1ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org/* 2ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * Copyright (c) 2013 The WebM project authors. All Rights Reserved. 3ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * 4ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * Use of this source code is governed by a BSD-style license 5ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * that can be found in the LICENSE file in the root of the source 6ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * tree. An additional intellectual property rights grant can be found 7ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * in the file PATENTS. All contributing project authors may 8ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * be found in the AUTHORS file in the root of the source tree. 9ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 10ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 11ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#include <assert.h> 12ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#include <stdio.h> 13ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 14ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#include "./vpx_config.h" 15ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#include "./vp9_rtcd.h" 16ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#include "vp9/common/vp9_common.h" 17ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#include "vp9/common/vp9_blockd.h" 18ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#include "vp9/common/vp9_idct.h" 19ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#include "vp9/common/mips/dspr2/vp9_common_dspr2.h" 20ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 21ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#if HAVE_DSPR2 2276e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.orgstatic void vp9_idct4_rows_dspr2(const int16_t *input, int16_t *output) { 23ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int16_t step_0, step_1, step_2, step_3; 24ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int Temp0, Temp1, Temp2, Temp3; 25ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org const int const_2_power_13 = 8192; 26ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int i; 27ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 28ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (i = 4; i--; ) { 29ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 30ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* 31ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp_1 = (input[0] + input[2]) * cospi_16_64; 32ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org step_0 = dct_const_round_shift(temp_1); 33ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 34ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp_2 = (input[0] - input[2]) * cospi_16_64; 35ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org step_1 = dct_const_round_shift(temp_2); 36ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 37ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lh %[Temp0], 0(%[input]) \n\t" 38ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lh %[Temp1], 4(%[input]) \n\t" 39ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mtlo %[const_2_power_13], $ac0 \n\t" 40ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mthi $zero, $ac0 \n\t" 41ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mtlo %[const_2_power_13], $ac1 \n\t" 42ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mthi $zero, $ac1 \n\t" 43ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp2], %[Temp0], %[Temp1] \n\t" 44ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sub %[Temp3], %[Temp0], %[Temp1] \n\t" 45ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac0, %[Temp2], %[cospi_16_64] \n\t" 46ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lh %[Temp0], 2(%[input]) \n\t" 47ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lh %[Temp1], 6(%[input]) \n\t" 48ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "extp %[step_0], $ac0, 31 \n\t" 49ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mtlo %[const_2_power_13], $ac0 \n\t" 50ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mthi $zero, $ac0 \n\t" 51ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 52ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac1, %[Temp3], %[cospi_16_64] \n\t" 53ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "extp %[step_1], $ac1, 31 \n\t" 54ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mtlo %[const_2_power_13], $ac1 \n\t" 55ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mthi $zero, $ac1 \n\t" 56ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 57ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* 58ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64; 59ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org step_2 = dct_const_round_shift(temp1); 60ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 61ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac0, %[Temp0], %[cospi_24_64] \n\t" 62ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "msub $ac0, %[Temp1], %[cospi_8_64] \n\t" 63ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "extp %[step_2], $ac0, 31 \n\t" 64ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 65ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* 66ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64; 67ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org step_3 = dct_const_round_shift(temp2); 68ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 69ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac1, %[Temp0], %[cospi_8_64] \n\t" 70ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac1, %[Temp1], %[cospi_24_64] \n\t" 71ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "extp %[step_3], $ac1, 31 \n\t" 72ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 73ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* 74ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[0] = step_0 + step_3; 75ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[4] = step_1 + step_2; 76ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[8] = step_1 - step_2; 77ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[12] = step_0 - step_3; 78ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 79ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp0], %[step_0], %[step_3] \n\t" 80ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sh %[Temp0], 0(%[output]) \n\t" 81ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 82ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp1], %[step_1], %[step_2] \n\t" 83ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sh %[Temp1], 8(%[output]) \n\t" 84ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 85ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sub %[Temp2], %[step_1], %[step_2] \n\t" 86ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sh %[Temp2], 16(%[output]) \n\t" 87ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 88ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sub %[Temp3], %[step_0], %[step_3] \n\t" 89ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sh %[Temp3], 24(%[output]) \n\t" 90ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 91ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [Temp0] "=&r" (Temp0), [Temp1] "=&r" (Temp1), 92ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3), 93ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [step_0] "=&r" (step_0), [step_1] "=&r" (step_1), 94ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [step_2] "=&r" (step_2), [step_3] "=&r" (step_3), 95ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [output] "+r" (output) 96ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [const_2_power_13] "r" (const_2_power_13), 97ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [cospi_8_64] "r" (cospi_8_64), [cospi_16_64] "r" (cospi_16_64), 98ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [cospi_24_64] "r" (cospi_24_64), 99ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [input] "r" (input) 100ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 101ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 102ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org input += 4; 103ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output += 1; 104ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 105ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org} 106ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 10776e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.orgstatic void vp9_idct4_columns_add_blk_dspr2(int16_t *input, uint8_t *dest, 108ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int dest_stride) { 109ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int16_t step_0, step_1, step_2, step_3; 110ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int Temp0, Temp1, Temp2, Temp3; 111ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org const int const_2_power_13 = 8192; 112ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int i; 113ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org uint8_t *dest_pix; 114ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org uint8_t *cm = vp9_ff_cropTbl; 115ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 116ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* prefetch vp9_ff_cropTbl */ 117ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org vp9_prefetch_load(vp9_ff_cropTbl); 118ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org vp9_prefetch_load(vp9_ff_cropTbl + 32); 119ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org vp9_prefetch_load(vp9_ff_cropTbl + 64); 120ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org vp9_prefetch_load(vp9_ff_cropTbl + 96); 121ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org vp9_prefetch_load(vp9_ff_cropTbl + 128); 122ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org vp9_prefetch_load(vp9_ff_cropTbl + 160); 123ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org vp9_prefetch_load(vp9_ff_cropTbl + 192); 124ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org vp9_prefetch_load(vp9_ff_cropTbl + 224); 125ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 126ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (i = 0; i < 4; ++i) { 127ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org dest_pix = (dest + i); 128ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 129ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 130ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* 131ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp_1 = (input[0] + input[2]) * cospi_16_64; 132ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org step_0 = dct_const_round_shift(temp_1); 133ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 134ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp_2 = (input[0] - input[2]) * cospi_16_64; 135ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org step_1 = dct_const_round_shift(temp_2); 136ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 137ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lh %[Temp0], 0(%[input]) \n\t" 138ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lh %[Temp1], 4(%[input]) \n\t" 139ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mtlo %[const_2_power_13], $ac0 \n\t" 140ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mthi $zero, $ac0 \n\t" 141ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mtlo %[const_2_power_13], $ac1 \n\t" 142ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mthi $zero, $ac1 \n\t" 143ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp2], %[Temp0], %[Temp1] \n\t" 144ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sub %[Temp3], %[Temp0], %[Temp1] \n\t" 145ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac0, %[Temp2], %[cospi_16_64] \n\t" 146ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lh %[Temp0], 2(%[input]) \n\t" 147ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lh %[Temp1], 6(%[input]) \n\t" 148ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "extp %[step_0], $ac0, 31 \n\t" 149ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mtlo %[const_2_power_13], $ac0 \n\t" 150ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mthi $zero, $ac0 \n\t" 151ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 152ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac1, %[Temp3], %[cospi_16_64] \n\t" 153ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "extp %[step_1], $ac1, 31 \n\t" 154ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mtlo %[const_2_power_13], $ac1 \n\t" 155ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "mthi $zero, $ac1 \n\t" 156ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 157ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* 158ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64; 159ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org step_2 = dct_const_round_shift(temp1); 160ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 161ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac0, %[Temp0], %[cospi_24_64] \n\t" 162ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "msub $ac0, %[Temp1], %[cospi_8_64] \n\t" 163ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "extp %[step_2], $ac0, 31 \n\t" 164ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 165ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* 166ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64; 167ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org step_3 = dct_const_round_shift(temp2); 168ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 169ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac1, %[Temp0], %[cospi_8_64] \n\t" 170ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "madd $ac1, %[Temp1], %[cospi_24_64] \n\t" 171ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "extp %[step_3], $ac1, 31 \n\t" 172ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 173ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* 174ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[0] = step_0 + step_3; 175ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[4] = step_1 + step_2; 176ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[8] = step_1 - step_2; 177ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[12] = step_0 - step_3; 178ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org */ 179ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp0], %[step_0], %[step_3] \n\t" 180ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addi %[Temp0], %[Temp0], 8 \n\t" 181ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sra %[Temp0], %[Temp0], 4 \n\t" 182ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lbu %[Temp1], 0(%[dest_pix]) \n\t" 183ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp1], %[Temp1], %[Temp0] \n\t" 184ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp0], %[step_1], %[step_2] \n\t" 185ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lbux %[Temp2], %[Temp1](%[cm]) \n\t" 186ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sb %[Temp2], 0(%[dest_pix]) \n\t" 187ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addu %[dest_pix], %[dest_pix], %[dest_stride] \n\t" 188ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 189ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addi %[Temp0], %[Temp0], 8 \n\t" 190ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sra %[Temp0], %[Temp0], 4 \n\t" 191ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lbu %[Temp1], 0(%[dest_pix]) \n\t" 192ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp1], %[Temp1], %[Temp0] \n\t" 193ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sub %[Temp0], %[step_1], %[step_2] \n\t" 194ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lbux %[Temp2], %[Temp1](%[cm]) \n\t" 195ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sb %[Temp2], 0(%[dest_pix]) \n\t" 196ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addu %[dest_pix], %[dest_pix], %[dest_stride] \n\t" 197ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 198ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addi %[Temp0], %[Temp0], 8 \n\t" 199ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sra %[Temp0], %[Temp0], 4 \n\t" 200ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lbu %[Temp1], 0(%[dest_pix]) \n\t" 201ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp1], %[Temp1], %[Temp0] \n\t" 202ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sub %[Temp0], %[step_0], %[step_3] \n\t" 203ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lbux %[Temp2], %[Temp1](%[cm]) \n\t" 204ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sb %[Temp2], 0(%[dest_pix]) \n\t" 205ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addu %[dest_pix], %[dest_pix], %[dest_stride] \n\t" 206ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 207ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addi %[Temp0], %[Temp0], 8 \n\t" 208ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sra %[Temp0], %[Temp0], 4 \n\t" 209ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lbu %[Temp1], 0(%[dest_pix]) \n\t" 210ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[Temp1], %[Temp1], %[Temp0] \n\t" 211ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lbux %[Temp2], %[Temp1](%[cm]) \n\t" 212ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sb %[Temp2], 0(%[dest_pix]) \n\t" 213ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 214ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [Temp0] "=&r" (Temp0), [Temp1] "=&r" (Temp1), 215ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3), 216ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [step_0] "=&r" (step_0), [step_1] "=&r" (step_1), 217ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [step_2] "=&r" (step_2), [step_3] "=&r" (step_3), 218ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [dest_pix] "+r" (dest_pix) 219ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [const_2_power_13] "r" (const_2_power_13), 220ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [cospi_8_64] "r" (cospi_8_64), [cospi_16_64] "r" (cospi_16_64), 221ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [cospi_24_64] "r" (cospi_24_64), 222ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [input] "r" (input), [cm] "r" (cm), [dest_stride] "r" (dest_stride) 223ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 224ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 225ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org input += 4; 226ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 227ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org} 228ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 229ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.orgvoid vp9_idct4x4_16_add_dspr2(const int16_t *input, uint8_t *dest, 230ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int dest_stride) { 231ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org DECLARE_ALIGNED(32, int16_t, out[4 * 4]); 232ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int16_t *outptr = out; 233ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org uint32_t pos = 45; 234ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 235ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* bit positon for extract from acc */ 236ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 237ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "wrdsp %[pos], 1 \n\t" 238ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : 239ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [pos] "r" (pos) 240ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 241ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 242ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org // Rows 24376e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org vp9_idct4_rows_dspr2(input, outptr); 244ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 245ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org // Columns 24676e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org vp9_idct4_columns_add_blk_dspr2(&out[0], dest, dest_stride); 247ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org} 248ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 249ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.orgvoid vp9_idct4x4_1_add_dspr2(const int16_t *input, uint8_t *dest, 250ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int dest_stride) { 251ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int a1, absa1; 252ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int r; 253ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int32_t out; 254ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int t2, vector_a1, vector_a; 255ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org uint32_t pos = 45; 256ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int16_t input_dc = input[0]; 257ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 258ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* bit positon for extract from acc */ 259ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 260ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "wrdsp %[pos], 1 \n\t" 261ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 262ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : 263ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [pos] "r" (pos) 264ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 265ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 266ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input_dc); 267ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 268ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addi %[out], %[out], 8 \n\t" 269ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sra %[a1], %[out], 4 \n\t" 270ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 271ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [out] "+r" (out), [a1] "=r" (a1) 272ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : 273ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 274ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 275ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org if (a1 < 0) { 276ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* use quad-byte 277ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * input and output memory are four byte aligned */ 278ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 279ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "abs %[absa1], %[a1] \n\t" 280ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "replv.qb %[vector_a1], %[absa1] \n\t" 281ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 282ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [absa1] "=r" (absa1), [vector_a1] "=r" (vector_a1) 283ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [a1] "r" (a1) 284ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 285ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 286ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (r = 4; r--;) { 287ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 288ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lw %[t2], 0(%[dest]) \n\t" 289ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "subu_s.qb %[vector_a], %[t2], %[vector_a1] \n\t" 290ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sw %[vector_a], 0(%[dest]) \n\t" 291ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[dest], %[dest], %[dest_stride] \n\t" 292ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 293ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [t2] "=&r" (t2), [vector_a] "=&r" (vector_a), 294ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [dest] "+&r" (dest) 295ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [dest_stride] "r" (dest_stride), [vector_a1] "r" (vector_a1) 296ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 297ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 298ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } else { 299ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* use quad-byte 300ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org * input and output memory are four byte aligned */ 301ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 302ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "replv.qb %[vector_a1], %[a1] \n\t" 303ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [vector_a1] "=r" (vector_a1) 304ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [a1] "r" (a1) 305ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 306ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 307ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (r = 4; r--;) { 308ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 309ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "lw %[t2], 0(%[dest]) \n\t" 310ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "addu_s.qb %[vector_a], %[t2], %[vector_a1] \n\t" 311ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "sw %[vector_a], 0(%[dest]) \n\t" 312ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "add %[dest], %[dest], %[dest_stride] \n\t" 313ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 314ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [t2] "=&r" (t2), [vector_a] "=&r" (vector_a), 315ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org [dest] "+&r" (dest) 316ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [dest_stride] "r" (dest_stride), [vector_a1] "r" (vector_a1) 317ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 318ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 319ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 320ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org} 321ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 32276e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.orgstatic void iadst4_dspr2(const int16_t *input, int16_t *output) { 323ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int s0, s1, s2, s3, s4, s5, s6, s7; 324ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int x0, x1, x2, x3; 325ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 326ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org x0 = input[0]; 327ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org x1 = input[1]; 328ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org x2 = input[2]; 329ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org x3 = input[3]; 330ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 331ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org if (!(x0 | x1 | x2 | x3)) { 332ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[0] = output[1] = output[2] = output[3] = 0; 333ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org return; 334ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 335ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 336ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s0 = sinpi_1_9 * x0; 337ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s1 = sinpi_2_9 * x0; 338ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s2 = sinpi_3_9 * x1; 339ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s3 = sinpi_4_9 * x2; 340ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s4 = sinpi_1_9 * x2; 341ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s5 = sinpi_2_9 * x3; 342ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s6 = sinpi_4_9 * x3; 343ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s7 = x0 - x2 + x3; 344ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 345ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org x0 = s0 + s3 + s5; 346ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org x1 = s1 - s4 - s6; 347ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org x2 = sinpi_3_9 * s7; 348ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org x3 = s2; 349ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 350ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s0 = x0 + x3; 351ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s1 = x1 + x3; 352ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s2 = x2; 353ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org s3 = x0 + x1 - x3; 354ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 355ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org // 1-D transform scaling factor is sqrt(2). 356ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org // The overall dynamic range is 14b (input) + 14b (multiplication scaling) 357ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org // + 1b (addition) = 29b. 358ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org // Hence the output bit depth is 15b. 359ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[0] = dct_const_round_shift(s0); 360ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[1] = dct_const_round_shift(s1); 361ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[2] = dct_const_round_shift(s2); 362ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org output[3] = dct_const_round_shift(s3); 363ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org} 364ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 365ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.orgvoid vp9_iht4x4_16_add_dspr2(const int16_t *input, uint8_t *dest, 366ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int dest_stride, int tx_type) { 367ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int i, j; 368ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org DECLARE_ALIGNED(32, int16_t, out[4 * 4]); 369ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int16_t *outptr = out; 370ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org int16_t temp_in[4 * 4], temp_out[4]; 371ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org uint32_t pos = 45; 372ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 373ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org /* bit positon for extract from acc */ 374ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org __asm__ __volatile__ ( 375ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org "wrdsp %[pos], 1 \n\t" 376ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : 377ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org : [pos] "r" (pos) 378ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org ); 379ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 380ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org switch (tx_type) { 381ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org case DCT_DCT: // DCT in both horizontal and vertical 38276e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org vp9_idct4_rows_dspr2(input, outptr); 38376e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org vp9_idct4_columns_add_blk_dspr2(&out[0], dest, dest_stride); 384ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org break; 385ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org case ADST_DCT: // ADST in vertical, DCT in horizontal 38676e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org vp9_idct4_rows_dspr2(input, outptr); 387ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 388ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org outptr = out; 389ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 390ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (i = 0; i < 4; ++i) { 39176e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org iadst4_dspr2(outptr, temp_out); 392ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 393ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (j = 0; j < 4; ++j) 394ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org dest[j * dest_stride + i] = 395ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 4) 396ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org + dest[j * dest_stride + i]); 397ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 398ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org outptr += 4; 399ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 400ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org break; 401ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org case DCT_ADST: // DCT in vertical, ADST in horizontal 402ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (i = 0; i < 4; ++i) { 40376e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org iadst4_dspr2(input, outptr); 404ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org input += 4; 405ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org outptr += 4; 406ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 407ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 408ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (i = 0; i < 4; ++i) { 409ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (j = 0; j < 4; ++j) { 410ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp_in[i * 4 + j] = out[j * 4 + i]; 411ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 412ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 41376e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org vp9_idct4_columns_add_blk_dspr2(&temp_in[0], dest, dest_stride); 414ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org break; 415ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org case ADST_ADST: // ADST in both directions 416ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (i = 0; i < 4; ++i) { 41776e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org iadst4_dspr2(input, outptr); 418ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org input += 4; 419ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org outptr += 4; 420ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 421ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 422ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (i = 0; i < 4; ++i) { 423ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (j = 0; j < 4; ++j) 424ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org temp_in[j] = out[j * 4 + i]; 42576e516e2154f353aa02c504bac88afb0f95fefa7johannkoenig@chromium.org iadst4_dspr2(temp_in, temp_out); 426ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org 427ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org for (j = 0; j < 4; ++j) 428ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org dest[j * dest_stride + i] = 429ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 4) 430ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org + dest[j * dest_stride + i]); 431ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 432ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org break; 433ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org default: 434ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org printf("vp9_short_iht4x4_add_dspr2 : Invalid tx_type\n"); 435ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org break; 436ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org } 437ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org} 438ecee051929d6ced19cf324688774acccc9ad4a0ajohannkoenig@chromium.org#endif // #if HAVE_DSPR2 439