apivovarov opened a new issue #4470: Conv2d_transpose kernel 2x2, strides (2,2) 
fails for CUDA - Cannot prove
URL: https://github.com/apache/incubator-tvm/issues/4470
 
 
   I found that conv2d_transpose op fails when kernel size is 2x2 and strides 
are (2,2).
   Error
   ```
   tvm/src/pass/loop_partition.cc:544: Cannot prove: ((((floordiv((((40 - 
(floordiv((dh + 1), 2)*8))*(5 - floordiv((dw + 1), 2))) + 63), 64) - 1) - (((40 
- (floordiv((dh + 1), 2)*8))*(5 - floordiv((dw + 1), 2))) - select((-63 <= ((40 
- (floordiv((dh + 1), 2)*8))*(5 - floordiv((dw + 1), 2)))), (floordiv((((40 - 
(floordiv((dh + 1), 2)*8))*(5 - floordiv((dw + 1), 2))) + 63), 64)*63), 0))) + 
1) >= 0), when generating the post doubt loop
   ```
   
   To reproduce the error
   ```
   import tvm
   from tvm import relay
   import tensorflow as tf
   input_tensor = "input_1"
   # NHWC
   input_shape=(1,16,16,8)
   x = tf.compat.v1.placeholder(tf.float32, shape=input_shape, 
name=input_tensor)
   # HWOI
   w2 = tf.compat.v1.placeholder(tf.float32, shape=(2,2,3,8))
   out_shape = tf.compat.v1.placeholder(tf.int32, shape=(4))
   deconv = tf.compat.v1.nn.conv2d_transpose(x, w2, out_shape, (1,2,2,1), 
padding='VALID')
   
   sess = tf.compat.v1.Session()
   graph_def = sess.graph_def
   
   mod, params = relay.frontend.from_tensorflow(graph_def, layout='NCHW', 
shape={input_tensor: input_shape})
   
   target = tvm.target.cuda()
   from tvm.autotvm.measure.measure_methods import set_cuda_target_arch
   set_cuda_target_arch('sm_70')
   
   with relay.build_config(opt_level=3):
       graph, lib, params = relay.build(mod, target, params=params)
   ```
   
   I tried other targets - llvm and arm_cpu - they are working fine. Only cuda 
fails.
   
   Related to [PR 4243](https://github.com/apache/incubator-tvm/pull/4243)
   
   

----------------------------------------------------------------
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

Reply via email to