icemelon9 commented on a change in pull request #4949: Conv3D ONNX supprot and
conv3D_ncdhw x86 schedules
URL: https://github.com/apache/incubator-tvm/pull/4949#discussion_r384804848
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -197,6 +273,92 @@ def _conv3d_ndhwc(cfg, data, kernel, strides, padding,
dilation, out_dtype):
tag='conv3d_ndhwc')
return conv_unpacked
+def _conv3d_ncdhw(cfg, data, kernel, strides, padding, dilation, layout,
out_dtype):
+ out_dtype = data.dtype if out_dtype is None else out_dtype
+
+ assert isinstance(dilation, int) or len(dilation) == 3
+ if isinstance(dilation, int):
+ dilation_d, dilation_h, dilation_w = (dilation, dilation, dilation)
+ else:
+ dilation_d, dilation_h, dilation_w = dilation
+
+ DSTR, HSTR, WSTR = strides
+ batch_size, in_channel, in_depth, in_height, in_width =
get_const_tuple(data.shape)
+ num_filter, _, kernel_depth, kernel_height, kernel_width =
get_const_tuple(kernel.shape)
+
+ dilated_kernel_d = (kernel_depth - 1) * dilation_d + 1
+ dilated_kernel_h = (kernel_height - 1) * dilation_h + 1
+ dilated_kernel_w = (kernel_width - 1) * dilation_w + 1
+
+ pad_front, pad_top, pad_left, pad_back, pad_down, pad_right =
get_pad_tuple3d(
+ padding, (dilated_kernel_d, dilated_kernel_h, dilated_kernel_w))
+
+ pad_d = pad_front + pad_back
+ pad_h = pad_top + pad_down
+ pad_w = pad_left + pad_right
+
+ pad_depth = in_depth + pad_d
+ pad_height = in_height + pad_h
+ pad_width = in_width + pad_w
+
+ out_depth = simplify((in_depth + pad_d - dilated_kernel_d) // DSTR + 1)
+ out_height = simplify((in_height + pad_h - dilated_kernel_h) // HSTR + 1)
+ out_width = simplify((in_width + pad_w - dilated_kernel_w) // WSTR + 1)
+
+ # pack data
+ DOPAD = (pad_d != 0 or pad_h != 0 or pad_w != 0)
+ if DOPAD:
+ data_pad = pad(data, (0, 0, pad_front, pad_top, pad_left),
+ (0, 0, pad_back, pad_down, pad_right), name="data_pad")
+ else:
+ data_pad = data
+
+ # fetch schedule
+ ic_bn, oc_bn = cfg["tile_ic"].size[-1], cfg["tile_oc"].size[-1]
+
+ shape = (batch_size, in_channel // ic_bn, pad_depth, pad_height, ic_bn,
pad_width)
+ data_vec = tvm.compute(shape,
+ lambda n, C, d, h, c, w: data_pad[n, C * ic_bn + c,
d, h, w],
+ name='data_vec')
+
+ # pack kernel
+ shape = (num_filter//oc_bn, in_channel//ic_bn,
+ kernel_depth, kernel_height, kernel_width, ic_bn, oc_bn)
Review comment:
Since this schedule packs the kernel, we should consider adding
`alter_op_layout` for conv3d to avoid the overhead of kernel packing.
----------------------------------------------------------------
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
