krishnaraj36 commented on code in PR #15786:
URL: https://github.com/apache/tvm/pull/15786#discussion_r1375672552
##########
python/tvm/relay/op/strategy/adreno.py:
##########
@@ -215,6 +215,58 @@ def
conv2d_winograd_without_weight_transform_strategy_adreno(attrs, inputs, out_
return strategy
+@conv2d_transpose_strategy.register("adreno")
+def conv2d_transpose_strategy_adreno(attrs, inputs, out_type, target):
+ """conv2d_transpose adreno strategy"""
+ strategy = _op.OpStrategy()
+ _, kernel = inputs
+ dilation = attrs.get_int_tuple("dilation")
+ groups = attrs.groups
+ data_layout = attrs.data_layout
+ kernel_layout = attrs.kernel_layout
+ assert dilation == (1, 1), "not support dilate now"
+
+ if (groups == 1) and (
+ (data_layout == "NCHW" and kernel_layout == "IOHW")
+ or (data_layout == "NCHW4c" and kernel_layout == "IOHW4o")
+ or (data_layout == "NCHW" and kernel_layout == "IOHW4o")
+ ):
+ if len(kernel.shape) == 4:
+ oc, _, _, _ = get_const_tuple(kernel.shape)
+ else:
+ oc, _, _, _, _ = get_const_tuple(kernel.shape)
+ # We cannot use textures for case than number of channels is less than
4.
+ # So, we use compute functions from cuda.
+ if len(kernel.shape) == 4 and oc < 4:
+ strategy.add_implementation(
+ wrap_compute_conv2d_transpose(topi.cuda.conv2d_transpose_nchw),
+ wrap_topi_schedule(topi.cuda.schedule_conv2d_transpose_nchw),
+ name="conv2d_transpose_nchw.cuda",
+ )
+ return strategy
Review Comment:
When the number of channels is less than 4, the repacking is not happening,
and it raises the runtime failures. This is known existing issue all conv2d
operators in tvm. So we are bypassing to cuda schedule.
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
