Add MMX SIMD implementation of computationally intensive routines.
git-svn-id: svn://svn.code.sf.net/p/tigervnc/code/trunk@3648 3789f03b-4d11-0410-bbf8-ca57d06f2519
diff --git a/common/jpeg/simd/jcqntmmx.asm b/common/jpeg/simd/jcqntmmx.asm
new file mode 100644
index 0000000..6096eaa
--- /dev/null
+++ b/common/jpeg/simd/jcqntmmx.asm
@@ -0,0 +1,271 @@
+;
+; jcqntmmx.asm - sample data conversion and quantization (MMX)
+;
+; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+;
+; Based on
+; 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).
+; NASM is available from http://nasm.sourceforge.net/ or
+; http://sourceforge.net/project/showfiles.php?group_id=6208
+;
+; [TAB8]
+
+%include "simd/jsimdext.inc"
+%include "simd/jdct.inc"
+
+; --------------------------------------------------------------------------
+ SECTION SEG_TEXT
+ BITS 32
+;
+; Load data into workspace, applying unsigned->signed conversion
+;
+; GLOBAL(void)
+; jsimd_convsamp_mmx (JSAMPARRAY sample_data, JDIMENSION start_col,
+; DCTELEM * workspace);
+;
+
+%define sample_data ebp+8 ; JSAMPARRAY sample_data
+%define start_col ebp+12 ; JDIMENSION start_col
+%define workspace ebp+16 ; DCTELEM * workspace
+
+ align 16
+ global EXTN(jsimd_convsamp_mmx)
+
+EXTN(jsimd_convsamp_mmx):
+ push ebp
+ mov ebp,esp
+ push ebx
+; push ecx ; need not be preserved
+; push edx ; need not be preserved
+ push esi
+ push edi
+
+ pxor mm6,mm6 ; mm6=(all 0's)
+ pcmpeqw mm7,mm7
+ psllw mm7,7 ; mm7={0xFF80 0xFF80 0xFF80 0xFF80}
+
+ mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *)
+ mov eax, JDIMENSION [start_col]
+ mov edi, POINTER [workspace] ; (DCTELEM *)
+ mov ecx, DCTSIZE/4
+ alignx 16,7
+.convloop:
+ mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+
+ movq mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE] ; mm0=(01234567)
+ movq mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE] ; mm1=(89ABCDEF)
+
+ mov ebx, JSAMPROW [esi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov edx, JSAMPROW [esi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+
+ movq mm2, MMWORD [ebx+eax*SIZEOF_JSAMPLE] ; mm2=(GHIJKLMN)
+ movq mm3, MMWORD [edx+eax*SIZEOF_JSAMPLE] ; mm3=(OPQRSTUV)
+
+ movq mm4,mm0
+ punpcklbw mm0,mm6 ; mm0=(0123)
+ punpckhbw mm4,mm6 ; mm4=(4567)
+ movq mm5,mm1
+ punpcklbw mm1,mm6 ; mm1=(89AB)
+ punpckhbw mm5,mm6 ; mm5=(CDEF)
+
+ paddw mm0,mm7
+ paddw mm4,mm7
+ paddw mm1,mm7
+ paddw mm5,mm7
+
+ movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_DCTELEM)], mm0
+ movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_DCTELEM)], mm4
+ movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_DCTELEM)], mm1
+ movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_DCTELEM)], mm5
+
+ movq mm0,mm2
+ punpcklbw mm2,mm6 ; mm2=(GHIJ)
+ punpckhbw mm0,mm6 ; mm0=(KLMN)
+ movq mm4,mm3
+ punpcklbw mm3,mm6 ; mm3=(OPQR)
+ punpckhbw mm4,mm6 ; mm4=(STUV)
+
+ paddw mm2,mm7
+ paddw mm0,mm7
+ paddw mm3,mm7
+ paddw mm4,mm7
+
+ movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_DCTELEM)], mm2
+ movq MMWORD [MMBLOCK(2,1,edi,SIZEOF_DCTELEM)], mm0
+ movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_DCTELEM)], mm3
+ movq MMWORD [MMBLOCK(3,1,edi,SIZEOF_DCTELEM)], mm4
+
+ add esi, byte 4*SIZEOF_JSAMPROW
+ add edi, byte 4*DCTSIZE*SIZEOF_DCTELEM
+ dec ecx
+ jnz short .convloop
+
+ emms ; empty MMX state
+
+ pop edi
+ pop esi
+; pop edx ; need not be preserved
+; pop ecx ; need not be preserved
+ pop ebx
+ pop ebp
+ ret
+
+; --------------------------------------------------------------------------
+;
+; Quantize/descale the coefficients, and store into coef_block
+;
+; This implementation is based on an algorithm described in
+; "How to optimize for the Pentium family of microprocessors"
+; (http://www.agner.org/assem/).
+;
+; GLOBAL(void)
+; jsimd_quantize_mmx (JCOEFPTR coef_block, DCTELEM * divisors,
+; DCTELEM * workspace);
+;
+
+%define RECIPROCAL(m,n,b) MMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM)
+%define CORRECTION(m,n,b) MMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM)
+%define SCALE(m,n,b) MMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM)
+%define SHIFT(m,n,b) MMBLOCK(DCTSIZE*3+(m),(n),(b),SIZEOF_DCTELEM)
+
+%define coef_block ebp+8 ; JCOEFPTR coef_block
+%define divisors ebp+12 ; DCTELEM * divisors
+%define workspace ebp+16 ; DCTELEM * workspace
+
+ align 16
+ global EXTN(jsimd_quantize_mmx)
+
+EXTN(jsimd_quantize_mmx):
+ push ebp
+ mov ebp,esp
+; push ebx ; unused
+; push ecx ; unused
+; push edx ; need not be preserved
+ push esi
+ push edi
+
+ mov esi, POINTER [workspace]
+ mov edx, POINTER [divisors]
+ mov edi, JCOEFPTR [coef_block]
+ mov ah, 2
+ alignx 16,7
+.quantloop1:
+ mov al, DCTSIZE2/8/2
+ alignx 16,7
+.quantloop2:
+ movq mm2, MMWORD [MMBLOCK(0,0,esi,SIZEOF_DCTELEM)]
+ movq mm3, MMWORD [MMBLOCK(0,1,esi,SIZEOF_DCTELEM)]
+
+ movq mm0,mm2
+ movq mm1,mm3
+
+ psraw mm2,(WORD_BIT-1) ; -1 if value < 0, 0 otherwise
+ psraw mm3,(WORD_BIT-1)
+
+ pxor mm0,mm2 ; val = -val
+ pxor mm1,mm3
+ psubw mm0,mm2
+ psubw mm1,mm3
+
+ ;
+ ; MMX is an annoyingly crappy instruction set. It has two
+ ; misfeatures that are causing problems here:
+ ;
+ ; - All multiplications are signed.
+ ;
+ ; - The second operand for the shifts is not treated as packed.
+ ;
+ ;
+ ; We work around the first problem by implementing this algorithm:
+ ;
+ ; unsigned long unsigned_multiply(unsigned short x, unsigned short y)
+ ; {
+ ; enum { SHORT_BIT = 16 };
+ ; signed short sx = (signed short) x;
+ ; signed short sy = (signed short) y;
+ ; signed long sz;
+ ;
+ ; sz = (long) sx * (long) sy; /* signed multiply */
+ ;
+ ; if (sx < 0) sz += (long) sy << SHORT_BIT;
+ ; if (sy < 0) sz += (long) sx << SHORT_BIT;
+ ;
+ ; return (unsigned long) sz;
+ ; }
+ ;
+ ; (note that a negative sx adds _sy_ and vice versa)
+ ;
+ ; For the second problem, we replace the shift by a multiplication.
+ ; Unfortunately that means we have to deal with the signed issue again.
+ ;
+
+ paddw mm0, MMWORD [CORRECTION(0,0,edx)] ; correction + roundfactor
+ paddw mm1, MMWORD [CORRECTION(0,1,edx)]
+
+ movq mm4,mm0 ; store current value for later
+ movq mm5,mm1
+ pmulhw mm0, MMWORD [RECIPROCAL(0,0,edx)] ; reciprocal
+ pmulhw mm1, MMWORD [RECIPROCAL(0,1,edx)]
+ paddw mm0,mm4 ; reciprocal is always negative (MSB=1),
+ paddw mm1,mm5 ; so we always need to add the initial value
+ ; (input value is never negative as we
+ ; inverted it at the start of this routine)
+
+ ; here it gets a bit tricky as both scale
+ ; and mm0/mm1 can be negative
+ movq mm6, MMWORD [SCALE(0,0,edx)] ; scale
+ movq mm7, MMWORD [SCALE(0,1,edx)]
+ movq mm4,mm0
+ movq mm5,mm1
+ pmulhw mm0,mm6
+ pmulhw mm1,mm7
+
+ psraw mm6,(WORD_BIT-1) ; determine if scale is negative
+ psraw mm7,(WORD_BIT-1)
+
+ pand mm6,mm4 ; and add input if it is
+ pand mm7,mm5
+ paddw mm0,mm6
+ paddw mm1,mm7
+
+ psraw mm4,(WORD_BIT-1) ; then check if negative input
+ psraw mm5,(WORD_BIT-1)
+
+ pand mm4, MMWORD [SCALE(0,0,edx)] ; and add scale if it is
+ pand mm5, MMWORD [SCALE(0,1,edx)]
+ paddw mm0,mm4
+ paddw mm1,mm5
+
+ pxor mm0,mm2 ; val = -val
+ pxor mm1,mm3
+ psubw mm0,mm2
+ psubw mm1,mm3
+
+ movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_DCTELEM)], mm0
+ movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_DCTELEM)], mm1
+
+ add esi, byte 8*SIZEOF_DCTELEM
+ add edx, byte 8*SIZEOF_DCTELEM
+ add edi, byte 8*SIZEOF_JCOEF
+ dec al
+ jnz near .quantloop2
+ dec ah
+ jnz near .quantloop1 ; to avoid branch misprediction
+
+ emms ; empty MMX state
+
+ pop edi
+ pop esi
+; pop edx ; need not be preserved
+; pop ecx ; unused
+; pop ebx ; unused
+ pop ebp
+ ret
+