[GitHub] [tvm] alanmacd commented on a diff in pull request #13752: [microTVM] Use QNN schedules to give SOTA performance

["alanmacd (via GitHub)" <gi...@apache.org>]

alanmacd commented on code in PR #13752:
URL: https://github.com/apache/tvm/pull/13752#discussion_r1087301065


##########
python/tvm/topi/arm_cpu/qnn.py:
##########
@@ -221,23 +261,21 @@ def qnn_conv2d(attrs, inputs, out_type):
     # Make a few checks to unpack the function arguments and ensure it was called with the right
     # arguments. Note that unlike most schedules, qnn_conv2d does not use a wrapper.
     assert len(inputs) == 11
-    data, kernel, _izp, _kzp, _iscale, _kscale, bias, scale = inputs[0:8]
-    output_layout = attrs.out_layout
-    assert output_layout == "NHWC"
+    assert not any(get_const_tuple(attrs.padding))
 
+    data, kernel, _izp, _kzp, _iscale, _kscale, bias, scale = inputs[0:8]
     _, height, width, in_channels = get_const_tuple(data.shape)
     out_channels, kernel_h, kernel_w, _ = get_const_tuple(kernel.shape)
-    y_stride, x_stride = get_const_tuple(attrs.strides)
 
+    y_stride, x_stride = get_const_tuple(attrs.strides)
     out_height = _compute_output_dim(height, kernel_h, y_stride)
     out_width = _compute_output_dim(width, kernel_w, x_stride)
 
     # Decide how many sums our function should have running at the same time. Doing
     # this lets us do "more work" for each memory load, but doing too many of them causes us to run
     # out of registers. Currently this is set to either 1 or 2, but autotuning this value would

Review Comment:
   Can this sentence in the comment be removed since (I think?) it is no longer true?
   ```suggestion
       # out of registers. 
   ```



-- 
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: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org