Integrate premultiply/unpremultiply operations into libjpeg-turbo....

third_party/libjpeg_turbo/simd/jcclrss2.asm

 ;
 ; jcclrss2.asm - colorspace conversion (SSE2)
 ;
 ; x86 SIMD extension for IJG JPEG library
 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
 ;
 ; This file should be assembled with NASM (Netwide Assembler),
 ; can *not* be assembled with Microsoft's MASM or any compatible
 ; assembler (including Borland's Turbo Assembler).
@@ skipping 275 matching lines @@
         punpcklbw xmmE,xmmF     ; xmmE=(20 22 24 26 28 2A 2C 2E)
         punpckhbw xmmG,xmmF     ; xmmG=(30 32 34 36 38 3A 3C 3E)
 
         punpcklbw xmmF,xmmH
         punpckhbw xmmH,xmmH
         psrlw     xmmF,BYTE_BIT ; xmmF=(21 23 25 27 29 2B 2D 2F)
         psrlw     xmmH,BYTE_BIT ; xmmH=(31 33 35 37 39 3B 3D 3F)
 
 %endif ; RGB_PIXELSIZE ; ---------------
 
+%if PREMULTIPLY == 1 ; ---------------
+        ; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE, xmm6=A(02468ACE)=AE
+        ; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO, xmm7=A(13579BDF)=AO
+
+        ; Unpremultiply even registers (i.e. xmm0, xmm2, xmm4, and xmm6).
+        movdqa [wk(0)], xmm1
+        movdqa [wk(1)], xmm3
+        movdqa [wk(2)], xmm5
+        movdqa [wk(3)], xmm7
+
+        ; for (int i = 0; i < 8; ++i)
+        ;   xmm6.word[i] = xmm6.word[i] ? xmm6.word[i] : 255; // a[i] ? a[i] : 255;
+        pxor xmm1, xmm1
+        pcmpeqw xmm1, xmm6
+        psrlw xmm1, 8
+        pmaxuw xmm6, xmm1
+
+        ; xmm0.dword[i] = r[i];   xmm2.dword[i] = g[i];   xmm4.dword[i] = b[i];   xmm6.dword[i] = a[i];
+        ; xmm1.dword[i] = r[4+i]; xmm3.dword[i] = g[4+i]; xmm5.dword[i] = b[4+i]; xmm7.dword[i] = a[4+i];
+        movdqa xmm1, xmm0
+        movdqa xmm3, xmm2
+        movdqa xmm5, xmm4
+        movdqa xmm7, xmm6
+
+        movdqa [wk(4)], xmm1
+        pxor xmm1, xmm1
+        punpcklwd xmm0, xmm1
+        punpcklwd xmm2, xmm1
+        punpcklwd xmm4, xmm1
+        punpcklwd xmm6, xmm1
+        movdqa xmm1, [wk(4)]
+
+        movdqa [wk(4)], xmm0
+        pxor xmm0, xmm0
+        punpckhwd xmm1, xmm0
+        punpckhwd xmm3, xmm0
+        punpckhwd xmm5, xmm0
+        punpckhwd xmm7, xmm0
+        movdqa xmm0, [wk(4)];
+
+        ; for (int i = 0; i < 4; ++i) {
+        ;   xmm0.float[i] = (float)xmm0.dword[i]; // (float)r[i];
+        ;   xmm2.float[i] = (float)xmm2.dword[i]; // (float)g[i];
+        ;   xmm4.float[i] = (float)xmm4.dword[i]; // (float)b[i];
+        ;   xmm6.float[i] = (float)xmm6.dword[i]; // (float)a[i];
+        ;   xmm1.float[i] = (float)xmm1.dword[i]; // (float)r[4+i];
+        ;   xmm3.float[i] = (float)xmm3.dword[i]; // (float)g[4+i];
+        ;   xmm5.float[i] = (float)xmm5.dword[i]; // (float)b[4+i];
+        ;   xmm7.float[i] = (float)xmm7.dword[i]; // (float)a[4+i];
+        ; }
+        cvtdq2ps xmm0, xmm0
+        cvtdq2ps xmm2, xmm2
+        cvtdq2ps xmm4, xmm4
+        cvtdq2ps xmm6, xmm6
+        cvtdq2ps xmm1, xmm1
+        cvtdq2ps xmm3, xmm3
+        cvtdq2ps xmm5, xmm5
+        cvtdq2ps xmm7, xmm7
+
+        ; for (int i = 0; i < 4; ++i) {
+        ;   xmm6.float[i] = 255.0 / xmm6.float[i]; // 255.0 / (float)a[i];
+        ;   xmm7.float[i] = 255.0 / xmm7.float[i]; // 255.0 / (float)a[4+i];
+        ; }
+        rcpps xmm6, xmm6
+        rcpps xmm7, xmm7
+        mulps xmm6, [GOTOFF(eax,PF_255)]
+        mulps xmm7, [GOTOFF(eax,PF_255)]
+
+        ; for (int i = 0; i < 4; ++i) {
+        ;   xmm0.word[i]   = (uint16)(xmm0.float[i] * xmm6.float[i]); // (uint16)((float)r[i] * 255.0 / (float)a[i]);
+        ;   xmm2.word[i]   = (uint16)(xmm2.float[i] * xmm6.float[i]); // (uint16)((float)g[i] * 255.0 / (float)a[i]);
+        ;   xmm4.word[i]   = (uint16)(xmm4.float[i] * xmm6.float[i]); // (uint16)((float)b[i] * 255.0 / (float)a[i]);
+        ;   xmm0.word[4+i] = (uint16)(xmm1.float[i] * xmm7.float[i]); // (uint16)((float)r[4+i] * 255.0 / (float)a[4+i]);
+        ;   xmm2.word[4+i] = (uint16)(xmm3.float[i] * xmm7.float[i]); // (uint16)((float)g[4+i] * 255.0 / (float)a[4+i]);
+        ;   xmm4.word[4+i] = (uint16)(xmm5.float[i] * xmm7.float[i]); // (uint16)((float)b[4+i] * 255.0 / (float)a[4+i]);
+        ; }
+        mulps xmm0, xmm6
+        cvtps2dq xmm0, xmm0
+        mulps xmm1, xmm7
+        cvtps2dq xmm1, xmm1
+        packusdw xmm0, xmm1
+
+        mulps xmm2, xmm6
+        cvtps2dq xmm2, xmm2
+        mulps xmm3, xmm7
+        cvtps2dq xmm3, xmm3
+        packusdw xmm2, xmm3
+
+        mulps xmm4, xmm6
+        cvtps2dq xmm4, xmm4
+        mulps xmm5, xmm7
+        cvtps2dq xmm5, xmm5
+        packusdw xmm4, xmm5
+
+        movdqa xmm1, [wk(0)]
+        movdqa xmm3, [wk(1)]
+        movdqa xmm5, [wk(2)]
+        movdqa xmm7, [wk(3)]
+
+        ; Unpremultiply odd registers (i.e. xmm1, xmm3, xmm5, and xmm7).
+        movdqa [wk(0)], xmm0
+        movdqa [wk(1)], xmm2
+        movdqa [wk(2)], xmm4
+        movdqa [wk(3)], xmm6
+
+        ; for (int i = 0; i < 8; ++i)
+        ;   xmm7.word[i] = xmm7.word[i] ? xmm7.word[i] : 255; // a[i] ? a[i] : 255;
+        pxor xmm0, xmm0
+        pcmpeqw xmm0, xmm7
+        psrlw xmm0, 8
+        pmaxuw xmm7, xmm0
+
+        ; xmm4.dword[i] = r[i];   xmm5.dword[i] = g[i];   xmm6.dword[i] = b[i];   xmm7.dword[i] = a[i];
+        ; xmm0.dword[i] = r[4+i]; xmm1.dword[i] = g[4+i]; xmm2.dword[i] = b[4+i]; xmm3.dword[i] = a[4+i];
+        movdqa xmm0, xmm1
+        movdqa xmm2, xmm3
+        movdqa xmm4, xmm5
+        movdqa xmm6, xmm7
+
+        movdqa [wk(4)], xmm0
+        pxor xmm0, xmm0
+        punpcklwd xmm1, xmm0
+        punpcklwd xmm3, xmm0
+        punpcklwd xmm5, xmm0
+        punpcklwd xmm7, xmm0
+        movdqa xmm0, [wk(4)]
+
+        movdqa [wk(4)], xmm1
+        pxor xmm1, xmm1
+        punpckhwd xmm0, xmm1
+        punpckhwd xmm2, xmm1
+        punpckhwd xmm4, xmm1
+        punpckhwd xmm6, xmm1
+        movdqa xmm1, [wk(4)];
+
+        ; for (int i = 0; i < 4; ++i) {
+        ;   xmm1.float[i] = (float)xmm1.dword[i]; // (float)r[i];
+        ;   xmm3.float[i] = (float)xmm3.dword[i]; // (float)g[i];
+        ;   xmm5.float[i] = (float)xmm5.dword[i]; // (float)b[i];
+        ;   xmm7.float[i] = (float)xmm7.dword[i]; // (float)a[i];
+        ;   xmm0.float[i] = (float)xmm0.dword[i]; // (float)r[4+i];
+        ;   xmm2.float[i] = (float)xmm2.dword[i]; // (float)g[4+i];
+        ;   xmm4.float[i] = (float)xmm4.dword[i]; // (float)b[4+i];
+        ;   xmm6.float[i] = (float)xmm6.dword[i]; // (float)a[4+i];
+        ; }
+        cvtdq2ps xmm1, xmm1
+        cvtdq2ps xmm3, xmm3
+        cvtdq2ps xmm5, xmm5
+        cvtdq2ps xmm7, xmm7
+        cvtdq2ps xmm0, xmm0
+        cvtdq2ps xmm2, xmm2
+        cvtdq2ps xmm4, xmm4
+        cvtdq2ps xmm6, xmm6
+
+        ; for (int i = 0; i < 4; ++i) {
+        ;   xmm7.float[i] = 255.0 / xmm7.float[i]; // 255.0 / (float)a[i];
+        ;   xmm6.float[i] = 255.0 / xmm6.float[i]; // 255.0 / (float)a[4+i];
+        ; }
+        rcpps xmm7, xmm7
+        rcpps xmm6, xmm6
+        mulps xmm7, [GOTOFF(eax,PF_255)]
+        mulps xmm6, [GOTOFF(eax,PF_255)]
+
+        ; for (int i = 0; i < 4; ++i) {
+        ;   xmm1.word[i]   = (uint16)(xmm1.float[i] * xmm7.float[i]); // (uint16)((float)r[i] * 255.0 / (float)a[i]);
+        ;   xmm3.word[i]   = (uint16)(xmm3.float[i] * xmm7.float[i]); // (uint16)((float)g[i] * 255.0 / (float)a[i]);
+        ;   xmm5.word[i]   = (uint16)(xmm5.float[i] * xmm7.float[i]); // (uint16)((float)b[i] * 255.0 / (float)a[i]);
+        ;   xmm1.word[4+i] = (uint16)(xmm0.float[i] * xmm6.float[i]); // (uint16)((float)r[4+i] * 255.0 / (float)a[4+i]);
+        ;   xmm3.word[4+i] = (uint16)(xmm0.float[i] * xmm6.float[i]); // (uint16)((float)g[4+i] * 255.0 / (float)a[4+i]);
+        ;   xmm5.word[4+i] = (uint16)(xmm0.float[i] * xmm6.float[i]); // (uint16)((float)b[4+i] * 255.0 / (float)a[4+i]);
+        ; }
+        mulps xmm1, xmm7
+        cvtps2dq xmm1, xmm1
+        mulps xmm0, xmm6
+        cvtps2dq xmm0, xmm0
+        packusdw xmm1, xmm0
+
+        mulps xmm3, xmm7
+        cvtps2dq xmm3, xmm3
+        mulps xmm2, xmm6
+        cvtps2dq xmm2, xmm2
+        packusdw xmm3, xmm2
+
+        mulps xmm5, xmm7
+        cvtps2dq xmm5, xmm5
+        mulps xmm4, xmm6
+        cvtps2dq xmm4, xmm4
+        packusdw xmm5, xmm4
+
+        movdqa xmm0, [wk(0)]
+        movdqa xmm2, [wk(1)]
+        movdqa xmm4, [wk(2)]
+        movdqa xmm6, [wk(3)]
+%endif ; PREMULTIPLY == 1
+
         ; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
         ; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
 
         ; (Original)
         ; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
         ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
         ; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
         ;
         ; (This implementation)
         ; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
@@ skipping 190 matching lines @@
 ;       pop     ecx             ; need not be preserved
         pop     ebx
         mov     esp,ebp         ; esp <- aligned ebp
         pop     esp             ; esp <- original ebp
         pop     ebp
         ret
 
 ; For some reason, the OS X linker does not honor the request to align the
 ; segment unless we do this.
         align   16