test_topi_winograd_accuracy.test_conv2d_nchw ... 
  Workload: (1, 64, 56, 64, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  NEW winograd
  Output shape:  (1, 64, 56, 56)
  ABSdiff:  2.4282570039177755e-05
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  OLD winograd
  Output shape:  (1, 64, 56, 56)
  ABSdiff:  5.9233052460490975e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (1, 64, 56, 56)
  ABSdiff:  2.874101823673433e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (1, 64, 56, 56)
  ABSdiff:  2.9824263463133686e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  NEW winograd
  Output shape:  (1, 64, 56, 56)
  ABSdiff:  2.4278512789239528e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  OLD winograd
  Output shape:  (1, 64, 56, 56)
  ABSdiff:  5.923121114905956e-05
  
  Workload: (1, 128, 28, 128, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  NEW winograd
  Output shape:  (1, 128, 28, 28)
  ABSdiff:  4.086823490468969e-05
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  OLD winograd
  Output shape:  (1, 128, 28, 28)
  ABSdiff:  4.295053575566694e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (1, 128, 28, 28)
  ABSdiff:  6.377403168790755e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (1, 128, 28, 28)
  ABSdiff:  6.615732833516908e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  NEW winograd
  Output shape:  (1, 128, 28, 28)
  ABSdiff:  5.610743231937542e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  OLD winograd
  Output shape:  (1, 128, 28, 28)
  ABSdiff:  0.00011906119297902599
  
  Workload: (1, 256, 14, 256, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  NEW winograd
  Output shape:  (1, 256, 14, 14)
  ABSdiff:  0.00011452753281651969
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  OLD winograd
  Output shape:  (1, 256, 14, 14)
  ABSdiff:  0.00011711756063344491
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (1, 256, 14, 14)
  ABSdiff:  0.00018067234542971906
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (1, 256, 14, 14)
  ABSdiff:  0.00018482461536374872
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  NEW winograd
  Output shape:  (1, 256, 14, 14)
  ABSdiff:  0.00011452753281651969
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  OLD winograd
  Output shape:  (1, 256, 14, 14)
  ABSdiff:  0.00011711756063344491
  
  Workload: (1, 512, 7, 512, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  NEW winograd
  Output shape:  (1, 512, 7, 7)
  ABSdiff:  0.00028562422717012646
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  OLD winograd
  Output shape:  (1, 512, 7, 7)
  ABSdiff:  0.00028898457758124026
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (1, 512, 7, 7)
  ABSdiff:  0.00035035983085944435
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (1, 512, 7, 7)
  ABSdiff:  0.0003539133945559881
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  NEW winograd
  Output shape:  (1, 512, 7, 7)
  ABSdiff:  0.00028562422717012646
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  OLD winograd
  Output shape:  (1, 512, 7, 7)
  ABSdiff:  0.00028898457758124026
  
  Workload: (2, 64, 56, 64, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.4295486720469287e-05
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  6.016031071156796e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.8758394443551472e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.9855934781453023e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.429293860432221e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  6.017109940054672e-05
  
  Workload: (2, 64, 56, 64, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.4376490881431137e-05
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  5.987216617158271e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.89711684989349e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  3.0101267847319094e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.437644818154733e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  5.983677381832141e-05
  
  Workload: (2, 64, 56, 64, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.467622402628273e-05
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  5.976416940094564e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.879268482485703e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.9990959888087458e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.467363882866482e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  5.976702062699013e-05
  
  Workload: (2, 64, 56, 64, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.4978470027669167e-05
  Running on target: cuda
  [DBG] [cuda] Tile size = 4
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  5.9295613778137555e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.870490887443283e-05
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.985468871842946e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  NEW winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  2.499327669831756e-05
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 4
  OLD winograd
  Output shape:  (2, 64, 56, 56)
  ABSdiff:  5.929320151113948e-05
  
  Workload: (1, 1, 1, 1, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  NEW winograd
  Output shape:  (1, 1, 1, 1)
  ABSdiff:  1.0126328220394498e-08
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  OLD winograd
  Output shape:  (1, 1, 1, 1)
  ABSdiff:  6.973097299578512e-08
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (1, 1, 1, 1)
  ABSdiff:  1.2933561777117575e-07
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (1, 1, 1, 1)
  ABSdiff:  1.2933561777117575e-07
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  NEW winograd
  Output shape:  (1, 1, 1, 1)
  ABSdiff:  1.0126328220394498e-08
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  OLD winograd
  Output shape:  (1, 1, 1, 1)
  ABSdiff:  6.973097299578512e-08
  
  Workload: (3, 3, 3, 3, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  NEW winograd
  Output shape:  (3, 3, 3, 3)
  ABSdiff:  2.538112472183633e-07
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  OLD winograd
  Output shape:  (3, 3, 3, 3)
  ABSdiff:  3.021806142704887e-07
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (3, 3, 3, 3)
  ABSdiff:  1.1635106824505584e-06
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (3, 3, 3, 3)
  ABSdiff:  1.3849481189530543e-06
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  NEW winograd
  Output shape:  (3, 3, 3, 3)
  ABSdiff:  2.538112472183633e-07
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  OLD winograd
  Output shape:  (3, 3, 3, 3)
  ABSdiff:  3.021806142704887e-07
  
  Workload: (2, 13, 71, 59, 3, 1, 1, 1)
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  NEW winograd
  Output shape:  (2, 59, 71, 71)
  ABSdiff:  1.5785498161180955e-06
  Running on target: cuda
  [DBG] [cuda] Tile size = 2
  OLD winograd
  Output shape:  (2, 59, 71, 71)
  ABSdiff:  1.8478274250846262e-06
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  NEW winograd
  Output shape:  (2, 59, 71, 71)
  ABSdiff:  6.105624850066014e-06
  Running on target: llvm -device=arm_cpu
  [DBG] [arm_cpu] Tile size = 6
  OLD winograd
  Output shape:  (2, 59, 71, 71)
  ABSdiff:  6.367171834081284e-06
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  NEW winograd
  Output shape:  (2, 59, 71, 71)
  ABSdiff:  1.5785498161180955e-06
  Running on target: opencl -device=mali
  [DBG] [mali] Tile size = 2
  OLD winograd
  Output shape:  (2, 59, 71, 71)
  ABSdiff:  1.8478274250846262e-06
  ok
  
  ----------------------------------------------------------------------
  Ran 1 test in 241.348s
  
  OK