| 1 |
commit: fad64b065493a4406304cf7cb4d4aa36a7876555 |
| 2 |
Author: Miezhiko <Miezhiko <AT> gmail <DOT> com> |
| 3 |
AuthorDate: Mon Dec 27 10:56:29 2021 +0000 |
| 4 |
Commit: Andrew Ammerlaan <andrewammerlaan <AT> gentoo <DOT> org> |
| 5 |
CommitDate: Mon Dec 27 11:01:04 2021 +0000 |
| 6 |
URL: https://gitweb.gentoo.org/proj/sci.git/commit/?id=fad64b06 |
| 7 |
|
| 8 |
sci-libs/pytorch: fixes for cuda USE flag, nccl update |
| 9 |
|
| 10 |
Signed-off-by: Miezhiko <Miezhiko <AT> gmail.com> |
| 11 |
Closes: https://github.com/gentoo/sci/pull/1132 |
| 12 |
Signed-off-by: Andrew Ammerlaan <andrewammerlaan <AT> gentoo.org> |
| 13 |
|
| 14 |
.../files/pytorch-1.10.1-nccl-nvccflags.patch | 37 ++++++++++++++++++++++ |
| 15 |
sci-libs/pytorch/pytorch-1.10.1.ebuild | 7 ++-- |
| 16 |
2 files changed, 41 insertions(+), 3 deletions(-) |
| 17 |
|
| 18 |
diff --git a/sci-libs/pytorch/files/pytorch-1.10.1-nccl-nvccflags.patch b/sci-libs/pytorch/files/pytorch-1.10.1-nccl-nvccflags.patch |
| 19 |
new file mode 100644 |
| 20 |
index 000000000..0ac12a1e7 |
| 21 |
--- /dev/null |
| 22 |
+++ b/sci-libs/pytorch/files/pytorch-1.10.1-nccl-nvccflags.patch |
| 23 |
@@ -0,0 +1,37 @@ |
| 24 |
+From 7b04cd503fa1a6265893e0320ee246bcc0bd98cf Mon Sep 17 00:00:00 2001 |
| 25 |
+From: Miezhiko <Miezhiko@×××××.com> |
| 26 |
+Date: Mon, 27 Dec 2021 14:48:12 +0400 |
| 27 |
+Subject: [PATCH] neutral flags |
| 28 |
+ |
| 29 |
+--- |
| 30 |
+ makefiles/common.mk | 10 +--------- |
| 31 |
+ 1 file changed, 1 insertion(+), 9 deletions(-) |
| 32 |
+ |
| 33 |
+diff --git a/makefiles/common.mk b/makefiles/common.mk |
| 34 |
+index 64f8d2dc6..7a979c90f 100644 |
| 35 |
+--- a/makefiles/common.mk |
| 36 |
++++ b/makefiles/common.mk |
| 37 |
+@@ -55,7 +55,7 @@ CXXFLAGS := -DCUDA_MAJOR=$(CUDA_MAJOR) -DCUDA_MINOR=$(CUDA_MINOR) -fPIC -fvisi |
| 38 |
+ # Maxrregcount needs to be set accordingly to NCCL_MAX_NTHREADS (otherwise it will cause kernel launch errors) |
| 39 |
+ # 512 : 120, 640 : 96, 768 : 80, 1024 : 60 |
| 40 |
+ # We would not have to set this if we used __launch_bounds__, but this only works on kernels, not on functions. |
| 41 |
+-NVCUFLAGS := -ccbin $(CXX) $(NVCC_GENCODE) -std=c++11 --expt-extended-lambda -Xptxas -maxrregcount=96 -Xfatbin -compress-all |
| 42 |
++NVCUFLAGS := $(NVCCFLAGS) $(NVCC_GENCODE) -std=c++11 --expt-extended-lambda -Xptxas -maxrregcount=96 -Xfatbin -compress-all |
| 43 |
+ # Use addprefix so that we can specify more than one path |
| 44 |
+ NVLDFLAGS := -L${CUDA_LIB} -lcudart -lrt |
| 45 |
+ |
| 46 |
+@@ -69,14 +69,6 @@ NVLDFLAGS += ${GCOV_FLAGS:%=-Xcompiler %} |
| 47 |
+ # $(warning GCOV_FLAGS=${GCOV_FLAGS}) |
| 48 |
+ ########## GCOV ########## |
| 49 |
+ |
| 50 |
+-ifeq ($(DEBUG), 0) |
| 51 |
+-NVCUFLAGS += -O3 |
| 52 |
+-CXXFLAGS += -O3 -g |
| 53 |
+-else |
| 54 |
+-NVCUFLAGS += -O0 -G -g |
| 55 |
+-CXXFLAGS += -O0 -g -ggdb3 |
| 56 |
+-endif |
| 57 |
+- |
| 58 |
+ ifneq ($(VERBOSE), 0) |
| 59 |
+ NVCUFLAGS += -Xptxas -v -Xcompiler -Wall,-Wextra,-Wno-unused-parameter |
| 60 |
+ CXXFLAGS += -Wall -Wextra |
| 61 |
|
| 62 |
diff --git a/sci-libs/pytorch/pytorch-1.10.1.ebuild b/sci-libs/pytorch/pytorch-1.10.1.ebuild |
| 63 |
index a7bbb62ac..948d595a7 100644 |
| 64 |
--- a/sci-libs/pytorch/pytorch-1.10.1.ebuild |
| 65 |
+++ b/sci-libs/pytorch/pytorch-1.10.1.ebuild |
| 66 |
@@ -28,7 +28,7 @@ https://github.com/facebookincubator/gloo/archive/c22a5cfba94edf8ea4f53a174d38aa |
| 67 |
https://github.com/google/googletest/archive/e2239ee6043f73722e7aa812a459f54a28552929.tar.gz -> googletest-e2239ee6043f73722e7aa812a459f54a28552929.tar.gz |
| 68 |
https://github.com/intel/ideep/archive/9ca27bbfd88fa1469cbf0467bd6f14cd1738fa40.tar.gz -> ideep-9ca27bbfd88fa1469cbf0467bd6f14cd1738fa40.tar.gz |
| 69 |
https://github.com/intel/mkl-dnn/archive/5ef631a0.tar.gz -> mkl-dnn-5ef631a0.tar.gz |
| 70 |
-cuda? ( https://github.com/NVIDIA/nccl/archive/033d7995.tar.gz -> nccl-033d7995.tar.gz ) |
| 71 |
+cuda? ( https://github.com/NVIDIA/nccl/archive/c5790b36722d5b41ee2a9b2bad69e364180ffd22.tar.gz -> nccl-c5790b36722d5b41ee2a9b2bad69e364180ffd22.tar.gz ) |
| 72 |
https://github.com/Maratyszcza/NNPACK/archive/c07e3a0400713d546e0dea2d5466dd22ea389c73.tar.gz -> NNPACK-c07e3a0400713d546e0dea2d5466dd22ea389c73.tar.gz |
| 73 |
https://github.com/onnx/onnx/archive/a82c6a70.tar.gz -> onnx-a82c6a70.tar.gz |
| 74 |
https://github.com/onnx/onnx-tensorrt/archive/c153211418a7c57ce071d9ce2a41f8d1c85a878f.tar.gz -> onnx-tensorrt-c153211418a7c57ce071d9ce2a41f8d1c85a878f.tar.gz |
| 75 |
@@ -168,7 +168,7 @@ src_prepare() { |
| 76 |
rmdir third_party/ideep/mkl-dnn || die |
| 77 |
ln -sv "${WORKDIR}"/mkl-dnn-5ef631a030a6f73131c77892041042805a06064f third_party/ideep/mkl-dnn || die |
| 78 |
rmdir third_party/nccl/nccl || die |
| 79 |
- ln -sv "${WORKDIR}"/nccl-033d799524fb97629af5ac2f609de367472b2696 third_party/nccl/nccl || die |
| 80 |
+ ln -sv "${WORKDIR}"/nccl-c5790b36722d5b41ee2a9b2bad69e364180ffd22 third_party/nccl/nccl || die |
| 81 |
rmdir third_party/NNPACK || die |
| 82 |
ln -sv "${WORKDIR}"/NNPACK-c07e3a0400713d546e0dea2d5466dd22ea389c73 third_party/NNPACK || die |
| 83 |
rmdir third_party/onnx || die |
| 84 |
@@ -208,9 +208,10 @@ src_prepare() { |
| 85 |
|
| 86 |
if use cuda; then |
| 87 |
cd third_party/nccl/nccl || die |
| 88 |
- eapply "${FILESDIR}"/${PN}-1.6.0-nccl-nvccflags.patch |
| 89 |
+ eapply "${FILESDIR}"/${PN}-1.10.1-nccl-nvccflags.patch |
| 90 |
cuda_src_prepare |
| 91 |
export CUDAHOSTCXX=$(cuda_gccdir)/g++ |
| 92 |
+ cd "${S}" |
| 93 |
fi |
| 94 |
|
| 95 |
if use rocm; then |
Archives