|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; jdmerge.asm - merged upsampling/color conversion (64-bit AVX2)
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB</ossman@cendio.se>
|
|
shun-iwasawa |
82a8f5 |
; Copyright (C) 2009, 2016, D. R. Commander.
|
|
shun-iwasawa |
82a8f5 |
; Copyright (C) 2015, Intel Corporation.
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; Based on the x86 SIMD extension for IJG JPEG library
|
|
shun-iwasawa |
82a8f5 |
; Copyright (C) 1999-2006, MIYASAKA Masaru.
|
|
shun-iwasawa |
82a8f5 |
; For conditions of distribution and use, see copyright notice in jsimdext.inc
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; This file should be assembled with NASM (Netwide Assembler),
|
|
shun-iwasawa |
82a8f5 |
; can *not* be assembled with Microsoft's MASM or any compatible
|
|
shun-iwasawa |
82a8f5 |
; assembler (including Borland's Turbo Assembler).
|
|
shun-iwasawa |
82a8f5 |
; NASM is available from http://nasm.sourceforge.net/ or
|
|
shun-iwasawa |
82a8f5 |
; http://sourceforge.net/project/showfiles.php?group_id=6208
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%include "jsimdext.inc"
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
; --------------------------------------------------------------------------
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%define SCALEBITS 16
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
F_0_344 equ 22554 ; FIX(0.34414)
|
|
shun-iwasawa |
82a8f5 |
F_0_714 equ 46802 ; FIX(0.71414)
|
|
shun-iwasawa |
82a8f5 |
F_1_402 equ 91881 ; FIX(1.40200)
|
|
shun-iwasawa |
82a8f5 |
F_1_772 equ 116130 ; FIX(1.77200)
|
|
shun-iwasawa |
82a8f5 |
F_0_402 equ (F_1_402 - 65536) ; FIX(1.40200) - FIX(1)
|
|
shun-iwasawa |
82a8f5 |
F_0_285 equ ( 65536 - F_0_714) ; FIX(1) - FIX(0.71414)
|
|
shun-iwasawa |
82a8f5 |
F_0_228 equ (131072 - F_1_772) ; FIX(2) - FIX(1.77200)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
; --------------------------------------------------------------------------
|
|
shun-iwasawa |
82a8f5 |
SECTION SEG_CONST
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
alignz 32
|
|
shun-iwasawa |
82a8f5 |
GLOBAL_DATA(jconst_merged_upsample_avx2)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
EXTN(jconst_merged_upsample_avx2):
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
PW_F0402 times 16 dw F_0_402
|
|
shun-iwasawa |
82a8f5 |
PW_MF0228 times 16 dw -F_0_228
|
|
shun-iwasawa |
82a8f5 |
PW_MF0344_F0285 times 8 dw -F_0_344, F_0_285
|
|
shun-iwasawa |
82a8f5 |
PW_ONE times 16 dw 1
|
|
shun-iwasawa |
82a8f5 |
PD_ONEHALF times 8 dd 1 << (SCALEBITS - 1)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
alignz 32
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
; --------------------------------------------------------------------------
|
|
shun-iwasawa |
82a8f5 |
SECTION SEG_TEXT
|
|
shun-iwasawa |
82a8f5 |
BITS 64
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%include "jdmrgext-avx2.asm"
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_RED
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_GREEN
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_BLUE
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_RED EXT_RGB_RED
|
|
shun-iwasawa |
82a8f5 |
%define RGB_GREEN EXT_RGB_GREEN
|
|
shun-iwasawa |
82a8f5 |
%define RGB_BLUE EXT_RGB_BLUE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_PIXELSIZE EXT_RGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v1_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v1_extrgb_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v2_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v2_extrgb_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%include "jdmrgext-avx2.asm"
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_RED
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_GREEN
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_BLUE
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_RED EXT_RGBX_RED
|
|
shun-iwasawa |
82a8f5 |
%define RGB_GREEN EXT_RGBX_GREEN
|
|
shun-iwasawa |
82a8f5 |
%define RGB_BLUE EXT_RGBX_BLUE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v1_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v1_extrgbx_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v2_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v2_extrgbx_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%include "jdmrgext-avx2.asm"
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_RED
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_GREEN
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_BLUE
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_RED EXT_BGR_RED
|
|
shun-iwasawa |
82a8f5 |
%define RGB_GREEN EXT_BGR_GREEN
|
|
shun-iwasawa |
82a8f5 |
%define RGB_BLUE EXT_BGR_BLUE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_PIXELSIZE EXT_BGR_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v1_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v1_extbgr_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v2_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v2_extbgr_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%include "jdmrgext-avx2.asm"
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_RED
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_GREEN
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_BLUE
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_RED EXT_BGRX_RED
|
|
shun-iwasawa |
82a8f5 |
%define RGB_GREEN EXT_BGRX_GREEN
|
|
shun-iwasawa |
82a8f5 |
%define RGB_BLUE EXT_BGRX_BLUE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v1_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v1_extbgrx_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v2_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v2_extbgrx_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%include "jdmrgext-avx2.asm"
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_RED
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_GREEN
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_BLUE
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_RED EXT_XBGR_RED
|
|
shun-iwasawa |
82a8f5 |
%define RGB_GREEN EXT_XBGR_GREEN
|
|
shun-iwasawa |
82a8f5 |
%define RGB_BLUE EXT_XBGR_BLUE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v1_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v1_extxbgr_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v2_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v2_extxbgr_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%include "jdmrgext-avx2.asm"
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_RED
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_GREEN
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_BLUE
|
|
shun-iwasawa |
82a8f5 |
%undef RGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_RED EXT_XRGB_RED
|
|
shun-iwasawa |
82a8f5 |
%define RGB_GREEN EXT_XRGB_GREEN
|
|
shun-iwasawa |
82a8f5 |
%define RGB_BLUE EXT_XRGB_BLUE
|
|
shun-iwasawa |
82a8f5 |
%define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v1_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v1_extxrgb_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%define jsimd_h2v2_merged_upsample_avx2 \
|
|
shun-iwasawa |
82a8f5 |
jsimd_h2v2_extxrgb_merged_upsample_avx2
|
|
shun-iwasawa |
82a8f5 |
%include "jdmrgext-avx2.asm"
|