• file
• revisions
• annotate
• diff
• comparison
• raw

media/libjpeg/simd/jfsseflt.asm@6474c204b198 (annotated)

media/libjpeg/simd/jfsseflt.asm

Wed, 31 Dec 2014 06:09:35 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Wed, 31 Dec 2014 06:09:35 +0100

changeset 0

6474c204b198

permissions

-rw-r--r--

Cloned upstream origin tor-browser at tor-browser-31.3.0esr-4.5-1-build1
revision ID fc1c9ff7c1b2defdbc039f12214767608f46423f for hacking purpose.

michael@0	1	;
michael@0	2	; jfsseflt.asm - floating-point FDCT (SSE)
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	; This file contains a floating-point implementation of the forward DCT
michael@0	18	; (Discrete Cosine Transform). The following code is based directly on
michael@0	19	; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
michael@0	20	;
michael@0	21	; [TAB8]
michael@0	22
michael@0	23	%include "jsimdext.inc"
michael@0	24	%include "jdct.inc"
michael@0	25
michael@0	26	; --------------------------------------------------------------------------
michael@0	27
michael@0	28	%macro unpcklps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
michael@0	29	shufps %1,%2,0x44
michael@0	30	%endmacro
michael@0	31
michael@0	32	%macro unpckhps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
michael@0	33	shufps %1,%2,0xEE
michael@0	34	%endmacro
michael@0	35
michael@0	36	; --------------------------------------------------------------------------
michael@0	37	SECTION SEG_CONST
michael@0	38
michael@0	39	alignz 16
michael@0	40	global EXTN(jconst_fdct_float_sse)
michael@0	41
michael@0	42	EXTN(jconst_fdct_float_sse):
michael@0	43
michael@0	44	PD_0_382 times 4 dd 0.382683432365089771728460
michael@0	45	PD_0_707 times 4 dd 0.707106781186547524400844
michael@0	46	PD_0_541 times 4 dd 0.541196100146196984399723
michael@0	47	PD_1_306 times 4 dd 1.306562964876376527856643
michael@0	48
michael@0	49	alignz 16
michael@0	50
michael@0	51	; --------------------------------------------------------------------------
michael@0	52	SECTION SEG_TEXT
michael@0	53	BITS 32
michael@0	54	;
michael@0	55	; Perform the forward DCT on one block of samples.
michael@0	56	;
michael@0	57	; GLOBAL(void)
michael@0	58	; jsimd_fdct_float_sse (FAST_FLOAT * data)
michael@0	59	;
michael@0	60
michael@0	61	%define data(b) (b)+8 ; FAST_FLOAT * data
michael@0	62
michael@0	63	%define original_ebp ebp+0
michael@0	64	%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
michael@0	65	%define WK_NUM 2
michael@0	66
michael@0	67	align 16
michael@0	68	global EXTN(jsimd_fdct_float_sse)
michael@0	69
michael@0	70	EXTN(jsimd_fdct_float_sse):
michael@0	71	push ebp
michael@0	72	mov eax,esp ; eax = original ebp
michael@0	73	sub esp, byte 4
michael@0	74	and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
michael@0	75	mov [esp],eax
michael@0	76	mov ebp,esp ; ebp = aligned ebp
michael@0	77	lea esp, [wk(0)]
michael@0	78	pushpic ebx
michael@0	79	; push ecx ; need not be preserved
michael@0	80	; push edx ; need not be preserved
michael@0	81	; push esi ; unused
michael@0	82	; push edi ; unused
michael@0	83
michael@0	84	get_GOT ebx ; get GOT address
michael@0	85
michael@0	86	; ---- Pass 1: process rows.
michael@0	87
michael@0	88	mov edx, POINTER [data(eax)] ; (FAST_FLOAT *)
michael@0	89	mov ecx, DCTSIZE/4
michael@0	90	alignx 16,7
michael@0	91	.rowloop:
michael@0	92
michael@0	93	movaps xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	94	movaps xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	95	movaps xmm2, XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)]
michael@0	96	movaps xmm3, XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)]
michael@0	97
michael@0	98	; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
michael@0	99	; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
michael@0	100
michael@0	101	movaps xmm4,xmm0 ; transpose coefficients(phase 1)
michael@0	102	unpcklps xmm0,xmm1 ; xmm0=(20 30 21 31)
michael@0	103	unpckhps xmm4,xmm1 ; xmm4=(22 32 23 33)
michael@0	104	movaps xmm5,xmm2 ; transpose coefficients(phase 1)
michael@0	105	unpcklps xmm2,xmm3 ; xmm2=(24 34 25 35)
michael@0	106	unpckhps xmm5,xmm3 ; xmm5=(26 36 27 37)
michael@0	107
michael@0	108	movaps xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	109	movaps xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	110	movaps xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
michael@0	111	movaps xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
michael@0	112
michael@0	113	; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
michael@0	114	; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
michael@0	115
michael@0	116	movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 32 23 33)
michael@0	117	movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(24 34 25 35)
michael@0	118
michael@0	119	movaps xmm4,xmm6 ; transpose coefficients(phase 1)
michael@0	120	unpcklps xmm6,xmm7 ; xmm6=(00 10 01 11)
michael@0	121	unpckhps xmm4,xmm7 ; xmm4=(02 12 03 13)
michael@0	122	movaps xmm2,xmm1 ; transpose coefficients(phase 1)
michael@0	123	unpcklps xmm1,xmm3 ; xmm1=(04 14 05 15)
michael@0	124	unpckhps xmm2,xmm3 ; xmm2=(06 16 07 17)
michael@0	125
michael@0	126	movaps xmm7,xmm6 ; transpose coefficients(phase 2)
michael@0	127	unpcklps2 xmm6,xmm0 ; xmm6=(00 10 20 30)=data0
michael@0	128	unpckhps2 xmm7,xmm0 ; xmm7=(01 11 21 31)=data1
michael@0	129	movaps xmm3,xmm2 ; transpose coefficients(phase 2)
michael@0	130	unpcklps2 xmm2,xmm5 ; xmm2=(06 16 26 36)=data6
michael@0	131	unpckhps2 xmm3,xmm5 ; xmm3=(07 17 27 37)=data7
michael@0	132
michael@0	133	movaps xmm0,xmm7
michael@0	134	movaps xmm5,xmm6
michael@0	135	subps xmm7,xmm2 ; xmm7=data1-data6=tmp6
michael@0	136	subps xmm6,xmm3 ; xmm6=data0-data7=tmp7
michael@0	137	addps xmm0,xmm2 ; xmm0=data1+data6=tmp1
michael@0	138	addps xmm5,xmm3 ; xmm5=data0+data7=tmp0
michael@0	139
michael@0	140	movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 32 23 33)
michael@0	141	movaps xmm3, XMMWORD [wk(1)] ; xmm3=(24 34 25 35)
michael@0	142	movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6
michael@0	143	movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
michael@0	144
michael@0	145	movaps xmm7,xmm4 ; transpose coefficients(phase 2)
michael@0	146	unpcklps2 xmm4,xmm2 ; xmm4=(02 12 22 32)=data2
michael@0	147	unpckhps2 xmm7,xmm2 ; xmm7=(03 13 23 33)=data3
michael@0	148	movaps xmm6,xmm1 ; transpose coefficients(phase 2)
michael@0	149	unpcklps2 xmm1,xmm3 ; xmm1=(04 14 24 34)=data4
michael@0	150	unpckhps2 xmm6,xmm3 ; xmm6=(05 15 25 35)=data5
michael@0	151
michael@0	152	movaps xmm2,xmm7
michael@0	153	movaps xmm3,xmm4
michael@0	154	addps xmm7,xmm1 ; xmm7=data3+data4=tmp3
michael@0	155	addps xmm4,xmm6 ; xmm4=data2+data5=tmp2
michael@0	156	subps xmm2,xmm1 ; xmm2=data3-data4=tmp4
michael@0	157	subps xmm3,xmm6 ; xmm3=data2-data5=tmp5
michael@0	158
michael@0	159	; -- Even part
michael@0	160
michael@0	161	movaps xmm1,xmm5
michael@0	162	movaps xmm6,xmm0
michael@0	163	subps xmm5,xmm7 ; xmm5=tmp13
michael@0	164	subps xmm0,xmm4 ; xmm0=tmp12
michael@0	165	addps xmm1,xmm7 ; xmm1=tmp10
michael@0	166	addps xmm6,xmm4 ; xmm6=tmp11
michael@0	167
michael@0	168	addps xmm0,xmm5
michael@0	169	mulps xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1
michael@0	170
michael@0	171	movaps xmm7,xmm1
michael@0	172	movaps xmm4,xmm5
michael@0	173	subps xmm1,xmm6 ; xmm1=data4
michael@0	174	subps xmm5,xmm0 ; xmm5=data6
michael@0	175	addps xmm7,xmm6 ; xmm7=data0
michael@0	176	addps xmm4,xmm0 ; xmm4=data2
michael@0	177
michael@0	178	movaps XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], xmm1
michael@0	179	movaps XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)], xmm5
michael@0	180	movaps XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7
michael@0	181	movaps XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4
michael@0	182
michael@0	183	; -- Odd part
michael@0	184
michael@0	185	movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6
michael@0	186	movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7
michael@0	187
michael@0	188	addps xmm2,xmm3 ; xmm2=tmp10
michael@0	189	addps xmm3,xmm6 ; xmm3=tmp11
michael@0	190	addps xmm6,xmm0 ; xmm6=tmp12, xmm0=tmp7
michael@0	191
michael@0	192	mulps xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3
michael@0	193
michael@0	194	movaps xmm1,xmm2 ; xmm1=tmp10
michael@0	195	subps xmm2,xmm6
michael@0	196	mulps xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5
michael@0	197	mulps xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
michael@0	198	mulps xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
michael@0	199	addps xmm1,xmm2 ; xmm1=z2
michael@0	200	addps xmm6,xmm2 ; xmm6=z4
michael@0	201
michael@0	202	movaps xmm5,xmm0
michael@0	203	subps xmm0,xmm3 ; xmm0=z13
michael@0	204	addps xmm5,xmm3 ; xmm5=z11
michael@0	205
michael@0	206	movaps xmm7,xmm0
michael@0	207	movaps xmm4,xmm5
michael@0	208	subps xmm0,xmm1 ; xmm0=data3
michael@0	209	subps xmm5,xmm6 ; xmm5=data7
michael@0	210	addps xmm7,xmm1 ; xmm7=data5
michael@0	211	addps xmm4,xmm6 ; xmm4=data1
michael@0	212
michael@0	213	movaps XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0
michael@0	214	movaps XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)], xmm5
michael@0	215	movaps XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], xmm7
michael@0	216	movaps XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4
michael@0	217
michael@0	218	add edx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
michael@0	219	dec ecx
michael@0	220	jnz near .rowloop
michael@0	221
michael@0	222	; ---- Pass 2: process columns.
michael@0	223
michael@0	224	mov edx, POINTER [data(eax)] ; (FAST_FLOAT *)
michael@0	225	mov ecx, DCTSIZE/4
michael@0	226	alignx 16,7
michael@0	227	.columnloop:
michael@0	228
michael@0	229	movaps xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	230	movaps xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	231	movaps xmm2, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	232	movaps xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	233
michael@0	234	; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
michael@0	235	; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
michael@0	236
michael@0	237	movaps xmm4,xmm0 ; transpose coefficients(phase 1)
michael@0	238	unpcklps xmm0,xmm1 ; xmm0=(02 03 12 13)
michael@0	239	unpckhps xmm4,xmm1 ; xmm4=(22 23 32 33)
michael@0	240	movaps xmm5,xmm2 ; transpose coefficients(phase 1)
michael@0	241	unpcklps xmm2,xmm3 ; xmm2=(42 43 52 53)
michael@0	242	unpckhps xmm5,xmm3 ; xmm5=(62 63 72 73)
michael@0	243
michael@0	244	movaps xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	245	movaps xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	246	movaps xmm1, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	247	movaps xmm3, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)]
michael@0	248
michael@0	249	; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
michael@0	250	; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
michael@0	251
michael@0	252	movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 23 32 33)
michael@0	253	movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(42 43 52 53)
michael@0	254
michael@0	255	movaps xmm4,xmm6 ; transpose coefficients(phase 1)
michael@0	256	unpcklps xmm6,xmm7 ; xmm6=(00 01 10 11)
michael@0	257	unpckhps xmm4,xmm7 ; xmm4=(20 21 30 31)
michael@0	258	movaps xmm2,xmm1 ; transpose coefficients(phase 1)
michael@0	259	unpcklps xmm1,xmm3 ; xmm1=(40 41 50 51)
michael@0	260	unpckhps xmm2,xmm3 ; xmm2=(60 61 70 71)
michael@0	261
michael@0	262	movaps xmm7,xmm6 ; transpose coefficients(phase 2)
michael@0	263	unpcklps2 xmm6,xmm0 ; xmm6=(00 01 02 03)=data0
michael@0	264	unpckhps2 xmm7,xmm0 ; xmm7=(10 11 12 13)=data1
michael@0	265	movaps xmm3,xmm2 ; transpose coefficients(phase 2)
michael@0	266	unpcklps2 xmm2,xmm5 ; xmm2=(60 61 62 63)=data6
michael@0	267	unpckhps2 xmm3,xmm5 ; xmm3=(70 71 72 73)=data7
michael@0	268
michael@0	269	movaps xmm0,xmm7
michael@0	270	movaps xmm5,xmm6
michael@0	271	subps xmm7,xmm2 ; xmm7=data1-data6=tmp6
michael@0	272	subps xmm6,xmm3 ; xmm6=data0-data7=tmp7
michael@0	273	addps xmm0,xmm2 ; xmm0=data1+data6=tmp1
michael@0	274	addps xmm5,xmm3 ; xmm5=data0+data7=tmp0
michael@0	275
michael@0	276	movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 23 32 33)
michael@0	277	movaps xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53)
michael@0	278	movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6
michael@0	279	movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
michael@0	280
michael@0	281	movaps xmm7,xmm4 ; transpose coefficients(phase 2)
michael@0	282	unpcklps2 xmm4,xmm2 ; xmm4=(20 21 22 23)=data2
michael@0	283	unpckhps2 xmm7,xmm2 ; xmm7=(30 31 32 33)=data3
michael@0	284	movaps xmm6,xmm1 ; transpose coefficients(phase 2)
michael@0	285	unpcklps2 xmm1,xmm3 ; xmm1=(40 41 42 43)=data4
michael@0	286	unpckhps2 xmm6,xmm3 ; xmm6=(50 51 52 53)=data5
michael@0	287
michael@0	288	movaps xmm2,xmm7
michael@0	289	movaps xmm3,xmm4
michael@0	290	addps xmm7,xmm1 ; xmm7=data3+data4=tmp3
michael@0	291	addps xmm4,xmm6 ; xmm4=data2+data5=tmp2
michael@0	292	subps xmm2,xmm1 ; xmm2=data3-data4=tmp4
michael@0	293	subps xmm3,xmm6 ; xmm3=data2-data5=tmp5
michael@0	294
michael@0	295	; -- Even part
michael@0	296
michael@0	297	movaps xmm1,xmm5
michael@0	298	movaps xmm6,xmm0
michael@0	299	subps xmm5,xmm7 ; xmm5=tmp13
michael@0	300	subps xmm0,xmm4 ; xmm0=tmp12
michael@0	301	addps xmm1,xmm7 ; xmm1=tmp10
michael@0	302	addps xmm6,xmm4 ; xmm6=tmp11
michael@0	303
michael@0	304	addps xmm0,xmm5
michael@0	305	mulps xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1
michael@0	306
michael@0	307	movaps xmm7,xmm1
michael@0	308	movaps xmm4,xmm5
michael@0	309	subps xmm1,xmm6 ; xmm1=data4
michael@0	310	subps xmm5,xmm0 ; xmm5=data6
michael@0	311	addps xmm7,xmm6 ; xmm7=data0
michael@0	312	addps xmm4,xmm0 ; xmm4=data2
michael@0	313
michael@0	314	movaps XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], xmm1
michael@0	315	movaps XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], xmm5
michael@0	316	movaps XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7
michael@0	317	movaps XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4
michael@0	318
michael@0	319	; -- Odd part
michael@0	320
michael@0	321	movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6
michael@0	322	movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7
michael@0	323
michael@0	324	addps xmm2,xmm3 ; xmm2=tmp10
michael@0	325	addps xmm3,xmm6 ; xmm3=tmp11
michael@0	326	addps xmm6,xmm0 ; xmm6=tmp12, xmm0=tmp7
michael@0	327
michael@0	328	mulps xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3
michael@0	329
michael@0	330	movaps xmm1,xmm2 ; xmm1=tmp10
michael@0	331	subps xmm2,xmm6
michael@0	332	mulps xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5
michael@0	333	mulps xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
michael@0	334	mulps xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
michael@0	335	addps xmm1,xmm2 ; xmm1=z2
michael@0	336	addps xmm6,xmm2 ; xmm6=z4
michael@0	337
michael@0	338	movaps xmm5,xmm0
michael@0	339	subps xmm0,xmm3 ; xmm0=z13
michael@0	340	addps xmm5,xmm3 ; xmm5=z11
michael@0	341
michael@0	342	movaps xmm7,xmm0
michael@0	343	movaps xmm4,xmm5
michael@0	344	subps xmm0,xmm1 ; xmm0=data3
michael@0	345	subps xmm5,xmm6 ; xmm5=data7
michael@0	346	addps xmm7,xmm1 ; xmm7=data5
michael@0	347	addps xmm4,xmm6 ; xmm4=data1
michael@0	348
michael@0	349	movaps XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0
michael@0	350	movaps XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], xmm5
michael@0	351	movaps XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], xmm7
michael@0	352	movaps XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4
michael@0	353
michael@0	354	add edx, byte 4*SIZEOF_FAST_FLOAT
michael@0	355	dec ecx
michael@0	356	jnz near .columnloop
michael@0	357
michael@0	358	; pop edi ; unused
michael@0	359	; pop esi ; unused
michael@0	360	; pop edx ; need not be preserved
michael@0	361	; pop ecx ; need not be preserved
michael@0	362	poppic ebx
michael@0	363	mov esp,ebp ; esp <- aligned ebp
michael@0	364	pop esp ; esp <- original ebp
michael@0	365	pop ebp
michael@0	366	ret
michael@0	367
michael@0	368	; For some reason, the OS X linker does not honor the request to align the
michael@0	369	; segment unless we do this.
michael@0	370	align 16