Quelle  jquanti-mmi.c

/*
 * Loongson MMI optimizations for libjpeg-turbo
 *
 * Copyright (C) 2016-2017, Loongson Technology Corporation Limited, BeiJing.
 *                          All Rights Reserved.
 * Authors:  ZhuChen     <zhuchen@loongson.cn>
 *           CaiWanwei   <caiwanwei@loongson.cn>
 *           SunZhangzhi <sunzhangzhi-cq@loongson.cn>
 * Copyright (C) 2018-2019, 2025, D. R. Commander.  All Rights Reserved.
 *
 * Based on the x86 SIMD extension for IJG JPEG library
 * Copyright (C) 1999-2006, MIYASAKA Masaru.
 *
 * This software is provided 'as-is', without any express or implied
 * warranty.  In no event will the authors be held liable for any damages
 * arising from the use of this software.
 *
 * Permission is granted to anyone to use this software for any purpose,
 * including commercial applications, and to alter it and redistribute it
 * freely, subject to the following restrictions:
 *
 * 1. The origin of this software must not be misrepresented; you must not
 *    claim that you wrote the original software. If you use this software
 *    in a product, an acknowledgment in the product documentation would be
 *    appreciated but is not required.
 * 2. Altered source versions must be plainly marked as such, and must not be
 *    misrepresented as being the original software.
 * 3. This notice may not be removed or altered from any source distribution.
 */

/* INTEGER QUANTIZATION AND SAMPLE CONVERSION */

#include "jsimd_mmi.h"


#define DO_QUANT() { \
  __m64 rowl, rowh, rowls, rowhs, rowlsave, rowhsave; \
  __m64 corrl, corrh, recipl, reciph, scalel, scaleh; \
  \
  rowl = _mm_load_si64((__m64 *)&workspace[0]); \
  rowh = _mm_load_si64((__m64 *)&workspace[4]); \
  \
  /* Branch-less absolute value */ \
  rowls = _mm_srai_pi16(rowl, (WORD_BIT - 1));  /* -1 if value < 0, */ \
                                                /* 0 otherwise */ \
  rowhs = _mm_srai_pi16(rowh, (WORD_BIT - 1)); \
  \
  rowl = _mm_xor_si64(rowl, rowls);           /* val = -val */ \
  rowh = _mm_xor_si64(rowh, rowhs); \
  rowl = _mm_sub_pi16(rowl, rowls); \
  rowh = _mm_sub_pi16(rowh, rowhs); \
  \
  corrl = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 1]);  /* correction */ \
  corrh = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 1 + 4]); \
  \
  rowlsave = rowl = _mm_add_pi16(rowl, corrl);  /* correction + roundfactor */ \
  rowhsave = rowh = _mm_add_pi16(rowh, corrh); \
  \
  recipl = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 0]);  /* reciprocal */ \
  reciph = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 0 + 4]); \
  \
  rowl = _mm_mulhi_pi16(rowl, recipl); \
  rowh = _mm_mulhi_pi16(rowh, reciph); \
  \
  /* reciprocal is always negative (MSB = 1), so we always need to add the \
   * initial value.  (The input value is never negative, as we inverted it at \
   * the start of this routine.) \
   */ \
  rowlsave = rowl = _mm_add_pi16(rowl, rowlsave); \
  rowhsave = rowh = _mm_add_pi16(rowh, rowhsave); \
  \
  scalel = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 2]);  /* scale */ \
  scaleh = _mm_load_si64((__m64 *)&divisors[DCTSIZE2 * 2 + 4]); \
  \
  rowl = _mm_mulhi_pi16(rowl, scalel); \
  rowh = _mm_mulhi_pi16(rowh, scaleh); \
  \
  /* determine if scale is negative */ \
  scalel = _mm_srai_pi16(scalel, (WORD_BIT - 1)); \
  scaleh = _mm_srai_pi16(scaleh, (WORD_BIT - 1)); \
  \
  /* and add input if it is */ \
  scalel = _mm_and_si64(scalel, rowlsave); \
  scaleh = _mm_and_si64(scaleh, rowhsave); \
  rowl = _mm_add_pi16(rowl, scalel); \
  rowh = _mm_add_pi16(rowh, scaleh); \
  \
  /* then check if negative input */ \
  rowlsave = _mm_srai_pi16(rowlsave, (WORD_BIT - 1)); \
  rowhsave = _mm_srai_pi16(rowhsave, (WORD_BIT - 1)); \
  \
  /* and add scale if it is */ \
  rowlsave = _mm_and_si64(rowlsave, scalel); \
  rowhsave = _mm_and_si64(rowhsave, scaleh); \
  rowl = _mm_add_pi16(rowl, rowlsave); \
  rowh = _mm_add_pi16(rowh, rowhsave); \
  \
  rowl = _mm_xor_si64(rowl, rowls);           /* val = -val */ \
  rowh = _mm_xor_si64(rowh, rowhs); \
  rowl = _mm_sub_pi16(rowl, rowls); \
  rowh = _mm_sub_pi16(rowh, rowhs); \
  \
  _mm_store_si64((__m64 *)&output_ptr[0], rowl); \
  _mm_store_si64((__m64 *)&output_ptr[4], rowh); \
  \
  workspace += DCTSIZE; \
  divisors += DCTSIZE; \
  output_ptr += DCTSIZE; \
}


void jsimd_quantize_mmi(JCOEFPTR coef_block, DCTELEM *divisors,
                        DCTELEM *workspace)
{
  JCOEFPTR output_ptr = coef_block;

  DO_QUANT()
  DO_QUANT()
  DO_QUANT()
  DO_QUANT()
  DO_QUANT()
  DO_QUANT()
  DO_QUANT()
  DO_QUANT()
}