Hi!
On 2023-09-08T11:04:24+0200, Tobias Burnus <tob...@codesourcery.com> wrote:
> On 02.08.23 19:00, Andrew Stubbs wrote:
>> The use of the PTX dynamic_smem_size feature means that low-latency allocator
>> will not work with the PTX 3.1 multilib.
Right: PTX '%dynamic_smem_size' was "Introduced in PTX ISA version 4.1",
and "Requires 'sm_20' or higher".
> - if I understand it correctly, our default build supports sm_30 and
> uses PTX ISA version 3.1 for it.
It's correct that we currently still build a '-march=sm_30', '-mptx=3.1'
multilib, but it's not the default one (the user has to explicitly
request it with '-mptx=3.1' on the GCC command line). The default one
had first been raised to '-mptx=6.3':
commit 8ff0669f6d1d6126b7c010da02fa6532abb5e1ca
Author: Tom de Vries <tdevr...@suse.de>
AuthorDate: Wed Jan 26 14:17:40 2022 +0100
Commit: Tom de Vries <tdevr...@suse.de>
CommitDate: Tue Feb 1 19:28:52 2022 +0100
[nvptx] Update default ptx isa to 6.3
[...]
--- gcc/config/nvptx/nvptx.opt
+++ gcc/config/nvptx/nvptx.opt
@@ -91,3 +91,3 @@ Enum(ptx_version) String(7.0) Value(PTX_VERSION_7_0)
mptx=
-Target RejectNegative ToLower Joined Enum(ptx_version)
Var(ptx_version_option) Init(PTX_VERSION_3_1)
+Target RejectNegative ToLower Joined Enum(ptx_version)
Var(ptx_version_option) Init(PTX_VERSION_6_3)
Specify the version of the ptx version to use.
[...]
..., and then again lowered to '-mptx=6.0':
commit decde11183bdccc46587d6614b75f3d56a2f2e4a
Author: Tom de Vries <tdevr...@suse.de>
AuthorDate: Fri Feb 4 08:53:52 2022 +0100
Commit: Tom de Vries <tdevr...@suse.de>
CommitDate: Tue Feb 8 13:55:23 2022 +0100
[nvptx] Choose -mptx default based on -misa
[...]
--- gcc/config/nvptx/nvptx.cc
+++ gcc/config/nvptx/nvptx.cc
[...]
+static enum ptx_version
+default_ptx_version_option (void)
+{
+ [...]
+ /* Pick at least 6.0, to enable using bar.warp.sync to have a way to
force
+ warp convergence. */
+ res = MAX (res, PTX_VERSION_6_0);
+ [...]
+}
[...]
--- gcc/config/nvptx/nvptx.opt
+++ gcc/config/nvptx/nvptx.opt
@@ -91,3 +91,3 @@ Enum(ptx_version) String(7.0) Value(PTX_VERSION_7_0)
mptx=
-Target RejectNegative ToLower Joined Enum(ptx_version)
Var(ptx_version_option) Init(PTX_VERSION_6_3)
+Target RejectNegative ToLower Joined Enum(ptx_version)
Var(ptx_version_option)
Specify the version of the ptx version to use.
> If so, I think we should mention that
> nvptx GCC has to be configured with with-arch=sm_... >= sm_53 (=
> supported version >=4.1) and, during compilation, no -march= < that
> configure-time value may be specified. (Cf. also
> https://gcc.gnu.org/install/specific.html#nvptx-x-none )
Given that GCC/nvptx generally supports 'sm_20', only the PTX ISA version
matters here, and that's all fine if just using GCC's defaults. OK to push
"Clarify libgomp nvptx 'omp_low_lat_mem_space' documentation", see
attached?
Grüße
Thomas
>From a6cdf1358c8b2f5279517ec7ebeb3336299ea928 Mon Sep 17 00:00:00 2001
From: Thomas Schwinge <tschwi...@baylibre.com>
Date: Tue, 12 Nov 2024 09:54:35 +0100
Subject: [PATCH] Clarify libgomp nvptx 'omp_low_lat_mem_space' documentation
PTX '%dynamic_smem_size' was "Introduced in PTX ISA version 4.1", and
"Requires 'sm_20' or higher". Given that GCC/nvptx generally supports
'sm_20', only the PTX ISA version matters here, and that's all fine if
just using GCC's defaults. Follow-up to
commit e9a19ead498fcc89186b724c6e76854f7751a89b
"openmp, nvptx: low-lat memory access traits".
libgomp/
* libgomp.texi: Clarify nvptx 'omp_low_lat_mem_space'
documentation.
---
libgomp/libgomp.texi | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/libgomp/libgomp.texi b/libgomp/libgomp.texi
index 6679f6da4b9..9602d70f26e 100644
--- a/libgomp/libgomp.texi
+++ b/libgomp/libgomp.texi
@@ -6968,8 +6968,10 @@ The implementation remark:
memory-copy functions of the CUDA library. Higher dimensions will
call those functions in a loop and are therefore supported.
@item Low-latency memory (@code{omp_low_lat_mem_space}) is supported when the
- the @code{access} trait is set to @code{cgroup}, the ISA is at least
- @code{sm_53}, and the PTX version is at least 4.1. The default pool size
+ the @code{access} trait is set to @code{cgroup}, and libgomp has
+ been built for PTX ISA version 4.1 or higher (such as in GCC's
+ default configuration). @c -mptx=4.1
+ The default pool size
is 8 kiB per team, but may be adjusted at runtime by setting environment
variable @code{GOMP_NVPTX_LOWLAT_POOL=@var{bytes}}. The maximum value is
limited by the available hardware, and care should be taken that the
--
2.34.1
