[linux.git] / lib / raid6 / sse1.c

/* -*- linux-c -*- ------------------------------------------------------- *
 *
 *   Copyright 2002 H. Peter Anvin - All Rights Reserved
 *
 *   This program is free software; you can redistribute it and/or modify
 *   it under the terms of the GNU General Public License as published by
 *   the Free Software Foundation, Inc., 53 Temple Place Ste 330,
 *   Boston MA 02111-1307, USA; either version 2 of the License, or
 *   (at your option) any later version; incorporated herein by reference.
 *
 * ----------------------------------------------------------------------- */

/*
 * raid6/sse1.c
 *
 * SSE-1/MMXEXT implementation of RAID-6 syndrome functions
 *
 * This is really an MMX implementation, but it requires SSE-1 or
 * AMD MMXEXT for prefetch support and a few other features.  The
 * support for nontemporal memory accesses is enough to make this
 * worthwhile as a separate implementation.
 */

#ifdef CONFIG_X86_32

#include <linux/raid/pq.h>
#include "x86.h"

/* Defined in raid6/mmx.c */
extern const struct raid6_mmx_constants {
	u64 x1d;
} raid6_mmx_constants;

static int raid6_have_sse1_or_mmxext(void)
{
	/* Not really boot_cpu but "all_cpus" */
	return boot_cpu_has(X86_FEATURE_MMX) &&
		(boot_cpu_has(X86_FEATURE_XMM) ||
		 boot_cpu_has(X86_FEATURE_MMXEXT));
}

/*
 * Plain SSE1 implementation
 */
static void raid6_sse11_gen_syndrome(int disks, size_t bytes, void **ptrs)
{
	u8 **dptr = (u8 **)ptrs;
	u8 *p, *q;
	int d, z, z0;

	z0 = disks - 3;		/* Highest data disk */
	p = dptr[z0+1];		/* XOR parity */
	q = dptr[z0+2];		/* RS syndrome */

	kernel_fpu_begin();

	asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
	asm volatile("pxor %mm5,%mm5");	/* Zero temp */

	for ( d = 0 ; d < bytes ; d += 8 ) {
		asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
		asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
		asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d]));
		asm volatile("movq %mm2,%mm4");	/* Q[0] */
		asm volatile("movq %0,%%mm6" : : "m" (dptr[z0-1][d]));
		for ( z = z0-2 ; z >= 0 ; z-- ) {
			asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
			asm volatile("pcmpgtb %mm4,%mm5");
			asm volatile("paddb %mm4,%mm4");
			asm volatile("pand %mm0,%mm5");
			asm volatile("pxor %mm5,%mm4");
			asm volatile("pxor %mm5,%mm5");
			asm volatile("pxor %mm6,%mm2");
			asm volatile("pxor %mm6,%mm4");
			asm volatile("movq %0,%%mm6" : : "m" (dptr[z][d]));
		}
		asm volatile("pcmpgtb %mm4,%mm5");
		asm volatile("paddb %mm4,%mm4");
		asm volatile("pand %mm0,%mm5");
		asm volatile("pxor %mm5,%mm4");
		asm volatile("pxor %mm5,%mm5");
		asm volatile("pxor %mm6,%mm2");
		asm volatile("pxor %mm6,%mm4");

		asm volatile("movntq %%mm2,%0" : "=m" (p[d]));
		asm volatile("movntq %%mm4,%0" : "=m" (q[d]));
	}

	asm volatile("sfence" : : : "memory");
	kernel_fpu_end();
}

const struct raid6_calls raid6_sse1x1 = {
	raid6_sse11_gen_syndrome,
	NULL,			/* XOR not yet implemented */
	raid6_have_sse1_or_mmxext,
	"sse1x1",
	1			/* Has cache hints */
};

/*
 * Unrolled-by-2 SSE1 implementation
 */
static void raid6_sse12_gen_syndrome(int disks, size_t bytes, void **ptrs)
{
	u8 **dptr = (u8 **)ptrs;
	u8 *p, *q;
	int d, z, z0;

	z0 = disks - 3;		/* Highest data disk */
	p = dptr[z0+1];		/* XOR parity */
	q = dptr[z0+2];		/* RS syndrome */

	kernel_fpu_begin();

	asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
	asm volatile("pxor %mm5,%mm5");	/* Zero temp */
	asm volatile("pxor %mm7,%mm7"); /* Zero temp */

	/* We uniformly assume a single prefetch covers at least 16 bytes */
	for ( d = 0 ; d < bytes ; d += 16 ) {
		asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
		asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
		asm volatile("movq %0,%%mm3" : : "m" (dptr[z0][d+8])); /* P[1] */
		asm volatile("movq %mm2,%mm4");	/* Q[0] */
		asm volatile("movq %mm3,%mm6"); /* Q[1] */
		for ( z = z0-1 ; z >= 0 ; z-- ) {
			asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
			asm volatile("pcmpgtb %mm4,%mm5");
			asm volatile("pcmpgtb %mm6,%mm7");
			asm volatile("paddb %mm4,%mm4");
			asm volatile("paddb %mm6,%mm6");
			asm volatile("pand %mm0,%mm5");
			asm volatile("pand %mm0,%mm7");
			asm volatile("pxor %mm5,%mm4");
			asm volatile("pxor %mm7,%mm6");
			asm volatile("movq %0,%%mm5" : : "m" (dptr[z][d]));
			asm volatile("movq %0,%%mm7" : : "m" (dptr[z][d+8]));
			asm volatile("pxor %mm5,%mm2");
			asm volatile("pxor %mm7,%mm3");
			asm volatile("pxor %mm5,%mm4");
			asm volatile("pxor %mm7,%mm6");
			asm volatile("pxor %mm5,%mm5");
			asm volatile("pxor %mm7,%mm7");
		}
		asm volatile("movntq %%mm2,%0" : "=m" (p[d]));
		asm volatile("movntq %%mm3,%0" : "=m" (p[d+8]));
		asm volatile("movntq %%mm4,%0" : "=m" (q[d]));
		asm volatile("movntq %%mm6,%0" : "=m" (q[d+8]));
	}

	asm volatile("sfence" : :: "memory");
	kernel_fpu_end();
}

const struct raid6_calls raid6_sse1x2 = {
	raid6_sse12_gen_syndrome,
	NULL,			/* XOR not yet implemented */
	raid6_have_sse1_or_mmxext,
	"sse1x2",
	1			/* Has cache hints */
};

#endif
Commit	Line	Data
1da177e4 LT	1	/* -- linux-c -- ------------------------------------------------------- *
	2	*
	3	* Copyright 2002 H. Peter Anvin - All Rights Reserved
	4	*
	5	* This program is free software; you can redistribute it and/or modify
	6	* it under the terms of the GNU General Public License as published by
	7	* the Free Software Foundation, Inc., 53 Temple Place Ste 330,
93ed05e2	8	* Boston MA 02111-1307, USA; either version 2 of the License, or
1da177e4 LT	9	* (at your option) any later version; incorporated herein by reference.
	10	*
	11	* ----------------------------------------------------------------------- */
	12
	13	/*
a8e026c7	14	* raid6/sse1.c
1da177e4 LT	15	*
	16	* SSE-1/MMXEXT implementation of RAID-6 syndrome functions
	17	*
	18	* This is really an MMX implementation, but it requires SSE-1 or
	19	* AMD MMXEXT for prefetch support and a few other features. The
	20	* support for nontemporal memory accesses is enough to make this
	21	* worthwhile as a separate implementation.
	22	*/
	23
4f8c55c5	24	#ifdef CONFIG_X86_32
1da177e4	25
f701d589	26	#include <linux/raid/pq.h>
a8e026c7	27	#include "x86.h"
1da177e4	28
a8e026c7	29	/* Defined in raid6/mmx.c */
1da177e4 LT	30	extern const struct raid6_mmx_constants {
	31	u64 x1d;
	32	} raid6_mmx_constants;
	33
	34	static int raid6_have_sse1_or_mmxext(void)
	35	{
1da177e4 LT	36	/* Not really boot_cpu but "all_cpus" */
	37	return boot_cpu_has(X86_FEATURE_MMX) &&
	38	(boot_cpu_has(X86_FEATURE_XMM) \|\|
	39	boot_cpu_has(X86_FEATURE_MMXEXT));
1da177e4 LT	40	}
	41
	42	/*
	43	* Plain SSE1 implementation
	44	*/
	45	static void raid6_sse11_gen_syndrome(int disks, size_t bytes, void **ptrs)
	46	{
	47	u8 dptr = (u8 )ptrs;
	48	u8 p, q;
	49	int d, z, z0;
1da177e4 LT	50
	51	z0 = disks - 3; /* Highest data disk */
	52	p = dptr[z0+1]; /* XOR parity */
	53	q = dptr[z0+2]; /* RS syndrome */
	54
a723406c	55	kernel_fpu_begin();
1da177e4 LT	56
	57	asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
	58	asm volatile("pxor %mm5,%mm5"); /* Zero temp */
	59
	60	for ( d = 0 ; d < bytes ; d += 8 ) {
	61	asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
	62	asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
	63	asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d]));
	64	asm volatile("movq %mm2,%mm4"); /* Q[0] */
	65	asm volatile("movq %0,%%mm6" : : "m" (dptr[z0-1][d]));
	66	for ( z = z0-2 ; z >= 0 ; z-- ) {
	67	asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
	68	asm volatile("pcmpgtb %mm4,%mm5");
	69	asm volatile("paddb %mm4,%mm4");
	70	asm volatile("pand %mm0,%mm5");
	71	asm volatile("pxor %mm5,%mm4");
	72	asm volatile("pxor %mm5,%mm5");
	73	asm volatile("pxor %mm6,%mm2");
	74	asm volatile("pxor %mm6,%mm4");
	75	asm volatile("movq %0,%%mm6" : : "m" (dptr[z][d]));
	76	}
	77	asm volatile("pcmpgtb %mm4,%mm5");
	78	asm volatile("paddb %mm4,%mm4");
	79	asm volatile("pand %mm0,%mm5");
	80	asm volatile("pxor %mm5,%mm4");
	81	asm volatile("pxor %mm5,%mm5");
	82	asm volatile("pxor %mm6,%mm2");
	83	asm volatile("pxor %mm6,%mm4");
	84
	85	asm volatile("movntq %%mm2,%0" : "=m" (p[d]));
	86	asm volatile("movntq %%mm4,%0" : "=m" (q[d]));
	87	}
	88
1da177e4	89	asm volatile("sfence" : : : "memory");
a723406c	90	kernel_fpu_end();
1da177e4 LT	91	}
	92
	93	const struct raid6_calls raid6_sse1x1 = {
	94	raid6_sse11_gen_syndrome,
fe5cbc6e	95	NULL, /* XOR not yet implemented */
1da177e4 LT	96	raid6_have_sse1_or_mmxext,
	97	"sse1x1",
	98	1 /* Has cache hints */
	99	};
	100
	101	/*
	102	* Unrolled-by-2 SSE1 implementation
	103	*/
	104	static void raid6_sse12_gen_syndrome(int disks, size_t bytes, void **ptrs)
	105	{
	106	u8 dptr = (u8 )ptrs;
	107	u8 p, q;
	108	int d, z, z0;
1da177e4 LT	109
	110	z0 = disks - 3; /* Highest data disk */
	111	p = dptr[z0+1]; /* XOR parity */
	112	q = dptr[z0+2]; /* RS syndrome */
	113
a723406c	114	kernel_fpu_begin();
1da177e4 LT	115
	116	asm volatile("movq %0,%%mm0" : : "m" (raid6_mmx_constants.x1d));
	117	asm volatile("pxor %mm5,%mm5"); /* Zero temp */
	118	asm volatile("pxor %mm7,%mm7"); /* Zero temp */
	119
	120	/* We uniformly assume a single prefetch covers at least 16 bytes */
	121	for ( d = 0 ; d < bytes ; d += 16 ) {
	122	asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
	123	asm volatile("movq %0,%%mm2" : : "m" (dptr[z0][d])); /* P[0] */
	124	asm volatile("movq %0,%%mm3" : : "m" (dptr[z0][d+8])); /* P[1] */
	125	asm volatile("movq %mm2,%mm4"); /* Q[0] */
	126	asm volatile("movq %mm3,%mm6"); /* Q[1] */
	127	for ( z = z0-1 ; z >= 0 ; z-- ) {
	128	asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
	129	asm volatile("pcmpgtb %mm4,%mm5");
	130	asm volatile("pcmpgtb %mm6,%mm7");
	131	asm volatile("paddb %mm4,%mm4");
	132	asm volatile("paddb %mm6,%mm6");
	133	asm volatile("pand %mm0,%mm5");
	134	asm volatile("pand %mm0,%mm7");
	135	asm volatile("pxor %mm5,%mm4");
	136	asm volatile("pxor %mm7,%mm6");
	137	asm volatile("movq %0,%%mm5" : : "m" (dptr[z][d]));
	138	asm volatile("movq %0,%%mm7" : : "m" (dptr[z][d+8]));
	139	asm volatile("pxor %mm5,%mm2");
	140	asm volatile("pxor %mm7,%mm3");
	141	asm volatile("pxor %mm5,%mm4");
	142	asm volatile("pxor %mm7,%mm6");
	143	asm volatile("pxor %mm5,%mm5");
	144	asm volatile("pxor %mm7,%mm7");
	145	}
	146	asm volatile("movntq %%mm2,%0" : "=m" (p[d]));
	147	asm volatile("movntq %%mm3,%0" : "=m" (p[d+8]));
	148	asm volatile("movntq %%mm4,%0" : "=m" (q[d]));
	149	asm volatile("movntq %%mm6,%0" : "=m" (q[d+8]));
	150	}
	151
1da177e4	152	asm volatile("sfence" : :: "memory");
a723406c	153	kernel_fpu_end();
1da177e4 LT	154	}
	155
	156	const struct raid6_calls raid6_sse1x2 = {
	157	raid6_sse12_gen_syndrome,
fe5cbc6e	158	NULL, /* XOR not yet implemented */
1da177e4 LT	159	raid6_have_sse1_or_mmxext,
	160	"sse1x2",
	161	1 /* Has cache hints */
	162	};
	163
	164	#endif