xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/common/pixelwise_mul_int.cl (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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#if defined(SATURATE)
27#define CONVERT_OP_INT_STR(x, type, size) (convert_##type##size##_sat(x))
28#else // SATURATE
29#define CONVERT_OP_INT_STR(x, type, size) (convert_##type##size(x))
30#endif // SATURATE
31#define CONVERT_OP_INT(x, type, size) CONVERT_OP_INT_STR(x, type, size)
32
33#define MUL_OP(x, y, scale, type, size) CONVERT_OP_INT((x) * (y) >> scale, type, size)
34
35#define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
36#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
37
38#if defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(ACC_DATA_TYPE) && defined(DATA_TYPE_OUT)
39
40#define VEC_ACC_TYPE VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE_OUT)
41#define VEC_OUT_TYPE VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
42
43/** Performs a pixelwise multiplication with integer scale of integer inputs.
44 *
45 * @attention The inputs and output data types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT:
46 * e.g. -DDATA_TYPE_IN1=uchar -DDATA_TYPE_IN2=ushort -DDATA_TYPE_OUT=short
47 * @attention The data_type of the intermediate result of the multiplication should passed as well using -DACC_DATA_TYPE.
48 * e.g. If one of inputs is S16 -DACC_DATA_TYPE=int should be passed else -DACC_DATA_TYPE=short.
49 *
50 * @param[in]  in1_ptr                           Pointer to the source image. Supported data types: U8/S16
51 * @param[in]  in1_stride_x                      Stride of the source image in X dimension (in bytes)
52 * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
53 * @param[in]  in1_stride_y                      Stride of the source image in Y dimension (in bytes)
54 * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
55 * @param[in]  in1_stride_z                      Stride of the source image in Y dimension (in bytes)
56 * @param[in]  in1_step_z                        in1_stride_z * number of elements along Y processed per workitem(in bytes)
57 * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source image
58 * @param[in]  in2_ptr                           Pointer to the source image. Supported data types: same as @p in1_ptr
59 * @param[in]  in2_stride_x                      Stride of the source image in X dimension (in bytes)
60 * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
61 * @param[in]  in2_stride_y                      Stride of the source image in Y dimension (in bytes)
62 * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
63 * @param[in]  in2_stride_z                      Stride of the source image in Y dimension (in bytes)
64 * @param[in]  in2_step_z                        in2_stride_z * number of elements along Y processed per workitem(in bytes)
65 * @param[in]  in2_offset_first_element_in_bytes The offset of the first element in the source image
66 * @param[out] out_ptr                           Pointer to the destination image. Supported data types: same as @p in1_ptr
67 * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
68 * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
69 * @param[in]  out_stride_y                      Stride of the destination image in Y dimension (in bytes)
70 * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
71 * @param[in]  out_stride_z                      Stride of the destination image in Y dimension (in bytes)
72 * @param[in]  out_step_z                        out_stride_z * number of elements along Y processed per workitem(in bytes)
73 * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination image
74 * @param[in]  scale                             Integer scaling factor. Supported data types: S32.
75 */
76__kernel void pixelwise_mul_int(
77    TENSOR3D_DECLARATION(in1),
78    TENSOR3D_DECLARATION(in2),
79#if !defined(IN_PLACE)
80    TENSOR3D_DECLARATION(out),
81#endif // !defined(IN_PLACE)
82    const uint scale)
83{
84    size_t x = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
85    size_t y = get_global_id(1);
86    size_t z = get_global_id(2);
87
88    __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + x * in1_stride_x + y * in1_stride_y + z * in1_stride_z;
89    __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + x * in2_stride_x + y * in2_stride_y + z * in2_stride_z;
90    __global        uchar *
91#if !defined(IN_PLACE)
92    out_addr = out_ptr + out_offset_first_element_in_bytes + x * out_stride_x + y * out_stride_y + z * out_stride_z;
93#else // !defined(IN_PLACE)
94#if defined(SRC1_IN_PLACE)
95    out_addr            = in1_addr;
96#else  //defined(SRC1_IN_PLACE)
97    out_addr = in2_addr;
98#endif //defined(SRC1_IN_PLACE)
99#endif // !defined(IN_PLACE)
100
101    // Load data
102    VEC_ACC_TYPE in1_data = CONVERT((VEC_DATA_TYPE(DATA_TYPE_IN1, VEC_SIZE_OUT))VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE_IN1 *)in1_addr), VEC_ACC_TYPE);
103    VEC_ACC_TYPE in2_data = CONVERT((VEC_DATA_TYPE(DATA_TYPE_IN2, VEC_SIZE_OUT))VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE_IN2 *)in2_addr), VEC_ACC_TYPE);
104    // Perform multiplication and store result
105    VEC_OUT_TYPE out_data0 = MUL_OP(in1_data, in2_data, scale, DATA_TYPE_OUT, VEC_SIZE_OUT);
106    STORE_VECTOR_SELECT(out_data, DATA_TYPE_OUT, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
107}
108#endif /* defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(ACC_DATA_TYPE) && defined(DATA_TYPE_OUT) */
109
110#if defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT) && defined(VEC_SIZE_OUT)
111
112#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE_OUT)
113#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE_OUT)
114#define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
115
116/** Performs a pixelwise multiplication with float scale of quantized inputs.
117 *
118 * @note The quantization offset of the first operand must be passed at compile time only if asymmetric using -DOFFSET_IN1, e.g. -DOFFSET_IN1=10
119 * @note The quantization offset of the second operand must be passed at compile time only if asymmetric using -DOFFSET_IN2, e.g. -DOFFSET_IN2=10
120 * @note The quantization offset of the output must be passed at compile time only if asymmetric using -DOFFSET_OUT, e.g. -DOFFSET_OUT=10
121 * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, e.g. -DSCALE_IN1=10
122 * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, e.g. -DSCALE_IN2=10
123 * @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, e.g. -DSCALE_OUT=10
124 * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
125 * @attention The data type must be passed at compile time using -DDATA_TYPE_OUT, i.e. -DDATA_TYPE_OUT=uchar
126 * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
127 *
128 * @param[in]  in1_ptr                           Pointer to the source image. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM16
129 * @param[in]  in1_stride_x                      Stride of the source image in X dimension (in bytes)
130 * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
131 * @param[in]  in1_stride_y                      Stride of the source image in Y dimension (in bytes)
132 * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
133 * @param[in]  in1_stride_z                      Stride of the source image in Y dimension (in bytes)
134 * @param[in]  in1_step_z                        in1_stride_z * number of elements along Y processed per workitem(in bytes)
135 * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source image
136 * @param[in]  in2_ptr                           Pointer to the source image. Supported data types: same as @p in1_ptr
137 * @param[in]  in2_stride_x                      Stride of the source image in X dimension (in bytes)
138 * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
139 * @param[in]  in2_stride_y                      Stride of the source image in Y dimension (in bytes)
140 * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
141 * @param[in]  in2_stride_z                      Stride of the source image in Y dimension (in bytes)
142 * @param[in]  in2_step_z                        in2_stride_z * number of elements along Y processed per workitem(in bytes)
143 * @param[in]  in2_offset_first_element_in_bytes The offset of the first element in the source image
144 * @param[out] out_ptr                           Pointer to the destination image. Supported data types: same as @p in1_ptr
145 * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
146 * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
147 * @param[in]  out_stride_y                      Stride of the destination image in Y dimension (in bytes)
148 * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
149 * @param[in]  out_stride_z                      Stride of the destination image in Y dimension (in bytes)
150 * @param[in]  out_step_z                        out_stride_z * number of elements along Y processed per workitem(in bytes)
151 * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination image
152 * @param[in]  scale                             Float scaling factor. Supported data types: F32
153 */
154__kernel void pixelwise_mul_quantized(
155    TENSOR3D_DECLARATION(in1),
156    TENSOR3D_DECLARATION(in2),
157#if !defined(IN_PLACE)
158    TENSOR3D_DECLARATION(out),
159#endif // !defined(IN_PLACE)
160    const float scale)
161{
162    size_t x = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
163    size_t y = get_global_id(1);
164    size_t z = get_global_id(2);
165
166    __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + x * in1_stride_x + y * in1_stride_y + z * in1_stride_z;
167    __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + x * in2_stride_x + y * in2_stride_y + z * in2_stride_z;
168    __global        uchar *
169#if !defined(IN_PLACE)
170    out_addr = out_ptr + out_offset_first_element_in_bytes + x * out_stride_x + y * out_stride_y + z * out_stride_z;
171#else // !defined(IN_PLACE)
172#if defined(SRC1_IN_PLACE)
173    out_addr            = in1_addr;
174#else  //defined(SRC1_IN_PLACE)
175    out_addr = in2_addr;
176#endif //defined(SRC1_IN_PLACE)
177#endif // !defined(IN_PLACE)
178
179    // Load data
180    VEC_INT in_a = CONVERT((VEC_TYPE)(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE_OUT *)in1_addr)), VEC_INT);
181    VEC_INT in_b = CONVERT((VEC_TYPE)(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE_OUT *)in2_addr)), VEC_INT);
182
183    // Dequantize
184#if defined(OFFSET_IN1)
185    in_a -= (VEC_INT)((int)OFFSET_IN1);
186#endif // defined(OFFSET_IN1)
187#if defined(OFFSET_IN2)
188    in_b -= (VEC_INT)((int)OFFSET_IN2);
189#endif // defined(OFFSET_IN2)
190    const VEC_FLOAT in1f32 = CONVERT(in_a, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN1);
191    const VEC_FLOAT in2f32 = CONVERT(in_b, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN2);
192
193#if defined(OFFSET_OUT)
194    const VEC_FLOAT qresf32 = (in1f32 * in2f32 * scale) / ((VEC_FLOAT)(float)SCALE_OUT) + ((VEC_FLOAT)((float)OFFSET_OUT));
195#else  // defined(OFFSET_OUT)
196    const VEC_FLOAT qresf32 = (in1f32 * in2f32 * scale) / ((VEC_FLOAT)(float)SCALE_OUT);
197#endif // defined(OFFSET_OUT)
198    const VEC_TYPE res0 = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE);
199
200    // Store result
201    STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
202}
203#endif /* defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT) && defined(VEC_SIZE_OUT) */
204