Shawn-Inspur commented on a change in pull request #5284: [Topi] Tensorcore
support for Conv3D
URL: https://github.com/apache/incubator-tvm/pull/5284#discussion_r406157545
##########
File path: python/tvm/relay/op/strategy/cuda.py
##########
@@ -268,11 +268,24 @@ def conv3d_strategy_cuda(attrs, inputs, out_type,
target):
wrap_topi_schedule(topi.cuda.schedule_conv3d_ncdhw_winograd),
name="conv3d_ncdhw_winograd.cuda",
plevel=5)
- else: # layout == "NDHWC":
-
strategy.add_implementation(wrap_compute_conv3d(topi.cuda.conv3d_ndhwc),
-
wrap_topi_schedule(topi.cuda.schedule_conv3d_ndhwc),
- name="conv3d_ndhwc.cuda",
- plevel=10)
+ else: # layout == "NDHWC":
+ strategy.add_implementation(
+ wrap_compute_conv3d(topi.cuda.conv3d_ndhwc),
+ wrap_topi_schedule(topi.cuda.schedule_conv3d_ndhwc),
+ name="conv3d_ndhwc.cuda",
+ plevel=10)
+ N, _, _, _, _ = get_const_tuple(data.shape)
+ _, _, _, CI, CO = get_const_tuple(kernel.shape)
+ if nvcc.have_tensorcore(tvm.gpu(0).compute_version):
Review comment:
Could you please check both the target_name and nvcc.have_tensorcore to skip
nvcc.have_tensorcore check when CUDA is not available?
Please refer to [Unable to get started with Metal on macOS 10.13 with Radeon
Pro 560 4096
MB](https://discuss.tvm.ai/t/unable-to-get-started-with-metal-on-macos-10-13-with-radeon-pro-560-4096-mb/6209/14)
for details.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
