svn commit: r315983 - in head/sys: conf libkern/x86
Bruce Evans
bde at FreeBSD.org
Sun Mar 26 10:31:50 UTC 2017
Author: bde
Date: Sun Mar 26 10:31:48 2017
New Revision: 315983
URL: https://svnweb.freebsd.org/changeset/base/315983
Log:
Use inline asm instead of unportable intrinsics for the SSE4 crc32
optimization.
This fixes building with gcc-4.2.1 (it doesn't support SSE4).
gas-2.17.50 [FreeBSD] supports SSE4 instructions, so this doesn't
need using .byte directives.
This fixes depending on host user headers in the kernel.
Fix user includes (don't depend on namespace pollution in <nmmintrin.h>
that is not included now).
The instrinsics had no advantages except to sometimes avoid compiler
pessimixations. clang understands them a bit better than inline asm,
and generates better looking code which also runs better for cem, but
for me it just at the same speed or slower by doing excessive
unrollowing in all the wrong places. gcc-4.2.1 also doesn't understand
what it is doing with unrolling, but with -O3 somehow it does more
unrolling that helps.
Reduce 1 of the the compiler pessimizations (copying a variable which
already satisfies an "rm" constraint in a good way by being in memory
and not used again, to different memory and accessing it there. Force
copying it to a register instead).
Try to optimize the inner loops significantly, so as to run at full
speed on smaller inputs. The algorithm is already very MD, and was
tuned for the throughput of 3 crc32 instructions per cycle found on
at least Sandybridge through Haswell. Now it is even more tuned for
this, so depends more on the compiler not rearranging or unrolling
things too much. The main inner loop for should have no difficulty
runing at full speed on these CPUs unless the compiler unrolls it too
much. However, the main inner loop wasn't even used for buffers smaller
than 24K. Now it is used for buffers larger than 384 bytes. Now it
is not so long, and the main outer loop is used more. The new
optimization is to try to arrange that the outer loop runs in parallel
with the next inner loop except for the final iteration; then reduce
the loop sizes significantly to take advantage of this.
Approved by: cem
Not tested in production by: bde
Modified:
head/sys/conf/files.amd64
head/sys/conf/files.i386
head/sys/libkern/x86/crc32_sse42.c
Modified: head/sys/conf/files.amd64
==============================================================================
--- head/sys/conf/files.amd64 Sun Mar 26 09:41:08 2017 (r315982)
+++ head/sys/conf/files.amd64 Sun Mar 26 10:31:48 2017 (r315983)
@@ -545,6 +545,9 @@ isa/syscons_isa.c optional sc
isa/vga_isa.c optional vga
kern/kern_clocksource.c standard
kern/link_elf_obj.c standard
+libkern/x86/crc32_sse42.c standard
+libkern/memmove.c standard
+libkern/memset.c standard
#
# IA32 binary support
#
@@ -602,14 +605,6 @@ compat/ndis/subr_pe.c optional ndisapi
compat/ndis/subr_usbd.c optional ndisapi pci
compat/ndis/winx64_wrap.S optional ndisapi pci
#
-crc32_sse42.o standard \
- dependency "$S/libkern/x86/crc32_sse42.c" \
- compile-with "${CC} -c ${CFLAGS:N-nostdinc} ${WERROR} ${PROF} -msse4 ${.IMPSRC}" \
- no-implicit-rule \
- clean "crc32_sse42.o"
-libkern/memmove.c standard
-libkern/memset.c standard
-#
# x86 real mode BIOS emulator, required by dpms/pci/vesa
#
compat/x86bios/x86bios.c optional x86bios | dpms | pci | vesa
Modified: head/sys/conf/files.i386
==============================================================================
--- head/sys/conf/files.i386 Sun Mar 26 09:41:08 2017 (r315982)
+++ head/sys/conf/files.i386 Sun Mar 26 10:31:48 2017 (r315983)
@@ -524,11 +524,6 @@ kern/kern_clocksource.c standard
kern/imgact_aout.c optional compat_aout
kern/imgact_gzip.c optional gzip
kern/subr_sfbuf.c standard
-crc32_sse42.o standard \
- dependency "$S/libkern/x86/crc32_sse42.c" \
- compile-with "${CC} -c ${CFLAGS:N-nostdinc} ${WERROR} ${PROF} -msse4 ${.IMPSRC}" \
- no-implicit-rule \
- clean "crc32_sse42.o"
libkern/divdi3.c standard
libkern/ffsll.c standard
libkern/flsll.c standard
@@ -539,6 +534,7 @@ libkern/qdivrem.c standard
libkern/ucmpdi2.c standard
libkern/udivdi3.c standard
libkern/umoddi3.c standard
+libkern/x86/crc32_sse42.c standard
i386/xbox/xbox.c optional xbox
i386/xbox/xboxfb.c optional xboxfb
dev/fb/boot_font.c optional xboxfb
Modified: head/sys/libkern/x86/crc32_sse42.c
==============================================================================
--- head/sys/libkern/x86/crc32_sse42.c Sun Mar 26 09:41:08 2017 (r315982)
+++ head/sys/libkern/x86/crc32_sse42.c Sun Mar 26 10:31:48 2017 (r315983)
@@ -31,14 +31,40 @@ __FBSDID("$FreeBSD$");
*/
#ifdef USERSPACE_TESTING
#include <stdint.h>
+#include <stdlib.h>
#else
#include <sys/param.h>
-#include <sys/kernel.h>
-#include <sys/libkern.h>
#include <sys/systm.h>
+#include <sys/kernel.h>
#endif
-#include <nmmintrin.h>
+static __inline uint32_t
+_mm_crc32_u8(uint32_t x, uint8_t y)
+{
+ /*
+ * clang (at least 3.9.[0-1]) pessimizes "rm" (y) and "m" (y)
+ * significantly and "r" (y) a lot by copying y to a different
+ * local variable (on the stack or in a register), so only use
+ * the latter. This costs a register and an instruction but
+ * not a uop.
+ */
+ __asm("crc32b %1,%0" : "+r" (x) : "r" (y));
+ return (x);
+}
+
+static __inline uint32_t
+_mm_crc32_u32(uint32_t x, uint32_t y)
+{
+ __asm("crc32l %1,%0" : "+r" (x) : "r" (y));
+ return (x);
+}
+
+static __inline uint64_t
+_mm_crc32_u64(uint64_t x, uint64_t y)
+{
+ __asm("crc32q %1,%0" : "+r" (x) : "r" (y));
+ return (x);
+}
/* CRC-32C (iSCSI) polynomial in reversed bit order. */
#define POLY 0x82f63b78
@@ -47,12 +73,18 @@ __FBSDID("$FreeBSD$");
* Block sizes for three-way parallel crc computation. LONG and SHORT must
* both be powers of two.
*/
-#define LONG 8192
-#define SHORT 256
+#define LONG 128
+#define SHORT 64
-/* Tables for hardware crc that shift a crc by LONG and SHORT zeros. */
+/*
+ * Tables for updating a crc for LONG, 2 * LONG, SHORT and 2 * SHORT bytes
+ * of value 0 later in the input stream, in the same way that the hardware
+ * would, but in software without calculating intermediate steps.
+ */
static uint32_t crc32c_long[4][256];
+static uint32_t crc32c_2long[4][256];
static uint32_t crc32c_short[4][256];
+static uint32_t crc32c_2short[4][256];
/*
* Multiply a matrix times a vector over the Galois field of two elements,
@@ -171,7 +203,9 @@ __attribute__((__constructor__))
crc32c_init_hw(void)
{
crc32c_zeros(crc32c_long, LONG);
+ crc32c_zeros(crc32c_2long, 2 * LONG);
crc32c_zeros(crc32c_short, SHORT);
+ crc32c_zeros(crc32c_2short, 2 * SHORT);
}
#ifdef _KERNEL
SYSINIT(crc32c_sse42, SI_SUB_LOCK, SI_ORDER_ANY, crc32c_init_hw, NULL);
@@ -190,7 +224,11 @@ sse42_crc32c(uint32_t crc, const unsigne
const size_t align = 4;
#endif
const unsigned char *next, *end;
- uint64_t crc0, crc1, crc2; /* need to be 64 bits for crc32q */
+#ifdef __amd64__
+ uint64_t crc0, crc1, crc2;
+#else
+ uint32_t crc0, crc1, crc2;
+#endif
next = buf;
crc0 = crc;
@@ -202,6 +240,7 @@ sse42_crc32c(uint32_t crc, const unsigne
len--;
}
+#if LONG > SHORT
/*
* Compute the crc on sets of LONG*3 bytes, executing three independent
* crc instructions, each on LONG bytes -- this is optimized for the
@@ -209,6 +248,7 @@ sse42_crc32c(uint32_t crc, const unsigne
* have a throughput of one crc per cycle, but a latency of three
* cycles.
*/
+ crc = 0;
while (len >= LONG * 3) {
crc1 = 0;
crc2 = 0;
@@ -229,16 +269,64 @@ sse42_crc32c(uint32_t crc, const unsigne
#endif
next += align;
} while (next < end);
- crc0 = crc32c_shift(crc32c_long, crc0) ^ crc1;
- crc0 = crc32c_shift(crc32c_long, crc0) ^ crc2;
+ /*-
+ * Update the crc. Try to do it in parallel with the inner
+ * loop. 'crc' is used to accumulate crc0 and crc1
+ * produced by the inner loop so that the next iteration
+ * of the loop doesn't depend on anything except crc2.
+ *
+ * The full expression for the update is:
+ * crc = S*S*S*crc + S*S*crc0 + S*crc1
+ * where the terms are polynomials modulo the CRC polynomial.
+ * We regroup this subtly as:
+ * crc = S*S * (S*crc + crc0) + S*crc1.
+ * This has an extra dependency which reduces possible
+ * parallelism for the expression, but it turns out to be
+ * best to intentionally delay evaluation of this expression
+ * so that it competes less with the inner loop.
+ *
+ * We also intentionally reduce parallelism by feedng back
+ * crc2 to the inner loop as crc0 instead of accumulating
+ * it in crc. This synchronizes the loop with crc update.
+ * CPU and/or compiler schedulers produced bad order without
+ * this.
+ *
+ * Shifts take about 12 cycles each, so 3 here with 2
+ * parallelizable take about 24 cycles and the crc update
+ * takes slightly longer. 8 dependent crc32 instructions
+ * can run in 24 cycles, so the 3-way blocking is worse
+ * than useless for sizes less than 8 * <word size> = 64
+ * on amd64. In practice, SHORT = 32 confirms these
+ * timing calculations by giving a small improvement
+ * starting at size 96. Then the inner loop takes about
+ * 12 cycles and the crc update about 24, but these are
+ * partly in parallel so the total time is less than the
+ * 36 cycles that 12 dependent crc32 instructions would
+ * take.
+ *
+ * To have a chance of completely hiding the overhead for
+ * the crc update, the inner loop must take considerably
+ * longer than 24 cycles. LONG = 64 makes the inner loop
+ * take about 24 cycles, so is not quite large enough.
+ * LONG = 128 works OK. Unhideable overheads are about
+ * 12 cycles per inner loop. All assuming timing like
+ * Haswell.
+ */
+ crc = crc32c_shift(crc32c_long, crc) ^ crc0;
+ crc1 = crc32c_shift(crc32c_long, crc1);
+ crc = crc32c_shift(crc32c_2long, crc) ^ crc1;
+ crc0 = crc2;
next += LONG * 2;
len -= LONG * 3;
}
+ crc0 ^= crc;
+#endif /* LONG > SHORT */
/*
* Do the same thing, but now on SHORT*3 blocks for the remaining data
* less than a LONG*3 block
*/
+ crc = 0;
while (len >= SHORT * 3) {
crc1 = 0;
crc2 = 0;
@@ -259,11 +347,14 @@ sse42_crc32c(uint32_t crc, const unsigne
#endif
next += align;
} while (next < end);
- crc0 = crc32c_shift(crc32c_short, crc0) ^ crc1;
- crc0 = crc32c_shift(crc32c_short, crc0) ^ crc2;
+ crc = crc32c_shift(crc32c_short, crc) ^ crc0;
+ crc1 = crc32c_shift(crc32c_short, crc1);
+ crc = crc32c_shift(crc32c_2short, crc) ^ crc1;
+ crc0 = crc2;
next += SHORT * 2;
len -= SHORT * 3;
}
+ crc0 ^= crc;
/* Compute the crc on the remaining bytes at native word size. */
end = next + (len - (len & (align - 1)));
More information about the svn-src-head
mailing list