1/* 2 * Copyright (c) 2018-2021 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#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)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) 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, 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, i.e. -DDATA_TYPE=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)( 88 TENSOR3D_DECLARATION(in1), 89 TENSOR3D_DECLARATION(in2) 90#if !defined(IN_PLACE) 91 , 92 TENSOR3D_DECLARATION(out) 93#endif // !defined(IN_PLACE) 94) 95{ 96#if VEC_SIZE_IN1 == 1 97 uint in1_x_offs = 0; 98#else // VEC_SIZE_IN1 == 1 99 uint in1_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN1 - (VEC_SIZE_IN1 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN1), 0); 100#endif // VEC_SIZE_IN1 == 1 101#if VEC_SIZE_IN2 == 1 102 uint in2_x_offs = 0; 103#else // VEC_SIZE_IN2 == 1 104 uint in2_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN2 - (VEC_SIZE_IN2 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN2), 0); 105#endif // VEC_SIZE_IN2 == 1 106#if !defined(IN_PLACE) 107 uint out_x_offs = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0); 108#endif // !defined(IN_PLACE) 109 110 // Get pixels pointer 111 __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + in1_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * in1_step_y + get_global_id(2) * in1_step_z; 112 __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + in2_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * in2_step_y + get_global_id(2) * in2_step_z; 113 __global uchar * 114#if !defined(IN_PLACE) 115 out_addr = out_ptr + out_offset_first_element_in_bytes + out_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * out_step_y + get_global_id(2) * out_step_z; 116#else // !defined(IN_PLACE) 117#if defined(SRC1_IN_PLACE) 118 out_addr = in1_addr; 119#else //defined(SRC1_IN_PLACE) 120 out_addr = in2_addr; 121#endif //defined(SRC1_IN_PLACE) 122#endif // !defined(IN_PLACE) 123 124 VEC_INT in_a = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE *)in1_addr)), VEC_INT); 125 VEC_INT in_b = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE *)in2_addr)), VEC_INT); 126 127 in_a = SUB(in_a, (VEC_INT)((int)OFFSET_IN1)); 128 in_b = SUB(in_b, (VEC_INT)((int)OFFSET_IN2)); 129 130 const VEC_FLOAT in1f32 = CONVERT(in_a, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN1); 131 const VEC_FLOAT in2f32 = CONVERT(in_b, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN2); 132 const VEC_FLOAT qresf32 = OP(in1f32, in2f32) / ((VEC_FLOAT)(float)SCALE_OUT) + ((VEC_FLOAT)((float)OFFSET_OUT)); 133 const VEC_TYPE res0 = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE); 134 135 // Store result 136 STORE_VECTOR_SELECT(res, DATA_TYPE, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) 137} 138#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) */ 139