1; 2; jquanti.asm - sample data conversion and quantization (64-bit AVX2) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; Copyright (C) 2009, 2016, 2018, D. R. Commander. 6; Copyright (C) 2016, Matthieu Darbois. 7; Copyright (C) 2018, Matthias Räncker. 8; 9; Based on the x86 SIMD extension for IJG JPEG library 10; Copyright (C) 1999-2006, MIYASAKA Masaru. 11; For conditions of distribution and use, see copyright notice in jsimdext.inc 12; 13; This file should be assembled with NASM (Netwide Assembler), 14; can *not* be assembled with Microsoft's MASM or any compatible 15; assembler (including Borland's Turbo Assembler). 16; NASM is available from http://nasm.sourceforge.net/ or 17; http://sourceforge.net/project/showfiles.php?group_id=6208 18 19%include "jsimdext.inc" 20%include "jdct.inc" 21 22; -------------------------------------------------------------------------- 23 SECTION SEG_TEXT 24 BITS 64 25; 26; Load data into workspace, applying unsigned->signed conversion 27; 28; GLOBAL(void) 29; jsimd_convsamp_avx2(JSAMPARRAY sample_data, JDIMENSION start_col, 30; DCTELEM *workspace); 31; 32 33; r10 = JSAMPARRAY sample_data 34; r11d = JDIMENSION start_col 35; r12 = DCTELEM *workspace 36 37 align 32 38 GLOBAL_FUNCTION(jsimd_convsamp_avx2) 39 40EXTN(jsimd_convsamp_avx2): 41 push rbp 42 mov rax, rsp 43 mov rbp, rsp 44 collect_args 3 45 46 mov eax, r11d 47 48 mov rsip, JSAMPROW [r10+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) 49 mov rdip, JSAMPROW [r10+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) 50 movq xmm0, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE] 51 pinsrq xmm0, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1 52 53 mov rsip, JSAMPROW [r10+2*SIZEOF_JSAMPROW] ; (JSAMPLE *) 54 mov rdip, JSAMPROW [r10+3*SIZEOF_JSAMPROW] ; (JSAMPLE *) 55 movq xmm1, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE] 56 pinsrq xmm1, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1 57 58 mov rsip, JSAMPROW [r10+4*SIZEOF_JSAMPROW] ; (JSAMPLE *) 59 mov rdip, JSAMPROW [r10+5*SIZEOF_JSAMPROW] ; (JSAMPLE *) 60 movq xmm2, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE] 61 pinsrq xmm2, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1 62 63 mov rsip, JSAMPROW [r10+6*SIZEOF_JSAMPROW] ; (JSAMPLE *) 64 mov rdip, JSAMPROW [r10+7*SIZEOF_JSAMPROW] ; (JSAMPLE *) 65 movq xmm3, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE] 66 pinsrq xmm3, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1 67 68 vpmovzxbw ymm0, xmm0 ; ymm0=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17) 69 vpmovzxbw ymm1, xmm1 ; ymm1=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37) 70 vpmovzxbw ymm2, xmm2 ; ymm2=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57) 71 vpmovzxbw ymm3, xmm3 ; ymm3=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77) 72 73 vpcmpeqw ymm7, ymm7, ymm7 74 vpsllw ymm7, ymm7, 7 ; ymm7={0xFF80 0xFF80 0xFF80 0xFF80 ..} 75 76 vpaddw ymm0, ymm0, ymm7 77 vpaddw ymm1, ymm1, ymm7 78 vpaddw ymm2, ymm2, ymm7 79 vpaddw ymm3, ymm3, ymm7 80 81 vmovdqu YMMWORD [YMMBLOCK(0,0,r12,SIZEOF_DCTELEM)], ymm0 82 vmovdqu YMMWORD [YMMBLOCK(2,0,r12,SIZEOF_DCTELEM)], ymm1 83 vmovdqu YMMWORD [YMMBLOCK(4,0,r12,SIZEOF_DCTELEM)], ymm2 84 vmovdqu YMMWORD [YMMBLOCK(6,0,r12,SIZEOF_DCTELEM)], ymm3 85 86 vzeroupper 87 uncollect_args 3 88 pop rbp 89 ret 90 91; -------------------------------------------------------------------------- 92; 93; Quantize/descale the coefficients, and store into coef_block 94; 95; This implementation is based on an algorithm described in 96; "How to optimize for the Pentium family of microprocessors" 97; (http://www.agner.org/assem/). 98; 99; GLOBAL(void) 100; jsimd_quantize_avx2(JCOEFPTR coef_block, DCTELEM *divisors, 101; DCTELEM *workspace); 102; 103 104%define RECIPROCAL(m, n, b) \ 105 YMMBLOCK(DCTSIZE * 0 + (m), (n), (b), SIZEOF_DCTELEM) 106%define CORRECTION(m, n, b) \ 107 YMMBLOCK(DCTSIZE * 1 + (m), (n), (b), SIZEOF_DCTELEM) 108%define SCALE(m, n, b) \ 109 YMMBLOCK(DCTSIZE * 2 + (m), (n), (b), SIZEOF_DCTELEM) 110 111; r10 = JCOEFPTR coef_block 112; r11 = DCTELEM *divisors 113; r12 = DCTELEM *workspace 114 115 align 32 116 GLOBAL_FUNCTION(jsimd_quantize_avx2) 117 118EXTN(jsimd_quantize_avx2): 119 push rbp 120 mov rax, rsp 121 mov rbp, rsp 122 collect_args 3 123 124 vmovdqu ymm4, [YMMBLOCK(0,0,r12,SIZEOF_DCTELEM)] 125 vmovdqu ymm5, [YMMBLOCK(2,0,r12,SIZEOF_DCTELEM)] 126 vmovdqu ymm6, [YMMBLOCK(4,0,r12,SIZEOF_DCTELEM)] 127 vmovdqu ymm7, [YMMBLOCK(6,0,r12,SIZEOF_DCTELEM)] 128 vpabsw ymm0, ymm4 129 vpabsw ymm1, ymm5 130 vpabsw ymm2, ymm6 131 vpabsw ymm3, ymm7 132 133 vpaddw ymm0, YMMWORD [CORRECTION(0,0,r11)] ; correction + roundfactor 134 vpaddw ymm1, YMMWORD [CORRECTION(2,0,r11)] 135 vpaddw ymm2, YMMWORD [CORRECTION(4,0,r11)] 136 vpaddw ymm3, YMMWORD [CORRECTION(6,0,r11)] 137 vpmulhuw ymm0, YMMWORD [RECIPROCAL(0,0,r11)] ; reciprocal 138 vpmulhuw ymm1, YMMWORD [RECIPROCAL(2,0,r11)] 139 vpmulhuw ymm2, YMMWORD [RECIPROCAL(4,0,r11)] 140 vpmulhuw ymm3, YMMWORD [RECIPROCAL(6,0,r11)] 141 vpmulhuw ymm0, YMMWORD [SCALE(0,0,r11)] ; scale 142 vpmulhuw ymm1, YMMWORD [SCALE(2,0,r11)] 143 vpmulhuw ymm2, YMMWORD [SCALE(4,0,r11)] 144 vpmulhuw ymm3, YMMWORD [SCALE(6,0,r11)] 145 146 vpsignw ymm0, ymm0, ymm4 147 vpsignw ymm1, ymm1, ymm5 148 vpsignw ymm2, ymm2, ymm6 149 vpsignw ymm3, ymm3, ymm7 150 151 vmovdqu [YMMBLOCK(0,0,r10,SIZEOF_DCTELEM)], ymm0 152 vmovdqu [YMMBLOCK(2,0,r10,SIZEOF_DCTELEM)], ymm1 153 vmovdqu [YMMBLOCK(4,0,r10,SIZEOF_DCTELEM)], ymm2 154 vmovdqu [YMMBLOCK(6,0,r10,SIZEOF_DCTELEM)], ymm3 155 156 vzeroupper 157 uncollect_args 3 158 pop rbp 159 ret 160 161; For some reason, the OS X linker does not honor the request to align the 162; segment unless we do this. 163 align 32 164