1;****************************************************************************** 2;* x86-SIMD-optimized IDCT for prores 3;* this is identical to "simple" IDCT written by Michael Niedermayer 4;* except for the clip range 5;* 6;* Copyright (c) 2011 Ronald S. Bultje <rsbultje@gmail.com> 7;* Copyright (c) 2015 Christophe Gisquet 8;* 9;* This file is part of FFmpeg. 10;* 11;* FFmpeg is free software; you can redistribute it and/or 12;* modify it under the terms of the GNU Lesser General Public 13;* License as published by the Free Software Foundation; either 14;* version 2.1 of the License, or (at your option) any later version. 15;* 16;* FFmpeg is distributed in the hope that it will be useful, 17;* but WITHOUT ANY WARRANTY; without even the implied warranty of 18;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 19;* Lesser General Public License for more details. 20;* 21;* You should have received a copy of the GNU Lesser General Public 22;* License along with FFmpeg; if not, write to the Free Software 23;* 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 24;****************************************************************************** 25 26%include "libavutil/x86/x86util.asm" 27 28%if ARCH_X86_64 29 30SECTION_RODATA 31 32cextern pw_2 33cextern pw_16 34cextern pw_32 35cextern pw_1023 36cextern pw_4095 37pd_round_11: times 4 dd 1<<(11-1) 38pd_round_12: times 4 dd 1<<(12-1) 39pd_round_15: times 4 dd 1<<(15-1) 40pd_round_19: times 4 dd 1<<(19-1) 41pd_round_20: times 4 dd 1<<(20-1) 42 43%macro CONST_DEC 3 44const %1 45times 4 dw %2, %3 46%endmacro 47 48%define W1sh2 22725 ; W1 = 90901 = 22725<<2 + 1 49%define W2sh2 21407 ; W2 = 85627 = 21407<<2 - 1 50%define W3sh2 19265 ; W3 = 77062 = 19265<<2 + 2 51%define W4sh2 16384 ; W4 = 65535 = 16384<<2 - 1 52%define W3sh2_lo 19266 53%define W4sh2_lo 16383 54%define W5sh2 12873 ; W5 = 51491 = 12873<<2 - 1 55%define W6sh2 8867 ; W6 = 35468 = 8867<<2 56%define W7sh2 4520 ; W7 = 18081 = 4520<<2 + 1 57 58CONST_DEC w4_plus_w2_hi, W4sh2, +W2sh2 59CONST_DEC w4_min_w2_hi, W4sh2, -W2sh2 60CONST_DEC w4_plus_w6_hi, W4sh2, +W6sh2 61CONST_DEC w4_min_w6_hi, W4sh2, -W6sh2 62CONST_DEC w1_plus_w3_hi, W1sh2, +W3sh2 63CONST_DEC w3_min_w1_hi, W3sh2, -W1sh2 64CONST_DEC w7_plus_w3_hi, W7sh2, +W3sh2 65CONST_DEC w3_min_w7_hi, W3sh2, -W7sh2 66CONST_DEC w1_plus_w5, W1sh2, +W5sh2 67CONST_DEC w5_min_w1, W5sh2, -W1sh2 68CONST_DEC w5_plus_w7, W5sh2, +W7sh2 69CONST_DEC w7_min_w5, W7sh2, -W5sh2 70CONST_DEC w4_plus_w2_lo, W4sh2_lo, +W2sh2 71CONST_DEC w4_min_w2_lo, W4sh2_lo, -W2sh2 72CONST_DEC w4_plus_w6_lo, W4sh2_lo, +W6sh2 73CONST_DEC w4_min_w6_lo, W4sh2_lo, -W6sh2 74CONST_DEC w1_plus_w3_lo, W1sh2, +W3sh2_lo 75CONST_DEC w3_min_w1_lo, W3sh2_lo, -W1sh2 76CONST_DEC w7_plus_w3_lo, W7sh2, +W3sh2_lo 77CONST_DEC w3_min_w7_lo, W3sh2_lo, -W7sh2 78 79%include "libavcodec/x86/simple_idct10_template.asm" 80 81SECTION .text 82 83%macro STORE_HI_LO 12 84 movq %1, %9 85 movq %3, %10 86 movq %5, %11 87 movq %7, %12 88 movhps %2, %9 89 movhps %4, %10 90 movhps %6, %11 91 movhps %8, %12 92%endmacro 93 94%macro LOAD_ZXBW_8 16 95 pmovzxbw %1, %9 96 pmovzxbw %2, %10 97 pmovzxbw %3, %11 98 pmovzxbw %4, %12 99 pmovzxbw %5, %13 100 pmovzxbw %6, %14 101 pmovzxbw %7, %15 102 pmovzxbw %8, %16 103%endmacro 104 105%macro LOAD_ZXBW_4 9 106 movh %1, %5 107 movh %2, %6 108 movh %3, %7 109 movh %4, %8 110 punpcklbw %1, %9 111 punpcklbw %2, %9 112 punpcklbw %3, %9 113 punpcklbw %4, %9 114%endmacro 115 116%define PASS4ROWS(base, stride, stride3) \ 117 [base], [base + stride], [base + 2*stride], [base + stride3] 118 119%macro idct_fn 0 120 121define_constants _lo 122 123cglobal simple_idct8, 1, 1, 16, 32, block 124 IDCT_FN "", 11, pw_32, 20, "store" 125RET 126 127cglobal simple_idct8_put, 3, 4, 16, 32, pixels, lsize, block 128 IDCT_FN "", 11, pw_32, 20 129 lea r3, [3*lsizeq] 130 lea r2, [pixelsq + r3] 131 packuswb m8, m0 132 packuswb m1, m2 133 packuswb m4, m11 134 packuswb m9, m10 135 STORE_HI_LO PASS8ROWS(pixelsq, r2, lsizeq, r3), m8, m1, m4, m9 136RET 137 138cglobal simple_idct8_add, 3, 4, 16, 32, pixels, lsize, block 139 IDCT_FN "", 11, pw_32, 20 140 lea r2, [3*lsizeq] 141 %if cpuflag(sse4) 142 lea r3, [pixelsq + r2] 143 LOAD_ZXBW_8 m3, m5, m6, m7, m12, m13, m14, m15, PASS8ROWS(pixelsq, r3, lsizeq, r2) 144 paddsw m8, m3 145 paddsw m0, m5 146 paddsw m1, m6 147 paddsw m2, m7 148 paddsw m4, m12 149 paddsw m11, m13 150 paddsw m9, m14 151 paddsw m10, m15 152 %else 153 pxor m12, m12 154 LOAD_ZXBW_4 m3, m5, m6, m7, PASS4ROWS(pixelsq, lsizeq, r2), m12 155 paddsw m8, m3 156 paddsw m0, m5 157 paddsw m1, m6 158 paddsw m2, m7 159 lea r3, [pixelsq + 4*lsizeq] 160 LOAD_ZXBW_4 m3, m5, m6, m7, PASS4ROWS(r3, lsizeq, r2), m12 161 paddsw m4, m3 162 paddsw m11, m5 163 paddsw m9, m6 164 paddsw m10, m7 165 lea r3, [pixelsq + r2] 166 %endif 167 packuswb m8, m0 168 packuswb m1, m2 169 packuswb m4, m11 170 packuswb m9, m10 171 STORE_HI_LO PASS8ROWS(pixelsq, r3, lsizeq, r2), m8, m1, m4, m9 172RET 173 174define_constants _hi 175 176cglobal simple_idct10, 1, 1, 16, block 177 IDCT_FN "", 12, "", 19, "store" 178 RET 179 180cglobal simple_idct10_put, 3, 3, 16, pixels, lsize, block 181 IDCT_FN "", 12, "", 19, "put", 0, pw_1023 182 RET 183 184cglobal simple_idct12, 1, 1, 16, block 185 ; coeffs are already 15bits, adding the offset would cause 186 ; overflow in the input 187 IDCT_FN "", 15, pw_2, 16, "store" 188 RET 189 190cglobal simple_idct12_put, 3, 3, 16, pixels, lsize, block 191 ; range isn't known, so the C simple_idct range is used 192 ; Also, using a bias on input overflows, so use the bias 193 ; on output of the first butterfly instead 194 IDCT_FN "", 15, pw_2, 16, "put", 0, pw_4095 195 RET 196%endmacro 197 198INIT_XMM sse2 199idct_fn 200%if HAVE_AVX_EXTERNAL 201INIT_XMM avx 202idct_fn 203%endif 204 205%endif 206