Source-Changes-HG archive
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]
[src/trunk]: src/common/lib/libc/arch/arm/string This is a working version of...
details: https://anonhg.NetBSD.org/src/rev/f4f4556a6d1f
branches: trunk
changeset: 783670:f4f4556a6d1f
user: matt <matt%NetBSD.org@localhost>
date: Thu Jan 03 09:34:44 2013 +0000
description:
This is a working version of memcpy implemented using NEON instructions.
Still needs tuning as it is still about 15% than the non-NEON version.
diffstat:
common/lib/libc/arch/arm/string/memcpy_neon.S | 277 ++++++++++++++++++++++++++
1 files changed, 277 insertions(+), 0 deletions(-)
diffs (281 lines):
diff -r ea690d1901b5 -r f4f4556a6d1f common/lib/libc/arch/arm/string/memcpy_neon.S
--- /dev/null Thu Jan 01 00:00:00 1970 +0000
+++ b/common/lib/libc/arch/arm/string/memcpy_neon.S Thu Jan 03 09:34:44 2013 +0000
@@ -0,0 +1,277 @@
+/*-
+ * Copyright (c) 2013 The NetBSD Foundation, Inc.
+ * All rights reserved.
+ *
+ * This code is derived from software contributed to The NetBSD Foundation
+ * by Matt Thomas of 3am Software Foundry.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
+ * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+ * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
+ * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+ * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+ * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ * POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <machine/asm.h>
+
+RCSID("$NetBSD: memcpy_neon.S,v 1.1 2013/01/03 09:34:44 matt Exp $")
+
+ .text
+ENTRY(memcpy)
+ teq r2, #0 /* 0 length? */
+ cmpne r0, r1 /* if not, does src == dst? */
+ RETc(eq) /* yes, (to either) return */
+
+ mov r3, r0 /* keep r0 unchanged */
+#if 0
+ cmp r2, #16 /* copy less than 8 bytes? */
+ bge .Ldst_aligner /* nope, do it the long way */
+
+1: ldrb ip, [r1], #1 /* load a byte from src */
+ subs r2, r2, #1 /* and more to transfer? */
+ strb ip, [r3], #1 /* save it to dst */
+ bne 1b /* yes, do next byte */
+ RET /* return */
+#endif
+
+.Ldst_aligner:
+ tst r3, #7 /* is dst pointer word aligned? */
+ beq .Lsrc_aligner /* yes, check src pointer */
+ /*
+ * Until the dst pointer is word aligned, read src and dst byte by
+ * byte until it is aligned or we've copied everything.
+ */
+ ldrb ip, [r1], #1 /* load a byte from src */
+ strb ip, [r3], #1 /* save the byte to dst */
+ subs r2, r2, #1 /* end of transfer? */
+ bne .Ldst_aligner /* no, try next byte */
+ RET /* yes, we're done! */
+
+.Lsrc_aligner:
+ push {r4-r5} /* save some registers */
+ add r4, r2, r3 /* keep a pointer to the end of src */
+ ands r5, r1, #7 /* get misalignment of src pointer */
+ beq .Lcongruent_main /* aligned, do it the fast way */
+
+ vdup.8 d1, r5 /* set offset for table */
+ rsb r5, r5, #8 /* calculate leftover of each word */
+ bic r1, r1, #7 /* dword align src pointer */
+
+ vldr d0, .Ltbl_value /* load table value */
+ vadd.u8 d0, d0, d1 /* add offset to it */
+
+ vld1.64 {d1}, [r1:64]! /* load a dword from src */
+
+ cmp r2, r5 /* do we already have enough? */
+ bgt .Lincongruent /* no, so read more */
+
+.Lincongruent_finish:
+ vtbl.8 d0, {d1-d2}, d0 /* merge last dwords */
+ cmp r2, #8 /* room for a full dword? */
+#ifdef __ARMEB__
+ vrev64.32 d0, d0 /* word swap to LE */
+#endif
+ blt .Lfinish /* no, write final partial dword */
+ vst1.32 {d0}, [r3:64] /* yes, write final full dword */
+ b .Ldone /* and we're done! */
+
+.Lincongruent:
+ vld1.64 {d2}, [r1:64]! /* load a dword */
+ cmp r2, #8 /* can we write a full dword? */
+ blt .Lincongruent_finish /* no, finish it. */
+ vtbl.8 d1, {d1-d2}, d0 /* reorder */
+ vst1.64 {d1}, [r3:64]! /* store a dword */
+ subs r2, r2, #8 /* have we written everything? */
+ beq .Ldone /* yes, we're done! */
+ vmov d1, d2 /* prepare for next dword */
+ tst r3, #63 /* are we 64-byte aligned? */
+ bne .Lincongruent /* no, load next dword */
+
+ /*
+ * We are now 64-byte aligneds so all writes should fill one or more
+ * cachelines. Even if d1 has 7 bytes cached, to write 32 bytes we
+ * still need to read 4 dwords (3 full dwords and 1 dword for that
+ * last byte).
+ */
+ cmp r2, #32 /* can we write 4 more dwords? */
+ blt .Lincongruent_dword /* no, handle dword by dword */
+ vld1.64 {d2-d5}, [r1:64]! /* read 4 dwords */
+ cmp r2, #64 /* can we write 4 more dwords? */
+ blt .Lincongruent_4dword /* no, handle it */
+
+1: vld1.64 {d7-d10}, [r1:64]! /* read 4 dwords */
+ vtbl.8 d1, {d1-d2}, d0 /* reorder */
+ vtbl.8 d2, {d2-d3}, d0 /* reorder */
+ vtbl.8 d3, {d3-d4}, d0 /* reorder */
+ vtbl.8 d4, {d4-d5}, d0 /* reorder */
+ vst1.64 {d1-d4}, [r3:64]! /* write 4 dwords */
+ vmov d6, d5 /* move out of the way the load */
+ cmp r2, #96 /* have 8+4 dwords to write? */
+ blt 2f /* no more data, skip the load */
+ vld1.64 {d2-d5}, [r1:64]! /* more data, load 4 dwords */
+2: vtbl.8 d6, {d6-d7}, d0 /* reorder */
+ vtbl.8 d7, {d7-d8}, d0 /* reorder */
+ vtbl.8 d8, {d8-d9}, d0 /* reorder */
+ vtbl.8 d9, {d9-d10}, d0 /* reorder */
+ vst1.64 {d6-d9}, [r3:64]! /* write 4 dwords */
+ subs r2, r2, #64
+ beq .Ldone
+ vmov d1, d10
+ cmp r2, #64
+ bge 1b
+
+ /*
+ * we have leftovers in d1 and new untranslated date in d2-d5.
+ */
+.Lincongruent_4dword:
+ cmp r2, #32
+ blt .Lincongruent_dword
+
+ vtbl.8 d1, {d1-d2}, d0 /* reorder */
+ vtbl.8 d2, {d2-d3}, d0 /* reorder */
+ vtbl.8 d3, {d3-d4}, d0 /* reorder */
+ vtbl.8 d4, {d4-d5}, d0 /* reorder */
+ vst1.64 {d1-d4}, [r3:64]! /* write 4 dwords */
+ vmov d1, d5 /* move leftovers */
+ subs r2, r2, #32
+ beq .Ldone
+
+.Lincongruent_dword:
+#if 0
+ cmp r2, r5 /* enough in leftovers? */
+ ble .Lincongruent_finish /* yes, finish it. */
+ vld1.64 {d2}, [r1:64]! /* load a dword */
+ cmp r2, #8 /* can we write a full dword? */
+ blt .Lincongruent_finish /* no, finish it. */
+ vtbl.8 d1, {d1-d2}, d0 /* reorder */
+ vst1.64 {d1}, [r3:64]! /* store a dword */
+ subs r2, r2, #8 /* have we written everything? */
+ beq .Ldone /* yes, we're done! */
+ b .Lincongruent_dword /* and go get it */
+#else
+ cmp r2, r5 /* are the bytes we have enough? */
+ ble .Lincongruent_finish /* yes, finish it. */
+ mov ip, r2 /* get remaining count */
+ bic ip, ip, #7 /* truncate to a dword */
+ rsb ip, ip, #32 /* subtract from 32 */
+ ands r2, r2, #7 /* count mod 8 */
+ add pc, pc, ip, lsl #1 /* and jump! */
+ nop
+ vld1.64 {d2}, [r1:64]! /* load a dword */
+ vtbl.8 d1, {d1-d2}, d0 /* reorder */
+ vst1.64 {d1}, [r3:64]! /* store a dword */
+ vmov d1, d2 /* prepare for next dword */
+ vld1.64 {d2}, [r1:64]! /* load a dword */
+ vtbl.8 d1, {d1-d2}, d0 /* reorder */
+ vst1.64 {d1}, [r3:64]! /* store a dword */
+ vmov d1, d2 /* prepare for next dword */
+ vld1.64 {d2}, [r1:64]! /* load a dword */
+ vtbl.8 d1, {d1-d2}, d0 /* reorder */
+ vst1.64 {d1}, [r3:64]! /* store a dword */
+ vmov d1, d2 /* prepare for next dword */
+ vld1.64 {d2}, [r1:64]! /* load a dword */
+ vtbl.8 d1, {d1-d2}, d0 /* reorder */
+ vst1.64 {d1}, [r3:64]! /* store a dword */
+ vmov d1, d2 /* prepare for next dword */
+ beq .Ldone
+ vld1.64 {d2}, [r1:64]! /* load a dword */
+ b .Lincongruent_finish /* write last partial dowrd */
+#endif
+
+.Lcongruent_main:
+ vld1.32 {d0}, [r1:64]! /* load next dword */
+ cmp r2, #8 /* compare current ptr against end */
+ blt .Lfinish /* greater so write final dword */
+ vst1.32 {d0}, [r3:64]! /* store dword */
+ subs r2, r2, #8 /* compare current ptr against end */
+ beq .Ldone /* equal? we're done! */
+ tst r3, #63 /* have we hit a 64-byte boundary? */
+ bne .Lcongruent_main /* no, write next word */
+
+ cmp r2, #64 /* can we write 4 dwords? */
+ blt .Lcongruent_loop /* no, this dword by dword */
+ vldm r1!, {d0-d7} /* load next 7 dwords */
+ cmp r2, #128 /* can we write 16 dwords */
+ blt 3f /* no, then deal with 8 dwords */
+
+ /*
+ * The following writes two 64-byte interleaving stores and loads.
+ */
+1: vldm r1!, {d8-d15} /* load next 8 dwords */
+ vstm r3!, {d0-d7} /* store 8 more dwords */
+ cmp r2, #192 /* can we write 16+8 dwords? */
+ blt 2f /* no, don't load the next 8 dwords */
+ vldm r1!, {d0-d7} /* yes, load next 8 dwords */
+2: vstm r3!, {d8-d15} /* store 8 more dwords */
+ sub r2, r2, #128 /* we just stored 16 (8+8) dwords */
+ beq .Ldone /* if 0, we're done! */
+ cmp r2, #128 /* can we write 16 dwords */
+ bge 1b /* yes, do it again */
+ cmp r2, #64 /* have we loaded 8 dwords? */
+ blt .Lcongruent_loop /* no, proceed to do it dword */
+
+ /*
+ * We now have 8 dwords we can write in d0-d7.
+ */
+3: vstm r3!, {d0-d7} /* store 8 more dwords */
+ subs r2, r2, #64 /* we wrote 8 dwords */
+ beq .Ldone /* if 0, we're done! */
+
+.Lcongruent_loop:
+ vld1.32 {d0}, [r1]! /* load dword from src */
+ cmp r2, #8 /* can we write a full dword? */
+ blt .Lfinish /* no, write last partial dword */
+.Lcongruent_loop_start:
+ vst1.32 {d0}, [r3]! /* store dword into dst */
+ subs r2, r2, #8 /* subtract it from length */
+ beq .Ldone /* if 0, we're done! */
+ vld1.32 {d0}, [r1]! /* load dword from src */
+ cmp r2, #8 /* can we write a full dword? */
+ bge .Lcongruent_loop_start /* yes, so do it */
+
+.Lfinish:
+ vmov r4, r5, d0 /* get last dword from NEON */
+ tst r2, #4 /* do we have at least 4 bytes left? */
+ strne r4, [r3], #4 /* store the 1st word */
+ movne r4, r5 /* move 2nd word into place */
+ tst r2, #2 /* do we have at least 2 bytes left? */
+#ifdef __ARMEB__
+ movne r4, r4, ror #16 /* yes, swap halfwords */
+#endif
+ strneh r4, [r3], #2 /* yes, store the halfword */
+#ifdef __ARMEL__
+ movne r4, r4, lsr #16 /* yes, discard just written bytes */
+#endif
+ tst r2, #1 /* do we have a final byte? */
+#ifdef __ARMEB__
+ movne r4, r4, lsr #24 /* yes, move MSB to LSB */
+#endif
+ strneb r4, [r3], #1 /* yes, store it */
+
+.Ldone:
+ pop {r4-r5} /* restore registers */
+ RET
+
+ .p2align 3
+.Ltbl_value:
+#ifdef __ARMEL__
+ .quad 0x0706050403020100
+#else
+ .quad 0x0001020304050607
+#endif
+END(memcpy)
Home |
Main Index |
Thread Index |
Old Index