Michele Di Giorgio | d63dfa2 | 2018-09-12 10:18:54 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018 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(DATA_TYPE) && defined(VEC_SIZE) && defined(OFFSET) && defined(SCALE) |
| 27 | |
| 28 | #define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 29 | #define OFFSET_FLT ((float)OFFSET) |
| 30 | #define SCALE_FLT ((float)SCALE) |
| 31 | |
| 32 | #if defined(NUM_CHANNELS) |
| 33 | |
| 34 | /** Apply normalize_planar_yuv layer on tensors with NCHW data layout. |
| 35 | * |
| 36 | * @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
| 37 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8 |
| 38 | * @note The depth of the input tensor should be given as a preprocessor argument using -DNUM_CHANNELS e.g. -DNUM_CHANNELS=8 |
| 39 | * @note The quantization offset should be given as a preprocessor argument using -DOFFSET e.g. -DOFFSET=8 |
| 40 | * @note The quantization scale should be given as a preprocessor argument using -DSCALE e.g. -DSCALE=8 |
| 41 | * |
| 42 | * @param[in] src_ptr Pointer to the first source tensor. Supported data types: QASYMM8 |
| 43 | * @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes) |
| 44 | * @param[in] src_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 45 | * @param[in] src_stride_y Stride of the first source tensor in Y dimension (in bytes) |
| 46 | * @param[in] src_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 47 | * @param[in] src_stride_z Stride of the first source tensor in Z dimension (in bytes) |
| 48 | * @param[in] src_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 49 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 50 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 51 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 52 | * @param[in] dst_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 53 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 54 | * @param[in] dst_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 55 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 56 | * @param[in] dst_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 57 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 58 | * @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p src_ptr |
| 59 | * @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes) |
| 60 | * @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes) |
| 61 | * @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor |
| 62 | * @param[in] std_ptr Pointer to the std tensor. Supported data types: same as @p src_ptr |
| 63 | * @param[in] std_stride_x Stride of the std tensor in X dimension (in bytes) |
| 64 | * @param[in] std_step_x std_stride_x * number of elements along X processed per workitem(in bytes) |
| 65 | * @param[in] std_offset_first_element_in_bytes The offset of the first element in the var source tensor |
| 66 | */ |
| 67 | __kernel void normalize_planar_yuv_layer_q8_nchw(TENSOR3D_DECLARATION(src), |
| 68 | TENSOR3D_DECLARATION(dst), |
| 69 | VECTOR_DECLARATION(mean), |
| 70 | VECTOR_DECLARATION(std)) |
| 71 | { |
| 72 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 73 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| 74 | Vector mean = CONVERT_TO_VECTOR_STRUCT(mean); |
| 75 | Vector std = CONVERT_TO_VECTOR_STRUCT(std); |
| 76 | |
| 77 | const uint current_slice = get_global_id(2) % NUM_CHANNELS; |
| 78 | |
| 79 | float16 curr_mean_flt = (float16)(*((__global DATA_TYPE *)(mean.ptr + current_slice * sizeof(DATA_TYPE)))); |
| 80 | curr_mean_flt = round(curr_mean_flt - OFFSET_FLT) * SCALE_FLT; |
| 81 | |
| 82 | float16 curr_std_flt = (float16)(*((__global DATA_TYPE *)(std.ptr + current_slice * sizeof(DATA_TYPE)))); |
| 83 | curr_std_flt = round(curr_std_flt - OFFSET_FLT) * SCALE_FLT; |
| 84 | |
| 85 | float16 data_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr), float16); |
| 86 | data_flt = round(data_flt - OFFSET_FLT) * SCALE_FLT; |
| 87 | |
| 88 | // Perform normalization |
| 89 | float16 res_flt = (data_flt - curr_mean_flt) / curr_std_flt; |
| 90 | |
| 91 | const TYPE res_u8 = CONVERT_SAT(round(res_flt / SCALE_FLT) + OFFSET_FLT, TYPE); |
| 92 | VSTORE(VEC_SIZE) |
| 93 | (res_u8, 0, (__global DATA_TYPE *)dst.ptr); |
| 94 | } |
| 95 | |
| 96 | #endif // defined(NUM_CHANNELS) |
| 97 | |
| 98 | /** Apply normalize_planar_yuv layer on tensors with NHWC data layout. |
| 99 | * |
| 100 | * @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
| 101 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8 |
| 102 | * @note The quantization offset should be given as a preprocessor argument using -DOFFSET e.g. -DOFFSET=8 |
| 103 | * @note The quantization scale should be given as a preprocessor argument using -DSCALE e.g. -DSCALE=8 |
| 104 | * |
| 105 | * @param[in] src_ptr Pointer to the first source tensor. Supported data types: QASYMM8 |
| 106 | * @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes) |
| 107 | * @param[in] src_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 108 | * @param[in] src_stride_y Stride of the first source tensor in Y dimension (in bytes) |
| 109 | * @param[in] src_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 110 | * @param[in] src_stride_z Stride of the first source tensor in Z dimension (in bytes) |
| 111 | * @param[in] src_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 112 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 113 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 114 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 115 | * @param[in] dst_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 116 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 117 | * @param[in] dst_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 118 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 119 | * @param[in] dst_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 120 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 121 | * @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p src_ptr |
| 122 | * @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes) |
| 123 | * @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes) |
| 124 | * @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor |
| 125 | * @param[in] std_ptr Pointer to the std tensor. Supported data types: same as @p src_ptr |
| 126 | * @param[in] std_stride_x Stride of the std tensor in X dimension (in bytes) |
| 127 | * @param[in] std_step_x std_stride_x * number of elements along X processed per workitem(in bytes) |
| 128 | * @param[in] std_offset_first_element_in_bytes The offset of the first element in the var source tensor |
| 129 | */ |
| 130 | __kernel void normalize_planar_yuv_layer_q8_nhwc(TENSOR3D_DECLARATION(src), |
| 131 | TENSOR3D_DECLARATION(dst), |
| 132 | VECTOR_DECLARATION(mean), |
| 133 | VECTOR_DECLARATION(std)) |
| 134 | { |
| 135 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 136 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| 137 | Vector mean = CONVERT_TO_VECTOR_STRUCT(mean); |
| 138 | Vector std = CONVERT_TO_VECTOR_STRUCT(std); |
| 139 | |
| 140 | const uint current_slice = get_global_id(0); |
| 141 | |
| 142 | float16 curr_mean_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(mean.ptr + current_slice * VEC_SIZE * sizeof(DATA_TYPE))), float16); |
| 143 | curr_mean_flt = round(curr_mean_flt - OFFSET_FLT) * SCALE_FLT; |
| 144 | |
| 145 | float16 curr_std_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(std.ptr + current_slice * VEC_SIZE * sizeof(DATA_TYPE))), float16); |
| 146 | curr_std_flt = round(curr_std_flt - OFFSET_FLT) * SCALE_FLT; |
| 147 | |
| 148 | float16 data_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr), float16); |
| 149 | data_flt = round(data_flt - OFFSET_FLT) * (SCALE_FLT); |
| 150 | |
| 151 | // Perform normalization |
| 152 | float16 res_flt = (data_flt - curr_mean_flt) / curr_std_flt; |
| 153 | |
| 154 | const TYPE res_u8 = CONVERT_SAT(round(res_flt / SCALE_FLT) + OFFSET_FLT, TYPE); |
| 155 | VSTORE(VEC_SIZE) |
| 156 | (res_u8, 0, (__global DATA_TYPE *)dst.ptr); |
| 157 | } |
| 158 | #endif // defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OFFSET) && defined(SCALE) |