[arch-commits] Commit in nccl/trunk (PKGBUILD)

[Konstantin Gizdov]

    Date: Wednesday, June 23, 2021 @ 15:30:02
  Author: kgizdov
Revision: 967074

upgpkg: nccl 2.9.9-2: remove unneeded PTX JIT compile units

Modified:
  nccl/trunk/PKGBUILD

----------+
 PKGBUILD |   20 ++++++++------------
 1 file changed, 8 insertions(+), 12 deletions(-)

Modified: PKGBUILD
===================================================================
--- PKGBUILD	2021-06-23 14:53:19 UTC (rev 967073)
+++ PKGBUILD	2021-06-23 15:30:02 UTC (rev 967074)
@@ -4,7 +4,7 @@
 pkgname=nccl
 pkgver=2.9.9
 _upstr_pkgrel=1
-pkgrel=1
+pkgrel=2
 pkgdesc='Library for NVIDIA multi-GPU and multi-node collective communication primitives'
 arch=('x86_64')
 url='https://developer.nvidia.com/nccl/'
@@ -30,26 +30,22 @@
 build() {
   cd "$pkgname"
 
+  # https://docs.nvidia.com/cuda/turing-compatibility-guide/index.html
+  # specify as follows:
+  ## PTX virtual arch: -gencode=arch=compute_XX,code=compute_XX
+  ## binary cubin arch: -gencode=arch=compute_XX,code=sm_XX
+  # we provide binary support on all arches, no PTX is needed
+  # drop PTX to improve performance and reduce linking issues
   export NVCC_GENCODE="-gencode=arch=compute_52,code=sm_52 \
-                       -gencode=arch=compute_52,code=compute_52 \
                        -gencode=arch=compute_53,code=sm_53 \
-                       -gencode=arch=compute_53,code=compute_53 \
                        -gencode=arch=compute_60,code=sm_60 \
-                       -gencode=arch=compute_60,code=compute_60 \
                        -gencode=arch=compute_61,code=sm_61 \
-                       -gencode=arch=compute_61,code=compute_61 \
                        -gencode=arch=compute_62,code=sm_62 \
-                       -gencode=arch=compute_62,code=compute_62 \
                        -gencode=arch=compute_70,code=sm_70 \
-                       -gencode=arch=compute_70,code=compute_70 \
                        -gencode=arch=compute_72,code=sm_72 \
-                       -gencode=arch=compute_72,code=compute_72 \
                        -gencode=arch=compute_75,code=sm_75 \
-                       -gencode=arch=compute_75,code=compute_75 \
                        -gencode=arch=compute_80,code=sm_80 \
-                       -gencode=arch=compute_80,code=compute_80 \
-                       -gencode=arch=compute_86,code=sm_86 \
-                       -gencode=arch=compute_86,code=compute_86"
+                       -gencode=arch=compute_86,code=sm_86"
 
   make CUDA_HOME=/opt/cuda PREFIX=/usr src.build
 }