ResNet50

June 4, 2020

(work in progress)

Model Architecture

The architecture of ResNet50 is described in the table below.

\begin{array}{c|c|c} \newcommand{\x}{\times} \hline \text{Layer #} & \text{Output Size} & \text{Operation} \\ \hline 1 & 112 \x 112 & 7 \x 7, 64, \text{ Stride } 2 \\ \hline & 56 \x 56 & 3 \x 3 \text{ MaxPool, Stride } 2 \\ \hline 2 & 56 \x 56 & \begin{bmatrix} 1\x1, 64 \\ 3\x3, 64 \\ 1\x1, 256 \\ \end{bmatrix} \x 3 \\ \hline 3 & 28 \x 28 & \begin{bmatrix} 1\x1, 128 \\ 3\x3, 128 \\ 1\x1, 512 \\ \end{bmatrix} \x 4 \\ \hline 4 & 14 \x 14 & \begin{bmatrix} 1\x1, 256 \\ 3\x3, 256 \\ 1\x1, 1024 \\ \end{bmatrix} \x 6 \\ \hline 5 & 7 \x 7 & \begin{bmatrix} 1\x1, 512 \\ 3\x3, 512 \\ 1\x1, 2048 \\ \end{bmatrix} \x 3 \\ \hline & 2048 & 7 \x 7 \text{ Average Pool} \\ \hline & 1000 & \text{Linear} \\ \hline \end{array}

defaults

ph=0 pw=0 U=1 V=1 dh=1 dw=1 g=1

GPU Kernels

Idx	Direction	Layer	Op	Params	Kernel
1	fprop	block_1	conv2d	N=32,C=3,H=224,W=224,K=64,P=112,Q=112,R=7,S=7,ph=3,pw=3,U=2,V=2	cudnn::gemm::computeOffsetsKernel
2	fprop	block_1	conv2d	N=32,C=3,H=224,W=224,K=64,P=112,Q=112,R=7,S=7,ph=3,pw=3,U=2,V=2	volta_fp16_scudnn_fp16_128x64_relu_medium_nn_v1
3	fprop	block_1	__add__	T=[(1,)]	legacy::elementwise_kernel
4	fprop	block_1	batch_norm	T=(32,64,112,112)	batch_norm_collect_statistics_kernel
5	fprop	block_1	batch_norm	T=(32,64,112,112)	batch_norm_transform_input_kernel
6	fprop	block_1	relu	T=(32,64,112,112)	modern::elementwise_kernel
7	fprop	block_1	max_pool2d	T=[(32,64,112,112)]	max_pool_forward_nchw

Idx	Direction	Layer	Op	Params	Kernel
8	fprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
9	fprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	volta_fp16_scudnn_fp16_128x64_relu_interior_nn_v1
10	fprop	2a:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
11	fprop	2a:BN1	batch_norm	T=(32,64,56,56)	batch_norm_collect_statistics_kernel
12	fprop	2a:BN1	batch_norm	T=(32,64,56,56)	batch_norm_transform_input_kernel
13	fprop	2a:ReLU1	relu	T=(32,64,56,56)	modern::elementwise_kernel
14	fprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
15	fprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
16	fprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
17	fprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_256x64_ldg8_relu_f2f_exp_small_nhwc_tn_v1
18	fprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
19	fprop	2a:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
20	fprop	2a:BN2	batch_norm	T=(32,64,56,56)	batch_norm_collect_statistics_kernel
21	fprop	2a:BN2	batch_norm	T=(32,64,56,56)	batch_norm_transform_input_kernel
22	fprop	2a:ReLU2	relu	T=(32,64,56,56)	modern::elementwise_kernel
23	fprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cask_cudnn::computeOffsetsKernel
24	fprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
25	fprop	2a:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
26	fprop	2a:BN3	batch_norm	T=(32,256,56,56)	batch_norm_collect_statistics_kernel
27	fprop	2a:BN3	batch_norm	T=(32,256,56,56)	batch_norm_transform_input_kernel
28	fprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cask_cudnn::computeOffsetsKernel
29	fprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
30	fprop	2a:Residual:Projection	__add__	T=[(1,)]	legacy::elementwise_kernel
31	fprop	2a:Residual:Projection	batch_norm	T=(32,256,56,56)	batch_norm_collect_statistics_kernel
32	fprop	2a:Residual:Projection	batch_norm	T=(32,256,56,56)	batch_norm_transform_input_kernel
33	fprop	2a:Residual	__iadd__	T=[(32,256,56,56),(32,256,56,56)]	modern::elementwise_kernel
34	fprop	2a:ReLU3	relu	T=(32,256,56,56)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
35	fprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cask_cudnn::computeOffsetsKernel
36	fprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
37	fprop	2b:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
38	fprop	2b:BN1	batch_norm	T=(32,64,56,56)	batch_norm_collect_statistics_kernel
39	fprop	2b:BN1	batch_norm	T=(32,64,56,56)	batch_norm_transform_input_kernel
40	fprop	2b:ReLU1	relu	T=(32,64,56,56)	modern::elementwise_kernel
41	fprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
42	fprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
43	fprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
44	fprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_256x64_ldg8_relu_f2f_exp_small_nhwc_tn_v1
45	fprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
46	fprop	2b:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
47	fprop	2b:BN2	batch_norm	T=(32,64,56,56)	batch_norm_collect_statistics_kernel
48	fprop	2b:BN2	batch_norm	T=(32,64,56,56)	batch_norm_transform_input_kernel
49	fprop	2b:ReLU2	relu	T=(32,64,56,56)	modern::elementwise_kernel
50	fprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cask_cudnn::computeOffsetsKernel
51	fprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
52	fprop	2b:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
53	fprop	2b:BN3	batch_norm	T=(32,256,56,56)	batch_norm_collect_statistics_kernel
54	fprop	2b:BN3	batch_norm	T=(32,256,56,56)	batch_norm_transform_input_kernel
55	fprop	2b:Residual	__iadd__	T=[(32,256,56,56),(32,256,56,56)]	modern::elementwise_kernel
56	fprop	2b:ReLU3	relu	T=(32,256,56,56)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
57	fprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cask_cudnn::computeOffsetsKernel
58	fprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
59	fprop	2c:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
60	fprop	2c:BN1	batch_norm	T=(32,64,56,56)	batch_norm_collect_statistics_kernel
61	fprop	2c:BN1	batch_norm	T=(32,64,56,56)	batch_norm_transform_input_kernel
62	fprop	2c:ReLU1	relu	T=(32,64,56,56)	modern::elementwise_kernel
63	fprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
64	fprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
65	fprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
66	fprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_256x64_ldg8_relu_f2f_exp_small_nhwc_tn_v1
67	fprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
68	fprop	2c:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
69	fprop	2c:BN2	batch_norm	T=(32,64,56,56)	batch_norm_collect_statistics_kernel
70	fprop	2c:BN2	batch_norm	T=(32,64,56,56)	batch_norm_transform_input_kernel
71	fprop	2c:ReLU2	relu	T=(32,64,56,56)	modern::elementwise_kernel
72	fprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cask_cudnn::computeOffsetsKernel
73	fprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
74	fprop	2c:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
75	fprop	2c:BN3	batch_norm	T=(32,256,56,56)	batch_norm_collect_statistics_kernel
76	fprop	2c:BN3	batch_norm	T=(32,256,56,56)	batch_norm_transform_input_kernel
77	fprop	2c:Residual	__iadd__	T=[(32,256,56,56),(32,256,56,56)]	modern::elementwise_kernel
78	fprop	2c:ReLU3	relu	T=(32,256,56,56)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
79	fprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	cask_cudnn::computeOffsetsKernel
80	fprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
81	fprop	3a:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
82	fprop	3a:BN1	batch_norm	T=(32,128,56,56)	batch_norm_collect_statistics_kernel
83	fprop	3a:BN1	batch_norm	T=(32,128,56,56)	batch_norm_transform_input_kernel
84	fprop	3a:ReLU1	relu	T=(32,128,56,56)	modern::elementwise_kernel
85	fprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
86	fprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
87	fprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	cudnn::gemm::computeOffsetsKernel
88	fprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
89	fprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
90	fprop	3a:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
91	fprop	3a:BN2	batch_norm	T=(32,128,28,28)	batch_norm_collect_statistics_kernel
92	fprop	3a:BN2	batch_norm	T=(32,128,28,28)	batch_norm_transform_input_kernel
93	fprop	3a:ReLU2	relu	T=(32,128,28,28)	modern::elementwise_kernel
94	fprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cask_cudnn::computeOffsetsKernel
95	fprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
96	fprop	3a:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
97	fprop	3a:BN3	batch_norm	T=(32,512,28,28)	batch_norm_collect_statistics_kernel
98	fprop	3a:BN3	batch_norm	T=(32,512,28,28)	batch_norm_transform_input_kernel
99	fprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	nchwToNhwcKernel
100	fprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	cudnn::gemm::computeOffsetsKernel
101	fprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
102	fprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	nhwcToNchwKernel
103	fprop	3a:Residual:Projection	__add__	T=[(1,)]	legacy::elementwise_kernel
104	fprop	3a:Residual:Projection	batch_norm	T=(32,512,28,28)	batch_norm_collect_statistics_kernel
105	fprop	3a:Residual:Projection	batch_norm	T=(32,512,28,28)	batch_norm_transform_input_kernel
106	fprop	3a:Residual	__iadd__	T=[(32,512,28,28),(32,512,28,28)]	modern::elementwise_kernel
107	fprop	3a:ReLU3	relu	T=(32,512,28,28)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
108	fprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cask_cudnn::computeOffsetsKernel
109	fprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
110	fprop	3b:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
111	fprop	3b:BN1	batch_norm	T=(32,128,28,28)	batch_norm_collect_statistics_kernel
112	fprop	3b:BN1	batch_norm	T=(32,128,28,28)	batch_norm_transform_input_kernel
113	fprop	3b:ReLU1	relu	T=(32,128,28,28)	modern::elementwise_kernel
114	fprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
115	fprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
116	fprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
117	fprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
118	fprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
119	fprop	3b:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
120	fprop	3b:BN2	batch_norm	T=(32,128,28,28)	batch_norm_collect_statistics_kernel
121	fprop	3b:BN2	batch_norm	T=(32,128,28,28)	batch_norm_transform_input_kernel
122	fprop	3b:ReLU2	relu	T=(32,128,28,28)	modern::elementwise_kernel
123	fprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cask_cudnn::computeOffsetsKernel
124	fprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
125	fprop	3b:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
126	fprop	3b:BN3	batch_norm	T=(32,512,28,28)	batch_norm_collect_statistics_kernel
127	fprop	3b:BN3	batch_norm	T=(32,512,28,28)	batch_norm_transform_input_kernel
128	fprop	3b:Residual	__iadd__	T=[(32,512,28,28),(32,512,28,28)]	modern::elementwise_kernel
129	fprop	3b:ReLU3	relu	T=(32,512,28,28)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
130	fprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cask_cudnn::computeOffsetsKernel
131	fprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
132	fprop	3c:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
133	fprop	3c:BN1	batch_norm	T=(32,128,28,28)	batch_norm_collect_statistics_kernel
134	fprop	3c:BN1	batch_norm	T=(32,128,28,28)	batch_norm_transform_input_kernel
135	fprop	3c:ReLU1	relu	T=(32,128,28,28)	modern::elementwise_kernel
136	fprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
137	fprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
138	fprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
139	fprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
140	fprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
141	fprop	3c:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
142	fprop	3c:BN2	batch_norm	T=(32,128,28,28)	batch_norm_collect_statistics_kernel
143	fprop	3c:BN2	batch_norm	T=(32,128,28,28)	batch_norm_transform_input_kernel
144	fprop	3c:ReLU2	relu	T=(32,128,28,28)	modern::elementwise_kernel
145	fprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cask_cudnn::computeOffsetsKernel
146	fprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
147	fprop	3c:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
148	fprop	3c:BN3	batch_norm	T=(32,512,28,28)	batch_norm_collect_statistics_kernel
149	fprop	3c:BN3	batch_norm	T=(32,512,28,28)	batch_norm_transform_input_kernel
150	fprop	3c:Residual	__iadd__	T=[(32,512,28,28),(32,512,28,28)]	modern::elementwise_kernel
151	fprop	3c:ReLU3	relu	T=(32,512,28,28)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
152	fprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cask_cudnn::computeOffsetsKernel
153	fprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
154	fprop	3d:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
155	fprop	3d:BN1	batch_norm	T=(32,128,28,28)	batch_norm_collect_statistics_kernel
156	fprop	3d:BN1	batch_norm	T=(32,128,28,28)	batch_norm_transform_input_kernel
157	fprop	3d:ReLU1	relu	T=(32,128,28,28)	modern::elementwise_kernel
158	fprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
159	fprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
160	fprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
161	fprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
162	fprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
163	fprop	3d:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
164	fprop	3d:BN2	batch_norm	T=(32,128,28,28)	batch_norm_collect_statistics_kernel
165	fprop	3d:BN2	batch_norm	T=(32,128,28,28)	batch_norm_transform_input_kernel
166	fprop	3d:ReLU2	relu	T=(32,128,28,28)	modern::elementwise_kernel
167	fprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cask_cudnn::computeOffsetsKernel
168	fprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
169	fprop	3d:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
170	fprop	3d:BN3	batch_norm	T=(32,512,28,28)	batch_norm_collect_statistics_kernel
171	fprop	3d:BN3	batch_norm	T=(32,512,28,28)	batch_norm_transform_input_kernel
172	fprop	3d:Residual	__iadd__	T=[(32,512,28,28),(32,512,28,28)]	modern::elementwise_kernel
173	fprop	3d:ReLU3	relu	T=(32,512,28,28)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
174	fprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	cask_cudnn::computeOffsetsKernel
175	fprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_filter1x1_stg8_interior_nchw_nn_v1
176	fprop	4a:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
177	fprop	4a:BN1	batch_norm	T=(32,256,28,28)	batch_norm_collect_statistics_kernel
178	fprop	4a:BN1	batch_norm	T=(32,256,28,28)	batch_norm_transform_input_kernel
179	fprop	4a:ReLU1	relu	T=(32,256,28,28)	modern::elementwise_kernel
180	fprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
181	fprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
182	fprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	cudnn::gemm::computeOffsetsKernel
183	fprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
184	fprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
185	fprop	4a:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
186	fprop	4a:BN2	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
187	fprop	4a:BN2	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
188	fprop	4a:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
189	fprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
190	fprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
191	fprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
192	fprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
193	fprop	4a:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
194	fprop	4a:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_collect_statistics_kernel
195	fprop	4a:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_transform_input_kernel
196	fprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	nchwToNhwcKernel
197	fprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	cudnn::gemm::computeOffsetsKernel
198	fprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
199	fprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	nhwcToNchwKernel
200	fprop	4a:Residual:Projection	__add__	T=[(1,)]	legacy::elementwise_kernel
201	fprop	4a:Residual:Projection	batch_norm	T=(32,1024,14,14)	batch_norm_collect_statistics_kernel
202	fprop	4a:Residual:Projection	batch_norm	T=(32,1024,14,14)	batch_norm_transform_input_kernel
203	fprop	4a:Residual	__iadd__	T=[(32,1024,14,14),(32,1024,14,14)]	modern::elementwise_kernel
204	fprop	4a:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
205	fprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
206	fprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
207	fprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
208	fprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
209	fprop	4b:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
210	fprop	4b:BN1	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
211	fprop	4b:BN1	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
212	fprop	4b:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
213	fprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
214	fprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
215	fprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
216	fprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
217	fprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
218	fprop	4b:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
219	fprop	4b:BN2	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
220	fprop	4b:BN2	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
221	fprop	4b:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
222	fprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
223	fprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
224	fprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
225	fprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
226	fprop	4b:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
227	fprop	4b:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_collect_statistics_kernel
228	fprop	4b:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_transform_input_kernel
229	fprop	4b:Residual	__iadd__	T=[(32,1024,14,14),(32,1024,14,14)]	modern::elementwise_kernel
230	fprop	4b:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
231	fprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
232	fprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
233	fprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
234	fprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
235	fprop	4c:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
236	fprop	4c:BN1	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
237	fprop	4c:BN1	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
238	fprop	4c:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
239	fprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
240	fprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
241	fprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
242	fprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
243	fprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
244	fprop	4c:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
245	fprop	4c:BN2	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
246	fprop	4c:BN2	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
247	fprop	4c:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
248	fprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
249	fprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
250	fprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
251	fprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
252	fprop	4c:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
253	fprop	4c:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_collect_statistics_kernel
254	fprop	4c:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_transform_input_kernel
255	fprop	4c:Residual	__iadd__	T=[(32,1024,14,14),(32,1024,14,14)]	modern::elementwise_kernel
256	fprop	4c:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
257	fprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
258	fprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
259	fprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
260	fprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
261	fprop	4d:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
262	fprop	4d:BN1	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
263	fprop	4d:BN1	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
264	fprop	4d:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
265	fprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
266	fprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
267	fprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
268	fprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
269	fprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
270	fprop	4d:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
271	fprop	4d:BN2	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
272	fprop	4d:BN2	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
273	fprop	4d:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
274	fprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
275	fprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
276	fprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
277	fprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
278	fprop	4d:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
279	fprop	4d:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_collect_statistics_kernel
280	fprop	4d:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_transform_input_kernel
281	fprop	4d:Residual	__iadd__	T=[(32,1024,14,14),(32,1024,14,14)]	modern::elementwise_kernel
282	fprop	4d:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
283	fprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
284	fprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
285	fprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
286	fprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
287	fprop	4e:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
288	fprop	4e:BN1	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
289	fprop	4e:BN1	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
290	fprop	4e:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
291	fprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
292	fprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
293	fprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
294	fprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
295	fprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
296	fprop	4e:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
297	fprop	4e:BN2	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
298	fprop	4e:BN2	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
299	fprop	4e:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
300	fprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
301	fprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
302	fprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
303	fprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
304	fprop	4e:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
305	fprop	4e:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_collect_statistics_kernel
306	fprop	4e:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_transform_input_kernel
307	fprop	4e:Residual	__iadd__	T=[(32,1024,14,14),(32,1024,14,14)]	modern::elementwise_kernel
308	fprop	4e:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
309	fprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
310	fprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
311	fprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
312	fprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
313	fprop	4f:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
314	fprop	4f:BN1	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
315	fprop	4f:BN1	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
316	fprop	4f:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
317	fprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
318	fprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
319	fprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
320	fprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
321	fprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
322	fprop	4f:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
323	fprop	4f:BN2	batch_norm	T=(32,256,14,14)	batch_norm_collect_statistics_kernel
324	fprop	4f:BN2	batch_norm	T=(32,256,14,14)	batch_norm_transform_input_kernel
325	fprop	4f:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
326	fprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
327	fprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
328	fprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
329	fprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
330	fprop	4f:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
331	fprop	4f:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_collect_statistics_kernel
332	fprop	4f:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_transform_input_kernel
333	fprop	4f:Residual	__iadd__	T=[(32,1024,14,14),(32,1024,14,14)]	modern::elementwise_kernel
334	fprop	4f:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
335	fprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
336	fprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
337	fprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
338	fprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
339	fprop	5a:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
340	fprop	5a:BN1	batch_norm	T=(32,512,14,14)	batch_norm_collect_statistics_kernel
341	fprop	5a:BN1	batch_norm	T=(32,512,14,14)	batch_norm_transform_input_kernel
342	fprop	5a:ReLU1	relu	T=(32,512,14,14)	modern::elementwise_kernel
343	fprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
344	fprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
345	fprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	cudnn::gemm::computeOffsetsKernel
346	fprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
347	fprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
348	fprop	5a:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
349	fprop	5a:BN2	batch_norm	T=(32,512,7,7)	batch_norm_collect_statistics_kernel
350	fprop	5a:BN2	batch_norm	T=(32,512,7,7)	batch_norm_transform_input_kernel
351	fprop	5a:ReLU2	relu	T=(32,512,7,7)	modern::elementwise_kernel
352	fprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
353	fprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
354	fprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
355	fprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
356	fprop	5a:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
357	fprop	5a:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_collect_statistics_kernel
358	fprop	5a:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_transform_input_kernel
359	fprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	nchwToNhwcKernel
360	fprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	cudnn::gemm::computeOffsetsKernel
361	fprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
362	fprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	nhwcToNchwKernel
363	fprop	5a:Residual:Projection	__add__	T=[(1,)]	legacy::elementwise_kernel
364	fprop	5a:Residual:Projection	batch_norm	T=(32,2048,7,7)	batch_norm_collect_statistics_kernel
365	fprop	5a:Residual:Projection	batch_norm	T=(32,2048,7,7)	batch_norm_transform_input_kernel
366	fprop	5a:Residual	__iadd__	T=[(32,2048,7,7),(32,2048,7,7)]	modern::elementwise_kernel
367	fprop	5a:ReLU3	relu	T=(32,2048,7,7)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
368	fprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
369	fprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
370	fprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
371	fprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
372	fprop	5b:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
373	fprop	5b:BN1	batch_norm	T=(32,512,7,7)	batch_norm_collect_statistics_kernel
374	fprop	5b:BN1	batch_norm	T=(32,512,7,7)	batch_norm_transform_input_kernel
375	fprop	5b:ReLU1	relu	T=(32,512,7,7)	modern::elementwise_kernel
376	fprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
377	fprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
378	fprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
379	fprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
380	fprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
381	fprop	5b:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
382	fprop	5b:BN2	batch_norm	T=(32,512,7,7)	batch_norm_collect_statistics_kernel
383	fprop	5b:BN2	batch_norm	T=(32,512,7,7)	batch_norm_transform_input_kernel
384	fprop	5b:ReLU2	relu	T=(32,512,7,7)	modern::elementwise_kernel
385	fprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
386	fprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
387	fprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
388	fprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
389	fprop	5b:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
390	fprop	5b:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_collect_statistics_kernel
391	fprop	5b:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_transform_input_kernel
392	fprop	5b:Residual	__iadd__	T=[(32,2048,7,7),(32,2048,7,7)]	modern::elementwise_kernel
393	fprop	5b:ReLU3	relu	T=(32,2048,7,7)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
394	fprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
395	fprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
396	fprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
397	fprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
398	fprop	5c:BN1	__add__	T=[(1,)]	legacy::elementwise_kernel
399	fprop	5c:BN1	batch_norm	T=(32,512,7,7)	batch_norm_collect_statistics_kernel
400	fprop	5c:BN1	batch_norm	T=(32,512,7,7)	batch_norm_transform_input_kernel
401	fprop	5c:ReLU1	relu	T=(32,512,7,7)	modern::elementwise_kernel
402	fprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
403	fprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
404	fprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
405	fprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	turing_fp16_s1688cudnn_fp16_128x128_ldg8_relu_f2f_exp_small_nhwc_tn_v1
406	fprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
407	fprop	5c:BN2	__add__	T=[(1,)]	legacy::elementwise_kernel
408	fprop	5c:BN2	batch_norm	T=(32,512,7,7)	batch_norm_collect_statistics_kernel
409	fprop	5c:BN2	batch_norm	T=(32,512,7,7)	batch_norm_transform_input_kernel
410	fprop	5c:ReLU2	relu	T=(32,512,7,7)	modern::elementwise_kernel
411	fprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
412	fprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
413	fprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	turing_fp16_s1688cudnn_fp16_256x128_ldg8_relu_f2f_exp_interior_nhwc_tn_v1
414	fprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
415	fprop	5c:BN3	__add__	T=[(1,)]	legacy::elementwise_kernel
416	fprop	5c:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_collect_statistics_kernel
417	fprop	5c:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_transform_input_kernel
418	fprop	5c:Residual	__iadd__	T=[(32,2048,7,7),(32,2048,7,7)]	modern::elementwise_kernel
419	fprop	5c:ReLU3	relu	T=(32,2048,7,7)	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
420	fprop	-	adaptive_avg_pool2d	T=[(32,2048,7,7)]	reduce_kernel
421	fprop	FC	bias	M=1000,N=32	legacy::elementwise_kernel
422	fprop	FC	linear	M=1000,N=32,K=2048	turing_fp16_s1688gemm_fp16_256x64_ldg8_f2f_tn
423	fprop	FC	linear	M=1000,N=32,K=2048	splitKreduce_kernel

Idx	Direction	Layer	Op	Params	Kernel
424	fprop	-	cross_entropy	T=[(32,1000),(32,)],[,]	softmax_warp_forward
425	fprop	-	cross_entropy	T=[(32,1000),(32,)],[,]	cunn_ClassNLLCriterion_updateOutput_kernel

Idx	Direction	Layer	Op	Params	Kernel
426	fprop	-	zero_	T=[(64,3,7,7)]	modern::elementwise_kernel
427	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
428	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
429	fprop	-	zero_	T=[(64,64,1,1)]	modern::elementwise_kernel
430	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
431	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
432	fprop	-	zero_	T=[(64,64,3,3)]	modern::elementwise_kernel
433	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
434	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
435	fprop	-	zero_	T=[(256,64,1,1)]	modern::elementwise_kernel
436	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
437	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
438	fprop	-	zero_	T=[(256,64,1,1)]	modern::elementwise_kernel
439	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
440	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
441	fprop	-	zero_	T=[(64,256,1,1)]	modern::elementwise_kernel
442	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
443	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
444	fprop	-	zero_	T=[(64,64,3,3)]	modern::elementwise_kernel
445	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
446	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
447	fprop	-	zero_	T=[(256,64,1,1)]	modern::elementwise_kernel
448	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
449	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
450	fprop	-	zero_	T=[(64,256,1,1)]	modern::elementwise_kernel
451	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
452	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
453	fprop	-	zero_	T=[(64,64,3,3)]	modern::elementwise_kernel
454	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
455	fprop	-	zero_	T=[(64,)]	modern::elementwise_kernel
456	fprop	-	zero_	T=[(256,64,1,1)]	modern::elementwise_kernel
457	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
458	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
459	fprop	-	zero_	T=[(128,256,1,1)]	modern::elementwise_kernel
460	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
461	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
462	fprop	-	zero_	T=[(128,128,3,3)]	modern::elementwise_kernel
463	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
464	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
465	fprop	-	zero_	T=[(512,128,1,1)]	modern::elementwise_kernel
466	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
467	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
468	fprop	-	zero_	T=[(512,256,1,1)]	modern::elementwise_kernel
469	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
470	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
471	fprop	-	zero_	T=[(128,512,1,1)]	modern::elementwise_kernel
472	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
473	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
474	fprop	-	zero_	T=[(128,128,3,3)]	modern::elementwise_kernel
475	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
476	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
477	fprop	-	zero_	T=[(512,128,1,1)]	modern::elementwise_kernel
478	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
479	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
480	fprop	-	zero_	T=[(128,512,1,1)]	modern::elementwise_kernel
481	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
482	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
483	fprop	-	zero_	T=[(128,128,3,3)]	modern::elementwise_kernel
484	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
485	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
486	fprop	-	zero_	T=[(512,128,1,1)]	modern::elementwise_kernel
487	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
488	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
489	fprop	-	zero_	T=[(128,512,1,1)]	modern::elementwise_kernel
490	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
491	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
492	fprop	-	zero_	T=[(128,128,3,3)]	modern::elementwise_kernel
493	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
494	fprop	-	zero_	T=[(128,)]	modern::elementwise_kernel
495	fprop	-	zero_	T=[(512,128,1,1)]	modern::elementwise_kernel
496	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
497	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
498	fprop	-	zero_	T=[(256,512,1,1)]	modern::elementwise_kernel
499	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
500	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
501	fprop	-	zero_	T=[(256,256,3,3)]	modern::elementwise_kernel
502	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
503	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
504	fprop	-	zero_	T=[(1024,256,1,1)]	modern::elementwise_kernel
505	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
506	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
507	fprop	-	zero_	T=[(1024,512,1,1)]	modern::elementwise_kernel
508	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
509	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
510	fprop	-	zero_	T=[(256,1024,1,1)]	modern::elementwise_kernel
511	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
512	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
513	fprop	-	zero_	T=[(256,256,3,3)]	modern::elementwise_kernel
514	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
515	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
516	fprop	-	zero_	T=[(1024,256,1,1)]	modern::elementwise_kernel
517	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
518	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
519	fprop	-	zero_	T=[(256,1024,1,1)]	modern::elementwise_kernel
520	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
521	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
522	fprop	-	zero_	T=[(256,256,3,3)]	modern::elementwise_kernel
523	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
524	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
525	fprop	-	zero_	T=[(1024,256,1,1)]	modern::elementwise_kernel
526	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
527	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
528	fprop	-	zero_	T=[(256,1024,1,1)]	modern::elementwise_kernel
529	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
530	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
531	fprop	-	zero_	T=[(256,256,3,3)]	modern::elementwise_kernel
532	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
533	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
534	fprop	-	zero_	T=[(1024,256,1,1)]	modern::elementwise_kernel
535	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
536	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
537	fprop	-	zero_	T=[(256,1024,1,1)]	modern::elementwise_kernel
538	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
539	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
540	fprop	-	zero_	T=[(256,256,3,3)]	modern::elementwise_kernel
541	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
542	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
543	fprop	-	zero_	T=[(1024,256,1,1)]	modern::elementwise_kernel
544	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
545	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
546	fprop	-	zero_	T=[(256,1024,1,1)]	modern::elementwise_kernel
547	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
548	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
549	fprop	-	zero_	T=[(256,256,3,3)]	modern::elementwise_kernel
550	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
551	fprop	-	zero_	T=[(256,)]	modern::elementwise_kernel
552	fprop	-	zero_	T=[(1024,256,1,1)]	modern::elementwise_kernel
553	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
554	fprop	-	zero_	T=[(1024,)]	modern::elementwise_kernel
555	fprop	-	zero_	T=[(512,1024,1,1)]	modern::elementwise_kernel
556	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
557	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
558	fprop	-	zero_	T=[(512,512,3,3)]	modern::elementwise_kernel
559	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
560	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
561	fprop	-	zero_	T=[(2048,512,1,1)]	modern::elementwise_kernel
562	fprop	-	zero_	T=[(2048,)]	modern::elementwise_kernel
563	fprop	-	zero_	T=[(2048,)]	modern::elementwise_kernel
564	fprop	-	zero_	T=[(2048,1024,1,1)]	modern::elementwise_kernel
565	fprop	-	zero_	T=[(2048,)]	modern::elementwise_kernel
566	fprop	-	zero_	T=[(2048,)]	modern::elementwise_kernel
567	fprop	-	zero_	T=[(512,2048,1,1)]	modern::elementwise_kernel
568	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
569	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
570	fprop	-	zero_	T=[(512,512,3,3)]	modern::elementwise_kernel
571	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
572	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
573	fprop	-	zero_	T=[(2048,512,1,1)]	modern::elementwise_kernel
574	fprop	-	zero_	T=[(2048,)]	modern::elementwise_kernel
575	fprop	-	zero_	T=[(2048,)]	modern::elementwise_kernel
576	fprop	-	zero_	T=[(512,2048,1,1)]	modern::elementwise_kernel
577	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
578	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
579	fprop	-	zero_	T=[(512,512,3,3)]	modern::elementwise_kernel
580	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
581	fprop	-	zero_	T=[(512,)]	modern::elementwise_kernel
582	fprop	-	zero_	T=[(2048,512,1,1)]	modern::elementwise_kernel
583	fprop	-	zero_	T=[(2048,)]	modern::elementwise_kernel
584	fprop	-	zero_	T=[(2048,)]	modern::elementwise_kernel
585	fprop	-	zero_	T=[(1000,2048)]	modern::elementwise_kernel
586	fprop	-	zero_	T=[(1000,)]	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
587	fprop	-	backward	T=,	legacy::elementwise_kernel
588	bprop	-	cross_entropy	T=[(32,1000),(32,)],[,]	cunn_ClassNLLCriterion_updateGradInput_kernel
589	bprop	-	cross_entropy	T=[(32,1000),(32,)],[,]	softmax_warp_backward
590	bprop	FC	linear	M=2048,N=32,K=1000	turing_fp16_s1688gemm_fp16_256x64_ldg8_f2f_nn
591	bprop	FC	linear	M=2048,N=1000,K=32	splitKreduce_kernel
592	bprop	FC	linear	X=(32,2048),W=(1000,2048)	turing_fp16_s1688gemm_fp16_128x128_ldg8_f2f_nt
593	fprop	-	sum	na	reduce_kernel
594	fprop	-	add_	na	modern::elementwise_kernel
595	fprop	-	add_	na	modern::elementwise_kernel
596	fprop	-	div	na	legacy::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
597	bprop	5c:ReLU3	relu	T=(32,2048,7,7)	modern::elementwise_kernel
598	bprop	5c:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_backward_kernel
599	fprop	-	add_	na	modern::elementwise_kernel
600	fprop	-	add_	na	modern::elementwise_kernel
601	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
602	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
603	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
604	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
605	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
606	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
607	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
608	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	scalePackedTensor_kernel
609	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
610	bprop	5c:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
611	fprop	-	add_	na	modern::elementwise_kernel
612	bprop	5c:ReLU2	relu	T=(32,512,7,7)	modern::elementwise_kernel
613	bprop	5c:BN2	batch_norm	T=(32,512,7,7)	batch_norm_backward_kernel
614	fprop	-	add_	na	modern::elementwise_kernel
615	fprop	-	add_	na	modern::elementwise_kernel
616	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
617	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
618	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
619	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
620	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
621	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
622	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
623	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
624	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
625	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
626	bprop	5c:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
627	fprop	-	add_	na	modern::elementwise_kernel
628	bprop	5c:ReLU1	relu	T=(32,512,7,7)	modern::elementwise_kernel
629	bprop	5c:BN1	batch_norm	T=(32,512,7,7)	batch_norm_backward_kernel
630	fprop	-	add_	na	modern::elementwise_kernel
631	fprop	-	add_	na	modern::elementwise_kernel
632	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
633	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
634	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
635	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
636	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
637	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
638	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
639	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	scalePackedTensor_kernel
640	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
641	bprop	5c:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
642	fprop	-	add	na	modern::elementwise_kernel
643	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
644	bprop	5b:ReLU3	relu	T=(32,2048,7,7)	modern::elementwise_kernel
645	bprop	5b:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_backward_kernel
646	fprop	-	add_	na	modern::elementwise_kernel
647	fprop	-	add_	na	modern::elementwise_kernel
648	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
649	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
650	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
651	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
652	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
653	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
654	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
655	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	scalePackedTensor_kernel
656	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
657	bprop	5b:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
658	fprop	-	add_	na	modern::elementwise_kernel
659	bprop	5b:ReLU2	relu	T=(32,512,7,7)	modern::elementwise_kernel
660	bprop	5b:BN2	batch_norm	T=(32,512,7,7)	batch_norm_backward_kernel
661	fprop	-	add_	na	modern::elementwise_kernel
662	fprop	-	add_	na	modern::elementwise_kernel
663	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
664	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
665	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
666	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
667	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
668	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
669	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
670	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
671	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
672	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
673	bprop	5b:Conv2	conv2d	N=32,C=512,H=7,W=7,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
674	fprop	-	add_	na	modern::elementwise_kernel
675	bprop	5b:ReLU1	relu	T=(32,512,7,7)	modern::elementwise_kernel
676	bprop	5b:BN1	batch_norm	T=(32,512,7,7)	batch_norm_backward_kernel
677	fprop	-	add_	na	modern::elementwise_kernel
678	fprop	-	add_	na	modern::elementwise_kernel
679	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
680	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
681	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
682	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
683	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
684	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
685	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
686	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	scalePackedTensor_kernel
687	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
688	bprop	5b:Conv1	conv2d	N=32,C=2048,H=7,W=7,K=512,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
689	fprop	-	add	na	modern::elementwise_kernel
690	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
691	bprop	5a:ReLU3	relu	T=(32,2048,7,7)	modern::elementwise_kernel
692	bprop	5a:Residual:Projection	batch_norm	T=(32,2048,7,7)	batch_norm_backward_kernel
693	fprop	-	add_	na	modern::elementwise_kernel
694	fprop	-	add_	na	modern::elementwise_kernel
695	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	nchwToNhwcKernel
696	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	dgrad_1x1_stride_2x2
697	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	nhwcToNchwKernel
698	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	nchwToNhwcKernel
699	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	nchwToNhwcKernel
700	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	cudnn::gemm::computeWgradOffsetsKernel
701	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	scalePackedTensor_kernel
702	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
703	bprop	5a:Residual:Projection	conv2d	N=32,C=1024,H=14,W=14,K=2048,P=7,Q=7,R=1,S=1,U=2,V=2	nhwcToNchwKernel
704	fprop	-	add_	na	modern::elementwise_kernel
705	bprop	5a:BN3	batch_norm	T=(32,2048,7,7)	batch_norm_backward_kernel
706	fprop	-	add_	na	modern::elementwise_kernel
707	fprop	-	add_	na	modern::elementwise_kernel
708	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
709	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeOffsetsKernel
710	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
711	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
712	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
713	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nchwToNhwcKernel
714	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
715	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	scalePackedTensor_kernel
716	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
717	bprop	5a:Conv3	conv2d	N=32,C=512,H=7,W=7,K=2048,P=7,Q=7,R=1,S=1	nhwcToNchwKernel
718	fprop	-	add_	na	modern::elementwise_kernel
719	bprop	5a:ReLU2	relu	T=(32,512,7,7)	modern::elementwise_kernel
720	bprop	5a:BN2	batch_norm	T=(32,512,7,7)	batch_norm_backward_kernel
721	fprop	-	add_	na	modern::elementwise_kernel
722	fprop	-	add_	na	modern::elementwise_kernel
723	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
724	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
725	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	dgrad_1d
726	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
727	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
728	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
729	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	cudnn::gemm::computeWgradOffsetsKernel
730	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	scalePackedTensor_kernel
731	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
732	bprop	5a:Conv2	conv2d	N=32,C=512,H=14,W=14,K=512,P=7,Q=7,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
733	fprop	-	add_	na	modern::elementwise_kernel
734	bprop	5a:ReLU1	relu	T=(32,512,14,14)	modern::elementwise_kernel
735	bprop	5a:BN1	batch_norm	T=(32,512,14,14)	batch_norm_backward_kernel
736	fprop	-	add_	na	modern::elementwise_kernel
737	fprop	-	add_	na	modern::elementwise_kernel
738	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
739	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
740	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
741	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
742	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
743	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
744	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
745	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
746	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
747	bprop	5a:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=512,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
748	fprop	-	add	na	modern::elementwise_kernel
749	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
750	bprop	4f:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel
751	bprop	4f:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_backward_kernel
752	fprop	-	add_	na	modern::elementwise_kernel
753	fprop	-	add_	na	modern::elementwise_kernel
754	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
755	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
756	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
757	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
758	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
759	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
760	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
761	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
762	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
763	bprop	4f:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
764	fprop	-	add_	na	modern::elementwise_kernel
765	bprop	4f:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
766	bprop	4f:BN2	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
767	fprop	-	add_	na	modern::elementwise_kernel
768	fprop	-	add_	na	modern::elementwise_kernel
769	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
770	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
771	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
772	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
773	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
774	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
775	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
776	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
777	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
778	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
779	bprop	4f:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
780	fprop	-	add_	na	modern::elementwise_kernel
781	bprop	4f:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
782	bprop	4f:BN1	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
783	fprop	-	add_	na	modern::elementwise_kernel
784	fprop	-	add_	na	modern::elementwise_kernel
785	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
786	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
787	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
788	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
789	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
790	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
791	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
792	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
793	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
794	bprop	4f:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
795	fprop	-	add	na	modern::elementwise_kernel
796	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
797	bprop	4e:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel
798	bprop	4e:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_backward_kernel
799	fprop	-	add_	na	modern::elementwise_kernel
800	fprop	-	add_	na	modern::elementwise_kernel
801	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
802	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
803	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
804	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
805	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
806	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
807	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
808	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
809	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
810	bprop	4e:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
811	fprop	-	add_	na	modern::elementwise_kernel
812	bprop	4e:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
813	bprop	4e:BN2	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
814	fprop	-	add_	na	modern::elementwise_kernel
815	fprop	-	add_	na	modern::elementwise_kernel
816	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
817	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
818	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
819	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
820	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
821	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
822	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
823	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
824	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
825	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
826	bprop	4e:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
827	fprop	-	add_	na	modern::elementwise_kernel
828	bprop	4e:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
829	bprop	4e:BN1	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
830	fprop	-	add_	na	modern::elementwise_kernel
831	fprop	-	add_	na	modern::elementwise_kernel
832	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
833	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
834	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
835	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
836	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
837	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
838	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
839	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
840	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
841	bprop	4e:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
842	fprop	-	add	na	modern::elementwise_kernel
843	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
844	bprop	4d:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel
845	bprop	4d:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_backward_kernel
846	fprop	-	add_	na	modern::elementwise_kernel
847	fprop	-	add_	na	modern::elementwise_kernel
848	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
849	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
850	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
851	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
852	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
853	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
854	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
855	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
856	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
857	bprop	4d:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
858	fprop	-	add_	na	modern::elementwise_kernel
859	bprop	4d:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
860	bprop	4d:BN2	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
861	fprop	-	add_	na	modern::elementwise_kernel
862	fprop	-	add_	na	modern::elementwise_kernel
863	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
864	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
865	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
866	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
867	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
868	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
869	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
870	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
871	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
872	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
873	bprop	4d:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
874	fprop	-	add_	na	modern::elementwise_kernel
875	bprop	4d:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
876	bprop	4d:BN1	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
877	fprop	-	add_	na	modern::elementwise_kernel
878	fprop	-	add_	na	modern::elementwise_kernel
879	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
880	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
881	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
882	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
883	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
884	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
885	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
886	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
887	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
888	bprop	4d:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
889	fprop	-	add	na	modern::elementwise_kernel
890	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
891	bprop	4c:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel
892	bprop	4c:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_backward_kernel
893	fprop	-	add_	na	modern::elementwise_kernel
894	fprop	-	add_	na	modern::elementwise_kernel
895	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
896	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
897	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
898	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
899	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
900	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
901	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
902	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
903	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
904	bprop	4c:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
905	fprop	-	add_	na	modern::elementwise_kernel
906	bprop	4c:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
907	bprop	4c:BN2	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
908	fprop	-	add_	na	modern::elementwise_kernel
909	fprop	-	add_	na	modern::elementwise_kernel
910	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
911	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
912	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
913	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
914	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
915	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
916	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
917	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
918	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
919	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
920	bprop	4c:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
921	fprop	-	add_	na	modern::elementwise_kernel
922	bprop	4c:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
923	bprop	4c:BN1	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
924	fprop	-	add_	na	modern::elementwise_kernel
925	fprop	-	add_	na	modern::elementwise_kernel
926	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
927	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
928	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
929	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
930	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
931	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
932	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
933	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
934	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
935	bprop	4c:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
936	fprop	-	add	na	modern::elementwise_kernel
937	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
938	bprop	4b:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel
939	bprop	4b:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_backward_kernel
940	fprop	-	add_	na	modern::elementwise_kernel
941	fprop	-	add_	na	modern::elementwise_kernel
942	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
943	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
944	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
945	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
946	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
947	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
948	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
949	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
950	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
951	bprop	4b:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
952	fprop	-	add_	na	modern::elementwise_kernel
953	bprop	4b:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
954	bprop	4b:BN2	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
955	fprop	-	add_	na	modern::elementwise_kernel
956	fprop	-	add_	na	modern::elementwise_kernel
957	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
958	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
959	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
960	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
961	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
962	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
963	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
964	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
965	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
966	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
967	bprop	4b:Conv2	conv2d	N=32,C=256,H=14,W=14,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
968	fprop	-	add_	na	modern::elementwise_kernel
969	bprop	4b:ReLU1	relu	T=(32,256,14,14)	modern::elementwise_kernel
970	bprop	4b:BN1	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
971	fprop	-	add_	na	modern::elementwise_kernel
972	fprop	-	add_	na	modern::elementwise_kernel
973	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
974	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
975	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
976	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
977	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
978	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
979	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
980	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
981	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
982	bprop	4b:Conv1	conv2d	N=32,C=1024,H=14,W=14,K=256,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
983	fprop	-	add	na	modern::elementwise_kernel
984	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
985	bprop	4a:ReLU3	relu	T=(32,1024,14,14)	modern::elementwise_kernel
986	bprop	4a:Residual:Projection	batch_norm	T=(32,1024,14,14)	batch_norm_backward_kernel
987	fprop	-	add_	na	modern::elementwise_kernel
988	fprop	-	add_	na	modern::elementwise_kernel
989	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	nchwToNhwcKernel
990	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	dgrad_1x1_stride_2x2
991	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	nhwcToNchwKernel
992	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	nchwToNhwcKernel
993	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	nchwToNhwcKernel
994	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	cudnn::gemm::computeWgradOffsetsKernel
995	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	scalePackedTensor_kernel
996	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
997	bprop	4a:Residual:Projection	conv2d	N=32,C=512,H=28,W=28,K=1024,P=14,Q=14,R=1,S=1,U=2,V=2	nhwcToNchwKernel
998	fprop	-	add_	na	modern::elementwise_kernel
999	bprop	4a:BN3	batch_norm	T=(32,1024,14,14)	batch_norm_backward_kernel
1000	fprop	-	add_	na	modern::elementwise_kernel
1001	fprop	-	add_	na	modern::elementwise_kernel
1002	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
1003	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1004	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1005	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1006	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
1007	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nchwToNhwcKernel
1008	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1009	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	scalePackedTensor_kernel
1010	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1011	bprop	4a:Conv3	conv2d	N=32,C=256,H=14,W=14,K=1024,P=14,Q=14,R=1,S=1	nhwcToNchwKernel
1012	fprop	-	add_	na	modern::elementwise_kernel
1013	bprop	4a:ReLU2	relu	T=(32,256,14,14)	modern::elementwise_kernel
1014	bprop	4a:BN2	batch_norm	T=(32,256,14,14)	batch_norm_backward_kernel
1015	fprop	-	add_	na	modern::elementwise_kernel
1016	fprop	-	add_	na	modern::elementwise_kernel
1017	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
1018	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
1019	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	dgrad_1d
1020	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
1021	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
1022	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
1023	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	cudnn::gemm::computeWgradOffsetsKernel
1024	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	scalePackedTensor_kernel
1025	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1026	bprop	4a:Conv2	conv2d	N=32,C=256,H=28,W=28,K=256,P=14,Q=14,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
1027	fprop	-	add_	na	modern::elementwise_kernel
1028	bprop	4a:ReLU1	relu	T=(32,256,28,28)	modern::elementwise_kernel
1029	bprop	4a:BN1	batch_norm	T=(32,256,28,28)	batch_norm_backward_kernel
1030	fprop	-	add_	na	modern::elementwise_kernel
1031	fprop	-	add_	na	modern::elementwise_kernel
1032	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1033	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1034	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1035	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1036	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1037	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1038	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1039	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	scalePackedTensor_kernel
1040	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1041	bprop	4a:Conv1	conv2d	N=32,C=512,H=28,W=28,K=256,P=28,Q=28,R=1,S=1	nhwcToNchwKernel
1042	fprop	-	add	na	modern::elementwise_kernel
1043	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1044	bprop	3d:ReLU3	relu	T=(32,512,28,28)	modern::elementwise_kernel
1045	bprop	3d:BN3	batch_norm	T=(32,512,28,28)	batch_norm_backward_kernel
1046	fprop	-	add_	na	modern::elementwise_kernel
1047	fprop	-	add_	na	modern::elementwise_kernel
1048	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1049	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1050	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1051	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1052	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1053	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1054	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1055	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	scalePackedTensor_kernel
1056	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1057	bprop	3d:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nhwcToNchwKernel
1058	fprop	-	add_	na	modern::elementwise_kernel
1059	bprop	3d:ReLU2	relu	T=(32,128,28,28)	modern::elementwise_kernel
1060	bprop	3d:BN2	batch_norm	T=(32,128,28,28)	batch_norm_backward_kernel
1061	fprop	-	add_	na	modern::elementwise_kernel
1062	fprop	-	add_	na	modern::elementwise_kernel
1063	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1064	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1065	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
1066	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
1067	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
1068	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1069	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1070	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
1071	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
1072	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1073	bprop	3d:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
1074	fprop	-	add_	na	modern::elementwise_kernel
1075	bprop	3d:ReLU1	relu	T=(32,128,28,28)	modern::elementwise_kernel
1076	bprop	3d:BN1	batch_norm	T=(32,128,28,28)	batch_norm_backward_kernel
1077	fprop	-	add_	na	modern::elementwise_kernel
1078	fprop	-	add_	na	modern::elementwise_kernel
1079	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1080	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1081	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1082	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1083	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1084	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1085	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1086	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	scalePackedTensor_kernel
1087	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1088	bprop	3d:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nhwcToNchwKernel
1089	fprop	-	add	na	modern::elementwise_kernel
1090	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1091	bprop	3c:ReLU3	relu	T=(32,512,28,28)	modern::elementwise_kernel
1092	bprop	3c:BN3	batch_norm	T=(32,512,28,28)	batch_norm_backward_kernel
1093	fprop	-	add_	na	modern::elementwise_kernel
1094	fprop	-	add_	na	modern::elementwise_kernel
1095	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1096	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1097	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1098	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1099	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1100	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1101	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1102	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	scalePackedTensor_kernel
1103	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1104	bprop	3c:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nhwcToNchwKernel
1105	fprop	-	add_	na	modern::elementwise_kernel
1106	bprop	3c:ReLU2	relu	T=(32,128,28,28)	modern::elementwise_kernel
1107	bprop	3c:BN2	batch_norm	T=(32,128,28,28)	batch_norm_backward_kernel
1108	fprop	-	add_	na	modern::elementwise_kernel
1109	fprop	-	add_	na	modern::elementwise_kernel
1110	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1111	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1112	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
1113	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
1114	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
1115	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1116	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1117	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
1118	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
1119	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1120	bprop	3c:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
1121	fprop	-	add_	na	modern::elementwise_kernel
1122	bprop	3c:ReLU1	relu	T=(32,128,28,28)	modern::elementwise_kernel
1123	bprop	3c:BN1	batch_norm	T=(32,128,28,28)	batch_norm_backward_kernel
1124	fprop	-	add_	na	modern::elementwise_kernel
1125	fprop	-	add_	na	modern::elementwise_kernel
1126	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1127	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1128	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1129	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1130	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1131	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1132	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1133	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	scalePackedTensor_kernel
1134	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1135	bprop	3c:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nhwcToNchwKernel
1136	fprop	-	add	na	modern::elementwise_kernel
1137	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1138	bprop	3b:ReLU3	relu	T=(32,512,28,28)	modern::elementwise_kernel
1139	bprop	3b:BN3	batch_norm	T=(32,512,28,28)	batch_norm_backward_kernel
1140	fprop	-	add_	na	modern::elementwise_kernel
1141	fprop	-	add_	na	modern::elementwise_kernel
1142	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1143	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1144	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1145	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1146	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1147	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1148	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1149	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	scalePackedTensor_kernel
1150	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1151	bprop	3b:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nhwcToNchwKernel
1152	fprop	-	add_	na	modern::elementwise_kernel
1153	bprop	3b:ReLU2	relu	T=(32,128,28,28)	modern::elementwise_kernel
1154	bprop	3b:BN2	batch_norm	T=(32,128,28,28)	batch_norm_backward_kernel
1155	fprop	-	add_	na	modern::elementwise_kernel
1156	fprop	-	add_	na	modern::elementwise_kernel
1157	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1158	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1159	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
1160	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
1161	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
1162	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1163	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1164	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
1165	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
1166	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1167	bprop	3b:Conv2	conv2d	N=32,C=128,H=28,W=28,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
1168	fprop	-	add_	na	modern::elementwise_kernel
1169	bprop	3b:ReLU1	relu	T=(32,128,28,28)	modern::elementwise_kernel
1170	bprop	3b:BN1	batch_norm	T=(32,128,28,28)	batch_norm_backward_kernel
1171	fprop	-	add_	na	modern::elementwise_kernel
1172	fprop	-	add_	na	modern::elementwise_kernel
1173	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1174	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1175	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1176	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1177	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1178	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1179	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1180	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	scalePackedTensor_kernel
1181	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1182	bprop	3b:Conv1	conv2d	N=32,C=512,H=28,W=28,K=128,P=28,Q=28,R=1,S=1	nhwcToNchwKernel
1183	fprop	-	add	na	modern::elementwise_kernel
1184	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1185	bprop	3a:ReLU3	relu	T=(32,512,28,28)	modern::elementwise_kernel
1186	bprop	3a:Residual:Projection	batch_norm	T=(32,512,28,28)	batch_norm_backward_kernel
1187	fprop	-	add_	na	modern::elementwise_kernel
1188	fprop	-	add_	na	modern::elementwise_kernel
1189	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	nchwToNhwcKernel
1190	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	dgrad_1x1_stride_2x2
1191	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	nhwcToNchwKernel
1192	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	nchwToNhwcKernel
1193	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	nchwToNhwcKernel
1194	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	cudnn::gemm::computeWgradOffsetsKernel
1195	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	scalePackedTensor_kernel
1196	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1197	bprop	3a:Residual:Projection	conv2d	N=32,C=256,H=56,W=56,K=512,P=28,Q=28,R=1,S=1,U=2,V=2	nhwcToNchwKernel
1198	fprop	-	add_	na	modern::elementwise_kernel
1199	bprop	3a:BN3	batch_norm	T=(32,512,28,28)	batch_norm_backward_kernel
1200	fprop	-	add_	na	modern::elementwise_kernel
1201	fprop	-	add_	na	modern::elementwise_kernel
1202	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1203	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1204	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1205	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	volta_fp16_s884cudnn_fp16_128x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1206	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1207	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nchwToNhwcKernel
1208	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1209	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	scalePackedTensor_kernel
1210	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1211	bprop	3a:Conv3	conv2d	N=32,C=128,H=28,W=28,K=512,P=28,Q=28,R=1,S=1	nhwcToNchwKernel
1212	fprop	-	add_	na	modern::elementwise_kernel
1213	bprop	3a:ReLU2	relu	T=(32,128,28,28)	modern::elementwise_kernel
1214	bprop	3a:BN2	batch_norm	T=(32,128,28,28)	batch_norm_backward_kernel
1215	fprop	-	add_	na	modern::elementwise_kernel
1216	fprop	-	add_	na	modern::elementwise_kernel
1217	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
1218	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
1219	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	dgrad_2d
1220	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
1221	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
1222	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nchwToNhwcKernel
1223	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	cudnn::gemm::computeWgradOffsetsKernel
1224	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	scalePackedTensor_kernel
1225	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1226	bprop	3a:Conv2	conv2d	N=32,C=128,H=56,W=56,K=128,P=28,Q=28,R=3,S=3,ph=1,pw=1,U=2,V=2	nhwcToNchwKernel
1227	fprop	-	add_	na	modern::elementwise_kernel
1228	bprop	3a:ReLU1	relu	T=(32,128,56,56)	modern::elementwise_kernel
1229	bprop	3a:BN1	batch_norm	T=(32,128,56,56)	batch_norm_backward_kernel
1230	fprop	-	add_	na	modern::elementwise_kernel
1231	fprop	-	add_	na	modern::elementwise_kernel
1232	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1233	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1234	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1235	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1236	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1237	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1238	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1239	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	scalePackedTensor_kernel
1240	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1241	bprop	3a:Conv1	conv2d	N=32,C=256,H=56,W=56,K=128,P=56,Q=56,R=1,S=1	nhwcToNchwKernel
1242	fprop	-	add	na	modern::elementwise_kernel
1243	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1244	bprop	2c:ReLU3	relu	T=(32,256,56,56)	modern::elementwise_kernel
1245	bprop	2c:BN3	batch_norm	T=(32,256,56,56)	batch_norm_backward_kernel
1246	fprop	-	add_	na	modern::elementwise_kernel
1247	fprop	-	add_	na	modern::elementwise_kernel
1248	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1249	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1250	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1251	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	volta_fp16_s884cudnn_fp16_256x64_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1252	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1253	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1254	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1255	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	scalePackedTensor_kernel
1256	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1257	bprop	2c:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nhwcToNchwKernel
1258	fprop	-	add_	na	modern::elementwise_kernel
1259	bprop	2c:ReLU2	relu	T=(32,64,56,56)	modern::elementwise_kernel
1260	bprop	2c:BN2	batch_norm	T=(32,64,56,56)	batch_norm_backward_kernel
1261	fprop	-	add_	na	modern::elementwise_kernel
1262	fprop	-	add_	na	modern::elementwise_kernel
1263	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1264	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1265	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
1266	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
1267	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_256x64_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
1268	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1269	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1270	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
1271	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
1272	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1273	bprop	2c:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
1274	fprop	-	add_	na	modern::elementwise_kernel
1275	bprop	2c:ReLU1	relu	T=(32,64,56,56)	modern::elementwise_kernel
1276	bprop	2c:BN1	batch_norm	T=(32,64,56,56)	batch_norm_backward_kernel
1277	fprop	-	add_	na	modern::elementwise_kernel
1278	fprop	-	add_	na	modern::elementwise_kernel
1279	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1280	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1281	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1282	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1283	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1284	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1285	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1286	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	scalePackedTensor_kernel
1287	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1288	bprop	2c:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nhwcToNchwKernel
1289	fprop	-	add	na	modern::elementwise_kernel
1290	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1291	bprop	2b:ReLU3	relu	T=(32,256,56,56)	modern::elementwise_kernel
1292	bprop	2b:BN3	batch_norm	T=(32,256,56,56)	batch_norm_backward_kernel
1293	fprop	-	add_	na	modern::elementwise_kernel
1294	fprop	-	add_	na	modern::elementwise_kernel
1295	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1296	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1297	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1298	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	volta_fp16_s884cudnn_fp16_256x64_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1299	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1300	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1301	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1302	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	scalePackedTensor_kernel
1303	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1304	bprop	2b:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nhwcToNchwKernel
1305	fprop	-	add_	na	modern::elementwise_kernel
1306	bprop	2b:ReLU2	relu	T=(32,64,56,56)	modern::elementwise_kernel
1307	bprop	2b:BN2	batch_norm	T=(32,64,56,56)	batch_norm_backward_kernel
1308	fprop	-	add_	na	modern::elementwise_kernel
1309	fprop	-	add_	na	modern::elementwise_kernel
1310	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1311	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1312	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
1313	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
1314	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_256x64_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
1315	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1316	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1317	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
1318	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
1319	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1320	bprop	2b:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
1321	fprop	-	add_	na	modern::elementwise_kernel
1322	bprop	2b:ReLU1	relu	T=(32,64,56,56)	modern::elementwise_kernel
1323	bprop	2b:BN1	batch_norm	T=(32,64,56,56)	batch_norm_backward_kernel
1324	fprop	-	add_	na	modern::elementwise_kernel
1325	fprop	-	add_	na	modern::elementwise_kernel
1326	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1327	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1328	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1329	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	volta_fp16_s884cudnn_fp16_256x128_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1330	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1331	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1332	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1333	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	scalePackedTensor_kernel
1334	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1335	bprop	2b:Conv1	conv2d	N=32,C=256,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nhwcToNchwKernel
1336	fprop	-	add	na	modern::elementwise_kernel
1337	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1338	bprop	2a:ReLU3	relu	T=(32,256,56,56)	modern::elementwise_kernel
1339	bprop	2a:Residual:Projection	batch_norm	T=(32,256,56,56)	batch_norm_backward_kernel
1340	fprop	-	add_	na	modern::elementwise_kernel
1341	fprop	-	add_	na	modern::elementwise_kernel
1342	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1343	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1344	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1345	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	volta_fp16_s884cudnn_fp16_256x64_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1346	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1347	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1348	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1349	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	scalePackedTensor_kernel
1350	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1351	bprop	2a:Residual:Projection	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nhwcToNchwKernel
1352	fprop	-	add_	na	modern::elementwise_kernel
1353	bprop	2a:BN3	batch_norm	T=(32,256,56,56)	batch_norm_backward_kernel
1354	fprop	-	add_	na	modern::elementwise_kernel
1355	fprop	-	add_	na	modern::elementwise_kernel
1356	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1357	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1358	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1359	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	volta_fp16_s884cudnn_fp16_256x64_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1360	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1361	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1362	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1363	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	scalePackedTensor_kernel
1364	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1365	bprop	2a:Conv3	conv2d	N=32,C=64,H=56,W=56,K=256,P=56,Q=56,R=1,S=1	nhwcToNchwKernel
1366	fprop	-	add_	na	modern::elementwise_kernel
1367	bprop	2a:ReLU2	relu	T=(32,64,56,56)	modern::elementwise_kernel
1368	bprop	2a:BN2	batch_norm	T=(32,64,56,56)	batch_norm_backward_kernel
1369	fprop	-	add_	na	modern::elementwise_kernel
1370	fprop	-	add_	na	modern::elementwise_kernel
1371	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1372	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1373	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeOffsetsKernel
1374	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeBOffsetsKernel
1375	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	volta_fp16_s884cudnn_fp16_256x64_ldg8_dgrad_f2f_exp_small_nhwc2nchw_tt_v1
1376	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1377	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nchwToNhwcKernel
1378	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	cudnn::gemm::computeWgradOffsetsKernel
1379	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	scalePackedTensor_kernel
1380	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1381	bprop	2a:Conv2	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=3,S=3,ph=1,pw=1	nhwcToNchwKernel
1382	fprop	-	add_	na	modern::elementwise_kernel
1383	bprop	2a:ReLU1	relu	T=(32,64,56,56)	modern::elementwise_kernel
1384	bprop	2a:BN1	batch_norm	T=(32,64,56,56)	batch_norm_backward_kernel
1385	fprop	-	add_	na	modern::elementwise_kernel
1386	fprop	-	add_	na	modern::elementwise_kernel
1387	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1388	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeOffsetsKernel
1389	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeBOffsetsKernel
1390	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	volta_fp16_s884cudnn_fp16_256x64_ldg8_dgrad_f2f_exp_interior_nhwc2nchw_tt_v1
1391	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1392	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nchwToNhwcKernel
1393	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	cudnn::gemm::computeWgradOffsetsKernel
1394	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	scalePackedTensor_kernel
1395	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	turing_s1688cudnn_fp16_128x128_ldg8_wgrad_idx_exp_interior_nhwc_nt_v1
1396	bprop	2a:Conv1	conv2d	N=32,C=64,H=56,W=56,K=64,P=56,Q=56,R=1,S=1	nhwcToNchwKernel
1397	fprop	-	add	na	modern::elementwise_kernel
1398	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1399	bprop	block_1	max_pool2d	T=[(32,64,112,112)]	modern::elementwise_kernel
1400	bprop	block_1	max_pool2d	T=[(32,64,112,112)]	max_pool_backward_nchw
1401	bprop	block_1	relu	T=(32,64,112,112)	modern::elementwise_kernel
1402	bprop	block_1	batch_norm	T=(32,64,112,112)	batch_norm_backward_kernel
1403	fprop	-	add_	na	modern::elementwise_kernel
1404	fprop	-	add_	na	modern::elementwise_kernel
1405	bprop	block_1	conv2d	N=32,C=3,H=224,W=224,K=64,P=112,Q=112,R=7,S=7,ph=3,pw=3,U=2,V=2	nchwToNhwc3To4Kernel
1406	bprop	block_1	conv2d	N=32,C=3,H=224,W=224,K=64,P=112,Q=112,R=7,S=7,ph=3,pw=3,U=2,V=2	nchwToNhwcKernel
1407	bprop	block_1	conv2d	N=32,C=3,H=224,W=224,K=64,P=112,Q=112,R=7,S=7,ph=3,pw=3,U=2,V=2	cask_cudnn::first_layer_wgrad_kernel
1408	bprop	block_1	conv2d	N=32,C=3,H=224,W=224,K=64,P=112,Q=112,R=7,S=7,ph=3,pw=3,U=2,V=2	generic4To3Channel_kernel
1409	fprop	-	add_	na	modern::elementwise_kernel

Idx	Direction	Layer	Op	Params	Kernel
1410	fprop	-	mul_	T=[(64,3,7,7)]	modern::elementwise_kernel
1411	fprop	-	add_	T=[(64,3,7,7),(64,3,7,7)]	modern::elementwise_kernel
1412	fprop	-	add_	T=[(64,3,7,7),(64,3,7,7)]	modern::elementwise_kernel
1413	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1414	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1415	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1416	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1417	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1418	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1419	fprop	-	mul_	T=[(64,64,1,1)]	modern::elementwise_kernel
1420	fprop	-	add_	T=[(64,64,1,1),(64,64,1,1)]	modern::elementwise_kernel
1421	fprop	-	add_	T=[(64,64,1,1),(64,64,1,1)]	modern::elementwise_kernel
1422	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1423	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1424	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1425	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1426	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1427	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1428	fprop	-	mul_	T=[(64,64,3,3)]	modern::elementwise_kernel
1429	fprop	-	add_	T=[(64,64,3,3),(64,64,3,3)]	modern::elementwise_kernel
1430	fprop	-	add_	T=[(64,64,3,3),(64,64,3,3)]	modern::elementwise_kernel
1431	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1432	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1433	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1434	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1435	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1436	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1437	fprop	-	mul_	T=[(256,64,1,1)]	modern::elementwise_kernel
1438	fprop	-	add_	T=[(256,64,1,1),(256,64,1,1)]	modern::elementwise_kernel
1439	fprop	-	add_	T=[(256,64,1,1),(256,64,1,1)]	modern::elementwise_kernel
1440	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1441	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1442	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1443	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1444	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1445	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1446	fprop	-	mul_	T=[(256,64,1,1)]	modern::elementwise_kernel
1447	fprop	-	add_	T=[(256,64,1,1),(256,64,1,1)]	modern::elementwise_kernel
1448	fprop	-	add_	T=[(256,64,1,1),(256,64,1,1)]	modern::elementwise_kernel
1449	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1450	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1451	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1452	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1453	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1454	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1455	fprop	-	mul_	T=[(64,256,1,1)]	modern::elementwise_kernel
1456	fprop	-	add_	T=[(64,256,1,1),(64,256,1,1)]	modern::elementwise_kernel
1457	fprop	-	add_	T=[(64,256,1,1),(64,256,1,1)]	modern::elementwise_kernel
1458	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1459	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1460	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1461	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1462	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1463	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1464	fprop	-	mul_	T=[(64,64,3,3)]	modern::elementwise_kernel
1465	fprop	-	add_	T=[(64,64,3,3),(64,64,3,3)]	modern::elementwise_kernel
1466	fprop	-	add_	T=[(64,64,3,3),(64,64,3,3)]	modern::elementwise_kernel
1467	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1468	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1469	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1470	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1471	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1472	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1473	fprop	-	mul_	T=[(256,64,1,1)]	modern::elementwise_kernel
1474	fprop	-	add_	T=[(256,64,1,1),(256,64,1,1)]	modern::elementwise_kernel
1475	fprop	-	add_	T=[(256,64,1,1),(256,64,1,1)]	modern::elementwise_kernel
1476	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1477	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1478	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1479	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1480	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1481	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1482	fprop	-	mul_	T=[(64,256,1,1)]	modern::elementwise_kernel
1483	fprop	-	add_	T=[(64,256,1,1),(64,256,1,1)]	modern::elementwise_kernel
1484	fprop	-	add_	T=[(64,256,1,1),(64,256,1,1)]	modern::elementwise_kernel
1485	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1486	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1487	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1488	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1489	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1490	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1491	fprop	-	mul_	T=[(64,64,3,3)]	modern::elementwise_kernel
1492	fprop	-	add_	T=[(64,64,3,3),(64,64,3,3)]	modern::elementwise_kernel
1493	fprop	-	add_	T=[(64,64,3,3),(64,64,3,3)]	modern::elementwise_kernel
1494	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1495	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1496	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1497	fprop	-	mul_	T=[(64,)]	modern::elementwise_kernel
1498	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1499	fprop	-	add_	T=[(64,),(64,)]	modern::elementwise_kernel
1500	fprop	-	mul_	T=[(256,64,1,1)]	modern::elementwise_kernel
1501	fprop	-	add_	T=[(256,64,1,1),(256,64,1,1)]	modern::elementwise_kernel
1502	fprop	-	add_	T=[(256,64,1,1),(256,64,1,1)]	modern::elementwise_kernel
1503	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1504	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1505	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1506	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1507	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1508	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1509	fprop	-	mul_	T=[(128,256,1,1)]	modern::elementwise_kernel
1510	fprop	-	add_	T=[(128,256,1,1),(128,256,1,1)]	modern::elementwise_kernel
1511	fprop	-	add_	T=[(128,256,1,1),(128,256,1,1)]	modern::elementwise_kernel
1512	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1513	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1514	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1515	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1516	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1517	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1518	fprop	-	mul_	T=[(128,128,3,3)]	modern::elementwise_kernel
1519	fprop	-	add_	T=[(128,128,3,3),(128,128,3,3)]	modern::elementwise_kernel
1520	fprop	-	add_	T=[(128,128,3,3),(128,128,3,3)]	modern::elementwise_kernel
1521	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1522	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1523	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1524	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1525	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1526	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1527	fprop	-	mul_	T=[(512,128,1,1)]	modern::elementwise_kernel
1528	fprop	-	add_	T=[(512,128,1,1),(512,128,1,1)]	modern::elementwise_kernel
1529	fprop	-	add_	T=[(512,128,1,1),(512,128,1,1)]	modern::elementwise_kernel
1530	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1531	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1532	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1533	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1534	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1535	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1536	fprop	-	mul_	T=[(512,256,1,1)]	modern::elementwise_kernel
1537	fprop	-	add_	T=[(512,256,1,1),(512,256,1,1)]	modern::elementwise_kernel
1538	fprop	-	add_	T=[(512,256,1,1),(512,256,1,1)]	modern::elementwise_kernel
1539	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1540	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1541	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1542	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1543	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1544	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1545	fprop	-	mul_	T=[(128,512,1,1)]	modern::elementwise_kernel
1546	fprop	-	add_	T=[(128,512,1,1),(128,512,1,1)]	modern::elementwise_kernel
1547	fprop	-	add_	T=[(128,512,1,1),(128,512,1,1)]	modern::elementwise_kernel
1548	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1549	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1550	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1551	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1552	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1553	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1554	fprop	-	mul_	T=[(128,128,3,3)]	modern::elementwise_kernel
1555	fprop	-	add_	T=[(128,128,3,3),(128,128,3,3)]	modern::elementwise_kernel
1556	fprop	-	add_	T=[(128,128,3,3),(128,128,3,3)]	modern::elementwise_kernel
1557	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1558	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1559	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1560	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1561	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1562	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1563	fprop	-	mul_	T=[(512,128,1,1)]	modern::elementwise_kernel
1564	fprop	-	add_	T=[(512,128,1,1),(512,128,1,1)]	modern::elementwise_kernel
1565	fprop	-	add_	T=[(512,128,1,1),(512,128,1,1)]	modern::elementwise_kernel
1566	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1567	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1568	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1569	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1570	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1571	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1572	fprop	-	mul_	T=[(128,512,1,1)]	modern::elementwise_kernel
1573	fprop	-	add_	T=[(128,512,1,1),(128,512,1,1)]	modern::elementwise_kernel
1574	fprop	-	add_	T=[(128,512,1,1),(128,512,1,1)]	modern::elementwise_kernel
1575	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1576	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1577	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1578	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1579	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1580	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1581	fprop	-	mul_	T=[(128,128,3,3)]	modern::elementwise_kernel
1582	fprop	-	add_	T=[(128,128,3,3),(128,128,3,3)]	modern::elementwise_kernel
1583	fprop	-	add_	T=[(128,128,3,3),(128,128,3,3)]	modern::elementwise_kernel
1584	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1585	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1586	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1587	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1588	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1589	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1590	fprop	-	mul_	T=[(512,128,1,1)]	modern::elementwise_kernel
1591	fprop	-	add_	T=[(512,128,1,1),(512,128,1,1)]	modern::elementwise_kernel
1592	fprop	-	add_	T=[(512,128,1,1),(512,128,1,1)]	modern::elementwise_kernel
1593	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1594	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1595	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1596	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1597	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1598	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1599	fprop	-	mul_	T=[(128,512,1,1)]	modern::elementwise_kernel
1600	fprop	-	add_	T=[(128,512,1,1),(128,512,1,1)]	modern::elementwise_kernel
1601	fprop	-	add_	T=[(128,512,1,1),(128,512,1,1)]	modern::elementwise_kernel
1602	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1603	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1604	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1605	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1606	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1607	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1608	fprop	-	mul_	T=[(128,128,3,3)]	modern::elementwise_kernel
1609	fprop	-	add_	T=[(128,128,3,3),(128,128,3,3)]	modern::elementwise_kernel
1610	fprop	-	add_	T=[(128,128,3,3),(128,128,3,3)]	modern::elementwise_kernel
1611	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1612	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1613	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1614	fprop	-	mul_	T=[(128,)]	modern::elementwise_kernel
1615	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1616	fprop	-	add_	T=[(128,),(128,)]	modern::elementwise_kernel
1617	fprop	-	mul_	T=[(512,128,1,1)]	modern::elementwise_kernel
1618	fprop	-	add_	T=[(512,128,1,1),(512,128,1,1)]	modern::elementwise_kernel
1619	fprop	-	add_	T=[(512,128,1,1),(512,128,1,1)]	modern::elementwise_kernel
1620	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1621	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1622	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1623	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1624	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1625	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1626	fprop	-	mul_	T=[(256,512,1,1)]	modern::elementwise_kernel
1627	fprop	-	add_	T=[(256,512,1,1),(256,512,1,1)]	modern::elementwise_kernel
1628	fprop	-	add_	T=[(256,512,1,1),(256,512,1,1)]	modern::elementwise_kernel
1629	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1630	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1631	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1632	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1633	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1634	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1635	fprop	-	mul_	T=[(256,256,3,3)]	modern::elementwise_kernel
1636	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1637	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1638	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1639	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1640	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1641	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1642	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1643	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1644	fprop	-	mul_	T=[(1024,256,1,1)]	modern::elementwise_kernel
1645	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1646	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1647	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1648	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1649	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1650	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1651	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1652	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1653	fprop	-	mul_	T=[(1024,512,1,1)]	modern::elementwise_kernel
1654	fprop	-	add_	T=[(1024,512,1,1),(1024,512,1,1)]	modern::elementwise_kernel
1655	fprop	-	add_	T=[(1024,512,1,1),(1024,512,1,1)]	modern::elementwise_kernel
1656	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1657	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1658	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1659	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1660	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1661	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1662	fprop	-	mul_	T=[(256,1024,1,1)]	modern::elementwise_kernel
1663	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1664	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1665	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1666	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1667	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1668	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1669	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1670	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1671	fprop	-	mul_	T=[(256,256,3,3)]	modern::elementwise_kernel
1672	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1673	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1674	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1675	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1676	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1677	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1678	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1679	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1680	fprop	-	mul_	T=[(1024,256,1,1)]	modern::elementwise_kernel
1681	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1682	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1683	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1684	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1685	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1686	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1687	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1688	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1689	fprop	-	mul_	T=[(256,1024,1,1)]	modern::elementwise_kernel
1690	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1691	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1692	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1693	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1694	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1695	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1696	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1697	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1698	fprop	-	mul_	T=[(256,256,3,3)]	modern::elementwise_kernel
1699	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1700	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1701	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1702	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1703	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1704	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1705	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1706	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1707	fprop	-	mul_	T=[(1024,256,1,1)]	modern::elementwise_kernel
1708	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1709	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1710	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1711	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1712	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1713	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1714	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1715	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1716	fprop	-	mul_	T=[(256,1024,1,1)]	modern::elementwise_kernel
1717	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1718	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1719	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1720	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1721	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1722	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1723	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1724	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1725	fprop	-	mul_	T=[(256,256,3,3)]	modern::elementwise_kernel
1726	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1727	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1728	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1729	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1730	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1731	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1732	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1733	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1734	fprop	-	mul_	T=[(1024,256,1,1)]	modern::elementwise_kernel
1735	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1736	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1737	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1738	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1739	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1740	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1741	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1742	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1743	fprop	-	mul_	T=[(256,1024,1,1)]	modern::elementwise_kernel
1744	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1745	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1746	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1747	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1748	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1749	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1750	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1751	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1752	fprop	-	mul_	T=[(256,256,3,3)]	modern::elementwise_kernel
1753	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1754	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1755	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1756	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1757	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1758	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1759	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1760	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1761	fprop	-	mul_	T=[(1024,256,1,1)]	modern::elementwise_kernel
1762	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1763	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1764	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1765	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1766	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1767	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1768	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1769	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1770	fprop	-	mul_	T=[(256,1024,1,1)]	modern::elementwise_kernel
1771	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1772	fprop	-	add_	T=[(256,1024,1,1),(256,1024,1,1)]	modern::elementwise_kernel
1773	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1774	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1775	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1776	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1777	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1778	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1779	fprop	-	mul_	T=[(256,256,3,3)]	modern::elementwise_kernel
1780	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1781	fprop	-	add_	T=[(256,256,3,3),(256,256,3,3)]	modern::elementwise_kernel
1782	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1783	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1784	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1785	fprop	-	mul_	T=[(256,)]	modern::elementwise_kernel
1786	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1787	fprop	-	add_	T=[(256,),(256,)]	modern::elementwise_kernel
1788	fprop	-	mul_	T=[(1024,256,1,1)]	modern::elementwise_kernel
1789	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1790	fprop	-	add_	T=[(1024,256,1,1),(1024,256,1,1)]	modern::elementwise_kernel
1791	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1792	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1793	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1794	fprop	-	mul_	T=[(1024,)]	modern::elementwise_kernel
1795	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1796	fprop	-	add_	T=[(1024,),(1024,)]	modern::elementwise_kernel
1797	fprop	-	mul_	T=[(512,1024,1,1)]	modern::elementwise_kernel
1798	fprop	-	add_	T=[(512,1024,1,1),(512,1024,1,1)]	modern::elementwise_kernel
1799	fprop	-	add_	T=[(512,1024,1,1),(512,1024,1,1)]	modern::elementwise_kernel
1800	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1801	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1802	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1803	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1804	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1805	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1806	fprop	-	mul_	T=[(512,512,3,3)]	modern::elementwise_kernel
1807	fprop	-	add_	T=[(512,512,3,3),(512,512,3,3)]	modern::elementwise_kernel
1808	fprop	-	add_	T=[(512,512,3,3),(512,512,3,3)]	modern::elementwise_kernel
1809	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1810	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1811	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1812	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1813	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1814	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1815	fprop	-	mul_	T=[(2048,512,1,1)]	modern::elementwise_kernel
1816	fprop	-	add_	T=[(2048,512,1,1),(2048,512,1,1)]	modern::elementwise_kernel
1817	fprop	-	add_	T=[(2048,512,1,1),(2048,512,1,1)]	modern::elementwise_kernel
1818	fprop	-	mul_	T=[(2048,)]	modern::elementwise_kernel
1819	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1820	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1821	fprop	-	mul_	T=[(2048,)]	modern::elementwise_kernel
1822	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1823	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1824	fprop	-	mul_	T=[(2048,1024,1,1)]	modern::elementwise_kernel
1825	fprop	-	add_	T=[(2048,1024,1,1),(2048,1024,1,1)]	modern::elementwise_kernel
1826	fprop	-	add_	T=[(2048,1024,1,1),(2048,1024,1,1)]	modern::elementwise_kernel
1827	fprop	-	mul_	T=[(2048,)]	modern::elementwise_kernel
1828	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1829	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1830	fprop	-	mul_	T=[(2048,)]	modern::elementwise_kernel
1831	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1832	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1833	fprop	-	mul_	T=[(512,2048,1,1)]	modern::elementwise_kernel
1834	fprop	-	add_	T=[(512,2048,1,1),(512,2048,1,1)]	modern::elementwise_kernel
1835	fprop	-	add_	T=[(512,2048,1,1),(512,2048,1,1)]	modern::elementwise_kernel
1836	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1837	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1838	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1839	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1840	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1841	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1842	fprop	-	mul_	T=[(512,512,3,3)]	modern::elementwise_kernel
1843	fprop	-	add_	T=[(512,512,3,3),(512,512,3,3)]	modern::elementwise_kernel
1844	fprop	-	add_	T=[(512,512,3,3),(512,512,3,3)]	modern::elementwise_kernel
1845	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1846	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1847	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1848	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1849	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1850	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1851	fprop	-	mul_	T=[(2048,512,1,1)]	modern::elementwise_kernel
1852	fprop	-	add_	T=[(2048,512,1,1),(2048,512,1,1)]	modern::elementwise_kernel
1853	fprop	-	add_	T=[(2048,512,1,1),(2048,512,1,1)]	modern::elementwise_kernel
1854	fprop	-	mul_	T=[(2048,)]	modern::elementwise_kernel
1855	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1856	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1857	fprop	-	mul_	T=[(2048,)]	modern::elementwise_kernel
1858	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1859	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1860	fprop	-	mul_	T=[(512,2048,1,1)]	modern::elementwise_kernel
1861	fprop	-	add_	T=[(512,2048,1,1),(512,2048,1,1)]	modern::elementwise_kernel
1862	fprop	-	add_	T=[(512,2048,1,1),(512,2048,1,1)]	modern::elementwise_kernel
1863	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1864	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1865	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1866	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1867	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1868	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1869	fprop	-	mul_	T=[(512,512,3,3)]	modern::elementwise_kernel
1870	fprop	-	add_	T=[(512,512,3,3),(512,512,3,3)]	modern::elementwise_kernel
1871	fprop	-	add_	T=[(512,512,3,3),(512,512,3,3)]	modern::elementwise_kernel
1872	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1873	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1874	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1875	fprop	-	mul_	T=[(512,)]	modern::elementwise_kernel
1876	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1877	fprop	-	add_	T=[(512,),(512,)]	modern::elementwise_kernel
1878	fprop	-	mul_	T=[(2048,512,1,1)]	modern::elementwise_kernel
1879	fprop	-	add_	T=[(2048,512,1,1),(2048,512,1,1)]	modern::elementwise_kernel
1880	fprop	-	add_	T=[(2048,512,1,1),(2048,512,1,1)]	modern::elementwise_kernel
1881	fprop	-	mul_	T=[(2048,)]	modern::elementwise_kernel
1882	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1883	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1884	fprop	-	mul_	T=[(2048,)]	modern::elementwise_kernel
1885	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1886	fprop	-	add_	T=[(2048,),(2048,)]	modern::elementwise_kernel
1887	fprop	-	mul_	T=[(1000,2048)]	modern::elementwise_kernel
1888	fprop	-	add_	T=[(1000,2048),(1000,2048)]	modern::elementwise_kernel
1889	fprop	-	add_	T=[(1000,2048),(1000,2048)]	modern::elementwise_kernel
1890	fprop	-	mul_	T=[(1000,)]	modern::elementwise_kernel
1891	fprop	-	add_	T=[(1000,),(1000,)]	modern::elementwise_kernel
1892	fprop	-	add_	T=[(1000,),(1000,)]	modern::elementwise_kernel