1b[31][i]; - LOAD_FROM_OUTPUT 17, 30, 31, q0, q1 - vadd.s16 q4, q2, q1 - vadd.s16 q5, q3, q0 - vsub.s16 q6, q3, q0 - vsub.s16 q7, q2, q1 - STORE_COMBINE_EXTREME_RESULTS - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;step1[2] = step1b[2][i] + step1b[13][i]; - ;step1[3] = step1b[3][i] + step1b[12][i]; - ;step1[12] = step1b[3][i] - step1b[12][i]; - ;step1[13] = step1b[2][i] - step1b[13][i]; - LOAD_FROM_OUTPUT 31, 12, 13, q0, q1 - vadd.s16 q2, q10, q1 - vadd.s16 q3, q11, q0 - vsub.s16 q4, q11, q0 - vsub.s16 q5, q10, q1 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[12 * 32] = step1b[12][i] + step1b[19][i]; - ;output[13 * 32] = step1b[13][i] + step1b[18][i]; - ;output[18 * 32] = step1b[13][i] - step1b[18][i]; - ;output[19 * 32] = step1b[12][i] - step1b[19][i]; - LOAD_FROM_OUTPUT 13, 18, 19, q0, q1 - vadd.s16 q8, q4, q1 - vadd.s16 q9, q5, q0 - vsub.s16 q6, q5, q0 - vsub.s16 q7, q4, q1 - STORE_COMBINE_CENTER_RESULTS - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 2 * 32] = step1b[2][i] + step1b[29][i]; - ;output[ 3 * 32] = step1b[3][i] + step1b[28][i]; - ;output[28 * 32] = step1b[3][i] - step1b[28][i]; - ;output[29 * 32] = step1b[2][i] - step1b[29][i]; - LOAD_FROM_OUTPUT 19, 28, 29, q0, q1 - vadd.s16 q4, q2, q1 - vadd.s16 q5, q3, q0 - vsub.s16 q6, q3, q0 - vsub.s16 q7, q2, q1 - STORE_COMBINE_EXTREME_RESULTS - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;step1[4] = step1b[4][i] + step1b[11][i]; - ;step1[5] = step1b[5][i] + step1b[10][i]; - ;step1[10] = step1b[5][i] - step1b[10][i]; - ;step1[11] = step1b[4][i] - step1b[11][i]; - LOAD_FROM_OUTPUT 29, 10, 11, q0, q1 - vadd.s16 q2, q12, q1 - vadd.s16 q3, q13, q0 - vsub.s16 q4, q13, q0 - vsub.s16 q5, q12, q1 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[10 * 32] = step1b[10][i] + step1b[21][i]; - ;output[11 * 32] = step1b[11][i] + step1b[20][i]; - ;output[20 * 32] = step1b[11][i] - step1b[20][i]; - ;output[21 * 32] = step1b[10][i] - step1b[21][i]; - LOAD_FROM_OUTPUT 11, 20, 21, q0, q1 - vadd.s16 q8, q4, q1 - vadd.s16 q9, q5, q0 - vsub.s16 q6, q5, q0 - vsub.s16 q7, q4, q1 - STORE_COMBINE_CENTER_RESULTS - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 4 * 32] = step1b[4][i] + step1b[27][i]; - ;output[ 5 * 32] = step1b[5][i] + step1b[26][i]; - ;output[26 * 32] = step1b[5][i] - step1b[26][i]; - ;output[27 * 32] = step1b[4][i] - step1b[27][i]; - LOAD_FROM_OUTPUT 21, 26, 27, q0, q1 - vadd.s16 q4, q2, q1 - vadd.s16 q5, q3, q0 - vsub.s16 q6, q3, q0 - vsub.s16 q7, q2, q1 - STORE_COMBINE_EXTREME_RESULTS - ; -------------------------------------------------------------------------- - ; part of stage 7 - ;step1[6] = step1b[6][i] + step1b[9][i]; - ;step1[7] = step1b[7][i] + step1b[8][i]; - ;step1[8] = step1b[7][i] - step1b[8][i]; - ;step1[9] = step1b[6][i] - step1b[9][i]; - LOAD_FROM_OUTPUT 27, 8, 9, q0, q1 - vadd.s16 q2, q14, q1 - vadd.s16 q3, q15, q0 - vsub.s16 q4, q15, q0 - vsub.s16 q5, q14, q1 - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 8 * 32] = step1b[8][i] + step1b[23][i]; - ;output[ 9 * 32] = step1b[9][i] + step1b[22][i]; - ;output[22 * 32] = step1b[9][i] - step1b[22][i]; - ;output[23 * 32] = step1b[8][i] - step1b[23][i]; - LOAD_FROM_OUTPUT 9, 22, 23, q0, q1 - vadd.s16 q8, q4, q1 - vadd.s16 q9, q5, q0 - vsub.s16 q6, q5, q0 - vsub.s16 q7, q4, q1 - STORE_COMBINE_CENTER_RESULTS_LAST - ; -------------------------------------------------------------------------- - ; part of final stage - ;output[ 6 * 32] = step1b[6][i] + step1b[25][i]; - ;output[ 7 * 32] = step1b[7][i] + step1b[24][i]; - ;output[24 * 32] = step1b[7][i] - step1b[24][i]; - ;output[25 * 32] = step1b[6][i] - step1b[25][i]; - LOAD_FROM_OUTPUT 23, 24, 25, q0, q1 - vadd.s16 q4, q2, q1 - vadd.s16 q5, q3, q0 - vsub.s16 q6, q3, q0 - vsub.s16 q7, q2, q1 - STORE_COMBINE_EXTREME_RESULTS_LAST - ; -------------------------------------------------------------------------- - ; restore pointers to their initial indices for next band pass by - ; removing/adding dest_stride * 8. The actual increment by eight - ; is taken care of within the _LAST macros. - add r6, r6, r2, lsl #3 - add r9, r9, r2, lsl #3 - sub r7, r7, r2, lsl #3 - sub r10, r10, r2, lsl #3 - - ; restore r0 by removing the last offset from the last - ; operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2 - sub r0, r0, #24*8*2 - ; restore r1 by removing the last offset from the last - ; operation (LOAD_FROM_OUTPUT 23, 24, 25) => 25*32*2 - ; advance by 8 columns => 8*2 - sub r1, r1, #25*32*2 - 8*2 - ; advance by 8 lines (8*32*2) - ; go back by the two pairs from the loop (32*2) - add r3, r3, #8*32*2 - 32*2 - - ; bands loop processing - subs r4, r4, #1 - bne idct32_bands_loop - - ; stack operation - add sp, sp, #512+2048+2048 - vpop {d8-d15} - pop {r4-r11} - bx lr - ENDP ; |aom_idct32x32_1024_add_neon| - END diff --git a/aom_dsp/arm/idct32x32_add_neon.c b/aom_dsp/arm/idct32x32_add_neon.c deleted file mode 100644 index a7562c7d5..000000000 --- a/aom_dsp/arm/idct32x32_add_neon.c +++ /dev/null @@ -1,686 +0,0 @@ -/* - * Copyright (c) 2016, 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 "./aom_config.h" -#include "aom_dsp/txfm_common.h" - -#define LOAD_FROM_TRANSPOSED(prev, first, second) \ - q14s16 = vld1q_s16(trans_buf + first * 8); \ - q13s16 = vld1q_s16(trans_buf + second * 8); - -#define LOAD_FROM_OUTPUT(prev, first, second, qA, qB) \ - qA = vld1q_s16(out + first * 32); \ - qB = vld1q_s16(out + second * 32); - -#define STORE_IN_OUTPUT(prev, first, second, qA, qB) \ - vst1q_s16(out + first * 32, qA); \ - vst1q_s16(out + second * 32, qB); - -#define STORE_COMBINE_CENTER_RESULTS(r10, r9) \ - __STORE_COMBINE_CENTER_RESULTS(r10, r9, stride, q6s16, q7s16, q8s16, q9s16); -static INLINE void __STORE_COMBINE_CENTER_RESULTS(uint8_t *p1, uint8_t *p2, - int stride, int16x8_t q6s16, - int16x8_t q7s16, - int16x8_t q8s16, - int16x8_t q9s16) { - int16x4_t d8s16, d9s16, d10s16, d11s16; - - d8s16 = vld1_s16((int16_t *)p1); - p1 += stride; - d11s16 = vld1_s16((int16_t *)p2); - p2 -= stride; - d9s16 = vld1_s16((int16_t *)p1); - d10s16 = vld1_s16((int16_t *)p2); - - q7s16 = vrshrq_n_s16(q7s16, 6); - q8s16 = vrshrq_n_s16(q8s16, 6); - q9s16 = vrshrq_n_s16(q9s16, 6); - q6s16 = vrshrq_n_s16(q6s16, 6); - - q7s16 = vreinterpretq_s16_u16( - vaddw_u8(vreinterpretq_u16_s16(q7s16), vreinterpret_u8_s16(d9s16))); - q8s16 = vreinterpretq_s16_u16( - vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_s16(d10s16))); - q9s16 = vreinterpretq_s16_u16( - vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_s16(d11s16))); - q6s16 = vreinterpretq_s16_u16( - vaddw_u8(vreinterpretq_u16_s16(q6s16), vreinterpret_u8_s16(d8s16))); - - d9s16 = vreinterpret_s16_u8(vqmovun_s16(q7s16)); - d10s16 = vreinterpret_s16_u8(vqmovun_s16(q8s16)); - d11s16 = vreinterpret_s16_u8(vqmovun_s16(q9s16)); - d8s16 = vreinterpret_s16_u8(vqmovun_s16(q6s16)); - - vst1_s16((int16_t *)p1, d9s16); - p1 -= stride; - vst1_s16((int16_t *)p2, d10s16); - p2 += stride; - vst1_s16((int16_t *)p1, d8s16); - vst1_s16((int16_t *)p2, d11s16); - return; -} - -#define STORE_COMBINE_EXTREME_RESULTS(r7, r6) \ - ; \ - __STORE_COMBINE_EXTREME_RESULTS(r7, r6, stride, q4s16, q5s16, q6s16, q7s16); -static INLINE void __STORE_COMBINE_EXTREME_RESULTS(uint8_t *p1, uint8_t *p2, - int stride, int16x8_t q4s16, - int16x8_t q5s16, - int16x8_t q6s16, - int16x8_t q7s16) { - int16x4_t d4s16, d5s16, d6s16, d7s16; - - d4s16 = vld1_s16((int16_t *)p1); - p1 += stride; - d7s16 = vld1_s16((int16_t *)p2); - p2 -= stride; - d5s16 = vld1_s16((int16_t *)p1); - d6s16 = vld1_s16((int16_t *)p2); - - q5s16 = vrshrq_n_s16(q5s16, 6); - q6s16 = vrshrq_n_s16(q6s16, 6); - q7s16 = vrshrq_n_s16(q7s16, 6); - q4s16 = vrshrq_n_s16(q4s16, 6); - - q5s16 = vreinterpretq_s16_u16( - vaddw_u8(vreinterpretq_u16_s16(q5s16), vreinterpret_u8_s16(d5s16))); - q6s16 = vreinterpretq_s16_u16( - vaddw_u8(vreinterpretq_u16_s16(q6s16), vreinterpret_u8_s16(d6s16))); - q7s16 = vreinterpretq_s16_u16( - vaddw_u8(vreinterpretq_u16_s16(q7s16), vreinterpret_u8_s16(d7s16))); - q4s16 = vreinterpretq_s16_u16( - vaddw_u8(vreinterpretq_u16_s16(q4s16), vreinterpret_u8_s16(d4s16))); - - d5s16 = vreinterpret_s16_u8(vqmovun_s16(q5s16)); - d6s16 = vreinterpret_s16_u8(vqmovun_s16(q6s16)); - d7s16 = vreinterpret_s16_u8(vqmovun_s16(q7s16)); - d4s16 = vreinterpret_s16_u8(vqmovun_s16(q4s16)); - - vst1_s16((int16_t *)p1, d5s16); - p1 -= stride; - vst1_s16((int16_t *)p2, d6s16); - p2 += stride; - vst1_s16((int16_t *)p2, d7s16); - vst1_s16((int16_t *)p1, d4s16); - return; -} - -#define DO_BUTTERFLY_STD(const_1, const_2, qA, qB) \ - DO_BUTTERFLY(q14s16, q13s16, const_1, const_2, qA, qB); -static INLINE void DO_BUTTERFLY(int16x8_t q14s16, int16x8_t q13s16, - int16_t first_const, int16_t second_const, - int16x8_t *qAs16, int16x8_t *qBs16) { - int16x4_t d30s16, d31s16; - int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q15s32; - int16x4_t dCs16, dDs16, dAs16, dBs16; - - dCs16 = vget_low_s16(q14s16); - dDs16 = vget_high_s16(q14s16); - dAs16 = vget_low_s16(q13s16); - dBs16 = vget_high_s16(q13s16); - - d30s16 = vdup_n_s16(first_const); - d31s16 = vdup_n_s16(second_const); - - q8s32 = vmull_s16(dCs16, d30s16); - q10s32 = vmull_s16(dAs16, d31s16); - q9s32 = vmull_s16(dDs16, d30s16); - q11s32 = vmull_s16(dBs16, d31s16); - q12s32 = vmull_s16(dCs16, d31s16); - - q8s32 = vsubq_s32(q8s32, q10s32); - q9s32 = vsubq_s32(q9s32, q11s32); - - q10s32 = vmull_s16(dDs16, d31s16); - q11s32 = vmull_s16(dAs16, d30s16); - q15s32 = vmull_s16(dBs16, d30s16); - - q11s32 = vaddq_s32(q12s32, q11s32); - q10s32 = vaddq_s32(q10s32, q15s32); - - *qAs16 = vcombine_s16(vqrshrn_n_s32(q8s32, 14), vqrshrn_n_s32(q9s32, 14)); - *qBs16 = vcombine_s16(vqrshrn_n_s32(q11s32, 14), vqrshrn_n_s32(q10s32, 14)); - return; -} - -static INLINE void idct32_transpose_pair(int16_t *input, int16_t *t_buf) { - int16_t *in; - int i; - const int stride = 32; - int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16; - int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16; - int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16; - int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32; - int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16; - - for (i = 0; i < 4; i++, input += 8) { - in = input; - q8s16 = vld1q_s16(in); - in += stride; - q9s16 = vld1q_s16(in); - in += stride; - q10s16 = vld1q_s16(in); - in += stride; - q11s16 = vld1q_s16(in); - in += stride; - q12s16 = vld1q_s16(in); - in += stride; - q13s16 = vld1q_s16(in); - in += stride; - q14s16 = vld1q_s16(in); - in += stride; - q15s16 = vld1q_s16(in); - - d16s16 = vget_low_s16(q8s16); - d17s16 = vget_high_s16(q8s16); - d18s16 = vget_low_s16(q9s16); - d19s16 = vget_high_s16(q9s16); - d20s16 = vget_low_s16(q10s16); - d21s16 = vget_high_s16(q10s16); - d22s16 = vget_low_s16(q11s16); - d23s16 = vget_high_s16(q11s16); - d24s16 = vget_low_s16(q12s16); - d25s16 = vget_high_s16(q12s16); - d26s16 = vget_low_s16(q13s16); - d27s16 = vget_high_s16(q13s16); - d28s16 = vget_low_s16(q14s16); - d29s16 = vget_high_s16(q14s16); - d30s16 = vget_low_s16(q15s16); - d31s16 = vget_high_s16(q15s16); - - q8s16 = vcombine_s16(d16s16, d24s16); // vswp d17, d24 - q9s16 = vcombine_s16(d18s16, d26s16); // vswp d19, d26 - q10s16 = vcombine_s16(d20s16, d28s16); // vswp d21, d28 - q11s16 = vcombine_s16(d22s16, d30s16); // vswp d23, d30 - q12s16 = vcombine_s16(d17s16, d25s16); - q13s16 = vcombine_s16(d19s16, d27s16); - q14s16 = vcombine_s16(d21s16, d29s16); - q15s16 = vcombine_s16(d23s16, d31s16); - - q0x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q10s16)); - q1x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(q9s16), vreinterpretq_s32_s16(q11s16)); - q2x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(q12s16), vreinterpretq_s32_s16(q14s16)); - q3x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(q13s16), vreinterpretq_s32_s16(q15s16)); - - q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]), // q8 - vreinterpretq_s16_s32(q1x2s32.val[0])); // q9 - q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]), // q10 - vreinterpretq_s16_s32(q1x2s32.val[1])); // q11 - q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]), // q12 - vreinterpretq_s16_s32(q3x2s32.val[0])); // q13 - q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]), // q14 - vreinterpretq_s16_s32(q3x2s32.val[1])); // q15 - - vst1q_s16(t_buf, q0x2s16.val[0]); - t_buf += 8; - vst1q_s16(t_buf, q0x2s16.val[1]); - t_buf += 8; - vst1q_s16(t_buf, q1x2s16.val[0]); - t_buf += 8; - vst1q_s16(t_buf, q1x2s16.val[1]); - t_buf += 8; - vst1q_s16(t_buf, q2x2s16.val[0]); - t_buf += 8; - vst1q_s16(t_buf, q2x2s16.val[1]); - t_buf += 8; - vst1q_s16(t_buf, q3x2s16.val[0]); - t_buf += 8; - vst1q_s16(t_buf, q3x2s16.val[1]); - t_buf += 8; - } - return; -} - -static INLINE void idct32_bands_end_1st_pass(int16_t *out, int16x8_t q2s16, - int16x8_t q3s16, int16x8_t q6s16, - int16x8_t q7s16, int16x8_t q8s16, - int16x8_t q9s16, int16x8_t q10s16, - int16x8_t q11s16, int16x8_t q12s16, - int16x8_t q13s16, int16x8_t q14s16, - int16x8_t q15s16) { - int16x8_t q0s16, q1s16, q4s16, q5s16; - - STORE_IN_OUTPUT(17, 16, 17, q6s16, q7s16); - STORE_IN_OUTPUT(17, 14, 15, q8s16, q9s16); - - LOAD_FROM_OUTPUT(15, 30, 31, q0s16, q1s16); - q4s16 = vaddq_s16(q2s16, q1s16); - q5s16 = vaddq_s16(q3s16, q0s16); - q6s16 = vsubq_s16(q3s16, q0s16); - q7s16 = vsubq_s16(q2s16, q1s16); - STORE_IN_OUTPUT(31, 30, 31, q6s16, q7s16); - STORE_IN_OUTPUT(31, 0, 1, q4s16, q5s16); - - LOAD_FROM_OUTPUT(1, 12, 13, q0s16, q1s16); - q2s16 = vaddq_s16(q10s16, q1s16); - q3s16 = vaddq_s16(q11s16, q0s16); - q4s16 = vsubq_s16(q11s16, q0s16); - q5s16 = vsubq_s16(q10s16, q1s16); - - LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16); - q8s16 = vaddq_s16(q4s16, q1s16); - q9s16 = vaddq_s16(q5s16, q0s16); - q6s16 = vsubq_s16(q5s16, q0s16); - q7s16 = vsubq_s16(q4s16, q1s16); - STORE_IN_OUTPUT(19, 18, 19, q6s16, q7s16); - STORE_IN_OUTPUT(19, 12, 13, q8s16, q9s16); - - LOAD_FROM_OUTPUT(13, 28, 29, q0s16, q1s16); - q4s16 = vaddq_s16(q2s16, q1s16); - q5s16 = vaddq_s16(q3s16, q0s16); - q6s16 = vsubq_s16(q3s16, q0s16); - q7s16 = vsubq_s16(q2s16, q1s16); - STORE_IN_OUTPUT(29, 28, 29, q6s16, q7s16); - STORE_IN_OUTPUT(29, 2, 3, q4s16, q5s16); - - LOAD_FROM_OUTPUT(3, 10, 11, q0s16, q1s16); - q2s16 = vaddq_s16(q12s16, q1s16); - q3s16 = vaddq_s16(q13s16, q0s16); - q4s16 = vsubq_s16(q13s16, q0s16); - q5s16 = vsubq_s16(q12s16, q1s16); - - LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16); - q8s16 = vaddq_s16(q4s16, q1s16); - q9s16 = vaddq_s16(q5s16, q0s16); - q6s16 = vsubq_s16(q5s16, q0s16); - q7s16 = vsubq_s16(q4s16, q1s16); - STORE_IN_OUTPUT(21, 20, 21, q6s16, q7s16); - STORE_IN_OUTPUT(21, 10, 11, q8s16, q9s16); - - LOAD_FROM_OUTPUT(11, 26, 27, q0s16, q1s16); - q4s16 = vaddq_s16(q2s16, q1s16); - q5s16 = vaddq_s16(q3s16, q0s16); - q6s16 = vsubq_s16(q3s16, q0s16); - q7s16 = vsubq_s16(q2s16, q1s16); - STORE_IN_OUTPUT(27, 26, 27, q6s16, q7s16); - STORE_IN_OUTPUT(27, 4, 5, q4s16, q5s16); - - LOAD_FROM_OUTPUT(5, 8, 9, q0s16, q1s16); - q2s16 = vaddq_s16(q14s16, q1s16); - q3s16 = vaddq_s16(q15s16, q0s16); - q4s16 = vsubq_s16(q15s16, q0s16); - q5s16 = vsubq_s16(q14s16, q1s16); - - LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16); - q8s16 = vaddq_s16(q4s16, q1s16); - q9s16 = vaddq_s16(q5s16, q0s16); - q6s16 = vsubq_s16(q5s16, q0s16); - q7s16 = vsubq_s16(q4s16, q1s16); - STORE_IN_OUTPUT(23, 22, 23, q6s16, q7s16); - STORE_IN_OUTPUT(23, 8, 9, q8s16, q9s16); - - LOAD_FROM_OUTPUT(9, 24, 25, q0s16, q1s16); - q4s16 = vaddq_s16(q2s16, q1s16); - q5s16 = vaddq_s16(q3s16, q0s16); - q6s16 = vsubq_s16(q3s16, q0s16); - q7s16 = vsubq_s16(q2s16, q1s16); - STORE_IN_OUTPUT(25, 24, 25, q6s16, q7s16); - STORE_IN_OUTPUT(25, 6, 7, q4s16, q5s16); - return; -} - -static INLINE void idct32_bands_end_2nd_pass( - int16_t *out, uint8_t *dest, int stride, int16x8_t q2s16, int16x8_t q3s16, - int16x8_t q6s16, int16x8_t q7s16, int16x8_t q8s16, int16x8_t q9s16, - int16x8_t q10s16, int16x8_t q11s16, int16x8_t q12s16, int16x8_t q13s16, - int16x8_t q14s16, int16x8_t q15s16) { - uint8_t *r6 = dest + 31 * stride; - uint8_t *r7 = dest /* + 0 * stride*/; - uint8_t *r9 = dest + 15 * stride; - uint8_t *r10 = dest + 16 * stride; - int str2 = stride << 1; - int16x8_t q0s16, q1s16, q4s16, q5s16; - - STORE_COMBINE_CENTER_RESULTS(r10, r9); - r10 += str2; - r9 -= str2; - - LOAD_FROM_OUTPUT(17, 30, 31, q0s16, q1s16) - q4s16 = vaddq_s16(q2s16, q1s16); - q5s16 = vaddq_s16(q3s16, q0s16); - q6s16 = vsubq_s16(q3s16, q0s16); - q7s16 = vsubq_s16(q2s16, q1s16); - STORE_COMBINE_EXTREME_RESULTS(r7, r6); - r7 += str2; - r6 -= str2; - - LOAD_FROM_OUTPUT(31, 12, 13, q0s16, q1s16) - q2s16 = vaddq_s16(q10s16, q1s16); - q3s16 = vaddq_s16(q11s16, q0s16); - q4s16 = vsubq_s16(q11s16, q0s16); - q5s16 = vsubq_s16(q10s16, q1s16); - - LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16) - q8s16 = vaddq_s16(q4s16, q1s16); - q9s16 = vaddq_s16(q5s16, q0s16); - q6s16 = vsubq_s16(q5s16, q0s16); - q7s16 = vsubq_s16(q4s16, q1s16); - STORE_COMBINE_CENTER_RESULTS(r10, r9); - r10 += str2; - r9 -= str2; - - LOAD_FROM_OUTPUT(19, 28, 29, q0s16, q1s16) - q4s16 = vaddq_s16(q2s16, q1s16); - q5s16 = vaddq_s16(q3s16, q0s16); - q6s16 = vsubq_s16(q3s16, q0s16); - q7s16 = vsubq_s16(q2s16, q1s16); - STORE_COMBINE_EXTREME_RESULTS(r7, r6); - r7 += str2; - r6 -= str2; - - LOAD_FROM_OUTPUT(29, 10, 11, q0s16, q1s16) - q2s16 = vaddq_s16(q12s16, q1s16); - q3s16 = vaddq_s16(q13s16, q0s16); - q4s16 = vsubq_s16(q13s16, q0s16); - q5s16 = vsubq_s16(q12s16, q1s16); - - LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16) - q8s16 = vaddq_s16(q4s16, q1s16); - q9s16 = vaddq_s16(q5s16, q0s16); - q6s16 = vsubq_s16(q5s16, q0s16); - q7s16 = vsubq_s16(q4s16, q1s16); - STORE_COMBINE_CENTER_RESULTS(r10, r9); - r10 += str2; - r9 -= str2; - - LOAD_FROM_OUTPUT(21, 26, 27, q0s16, q1s16) - q4s16 = vaddq_s16(q2s16, q1s16); - q5s16 = vaddq_s16(q3s16, q0s16); - q6s16 = vsubq_s16(q3s16, q0s16); - q7s16 = vsubq_s16(q2s16, q1s16); - STORE_COMBINE_EXTREME_RESULTS(r7, r6); - r7 += str2; - r6 -= str2; - - LOAD_FROM_OUTPUT(27, 8, 9, q0s16, q1s16) - q2s16 = vaddq_s16(q14s16, q1s16); - q3s16 = vaddq_s16(q15s16, q0s16); - q4s16 = vsubq_s16(q15s16, q0s16); - q5s16 = vsubq_s16(q14s16, q1s16); - - LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16) - q8s16 = vaddq_s16(q4s16, q1s16); - q9s16 = vaddq_s16(q5s16, q0s16); - q6s16 = vsubq_s16(q5s16, q0s16); - q7s16 = vsubq_s16(q4s16, q1s16); - STORE_COMBINE_CENTER_RESULTS(r10, r9); - - LOAD_FROM_OUTPUT(23, 24, 25, q0s16, q1s16) - q4s16 = vaddq_s16(q2s16, q1s16); - q5s16 = vaddq_s16(q3s16, q0s16); - q6s16 = vsubq_s16(q3s16, q0s16); - q7s16 = vsubq_s16(q2s16, q1s16); - STORE_COMBINE_EXTREME_RESULTS(r7, r6); - return; -} - -void aom_idct32x32_1024_add_neon(int16_t *input, uint8_t *dest, int stride) { - int i, idct32_pass_loop; - int16_t trans_buf[32 * 8]; - int16_t pass1[32 * 32]; - int16_t pass2[32 * 32]; - int16_t *out; - int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16; - int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16; - - for (idct32_pass_loop = 0, out = pass1; idct32_pass_loop < 2; - idct32_pass_loop++, - input = pass1, // the input of pass2 is the result of pass1 - out = pass2) { - for (i = 0; i < 4; i++, input += 32 * 8, out += 8) { // idct32_bands_loop - idct32_transpose_pair(input, trans_buf); - - // ----------------------------------------- - // BLOCK A: 16-19,28-31 - // ----------------------------------------- - // generate 16,17,30,31 - // part of stage 1 - LOAD_FROM_TRANSPOSED(0, 1, 31) - DO_BUTTERFLY_STD(cospi_31_64, cospi_1_64, &q0s16, &q2s16) - LOAD_FROM_TRANSPOSED(31, 17, 15) - DO_BUTTERFLY_STD(cospi_15_64, cospi_17_64, &q1s16, &q3s16) - // part of stage 2 - q4s16 = vaddq_s16(q0s16, q1s16); - q13s16 = vsubq_s16(q0s16, q1s16); - q6s16 = vaddq_s16(q2s16, q3s16); - q14s16 = vsubq_s16(q2s16, q3s16); - // part of stage 3 - DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q5s16, &q7s16) - - // generate 18,19,28,29 - // part of stage 1 - LOAD_FROM_TRANSPOSED(15, 9, 23) - DO_BUTTERFLY_STD(cospi_23_64, cospi_9_64, &q0s16, &q2s16) - LOAD_FROM_TRANSPOSED(23, 25, 7) - DO_BUTTERFLY_STD(cospi_7_64, cospi_25_64, &q1s16, &q3s16) - // part of stage 2 - q13s16 = vsubq_s16(q3s16, q2s16); - q3s16 = vaddq_s16(q3s16, q2s16); - q14s16 = vsubq_s16(q1s16, q0s16); - q2s16 = vaddq_s16(q1s16, q0s16); - // part of stage 3 - DO_BUTTERFLY_STD(-cospi_4_64, -cospi_28_64, &q1s16, &q0s16) - // part of stage 4 - q8s16 = vaddq_s16(q4s16, q2s16); - q9s16 = vaddq_s16(q5s16, q0s16); - q10s16 = vaddq_s16(q7s16, q1s16); - q15s16 = vaddq_s16(q6s16, q3s16); - q13s16 = vsubq_s16(q5s16, q0s16); - q14s16 = vsubq_s16(q7s16, q1s16); - STORE_IN_OUTPUT(0, 16, 31, q8s16, q15s16) - STORE_IN_OUTPUT(31, 17, 30, q9s16, q10s16) - // part of stage 5 - DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q0s16, &q1s16) - STORE_IN_OUTPUT(30, 29, 18, q1s16, q0s16) - // part of stage 4 - q13s16 = vsubq_s16(q4s16, q2s16); - q14s16 = vsubq_s16(q6s16, q3s16); - // part of stage 5 - DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q4s16, &q6s16) - STORE_IN_OUTPUT(18, 19, 28, q4s16, q6s16) - - // ----------------------------------------- - // BLOCK B: 20-23,24-27 - // ----------------------------------------- - // generate 20,21,26,27 - // part of stage 1 - LOAD_FROM_TRANSPOSED(7, 5, 27) - DO_BUTTERFLY_STD(cospi_27_64, cospi_5_64, &q0s16, &q2s16) - LOAD_FROM_TRANSPOSED(27, 21, 11) - DO_BUTTERFLY_STD(cospi_11_64, cospi_21_64, &q1s16, &q3s16) - // part of stage 2 - q13s16 = vsubq_s16(q0s16, q1s16); - q0s16 = vaddq_s16(q0s16, q1s16); - q14s16 = vsubq_s16(q2s16, q3s16); - q2s16 = vaddq_s16(q2s16, q3s16); - // part of stage 3 - DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16) - - // generate 22,23,24,25 - // part of stage 1 - LOAD_FROM_TRANSPOSED(11, 13, 19) - DO_BUTTERFLY_STD(cospi_19_64, cospi_13_64, &q5s16, &q7s16) - LOAD_FROM_TRANSPOSED(19, 29, 3) - DO_BUTTERFLY_STD(cospi_3_64, cospi_29_64, &q4s16, &q6s16) - // part of stage 2 - q14s16 = vsubq_s16(q4s16, q5s16); - q5s16 = vaddq_s16(q4s16, q5s16); - q13s16 = vsubq_s16(q6s16, q7s16); - q6s16 = vaddq_s16(q6s16, q7s16); - // part of stage 3 - DO_BUTTERFLY_STD(-cospi_20_64, -cospi_12_64, &q4s16, &q7s16) - // part of stage 4 - q10s16 = vaddq_s16(q7s16, q1s16); - q11s16 = vaddq_s16(q5s16, q0s16); - q12s16 = vaddq_s16(q6s16, q2s16); - q15s16 = vaddq_s16(q4s16, q3s16); - // part of stage 6 - LOAD_FROM_OUTPUT(28, 16, 17, q14s16, q13s16) - q8s16 = vaddq_s16(q14s16, q11s16); - q9s16 = vaddq_s16(q13s16, q10s16); - q13s16 = vsubq_s16(q13s16, q10s16); - q11s16 = vsubq_s16(q14s16, q11s16); - STORE_IN_OUTPUT(17, 17, 16, q9s16, q8s16) - LOAD_FROM_OUTPUT(16, 30, 31, q14s16, q9s16) - q8s16 = vsubq_s16(q9s16, q12s16); - q10s16 = vaddq_s16(q14s16, q15s16); - q14s16 = vsubq_s16(q14s16, q15s16); - q12s16 = vaddq_s16(q9s16, q12s16); - STORE_IN_OUTPUT(31, 30, 31, q10s16, q12s16) - // part of stage 7 - DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16) - STORE_IN_OUTPUT(31, 25, 22, q14s16, q13s16) - q13s16 = q11s16; - q14s16 = q8s16; - DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16) - STORE_IN_OUTPUT(22, 24, 23, q14s16, q13s16) - // part of stage 4 - q14s16 = vsubq_s16(q5s16, q0s16); - q13s16 = vsubq_s16(q6s16, q2s16); - DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q5s16, &q6s16); - q14s16 = vsubq_s16(q7s16, q1s16); - q13s16 = vsubq_s16(q4s16, q3s16); - DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q0s16, &q1s16); - // part of stage 6 - LOAD_FROM_OUTPUT(23, 18, 19, q14s16, q13s16) - q8s16 = vaddq_s16(q14s16, q1s16); - q9s16 = vaddq_s16(q13s16, q6s16); - q13s16 = vsubq_s16(q13s16, q6s16); - q1s16 = vsubq_s16(q14s16, q1s16); - STORE_IN_OUTPUT(19, 18, 19, q8s16, q9s16) - LOAD_FROM_OUTPUT(19, 28, 29, q8s16, q9s16) - q14s16 = vsubq_s16(q8s16, q5s16); - q10s16 = vaddq_s16(q8s16, q5s16); - q11s16 = vaddq_s16(q9s16, q0s16); - q0s16 = vsubq_s16(q9s16, q0s16); - STORE_IN_OUTPUT(29, 28, 29, q10s16, q11s16) - // part of stage 7 - DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16) - STORE_IN_OUTPUT(29, 20, 27, q13s16, q14s16) - DO_BUTTERFLY(q0s16, q1s16, cospi_16_64, cospi_16_64, &q1s16, &q0s16); - STORE_IN_OUTPUT(27, 21, 26, q1s16, q0s16) - - // ----------------------------------------- - // BLOCK C: 8-10,11-15 - // ----------------------------------------- - // generate 8,9,14,15 - // part of stage 2 - LOAD_FROM_TRANSPOSED(3, 2, 30) - DO_BUTTERFLY_STD(cospi_30_64, cospi_2_64, &q0s16, &q2s16) - LOAD_FROM_TRANSPOSED(30, 18, 14) - DO_BUTTERFLY_STD(cospi_14_64, cospi_18_64, &q1s16, &q3s16) - // part of stage 3 - q13s16 = vsubq_s16(q0s16, q1s16); - q0s16 = vaddq_s16(q0s16, q1s16); - q14s16 = vsubq_s16(q2s16, q3s16); - q2s16 = vaddq_s16(q2s16, q3s16); - // part of stage 4 - DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q1s16, &q3s16) - - // generate 10,11,12,13 - // part of stage 2 - LOAD_FROM_TRANSPOSED(14, 10, 22) - DO_BUTTERFLY_STD(cospi_22_64, cospi_10_64, &q5s16, &q7s16) - LOAD_FROM_TRANSPOSED(22, 26, 6) - DO_BUTTERFLY_STD(cospi_6_64, cospi_26_64, &q4s16, &q6s16) - // part of stage 3 - q14s16 = vsubq_s16(q4s16, q5s16); - q5s16 = vaddq_s16(q4s16, q5s16); - q13s16 = vsubq_s16(q6s16, q7s16); - q6s16 = vaddq_s16(q6s16, q7s16); - // part of stage 4 - DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q4s16, &q7s16) - // part of stage 5 - q8s16 = vaddq_s16(q0s16, q5s16); - q9s16 = vaddq_s16(q1s16, q7s16); - q13s16 = vsubq_s16(q1s16, q7s16); - q14s16 = vsubq_s16(q3s16, q4s16); - q10s16 = vaddq_s16(q3s16, q4s16); - q15s16 = vaddq_s16(q2s16, q6s16); - STORE_IN_OUTPUT(26, 8, 15, q8s16, q15s16) - STORE_IN_OUTPUT(15, 9, 14, q9s16, q10s16) - // part of stage 6 - DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16) - STORE_IN_OUTPUT(14, 13, 10, q3s16, q1s16) - q13s16 = vsubq_s16(q0s16, q5s16); - q14s16 = vsubq_s16(q2s16, q6s16); - DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16) - STORE_IN_OUTPUT(10, 11, 12, q1s16, q3s16) - - // ----------------------------------------- - // BLOCK D: 0-3,4-7 - // ----------------------------------------- - // generate 4,5,6,7 - // part of stage 3 - LOAD_FROM_TRANSPOSED(6, 4, 28) - DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q0s16, &q2s16) - LOAD_FROM_TRANSPOSED(28, 20, 12) - DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16) - // part of stage 4 - q13s16 = vsubq_s16(q0s16, q1s16); - q0s16 = vaddq_s16(q0s16, q1s16); - q14s16 = vsubq_s16(q2s16, q3s16); - q2s16 = vaddq_s16(q2s16, q3s16); - // part of stage 5 - DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16) - - // generate 0,1,2,3 - // part of stage 4 - LOAD_FROM_TRANSPOSED(12, 0, 16) - DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q5s16, &q7s16) - LOAD_FROM_TRANSPOSED(16, 8, 24) - DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q14s16, &q6s16) - // part of stage 5 - q4s16 = vaddq_s16(q7s16, q6s16); - q7s16 = vsubq_s16(q7s16, q6s16); - q6s16 = vsubq_s16(q5s16, q14s16); - q5s16 = vaddq_s16(q5s16, q14s16); - // part of stage 6 - q8s16 = vaddq_s16(q4s16, q2s16); - q9s16 = vaddq_s16(q5s16, q3s16); - q10s16 = vaddq_s16(q6s16, q1s16); - q11s16 = vaddq_s16(q7s16, q0s16); - q12s16 = vsubq_s16(q7s16, q0s16); - q13s16 = vsubq_s16(q6s16, q1s16); - q14s16 = vsubq_s16(q5s16, q3s16); - q15s16 = vsubq_s16(q4s16, q2s16); - // part of stage 7 - LOAD_FROM_OUTPUT(12, 14, 15, q0s16, q1s16) - q2s16 = vaddq_s16(q8s16, q1s16); - q3s16 = vaddq_s16(q9s16, q0s16); - q4s16 = vsubq_s16(q9s16, q0s16); - q5s16 = vsubq_s16(q8s16, q1s16); - LOAD_FROM_OUTPUT(15, 16, 17, q0s16, q1s16) - q8s16 = vaddq_s16(q4s16, q1s16); - q9s16 = vaddq_s16(q5s16, q0s16); - q6s16 = vsubq_s16(q5s16, q0s16); - q7s16 = vsubq_s16(q4s16, q1s16); - - if (idct32_pass_loop == 0) { - idct32_bands_end_1st_pass(out, q2s16, q3s16, q6s16, q7s16, q8s16, q9s16, - q10s16, q11s16, q12s16, q13s16, q14s16, - q15s16); - } else { - idct32_bands_end_2nd_pass(out, dest, stride, q2s16, q3s16, q6s16, q7s16, - q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, - q14s16, q15s16); - dest += 8; - } - } - } - return; -} diff --git a/aom_dsp/arm/idct4x4_1_add_neon.asm b/aom_dsp/arm/idct4x4_1_add_neon.asm deleted file mode 100644 index 6bd733d5d..000000000 --- a/aom_dsp/arm/idct4x4_1_add_neon.asm +++ /dev/null @@ -1,71 +0,0 @@ -; -; Copyright (c) 2016, 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. -; - - - - EXPORT |aom_idct4x4_1_add_neon| - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - -;void aom_idct4x4_1_add_neon(int16_t *input, uint8_t *dest, -; int dest_stride) -; -; r0 int16_t input -; r1 uint8_t *dest -; r2 int dest_stride) - -|aom_idct4x4_1_add_neon| PROC - ldrsh r0, [r0] - - ; generate cospi_16_64 = 11585 - mov r12, #0x2d00 - add r12, #0x41 - - ; out = dct_const_round_shift(input[0] * cospi_16_64) - mul r0, r0, r12 ; input[0] * cospi_16_64 - add r0, r0, #0x2000 ; +(1 << ((DCT_CONST_BITS) - 1)) - asr r0, r0, #14 ; >> DCT_CONST_BITS - - ; out = dct_const_round_shift(out * cospi_16_64) - mul r0, r0, r12 ; out * cospi_16_64 - mov r12, r1 ; save dest - add r0, r0, #0x2000 ; +(1 << ((DCT_CONST_BITS) - 1)) - asr r0, r0, #14 ; >> DCT_CONST_BITS - - ; a1 = ROUND_POWER_OF_TWO(out, 4) - add r0, r0, #8 ; + (1 <<((4) - 1)) - asr r0, r0, #4 ; >> 4 - - vdup.s16 q0, r0 ; duplicate a1 - - vld1.32 {d2[0]}, [r1], r2 - vld1.32 {d2[1]}, [r1], r2 - vld1.32 {d4[0]}, [r1], r2 - vld1.32 {d4[1]}, [r1] - - vaddw.u8 q8, q0, d2 ; dest[x] + a1 - vaddw.u8 q9, q0, d4 - - vqmovun.s16 d6, q8 ; clip_pixel - vqmovun.s16 d7, q9 - - vst1.32 {d6[0]}, [r12], r2 - vst1.32 {d6[1]}, [r12], r2 - vst1.32 {d7[0]}, [r12], r2 - vst1.32 {d7[1]}, [r12] - - bx lr - ENDP ; |aom_idct4x4_1_add_neon| - - END diff --git a/aom_dsp/arm/idct4x4_1_add_neon.c b/aom_dsp/arm/idct4x4_1_add_neon.c deleted file mode 100644 index 3df7a901b..000000000 --- a/aom_dsp/arm/idct4x4_1_add_neon.c +++ /dev/null @@ -1,47 +0,0 @@ -/* - * Copyright (c) 2016, 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 "aom_dsp/inv_txfm.h" -#include "aom_ports/mem.h" - -void aom_idct4x4_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) { - uint8x8_t d6u8; - uint32x2_t d2u32 = vdup_n_u32(0); - uint16x8_t q8u16; - int16x8_t q0s16; - uint8_t *d1, *d2; - int16_t i, a1; - int16_t out = dct_const_round_shift(input[0] * cospi_16_64); - out = dct_const_round_shift(out * cospi_16_64); - a1 = ROUND_POWER_OF_TWO(out, 4); - - q0s16 = vdupq_n_s16(a1); - - // dc_only_idct_add - d1 = d2 = dest; - for (i = 0; i < 2; i++) { - d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 0); - d1 += dest_stride; - d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 1); - d1 += dest_stride; - - q8u16 = vaddw_u8(vreinterpretq_u16_s16(q0s16), vreinterpret_u8_u32(d2u32)); - d6u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16)); - - vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 0); - d2 += dest_stride; - vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 1); - d2 += dest_stride; - } - return; -} diff --git a/aom_dsp/arm/idct4x4_add_neon.asm b/aom_dsp/arm/idct4x4_add_neon.asm deleted file mode 100644 index 127acf614..000000000 --- a/aom_dsp/arm/idct4x4_add_neon.asm +++ /dev/null @@ -1,193 +0,0 @@ -; -; Copyright (c) 2016, 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. -; - -; - - EXPORT |aom_idct4x4_16_add_neon| - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - - AREA Block, CODE, READONLY ; name this block of code -;void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride) -; -; r0 int16_t input -; r1 uint8_t *dest -; r2 int dest_stride) - -|aom_idct4x4_16_add_neon| PROC - - ; The 2D transform is done with two passes which are actually pretty - ; similar. We first transform the rows. This is done by transposing - ; the inputs, doing an SIMD column transform (the columns are the - ; transposed rows) and then transpose the results (so that it goes back - ; in normal/row positions). Then, we transform the columns by doing - ; another SIMD column transform. - ; So, two passes of a transpose followed by a column transform. - - ; load the inputs into q8-q9, d16-d19 - vld1.s16 {q8,q9}, [r0]! - - ; generate scalar constants - ; cospi_8_64 = 15137 = 0x3b21 - mov r0, #0x3b00 - add r0, #0x21 - ; cospi_16_64 = 11585 = 0x2d41 - mov r3, #0x2d00 - add r3, #0x41 - ; cospi_24_64 = 6270 = 0x 187e - mov r12, #0x1800 - add r12, #0x7e - - ; transpose the input data - ; 00 01 02 03 d16 - ; 10 11 12 13 d17 - ; 20 21 22 23 d18 - ; 30 31 32 33 d19 - vtrn.16 d16, d17 - vtrn.16 d18, d19 - - ; generate constant vectors - vdup.16 d20, r0 ; replicate cospi_8_64 - vdup.16 d21, r3 ; replicate cospi_16_64 - - ; 00 10 02 12 d16 - ; 01 11 03 13 d17 - ; 20 30 22 32 d18 - ; 21 31 23 33 d19 - vtrn.32 q8, q9 - ; 00 10 20 30 d16 - ; 01 11 21 31 d17 - ; 02 12 22 32 d18 - ; 03 13 23 33 d19 - - vdup.16 d22, r12 ; replicate cospi_24_64 - - ; do the transform on transposed rows - - ; stage 1 - vadd.s16 d23, d16, d18 ; (input[0] + input[2]) - vsub.s16 d24, d16, d18 ; (input[0] - input[2]) - - vmull.s16 q15, d17, d22 ; input[1] * cospi_24_64 - vmull.s16 q1, d17, d20 ; input[1] * cospi_8_64 - - ; (input[0] + input[2]) * cospi_16_64; - ; (input[0] - input[2]) * cospi_16_64; - vmull.s16 q13, d23, d21 - vmull.s16 q14, d24, d21 - - ; input[1] * cospi_24_64 - input[3] * cospi_8_64; - ; input[1] * cospi_8_64 + input[3] * cospi_24_64; - vmlsl.s16 q15, d19, d20 - vmlal.s16 q1, d19, d22 - - ; dct_const_round_shift - vqrshrn.s32 d26, q13, #14 - vqrshrn.s32 d27, q14, #14 - vqrshrn.s32 d29, q15, #14 - vqrshrn.s32 d28, q1, #14 - - ; stage 2 - ; output[0] = step[0] + step[3]; - ; output[1] = step[1] + step[2]; - ; output[3] = step[0] - step[3]; - ; output[2] = step[1] - step[2]; - vadd.s16 q8, q13, q14 - vsub.s16 q9, q13, q14 - vswp d18, d19 - - ; transpose the results - ; 00 01 02 03 d16 - ; 10 11 12 13 d17 - ; 20 21 22 23 d18 - ; 30 31 32 33 d19 - vtrn.16 d16, d17 - vtrn.16 d18, d19 - ; 00 10 02 12 d16 - ; 01 11 03 13 d17 - ; 20 30 22 32 d18 - ; 21 31 23 33 d19 - vtrn.32 q8, q9 - ; 00 10 20 30 d16 - ; 01 11 21 31 d17 - ; 02 12 22 32 d18 - ; 03 13 23 33 d19 - - ; do the transform on columns - - ; stage 1 - vadd.s16 d23, d16, d18 ; (input[0] + input[2]) - vsub.s16 d24, d16, d18 ; (input[0] - input[2]) - - vmull.s16 q15, d17, d22 ; input[1] * cospi_24_64 - vmull.s16 q1, d17, d20 ; input[1] * cospi_8_64 - - ; (input[0] + input[2]) * cospi_16_64; - ; (input[0] - input[2]) * cospi_16_64; - vmull.s16 q13, d23, d21 - vmull.s16 q14, d24, d21 - - ; input[1] * cospi_24_64 - input[3] * cospi_8_64; - ; input[1] * cospi_8_64 + input[3] * cospi_24_64; - vmlsl.s16 q15, d19, d20 - vmlal.s16 q1, d19, d22 - - ; dct_const_round_shift - vqrshrn.s32 d26, q13, #14 - vqrshrn.s32 d27, q14, #14 - vqrshrn.s32 d29, q15, #14 - vqrshrn.s32 d28, q1, #14 - - ; stage 2 - ; output[0] = step[0] + step[3]; - ; output[1] = step[1] + step[2]; - ; output[3] = step[0] - step[3]; - ; output[2] = step[1] - step[2]; - vadd.s16 q8, q13, q14 - vsub.s16 q9, q13, q14 - - ; The results are in two registers, one of them being swapped. This will - ; be taken care of by loading the 'dest' value in a swapped fashion and - ; also storing them in the same swapped fashion. - ; temp_out[0, 1] = d16, d17 = q8 - ; temp_out[2, 3] = d19, d18 = q9 swapped - - ; ROUND_POWER_OF_TWO(temp_out[j], 4) - vrshr.s16 q8, q8, #4 - vrshr.s16 q9, q9, #4 - - vld1.32 {d26[0]}, [r1], r2 - vld1.32 {d26[1]}, [r1], r2 - vld1.32 {d27[1]}, [r1], r2 - vld1.32 {d27[0]}, [r1] ; no post-increment - - ; ROUND_POWER_OF_TWO(temp_out[j], 4) + dest[j * dest_stride + i] - vaddw.u8 q8, q8, d26 - vaddw.u8 q9, q9, d27 - - ; clip_pixel - vqmovun.s16 d26, q8 - vqmovun.s16 d27, q9 - - ; do the stores in reverse order with negative post-increment, by changing - ; the sign of the stride - rsb r2, r2, #0 - vst1.32 {d27[0]}, [r1], r2 - vst1.32 {d27[1]}, [r1], r2 - vst1.32 {d26[1]}, [r1], r2 - vst1.32 {d26[0]}, [r1] ; no post-increment - bx lr - ENDP ; |aom_idct4x4_16_add_neon| - - END diff --git a/aom_dsp/arm/idct4x4_add_neon.c b/aom_dsp/arm/idct4x4_add_neon.c deleted file mode 100644 index 763be1ab0..000000000 --- a/aom_dsp/arm/idct4x4_add_neon.c +++ /dev/null @@ -1,146 +0,0 @@ -/* - * Copyright (c) 2016, 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 "aom_dsp/txfm_common.h" - -void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride) { - uint8x8_t d26u8, d27u8; - uint32x2_t d26u32, d27u32; - uint16x8_t q8u16, q9u16; - int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16; - int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16; - int16x8_t q8s16, q9s16, q13s16, q14s16; - int32x4_t q1s32, q13s32, q14s32, q15s32; - int16x4x2_t d0x2s16, d1x2s16; - int32x4x2_t q0x2s32; - uint8_t *d; - - d26u32 = d27u32 = vdup_n_u32(0); - - q8s16 = vld1q_s16(input); - q9s16 = vld1q_s16(input + 8); - - d16s16 = vget_low_s16(q8s16); - d17s16 = vget_high_s16(q8s16); - d18s16 = vget_low_s16(q9s16); - d19s16 = vget_high_s16(q9s16); - - d0x2s16 = vtrn_s16(d16s16, d17s16); - d1x2s16 = vtrn_s16(d18s16, d19s16); - q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]); - q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]); - - d20s16 = vdup_n_s16((int16_t)cospi_8_64); - d21s16 = vdup_n_s16((int16_t)cospi_16_64); - - q0x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16)); - d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0])); - d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0])); - d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1])); - d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1])); - - d22s16 = vdup_n_s16((int16_t)cospi_24_64); - - // stage 1 - d23s16 = vadd_s16(d16s16, d18s16); - d24s16 = vsub_s16(d16s16, d18s16); - - q15s32 = vmull_s16(d17s16, d22s16); - q1s32 = vmull_s16(d17s16, d20s16); - q13s32 = vmull_s16(d23s16, d21s16); - q14s32 = vmull_s16(d24s16, d21s16); - - q15s32 = vmlsl_s16(q15s32, d19s16, d20s16); - q1s32 = vmlal_s16(q1s32, d19s16, d22s16); - - d26s16 = vqrshrn_n_s32(q13s32, 14); - d27s16 = vqrshrn_n_s32(q14s32, 14); - d29s16 = vqrshrn_n_s32(q15s32, 14); - d28s16 = vqrshrn_n_s32(q1s32, 14); - q13s16 = vcombine_s16(d26s16, d27s16); - q14s16 = vcombine_s16(d28s16, d29s16); - - // stage 2 - q8s16 = vaddq_s16(q13s16, q14s16); - q9s16 = vsubq_s16(q13s16, q14s16); - - d16s16 = vget_low_s16(q8s16); - d17s16 = vget_high_s16(q8s16); - d18s16 = vget_high_s16(q9s16); // vswp d18 d19 - d19s16 = vget_low_s16(q9s16); - - d0x2s16 = vtrn_s16(d16s16, d17s16); - d1x2s16 = vtrn_s16(d18s16, d19s16); - q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]); - q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]); - - q0x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16)); - d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0])); - d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0])); - d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1])); - d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1])); - - // do the transform on columns - // stage 1 - d23s16 = vadd_s16(d16s16, d18s16); - d24s16 = vsub_s16(d16s16, d18s16); - - q15s32 = vmull_s16(d17s16, d22s16); - q1s32 = vmull_s16(d17s16, d20s16); - q13s32 = vmull_s16(d23s16, d21s16); - q14s32 = vmull_s16(d24s16, d21s16); - - q15s32 = vmlsl_s16(q15s32, d19s16, d20s16); - q1s32 = vmlal_s16(q1s32, d19s16, d22s16); - - d26s16 = vqrshrn_n_s32(q13s32, 14); - d27s16 = vqrshrn_n_s32(q14s32, 14); - d29s16 = vqrshrn_n_s32(q15s32, 14); - d28s16 = vqrshrn_n_s32(q1s32, 14); - q13s16 = vcombine_s16(d26s16, d27s16); - q14s16 = vcombine_s16(d28s16, d29s16); - - // stage 2 - q8s16 = vaddq_s16(q13s16, q14s16); - q9s16 = vsubq_s16(q13s16, q14s16); - - q8s16 = vrshrq_n_s16(q8s16, 4); - q9s16 = vrshrq_n_s16(q9s16, 4); - - d = dest; - d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0); - d += dest_stride; - d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1); - d += dest_stride; - d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1); - d += dest_stride; - d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0); - - q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32)); - q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32)); - - d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16)); - d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16)); - - d = dest; - vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0); - d += dest_stride; - vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1); - d += dest_stride; - vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1); - d += dest_stride; - vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0); - return; -} diff --git a/aom_dsp/arm/idct8x8_1_add_neon.asm b/aom_dsp/arm/idct8x8_1_add_neon.asm deleted file mode 100644 index ec07e2053..000000000 --- a/aom_dsp/arm/idct8x8_1_add_neon.asm +++ /dev/null @@ -1,91 +0,0 @@ -; -; Copyright (c) 2016, 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. -; - - - - EXPORT |aom_idct8x8_1_add_neon| - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - -;void aom_idct8x8_1_add_neon(int16_t *input, uint8_t *dest, -; int dest_stride) -; -; r0 int16_t input -; r1 uint8_t *dest -; r2 int dest_stride) - -|aom_idct8x8_1_add_neon| PROC - ldrsh r0, [r0] - - ; generate cospi_16_64 = 11585 - mov r12, #0x2d00 - add r12, #0x41 - - ; out = dct_const_round_shift(input[0] * cospi_16_64) - mul r0, r0, r12 ; input[0] * cospi_16_64 - add r0, r0, #0x2000 ; +(1 << ((DCT_CONST_BITS) - 1)) - asr r0, r0, #14 ; >> DCT_CONST_BITS - - ; out = dct_const_round_shift(out * cospi_16_64) - mul r0, r0, r12 ; out * cospi_16_64 - mov r12, r1 ; save dest - add r0, r0, #0x2000 ; +(1 << ((DCT_CONST_BITS) - 1)) - asr r0, r0, #14 ; >> DCT_CONST_BITS - - ; a1 = ROUND_POWER_OF_TWO(out, 5) - add r0, r0, #16 ; + (1 <<((5) - 1)) - asr r0, r0, #5 ; >> 5 - - vdup.s16 q0, r0 ; duplicate a1 - - ; load destination data - vld1.64 {d2}, [r1], r2 - vld1.64 {d3}, [r1], r2 - vld1.64 {d4}, [r1], r2 - vld1.64 {d5}, [r1], r2 - vld1.64 {d6}, [r1], r2 - vld1.64 {d7}, [r1], r2 - vld1.64 {d16}, [r1], r2 - vld1.64 {d17}, [r1] - - vaddw.u8 q9, q0, d2 ; dest[x] + a1 - vaddw.u8 q10, q0, d3 ; dest[x] + a1 - vaddw.u8 q11, q0, d4 ; dest[x] + a1 - vaddw.u8 q12, q0, d5 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r2 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r2 - vst1.64 {d31}, [r12], r2 - - vaddw.u8 q9, q0, d6 ; dest[x] + a1 - vaddw.u8 q10, q0, d7 ; dest[x] + a1 - vaddw.u8 q11, q0, d16 ; dest[x] + a1 - vaddw.u8 q12, q0, d17 ; dest[x] + a1 - vqmovun.s16 d2, q9 ; clip_pixel - vqmovun.s16 d3, q10 ; clip_pixel - vqmovun.s16 d30, q11 ; clip_pixel - vqmovun.s16 d31, q12 ; clip_pixel - vst1.64 {d2}, [r12], r2 - vst1.64 {d3}, [r12], r2 - vst1.64 {d30}, [r12], r2 - vst1.64 {d31}, [r12], r2 - - bx lr - ENDP ; |aom_idct8x8_1_add_neon| - - END diff --git a/aom_dsp/arm/idct8x8_1_add_neon.c b/aom_dsp/arm/idct8x8_1_add_neon.c deleted file mode 100644 index c7926f9e4..000000000 --- a/aom_dsp/arm/idct8x8_1_add_neon.c +++ /dev/null @@ -1,62 +0,0 @@ -/* - * Copyright (c) 2016, 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 "aom_dsp/inv_txfm.h" -#include "aom_ports/mem.h" - -void aom_idct8x8_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) { - uint8x8_t d2u8, d3u8, d30u8, d31u8; - uint64x1_t d2u64, d3u64, d4u64, d5u64; - uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16; - int16x8_t q0s16; - uint8_t *d1, *d2; - int16_t i, a1; - int16_t out = dct_const_round_shift(input[0] * cospi_16_64); - out = dct_const_round_shift(out * cospi_16_64); - a1 = ROUND_POWER_OF_TWO(out, 5); - - q0s16 = vdupq_n_s16(a1); - q0u16 = vreinterpretq_u16_s16(q0s16); - - d1 = d2 = dest; - for (i = 0; i < 2; i++) { - d2u64 = vld1_u64((const uint64_t *)d1); - d1 += dest_stride; - d3u64 = vld1_u64((const uint64_t *)d1); - d1 += dest_stride; - d4u64 = vld1_u64((const uint64_t *)d1); - d1 += dest_stride; - d5u64 = vld1_u64((const uint64_t *)d1); - d1 += dest_stride; - - q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64)); - q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64)); - q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64)); - q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64)); - - d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16)); - d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16)); - d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16)); - d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16)); - - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d31u8)); - d2 += dest_stride; - } - return; -} diff --git a/aom_dsp/arm/idct8x8_add_neon.asm b/aom_dsp/arm/idct8x8_add_neon.asm deleted file mode 100644 index f3d5f246d..000000000 --- a/aom_dsp/arm/idct8x8_add_neon.asm +++ /dev/null @@ -1,522 +0,0 @@ -; -; Copyright (c) 2016, 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. -; - -; - - EXPORT |aom_idct8x8_64_add_neon| - EXPORT |aom_idct8x8_12_add_neon| - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - - ; Parallel 1D IDCT on all the columns of a 8x8 16bit data matrix which are - ; loaded in q8-q15. The output will be stored back into q8-q15 registers. - ; This macro will touch q0-q7 registers and use them as buffer during - ; calculation. - MACRO - IDCT8x8_1D - ; stage 1 - vdup.16 d0, r3 ; duplicate cospi_28_64 - vdup.16 d1, r4 ; duplicate cospi_4_64 - vdup.16 d2, r5 ; duplicate cospi_12_64 - vdup.16 d3, r6 ; duplicate cospi_20_64 - - ; input[1] * cospi_28_64 - vmull.s16 q2, d18, d0 - vmull.s16 q3, d19, d0 - - ; input[5] * cospi_12_64 - vmull.s16 q5, d26, d2 - vmull.s16 q6, d27, d2 - - ; input[1]*cospi_28_64-input[7]*cospi_4_64 - vmlsl.s16 q2, d30, d1 - vmlsl.s16 q3, d31, d1 - - ; input[5] * cospi_12_64 - input[3] * cospi_20_64 - vmlsl.s16 q5, d22, d3 - vmlsl.s16 q6, d23, d3 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d8, q2, #14 ; >> 14 - vqrshrn.s32 d9, q3, #14 ; >> 14 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d10, q5, #14 ; >> 14 - vqrshrn.s32 d11, q6, #14 ; >> 14 - - ; input[1] * cospi_4_64 - vmull.s16 q2, d18, d1 - vmull.s16 q3, d19, d1 - - ; input[5] * cospi_20_64 - vmull.s16 q9, d26, d3 - vmull.s16 q13, d27, d3 - - ; input[1]*cospi_4_64+input[7]*cospi_28_64 - vmlal.s16 q2, d30, d0 - vmlal.s16 q3, d31, d0 - - ; input[5] * cospi_20_64 + input[3] * cospi_12_64 - vmlal.s16 q9, d22, d2 - vmlal.s16 q13, d23, d2 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d14, q2, #14 ; >> 14 - vqrshrn.s32 d15, q3, #14 ; >> 14 - - ; stage 2 & stage 3 - even half - vdup.16 d0, r7 ; duplicate cospi_16_64 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d12, q9, #14 ; >> 14 - vqrshrn.s32 d13, q13, #14 ; >> 14 - - ; input[0] * cospi_16_64 - vmull.s16 q2, d16, d0 - vmull.s16 q3, d17, d0 - - ; input[0] * cospi_16_64 - vmull.s16 q13, d16, d0 - vmull.s16 q15, d17, d0 - - ; (input[0] + input[2]) * cospi_16_64 - vmlal.s16 q2, d24, d0 - vmlal.s16 q3, d25, d0 - - ; (input[0] - input[2]) * cospi_16_64 - vmlsl.s16 q13, d24, d0 - vmlsl.s16 q15, d25, d0 - - vdup.16 d0, r8 ; duplicate cospi_24_64 - vdup.16 d1, r9 ; duplicate cospi_8_64 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d18, q2, #14 ; >> 14 - vqrshrn.s32 d19, q3, #14 ; >> 14 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d22, q13, #14 ; >> 14 - vqrshrn.s32 d23, q15, #14 ; >> 14 - - ; input[1] * cospi_24_64 - input[3] * cospi_8_64 - ; input[1] * cospi_24_64 - vmull.s16 q2, d20, d0 - vmull.s16 q3, d21, d0 - - ; input[1] * cospi_8_64 - vmull.s16 q8, d20, d1 - vmull.s16 q12, d21, d1 - - ; input[1] * cospi_24_64 - input[3] * cospi_8_64 - vmlsl.s16 q2, d28, d1 - vmlsl.s16 q3, d29, d1 - - ; input[1] * cospi_8_64 + input[3] * cospi_24_64 - vmlal.s16 q8, d28, d0 - vmlal.s16 q12, d29, d0 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d26, q2, #14 ; >> 14 - vqrshrn.s32 d27, q3, #14 ; >> 14 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d30, q8, #14 ; >> 14 - vqrshrn.s32 d31, q12, #14 ; >> 14 - - vadd.s16 q0, q9, q15 ; output[0] = step[0] + step[3] - vadd.s16 q1, q11, q13 ; output[1] = step[1] + step[2] - vsub.s16 q2, q11, q13 ; output[2] = step[1] - step[2] - vsub.s16 q3, q9, q15 ; output[3] = step[0] - step[3] - - ; stage 3 -odd half - vdup.16 d16, r7 ; duplicate cospi_16_64 - - ; stage 2 - odd half - vsub.s16 q13, q4, q5 ; step2[5] = step1[4] - step1[5] - vadd.s16 q4, q4, q5 ; step2[4] = step1[4] + step1[5] - vsub.s16 q14, q7, q6 ; step2[6] = -step1[6] + step1[7] - vadd.s16 q7, q7, q6 ; step2[7] = step1[6] + step1[7] - - ; step2[6] * cospi_16_64 - vmull.s16 q9, d28, d16 - vmull.s16 q10, d29, d16 - - ; step2[6] * cospi_16_64 - vmull.s16 q11, d28, d16 - vmull.s16 q12, d29, d16 - - ; (step2[6] - step2[5]) * cospi_16_64 - vmlsl.s16 q9, d26, d16 - vmlsl.s16 q10, d27, d16 - - ; (step2[5] + step2[6]) * cospi_16_64 - vmlal.s16 q11, d26, d16 - vmlal.s16 q12, d27, d16 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d10, q9, #14 ; >> 14 - vqrshrn.s32 d11, q10, #14 ; >> 14 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d12, q11, #14 ; >> 14 - vqrshrn.s32 d13, q12, #14 ; >> 14 - - ; stage 4 - vadd.s16 q8, q0, q7 ; output[0] = step1[0] + step1[7]; - vadd.s16 q9, q1, q6 ; output[1] = step1[1] + step1[6]; - vadd.s16 q10, q2, q5 ; output[2] = step1[2] + step1[5]; - vadd.s16 q11, q3, q4 ; output[3] = step1[3] + step1[4]; - vsub.s16 q12, q3, q4 ; output[4] = step1[3] - step1[4]; - vsub.s16 q13, q2, q5 ; output[5] = step1[2] - step1[5]; - vsub.s16 q14, q1, q6 ; output[6] = step1[1] - step1[6]; - vsub.s16 q15, q0, q7 ; output[7] = step1[0] - step1[7]; - MEND - - ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15. - MACRO - TRANSPOSE8X8 - vswp d17, d24 - vswp d23, d30 - vswp d21, d28 - vswp d19, d26 - vtrn.32 q8, q10 - vtrn.32 q9, q11 - vtrn.32 q12, q14 - vtrn.32 q13, q15 - vtrn.16 q8, q9 - vtrn.16 q10, q11 - vtrn.16 q12, q13 - vtrn.16 q14, q15 - MEND - - AREA Block, CODE, READONLY ; name this block of code -;void aom_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride) -; -; r0 int16_t input -; r1 uint8_t *dest -; r2 int dest_stride) - -|aom_idct8x8_64_add_neon| PROC - push {r4-r9} - vpush {d8-d15} - vld1.s16 {q8,q9}, [r0]! - vld1.s16 {q10,q11}, [r0]! - vld1.s16 {q12,q13}, [r0]! - vld1.s16 {q14,q15}, [r0]! - - ; transpose the input data - TRANSPOSE8X8 - - ; generate cospi_28_64 = 3196 - mov r3, #0x0c00 - add r3, #0x7c - - ; generate cospi_4_64 = 16069 - mov r4, #0x3e00 - add r4, #0xc5 - - ; generate cospi_12_64 = 13623 - mov r5, #0x3500 - add r5, #0x37 - - ; generate cospi_20_64 = 9102 - mov r6, #0x2300 - add r6, #0x8e - - ; generate cospi_16_64 = 11585 - mov r7, #0x2d00 - add r7, #0x41 - - ; generate cospi_24_64 = 6270 - mov r8, #0x1800 - add r8, #0x7e - - ; generate cospi_8_64 = 15137 - mov r9, #0x3b00 - add r9, #0x21 - - ; First transform rows - IDCT8x8_1D - - ; Transpose the matrix - TRANSPOSE8X8 - - ; Then transform columns - IDCT8x8_1D - - ; ROUND_POWER_OF_TWO(temp_out[j], 5) - vrshr.s16 q8, q8, #5 - vrshr.s16 q9, q9, #5 - vrshr.s16 q10, q10, #5 - vrshr.s16 q11, q11, #5 - vrshr.s16 q12, q12, #5 - vrshr.s16 q13, q13, #5 - vrshr.s16 q14, q14, #5 - vrshr.s16 q15, q15, #5 - - ; save dest pointer - mov r0, r1 - - ; load destination data - vld1.64 {d0}, [r1], r2 - vld1.64 {d1}, [r1], r2 - vld1.64 {d2}, [r1], r2 - vld1.64 {d3}, [r1], r2 - vld1.64 {d4}, [r1], r2 - vld1.64 {d5}, [r1], r2 - vld1.64 {d6}, [r1], r2 - vld1.64 {d7}, [r1] - - ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i] - vaddw.u8 q8, q8, d0 - vaddw.u8 q9, q9, d1 - vaddw.u8 q10, q10, d2 - vaddw.u8 q11, q11, d3 - vaddw.u8 q12, q12, d4 - vaddw.u8 q13, q13, d5 - vaddw.u8 q14, q14, d6 - vaddw.u8 q15, q15, d7 - - ; clip_pixel - vqmovun.s16 d0, q8 - vqmovun.s16 d1, q9 - vqmovun.s16 d2, q10 - vqmovun.s16 d3, q11 - vqmovun.s16 d4, q12 - vqmovun.s16 d5, q13 - vqmovun.s16 d6, q14 - vqmovun.s16 d7, q15 - - ; store the data - vst1.64 {d0}, [r0], r2 - vst1.64 {d1}, [r0], r2 - vst1.64 {d2}, [r0], r2 - vst1.64 {d3}, [r0], r2 - vst1.64 {d4}, [r0], r2 - vst1.64 {d5}, [r0], r2 - vst1.64 {d6}, [r0], r2 - vst1.64 {d7}, [r0], r2 - - vpop {d8-d15} - pop {r4-r9} - bx lr - ENDP ; |aom_idct8x8_64_add_neon| - -;void aom_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride) -; -; r0 int16_t input -; r1 uint8_t *dest -; r2 int dest_stride) - -|aom_idct8x8_12_add_neon| PROC - push {r4-r9} - vpush {d8-d15} - vld1.s16 {q8,q9}, [r0]! - vld1.s16 {q10,q11}, [r0]! - vld1.s16 {q12,q13}, [r0]! - vld1.s16 {q14,q15}, [r0]! - - ; transpose the input data - TRANSPOSE8X8 - - ; generate cospi_28_64 = 3196 - mov r3, #0x0c00 - add r3, #0x7c - - ; generate cospi_4_64 = 16069 - mov r4, #0x3e00 - add r4, #0xc5 - - ; generate cospi_12_64 = 13623 - mov r5, #0x3500 - add r5, #0x37 - - ; generate cospi_20_64 = 9102 - mov r6, #0x2300 - add r6, #0x8e - - ; generate cospi_16_64 = 11585 - mov r7, #0x2d00 - add r7, #0x41 - - ; generate cospi_24_64 = 6270 - mov r8, #0x1800 - add r8, #0x7e - - ; generate cospi_8_64 = 15137 - mov r9, #0x3b00 - add r9, #0x21 - - ; First transform rows - ; stage 1 - ; The following instructions use vqrdmulh to do the - ; dct_const_round_shift(input[1] * cospi_28_64). vqrdmulh will do doubling - ; multiply and shift the result by 16 bits instead of 14 bits. So we need - ; to double the constants before multiplying to compensate this. - mov r12, r3, lsl #1 - vdup.16 q0, r12 ; duplicate cospi_28_64*2 - mov r12, r4, lsl #1 - vdup.16 q1, r12 ; duplicate cospi_4_64*2 - - ; dct_const_round_shift(input[1] * cospi_28_64) - vqrdmulh.s16 q4, q9, q0 - - mov r12, r6, lsl #1 - rsb r12, #0 - vdup.16 q0, r12 ; duplicate -cospi_20_64*2 - - ; dct_const_round_shift(input[1] * cospi_4_64) - vqrdmulh.s16 q7, q9, q1 - - mov r12, r5, lsl #1 - vdup.16 q1, r12 ; duplicate cospi_12_64*2 - - ; dct_const_round_shift(- input[3] * cospi_20_64) - vqrdmulh.s16 q5, q11, q0 - - mov r12, r7, lsl #1 - vdup.16 q0, r12 ; duplicate cospi_16_64*2 - - ; dct_const_round_shift(input[3] * cospi_12_64) - vqrdmulh.s16 q6, q11, q1 - - ; stage 2 & stage 3 - even half - mov r12, r8, lsl #1 - vdup.16 q1, r12 ; duplicate cospi_24_64*2 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrdmulh.s16 q9, q8, q0 - - mov r12, r9, lsl #1 - vdup.16 q0, r12 ; duplicate cospi_8_64*2 - - ; dct_const_round_shift(input[1] * cospi_24_64) - vqrdmulh.s16 q13, q10, q1 - - ; dct_const_round_shift(input[1] * cospi_8_64) - vqrdmulh.s16 q15, q10, q0 - - ; stage 3 -odd half - vdup.16 d16, r7 ; duplicate cospi_16_64 - - vadd.s16 q0, q9, q15 ; output[0] = step[0] + step[3] - vadd.s16 q1, q9, q13 ; output[1] = step[1] + step[2] - vsub.s16 q2, q9, q13 ; output[2] = step[1] - step[2] - vsub.s16 q3, q9, q15 ; output[3] = step[0] - step[3] - - ; stage 2 - odd half - vsub.s16 q13, q4, q5 ; step2[5] = step1[4] - step1[5] - vadd.s16 q4, q4, q5 ; step2[4] = step1[4] + step1[5] - vsub.s16 q14, q7, q6 ; step2[6] = -step1[6] + step1[7] - vadd.s16 q7, q7, q6 ; step2[7] = step1[6] + step1[7] - - ; step2[6] * cospi_16_64 - vmull.s16 q9, d28, d16 - vmull.s16 q10, d29, d16 - - ; step2[6] * cospi_16_64 - vmull.s16 q11, d28, d16 - vmull.s16 q12, d29, d16 - - ; (step2[6] - step2[5]) * cospi_16_64 - vmlsl.s16 q9, d26, d16 - vmlsl.s16 q10, d27, d16 - - ; (step2[5] + step2[6]) * cospi_16_64 - vmlal.s16 q11, d26, d16 - vmlal.s16 q12, d27, d16 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d10, q9, #14 ; >> 14 - vqrshrn.s32 d11, q10, #14 ; >> 14 - - ; dct_const_round_shift(input_dc * cospi_16_64) - vqrshrn.s32 d12, q11, #14 ; >> 14 - vqrshrn.s32 d13, q12, #14 ; >> 14 - - ; stage 4 - vadd.s16 q8, q0, q7 ; output[0] = step1[0] + step1[7]; - vadd.s16 q9, q1, q6 ; output[1] = step1[1] + step1[6]; - vadd.s16 q10, q2, q5 ; output[2] = step1[2] + step1[5]; - vadd.s16 q11, q3, q4 ; output[3] = step1[3] + step1[4]; - vsub.s16 q12, q3, q4 ; output[4] = step1[3] - step1[4]; - vsub.s16 q13, q2, q5 ; output[5] = step1[2] - step1[5]; - vsub.s16 q14, q1, q6 ; output[6] = step1[1] - step1[6]; - vsub.s16 q15, q0, q7 ; output[7] = step1[0] - step1[7]; - - ; Transpose the matrix - TRANSPOSE8X8 - - ; Then transform columns - IDCT8x8_1D - - ; ROUND_POWER_OF_TWO(temp_out[j], 5) - vrshr.s16 q8, q8, #5 - vrshr.s16 q9, q9, #5 - vrshr.s16 q10, q10, #5 - vrshr.s16 q11, q11, #5 - vrshr.s16 q12, q12, #5 - vrshr.s16 q13, q13, #5 - vrshr.s16 q14, q14, #5 - vrshr.s16 q15, q15, #5 - - ; save dest pointer - mov r0, r1 - - ; load destination data - vld1.64 {d0}, [r1], r2 - vld1.64 {d1}, [r1], r2 - vld1.64 {d2}, [r1], r2 - vld1.64 {d3}, [r1], r2 - vld1.64 {d4}, [r1], r2 - vld1.64 {d5}, [r1], r2 - vld1.64 {d6}, [r1], r2 - vld1.64 {d7}, [r1] - - ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i] - vaddw.u8 q8, q8, d0 - vaddw.u8 q9, q9, d1 - vaddw.u8 q10, q10, d2 - vaddw.u8 q11, q11, d3 - vaddw.u8 q12, q12, d4 - vaddw.u8 q13, q13, d5 - vaddw.u8 q14, q14, d6 - vaddw.u8 q15, q15, d7 - - ; clip_pixel - vqmovun.s16 d0, q8 - vqmovun.s16 d1, q9 - vqmovun.s16 d2, q10 - vqmovun.s16 d3, q11 - vqmovun.s16 d4, q12 - vqmovun.s16 d5, q13 - vqmovun.s16 d6, q14 - vqmovun.s16 d7, q15 - - ; store the data - vst1.64 {d0}, [r0], r2 - vst1.64 {d1}, [r0], r2 - vst1.64 {d2}, [r0], r2 - vst1.64 {d3}, [r0], r2 - vst1.64 {d4}, [r0], r2 - vst1.64 {d5}, [r0], r2 - vst1.64 {d6}, [r0], r2 - vst1.64 {d7}, [r0], r2 - - vpop {d8-d15} - pop {r4-r9} - bx lr - ENDP ; |aom_idct8x8_12_add_neon| - - END diff --git a/aom_dsp/arm/idct8x8_add_neon.c b/aom_dsp/arm/idct8x8_add_neon.c deleted file mode 100644 index 8ad70862d..000000000 --- a/aom_dsp/arm/idct8x8_add_neon.c +++ /dev/null @@ -1,509 +0,0 @@ -/* - * Copyright (c) 2016, 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 "./aom_config.h" -#include "aom_dsp/txfm_common.h" - -static INLINE void TRANSPOSE8X8(int16x8_t *q8s16, int16x8_t *q9s16, - int16x8_t *q10s16, int16x8_t *q11s16, - int16x8_t *q12s16, int16x8_t *q13s16, - int16x8_t *q14s16, int16x8_t *q15s16) { - int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16; - int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16; - int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32; - int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16; - - d16s16 = vget_low_s16(*q8s16); - d17s16 = vget_high_s16(*q8s16); - d18s16 = vget_low_s16(*q9s16); - d19s16 = vget_high_s16(*q9s16); - d20s16 = vget_low_s16(*q10s16); - d21s16 = vget_high_s16(*q10s16); - d22s16 = vget_low_s16(*q11s16); - d23s16 = vget_high_s16(*q11s16); - d24s16 = vget_low_s16(*q12s16); - d25s16 = vget_high_s16(*q12s16); - d26s16 = vget_low_s16(*q13s16); - d27s16 = vget_high_s16(*q13s16); - d28s16 = vget_low_s16(*q14s16); - d29s16 = vget_high_s16(*q14s16); - d30s16 = vget_low_s16(*q15s16); - d31s16 = vget_high_s16(*q15s16); - - *q8s16 = vcombine_s16(d16s16, d24s16); // vswp d17, d24 - *q9s16 = vcombine_s16(d18s16, d26s16); // vswp d19, d26 - *q10s16 = vcombine_s16(d20s16, d28s16); // vswp d21, d28 - *q11s16 = vcombine_s16(d22s16, d30s16); // vswp d23, d30 - *q12s16 = vcombine_s16(d17s16, d25s16); - *q13s16 = vcombine_s16(d19s16, d27s16); - *q14s16 = vcombine_s16(d21s16, d29s16); - *q15s16 = vcombine_s16(d23s16, d31s16); - - q0x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(*q8s16), vreinterpretq_s32_s16(*q10s16)); - q1x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(*q9s16), vreinterpretq_s32_s16(*q11s16)); - q2x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(*q12s16), vreinterpretq_s32_s16(*q14s16)); - q3x2s32 = - vtrnq_s32(vreinterpretq_s32_s16(*q13s16), vreinterpretq_s32_s16(*q15s16)); - - q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]), // q8 - vreinterpretq_s16_s32(q1x2s32.val[0])); // q9 - q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]), // q10 - vreinterpretq_s16_s32(q1x2s32.val[1])); // q11 - q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]), // q12 - vreinterpretq_s16_s32(q3x2s32.val[0])); // q13 - q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]), // q14 - vreinterpretq_s16_s32(q3x2s32.val[1])); // q15 - - *q8s16 = q0x2s16.val[0]; - *q9s16 = q0x2s16.val[1]; - *q10s16 = q1x2s16.val[0]; - *q11s16 = q1x2s16.val[1]; - *q12s16 = q2x2s16.val[0]; - *q13s16 = q2x2s16.val[1]; - *q14s16 = q3x2s16.val[0]; - *q15s16 = q3x2s16.val[1]; - return; -} - -static INLINE void IDCT8x8_1D(int16x8_t *q8s16, int16x8_t *q9s16, - int16x8_t *q10s16, int16x8_t *q11s16, - int16x8_t *q12s16, int16x8_t *q13s16, - int16x8_t *q14s16, int16x8_t *q15s16) { - int16x4_t d0s16, d1s16, d2s16, d3s16; - int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16; - int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16; - int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16; - int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16; - int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32; - int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32; - - d0s16 = vdup_n_s16((int16_t)cospi_28_64); - d1s16 = vdup_n_s16((int16_t)cospi_4_64); - d2s16 = vdup_n_s16((int16_t)cospi_12_64); - d3s16 = vdup_n_s16((int16_t)cospi_20_64); - - d16s16 = vget_low_s16(*q8s16); - d17s16 = vget_high_s16(*q8s16); - d18s16 = vget_low_s16(*q9s16); - d19s16 = vget_high_s16(*q9s16); - d20s16 = vget_low_s16(*q10s16); - d21s16 = vget_high_s16(*q10s16); - d22s16 = vget_low_s16(*q11s16); - d23s16 = vget_high_s16(*q11s16); - d24s16 = vget_low_s16(*q12s16); - d25s16 = vget_high_s16(*q12s16); - d26s16 = vget_low_s16(*q13s16); - d27s16 = vget_high_s16(*q13s16); - d28s16 = vget_low_s16(*q14s16); - d29s16 = vget_high_s16(*q14s16); - d30s16 = vget_low_s16(*q15s16); - d31s16 = vget_high_s16(*q15s16); - - q2s32 = vmull_s16(d18s16, d0s16); - q3s32 = vmull_s16(d19s16, d0s16); - q5s32 = vmull_s16(d26s16, d2s16); - q6s32 = vmull_s16(d27s16, d2s16); - - q2s32 = vmlsl_s16(q2s32, d30s16, d1s16); - q3s32 = vmlsl_s16(q3s32, d31s16, d1s16); - q5s32 = vmlsl_s16(q5s32, d22s16, d3s16); - q6s32 = vmlsl_s16(q6s32, d23s16, d3s16); - - d8s16 = vqrshrn_n_s32(q2s32, 14); - d9s16 = vqrshrn_n_s32(q3s32, 14); - d10s16 = vqrshrn_n_s32(q5s32, 14); - d11s16 = vqrshrn_n_s32(q6s32, 14); - q4s16 = vcombine_s16(d8s16, d9s16); - q5s16 = vcombine_s16(d10s16, d11s16); - - q2s32 = vmull_s16(d18s16, d1s16); - q3s32 = vmull_s16(d19s16, d1s16); - q9s32 = vmull_s16(d26s16, d3s16); - q13s32 = vmull_s16(d27s16, d3s16); - - q2s32 = vmlal_s16(q2s32, d30s16, d0s16); - q3s32 = vmlal_s16(q3s32, d31s16, d0s16); - q9s32 = vmlal_s16(q9s32, d22s16, d2s16); - q13s32 = vmlal_s16(q13s32, d23s16, d2s16); - - d14s16 = vqrshrn_n_s32(q2s32, 14); - d15s16 = vqrshrn_n_s32(q3s32, 14); - d12s16 = vqrshrn_n_s32(q9s32, 14); - d13s16 = vqrshrn_n_s32(q13s32, 14); - q6s16 = vcombine_s16(d12s16, d13s16); - q7s16 = vcombine_s16(d14s16, d15s16); - - d0s16 = vdup_n_s16((int16_t)cospi_16_64); - - q2s32 = vmull_s16(d16s16, d0s16); - q3s32 = vmull_s16(d17s16, d0s16); - q13s32 = vmull_s16(d16s16, d0s16); - q15s32 = vmull_s16(d17s16, d0s16); - - q2s32 = vmlal_s16(q2s32, d24s16, d0s16); - q3s32 = vmlal_s16(q3s32, d25s16, d0s16); - q13s32 = vmlsl_s16(q13s32, d24s16, d0s16); - q15s32 = vmlsl_s16(q15s32, d25s16, d0s16); - - d0s16 = vdup_n_s16((int16_t)cospi_24_64); - d1s16 = vdup_n_s16((int16_t)cospi_8_64); - - d18s16 = vqrshrn_n_s32(q2s32, 14); - d19s16 = vqrshrn_n_s32(q3s32, 14); - d22s16 = vqrshrn_n_s32(q13s32, 14); - d23s16 = vqrshrn_n_s32(q15s32, 14); - *q9s16 = vcombine_s16(d18s16, d19s16); - *q11s16 = vcombine_s16(d22s16, d23s16); - - q2s32 = vmull_s16(d20s16, d0s16); - q3s32 = vmull_s16(d21s16, d0s16); - q8s32 = vmull_s16(d20s16, d1s16); - q12s32 = vmull_s16(d21s16, d1s16); - - q2s32 = vmlsl_s16(q2s32, d28s16, d1s16); - q3s32 = vmlsl_s16(q3s32, d29s16, d1s16); - q8s32 = vmlal_s16(q8s32, d28s16, d0s16); - q12s32 = vmlal_s16(q12s32, d29s16, d0s16); - - d26s16 = vqrshrn_n_s32(q2s32, 14); - d27s16 = vqrshrn_n_s32(q3s32, 14); - d30s16 = vqrshrn_n_s32(q8s32, 14); - d31s16 = vqrshrn_n_s32(q12s32, 14); - *q13s16 = vcombine_s16(d26s16, d27s16); - *q15s16 = vcombine_s16(d30s16, d31s16); - - q0s16 = vaddq_s16(*q9s16, *q15s16); - q1s16 = vaddq_s16(*q11s16, *q13s16); - q2s16 = vsubq_s16(*q11s16, *q13s16); - q3s16 = vsubq_s16(*q9s16, *q15s16); - - *q13s16 = vsubq_s16(q4s16, q5s16); - q4s16 = vaddq_s16(q4s16, q5s16); - *q14s16 = vsubq_s16(q7s16, q6s16); - q7s16 = vaddq_s16(q7s16, q6s16); - d26s16 = vget_low_s16(*q13s16); - d27s16 = vget_high_s16(*q13s16); - d28s16 = vget_low_s16(*q14s16); - d29s16 = vget_high_s16(*q14s16); - - d16s16 = vdup_n_s16((int16_t)cospi_16_64); - - q9s32 = vmull_s16(d28s16, d16s16); - q10s32 = vmull_s16(d29s16, d16s16); - q11s32 = vmull_s16(d28s16, d16s16); - q12s32 = vmull_s16(d29s16, d16s16); - - q9s32 = vmlsl_s16(q9s32, d26s16, d16s16); - q10s32 = vmlsl_s16(q10s32, d27s16, d16s16); - q11s32 = vmlal_s16(q11s32, d26s16, d16s16); - q12s32 = vmlal_s16(q12s32, d27s16, d16s16); - - d10s16 = vqrshrn_n_s32(q9s32, 14); - d11s16 = vqrshrn_n_s32(q10s32, 14); - d12s16 = vqrshrn_n_s32(q11s32, 14); - d13s16 = vqrshrn_n_s32(q12s32, 14); - q5s16 = vcombine_s16(d10s16, d11s16); - q6s16 = vcombine_s16(d12s16, d13s16); - - *q8s16 = vaddq_s16(q0s16, q7s16); - *q9s16 = vaddq_s16(q1s16, q6s16); - *q10s16 = vaddq_s16(q2s16, q5s16); - *q11s16 = vaddq_s16(q3s16, q4s16); - *q12s16 = vsubq_s16(q3s16, q4s16); - *q13s16 = vsubq_s16(q2s16, q5s16); - *q14s16 = vsubq_s16(q1s16, q6s16); - *q15s16 = vsubq_s16(q0s16, q7s16); - return; -} - -void aom_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride) { - uint8_t *d1, *d2; - uint8x8_t d0u8, d1u8, d2u8, d3u8; - uint64x1_t d0u64, d1u64, d2u64, d3u64; - int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16; - uint16x8_t q8u16, q9u16, q10u16, q11u16; - - q8s16 = vld1q_s16(input); - q9s16 = vld1q_s16(input + 8); - q10s16 = vld1q_s16(input + 16); - q11s16 = vld1q_s16(input + 24); - q12s16 = vld1q_s16(input + 32); - q13s16 = vld1q_s16(input + 40); - q14s16 = vld1q_s16(input + 48); - q15s16 = vld1q_s16(input + 56); - - TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - q8s16 = vrshrq_n_s16(q8s16, 5); - q9s16 = vrshrq_n_s16(q9s16, 5); - q10s16 = vrshrq_n_s16(q10s16, 5); - q11s16 = vrshrq_n_s16(q11s16, 5); - q12s16 = vrshrq_n_s16(q12s16, 5); - q13s16 = vrshrq_n_s16(q13s16, 5); - q14s16 = vrshrq_n_s16(q14s16, 5); - q15s16 = vrshrq_n_s16(q15s16, 5); - - d1 = d2 = dest; - - d0u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d1u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d2u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d3u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - - q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64)); - q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64)); - q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64)); - q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64)); - - d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16)); - d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16)); - d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16)); - d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16)); - - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8)); - d2 += dest_stride; - - q8s16 = q12s16; - q9s16 = q13s16; - q10s16 = q14s16; - q11s16 = q15s16; - - d0u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d1u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d2u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d3u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - - q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64)); - q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64)); - q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64)); - q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64)); - - d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16)); - d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16)); - d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16)); - d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16)); - - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8)); - d2 += dest_stride; - return; -} - -void aom_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride) { - uint8_t *d1, *d2; - uint8x8_t d0u8, d1u8, d2u8, d3u8; - int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16; - int16x4_t d26s16, d27s16, d28s16, d29s16; - uint64x1_t d0u64, d1u64, d2u64, d3u64; - int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16; - int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16; - uint16x8_t q8u16, q9u16, q10u16, q11u16; - int32x4_t q9s32, q10s32, q11s32, q12s32; - - q8s16 = vld1q_s16(input); - q9s16 = vld1q_s16(input + 8); - q10s16 = vld1q_s16(input + 16); - q11s16 = vld1q_s16(input + 24); - q12s16 = vld1q_s16(input + 32); - q13s16 = vld1q_s16(input + 40); - q14s16 = vld1q_s16(input + 48); - q15s16 = vld1q_s16(input + 56); - - TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - // First transform rows - // stage 1 - q0s16 = vdupq_n_s16((int16_t)cospi_28_64 * 2); - q1s16 = vdupq_n_s16((int16_t)cospi_4_64 * 2); - - q4s16 = vqrdmulhq_s16(q9s16, q0s16); - - q0s16 = vdupq_n_s16(-(int16_t)cospi_20_64 * 2); - - q7s16 = vqrdmulhq_s16(q9s16, q1s16); - - q1s16 = vdupq_n_s16((int16_t)cospi_12_64 * 2); - - q5s16 = vqrdmulhq_s16(q11s16, q0s16); - - q0s16 = vdupq_n_s16((int16_t)cospi_16_64 * 2); - - q6s16 = vqrdmulhq_s16(q11s16, q1s16); - - // stage 2 & stage 3 - even half - q1s16 = vdupq_n_s16((int16_t)cospi_24_64 * 2); - - q9s16 = vqrdmulhq_s16(q8s16, q0s16); - - q0s16 = vdupq_n_s16((int16_t)cospi_8_64 * 2); - - q13s16 = vqrdmulhq_s16(q10s16, q1s16); - - q15s16 = vqrdmulhq_s16(q10s16, q0s16); - - // stage 3 -odd half - q0s16 = vaddq_s16(q9s16, q15s16); - q1s16 = vaddq_s16(q9s16, q13s16); - q2s16 = vsubq_s16(q9s16, q13s16); - q3s16 = vsubq_s16(q9s16, q15s16); - - // stage 2 - odd half - q13s16 = vsubq_s16(q4s16, q5s16); - q4s16 = vaddq_s16(q4s16, q5s16); - q14s16 = vsubq_s16(q7s16, q6s16); - q7s16 = vaddq_s16(q7s16, q6s16); - d26s16 = vget_low_s16(q13s16); - d27s16 = vget_high_s16(q13s16); - d28s16 = vget_low_s16(q14s16); - d29s16 = vget_high_s16(q14s16); - - d16s16 = vdup_n_s16((int16_t)cospi_16_64); - q9s32 = vmull_s16(d28s16, d16s16); - q10s32 = vmull_s16(d29s16, d16s16); - q11s32 = vmull_s16(d28s16, d16s16); - q12s32 = vmull_s16(d29s16, d16s16); - - q9s32 = vmlsl_s16(q9s32, d26s16, d16s16); - q10s32 = vmlsl_s16(q10s32, d27s16, d16s16); - q11s32 = vmlal_s16(q11s32, d26s16, d16s16); - q12s32 = vmlal_s16(q12s32, d27s16, d16s16); - - d10s16 = vqrshrn_n_s32(q9s32, 14); - d11s16 = vqrshrn_n_s32(q10s32, 14); - d12s16 = vqrshrn_n_s32(q11s32, 14); - d13s16 = vqrshrn_n_s32(q12s32, 14); - q5s16 = vcombine_s16(d10s16, d11s16); - q6s16 = vcombine_s16(d12s16, d13s16); - - // stage 4 - q8s16 = vaddq_s16(q0s16, q7s16); - q9s16 = vaddq_s16(q1s16, q6s16); - q10s16 = vaddq_s16(q2s16, q5s16); - q11s16 = vaddq_s16(q3s16, q4s16); - q12s16 = vsubq_s16(q3s16, q4s16); - q13s16 = vsubq_s16(q2s16, q5s16); - q14s16 = vsubq_s16(q1s16, q6s16); - q15s16 = vsubq_s16(q0s16, q7s16); - - TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16, - &q15s16); - - q8s16 = vrshrq_n_s16(q8s16, 5); - q9s16 = vrshrq_n_s16(q9s16, 5); - q10s16 = vrshrq_n_s16(q10s16, 5); - q11s16 = vrshrq_n_s16(q11s16, 5); - q12s16 = vrshrq_n_s16(q12s16, 5); - q13s16 = vrshrq_n_s16(q13s16, 5); - q14s16 = vrshrq_n_s16(q14s16, 5); - q15s16 = vrshrq_n_s16(q15s16, 5); - - d1 = d2 = dest; - - d0u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d1u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d2u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d3u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - - q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64)); - q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64)); - q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64)); - q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64)); - - d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16)); - d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16)); - d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16)); - d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16)); - - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8)); - d2 += dest_stride; - - q8s16 = q12s16; - q9s16 = q13s16; - q10s16 = q14s16; - q11s16 = q15s16; - - d0u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d1u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d2u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - d3u64 = vld1_u64((uint64_t *)d1); - d1 += dest_stride; - - q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64)); - q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64)); - q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64)); - q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64)); - - d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16)); - d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16)); - d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16)); - d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16)); - - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8)); - d2 += dest_stride; - vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8)); - d2 += dest_stride; - return; -} |