summaryrefslogtreecommitdiff
path: root/libs/ARMComputeEx/src/core/CL/cl_kernels/prelu_quantized.cl
blob: 7e97b7ed6ca81c3d00fb63431d9f951fcd31d65b (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
/*
 * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved
 * Copyright (c) 2016-2018 ARM Limited.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
#include "helpers.h"
#define SUB(x, y) (x) - (y)

#if defined(OFF_IN1) && defined(OFF_IN2) && defined(OFF_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(VEC_SIZE)

#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
#define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE)
#define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)

/** Returns result of prelu function implemented as below:
 * f(input) = alpha * input for input < 0, f(input) = input for input >= 0.
 *
 * @attention Data type can be passed using the -DDATA_TYPE_IN compile flag, e.g. -DDATA_TYPE_IN=uchar
 * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
 * @note Can only take uchar data types.
 *
 * @param[in]  input1_ptr                            Pointer to the source image. Supported Data types : QASYMM8
 * @param[in]  input1_stride_x                       Stride of the source image in X dimension (in bytes)
 * @param[in]  input1_step_x                         input1_stride_x * number of elements along X processed per workitem(in bytes)
 * @param[in]  input1_stride_y                       Stride of the source image in Y dimension (in bytes)
 * @param[in]  input1_step_y                         input1_stride_y * number of elements along Y processed per workitem(in bytes)
 * @param[in]  input1_stride_z                       Stride of the source tensor in Z dimension (in bytes)
 * @param[in]  input1_step_z                         input1_stride_z * number of elements along Z processed per workitem(in bytes)
 * @param[in]  input1_offset_first_element_in_bytes  The offset of the first element in the source image
 *
 * @param[in]  alpha_ptr                            Pointer to the source image. Supported Data types : QASYMM8
 * @param[in]  alpha_stride_x                       Stride of the source image in X dimension (in bytes)
 * @param[in]  alpha_step_x                         input2_stride_x * number of elements along X processed per workitem(in bytes)
 * @param[in]  alpha_stride_y                       Stride of the source image in Y dimension (in bytes)
 * @param[in]  alpha_step_y                         input2_stride_y * number of elements along Y processed per workitem(in bytes)
 * @param[in]  alpha_stride_z                       Stride of the source tensor in Z dimension (in bytes)
 * @param[in]  alpha_step_z                         input2_stride_z * number of elements along Z processed per workitem(in bytes)
 * @param[in]  alpha_offset_first_element_in_bytes  The offset of the first element in the source image
 *
 * @param[out] output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
 * @param[in]  output_stride_x                      Stride of the destination image in X dimension (in bytes)
 * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
 * @param[in]  output_stride_y                      Stride of the destination image in Y dimension (in bytes)
 * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
 * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
 * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
 * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination image
 */
__kernel void prelu_qasymm8(
    TENSOR3D_DECLARATION(input),
    TENSOR3D_DECLARATION(alpha),
    TENSOR3D_DECLARATION(output))
{
    // Get pixels pointer
    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
    Tensor3D alpha = CONVERT_TO_TENSOR3D_STRUCT(alpha);
    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);

    VEC_INT in_a = CONVERT(VLOAD(VEC_SIZE)(0, (__global uchar *)input.ptr), VEC_INT);
    VEC_INT in_b = CONVERT(VLOAD(VEC_SIZE)(0, (__global uchar *)alpha.ptr), VEC_INT);

    in_a = SUB(in_a, (VEC_INT)((int)OFF_IN1));
    in_b = SUB(in_b, (VEC_INT)((int)OFF_IN2));

    const VEC_FLOAT in1f32  = CONVERT(in_a, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN1);
    const VEC_FLOAT in2f32  = CONVERT(in_b, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN2);
    const VEC_FLOAT outf32 = in1f32 < 0 ? in1f32 * in2f32 : in1f32;
    const VEC_FLOAT qresf32 = outf32 / ((VEC_FLOAT)(float)SCALE_OUT) + ((VEC_FLOAT)((float)OFF_OUT));
    const VEC_UCHAR res     = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_UCHAR);

    VSTORE(VEC_SIZE)
    (res, 0, (__global uchar *)output.ptr);
}

#endif // defined(OFF_IN1) && defined(OFF_IN2) && defined(OFF_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(VEC_SIZE)