Compute Library
 23.05
cast.cl
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #include "helpers.h"
25 
26 #ifdef SATURATE
27 #define CONVERT_DOWN(x, type) CONVERT_SAT(x, type)
28 #else /* SATURATE */
29 #define CONVERT_DOWN(x, type) CONVERT(x, type)
30 #endif /* SATURATE */
31 
32 #define CONVERT_UP(x, type) CONVERT(x, type)
33 
34 /** This function performs a down-casting
35  *
36  * @attention For QSYMM8_PER_CHANNEL -> QASYMM8, it is user's responsibility to keep track of the quantization info.
37  *
38  * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
39  * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
40  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
41  * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
42  *
43  * @param[in] in_ptr Pointer to the source image. Supported data types: U8/S8/QSYMM8_PER_CHANNEL/U16/S16/U32/S32/F16/F32
44  * @param[in] in_stride_x Stride of the source image in X dimension (in bytes)
45  * @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes)
46  * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes)
47  * @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes)
48  * @param[in] in_stride_z Stride of the source tensor in Z dimension (in bytes)
49  * @param[in] in_step_z in_stride_z * number of elements along Z processed per workitem(in bytes)
50  * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image
51  * @param[out] out_ptr Pointer to the destination image. Supported data types: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32
52  * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes)
53  * @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes)
54  * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes)
55  * @param[in] out_step_y out_stride_y * number of elements along Y processed per workitem(in bytes)
56  * @param[in] out_stride_z Stride of the source tensor in Z dimension (in bytes)
57  * @param[in] out_step_z out_stride_z * number of elements along Z processed per workitem(in bytes)
58  * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image
59  */
60 __kernel void cast_down(
61  TENSOR3D_DECLARATION(in),
62  TENSOR3D_DECLARATION(out))
63 {
64  int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
65 
66  __global uchar *in_addr = in_ptr + in_offset_first_element_in_bytes + sizeof(DATA_TYPE_IN) * x_offs + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z;
67  __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + sizeof(DATA_TYPE_OUT) * x_offs + get_global_id(1) * out_stride_y + get_global_id(2) * out_stride_z;
68 
69  // Load data
70  VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE)
71  in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in_addr);
72 
73 #if defined(IS_DATA_TYPE_QUANTIZED)
74  in_data ^= (VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE))0x80;
75 #endif // defined(IS_DATA_TYPE_QUANTIZED)
76 
77 #if defined(IS_DATA_TYPE_FLOAT)
78  VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
79  res0 = CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE));
80  STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
81 #else /* defined(IS_DATA_TYPE_FLOAT) */
82  VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
83  res0 = CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE));
84  STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
85 #endif /* defined(IS_DATA_TYPE_FLOAT) */
86 }
87 
88 /** This function performs a up-casting
89  *
90  * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
91  * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
92  * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
93  * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
94  *
95  * @param[in] in_ptr Pointer to the source image. Supported data types: U8/S8/U16/S16/U32/S32/F16/F32
96  * @param[in] in_stride_x Stride of the source image in X dimension (in bytes)
97  * @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes)
98  * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes)
99  * @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes)
100  * @param[in] in_stride_z Stride of the source tensor in Z dimension (in bytes)
101  * @param[in] in_step_z in_stride_z * number of elements along Z processed per workitem(in bytes)
102  * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image
103  * @param[out] out_ptr Pointer to the destination image. Supported data types: U8/U16/S16/U32/S32/F16/F32
104  * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes)
105  * @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes)
106  * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes)
107  * @param[in] out_step_y out_stride_y * number of elements along Y processed per workitem(in bytes)
108  * @param[in] out_stride_z Stride of the source tensor in Z dimension (in bytes)
109  * @param[in] out_step_z out_stride_z * number of elements along Z processed per workitem(in bytes)
110  * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image
111  */
112 __kernel void cast_up(
113  TENSOR3D_DECLARATION(in),
114  TENSOR3D_DECLARATION(out))
115 {
116  int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0);
117 
118  __global uchar *in_addr = in_ptr + in_offset_first_element_in_bytes + sizeof(DATA_TYPE_IN) * x_offs + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z;
119  __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + sizeof(DATA_TYPE_OUT) * x_offs + get_global_id(1) * out_stride_y + get_global_id(2) * out_stride_z;
120 
121  // Load data
122  VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE)
123  in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in_addr);
124 
125 #if defined(IS_DATA_TYPE_FLOAT)
126  VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
127  res0 = CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE));
128  STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
129 #else /* defined(IS_DATA_TYPE_FLOAT) */
130  VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
131  res0 = CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE));
132  STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
133 #endif /* defined(IS_DATA_TYPE_FLOAT) */
134 }
VEC_SIZE
#define VEC_SIZE
Definition: qlstm_layer_normalization.cl:54
cast_down
__kernel void cast_down(__global uchar *in_ptr, uint in_stride_x, uint in_step_x, uint in_stride_y, uint in_step_y, uint in_stride_z, uint in_step_z, uint in_offset_first_element_in_bytes, __global uchar *out_ptr, uint out_stride_x, uint out_step_x, uint out_stride_y, uint out_step_y, uint out_stride_z, uint out_step_z, uint out_offset_first_element_in_bytes)
This function performs a down-casting.
Definition: cast.cl:60
CONVERT_UP
#define CONVERT_UP(x, type)
Definition: cast.cl:32
cast_up
__kernel void cast_up(__global uchar *in_ptr, uint in_stride_x, uint in_step_x, uint in_stride_y, uint in_step_y, uint in_stride_z, uint in_step_z, uint in_offset_first_element_in_bytes, __global uchar *out_ptr, uint out_stride_x, uint out_step_x, uint out_stride_y, uint out_step_y, uint out_stride_z, uint out_step_z, uint out_offset_first_element_in_bytes)
This function performs a up-casting.
Definition: cast.cl:112
STORE_VECTOR_SELECT
#define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond)
Store a vector that can only be partial in x.
Definition: load_store_utility.h:584
CONVERT_DOWN
#define CONVERT_DOWN(x, type)
Definition: cast.cl:29
helpers.h
VLOAD
#define VLOAD(size)
Definition: helpers.h:204
TENSOR3D_DECLARATION
#define TENSOR3D_DECLARATION(name)
Definition: helpers.h:839
VEC_DATA_TYPE
#define VEC_DATA_TYPE(type, size)
Definition: helpers.h:754