Commit 752614cab8 ("target/riscv: rvv: Add tail agnostic for vector
load / store instructions") added an extra check for LMUL fragmentation,
intended for setting the "rest tail elements" in the last register for a
segment load insn.
Actually, the max_elements derived in vext_ld*() won't be a fraction of
vector register size, since the lmul encoded in desc is emul, which has
already been adjusted to 1 for LMUL fragmentation case by vext_get_emul()
in trans_rvv.c.inc, for ld_stride(), ld_us(), ld_index() and ldff().
Besides, vext_get_emul() has also taken EEW/SEW into consideration, so no
need to call vext_get_total_elems() which would base on the emul to derive
another emul, the second emul would be incorrect when esz differs from sew.
Thus this patch removes the check for extra tail elements.
Fixes: 752614cab8 ("target/riscv: rvv: Add tail agnostic for vector load / store instructions")
Signed-off-by: Xiao Wang <xiao.w.wang@intel.com>
Reviewed-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com>
Reviewed-by: Weiwei Li <liweiwei@iscas.ac.cn>
Message-Id: <20230607091646.4049428-1-xiao.w.wang@intel.com>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
The function is a no-op if 'vta' is zero but we're still doing a lot of
stuff in this function regardless. vext_set_elems_1s() will ignore every
single time (since vta is zero) and we just wasted time.
Skip it altogether in this case. Aside from the code simplification
there's a noticeable emulation performance gain by doing it. For a
regular C binary that does a vectors operation like this:
=======
#define SZ 10000000
int main ()
{
int *a = malloc (SZ * sizeof (int));
int *b = malloc (SZ * sizeof (int));
int *c = malloc (SZ * sizeof (int));
for (int i = 0; i < SZ; i++)
c[i] = a[i] + b[i];
return c[SZ - 1];
}
=======
Emulating it with qemu-riscv64 and RVV takes ~0.3 sec:
$ time ~/work/qemu/build/qemu-riscv64 \
-cpu rv64,debug=false,vext_spec=v1.0,v=true,vlen=128 ./foo.out
real 0m0.303s
user 0m0.281s
sys 0m0.023s
With this skip we take ~0.275 sec:
$ time ~/work/qemu/build/qemu-riscv64 \
-cpu rv64,debug=false,vext_spec=v1.0,v=true,vlen=128 ./foo.out
real 0m0.274s
user 0m0.252s
sys 0m0.019s
This performance gain adds up fast when executing heavy benchmarks like
SPEC.
Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com>
Acked-by: Alistair Francis <alistair.francis@wdc.com>
Reviewed-by: Palmer Dabbelt <palmer@rivosinc.com>
Reviewed-by: Weiwei Li <liweiwei@iscas.ac.cn>
Message-Id: <20230427205708.246679-2-dbarboza@ventanamicro.com>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
Commit 752614cab8 ("target/riscv: rvv: Add tail agnostic for vector
load / store instructions") added code to set the tail elements to 1 in
the end of vext_ldst_stride(), vext_ldst_us(), vext_ldst_index() and
vext_ldff(). Aside from a env->vl versus an evl value being used in the
first loop, the code is being repeated 4 times.
Create a helper to avoid code repetition in all those functions.
Arguments that are used in the callers (nf, esz and max_elems) are
passed as arguments. All other values are being derived inside the
helper.
Reviewed-by: Weiwei Li <liweiwei@iscas.ac.cn>
Reviewed-by: Frank Chang <frank.chang@sifive.com>
Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com>
Message-ID: <20230226170514.588071-2-dbarboza@ventanamicro.com>
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
Tweak the semantic patch to drop redundant parenthesis around the
return expression.
Coccinelle drops a comment in hw/rdma/vmw/pvrdma_cmd.c; restored
manually.
Coccinelle messes up vmdk_co_create(), not sure why. Change dropped,
will be done manually in the next commit.
Line breaks in target/avr/cpu.h and hw/rdma/vmw/pvrdma_cmd.c tidied up
manually.
Whitespace in tools/virtiofsd/fuse_lowlevel.c tidied up manually.
checkpatch.pl complains "return of an errno should typically be -ve"
two times for hw/9pfs/9p-synth.c. Preexisting, the patch merely makes
it visible to checkpatch.pl.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20221122134917.1217307-2-armbru@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Acked-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
According to v-spec, mask agnostic behavior can be either kept as
undisturbed or set elements' bits to all 1s. To distinguish the
difference of mask policies, QEMU should be able to simulate the mask
agnostic behavior as "set mask elements' bits to all 1s".
There are multiple possibility for agnostic elements according to
v-spec. The main intent of this patch-set tries to add option that
can distinguish between mask policies. Setting agnostic elements to
all 1s allows QEMU to express this.
This is the first commit regarding the optional mask agnostic
behavior. Follow-up commits will add this optional behavior
for all rvv instructions.
Signed-off-by: eop Chen <eop.chen@sifive.com>
Reviewed-by: Frank Chang <frank.chang@sifive.com>
Reviewed-by: Weiwei Li <liweiwei@iscas.ac.cn>
Reviewed-by: Alistair Francis <alistair.francis@wdc.com>
Message-Id: <165570784143.17634.35095816584573691-1@git.sr.ht>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
According to v-spec, tail agnostic behavior can be either kept as
undisturbed or set elements' bits to all 1s. To distinguish the
difference of tail policies, QEMU should be able to simulate the tail
agnostic behavior as "set tail elements' bits to all 1s".
There are multiple possibility for agnostic elements according to
v-spec. The main intent of this patch-set tries to add option that
can distinguish between tail policies. Setting agnostic elements to
all 1s allows QEMU to express this.
This is the first commit regarding the optional tail agnostic
behavior. Follow-up commits will add this optional behavior
for all rvv instructions.
Signed-off-by: eop Chen <eop.chen@sifive.com>
Reviewed-by: Frank Chang <frank.chang@sifive.com>
Reviewed-by: Weiwei Li <liweiwei@iscas.ac.cn>
Acked-by: Alistair Francis <alistair.francis@wdc.com>
Message-Id: <165449614532.19704.7000832880482980398-5@git.sr.ht>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
Replace a config-time define with a compile time condition
define (compatible with clang and gcc) that must be declared prior to
its usage. This avoids having a global configure time define, but also
prevents from bad usage, if the config header wasn't included before.
This can help to make some code independent from qemu too.
gcc supports __BYTE_ORDER__ from about 4.6 and clang from 3.2.
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
[ For the s390x parts I'm involved in ]
Acked-by: Halil Pasic <pasic@linux.ibm.com>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20220323155743.1585078-7-marcandre.lureau@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add supports of Vector unit-stride mask load/store instructions
(vlm.v, vsm.v), which has:
evl (effective vector length) = ceil(env->vl / 8).
The new instructions operate the same as unmasked byte loads and stores.
Add evl parameter to reuse vext_ldst_us().
Signed-off-by: Frank Chang <frank.chang@sifive.com>
Reviewed-by: Alistair Francis <alistair.francis@wdc.com>
Message-Id: <20211210075704.23951-74-frank.chang@sifive.com>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
