Compute Library
 21.02
elementwise_operation_quantized.cl
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018-2020 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 #define SUB(x, y) (x - y)
27 #define ADD(x, y) (x + y)
28 #define MAX(x, y) max((x), (y))
29 #define MIN(x, y) min((x), (y))
30 #define SQUARED_DIFF(x, y) (x - y) * (x - y)
31 #define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE_OUT)0), SELECT_VEC_DATA_TYPE(float, VEC_SIZE_OUT))))
32 #define DIV(x, y) (x / y)
33 
34 #define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
35 #define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
36 
37 #define OP_FUN_NAME_STR(op) elementwise_operation_##op##_quantized
38 #define OP_FUN_NAME(op) OP_FUN_NAME_STR(op)
39 
40 #if defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT)
41 
42 #define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE_OUT)
43 #define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE_OUT)
44 #define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
45 
46 /** This function executes an element-wise operation among two tensors.
47  *
48  * @note Vector sizes of inputs and output have to be passed at compile time using -DVEC_SIZE_IN1, -DVEC_SIZE_IN2, -DVEC_SIZE_OUT.
49  * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
50  * @note In case of broadcasting along the X dimension the proper preprocessor argument should be passed depending on the input (e.g. -DIS_IN1_X_BROADCASTING, -DIS_IN2_X_BROADCASTING)
51  * @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10
52  * @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10
53  * @note The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, i.e. -DOFFSET_OUT=10
54  * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10
55  * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10
56  * @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, i.e. -DSCALE_OUT=10
57  * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
58  * @note The element-wise operation to be executed has to be passed at compile time using -DOP (e.g., -DOP=ADD)
59  * @note For QSYMM16 operations OFFSET_IN1, OFFSET_IN2 and OFFSET_OUT must be set to zero
60  * @note The data type must be passed at compile time using -DDATA_TYPE_OUT, i.e. -DDATA_TYPE_OUT=uchar
61  *
62  * @param[in] in1_ptr Pointer to the source tensor. Supported data types: QASYMM8/QSYMM16
63  * @param[in] in1_stride_x Stride of the source tensor in X dimension (in bytes)
64  * @param[in] in1_step_x in1_stride_x * number of elements along X processed per workitem(in bytes)
65  * @param[in] in1_stride_y Stride of the source tensor in Y dimension (in bytes)
66  * @param[in] in1_step_y in1_stride_y * number of elements along Y processed per workitem(in bytes)
67  * @param[in] in1_stride_z Stride of the source tensor in Z dimension (in bytes)
68  * @param[in] in1_step_z in1_stride_z * number of elements along Z processed per workitem(in bytes)
69  * @param[in] in1_offset_first_element_in_bytes The offset of the first element in the source tensor
70  * @param[in] in2_ptr Pointer to the source tensor. Supported data types: same as @p in1_ptr
71  * @param[in] in2_stride_x Stride of the source tensor in X dimension (in bytes)
72  * @param[in] in2_step_x in2_stride_x * number of elements along X processed per workitem(in bytes)
73  * @param[in] in2_stride_y Stride of the source tensor in Y dimension (in bytes)
74  * @param[in] in2_step_y in2_stride_y * number of elements along Y processed per workitem(in bytes)
75  * @param[in] in2_stride_z Stride of the source tensor in Z dimension (in bytes)
76  * @param[in] in2_step_z in2_stride_z * number of elements along Z processed per workitem(in bytes)
77  * @param[in] in2_offset_first_element_in_bytes The offset of the first element in the source tensor
78  * @param[out] out_ptr Pointer to the destination tensor. Supported data types: same as @p in1_ptr
79  * @param[in] out_stride_x Stride of the destination tensor in X dimension (in bytes)
80  * @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes)
81  * @param[in] out_stride_y Stride of the destination tensor in Y dimension (in bytes)
82  * @param[in] out_step_y out_stride_y * number of elements along Y processed per workitem(in bytes)
83  * @param[in] out_stride_z Stride of the source tensor in Z dimension (in bytes)
84  * @param[in] out_step_z out_stride_z * number of elements along Z processed per workitem(in bytes)
85  * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination tensor
86  */
87 __kernel void OP_FUN_NAME(OP)(
91 {
92 #if VEC_SIZE_IN1 == 1
93  uint in1_x_offs = 0;
94 #else // VEC_SIZE_IN1 == 1
95  uint in1_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN1 - (VEC_SIZE_IN1 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN1), 0);
96 #endif // VEC_SIZE_IN1 == 1
97 #if VEC_SIZE_IN2 == 1
98  uint in2_x_offs = 0;
99 #else // VEC_SIZE_IN2 == 1
100  uint in2_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN2 - (VEC_SIZE_IN2 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN2), 0);
101 #endif // VEC_SIZE_IN2 == 1
102  uint out_x_offs = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
103 
104  // Get pixels pointer
105  __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + in1_x_offs * sizeof(DATA_TYPE_OUT) + get_global_id(1) * in1_step_y + get_global_id(2) * in1_step_z;
106  __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + in2_x_offs * sizeof(DATA_TYPE_OUT) + get_global_id(1) * in2_step_y + get_global_id(2) * in2_step_z;
107  __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + out_x_offs * sizeof(DATA_TYPE_OUT) + get_global_id(1) * out_step_y + get_global_id(2) * out_step_z;
108 
109  VEC_INT in_a = CONVERT((VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE_OUT *)in1_addr)), VEC_INT);
110  VEC_INT in_b = CONVERT((VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE_OUT *)in2_addr)), VEC_INT);
111 
112  in_a = SUB(in_a, (VEC_INT)((int)OFFSET_IN1));
113  in_b = SUB(in_b, (VEC_INT)((int)OFFSET_IN2));
114 
115  const VEC_FLOAT in1f32 = CONVERT(in_a, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN1);
116  const VEC_FLOAT in2f32 = CONVERT(in_b, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN2);
117  const VEC_FLOAT qresf32 = OP(in1f32, in2f32) / ((VEC_FLOAT)(float)SCALE_OUT) + ((VEC_FLOAT)((float)OFFSET_OUT));
118  const VEC_TYPE res0 = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE);
119 
120  // Store result
121  STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
122 }
123 #endif /* defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT) */
#define CONVERT(x, type)
Definition: helpers.h:517
#define CONVERT_SAT(a, b)
#define VEC_FLOAT
#define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond)
Store a vector that can only be partial in x.
#define CONVERT_DOWN(x, type)
#define DATA_TYPE_OUT
#define SUB(x, y)
#define OP_FUN_NAME(op)
#define VLOAD(size)
Definition: helpers.h:203
#define TENSOR3D_DECLARATION(name)
Definition: helpers.h:589
#define VEC_DATA_TYPE(type, size)
Definition: helpers.h:514