Split out host/bufferiszero.h.inc for x86, aarch64 and generic
in order to avoid an overlong ifdef ladder.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
QEMU now requires an x86-64-v2 host, which has SSE2.
Use it freely in buffer_is_zero.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Because non-embedded aarch64 is expected to have AdvSIMD enabled, merely
double-check with the compiler flags for __ARM_NEON and don't bother with
a runtime check. Otherwise, model the loop after the x86 SSE2 function.
Use UMAXV for the vector reduction. This is 3 cycles on cortex-a76 and
2 cycles on neoverse-n1.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Because the three alternatives are monotonic, we don't need
to keep a couple of bitmasks, just identify the strongest
alternative at startup.
Generalize test_buffer_is_zero_next_accel and init_accel
by always defining an accel_table array.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Split less-than and greater-than 256 cases.
Use unaligned accesses for head and tail.
Avoid using out-of-bounds pointers in loop boundary conditions.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Increase unroll factor in SIMD loops from 4x to 8x in order to move
their bottlenecks from ALU port contention to load issue rate (two loads
per cycle on popular x86 implementations).
Avoid using out-of-bounds pointers in loop boundary conditions.
Follow SSE2 implementation strategy in the AVX2 variant. Avoid use of
PTEST, which is not profitable there (like in the removed SSE4 variant).
Signed-off-by: Alexander Monakov <amonakov@ispras.ru>
Signed-off-by: Mikhail Romanov <mmromanov@ispras.ru>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20240206204809.9859-6-amonakov@ispras.ru>
Use of prefetching in bufferiszero.c is quite questionable:
- prefetches are issued just a few CPU cycles before the corresponding
line would be hit by demand loads;
- they are done for simple access patterns, i.e. where hardware
prefetchers can perform better;
- they compete for load ports in loops that should be limited by load
port throughput rather than ALU throughput.
Signed-off-by: Alexander Monakov <amonakov@ispras.ru>
Signed-off-by: Mikhail Romanov <mmromanov@ispras.ru>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20240206204809.9859-5-amonakov@ispras.ru>
Test for length >= 256 inline, where is is often a constant.
Before calling into the accelerated routine, sample three bytes
from the buffer, which handles most non-zero buffers.
Signed-off-by: Alexander Monakov <amonakov@ispras.ru>
Signed-off-by: Mikhail Romanov <mmromanov@ispras.ru>
Message-Id: <20240206204809.9859-3-amonakov@ispras.ru>
[rth: Use __builtin_constant_p; move the indirect call out of line.]
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Thanks to early checks in the inline buffer_is_zero wrapper, the SIMD
routines are invoked much more rarely in normal use when most buffers
are non-zero. This makes use of AVX512 unprofitable, as it incurs extra
frequency and voltage transition periods during which the CPU operates
at reduced performance, as described in
https://travisdowns.github.io/blog/2020/01/17/avxfreq1.html
Signed-off-by: Mikhail Romanov <mmromanov@ispras.ru>
Signed-off-by: Alexander Monakov <amonakov@ispras.ru>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20240206204809.9859-4-amonakov@ispras.ru>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
The SSE4.1 variant is virtually identical to the SSE2 variant, except
for using 'PTEST+JNZ' in place of 'PCMPEQB+PMOVMSKB+CMP+JNE' for testing
if an SSE register is all zeroes. The PTEST instruction decodes to two
uops, so it can be handled only by the complex decoder, and since
CMP+JNE are macro-fused, both sequences decode to three uops. The uops
comprising the PTEST instruction dispatch to p0 and p5 on Intel CPUs, so
PCMPEQB+PMOVMSKB is comparatively more flexible from dispatch
standpoint.
Hence, the use of PTEST brings no benefit from throughput standpoint.
Its latency is not important, since it feeds only a conditional jump,
which terminates the dependency chain.
I never observed PTEST variants to be faster on real hardware.
Signed-off-by: Alexander Monakov <amonakov@ispras.ru>
Signed-off-by: Mikhail Romanov <mmromanov@ispras.ru>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20240206204809.9859-2-amonakov@ispras.ru>
Use cpuinfo_init() during init_accel(), and the variable cpuinfo
during test_buffer_is_zero_next_accel(). Adjust the logic that
cycles through the set of accelerators for testing.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Replace the two uses of asm to expand xgetbv with an inline function.
Since one of the two has been using the mnemonic, assume that the
comment about "older versions of the assember" is obsolete, as even
that is 4 years old.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Use the attribute, which is supported by clang, instead of
the #pragma, which is not supported and, for some reason,
also not detected by the meson probe, so we fail by -Werror.
Include only <immintrin.h> as that is the outermost "official"
header for these intrinsics -- emmintrin.h and smmintrin -- are
older SSE2 and SSE4 specific headers, while the immintrin.h
includes all of the Intel intrinsics.
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
__get_cpuid_max returns an unsigned value.
For consistency, store the result in an unsigned variable.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Richard Henderson <rth@twiddle.net>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
By increasing avx2 length_to_accel to 128, we can simplify its logic and reduce a
branch.
The authorship of this patch actually belongs to Richard Henderson
<richard.henderson@linaro.org>, I just fixed a boundary case on his
original patch.
Suggested-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Robert Hoo <robert.hu@linux.intel.com>
Message-Id: <1585119021-46593-2-git-send-email-robert.hu@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Because in unit test, init_accel() will be called several times, each with
different accelerator type.
Signed-off-by: Robert Hoo <robert.hu@linux.intel.com>
Message-Id: <1585119021-46593-1-git-send-email-robert.hu@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
And intialize buffer_is_zero() with it, when Intel AVX512F is
available on host.
This function utilizes Intel AVX512 fundamental instructions which
is faster than its implementation with AVX2 (in my unit test, with
4K buffer, on CascadeLake SP, ~36% faster, buffer_zero_avx512() V.S.
buffer_zero_avx2()).
Signed-off-by: Robert Hoo <robert.hu@linux.intel.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
No header includes qemu-common.h after this commit, as prescribed by
qemu-common.h's file comment.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20190523143508.25387-5-armbru@redhat.com>
[Rebased with conflicts resolved automatically, except for
include/hw/arm/xlnx-zynqmp.h hw/arm/nrf51_soc.c hw/arm/msf2-soc.c
block/qcow2-refcount.c block/qcow2-cluster.c block/qcow2-cache.c
target/arm/cpu.h target/lm32/cpu.h target/m68k/cpu.h target/mips/cpu.h
target/moxie/cpu.h target/nios2/cpu.h target/openrisc/cpu.h
target/riscv/cpu.h target/tilegx/cpu.h target/tricore/cpu.h
target/unicore32/cpu.h target/xtensa/cpu.h; bsd-user/main.c and
net/tap-bsd.c fixed up]
Clang 3.9 passes the CONFIG_AVX2_OPT configure test. However, the
supplied <cpuid.h> does not contain the bit_AVX2 define that we use
when detecting whether the routine can be enabled.
Introduce a qemu-specific header that uses the compiler's definition
of __cpuid et al, but supplies any missing bit_* definitions needed.
This avoids introducing any extra ifdefs to util/bufferiszero.c, and
allows quite a few to be removed from tcg/i386/tcg-target.inc.c.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Message-id: 20170719044018.18063-1-rth@twiddle.net
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Handle alignment of buffers, so that the vector paths
can be used more often.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <1473800239-13841-1-git-send-email-rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
There's no real knowledge of the cacheline size,
just prefetching one loop ahead.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Message-Id: <1472496380-19706-7-git-send-email-rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Richard Henderson <rth@twiddle.net>
Message-Id: <1472496380-19706-6-git-send-email-rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
For ppc64le, gcc6 does extremely poorly with the Altivec code.
Moreover, on POWER7 and POWER8, a hand-optimized Altivec version
turns out to be no faster than the revised integer version, and
therefore not worth the effort.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The revised integer version is 4 times faster than the neon version
on an AppliedMicro Mustang. Even with hand scheduling and additional
unrolling I cannot make any neon version run as fast as the integer.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Allow selection of several acceleration functions
based on the size and alignment of the buffer.
Do not require ifunc support for AVX2 acceleration.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Message-Id: <1472496380-19706-5-git-send-email-rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Since the two users don't make use of the returned offset,
beyond ensuring that the entire buffer is zero, consider the
can_use_buffer_find_nonzero_offset and buffer_find_nonzero_offset
functions internal.
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Signed-off-by: Richard Henderson <rth@twiddle.net>
Message-Id: <1472496380-19706-4-git-send-email-rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This is unused and complicates the vector interface.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Message-Id: <1472496380-19706-3-git-send-email-rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Richard Henderson <rth@twiddle.net>
Message-Id: <1472496380-19706-2-git-send-email-rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
