cuda: update ptx code with cuda7.5 nvcc

kidandcat · Nov 2, 2015 · 8d67f79 · 8d67f79
1 parent ed53ee1
commit 8d67f79
Show file tree

Hide file tree

Showing 39 changed files with 21,929 additions and 10,658 deletions.
diff --git a/cuda/copypadmul_wrapper.go b/cuda/copypadmul_wrapper.go
diff --git a/cuda/copyunpad_wrapper.go b/cuda/copyunpad_wrapper.go
diff --git a/cuda/crop_wrapper.go b/cuda/crop_wrapper.go
diff --git a/cuda/cubicanisotropy_wrapper.go b/cuda/cubicanisotropy_wrapper.go
diff --git a/cuda/dmi_wrapper.go b/cuda/dmi_wrapper.go
diff --git a/cuda/dmibulk_wrapper.go b/cuda/dmibulk_wrapper.go
diff --git a/cuda/dotproduct_wrapper.go b/cuda/dotproduct_wrapper.go
diff --git a/cuda/exchange_wrapper.go b/cuda/exchange_wrapper.go
diff --git a/cuda/exchangedecode_wrapper.go b/cuda/exchangedecode_wrapper.go
diff --git a/cuda/kernmulc_wrapper.go b/cuda/kernmulc_wrapper.go
diff --git a/cuda/kernmulrsymm2dxy_wrapper.go b/cuda/kernmulrsymm2dxy_wrapper.go
diff --git a/cuda/kernmulrsymm2dz_wrapper.go b/cuda/kernmulrsymm2dz_wrapper.go
diff --git a/cuda/kernmulrsymm3d_wrapper.go b/cuda/kernmulrsymm3d_wrapper.go
diff --git a/cuda/llnoprecess_wrapper.go b/cuda/llnoprecess_wrapper.go
diff --git a/cuda/lltorque_wrapper.go b/cuda/lltorque_wrapper.go
diff --git a/cuda/madd2_wrapper.go b/cuda/madd2_wrapper.go
diff --git a/cuda/madd3_wrapper.go b/cuda/madd3_wrapper.go
diff --git a/cuda/maxangle_wrapper.go b/cuda/maxangle_wrapper.go
diff --git a/cuda/minimize_wrapper.go b/cuda/minimize_wrapper.go
diff --git a/cuda/mul_wrapper.go b/cuda/mul_wrapper.go
diff --git a/cuda/normalize_wrapper.go b/cuda/normalize_wrapper.go
diff --git a/cuda/reducedot_wrapper.go b/cuda/reducedot_wrapper.go
diff --git a/cuda/reducemaxabs_wrapper.go b/cuda/reducemaxabs_wrapper.go
diff --git a/cuda/reducemaxdiff_wrapper.go b/cuda/reducemaxdiff_wrapper.go
diff --git a/cuda/reducemaxvecdiff2_wrapper.go b/cuda/reducemaxvecdiff2_wrapper.go
diff --git a/cuda/reducemaxvecnorm2_wrapper.go b/cuda/reducemaxvecnorm2_wrapper.go
diff --git a/cuda/reducesum_wrapper.go b/cuda/reducesum_wrapper.go
diff --git a/cuda/regionaddv_wrapper.go b/cuda/regionaddv_wrapper.go
diff --git a/cuda/regiondecode_wrapper.go b/cuda/regiondecode_wrapper.go
diff --git a/cuda/regionselect_wrapper.go b/cuda/regionselect_wrapper.go
diff --git a/cuda/resize_wrapper.go b/cuda/resize_wrapper.go
diff --git a/cuda/shiftbytes_wrapper.go b/cuda/shiftbytes_wrapper.go
diff --git a/cuda/shiftx_wrapper.go b/cuda/shiftx_wrapper.go
diff --git a/cuda/slonczewski_wrapper.go b/cuda/slonczewski_wrapper.go
diff --git a/cuda/temperature_wrapper.go b/cuda/temperature_wrapper.go
diff --git a/cuda/topologicalcharge_wrapper.go b/cuda/topologicalcharge_wrapper.go
diff --git a/cuda/uniaxialanisotropy_wrapper.go b/cuda/uniaxialanisotropy_wrapper.go
diff --git a/cuda/zeromask_wrapper.go b/cuda/zeromask_wrapper.go
diff --git a/cuda/zhangli_wrapper.go b/cuda/zhangli_wrapper.go