Specify the Name Mangling Rule for Vector Function Decorated with OpenMP #455
Conversation
|
@kito-cheng Hi, I create a pr and I think we can discuss the name mangling for vector function in here. |
Introduce how to name vector function associated to the scalar function decorated with the OpenMP `declare simd` pragma. These vector functions can be used to process multiple instances in a single invocation to improve performance. Both library vendors and compilers must obey this rule to interface with each other.
zengdage
force-pushed
the
name-mangling-vector-function-openmp
branch
from
4fc6edb to
ffd686c
Compare
1. Rename LIBMVEC_X86 into LIBMVEC to support libmvec in risc-v.
2. Add RVVM1/2/4/8 in VFISAKind to distingusih the LMUL value.
3. Declare some RVV vector math functions in VecFuncs.def.
In VecFuncs.def, I add the LI_DEFINE_VECFUNC of LIBMVEC_RVV as follow:
'''
TLI_DEFINE_VECFUNC("sin", "_ZGV1Nxv_sin", SCALABLE(1), "_ZGVr1Nxv")
TLI_DEFINE_VECFUNC("sin", "_ZGV2Nxv_sin", SCALABLE(2), "_ZGVr2Nxv")
TLI_DEFINE_VECFUNC("llvm.exp.f32", "_ZGV1Nxv_expf", SCALABLE(2), "_ZGVr1Nxv")
TLI_DEFINE_VECFUNC("llvm.exp.f32", "_ZGV2Nxv_expf", SCALABLE(4), "_ZGVr2Nxv")
'''
The 'VEC' of TLI_DEFINE_VECFUNC is like '_ZGV2Nxv_sin', its name mangling rules
defined in
riscv-non-isa/riscv-elf-psabi-doc#455
The 'VF' of TLI_DEFINE_VECFUNC is like 'SCALABLE(2)', it should be
'vscale x (LMUL * 64 / sizeof(Type)'.
The 'VABI_PREFIX' of TLI_DEFINE_VECFUNC is like '_ZGVr1Nxv', 'r' means RISC-V
vector extension, '1' is the LMUL value.
'''
_ZGVr1Nxv --> RISC-V Vector Extension with LMUL=1
_ZGVr2Nxv --> RISC-V Vector Extension with LMUL=2
_ZGVr4Nxv --> RISC-V Vector Extension with LMUL=4
_ZGVr8Nxv --> RISC-V Vector Extension with LMUL=8
'''
1. Rename LIBMVEC_X86 into LIBMVEC to support libmvec in risc-v.
2. Add RVVM1/2/4/8 in VFISAKind to distingusih the LMUL value.
3. Declare some RVV vector math functions in VecFuncs.def.
In VecFuncs.def, I add the LI_DEFINE_VECFUNC of LIBMVEC_RVV as follow:
```
TLI_DEFINE_VECFUNC("sin", "_ZGV1Nxv_sin", SCALABLE(1), "_ZGVr1Nxv")
TLI_DEFINE_VECFUNC("sin", "_ZGV2Nxv_sin", SCALABLE(2), "_ZGVr2Nxv")
TLI_DEFINE_VECFUNC("llvm.exp.f32", "_ZGV1Nxv_expf", SCALABLE(2), "_ZGVr1Nxv")
TLI_DEFINE_VECFUNC("llvm.exp.f32", "_ZGV2Nxv_expf", SCALABLE(4), "_ZGVr2Nxv")
```
The `VEC` of TLI_DEFINE_VECFUNC (e.g., `_ZGV2Nxv_sin`), its name mangling rules
defined in
riscv-non-isa/riscv-elf-psabi-doc#455
Now it's still under review.
The `VF` (e.g., `SCALABLE(2)`), it should be `vscale x (LMUL * 64 / sizeof(Type)`.
The `VABI_PREFIX` (e.g., `_ZGVr1Nxv`), `r` means RISC-V vector extension, `1` is
the LMUL value.
```
_ZGVr1Nxv --> RISC-V Vector Extension with LMUL=1
_ZGVr2Nxv --> RISC-V Vector Extension with LMUL=2
_ZGVr4Nxv --> RISC-V Vector Extension with LMUL=4
_ZGVr8Nxv --> RISC-V Vector Extension with LMUL=8
```
|
ping @kito-cheng , @topperc |
riscv-elf.adoc
Outdated
|
|
||
| [source,abnf] | ||
| ---- | ||
| mangled-vector-name := "_ZGV" <lmul> <mask> <len> <parameters> "_" <func-name> |
There was a problem hiding this comment.
I also have considered using "_ZGVr", but it will lack LMUL information and so just support a single fixed LMUL value. _ZGV_xxx(vfloat64m4_t) and __ZGV_xxx(vfloat64m2_t) have different computing power. I think if we want to take full advantage of RVV computing power, and need to attach LMUL information into vector function name.
There was a problem hiding this comment.
I just meant an r before the LMUL. There's an r in this patch right? llvm/llvm-project#119844
There was a problem hiding this comment.
The reason I chose "_ZGV" in the first place is that GCC take a single character as isa name in simd_clone_mangle (vecsize_mangle in struct cgraph_simd_clone), so I delete "r" from "r1/2/4/8" to meet gcc requirements.
|
I’m not sure we should let a gcc implementation detail constrain the spec.
Is it possible to encode the LMUL in the length field maybe by putting a
number after the 'x'?
…On Fri, Dec 13, 2024 at 7:11 PM Zhijin Zeng ***@***.***> wrote:
***@***.**** commented on this pull request.
------------------------------
In riscv-elf.adoc
<#455 (comment)>
:
> @@ -234,6 +234,116 @@ type-name = identifier-nondigit *identifier-char
identifier-nondigit = ALPHA / "_"
identifier-char = identifier-nondigit / "_"
----
+
+== Name Mangling for Vector Function
+
+This section describes how to name vector functions corresponding to scalar
+functions which are decorated with `#pragma omp declare simd`. The order of the
+vector function parameters should be the same as that of the original scalar
+function.
+
+Name mangling rule for vector functions.
+
+[source,abnf]
+----
+mangled-vector-name := "_ZGV" <lmul> <mask> <len> <parameters> "_" <func-name>
I also have considered using "_ZGVr", but it will lack LMUL information
and so just support a fixed LMUL value. _ZGV_xxx(vfloat64m4_t) and
__ZGV_xxx(vfloat64m2_t) have different computing power. I think if we
want to take full advantage of RVV computing power, and need to attach LMUL
information into vector function name.
—
Reply to this email directly, view it on GitHub
<#455 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFMFNKXMIATBZMZQRAO6IQT2FOOU7AVCNFSM6AAAAABSUO6PU2VHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZDKMBTGQ4TONRYHA>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
|
This is more complicated in GCC and LLVM. Because GCC just support taking constant value as length, and LLVM need to use |
This pr is to introduce how to name the vector function associated to the scalar function decorated with an OpenMP
declare simddirective. It' s important because library vendor and compilers must follow this rule to interface with each other. For example, now risc-v doesn't support libmvec in glibc, if we want to support it, we need to know how to name the vectorized math functions.I have referenced the aarch64 name manling rule
vfabia64in https://github.com/ARM-software/abi-aa/blob/main/vfabia64/vfabia64.rst#vector-function-signature. The main difference is that thelmulin risc-v replace theisain aarch64.