cbalint13 commented on a change in pull request #5363: [Topi, ARM] Disbale
Winograd for quantized tensors.
URL: https://github.com/apache/incubator-tvm/pull/5363#discussion_r410686770
##########
File path: python/tvm/relay/op/strategy/arm_cpu.py
##########
@@ -59,16 +59,21 @@ def conv2d_strategy_arm_cpu(attrs, inputs, out_type,
target):
wrap_compute_conv2d(topi.arm_cpu.conv2d_nchw_spatial_pack),
wrap_topi_schedule(topi.arm_cpu.schedule_conv2d_nchw_spatial_pack),
name="conv2d_nchw_spatial_pack.arm_cpu")
+
# Intel x86 conv2d schedule.
strategy.add_implementation(
wrap_compute_conv2d(topi.x86.conv2d_nchw),
wrap_topi_schedule(topi.x86.schedule_conv2d_nchw),
name="conv2d_nchw.x86")
+
# check if winograd algorithm is applicable
_, _, kh, kw = get_const_tuple(kernel.shape)
pt, pl, pb, pr = topi.nn.get_pad_tuple(padding, (kh, kw))
- if kh == 3 and kw == 3 and stride_h == 1 and stride_w == 1 and
\
- dilation_h == 1 and dilation_w == 1:
+ is_dtype_fp32 = data.dtype == "float32" and kernel.dtype ==
"float32"
Review comment:
Float16 Too ? (perhaps double aka float64 too)
I get good results with it (e.g yolov3-tiny on mali).
Wouldn't be better to deny int8/int16/int32 ?
----------------------------------------------------------------
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
