提交 3180ec4d authored 作者: Pascal Lamblin's avatar Pascal Lamblin

Merge pull request #3662 from nouiz/noinplace_gpu

Allow optimizer_excluding=inplace to keep the move to GPU.
...@@ -522,11 +522,10 @@ def unuse(): ...@@ -522,11 +522,10 @@ def unuse():
cuda_enabled = False cuda_enabled = False
handle_shared_float32(False) handle_shared_float32(False)
optdb.remove_tags('gpu_opt', optdb.remove_tags('gpu_opt',
'fast_run', 'fast_compile',
'inplace') 'fast_run')
optdb.remove_tags('gpu_after_fusion', optdb.remove_tags('gpu_after_fusion',
'fast_run', 'fast_run')
'inplace')
def handle_shared_float32(tf): def handle_shared_float32(tf):
......
...@@ -92,7 +92,7 @@ except ImportError: ...@@ -92,7 +92,7 @@ except ImportError:
gpu_cut_copies = EquilibriumDB() gpu_cut_copies = EquilibriumDB()
gpu_seqopt.register('gpu_local_optimizations', gpu_optimizer, 1, gpu_seqopt.register('gpu_local_optimizations', gpu_optimizer, 1,
'fast_run', 'fast_compile', 'inplace', 'gpu') 'fast_run', 'fast_compile', 'gpu')
gpu_seqopt.register('gpu_cut_transfers', gpu_cut_copies, 2, gpu_seqopt.register('gpu_cut_transfers', gpu_cut_copies, 2,
'fast_run', 'fast_compile', 'gpu') 'fast_run', 'fast_compile', 'gpu')
# DO NOT PUT fast_run or fast_compile in gpu_opt! This will ALWAYS # DO NOT PUT fast_run or fast_compile in gpu_opt! This will ALWAYS
......
...@@ -51,7 +51,7 @@ conv_groupopt = LocalGroupDB() ...@@ -51,7 +51,7 @@ conv_groupopt = LocalGroupDB()
conv_groupopt.__name__ = "gpua_conv_opts" conv_groupopt.__name__ = "gpua_conv_opts"
gpu_seqopt.register('gpuarray_local_optimiziations', gpu_optimizer, 1, gpu_seqopt.register('gpuarray_local_optimiziations', gpu_optimizer, 1,
'fast_compile', 'fast_run', 'inplace', 'gpuarray') 'fast_compile', 'fast_run', 'gpuarray')
gpu_seqopt.register('gpuarray_cut_transfers', gpu_cut_copies, 2, gpu_seqopt.register('gpuarray_cut_transfers', gpu_cut_copies, 2,
'fast_compile', 'fast_run', 'gpuarray') 'fast_compile', 'fast_run', 'gpuarray')
......
Markdown 格式
0%
您添加了 0 到此讨论。请谨慎行事。
请先完成此评论的编辑!
注册 或者 后发表评论