diff mbox series

[2/4] riscv: vectorized mem* functions

Message ID 20230301153247.1499566-3-vincent.chen@sifive.com
State New
Headers
Series riscv: Vectorized mem*/str* function |

Checks

Context Check Description
dj/TryBot-apply_patch success Patch applied to master at the time it was sent

Commit Message

Vincent Chen March 1, 2023, 3:32 p.m. UTC 
  From: Jerry Shih <jerry.shih@sifive.com>

This patch proposes implementations of memchr, memcmp, memcpy, memmove,
and memset that leverage the RISC-V V extension (RVV), version 1.0.
These routines assumes VLEN is at least 32 bits, as is required by all
currently defined vector extensions, and they support arbitrarily large
VLEN. All implementations work for both RV32 and RV64 platforms, and
make no assumptions about page size.
---
 sysdeps/riscv/rvv/memchr.S  | 63 +++++++++++++++++++++++++++++++
 sysdeps/riscv/rvv/memcmp.S  | 75 +++++++++++++++++++++++++++++++++++++
 sysdeps/riscv/rvv/memcpy.S  | 51 +++++++++++++++++++++++++
 sysdeps/riscv/rvv/memmove.S | 72 +++++++++++++++++++++++++++++++++++
 sysdeps/riscv/rvv/memset.S  | 51 +++++++++++++++++++++++++
 5 files changed, 312 insertions(+)
 create mode 100644 sysdeps/riscv/rvv/memchr.S
 create mode 100644 sysdeps/riscv/rvv/memcmp.S
 create mode 100644 sysdeps/riscv/rvv/memcpy.S
 create mode 100644 sysdeps/riscv/rvv/memmove.S
 create mode 100644 sysdeps/riscv/rvv/memset.S

Comments

Sergei Lewis March 2, 2023, 10:37 a.m. UTC | #1 
I note this memcpy implementation performs a vsetvli, load, store, three
ALU ops and a branch per loop.

If, instead, you start by copying just enough bytes to align the
destination pointer to vlen*lmul, you can pull the vsetvli out of the loop,
since you know the number of bytes remaining is a multiple of the operation
size - essentially, you've done the tail handling up front. Then, since you
no longer need to track the number of bytes remaining in order to feed it
to vsetvli, you can also calculate the expected buffer end outside the
loop, have the conditional branch compare current destination with that
instead of comparing number of bytes remaining with 0, and therefore also
lose an ALU op from the loop.

You can see this pattern in the vectorised mem* patchset I posted a while
back. In my testing I found this can improve performance by 10-15% compared
to the naive approach (depending on chosen LMUL, operation size and L1
size).

The other mem* functions can also similarly benefit from this pattern.
Yun Hsiang May 3, 2023, 6:56 a.m. UTC | #2 
On Thu, Mar 02, 2023 at 10:37:04AM +0000, Sergei Lewis wrote:
> 
> I note this memcpy implementation performs a vsetvli, load, store, three
> ALU ops and a branch per loop.
> 
> If, instead, you start by copying just enough bytes to align the
> destination pointer to vlen*lmul, you can pull the vsetvli out of the loop,
> since you know the number of bytes remaining is a multiple of the operation
> size - essentially, you've done the tail handling up front. Then, since you
> no longer need to track the number of bytes remaining in order to feed it
> to vsetvli, you can also calculate the expected buffer end outside the
> loop, have the conditional branch compare current destination with that
> instead of comparing number of bytes remaining with 0, and therefore also
> lose an ALU op from the loop.
> 
> You can see this pattern in the vectorised mem* patchset I posted a while
> back. In my testing I found this can improve performance by 10-15% compared
> to the naive approach (depending on chosen LMUL, operation size and L1
> size).
I've run glibc benchtests and rtl simulation to evaluate the performance between https://patchwork.sourceware.org/project/glibc/patch/20230201095232.15942-2-slewis@rivosinc.com/ and this implementation. In my test, the implementations in this thread have less instruction/cycle count in most test cases.
> 
> The other mem* functions can also similarly benefit from this pattern.
diff mbox series

Patch

diff --git a/sysdeps/riscv/rvv/memchr.S b/sysdeps/riscv/rvv/memchr.S
new file mode 100644
index 0000000000..6981a9f8b0
--- /dev/null
+++ b/sysdeps/riscv/rvv/memchr.S
@@ -0,0 +1,63 @@ 
+/* RVV versions memchr.  RISC-V version.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Jerry Shih <jerry.shih@sifive.com>.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+#include <sys/asm.h>
+
+#define iResult a0
+
+#define pSrc a0
+#define iValue a1
+#define iNum a2
+
+#define iVL a3
+#define iTemp a4
+
+#define ELEM_LMUL_SETTING m8
+#define vData v0
+#define vMask v8
+
+ENTRY(memchr)
+
+L(loop):
+    vsetvli zero, iNum, e8, ELEM_LMUL_SETTING, ta, ma
+
+    vle8ff.v vData, (pSrc)
+    /* Find the iValue inside the loaded data.  */
+    vmseq.vx vMask, vData, iValue
+    vfirst.m iTemp, vMask
+
+    /* Skip the loop if we find the matched value.  */
+    bgez iTemp, L(found)
+
+    csrr iVL, vl
+    sub iNum, iNum, iVL
+    add pSrc, pSrc, iVL
+
+    bnez iNum, L(loop)
+
+    li iResult, 0
+    ret
+
+L(found):
+    add iResult, pSrc, iTemp
+    ret
+
+END(memchr)
+libc_hidden_builtin_def (memchr)
diff --git a/sysdeps/riscv/rvv/memcmp.S b/sysdeps/riscv/rvv/memcmp.S
new file mode 100644
index 0000000000..b156ec524c
--- /dev/null
+++ b/sysdeps/riscv/rvv/memcmp.S
@@ -0,0 +1,75 @@ 
+/* RVV versions memcmp.  RISC-V version.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Jerry Shih <jerry.shih@sifive.com>.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+#include <sys/asm.h>
+
+#define iResult a0
+
+#define pSrc1 a0
+#define pSrc2 a1
+#define iNum a2
+
+#define iVL a3
+#define iTemp a4
+#define iTemp1 a5
+#define iTemp2 a6
+
+#define ELEM_LMUL_SETTING m8
+#define vData1 v0
+#define vData2 v8
+#define vMask v16
+
+ENTRY(memcmp)
+
+L(loop):
+    vsetvli iVL, iNum, e8, ELEM_LMUL_SETTING, ta, ma
+
+    vle8.v vData1, (pSrc1)
+    vle8.v vData2, (pSrc2)
+
+    vmsne.vv vMask, vData1, vData2
+    sub iNum, iNum, iVL
+    vfirst.m iTemp, vMask
+
+    /* Skip the loop if we find the different value between pSrc1 and pSrc2.  */
+    bgez iTemp, L(found)
+
+    add pSrc1, pSrc1, iVL
+    add pSrc2, pSrc2, iVL
+
+    bnez iNum, L(loop)
+
+    li iResult, 0
+    ret
+
+L(found):
+    add pSrc1, pSrc1, iTemp
+    add pSrc2, pSrc2, iTemp
+    lbu iTemp1, 0(pSrc1)
+    lbu iTemp2, 0(pSrc2)
+    sub iResult, iTemp1, iTemp2
+    ret
+
+END(memcmp)
+libc_hidden_builtin_def (memcmp)
+weak_alias (memcmp,bcmp)
+strong_alias (memcmp, __memcmpeq)
+libc_hidden_def (__memcmpeq)
+
diff --git a/sysdeps/riscv/rvv/memcpy.S b/sysdeps/riscv/rvv/memcpy.S
new file mode 100644
index 0000000000..de790fbe51
--- /dev/null
+++ b/sysdeps/riscv/rvv/memcpy.S
@@ -0,0 +1,51 @@ 
+/* RVV versions memcpy.  RISC-V version.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Jerry Shih <jerry.shih@sifive.com>.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+#include <sys/asm.h>
+
+#define pDst a0
+#define pSrc a1
+#define iNum a2
+
+#define iVL a3
+#define pDstPtr a4
+
+#define ELEM_LMUL_SETTING m8
+#define vData v0
+
+ENTRY(memcpy)
+
+    mv pDstPtr, pDst
+
+L(loop):
+    vsetvli iVL, iNum, e8, ELEM_LMUL_SETTING, ta, ma
+
+    vle8.v vData, (pSrc)
+    sub iNum, iNum, iVL
+    add pSrc, pSrc, iVL
+    vse8.v vData, (pDstPtr)
+    add pDstPtr, pDstPtr, iVL
+
+    bnez iNum, L(loop)
+
+    ret
+
+END(memcpy)
+libc_hidden_builtin_def (memcpy)
diff --git a/sysdeps/riscv/rvv/memmove.S b/sysdeps/riscv/rvv/memmove.S
new file mode 100644
index 0000000000..ed12744064
--- /dev/null
+++ b/sysdeps/riscv/rvv/memmove.S
@@ -0,0 +1,72 @@ 
+/* RVV versions memmove.  RISC-V version.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Jerry Shih <jerry.shih@sifive.com>.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+#include <sys/asm.h>
+
+#define pDst a0
+#define pSrc a1
+#define iNum a2
+
+#define iVL a3
+#define pDstPtr a4
+#define pSrcBackwardPtr a5
+#define pDstBackwardPtr a6
+
+#define ELEM_LMUL_SETTING m8
+#define vData v0
+
+ENTRY(memmove)
+
+    mv pDstPtr, pDst
+
+    /* If pSrc is equal or after pDst, all data in pSrc will be loaded before
+       overwrited for the overlapping case. We could use faster `forward-copy`.  */
+    bgeu pSrc, pDst, L(forward_copy_loop)
+    add pSrcBackwardPtr, pSrc, iNum
+    add pDstBackwardPtr, pDst, iNum
+    /* If pDst inside source data range, we need to use `backward_copy_loop` to
+       handle the overlapping issue.  */
+    bltu pDst, pSrcBackwardPtr, L(backward_copy_loop)
+
+L(forward_copy_loop):
+    vsetvli iVL, iNum, e8, ELEM_LMUL_SETTING, ta, ma
+
+    vle8.v vData, (pSrc)
+    sub iNum, iNum, iVL
+    add pSrc, pSrc, iVL
+    vse8.v vData, (pDstPtr)
+    add pDstPtr, pDstPtr, iVL
+
+    bnez iNum, L(forward_copy_loop)
+    ret
+
+L(backward_copy_loop):
+    vsetvli iVL, iNum, e8, ELEM_LMUL_SETTING, ta, ma
+
+    sub pSrcBackwardPtr, pSrcBackwardPtr, iVL
+    vle8.v vData, (pSrcBackwardPtr)
+    sub iNum, iNum, iVL
+    sub pDstBackwardPtr, pDstBackwardPtr, iVL
+    vse8.v vData, (pDstBackwardPtr)
+    bnez iNum, L(backward_copy_loop)
+    ret
+
+END(memmove)
+libc_hidden_builtin_def (memmove)
diff --git a/sysdeps/riscv/rvv/memset.S b/sysdeps/riscv/rvv/memset.S
new file mode 100644
index 0000000000..3a6c3d0afd
--- /dev/null
+++ b/sysdeps/riscv/rvv/memset.S
@@ -0,0 +1,51 @@ 
+/* RVV versions memset.  RISC-V version.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Jerry Shih <jerry.shih@sifive.com>.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+#include <sys/asm.h>
+
+#define pDst a0
+#define iValue a1
+#define iNum a2
+
+#define iVL a3
+#define iTemp a4
+#define pDstPtr a5
+
+#define ELEM_LMUL_SETTING m8
+#define vData v0
+
+ENTRY(memset)
+
+    mv pDstPtr, pDst
+
+    vsetvli iVL, iNum, e8, ELEM_LMUL_SETTING, ta, ma
+    vmv.v.x vData, iValue
+
+L(loop):
+    vse8.v vData, (pDstPtr)
+    sub iNum, iNum, iVL
+    add pDstPtr, pDstPtr, iVL
+    vsetvli iVL, iNum, e8, ELEM_LMUL_SETTING, ta, ma
+    bnez iNum, L(loop)
+
+    ret
+
+END(memset)
+libc_hidden_builtin_def (memset)