1 /* 2 * Loongson MMI optimizations for libjpeg-turbo 3 * 4 * Copyright (C) 2016-2017, Loongson Technology Corporation Limited, BeiJing. 5 * All Rights Reserved. 6 * Authors: ZhuChen <zhuchen (at) loongson.cn> 7 * CaiWanwei <caiwanwei (at) loongson.cn> 8 * SunZhangzhi <sunzhangzhi-cq (at) loongson.cn> 9 * Copyright (C) 2018, D. R. Commander. All Rights Reserved. 10 * 11 * Based on the x86 SIMD extension for IJG JPEG library 12 * Copyright (C) 1999-2006, MIYASAKA Masaru. 13 * 14 * This software is provided 'as-is', without any express or implied 15 * warranty. In no event will the authors be held liable for any damages 16 * arising from the use of this software. 17 * 18 * Permission is granted to anyone to use this software for any purpose, 19 * including commercial applications, and to alter it and redistribute it 20 * freely, subject to the following restrictions: 21 * 22 * 1. The origin of this software must not be misrepresented; you must not 23 * claim that you wrote the original software. If you use this software 24 * in a product, an acknowledgment in the product documentation would be 25 * appreciated but is not required. 26 * 2. Altered source versions must be plainly marked as such, and must not be 27 * misrepresented as being the original software. 28 * 3. This notice may not be removed or altered from any source distribution. 29 */ 30 31 /* INTEGER QUANTIZATION AND SAMPLE CONVERSION */ 32 33 #include "jsimd_mmi.h" 34 35 36 #define DO_QUANT() { \ 37 mm2 = _mm_load_si64((__m64 *)&workspace[0]); \ 38 mm3 = _mm_load_si64((__m64 *)&workspace[4]); \ 39 \ 40 mm0 = mm2; \ 41 mm1 = mm3; \ 42 \ 43 mm2 = _mm_srai_pi16(mm2, (WORD_BIT - 1)); /* -1 if value < 0, */ \ 44 /* 0 otherwise */ \ 45 mm3 = _mm_srai_pi16(mm3, (WORD_BIT - 1)); \ 46 \ 47 mm0 = _mm_xor_si64(mm0, mm2); /* val = -val */ \ 48 mm1 = _mm_xor_si64(mm1, mm3); \ 49 mm0 = _mm_sub_pi16(mm0, mm2); \ 50 mm1 = _mm_sub_pi16(mm1, mm3); \ 51 \ 52 corr0 = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 1]); /* correction */ \ 53 corr1 = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 1 + 4]); \ 54 \ 55 mm0 = _mm_add_pi16(mm0, corr0); /* correction + roundfactor */ \ 56 mm1 = _mm_add_pi16(mm1, corr1); \ 57 \ 58 mm4 = mm0; \ 59 mm5 = mm1; \ 60 \ 61 recip0 = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 0]); /* reciprocal */ \ 62 recip1 = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 0 + 4]); \ 63 \ 64 mm0 = _mm_mulhi_pi16(mm0, recip0); \ 65 mm1 = _mm_mulhi_pi16(mm1, recip1); \ 66 \ 67 mm0 = _mm_add_pi16(mm0, mm4); /* reciprocal is always negative */ \ 68 mm1 = _mm_add_pi16(mm1, mm5); /* (MSB=1), so we always need to add the */ \ 69 /* initial value (input value is never */ \ 70 /* negative as we inverted it at the */ \ 71 /* start of this routine) */ \ 72 \ 73 scale0 = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 2]); /* scale */ \ 74 scale1 = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 2 + 4]); \ 75 \ 76 mm6 = scale0; \ 77 mm7 = scale1; \ 78 mm4 = mm0; \ 79 mm5 = mm1; \ 80 \ 81 mm0 = _mm_mulhi_pi16(mm0, mm6); \ 82 mm1 = _mm_mulhi_pi16(mm1, mm7); \ 83 \ 84 mm6 = _mm_srai_pi16(mm6, (WORD_BIT - 1)); /* determine if scale... */ \ 85 /* is negative */ \ 86 mm7 = _mm_srai_pi16(mm7, (WORD_BIT - 1)); \ 87 \ 88 mm6 = _mm_and_si64(mm6, mm4); /* and add input if it is */ \ 89 mm7 = _mm_and_si64(mm7, mm5); \ 90 mm0 = _mm_add_pi16(mm0, mm6); \ 91 mm1 = _mm_add_pi16(mm1, mm7); \ 92 \ 93 mm4 = _mm_srai_pi16(mm4, (WORD_BIT - 1)); /* then check if... */ \ 94 mm5 = _mm_srai_pi16(mm5, (WORD_BIT - 1)); /* negative input */ \ 95 \ 96 mm4 = _mm_and_si64(mm4, scale0); /* and add scale if it is */ \ 97 mm5 = _mm_and_si64(mm5, scale1); \ 98 mm0 = _mm_add_pi16(mm0, mm4); \ 99 mm1 = _mm_add_pi16(mm1, mm5); \ 100 \ 101 mm0 = _mm_xor_si64(mm0, mm2); /* val = -val */ \ 102 mm1 = _mm_xor_si64(mm1, mm3); \ 103 mm0 = _mm_sub_pi16(mm0, mm2); \ 104 mm1 = _mm_sub_pi16(mm1, mm3); \ 105 \ 106 _mm_store_si64((__m64 *)&output_ptr[0], mm0); \ 107 _mm_store_si64((__m64 *)&output_ptr[4], mm1); \ 108 \ 109 workspace += DCTSIZE; \ 110 divisors += DCTSIZE; \ 111 output_ptr += DCTSIZE; \ 112 } 113 114 115 void jsimd_quantize_mmi(JCOEFPTR coef_block, DCTELEM *divisors, 116 DCTELEM *workspace) 117 { 118 JCOEFPTR output_ptr = coef_block; 119 __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7; 120 __m64 corr0, corr1, recip0, recip1, scale0, scale1; 121 122 DO_QUANT() 123 DO_QUANT() 124 DO_QUANT() 125 DO_QUANT() 126 DO_QUANT() 127 DO_QUANT() 128 DO_QUANT() 129 DO_QUANT() 130 } 131
