| giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 1 | /* |
| giuros01 | 1e6e1b8 | 2019-05-14 16:12:53 +0100 | [diff] [blame] | 2 | * Copyright (c) 2018-2019 ARM Limited. |
| giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 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) |
| giuros01 | 1e6e1b8 | 2019-05-14 16:12:53 +0100 | [diff] [blame] | 31 | #define PRELU(x, y) (select(y * x, x, x > (DATA_TYPE_OUT)0)) |
| giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 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) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) |
| 41 | |
| 42 | #define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE) |
| 43 | #define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE) |
| 44 | #define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE) |
| 45 | |
| 46 | /** This function executes an element-wise operation among two tensors. |
| 47 | * |
| 48 | * @attention The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10 |
| 49 | * @attention The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10 |
| 50 | * @attention The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, i.e. -DOFFSET_OUT=10 |
| 51 | * @attention The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10 |
| 52 | * @attention The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10 |
| 53 | * @attention The quantization scale of the output must be passed at compile time using -DSCALE_OUT, i.e. -DSCALE_OUT=10 |
| 54 | * @attention To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used. |
| 55 | * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 |
| 56 | * @attention The element-wise operation to be executed has to be passed at compile time using -DOP (e.g., -DOP=ADD) |
| 57 | * |
| 58 | * @param[in] in1_ptr Pointer to the source tensor. Supported data types: QASYMM8 |
| 59 | * @param[in] in1_stride_x Stride of the source tensor in X dimension (in bytes) |
| 60 | * @param[in] in1_step_x in1_stride_x * number of elements along X processed per workitem(in bytes) |
| 61 | * @param[in] in1_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 62 | * @param[in] in1_step_y in1_stride_y * number of elements along Y processed per workitem(in bytes) |
| 63 | * @param[in] in1_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 64 | * @param[in] in1_step_z in1_stride_z * number of elements along Z processed per workitem(in bytes) |
| 65 | * @param[in] in1_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 66 | * @param[in] in2_ptr Pointer to the source tensor. Supported data types: same as @p in1_ptr |
| 67 | * @param[in] in2_stride_x Stride of the source tensor in X dimension (in bytes) |
| 68 | * @param[in] in2_step_x in2_stride_x * number of elements along X processed per workitem(in bytes) |
| 69 | * @param[in] in2_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 70 | * @param[in] in2_step_y in2_stride_y * number of elements along Y processed per workitem(in bytes) |
| 71 | * @param[in] in2_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 72 | * @param[in] in2_step_z in2_stride_z * number of elements along Z processed per workitem(in bytes) |
| 73 | * @param[in] in2_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 74 | * @param[out] out_ptr Pointer to the destination tensor. Supported data types: same as @p in1_ptr |
| 75 | * @param[in] out_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 76 | * @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes) |
| 77 | * @param[in] out_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 78 | * @param[in] out_step_y out_stride_y * number of elements along Y processed per workitem(in bytes) |
| 79 | * @param[in] out_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 80 | * @param[in] out_step_z out_stride_z * number of elements along Z processed per workitem(in bytes) |
| 81 | * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 82 | */ |
| 83 | __kernel void OP_FUN_NAME(OP)( |
| 84 | TENSOR3D_DECLARATION(in1), |
| 85 | TENSOR3D_DECLARATION(in2), |
| 86 | TENSOR3D_DECLARATION(out)) |
| 87 | { |
| 88 | // Get pixels pointer |
| 89 | Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1); |
| 90 | Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2); |
| 91 | Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out); |
| 92 | |
| 93 | VEC_INT in_a = CONVERT(VLOAD(VEC_SIZE)(0, (__global uchar *)in1.ptr), VEC_INT); |
| 94 | VEC_INT in_b = CONVERT(VLOAD(VEC_SIZE)(0, (__global uchar *)in2.ptr), VEC_INT); |
| 95 | |
| 96 | in_a = SUB(in_a, (VEC_INT)((int)OFFSET_IN1)); |
| 97 | in_b = SUB(in_b, (VEC_INT)((int)OFFSET_IN2)); |
| 98 | |
| 99 | const VEC_FLOAT in1f32 = CONVERT(in_a, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN1); |
| 100 | const VEC_FLOAT in2f32 = CONVERT(in_b, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN2); |
| 101 | const VEC_FLOAT qresf32 = OP(in1f32, in2f32) / ((VEC_FLOAT)(float)SCALE_OUT) + ((VEC_FLOAT)((float)OFFSET_OUT)); |
| 102 | const VEC_UCHAR res = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_UCHAR); |
| 103 | |
| 104 | // Store result |
| 105 | VSTORE(VEC_SIZE) |
| 106 | (res, 0, (__global uchar *)out.ptr); |
| 107 | } |
| 108 | #endif /* defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) */ |