Merge pull request #16218 from JulienMaille:cuda-dnn-for-older-gpus
authorJulien <182520+JulienMaille@users.noreply.github.com>
Wed, 15 Jan 2020 15:28:37 +0000 (16:28 +0100)
committerAlexander Alekhin <alexander.a.alekhin@gmail.com>
Wed, 15 Jan 2020 15:28:37 +0000 (18:28 +0300)
commit4e2ef8c8f57644ccb8e762a37f70a61007c6be1c
tree8c96abfa291b488bbed4b8384b1a567488d81713
parent4cb9faf6c9c012d603217bc09313d325bb9fb04a
Merge pull request #16218 from JulienMaille:cuda-dnn-for-older-gpus

Enable cuda4dnn on hardware without support for __half

* Enable cuda4dnn on hardware without support for half (ie. compute capability < 5.3)

Update CMakeLists.txt

Lowered minimum CC to 3.0

* UPD: added ifdef on new copy kernel

* added fp16 support detection at runtime

* Clarified #if condition on atomicAdd definition

* More explicit CMake error message
22 files changed:
modules/dnn/CMakeLists.txt
modules/dnn/src/cuda/activations.cu
modules/dnn/src/cuda/atomics.hpp
modules/dnn/src/cuda/bias_activation.cu
modules/dnn/src/cuda/concat.cu
modules/dnn/src/cuda/crop_and_resize.cu
modules/dnn/src/cuda/eltwise_ops.cu
modules/dnn/src/cuda/fill_copy.cu
modules/dnn/src/cuda/limits.hpp
modules/dnn/src/cuda/math.hpp
modules/dnn/src/cuda/max_unpooling.cu
modules/dnn/src/cuda/normalize.cu
modules/dnn/src/cuda/padding.cu
modules/dnn/src/cuda/permute.cu
modules/dnn/src/cuda/prior_box.cu
modules/dnn/src/cuda/region.cu
modules/dnn/src/cuda/resize.cu
modules/dnn/src/cuda/roi_pooling.cu
modules/dnn/src/cuda/scale_shift.cu
modules/dnn/src/cuda/slice.cu
modules/dnn/src/cuda4dnn/csl/fp16.hpp
modules/dnn/src/dnn.cpp