jquant-sse.asm (7368B)
1 ; 2 ; jquant.asm - sample data conversion and quantization (SSE & MMX) 3 ; 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5 ; Copyright (C) 2016, 2024, D. R. Commander. 6 ; 7 ; Based on the x86 SIMD extension for IJG JPEG library 8 ; Copyright (C) 1999-2006, MIYASAKA Masaru. 9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc 10 ; 11 ; This file should be assembled with NASM (Netwide Assembler) or Yasm. 12 13 %include "jsimdext.inc" 14 %include "jdct.inc" 15 16 ; -------------------------------------------------------------------------- 17 SECTION SEG_TEXT 18 BITS 32 19 ; 20 ; Load data into workspace, applying unsigned->signed conversion 21 ; 22 ; GLOBAL(void) 23 ; jsimd_convsamp_float_sse(JSAMPARRAY sample_data, JDIMENSION start_col, 24 ; FAST_FLOAT *workspace); 25 ; 26 27 %define sample_data ebp + 8 ; JSAMPARRAY sample_data 28 %define start_col ebp + 12 ; JDIMENSION start_col 29 %define workspace ebp + 16 ; FAST_FLOAT *workspace 30 31 align 32 32 GLOBAL_FUNCTION(jsimd_convsamp_float_sse) 33 34 EXTN(jsimd_convsamp_float_sse): 35 push ebp 36 mov ebp, esp 37 push ebx 38 ; push ecx ; need not be preserved 39 ; push edx ; need not be preserved 40 push esi 41 push edi 42 43 pcmpeqw mm7, mm7 44 psllw mm7, 7 45 packsswb mm7, mm7 ; mm7 = PB_CENTERJSAMPLE (0x808080..) 46 47 mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *) 48 mov eax, JDIMENSION [start_col] 49 mov edi, POINTER [workspace] ; (DCTELEM *) 50 mov ecx, DCTSIZE/2 51 ALIGNX 16, 7 52 .convloop: 53 mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) 54 mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) 55 56 movq mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE] 57 movq mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE] 58 59 psubb mm0, mm7 ; mm0=(01234567) 60 psubb mm1, mm7 ; mm1=(89ABCDEF) 61 62 punpcklbw mm2, mm0 ; mm2=(*0*1*2*3) 63 punpckhbw mm0, mm0 ; mm0=(*4*5*6*7) 64 punpcklbw mm3, mm1 ; mm3=(*8*9*A*B) 65 punpckhbw mm1, mm1 ; mm1=(*C*D*E*F) 66 67 punpcklwd mm4, mm2 ; mm4=(***0***1) 68 punpckhwd mm2, mm2 ; mm2=(***2***3) 69 punpcklwd mm5, mm0 ; mm5=(***4***5) 70 punpckhwd mm0, mm0 ; mm0=(***6***7) 71 72 psrad mm4, (DWORD_BIT-BYTE_BIT) ; mm4=(01) 73 psrad mm2, (DWORD_BIT-BYTE_BIT) ; mm2=(23) 74 cvtpi2ps xmm0, mm4 ; xmm0=(01**) 75 cvtpi2ps xmm1, mm2 ; xmm1=(23**) 76 psrad mm5, (DWORD_BIT-BYTE_BIT) ; mm5=(45) 77 psrad mm0, (DWORD_BIT-BYTE_BIT) ; mm0=(67) 78 cvtpi2ps xmm2, mm5 ; xmm2=(45**) 79 cvtpi2ps xmm3, mm0 ; xmm3=(67**) 80 81 punpcklwd mm6, mm3 ; mm6=(***8***9) 82 punpckhwd mm3, mm3 ; mm3=(***A***B) 83 punpcklwd mm4, mm1 ; mm4=(***C***D) 84 punpckhwd mm1, mm1 ; mm1=(***E***F) 85 86 psrad mm6, (DWORD_BIT-BYTE_BIT) ; mm6=(89) 87 psrad mm3, (DWORD_BIT-BYTE_BIT) ; mm3=(AB) 88 cvtpi2ps xmm4, mm6 ; xmm4=(89**) 89 cvtpi2ps xmm5, mm3 ; xmm5=(AB**) 90 psrad mm4, (DWORD_BIT-BYTE_BIT) ; mm4=(CD) 91 psrad mm1, (DWORD_BIT-BYTE_BIT) ; mm1=(EF) 92 cvtpi2ps xmm6, mm4 ; xmm6=(CD**) 93 cvtpi2ps xmm7, mm1 ; xmm7=(EF**) 94 95 movlhps xmm0, xmm1 ; xmm0=(0123) 96 movlhps xmm2, xmm3 ; xmm2=(4567) 97 movlhps xmm4, xmm5 ; xmm4=(89AB) 98 movlhps xmm6, xmm7 ; xmm6=(CDEF) 99 100 movaps XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0 101 movaps XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm2 102 movaps XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm4 103 movaps XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6 104 105 add esi, byte 2*SIZEOF_JSAMPROW 106 add edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT 107 dec ecx 108 jnz near .convloop 109 110 emms ; empty MMX state 111 112 pop edi 113 pop esi 114 ; pop edx ; need not be preserved 115 ; pop ecx ; need not be preserved 116 pop ebx 117 pop ebp 118 ret 119 120 ; -------------------------------------------------------------------------- 121 ; 122 ; Quantize/descale the coefficients, and store into coef_block 123 ; 124 ; GLOBAL(void) 125 ; jsimd_quantize_float_sse(JCOEFPTR coef_block, FAST_FLOAT *divisors, 126 ; FAST_FLOAT *workspace); 127 ; 128 129 %define coef_block ebp + 8 ; JCOEFPTR coef_block 130 %define divisors ebp + 12 ; FAST_FLOAT *divisors 131 %define workspace ebp + 16 ; FAST_FLOAT *workspace 132 133 align 32 134 GLOBAL_FUNCTION(jsimd_quantize_float_sse) 135 136 EXTN(jsimd_quantize_float_sse): 137 push ebp 138 mov ebp, esp 139 ; push ebx ; unused 140 ; push ecx ; unused 141 ; push edx ; need not be preserved 142 push esi 143 push edi 144 145 mov esi, POINTER [workspace] 146 mov edx, POINTER [divisors] 147 mov edi, JCOEFPTR [coef_block] 148 mov eax, DCTSIZE2/16 149 ALIGNX 16, 7 150 .quantloop: 151 movaps xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)] 152 movaps xmm1, XMMWORD [XMMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)] 153 mulps xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)] 154 mulps xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)] 155 movaps xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)] 156 movaps xmm3, XMMWORD [XMMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)] 157 mulps xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)] 158 mulps xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)] 159 160 movhlps xmm4, xmm0 161 movhlps xmm5, xmm1 162 163 cvtps2pi mm0, xmm0 164 cvtps2pi mm1, xmm1 165 cvtps2pi mm4, xmm4 166 cvtps2pi mm5, xmm5 167 168 movhlps xmm6, xmm2 169 movhlps xmm7, xmm3 170 171 cvtps2pi mm2, xmm2 172 cvtps2pi mm3, xmm3 173 cvtps2pi mm6, xmm6 174 cvtps2pi mm7, xmm7 175 176 packssdw mm0, mm4 177 packssdw mm1, mm5 178 packssdw mm2, mm6 179 packssdw mm3, mm7 180 181 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0 182 movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm1 183 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm2 184 movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm3 185 186 add esi, byte 16*SIZEOF_FAST_FLOAT 187 add edx, byte 16*SIZEOF_FAST_FLOAT 188 add edi, byte 16*SIZEOF_JCOEF 189 dec eax 190 jnz short .quantloop 191 192 emms ; empty MMX state 193 194 pop edi 195 pop esi 196 ; pop edx ; need not be preserved 197 ; pop ecx ; unused 198 ; pop ebx ; unused 199 pop ebp 200 ret 201 202 ; For some reason, the OS X linker does not honor the request to align the 203 ; segment unless we do this. 204 align 32