/home/liu/actions-runner/_work/ccv/ccv/lib/nnc/cmd/pad/ccv_nnc_pad.c
Line | Count | Source (jump to first uncovered line) |
1 | | #include "ccv.h" |
2 | | #include "nnc/ccv_nnc.h" |
3 | | #include "nnc/ccv_nnc_easy.h" |
4 | | #include "nnc/ccv_nnc_internal.h" |
5 | | |
6 | | static int _ccv_nnc_pad_forw_bitmask(const ccv_nnc_cmd_param_t cmd, const int input_size, const int output_size, const uint64_t* const input_bitmasks, const int input_bitmask_size, const uint64_t* const output_bitmasks, const int output_bitmask_size) |
7 | 1 | { |
8 | 1 | if ((input_bitmasks[0] & 1u) == 1u && output_bitmasks[0] == 1u) |
9 | 1 | return 1; |
10 | 0 | return 0; |
11 | 1 | } |
12 | | |
13 | | static int _ccv_nnc_pad_back_bitmask(const ccv_nnc_cmd_param_t cmd, const int input_size, const int output_size, const uint64_t* const input_bitmasks, const int input_bitmask_size, const uint64_t* const output_bitmasks, const int output_bitmask_size) |
14 | 0 | { |
15 | | // We don't need the original input since roi align does averaging. |
16 | | // We do, however, need the coordinate. |
17 | 0 | if ((input_bitmasks[0] & 1u) == (1u << 0) && output_bitmasks[0] == 1u) |
18 | 0 | return 1; |
19 | 0 | return 0; |
20 | 0 | } |
21 | | |
22 | | static void _ccv_nnc_pad_tensor_auto_forw(const ccv_nnc_cmd_param_t cmd, const ccv_nnc_tensor_param_t* inputs, const int input_size, const ccv_nnc_hint_t hint, ccv_nnc_tensor_param_t* outputs, const int output_size) |
23 | 3 | { |
24 | 3 | assert(output_size == 1); |
25 | 3 | outputs[0] = inputs[0]; |
26 | 3 | const int nd = ccv_nnc_tensor_nd(outputs[0].dim); |
27 | 3 | int i; |
28 | 15 | for (i = 0; i < nd; i++12 ) |
29 | 12 | outputs[0].dim[i] += cmd.size.dim[i] + cmd.pad.end[i]; |
30 | 3 | } |
31 | | |
32 | | static void _ccv_nnc_pad_tensor_auto_back(const ccv_nnc_cmd_param_t cmd, const ccv_nnc_tensor_param_t* inputs, const int input_size, const ccv_nnc_hint_t hint, ccv_nnc_tensor_param_t* outputs, const int output_size) |
33 | 0 | { |
34 | 0 | assert(output_size == 1); |
35 | 0 | outputs[0] = inputs[1]; |
36 | 0 | } |
37 | | |
38 | | REGISTER_COMMAND(CCV_NNC_PAD_FORWARD)(ccv_nnc_cmd_registry_t* const registry) |
39 | | FIND_BACKEND(ccv_nnc_pad_cpu_ref.c, gpu/ccv_nnc_pad_gpu_ref.cu, mps/ccv_nnc_pad_mps.m) |
40 | 1 | { |
41 | 1 | registry->bitmask = _ccv_nnc_pad_forw_bitmask; |
42 | 1 | registry->tensor_auto = _ccv_nnc_pad_tensor_auto_forw; |
43 | 1 | } |
44 | | |
45 | | REGISTER_COMMAND(CCV_NNC_PAD_BACKWARD)(ccv_nnc_cmd_registry_t* const registry) |
46 | | FIND_BACKEND(ccv_nnc_pad_cpu_ref.c, gpu/ccv_nnc_pad_gpu_ref.cu, mps/ccv_nnc_pad_mps.m) |
47 | 1 | { |
48 | 1 | registry->bitmask = _ccv_nnc_pad_back_bitmask; |
49 | 1 | registry->tensor_auto = _ccv_nnc_pad_tensor_auto_back; |
50 | 1 | } |
51 | | |
52 | | //@REGISTER_EASY_COMMAND_MACRO(CCV_NNC_PAD_FORWARD) |
53 | | #define CMD_PAD_FORWARD(_type, _begin, _end) ccv_nnc_cmd(CCV_NNC_PAD_FORWARD, 0, ((ccv_nnc_cmd_param_t){.size={.dim={ESCAPE_X _begin}},.pad={.type=_type,.end={ESCAPE_X _end}}}), 0) |
54 | | //@REGISTER_EASY_COMMAND_MACRO(CCV_NNC_PAD_BACKWARD) |
55 | | #define CMD_PAD_BACKWARD(_type, _begin, _end) ccv_nnc_cmd(CCV_NNC_PAD_BACKWARD, 0, ((ccv_nnc_cmd_param_t){.size={.dim={ESCAPE_X _begin}},.pad={.type=_type,.end={ESCAPE_X _end}}}), 0) |