CoCalc -- default_kernel

GitHub Repository: Tetragramm/opencv
Path: blob/master/modules/dnn/src/ocl4dnn/include/default_kernel_config.hpp
¹⁶³⁴⁸ views
1
#ifndef _OPENCV_OCL4DNN_DEFAULT_KERNEL_CONFIG_HPP_
2
#define _OPENCV_OCL4DNN_DEFAULT_KERNEL_CONFIG_HPP_
3
const char *default_kernel_config_intel_fp32[] = {
4
// Below is the information for OpenCL based on which these configurations tuned
5
/*******************************************************************************
6
Number of platforms                               1
7
  Platform Name                                   Intel(R) OpenCL HD Graphics
8
  Platform Vendor                                 Intel(R) Corporation
9
  Platform Version                                OpenCL 2.1
10
  Platform Profile                                FULL_PROFILE
11
  Platform Extensions                             cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_khr_fp64 cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation
12
  Platform Host timer resolution                  1ns
13
  Platform Extensions function suffix             INTEL
14

15
  Platform Name                                   Intel(R) OpenCL HD Graphics
16
Number of devices                                 1
17
  Device Name                                     Intel(R) Gen9 HD Graphics NEO
18
  Device Vendor                                   Intel(R) Corporation
19
  Device Vendor ID                                0x8086
20
  Device Version                                  OpenCL 2.1 NEO
21
  Driver Version                                  2018ww15-010713
22
  Device OpenCL C Version                         OpenCL C 2.0
23
  Device Type                                     GPU
24
  Device Profile                                  FULL_PROFILE
25
  Max compute units                               72
26
  Max clock frequency                             950MHz
27
  Device Partition                                (core)
28
    Max number of sub-devices                     0
29
    Supported partition types                     None
30
  Max work item dimensions                        3
31
  Max work item sizes                             256x256x256
32
  Max work group size                             256
33
  Preferred work group size multiple              32
34
  Max sub-groups per work group                   32
35
  Preferred / native vector sizes
36
    char                                                16 / 16
37
    short                                                8 / 8
38
    int                                                  4 / 4
39
    long                                                 1 / 1
40
    half                                                 8 / 8        (cl_khr_fp16)
41
    float                                                1 / 1
42
    double                                               1 / 1        (cl_khr_fp64)
43
  Half-precision Floating-point support           (cl_khr_fp16)
44
    Denormals                                     Yes
45
    Infinity and NANs                             Yes
46
    Round to nearest                              Yes
47
    Round to zero                                 Yes
48
    Round to infinity                             Yes
49
    IEEE754-2008 fused multiply-add               Yes
50
    Support is emulated in software               No
51
    Correctly-rounded divide and sqrt operations  No
52
  Single-precision Floating-point support         (core)
53
    Denormals                                     Yes
54
    Infinity and NANs                             Yes
55
    Round to nearest                              Yes
56
    Round to zero                                 Yes
57
    Round to infinity                             Yes
58
    IEEE754-2008 fused multiply-add               Yes
59
    Support is emulated in software               No
60
    Correctly-rounded divide and sqrt operations  Yes
61
  Double-precision Floating-point support         (cl_khr_fp64)
62
    Denormals                                     Yes
63
    Infinity and NANs                             Yes
64
    Round to nearest                              Yes
65
    Round to zero                                 Yes
66
    Round to infinity                             Yes
67
    IEEE754-2008 fused multiply-add               Yes
68
    Support is emulated in software               No
69
    Correctly-rounded divide and sqrt operations  No
70
  Address bits                                    64, Little-Endian
71
  Global memory size                              26892222464 (25.05GiB)
72
  Error Correction support                        No
73
  Max memory allocation                           4294959104 (4GiB)
74
  Unified memory for Host and Device              Yes
75
  Shared Virtual Memory (SVM) capabilities        (core)
76
    Coarse-grained buffer sharing                 Yes
77
    Fine-grained buffer sharing                   No
78
    Fine-grained system sharing                   No
79
    Atomics                                       No
80
  Minimum alignment for any data type             128 bytes
81
  Alignment of base address                       1024 bits (128 bytes)
82
  Preferred alignment for atomics
83
    SVM                                           64 bytes
84
    Global                                        64 bytes
85
    Local                                         64 bytes
86
  Max size for global variable                    65536 (64KiB)
87
  Preferred total size of global vars             4294959104 (4GiB)
88
  Global Memory cache type                        Read/Write
89
  Global Memory cache size                        1572864
90
  Global Memory cache line                        64 bytes
91
  Image support                                   Yes
92
    Max number of samplers per kernel             16
93
    Max size for 1D images from buffer            268434944 pixels
94
    Max 1D or 2D image array size                 2048 images
95
    Base address alignment for 2D image buffers   4 bytes
96
    Pitch alignment for 2D image buffers          4 bytes
97
    Max 2D image size                             16384x16384 pixels
98
    Max 3D image size                             16384x16384x2048 pixels
99
    Max number of read image args                 128
100
    Max number of write image args                128
101
    Max number of read/write image args           128
102
  Max number of pipe args                         16
103
  Max active pipe reservations                    1
104
  Max pipe packet size                            1024
105
  Local memory type                               Local
106
  Local memory size                               65536 (64KiB)
107
  Max constant buffer size                        4294959104 (4GiB)
108
  Max number of constant args                     8
109
  Max size of kernel argument                     1024
110
  Queue properties (on host)
111
    Out-of-order execution                        Yes
112
    Profiling                                     Yes
113
  Queue properties (on device)
114
    Out-of-order execution                        Yes
115
    Profiling                                     Yes
116
    Preferred size                                131072 (128KiB)
117
    Max size                                      67108864 (64MiB)
118
  Max queues on device                            1
119
  Max events on device                            1024
120
  Prefer user sync for interop                    Yes
121
  Profiling timer resolution                      83ns
122
  Execution capabilities
123
    Run OpenCL kernels                            Yes
124
    Run native kernels                            No
125
    Sub-group independent forward progress        Yes
126
    IL version                                    SPIR-V_1.0
127
    SPIR versions                                 1.2
128
  printf() buffer size                            4194304 (4MiB)
129
  Built-in kernels                                block_motion_estimate_intel;block_advanced_motion_estimate_check_intel;block_advanced_motion_estimate_bidirectional_check_intel;
130
  Motion Estimation accelerator version	(Intel)   2
131
  Device Available                                Yes
132
  Compiler Available                              Yes
133
  Linker Available                                Yes
134
  Device Extensions                               cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_khr_fp64 cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation
135

136
NULL platform behavior
137
  clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...)  Intel(R) OpenCL HD Graphics
138
  clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...)   Success [INTEL]
139
  clCreateContext(NULL, ...) [default]            Success [INTEL]
140
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU)  No devices found in platform
141
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  Success (1)
142
    Platform Name                                 Intel(R) OpenCL HD Graphics
143
    Device Name                                   Intel(R) Gen9 HD Graphics NEO
144
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR)  No devices found in platform
145
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM)  No devices found in platform
146
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL)  Success (1)
147
    Platform Name                                 Intel(R) OpenCL HD Graphics
148
    Device Name                                   Intel(R) Gen9 HD Graphics NEO
149

150
ICD loader properties
151
  ICD loader Name                                 OpenCL ICD Loader
152
  ICD loader Vendor                               OCL Icd free software
153
  ICD loader Version                              2.2.8
154
  ICD loader Profile                              OpenCL 1.2
155
      NOTE:	your OpenCL library declares to support OpenCL 1.2,
156
                but it seems to support up to OpenCL 2.1 too.
157
********************************************************************************/
158

159
"EU72_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
160
"EU72_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ5_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
161
"EU72_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M126_activ0_eltwise0_FP32", "6 1 16 2 1 1 16 1 0",
162
"EU72_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ0_eltwise0_FP32", "2 1 8 2 1 1 8 1 0",
163
"EU72_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
164
"EU72_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ5_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
165
"EU72_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M546_activ0_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
166
"EU72_k1x1_cn128_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M126_activ0_eltwise0_FP32", "1 1 16 2 1 1 16 1 0",
167
"EU72_k1x1_cn128_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ0_eltwise0_FP32", "1 1 16 2 1 1 16 1 0",
168
"EU72_k1x1_cn128_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M546_activ0_eltwise0_FP32", "1 1 16 2 1 1 16 1 0",
169
"EU72_k1x1_cn128_g1_s1x1_d1x1_b1_in48x48_p0x0_num1_M256_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
170
"EU72_k1x1_cn128_g1_s1x1_d1x1_b1_in48x48_p0x0_num1_M256_activ5_eltwise0_FP32", "12 2 16 2 1 1 16 1 0",
171
"EU72_k1x1_cn128_g1_s1x1_d1x1_b1_in80x80_p0x0_num1_M128_activ1_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
172
"EU72_k1x1_cn128_g1_s1x1_d1x1_b1_in80x80_p0x0_num1_M128_activ5_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
173
"EU72_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M16_activ1_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
174
"EU72_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP32", "8 1 8 2 1 1 8 1 0",
175
"EU72_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
176
"EU72_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M96_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
177
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M126_activ0_eltwise0_FP32", "1 3 8 2 1 1 8 1 0",
178
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "2 1 8 2 1 1 8 1 0",
179
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ5_eltwise0_FP32", "2 1 16 2 1 1 16 1 0",
180
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ0_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
181
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M546_activ0_eltwise0_FP32", "3 1 16 2 1 1 16 1 0",
182
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
183
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ5_eltwise0_FP32", "2 1 8 2 1 1 8 1 0",
184
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M128_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
185
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP32", "8 1 8 2 1 1 8 1 0",
186
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
187
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ5_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
188
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP32", "8 1 8 2 1 1 8 1 0",
189
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in48x48_p0x0_num1_M256_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
190
"EU72_k1x1_cn256_g1_s1x1_d1x1_b1_in48x48_p0x0_num1_M256_activ5_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
191
"EU72_k1x1_cn32_g1_s1x1_d1x1_b1_in160x160_p0x0_num1_M64_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
192
"EU72_k1x1_cn32_g1_s1x1_d1x1_b1_in160x160_p0x0_num1_M64_activ5_eltwise0_FP32", "1 16 32 5 1 16 1 1 0",
193
"EU72_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M16_activ1_eltwise0_FP32", "5 1 8 2 1 1 8 1 0",
194
"EU72_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
195
"EU72_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP32", "7 1 8 2 1 1 8 1 0",
196
"EU72_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M96_activ1_eltwise0_FP32", "7 2 8 2 1 1 8 1 0",
197
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ1_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
198
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ5_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
199
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M112_activ1_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
200
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M126_activ0_eltwise0_FP32", "5 1 8 2 1 1 8 1 0",
201
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "8 2 8 2 1 1 8 1 0",
202
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ5_eltwise0_FP32", "5 1 8 2 1 1 8 1 0",
203
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M144_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
204
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "8 2 8 2 1 1 8 1 0",
205
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ0_eltwise0_FP32", "3 1 8 2 1 1 8 1 0",
206
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ1_eltwise0_FP32", "4 1 16 2 1 1 16 1 0",
207
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "7 1 8 2 1 1 8 1 0",
208
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M546_activ0_eltwise0_FP32", "3 1 16 2 1 1 16 1 0",
209
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP32", "8 1 8 2 1 1 8 1 0",
210
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M12_activ0_eltwise0_FP32", "6 1 16 2 1 1 16 1 0",
211
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M273_activ0_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
212
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ1_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
213
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ5_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
214
"EU72_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M63_activ0_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
215
"EU72_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
216
"EU72_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "8 2 8 2 1 1 8 1 0",
217
"EU72_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
218
"EU72_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "8 1 8 2 1 1 8 1 0",
219
"EU72_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M64_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
220
"EU72_k1x1_cn64_g1_s1x1_d1x1_b1_in80x80_p0x0_num1_M128_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
221
"EU72_k1x1_cn64_g1_s1x1_d1x1_b1_in80x80_p0x0_num1_M128_activ5_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
222
"EU72_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "7 1 8 2 1 1 8 1 0",
223
"EU72_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "7 1 8 2 1 1 8 1 0",
224
"EU72_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP32", "7 1 8 2 1 1 8 1 0",
225
"EU72_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "7 1 8 2 1 1 8 1 0",
226
"EU72_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "5 1 8 2 1 1 8 1 0",
227
"EU72_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M384_activ1_eltwise0_FP32", "7 2 8 2 1 1 8 1 0",
228
"EU72_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M48_activ1_eltwise0_FP32", "5 1 8 2 1 1 8 1 0",
229
"EU72_k3x3_cn1024_g1024_s1x1_d1x1_b1_in16x16_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
230
"EU72_k3x3_cn1024_g1024_s1x1_d1x1_b1_in16x16_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
231
"EU72_k3x3_cn112_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M224_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
232
"EU72_k3x3_cn128_g128_s1x1_d1x1_b1_in80x80_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
233
"EU72_k3x3_cn128_g128_s1x1_d1x1_b1_in80x80_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
234
"EU72_k3x3_cn128_g128_s2x2_d1x1_b1_in80x80_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
235
"EU72_k3x3_cn128_g128_s2x2_d1x1_b1_in80x80_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
236
"EU72_k3x3_cn128_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
237
"EU72_k3x3_cn128_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M192_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
238
"EU72_k3x3_cn128_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M256_activ5_eltwise0_FP32", "3 1 8 2 1 1 8 1 0",
239
"EU72_k3x3_cn128_g1_s2x2_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP32", "1 1 16 2 1 1 16 1 0",
240
"EU72_k3x3_cn144_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M288_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
241
"EU72_k3x3_cn160_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M320_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
242
"EU72_k3x3_cn192_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
243
"EU72_k3x3_cn256_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M512_activ5_eltwise0_FP32", "3 1 16 2 1 1 16 1 0",
244
"EU72_k3x3_cn256_g1_s2x2_d1x1_b1_in16x16_p1x1_num1_M512_activ1_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
245
"EU72_k3x3_cn256_g256_s1x1_d1x1_b1_in48x48_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
246
"EU72_k3x3_cn256_g256_s1x1_d1x1_b1_in48x48_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
247
"EU72_k3x3_cn256_g256_s2x2_d1x1_b1_in48x48_p0x0_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
248
"EU72_k3x3_cn256_g256_s2x2_d1x1_b1_in48x48_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
249
"EU72_k3x3_cn32_g32_s1x1_d1x1_b1_in160x160_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
250
"EU72_k3x3_cn32_g32_s1x1_d1x1_b1_in160x160_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
251
"EU72_k3x3_cn3_g1_s2x2_d1x1_b1_in256x256_p0x0_num1_M32_activ5_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
252
"EU72_k3x3_cn3_g1_s2x2_d1x1_b1_in256x256_p1x1_num1_M32_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
253
"EU72_k3x3_cn512_g512_s1x1_d1x1_b1_in32x32_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
254
"EU72_k3x3_cn512_g512_s1x1_d1x1_b1_in32x32_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
255
"EU72_k3x3_cn512_g512_s2x2_d1x1_b1_in32x32_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
256
"EU72_k3x3_cn512_g512_s2x2_d1x1_b1_in32x32_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
257
"EU72_k3x3_cn64_g1_s1x1_d1x1_b1_in64x64_p1x1_num1_M192_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
258
"EU72_k3x3_cn64_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M128_activ5_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
259
"EU72_k3x3_cn64_g1_s2x2_d1x1_b1_in16x16_p1x1_num1_M128_activ1_eltwise0_FP32", "1 1 16 2 1 1 16 1 0",
260
"EU72_k3x3_cn64_g64_s2x2_d1x1_b1_in160x160_p0x0_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
261
"EU72_k3x3_cn64_g64_s2x2_d1x1_b1_in160x160_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
262
"EU72_k3x3_cn96_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M208_activ1_eltwise0_FP32", "7 2 16 2 1 1 16 1 0",
263
"EU72_k3x3_cn96_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M128_activ1_eltwise0_FP32", "7 2 16 2 1 1 16 1 0",
264
"EU72_k5x5_cn16_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M48_activ1_eltwise0_FP32", "4 2 8 2 1 1 8 1 0",
265
"EU72_k5x5_cn16_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M32_activ1_eltwise0_FP32", "6 1 16 2 1 1 16 1 0",
266
"EU72_k5x5_cn24_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
267
"EU72_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
268
"EU72_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP32", "6 1 16 2 1 1 16 1 0",
269
"EU72_k5x5_cn32_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M96_activ1_eltwise0_FP32", "7 2 16 2 1 1 16 1 0",
270
"EU72_k5x5_cn48_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP32", "4 1 16 2 1 1 16 1 0",
271
"EU72_k7x7_cn3_g1_s2x2_d1x1_b1_in224x224_p3x3_num1_M64_activ1_eltwise0_FP32", "4 3 16 2 1 1 16 1 0",
272
// Below is the information for OpenCL based on which these configurations tuned
273
/*******************************************************************************
274
Number of platforms                               1
275
  Platform Name                                   Intel(R) OpenCL HD Graphics
276
  Platform Vendor                                 Intel(R) Corporation
277
  Platform Version                                OpenCL 2.1
278
  Platform Profile                                FULL_PROFILE
279
  Platform Extensions                             cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_khr_fp64 cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation
280
  Platform Host timer resolution                  1ns
281
  Platform Extensions function suffix             INTEL
282

283
  Platform Name                                   Intel(R) OpenCL HD Graphics
284
Number of devices                                 1
285
  Device Name                                     Intel(R) Gen9 HD Graphics NEO
286
  Device Vendor                                   Intel(R) Corporation
287
  Device Vendor ID                                0x8086
288
  Device Version                                  OpenCL 2.1 NEO
289
  Driver Version                                  18.21.10858
290
  Device OpenCL C Version                         OpenCL C 2.0
291
  Device Type                                     GPU
292
  Device Profile                                  FULL_PROFILE
293
  Max compute units                               48
294
  Max clock frequency                             950MHz
295
  Device Partition                                (core)
296
    Max number of sub-devices                     0
297
    Supported partition types                     None
298
  Max work item dimensions                        3
299
  Max work item sizes                             256x256x256
300
  Max work group size                             256
301
  Preferred work group size multiple              32
302
  Max sub-groups per work group                   32
303
  Preferred / native vector sizes
304
    char                                                16 / 16
305
    short                                                8 / 8
306
    int                                                  4 / 4
307
    long                                                 1 / 1
308
    half                                                 8 / 8        (cl_khr_fp16)
309
    float                                                1 / 1
310
    double                                               1 / 1        (cl_khr_fp64)
311
  Half-precision Floating-point support           (cl_khr_fp16)
312
    Denormals                                     Yes
313
    Infinity and NANs                             Yes
314
    Round to nearest                              Yes
315
    Round to zero                                 Yes
316
    Round to infinity                             Yes
317
    IEEE754-2008 fused multiply-add               Yes
318
    Support is emulated in software               No
319
    Correctly-rounded divide and sqrt operations  No
320
  Single-precision Floating-point support         (core)
321
    Denormals                                     Yes
322
    Infinity and NANs                             Yes
323
    Round to nearest                              Yes
324
    Round to zero                                 Yes
325
    Round to infinity                             Yes
326
    IEEE754-2008 fused multiply-add               Yes
327
    Support is emulated in software               No
328
    Correctly-rounded divide and sqrt operations  Yes
329
  Double-precision Floating-point support         (cl_khr_fp64)
330
    Denormals                                     Yes
331
    Infinity and NANs                             Yes
332
    Round to nearest                              Yes
333
    Round to zero                                 Yes
334
    Round to infinity                             Yes
335
    IEEE754-2008 fused multiply-add               Yes
336
    Support is emulated in software               No
337
    Correctly-rounded divide and sqrt operations  No
338
  Address bits                                    64, Little-Endian
339
  Global memory size                              13364170752 (12.45GiB)
340
  Error Correction support                        No
341
  Max memory allocation                           4294959104 (4GiB)
342
  Unified memory for Host and Device              Yes
343
  Shared Virtual Memory (SVM) capabilities        (core)
344
    Coarse-grained buffer sharing                 Yes
345
    Fine-grained buffer sharing                   No
346
    Fine-grained system sharing                   No
347
    Atomics                                       No
348
  Minimum alignment for any data type             128 bytes
349
  Alignment of base address                       1024 bits (128 bytes)
350
  Preferred alignment for atomics
351
    SVM                                           64 bytes
352
    Global                                        64 bytes
353
    Local                                         64 bytes
354
  Max size for global variable                    65536 (64KiB)
355
  Preferred total size of global vars             4294959104 (4GiB)
356
  Global Memory cache type                        Read/Write
357
  Global Memory cache size                        1048576
358
  Global Memory cache line                        64 bytes
359
  Image support                                   Yes
360
    Max number of samplers per kernel             16
361
    Max size for 1D images from buffer            268434944 pixels
362
    Max 1D or 2D image array size                 2048 images
363
    Base address alignment for 2D image buffers   4 bytes
364
    Pitch alignment for 2D image buffers          4 bytes
365
    Max 2D image size                             16384x16384 pixels
366
    Max 3D image size                             16384x16384x2048 pixels
367
    Max number of read image args                 128
368
    Max number of write image args                128
369
    Max number of read/write image args           128
370
  Max number of pipe args                         16
371
  Max active pipe reservations                    1
372
  Max pipe packet size                            1024
373
  Local memory type                               Local
374
  Local memory size                               65536 (64KiB)
375
  Max constant buffer size                        4294959104 (4GiB)
376
  Max number of constant args                     8
377
  Max size of kernel argument                     1024
378
  Queue properties (on host)
379
    Out-of-order execution                        Yes
380
    Profiling                                     Yes
381
  Queue properties (on device)
382
    Out-of-order execution                        Yes
383
    Profiling                                     Yes
384
    Preferred size                                131072 (128KiB)
385
    Max size                                      67108864 (64MiB)
386
  Max queues on device                            1
387
  Max events on device                            1024
388
  Prefer user sync for interop                    Yes
389
  Profiling timer resolution                      83ns
390
  Execution capabilities
391
    Run OpenCL kernels                            Yes
392
    Run native kernels                            No
393
    Sub-group independent forward progress        Yes
394
    IL version                                    SPIR-V_1.0
395
    SPIR versions                                 1.2
396
  printf() buffer size                            4194304 (4MiB)
397
  Built-in kernels                                block_motion_estimate_intel;block_advanced_motion_estimate_check_intel;block_advanced_motion_estimate_bidirectional_check_intel;
398
  Motion Estimation accelerator version	(Intel)   2
399
  Device Available                                Yes
400
  Compiler Available                              Yes
401
  Linker Available                                Yes
402
  Device Extensions                               cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_khr_fp64 cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation
403

404
NULL platform behavior
405
  clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...)  No platform
406
  clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...)   No platform
407
  clCreateContext(NULL, ...) [default]            No platform
408
  clCreateContext(NULL, ...) [other]              Success [INTEL]
409
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU)  No platform
410
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  No platform
411
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR)  No platform
412
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM)  No platform
413
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL)  No platform
414
********************************************************************************/
415
"EU48_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ1_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
416
"EU48_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ5_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
417
"EU48_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M126_activ0_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
418
"EU48_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ0_eltwise0_FP32", "5 1 8 2 1 1 8 1 0",
419
"EU48_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
420
"EU48_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ5_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
421
"EU48_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M546_activ0_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
422
"EU48_k1x1_cn128_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M126_activ0_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
423
"EU48_k1x1_cn128_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ0_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
424
"EU48_k1x1_cn128_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M546_activ0_eltwise0_FP32", "1 1 16 2 1 1 16 1 0",
425
"EU48_k1x1_cn128_g1_s1x1_d1x1_b1_in48x48_p0x0_num1_M256_activ1_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
426
"EU48_k1x1_cn128_g1_s1x1_d1x1_b1_in48x48_p0x0_num1_M256_activ5_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
427
"EU48_k1x1_cn128_g1_s1x1_d1x1_b1_in80x80_p0x0_num1_M128_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
428
"EU48_k1x1_cn128_g1_s1x1_d1x1_b1_in80x80_p0x0_num1_M128_activ5_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
429
"EU48_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M16_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
430
"EU48_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
431
"EU48_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
432
"EU48_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M96_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
433
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M126_activ0_eltwise0_FP32", "3 1 8 2 1 1 8 1 0",
434
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "3 1 8 2 1 1 8 1 0",
435
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ5_eltwise0_FP32", "3 1 8 2 1 1 8 1 0",
436
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ0_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
437
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M546_activ0_eltwise0_FP32", "3 1 16 2 1 1 16 1 0",
438
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP32", "2 1 8 2 1 1 8 1 0",
439
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ5_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
440
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M128_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
441
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
442
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ1_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
443
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ5_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
444
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
445
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in48x48_p0x0_num1_M256_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
446
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in48x48_p0x0_num1_M256_activ5_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
447
"EU48_k1x1_cn32_g1_s1x1_d1x1_b1_in160x160_p0x0_num1_M64_activ1_eltwise0_FP32", "1 16 32 5 1 16 1 1 0",
448
"EU48_k1x1_cn32_g1_s1x1_d1x1_b1_in160x160_p0x0_num1_M64_activ5_eltwise0_FP32", "1 16 32 5 1 16 1 1 0",
449
"EU48_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M16_activ1_eltwise0_FP32", "8 1 8 2 1 1 8 1 0",
450
"EU48_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
451
"EU48_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
452
"EU48_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M96_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
453
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ1_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
454
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ5_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
455
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M112_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
456
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M126_activ0_eltwise0_FP32", "5 1 8 2 1 1 8 1 0",
457
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
458
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ5_eltwise0_FP32", "5 1 8 2 1 1 8 1 0",
459
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M144_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
460
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
461
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ0_eltwise0_FP32", "4 1 8 2 1 1 8 1 0",
462
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ1_eltwise0_FP32", "8 1 8 2 1 1 8 1 0",
463
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "8 1 8 2 1 1 8 1 0",
464
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M546_activ0_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
465
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
466
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M12_activ0_eltwise0_FP32", "9 1 16 2 1 1 16 1 0",
467
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M273_activ0_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
468
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
469
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ5_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
470
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M63_activ0_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
471
"EU48_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
472
"EU48_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
473
"EU48_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
474
"EU48_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
475
"EU48_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M64_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
476
"EU48_k1x1_cn64_g1_s1x1_d1x1_b1_in80x80_p0x0_num1_M128_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
477
"EU48_k1x1_cn64_g1_s1x1_d1x1_b1_in80x80_p0x0_num1_M128_activ5_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
478
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "7 2 8 2 1 1 8 1 0",
479
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
480
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
481
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
482
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "4 1 8 2 1 1 8 1 0",
483
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M384_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
484
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M48_activ1_eltwise0_FP32", "4 1 8 2 1 1 8 1 0",
485
"EU48_k3x3_cn1024_g1024_s1x1_d1x1_b1_in16x16_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
486
"EU48_k3x3_cn1024_g1024_s1x1_d1x1_b1_in16x16_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
487
"EU48_k3x3_cn112_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M224_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
488
"EU48_k3x3_cn128_g128_s1x1_d1x1_b1_in80x80_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
489
"EU48_k3x3_cn128_g128_s1x1_d1x1_b1_in80x80_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
490
"EU48_k3x3_cn128_g128_s2x2_d1x1_b1_in80x80_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
491
"EU48_k3x3_cn128_g128_s2x2_d1x1_b1_in80x80_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
492
"EU48_k3x3_cn128_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
493
"EU48_k3x3_cn128_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M192_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
494
"EU48_k3x3_cn128_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M256_activ5_eltwise0_FP32", "2 1 16 2 1 1 16 1 0",
495
"EU48_k3x3_cn128_g1_s2x2_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP32", "3 1 8 2 1 1 8 1 0",
496
"EU48_k3x3_cn144_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M288_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
497
"EU48_k3x3_cn160_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M320_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
498
"EU48_k3x3_cn192_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
499
"EU48_k3x3_cn256_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M512_activ5_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
500
"EU48_k3x3_cn256_g1_s2x2_d1x1_b1_in16x16_p1x1_num1_M512_activ1_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
501
"EU48_k3x3_cn256_g256_s1x1_d1x1_b1_in48x48_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
502
"EU48_k3x3_cn256_g256_s1x1_d1x1_b1_in48x48_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
503
"EU48_k3x3_cn256_g256_s2x2_d1x1_b1_in48x48_p0x0_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
504
"EU48_k3x3_cn256_g256_s2x2_d1x1_b1_in48x48_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
505
"EU48_k3x3_cn32_g32_s1x1_d1x1_b1_in160x160_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
506
"EU48_k3x3_cn32_g32_s1x1_d1x1_b1_in160x160_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
507
"EU48_k3x3_cn3_g1_s2x2_d1x1_b1_in256x256_p0x0_num1_M32_activ5_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
508
"EU48_k3x3_cn3_g1_s2x2_d1x1_b1_in256x256_p1x1_num1_M32_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
509
"EU48_k3x3_cn512_g512_s1x1_d1x1_b1_in32x32_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
510
"EU48_k3x3_cn512_g512_s1x1_d1x1_b1_in32x32_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
511
"EU48_k3x3_cn512_g512_s2x2_d1x1_b1_in32x32_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
512
"EU48_k3x3_cn512_g512_s2x2_d1x1_b1_in32x32_p1x1_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
513
"EU48_k3x3_cn64_g1_s1x1_d1x1_b1_in64x64_p1x1_num1_M192_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
514
"EU48_k3x3_cn64_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M128_activ5_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
515
"EU48_k3x3_cn64_g1_s2x2_d1x1_b1_in16x16_p1x1_num1_M128_activ1_eltwise0_FP32", "1 1 8 2 1 1 8 1 0",
516
"EU48_k3x3_cn64_g64_s2x2_d1x1_b1_in160x160_p0x0_num1_M1_activ5_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
517
"EU48_k3x3_cn64_g64_s2x2_d1x1_b1_in160x160_p1x1_num1_M1_activ1_eltwise0_FP32", "1 1 1 6 1 1 1 0 0",
518
"EU48_k3x3_cn96_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M208_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
519
"EU48_k3x3_cn96_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M128_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
520
"EU48_k5x5_cn16_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M48_activ1_eltwise0_FP32", "4 2 8 2 1 1 8 1 0",
521
"EU48_k5x5_cn16_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M32_activ1_eltwise0_FP32", "4 3 8 2 1 1 8 1 0",
522
"EU48_k5x5_cn24_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP32", "4 1 16 2 1 1 16 1 0",
523
"EU48_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
524
"EU48_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP32", "4 2 8 2 1 1 8 1 0",
525
"EU48_k5x5_cn32_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M96_activ1_eltwise0_FP32", "4 7 8 2 1 1 8 1 0",
526
"EU48_k5x5_cn48_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP32", "4 1 16 2 1 1 16 1 0",
527
"EU48_k7x7_cn3_g1_s2x2_d1x1_b1_in224x224_p3x3_num1_M64_activ1_eltwise0_FP32", "5 4 16 2 1 1 16 1 0",
528
"EU48_k11x11_cn3_g1_s4x4_d1x1_b1_in240x240_p0x0_num1_M96_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
529
"EU48_k3x3_cn256_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP32", "13 1 16 2 1 1 16 1 0",
530
"EU48_k3x3_cn384_g2_s1x1_d1x1_b1_in16x16_p1x1_num1_M128_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
531
"EU48_k3x3_cn384_g2_s1x1_d1x1_b1_in16x16_p1x1_num1_M192_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
532
"EU48_k5x5_cn96_g2_s1x1_d1x1_b1_in32x32_p2x2_num1_M128_activ1_eltwise0_FP32", "3 9 8 2 1 1 8 1 0",
533
// Below is the information for OpenCL based on which these configurations tuned
534
/*******************************************************************************
535
Number of platforms                               1
536
  Platform Name                                   Intel(R) OpenCL HD Graphics
537
  Platform Vendor                                 Intel(R) Corporation
538
  Platform Version                                OpenCL 2.1
539
  Platform Profile                                FULL_PROFILE
540
  Platform Extensions                             cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_fp64 cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation cl_intel_va_api_media_sharing
541
  Platform Host timer resolution                  1ns
542
  Platform Extensions function suffix             INTEL
543

544
  Platform Name                                   Intel(R) OpenCL HD Graphics
545
Number of devices                                 1
546
  Device Name                                     Intel(R) Gen9 HD Graphics NEO
547
  Device Vendor                                   Intel(R) Corporation
548
  Device Vendor ID                                0x8086
549
  Device Version                                  OpenCL 2.1 NEO
550
  Driver Version                                  18.23.10915
551
  Device OpenCL C Version                         OpenCL C 2.0
552
  Device Type                                     GPU
553
  Device Profile                                  FULL_PROFILE
554
  Max compute units                               24
555
  Max clock frequency                             1150MHz
556
  Device Partition                                (core)
557
    Max number of sub-devices                     0
558
    Supported partition types                     None
559
  Max work item dimensions                        3
560
  Max work item sizes                             256x256x256
561
  Max work group size                             256
562
  Preferred work group size multiple              32
563
  Max sub-groups per work group                   32
564
  Preferred / native vector sizes
565
    char                                                16 / 16
566
    short                                                8 / 8
567
    int                                                  4 / 4
568
    long                                                 1 / 1
569
    half                                                 8 / 8        (cl_khr_fp16)
570
    float                                                1 / 1
571
    double                                               1 / 1        (cl_khr_fp64)
572
  Half-precision Floating-point support           (cl_khr_fp16)
573
    Denormals                                     Yes
574
    Infinity and NANs                             Yes
575
    Round to nearest                              Yes
576
    Round to zero                                 Yes
577
    Round to infinity                             Yes
578
    IEEE754-2008 fused multiply-add               Yes
579
    Support is emulated in software               No
580
    Correctly-rounded divide and sqrt operations  No
581
  Single-precision Floating-point support         (core)
582
    Denormals                                     Yes
583
    Infinity and NANs                             Yes
584
    Round to nearest                              Yes
585
    Round to zero                                 Yes
586
    Round to infinity                             Yes
587
    IEEE754-2008 fused multiply-add               Yes
588
    Support is emulated in software               No
589
    Correctly-rounded divide and sqrt operations  Yes
590
  Double-precision Floating-point support         (cl_khr_fp64)
591
    Denormals                                     Yes
592
    Infinity and NANs                             Yes
593
    Round to nearest                              Yes
594
    Round to zero                                 Yes
595
    Round to infinity                             Yes
596
    IEEE754-2008 fused multiply-add               Yes
597
    Support is emulated in software               No
598
    Correctly-rounded divide and sqrt operations  No
599
  Address bits                                    64, Little-Endian
600
  Global memory size                              6575288320 (6.124GiB)
601
  Error Correction support                        No
602
  Max memory allocation                           3287644160 (3.062GiB)
603
  Unified memory for Host and Device              Yes
604
  Shared Virtual Memory (SVM) capabilities        (core)
605
    Coarse-grained buffer sharing                 Yes
606
    Fine-grained buffer sharing                   No
607
    Fine-grained system sharing                   No
608
    Atomics                                       No
609
  Minimum alignment for any data type             128 bytes
610
  Alignment of base address                       1024 bits (128 bytes)
611
  Preferred alignment for atomics
612
    SVM                                           64 bytes
613
    Global                                        64 bytes
614
    Local                                         64 bytes
615
  Max size for global variable                    65536 (64KiB)
616
  Preferred total size of global vars             3287644160 (3.062GiB)
617
  Global Memory cache type                        Read/Write
618
  Global Memory cache size                        524288
619
  Global Memory cache line                        64 bytes
620
  Image support                                   Yes
621
    Max number of samplers per kernel             16
622
    Max size for 1D images from buffer            205477760 pixels
623
    Max 1D or 2D image array size                 2048 images
624
    Base address alignment for 2D image buffers   4 bytes
625
    Pitch alignment for 2D image buffers          4 bytes
626
    Max 2D image size                             16384x16384 pixels
627
    Max 3D image size                             16384x16384x2048 pixels
628
    Max number of read image args                 128
629
    Max number of write image args                128
630
    Max number of read/write image args           128
631
  Max number of pipe args                         16
632
  Max active pipe reservations                    1
633
  Max pipe packet size                            1024
634
  Local memory type                               Local
635
  Local memory size                               65536 (64KiB)
636
  Max constant buffer size                        3287644160 (3.062GiB)
637
  Max number of constant args                     8
638
  Max size of kernel argument                     1024
639
  Queue properties (on host)
640
    Out-of-order execution                        Yes
641
    Profiling                                     Yes
642
  Queue properties (on device)
643
    Out-of-order execution                        Yes
644
    Profiling                                     Yes
645
    Preferred size                                131072 (128KiB)
646
    Max size                                      67108864 (64MiB)
647
  Max queues on device                            1
648
  Max events on device                            1024
649
  Prefer user sync for interop                    Yes
650
  Profiling timer resolution                      83ns
651
  Execution capabilities
652
    Run OpenCL kernels                            Yes
653
    Run native kernels                            No
654
    Sub-group independent forward progress        Yes
655
    IL version                                    SPIR-V_1.0
656
    SPIR versions                                 1.2
657
  printf() buffer size                            4194304 (4MiB)
658
  Built-in kernels                                block_motion_estimate_intel;block_advanced_motion_estimate_check_intel;block_advanced_motion_estimate_bidirectional_check_intel;
659
  Motion Estimation accelerator version	(Intel)   2
660
  Device Available                                Yes
661
  Compiler Available                              Yes
662
  Linker Available                                Yes
663
  Device Extensions                               cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_fp64 cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation cl_intel_va_api_media_sharing
664

665
NULL platform behavior
666
  clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...)  Intel(R) OpenCL HD Graphics
667
  clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...)   Success [INTEL]
668
  clCreateContext(NULL, ...) [default]            Success [INTEL]
669
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU)  No devices found in platform
670
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  Success (1)
671
    Platform Name                                 Intel(R) OpenCL HD Graphics
672
    Device Name                                   Intel(R) Gen9 HD Graphics NEO
673
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR)  No devices found in platform
674
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM)  No devices found in platform
675
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL)  Success (1)
676
    Platform Name                                 Intel(R) OpenCL HD Graphics
677
    Device Name                                   Intel(R) Gen9 HD Graphics NEO
678

679
ICD loader properties
680
  ICD loader Name                                 OpenCL ICD Loader
681
  ICD loader Vendor                               OCL Icd free software
682
  ICD loader Version                              2.2.8
683
  ICD loader Profile                              OpenCL 1.2
684
        NOTE:	your OpenCL library declares to support OpenCL 1.2,
685
                but it seems to support up to OpenCL 2.1 too.
686
********************************************************************************/
687
"EU24_k11x11_cn3_g1_s4x4_d1x1_b1_in240x240_p0x0_num1_M96_activ1_eltwise0_FP32", "2 5 16 2 1 1 16 1 0",
688
"EU24_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
689
"EU24_k1x1_cn1024_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M2048_activ0_eltwise0_FP32", "7 4 16 2 1 1 16 1 0",
690
"EU24_k1x1_cn1024_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M512_activ1_eltwise0_FP32", "7 2 16 2 1 1 16 1 0",
691
"EU24_k1x1_cn128_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ1_eltwise1_FP32", "2 8 32 5 1 8 1 1 0",
692
"EU24_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M16_activ1_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
693
"EU24_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
694
"EU24_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP32", "10 2 16 2 1 1 16 1 0",
695
"EU24_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M96_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
696
"EU24_k1x1_cn2048_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M512_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
697
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ1_eltwise1_FP32", "14 1 16 2 1 1 16 1 0",
698
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M128_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
699
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP32", "10 1 16 2 1 1 16 1 0",
700
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
701
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M64_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
702
"EU24_k1x1_cn256_g1_s2x2_d1x1_b1_in64x64_p0x0_num1_M128_activ1_eltwise0_FP32", "1 8 32 5 1 8 1 1 0",
703
"EU24_k1x1_cn256_g1_s2x2_d1x1_b1_in64x64_p0x0_num1_M512_activ0_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
704
"EU24_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M16_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
705
"EU24_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
706
"EU24_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
707
"EU24_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M96_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
708
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M112_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
709
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "8 1 16 2 1 1 16 1 0",
710
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M144_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
711
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
712
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M2048_activ1_eltwise1_FP32", "7 1 16 2 1 1 16 1 0",
713
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
714
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
715
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
716
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M128_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
717
"EU24_k1x1_cn512_g1_s2x2_d1x1_b1_in32x32_p0x0_num1_M1024_activ0_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
718
"EU24_k1x1_cn512_g1_s2x2_d1x1_b1_in32x32_p0x0_num1_M256_activ1_eltwise0_FP32", "7 3 16 2 1 1 16 1 0",
719
"EU24_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
720
"EU24_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
721
"EU24_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
722
"EU24_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
723
"EU24_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M256_activ0_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
724
"EU24_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M256_activ1_eltwise1_FP32", "1 16 32 5 1 16 1 1 0",
725
"EU24_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M64_activ1_eltwise0_FP32", "2 8 32 5 1 8 1 1 0",
726
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
727
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
728
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP32", "7 2 16 2 1 1 16 1 0",
729
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
730
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP32", "7 1 8 2 1 1 8 1 0",
731
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M384_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
732
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M48_activ1_eltwise0_FP32", "4 1 16 2 1 1 16 1 0",
733
"EU24_k3x3_cn112_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M224_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
734
"EU24_k3x3_cn128_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
735
"EU24_k3x3_cn128_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M128_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
736
"EU24_k3x3_cn128_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M192_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
737
"EU24_k3x3_cn144_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M288_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
738
"EU24_k3x3_cn160_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M320_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
739
"EU24_k3x3_cn192_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
740
"EU24_k3x3_cn256_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
741
"EU24_k3x3_cn256_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP32", "13 1 16 2 1 1 16 1 0",
742
"EU24_k3x3_cn384_g2_s1x1_d1x1_b1_in16x16_p1x1_num1_M128_activ1_eltwise0_FP32", "13 1 16 2 1 1 16 1 0",
743
"EU24_k3x3_cn384_g2_s1x1_d1x1_b1_in16x16_p1x1_num1_M192_activ1_eltwise0_FP32", "13 1 16 2 1 1 16 1 0",
744
"EU24_k3x3_cn512_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M512_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
745
"EU24_k3x3_cn64_g1_s1x1_d1x1_b1_in64x64_p1x1_num1_M192_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
746
"EU24_k3x3_cn64_g1_s1x1_d1x1_b1_in64x64_p1x1_num1_M64_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
747
"EU24_k3x3_cn96_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M208_activ1_eltwise0_FP32", "14 2 16 2 1 1 16 1 0",
748
"EU24_k3x3_cn96_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M128_activ1_eltwise0_FP32", "14 1 16 2 1 1 16 1 0",
749
"EU24_k5x5_cn16_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M48_activ1_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
750
"EU24_k5x5_cn16_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M32_activ1_eltwise0_FP32", "4 3 8 2 1 1 8 1 0",
751
"EU24_k5x5_cn24_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP32", "5 1 16 2 1 1 16 1 0",
752
"EU24_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP32", "4 1 16 2 1 1 16 1 0",
753
"EU24_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP32", "7 1 16 2 1 1 16 1 0",
754
"EU24_k5x5_cn32_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M96_activ1_eltwise0_FP32", "7 2 16 2 1 1 16 1 0",
755
"EU24_k5x5_cn48_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP32", "3 7 8 2 1 1 8 1 0",
756
"EU24_k5x5_cn96_g2_s1x1_d1x1_b1_in32x32_p2x2_num1_M128_activ1_eltwise0_FP32", "9 3 16 2 1 1 16 1 0",
757
"EU24_k7x7_cn3_g1_s2x2_d1x1_b1_in224x224_p3x3_num1_M64_activ1_eltwise0_FP32", "4 4 16 2 1 1 16 1 0",
758
};
759

760
const char *default_kernel_config_intel_fp16[] = {
761
// Below is the information for OpenCL based on which these configurations tuned
762
/*******************************************************************************
763
Number of platforms                               1
764
  Platform Name                                   Intel(R) OpenCL HD Graphics
765
  Platform Vendor                                 Intel(R) Corporation
766
  Platform Version                                OpenCL 2.1
767
  Platform Profile                                FULL_PROFILE
768
  Platform Extensions                             cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_khr_fp64 cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation
769
  Platform Host timer resolution                  1ns
770
  Platform Extensions function suffix             INTEL
771

772
  Platform Name                                   Intel(R) OpenCL HD Graphics
773
Number of devices                                 1
774
  Device Name                                     Intel(R) Gen9 HD Graphics NEO
775
  Device Vendor                                   Intel(R) Corporation
776
  Device Vendor ID                                0x8086
777
  Device Version                                  OpenCL 2.1 NEO
778
  Driver Version                                  18.21.10858
779
  Device OpenCL C Version                         OpenCL C 2.0
780
  Device Type                                     GPU
781
  Device Profile                                  FULL_PROFILE
782
  Max compute units                               48
783
  Max clock frequency                             950MHz
784
  Device Partition                                (core)
785
    Max number of sub-devices                     0
786
    Supported partition types                     None
787
  Max work item dimensions                        3
788
  Max work item sizes                             256x256x256
789
  Max work group size                             256
790
  Preferred work group size multiple              32
791
  Max sub-groups per work group                   32
792
  Preferred / native vector sizes
793
    char                                                16 / 16
794
    short                                                8 / 8
795
    int                                                  4 / 4
796
    long                                                 1 / 1
797
    half                                                 8 / 8        (cl_khr_fp16)
798
    float                                                1 / 1
799
    double                                               1 / 1        (cl_khr_fp64)
800
  Half-precision Floating-point support           (cl_khr_fp16)
801
    Denormals                                     Yes
802
    Infinity and NANs                             Yes
803
    Round to nearest                              Yes
804
    Round to zero                                 Yes
805
    Round to infinity                             Yes
806
    IEEE754-2008 fused multiply-add               Yes
807
    Support is emulated in software               No
808
    Correctly-rounded divide and sqrt operations  No
809
  Single-precision Floating-point support         (core)
810
    Denormals                                     Yes
811
    Infinity and NANs                             Yes
812
    Round to nearest                              Yes
813
    Round to zero                                 Yes
814
    Round to infinity                             Yes
815
    IEEE754-2008 fused multiply-add               Yes
816
    Support is emulated in software               No
817
    Correctly-rounded divide and sqrt operations  Yes
818
  Double-precision Floating-point support         (cl_khr_fp64)
819
    Denormals                                     Yes
820
    Infinity and NANs                             Yes
821
    Round to nearest                              Yes
822
    Round to zero                                 Yes
823
    Round to infinity                             Yes
824
    IEEE754-2008 fused multiply-add               Yes
825
    Support is emulated in software               No
826
    Correctly-rounded divide and sqrt operations  No
827
  Address bits                                    64, Little-Endian
828
  Global memory size                              13364170752 (12.45GiB)
829
  Error Correction support                        No
830
  Max memory allocation                           4294959104 (4GiB)
831
  Unified memory for Host and Device              Yes
832
  Shared Virtual Memory (SVM) capabilities        (core)
833
    Coarse-grained buffer sharing                 Yes
834
    Fine-grained buffer sharing                   No
835
    Fine-grained system sharing                   No
836
    Atomics                                       No
837
  Minimum alignment for any data type             128 bytes
838
  Alignment of base address                       1024 bits (128 bytes)
839
  Preferred alignment for atomics
840
    SVM                                           64 bytes
841
    Global                                        64 bytes
842
    Local                                         64 bytes
843
  Max size for global variable                    65536 (64KiB)
844
  Preferred total size of global vars             4294959104 (4GiB)
845
  Global Memory cache type                        Read/Write
846
  Global Memory cache size                        1048576
847
  Global Memory cache line                        64 bytes
848
  Image support                                   Yes
849
    Max number of samplers per kernel             16
850
    Max size for 1D images from buffer            268434944 pixels
851
    Max 1D or 2D image array size                 2048 images
852
    Base address alignment for 2D image buffers   4 bytes
853
    Pitch alignment for 2D image buffers          4 bytes
854
    Max 2D image size                             16384x16384 pixels
855
    Max 3D image size                             16384x16384x2048 pixels
856
    Max number of read image args                 128
857
    Max number of write image args                128
858
    Max number of read/write image args           128
859
  Max number of pipe args                         16
860
  Max active pipe reservations                    1
861
  Max pipe packet size                            1024
862
  Local memory type                               Local
863
  Local memory size                               65536 (64KiB)
864
  Max constant buffer size                        4294959104 (4GiB)
865
  Max number of constant args                     8
866
  Max size of kernel argument                     1024
867
  Queue properties (on host)
868
    Out-of-order execution                        Yes
869
    Profiling                                     Yes
870
  Queue properties (on device)
871
    Out-of-order execution                        Yes
872
    Profiling                                     Yes
873
    Preferred size                                131072 (128KiB)
874
    Max size                                      67108864 (64MiB)
875
  Max queues on device                            1
876
  Max events on device                            1024
877
  Prefer user sync for interop                    Yes
878
  Profiling timer resolution                      83ns
879
  Execution capabilities
880
    Run OpenCL kernels                            Yes
881
    Run native kernels                            No
882
    Sub-group independent forward progress        Yes
883
    IL version                                    SPIR-V_1.0
884
    SPIR versions                                 1.2
885
  printf() buffer size                            4194304 (4MiB)
886
  Built-in kernels                                block_motion_estimate_intel;block_advanced_motion_estimate_check_intel;block_advanced_motion_estimate_bidirectional_check_intel;
887
  Motion Estimation accelerator version	(Intel)   2
888
  Device Available                                Yes
889
  Compiler Available                              Yes
890
  Linker Available                                Yes
891
  Device Extensions                               cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_khr_fp64 cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation
892

893
NULL platform behavior
894
  clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...)  No platform
895
  clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...)   No platform
896
  clCreateContext(NULL, ...) [default]            No platform
897
  clCreateContext(NULL, ...) [other]              Success [INTEL]
898
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU)  No platform
899
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  No platform
900
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR)  No platform
901
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM)  No platform
902
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL)  No platform
903
********************************************************************************/
904
"EU48_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M16_activ1_eltwise0_FP16", "11 1 16 2 1 1 16 1 0",
905
"EU48_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
906
"EU48_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
907
"EU48_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M96_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
908
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M128_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
909
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
910
"EU48_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
911
"EU48_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M16_activ1_eltwise0_FP16", "8 1 8 2 1 1 8 1 0",
912
"EU48_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
913
"EU48_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
914
"EU48_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M96_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
915
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M112_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
916
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
917
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M144_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
918
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
919
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ1_eltwise0_FP16", "7 1 8 2 1 1 8 1 0",
920
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP16", "8 1 16 2 1 1 16 1 0",
921
"EU48_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
922
"EU48_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
923
"EU48_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
924
"EU48_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
925
"EU48_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP16", "6 1 16 2 1 1 16 1 0",
926
"EU48_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M64_activ1_eltwise0_FP16", "1 16 32 5 1 16 1 1 0",
927
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
928
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
929
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
930
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
931
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP16", "4 1 8 2 1 1 8 1 0",
932
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M384_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
933
"EU48_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M48_activ1_eltwise0_FP16", "7 1 8 2 1 1 8 1 0",
934
"EU48_k3x3_cn112_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M224_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
935
"EU48_k3x3_cn128_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
936
"EU48_k3x3_cn128_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M192_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
937
"EU48_k3x3_cn144_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M288_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
938
"EU48_k3x3_cn160_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M320_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
939
"EU48_k3x3_cn192_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
940
"EU48_k3x3_cn64_g1_s1x1_d1x1_b1_in64x64_p1x1_num1_M192_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
941
"EU48_k3x3_cn96_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M208_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
942
"EU48_k3x3_cn96_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M128_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
943
"EU48_k5x5_cn16_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M48_activ1_eltwise0_FP16", "5 1 16 2 1 1 16 1 0",
944
"EU48_k5x5_cn16_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M32_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
945
"EU48_k5x5_cn24_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP16", "5 1 16 2 1 1 16 1 0",
946
"EU48_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP16", "8 1 16 2 1 1 16 1 0",
947
"EU48_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
948
"EU48_k5x5_cn32_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M96_activ1_eltwise0_FP16", "10 2 16 2 1 1 16 1 0",
949
"EU48_k5x5_cn48_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP16", "5 1 16 2 1 1 16 1 0",
950
"EU48_k7x7_cn3_g1_s2x2_d1x1_b1_in224x224_p3x3_num1_M64_activ1_eltwise0_FP16", "5 6 16 2 1 1 16 1 0",
951
"EU48_k11x11_cn3_g1_s4x4_d1x1_b1_in240x240_p0x0_num1_M96_activ1_eltwise0_FP16", "2 8 16 2 1 1 16 1 0",
952
"EU48_k3x3_cn256_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP16", "13 1 16 2 1 1 16 1 0",
953
"EU48_k3x3_cn384_g2_s1x1_d1x1_b1_in16x16_p1x1_num1_M128_activ1_eltwise0_FP16", "13 1 16 2 1 1 16 1 0",
954
"EU48_k3x3_cn384_g2_s1x1_d1x1_b1_in16x16_p1x1_num1_M192_activ1_eltwise0_FP16", "13 1 16 2 1 1 16 1 0",
955
"EU48_k5x5_cn96_g2_s1x1_d1x1_b1_in32x32_p2x2_num1_M128_activ1_eltwise0_FP16", "9 2 16 2 1 1 16 1 0",
956
// Below is the information for OpenCL based on which these configurations tuned
957
/*******************************************************************************
958
Number of platforms                               1
959
  Platform Name                                   Intel(R) OpenCL HD Graphics
960
  Platform Vendor                                 Intel(R) Corporation
961
  Platform Version                                OpenCL 2.1
962
  Platform Profile                                FULL_PROFILE
963
  Platform Extensions                             cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_fp64 cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation cl_intel_va_api_media_sharing
964
  Platform Host timer resolution                  1ns
965
  Platform Extensions function suffix             INTEL
966

967
  Platform Name                                   Intel(R) OpenCL HD Graphics
968
Number of devices                                 1
969
  Device Name                                     Intel(R) Gen9 HD Graphics NEO
970
  Device Vendor                                   Intel(R) Corporation
971
  Device Vendor ID                                0x8086
972
  Device Version                                  OpenCL 2.1 NEO
973
  Driver Version                                  18.23.10915
974
  Device OpenCL C Version                         OpenCL C 2.0
975
  Device Type                                     GPU
976
  Device Profile                                  FULL_PROFILE
977
  Max compute units                               24
978
  Max clock frequency                             1150MHz
979
  Device Partition                                (core)
980
    Max number of sub-devices                     0
981
    Supported partition types                     None
982
  Max work item dimensions                        3
983
  Max work item sizes                             256x256x256
984
  Max work group size                             256
985
  Preferred work group size multiple              32
986
  Max sub-groups per work group                   32
987
  Preferred / native vector sizes
988
    char                                                16 / 16
989
    short                                                8 / 8
990
    int                                                  4 / 4
991
    long                                                 1 / 1
992
    half                                                 8 / 8        (cl_khr_fp16)
993
    float                                                1 / 1
994
    double                                               1 / 1        (cl_khr_fp64)
995
  Half-precision Floating-point support           (cl_khr_fp16)
996
    Denormals                                     Yes
997
    Infinity and NANs                             Yes
998
    Round to nearest                              Yes
999
    Round to zero                                 Yes
1000
    Round to infinity                             Yes
1001
    IEEE754-2008 fused multiply-add               Yes
1002
    Support is emulated in software               No
1003
    Correctly-rounded divide and sqrt operations  No
1004
  Single-precision Floating-point support         (core)
1005
    Denormals                                     Yes
1006
    Infinity and NANs                             Yes
1007
    Round to nearest                              Yes
1008
    Round to zero                                 Yes
1009
    Round to infinity                             Yes
1010
    IEEE754-2008 fused multiply-add               Yes
1011
    Support is emulated in software               No
1012
    Correctly-rounded divide and sqrt operations  Yes
1013
  Double-precision Floating-point support         (cl_khr_fp64)
1014
    Denormals                                     Yes
1015
    Infinity and NANs                             Yes
1016
    Round to nearest                              Yes
1017
    Round to zero                                 Yes
1018
    Round to infinity                             Yes
1019
    IEEE754-2008 fused multiply-add               Yes
1020
    Support is emulated in software               No
1021
    Correctly-rounded divide and sqrt operations  No
1022
  Address bits                                    64, Little-Endian
1023
  Global memory size                              6575288320 (6.124GiB)
1024
  Error Correction support                        No
1025
  Max memory allocation                           3287644160 (3.062GiB)
1026
  Unified memory for Host and Device              Yes
1027
  Shared Virtual Memory (SVM) capabilities        (core)
1028
    Coarse-grained buffer sharing                 Yes
1029
    Fine-grained buffer sharing                   No
1030
    Fine-grained system sharing                   No
1031
    Atomics                                       No
1032
  Minimum alignment for any data type             128 bytes
1033
  Alignment of base address                       1024 bits (128 bytes)
1034
  Preferred alignment for atomics
1035
    SVM                                           64 bytes
1036
    Global                                        64 bytes
1037
    Local                                         64 bytes
1038
  Max size for global variable                    65536 (64KiB)
1039
  Preferred total size of global vars             3287644160 (3.062GiB)
1040
  Global Memory cache type                        Read/Write
1041
  Global Memory cache size                        524288
1042
  Global Memory cache line                        64 bytes
1043
  Image support                                   Yes
1044
    Max number of samplers per kernel             16
1045
    Max size for 1D images from buffer            205477760 pixels
1046
    Max 1D or 2D image array size                 2048 images
1047
    Base address alignment for 2D image buffers   4 bytes
1048
    Pitch alignment for 2D image buffers          4 bytes
1049
    Max 2D image size                             16384x16384 pixels
1050
    Max 3D image size                             16384x16384x2048 pixels
1051
    Max number of read image args                 128
1052
    Max number of write image args                128
1053
    Max number of read/write image args           128
1054
  Max number of pipe args                         16
1055
  Max active pipe reservations                    1
1056
  Max pipe packet size                            1024
1057
  Local memory type                               Local
1058
  Local memory size                               65536 (64KiB)
1059
  Max constant buffer size                        3287644160 (3.062GiB)
1060
  Max number of constant args                     8
1061
  Max size of kernel argument                     1024
1062
  Queue properties (on host)
1063
    Out-of-order execution                        Yes
1064
    Profiling                                     Yes
1065
  Queue properties (on device)
1066
    Out-of-order execution                        Yes
1067
    Profiling                                     Yes
1068
    Preferred size                                131072 (128KiB)
1069
    Max size                                      67108864 (64MiB)
1070
  Max queues on device                            1
1071
  Max events on device                            1024
1072
  Prefer user sync for interop                    Yes
1073
  Profiling timer resolution                      83ns
1074
  Execution capabilities
1075
    Run OpenCL kernels                            Yes
1076
    Run native kernels                            No
1077
    Sub-group independent forward progress        Yes
1078
    IL version                                    SPIR-V_1.0
1079
    SPIR versions                                 1.2
1080
  printf() buffer size                            4194304 (4MiB)
1081
  Built-in kernels                                block_motion_estimate_intel;block_advanced_motion_estimate_check_intel;block_advanced_motion_estimate_bidirectional_check_intel;
1082
  Motion Estimation accelerator version	(Intel)   2
1083
  Device Available                                Yes
1084
  Compiler Available                              Yes
1085
  Linker Available                                Yes
1086
  Device Extensions                               cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_depth_images cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_image2d_from_buffer cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_media_block_io cl_intel_driver_diagnostics cl_intel_device_side_avc_motion_estimation cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_khr_fp64 cl_khr_subgroups cl_khr_il_program cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_intel_planar_yuv cl_intel_packed_yuv cl_intel_motion_estimation cl_intel_advanced_motion_estimation cl_intel_va_api_media_sharing
1087

1088
NULL platform behavior
1089
  clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...)  Intel(R) OpenCL HD Graphics
1090
  clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...)   Success [INTEL]
1091
  clCreateContext(NULL, ...) [default]            Success [INTEL]
1092
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU)  No devices found in platform
1093
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  Success (1)
1094
    Platform Name                                 Intel(R) OpenCL HD Graphics
1095
    Device Name                                   Intel(R) Gen9 HD Graphics NEO
1096
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR)  No devices found in platform
1097
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM)  No devices found in platform
1098
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL)  Success (1)
1099
    Platform Name                                 Intel(R) OpenCL HD Graphics
1100
    Device Name                                   Intel(R) Gen9 HD Graphics NEO
1101

1102
ICD loader properties
1103
  ICD loader Name                                 OpenCL ICD Loader
1104
  ICD loader Vendor                               OCL Icd free software
1105
  ICD loader Version                              2.2.8
1106
  ICD loader Profile                              OpenCL 1.2
1107
        NOTE:	your OpenCL library declares to support OpenCL 1.2,
1108
                but it seems to support up to OpenCL 2.1 too.
1109
********************************************************************************/
1110
"EU24_k11x11_cn3_g1_s4x4_d1x1_b1_in240x240_p0x0_num1_M96_activ1_eltwise0_FP16", "2 7 16 2 1 1 16 1 0",
1111
"EU24_k1x1_cn1024_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1112
"EU24_k1x1_cn1024_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M2048_activ0_eltwise0_FP16", "7 4 16 2 1 1 16 1 0",
1113
"EU24_k1x1_cn1024_g1_s2x2_d1x1_b1_in16x16_p0x0_num1_M512_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1114
"EU24_k1x1_cn128_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M512_activ1_eltwise1_FP16", "1 16 32 5 1 16 1 1 0",
1115
"EU24_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M16_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1116
"EU24_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1117
"EU24_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP16", "10 3 16 2 1 1 16 1 0",
1118
"EU24_k1x1_cn192_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M96_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1119
"EU24_k1x1_cn2048_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M512_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1120
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M1024_activ1_eltwise1_FP16", "14 2 16 2 1 1 16 1 0",
1121
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M128_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1122
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M32_activ1_eltwise0_FP16", "10 2 16 2 1 1 16 1 0",
1123
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M64_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1124
"EU24_k1x1_cn256_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M64_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1125
"EU24_k1x1_cn256_g1_s2x2_d1x1_b1_in64x64_p0x0_num1_M128_activ1_eltwise0_FP16", "8 4 16 2 1 1 16 1 0",
1126
"EU24_k1x1_cn256_g1_s2x2_d1x1_b1_in64x64_p0x0_num1_M512_activ0_eltwise0_FP16", "1 16 32 5 1 16 1 1 0",
1127
"EU24_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M16_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1128
"EU24_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1129
"EU24_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1130
"EU24_k1x1_cn480_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M96_activ1_eltwise0_FP16", "7 2 16 2 1 1 16 1 0",
1131
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M112_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1132
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1133
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M144_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1134
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1135
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M2048_activ1_eltwise1_FP16", "7 1 16 2 1 1 16 1 0",
1136
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M24_activ1_eltwise0_FP16", "10 1 16 2 1 1 16 1 0",
1137
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP16", "9 1 16 2 1 1 16 1 0",
1138
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M64_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1139
"EU24_k1x1_cn512_g1_s1x1_d1x1_b1_in32x32_p0x0_num1_M128_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1140
"EU24_k1x1_cn512_g1_s2x2_d1x1_b1_in32x32_p0x0_num1_M1024_activ0_eltwise0_FP16", "8 3 16 2 1 1 16 1 0",
1141
"EU24_k1x1_cn512_g1_s2x2_d1x1_b1_in32x32_p0x0_num1_M256_activ1_eltwise0_FP16", "7 3 16 2 1 1 16 1 0",
1142
"EU24_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1143
"EU24_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1144
"EU24_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1145
"EU24_k1x1_cn528_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP16", "8 1 16 2 1 1 16 1 0",
1146
"EU24_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M256_activ0_eltwise0_FP16", "1 16 32 5 1 16 1 1 0",
1147
"EU24_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M256_activ1_eltwise1_FP16", "1 16 32 5 1 16 1 1 0",
1148
"EU24_k1x1_cn64_g1_s1x1_d1x1_b1_in64x64_p0x0_num1_M64_activ1_eltwise0_FP16", "1 16 32 5 1 16 1 1 0",
1149
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M128_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1150
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M160_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1151
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M192_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1152
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M256_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1153
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M32_activ1_eltwise0_FP16", "6 1 16 2 1 1 16 1 0",
1154
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M384_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1155
"EU24_k1x1_cn832_g1_s1x1_d1x1_b1_in16x16_p0x0_num1_M48_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1156
"EU24_k3x3_cn112_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M224_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1157
"EU24_k3x3_cn128_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1158
"EU24_k3x3_cn128_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M128_activ1_eltwise0_FP16", "10 2 16 2 1 1 16 1 0",
1159
"EU24_k3x3_cn128_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M192_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1160
"EU24_k3x3_cn144_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M288_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1161
"EU24_k3x3_cn160_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M320_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1162
"EU24_k3x3_cn192_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1163
"EU24_k3x3_cn256_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M256_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1164
"EU24_k3x3_cn256_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M384_activ1_eltwise0_FP16", "13 2 16 2 1 1 16 1 0",
1165
"EU24_k3x3_cn384_g2_s1x1_d1x1_b1_in16x16_p1x1_num1_M128_activ1_eltwise0_FP16", "13 1 16 2 1 1 16 1 0",
1166
"EU24_k3x3_cn384_g2_s1x1_d1x1_b1_in16x16_p1x1_num1_M192_activ1_eltwise0_FP16", "13 1 16 2 1 1 16 1 0",
1167
"EU24_k3x3_cn512_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M512_activ1_eltwise0_FP16", "7 2 16 2 1 1 16 1 0",
1168
"EU24_k3x3_cn64_g1_s1x1_d1x1_b1_in64x64_p1x1_num1_M192_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1169
"EU24_k3x3_cn64_g1_s1x1_d1x1_b1_in64x64_p1x1_num1_M64_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1170
"EU24_k3x3_cn96_g1_s1x1_d1x1_b1_in16x16_p1x1_num1_M208_activ1_eltwise0_FP16", "14 2 16 2 1 1 16 1 0",
1171
"EU24_k3x3_cn96_g1_s1x1_d1x1_b1_in32x32_p1x1_num1_M128_activ1_eltwise0_FP16", "14 1 16 2 1 1 16 1 0",
1172
"EU24_k5x5_cn16_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M48_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1173
"EU24_k5x5_cn16_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M32_activ1_eltwise0_FP16", "7 2 16 2 1 1 16 1 0",
1174
"EU24_k5x5_cn24_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1175
"EU24_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP16", "7 2 16 2 1 1 16 1 0",
1176
"EU24_k5x5_cn32_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M64_activ1_eltwise0_FP16", "7 1 16 2 1 1 16 1 0",
1177
"EU24_k5x5_cn32_g1_s1x1_d1x1_b1_in32x32_p2x2_num1_M96_activ1_eltwise0_FP16", "7 2 16 2 1 1 16 1 0",
1178
"EU24_k5x5_cn48_g1_s1x1_d1x1_b1_in16x16_p2x2_num1_M128_activ1_eltwise0_FP16", "4 1 16 2 1 1 16 1 0",
1179
"EU24_k5x5_cn96_g2_s1x1_d1x1_b1_in32x32_p2x2_num1_M128_activ1_eltwise0_FP16", "7 3 16 2 1 1 16 1 0",
1180
"EU24_k7x7_cn3_g1_s2x2_d1x1_b1_in224x224_p3x3_num1_M64_activ1_eltwise0_FP16", "4 7 16 2 1 1 16 1 0",
1181
};
1182
#endif
1183

1184
Product

Resources

Company
