Thu, 22 Jan 2015 13:21:57 +0100
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
michael@0 | 1 | ; |
michael@0 | 2 | ; jcqnts2f.asm - sample data conversion and quantization (SSE & SSE2) |
michael@0 | 3 | ; |
michael@0 | 4 | ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB |
michael@0 | 5 | ; |
michael@0 | 6 | ; Based on |
michael@0 | 7 | ; x86 SIMD extension for IJG JPEG library |
michael@0 | 8 | ; Copyright (C) 1999-2006, MIYASAKA Masaru. |
michael@0 | 9 | ; For conditions of distribution and use, see copyright notice in jsimdext.inc |
michael@0 | 10 | ; |
michael@0 | 11 | ; This file should be assembled with NASM (Netwide Assembler), |
michael@0 | 12 | ; can *not* be assembled with Microsoft's MASM or any compatible |
michael@0 | 13 | ; assembler (including Borland's Turbo Assembler). |
michael@0 | 14 | ; NASM is available from http://nasm.sourceforge.net/ or |
michael@0 | 15 | ; http://sourceforge.net/project/showfiles.php?group_id=6208 |
michael@0 | 16 | ; |
michael@0 | 17 | ; [TAB8] |
michael@0 | 18 | |
michael@0 | 19 | %include "jsimdext.inc" |
michael@0 | 20 | %include "jdct.inc" |
michael@0 | 21 | |
michael@0 | 22 | ; -------------------------------------------------------------------------- |
michael@0 | 23 | SECTION SEG_TEXT |
michael@0 | 24 | BITS 32 |
michael@0 | 25 | ; |
michael@0 | 26 | ; Load data into workspace, applying unsigned->signed conversion |
michael@0 | 27 | ; |
michael@0 | 28 | ; GLOBAL(void) |
michael@0 | 29 | ; jsimd_convsamp_float_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col, |
michael@0 | 30 | ; FAST_FLOAT * workspace); |
michael@0 | 31 | ; |
michael@0 | 32 | |
michael@0 | 33 | %define sample_data ebp+8 ; JSAMPARRAY sample_data |
michael@0 | 34 | %define start_col ebp+12 ; JDIMENSION start_col |
michael@0 | 35 | %define workspace ebp+16 ; FAST_FLOAT * workspace |
michael@0 | 36 | |
michael@0 | 37 | align 16 |
michael@0 | 38 | global EXTN(jsimd_convsamp_float_sse2) |
michael@0 | 39 | |
michael@0 | 40 | EXTN(jsimd_convsamp_float_sse2): |
michael@0 | 41 | push ebp |
michael@0 | 42 | mov ebp,esp |
michael@0 | 43 | push ebx |
michael@0 | 44 | ; push ecx ; need not be preserved |
michael@0 | 45 | ; push edx ; need not be preserved |
michael@0 | 46 | push esi |
michael@0 | 47 | push edi |
michael@0 | 48 | |
michael@0 | 49 | pcmpeqw xmm7,xmm7 |
michael@0 | 50 | psllw xmm7,7 |
michael@0 | 51 | packsswb xmm7,xmm7 ; xmm7 = PB_CENTERJSAMPLE (0x808080..) |
michael@0 | 52 | |
michael@0 | 53 | mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *) |
michael@0 | 54 | mov eax, JDIMENSION [start_col] |
michael@0 | 55 | mov edi, POINTER [workspace] ; (DCTELEM *) |
michael@0 | 56 | mov ecx, DCTSIZE/2 |
michael@0 | 57 | alignx 16,7 |
michael@0 | 58 | .convloop: |
michael@0 | 59 | mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) |
michael@0 | 60 | mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) |
michael@0 | 61 | |
michael@0 | 62 | movq xmm0, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE] |
michael@0 | 63 | movq xmm1, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE] |
michael@0 | 64 | |
michael@0 | 65 | psubb xmm0,xmm7 ; xmm0=(01234567) |
michael@0 | 66 | psubb xmm1,xmm7 ; xmm1=(89ABCDEF) |
michael@0 | 67 | |
michael@0 | 68 | punpcklbw xmm0,xmm0 ; xmm0=(*0*1*2*3*4*5*6*7) |
michael@0 | 69 | punpcklbw xmm1,xmm1 ; xmm1=(*8*9*A*B*C*D*E*F) |
michael@0 | 70 | |
michael@0 | 71 | punpcklwd xmm2,xmm0 ; xmm2=(***0***1***2***3) |
michael@0 | 72 | punpckhwd xmm0,xmm0 ; xmm0=(***4***5***6***7) |
michael@0 | 73 | punpcklwd xmm3,xmm1 ; xmm3=(***8***9***A***B) |
michael@0 | 74 | punpckhwd xmm1,xmm1 ; xmm1=(***C***D***E***F) |
michael@0 | 75 | |
michael@0 | 76 | psrad xmm2,(DWORD_BIT-BYTE_BIT) ; xmm2=(0123) |
michael@0 | 77 | psrad xmm0,(DWORD_BIT-BYTE_BIT) ; xmm0=(4567) |
michael@0 | 78 | cvtdq2ps xmm2,xmm2 ; xmm2=(0123) |
michael@0 | 79 | cvtdq2ps xmm0,xmm0 ; xmm0=(4567) |
michael@0 | 80 | psrad xmm3,(DWORD_BIT-BYTE_BIT) ; xmm3=(89AB) |
michael@0 | 81 | psrad xmm1,(DWORD_BIT-BYTE_BIT) ; xmm1=(CDEF) |
michael@0 | 82 | cvtdq2ps xmm3,xmm3 ; xmm3=(89AB) |
michael@0 | 83 | cvtdq2ps xmm1,xmm1 ; xmm1=(CDEF) |
michael@0 | 84 | |
michael@0 | 85 | movaps XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm2 |
michael@0 | 86 | movaps XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0 |
michael@0 | 87 | movaps XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3 |
michael@0 | 88 | movaps XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1 |
michael@0 | 89 | |
michael@0 | 90 | add esi, byte 2*SIZEOF_JSAMPROW |
michael@0 | 91 | add edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT |
michael@0 | 92 | dec ecx |
michael@0 | 93 | jnz short .convloop |
michael@0 | 94 | |
michael@0 | 95 | pop edi |
michael@0 | 96 | pop esi |
michael@0 | 97 | ; pop edx ; need not be preserved |
michael@0 | 98 | ; pop ecx ; need not be preserved |
michael@0 | 99 | pop ebx |
michael@0 | 100 | pop ebp |
michael@0 | 101 | ret |
michael@0 | 102 | |
michael@0 | 103 | |
michael@0 | 104 | ; -------------------------------------------------------------------------- |
michael@0 | 105 | ; |
michael@0 | 106 | ; Quantize/descale the coefficients, and store into coef_block |
michael@0 | 107 | ; |
michael@0 | 108 | ; GLOBAL(void) |
michael@0 | 109 | ; jsimd_quantize_float_sse2 (JCOEFPTR coef_block, FAST_FLOAT * divisors, |
michael@0 | 110 | ; FAST_FLOAT * workspace); |
michael@0 | 111 | ; |
michael@0 | 112 | |
michael@0 | 113 | %define coef_block ebp+8 ; JCOEFPTR coef_block |
michael@0 | 114 | %define divisors ebp+12 ; FAST_FLOAT * divisors |
michael@0 | 115 | %define workspace ebp+16 ; FAST_FLOAT * workspace |
michael@0 | 116 | |
michael@0 | 117 | align 16 |
michael@0 | 118 | global EXTN(jsimd_quantize_float_sse2) |
michael@0 | 119 | |
michael@0 | 120 | EXTN(jsimd_quantize_float_sse2): |
michael@0 | 121 | push ebp |
michael@0 | 122 | mov ebp,esp |
michael@0 | 123 | ; push ebx ; unused |
michael@0 | 124 | ; push ecx ; unused |
michael@0 | 125 | ; push edx ; need not be preserved |
michael@0 | 126 | push esi |
michael@0 | 127 | push edi |
michael@0 | 128 | |
michael@0 | 129 | mov esi, POINTER [workspace] |
michael@0 | 130 | mov edx, POINTER [divisors] |
michael@0 | 131 | mov edi, JCOEFPTR [coef_block] |
michael@0 | 132 | mov eax, DCTSIZE2/16 |
michael@0 | 133 | alignx 16,7 |
michael@0 | 134 | .quantloop: |
michael@0 | 135 | movaps xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)] |
michael@0 | 136 | movaps xmm1, XMMWORD [XMMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)] |
michael@0 | 137 | mulps xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)] |
michael@0 | 138 | mulps xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)] |
michael@0 | 139 | movaps xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)] |
michael@0 | 140 | movaps xmm3, XMMWORD [XMMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)] |
michael@0 | 141 | mulps xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)] |
michael@0 | 142 | mulps xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)] |
michael@0 | 143 | |
michael@0 | 144 | cvtps2dq xmm0,xmm0 |
michael@0 | 145 | cvtps2dq xmm1,xmm1 |
michael@0 | 146 | cvtps2dq xmm2,xmm2 |
michael@0 | 147 | cvtps2dq xmm3,xmm3 |
michael@0 | 148 | |
michael@0 | 149 | packssdw xmm0,xmm1 |
michael@0 | 150 | packssdw xmm2,xmm3 |
michael@0 | 151 | |
michael@0 | 152 | movdqa XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_JCOEF)], xmm0 |
michael@0 | 153 | movdqa XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_JCOEF)], xmm2 |
michael@0 | 154 | |
michael@0 | 155 | add esi, byte 16*SIZEOF_FAST_FLOAT |
michael@0 | 156 | add edx, byte 16*SIZEOF_FAST_FLOAT |
michael@0 | 157 | add edi, byte 16*SIZEOF_JCOEF |
michael@0 | 158 | dec eax |
michael@0 | 159 | jnz short .quantloop |
michael@0 | 160 | |
michael@0 | 161 | pop edi |
michael@0 | 162 | pop esi |
michael@0 | 163 | ; pop edx ; need not be preserved |
michael@0 | 164 | ; pop ecx ; unused |
michael@0 | 165 | ; pop ebx ; unused |
michael@0 | 166 | pop ebp |
michael@0 | 167 | ret |
michael@0 | 168 | |
michael@0 | 169 | ; For some reason, the OS X linker does not honor the request to align the |
michael@0 | 170 | ; segment unless we do this. |
michael@0 | 171 | align 16 |