Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [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 | |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 26 | #define CONVERT_RTE(x, type) (convert_##type##_rte((x))) |
| 27 | #define CONVERT_RTE_VEC_STR(x, type, size) (convert_##type##size##_rte((x))) |
| 28 | #define CONVERT_RTE_VEC(x, type, size) CONVERT_RTE_VEC_STR(x, type, size) |
| 29 | |
Michele Di Giorgio | d87a7b2 | 2019-09-10 10:42:27 +0100 | [diff] [blame] | 30 | #if defined(VEC_SIZE) && defined(DATA_TYPE_IN) && defined(DATA_TYPE_OUT) && defined(SCALE) && defined(OFFSET) && defined(MIN_QUANT_VAL) && defined(MAX_QUANT_VAL) |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 31 | |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 32 | /** This performs the quantization of floating point inputs or 8-bit quantized integers to 8-bit integers. |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 33 | * |
Michele Di Giorgio | d87a7b2 | 2019-09-10 10:42:27 +0100 | [diff] [blame] | 34 | * @note Input data type should be given as a preprocessor argument using -DDATA_TYPE_IN=type. e.g. -DDATA_TYPE=short |
| 35 | * @note Output data type should be given as a preprocessor argument using -DDATA_TYPE_OUT=type. e.g. -DDATA_TYPE=short |
| 36 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 |
| 37 | * @note Quantization scale should be given as a preprocessor argument using -DSCALE=scale. e.g. -DSCALE=0.125 |
| 38 | * @note Quantization offset should be given as a preprocessor argument using -DOFFSET=offset. e.g. -DOFFSET=125 |
| 39 | * @note Minimum value for quantized type should be given as a preprocessor argument using -DMIN_QUANT_VAL=value. e.g. -DMIN_QUANT_VAL=0 |
| 40 | * @note Maximum value for quantized type should be given as a preprocessor argument using -DMAX_QUANT_VAL=value. e.g. -DMAXIN_QUANT_VAL=255 |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 41 | * @note If the input data type if a floating point (F16 or F32) the preprocessor argument should be give as -DIS_FLOAT |
Michele Di Giorgio | d87a7b2 | 2019-09-10 10:42:27 +0100 | [diff] [blame] | 42 | * |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 43 | * @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/F16/F32 |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 44 | * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) |
| 45 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 46 | * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 47 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 48 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 49 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 50 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 51 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: QASYMM8/QASYMM8_SIGNED |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 52 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 53 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 54 | * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 55 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 56 | * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 57 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 58 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 59 | */ |
| 60 | __kernel void quantization_layer( |
| 61 | TENSOR3D_DECLARATION(input), |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 62 | TENSOR3D_DECLARATION(output)) |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 63 | { |
| 64 | // Get pixels pointer |
| 65 | Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 66 | Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); |
| 67 | |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 68 | #if defined(VEC_SIZE) && defined(LAST_ACCESSED_X) |
| 69 | // Check if access on width gets out of bounds |
| 70 | // If it does shift access vector to access elements within bounds |
| 71 | const int xi = (int)(get_global_id(0) * VEC_SIZE); |
| 72 | input.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * input_stride_x; |
| 73 | output.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * output_stride_x; |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 74 | |
| 75 | // Load data |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 76 | #if defined(IS_FLOAT) |
| 77 | // Load data |
Michele Di Giorgio | d87a7b2 | 2019-09-10 10:42:27 +0100 | [diff] [blame] | 78 | VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 79 | val_float = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)input.ptr); |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 80 | |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 81 | // Create scale and offset vectors |
Michele Di Giorgio | d87a7b2 | 2019-09-10 10:42:27 +0100 | [diff] [blame] | 82 | const VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) vscale = SCALE; |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 83 | const VEC_DATA_TYPE(int, VEC_SIZE) voffset = OFFSET; |
| 84 | #else // defined(IS_FLOAT) |
| 85 | // Load data |
| 86 | VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) |
| 87 | val = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)input.ptr); |
| 88 | |
| 89 | const VEC_DATA_TYPE(float, VEC_SIZE) |
| 90 | val_float = CONVERT(val, VEC_DATA_TYPE(float, VEC_SIZE)); |
| 91 | |
| 92 | // Create scale and offset vectors |
| 93 | const VEC_DATA_TYPE(float, VEC_SIZE) vscale = SCALE; |
| 94 | const VEC_DATA_TYPE(int, VEC_SIZE) voffset = OFFSET; |
| 95 | #endif // defined(IS_FLOAT) |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 96 | |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 97 | // Quantize |
| 98 | VEC_DATA_TYPE(int, VEC_SIZE) |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 99 | res = CLAMP(CONVERT_RTE_VEC(val_float / vscale, int, VEC_SIZE) + voffset, MIN_QUANT_VAL, MAX_QUANT_VAL); |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 100 | |
Michele Di Giorgio | d87a7b2 | 2019-09-10 10:42:27 +0100 | [diff] [blame] | 101 | // Store result |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 102 | VSTORE(VEC_SIZE) |
Michele Di Giorgio | d87a7b2 | 2019-09-10 10:42:27 +0100 | [diff] [blame] | 103 | (CONVERT(res, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)), 0, (__global DATA_TYPE_OUT *)output.ptr); |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 104 | #else //!defined(VEC_SIZE) || !defined(LAST_ACCESSED_X) |
Michele Di Giorgio | d87a7b2 | 2019-09-10 10:42:27 +0100 | [diff] [blame] | 105 | *((__global DATA_TYPE_OUT *)(output.ptr)) = (DATA_TYPE_OUT)CLAMP(CONVERT_RTE(((float) * (__global DATA_TYPE_IN *)input.ptr) / ((float)SCALE), int) + (int)OFFSET, MIN_QUANT_VAL, MAX_QUANT_VAL); |
Usama Arif | e03802e | 2019-03-11 12:20:20 +0000 | [diff] [blame] | 106 | #endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X) |
Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 107 | } |
Manuel Bottini | 2f60221 | 2020-01-30 17:30:32 +0000 | [diff] [blame] | 108 | #endif // defined(VEC_SIZE) && defined(DATA_TYPE_IN) && defined(DATA_TYPE_OUT) && defined(SCALE) && defined(OFFSET) && defined(MIN_QUANT_VAL) && defined(MAX_QUANT_VAL) |