Michele Di Giorgio | 7217563 | 2018-05-01 16:52:00 +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(BLOCK_SIZE) && defined(NUM_GROUPS) && defined(K) |
| 27 | |
| 28 | // Check valid BLOCK_SIZES |
| 29 | #if BLOCK_SIZE != 4 && BLOCK_SIZE != 8 && BLOCK_SIZE != 16 |
| 30 | #error "Only block sizes 4, 8 and 16 are supported" |
| 31 | #endif /* BLOCK_SIZE != 4 && BLOCK_SIZE != 8 && BLOCK_SIZE != 16 */ |
| 32 | |
| 33 | #define TYPE VEC_DATA_TYPE(DATA_TYPE, BLOCK_SIZE) |
| 34 | |
| 35 | /** Perfoms channel shuffle see https://arxiv.org/pdf/1707.01083.pdf for details. |
| 36 | * |
| 37 | * @note The number of groups should be given as a preprocessor argument using -DNUM_GROUPS=num_groups. e.g. -DNUM_GROUPS=2 |
| 38 | * @note The number of channels in each group should be given as a preprocessor argument using -DK=num. e.g. -DK=1 |
| 39 | * K is equal to num_channels / num_groups. |
| 40 | * |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 41 | * @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 |
Michele Di Giorgio | 7217563 | 2018-05-01 16:52:00 +0100 | [diff] [blame] | 42 | * @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes) |
| 43 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 44 | * @param[in] src_stride_y Stride of the first source tensor in Y dimension (in bytes) |
| 45 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 46 | * @param[in] src_stride_z Stride of the first source tensor in Z dimension (in bytes) |
| 47 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 48 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 49 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 50 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 51 | * @param[in] dst_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 52 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 53 | * @param[in] dst_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 54 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 55 | * @param[in] dst_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 56 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 57 | */ |
| 58 | __kernel void channel_shuffle_nchw(TENSOR3D_DECLARATION(src), |
| 59 | TENSOR3D_DECLARATION(dst)) |
| 60 | { |
| 61 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 62 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(dst); |
| 63 | |
Michele Di Giorgio | efac7c6 | 2018-05-16 00:02:35 +0100 | [diff] [blame] | 64 | const uint curr_channel = get_global_id(2); // channel id of input |
| 65 | const uint group_id = curr_channel / K; // group id |
| 66 | const uint channel_id = curr_channel % K; // channel id within the group |
Michele Di Giorgio | 7217563 | 2018-05-01 16:52:00 +0100 | [diff] [blame] | 67 | |
| 68 | const uint x = get_global_id(0) * BLOCK_SIZE; |
| 69 | const uint y = get_global_id(1) * BLOCK_SIZE; |
Michele Di Giorgio | efac7c6 | 2018-05-16 00:02:35 +0100 | [diff] [blame] | 70 | const uint z = channel_id * NUM_GROUPS + group_id; |
Michele Di Giorgio | 7217563 | 2018-05-01 16:52:00 +0100 | [diff] [blame] | 71 | |
| 72 | // Load the NxN block |
| 73 | TYPE u0 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 0, 0)); |
| 74 | TYPE u1 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 1, 0)); |
| 75 | TYPE u2 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 2, 0)); |
| 76 | TYPE u3 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 3, 0)); |
| 77 | #if BLOCK_SIZE > 4 |
| 78 | TYPE u4 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 4, 0)); |
| 79 | TYPE u5 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 5, 0)); |
| 80 | TYPE u6 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 6, 0)); |
| 81 | TYPE u7 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 7, 0)); |
| 82 | #if BLOCK_SIZE == 16 |
| 83 | TYPE u8 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 8, 0)); |
| 84 | TYPE u9 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 9, 0)); |
| 85 | TYPE u10 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 10, 0)); |
| 86 | TYPE u11 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 11, 0)); |
| 87 | TYPE u12 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 12, 0)); |
| 88 | TYPE u13 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 13, 0)); |
| 89 | TYPE u14 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 14, 0)); |
| 90 | TYPE u15 = VLOAD(BLOCK_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, 0, 15, 0)); |
| 91 | #endif /* BLOCK_SIZE == 16 */ |
| 92 | #endif /* BLOCK_SIZE > 4 */ |
| 93 | |
| 94 | // Store blocks |
| 95 | VSTORE(BLOCK_SIZE) |
| 96 | (u0, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 0, z)); |
| 97 | VSTORE(BLOCK_SIZE) |
| 98 | (u1, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 1, z)); |
| 99 | VSTORE(BLOCK_SIZE) |
| 100 | (u2, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 2, z)); |
| 101 | VSTORE(BLOCK_SIZE) |
| 102 | (u3, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 3, z)); |
| 103 | #if BLOCK_SIZE > 4 |
| 104 | VSTORE(BLOCK_SIZE) |
| 105 | (u4, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 4, z)); |
| 106 | VSTORE(BLOCK_SIZE) |
| 107 | (u5, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 5, z)); |
| 108 | VSTORE(BLOCK_SIZE) |
| 109 | (u6, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 6, z)); |
| 110 | VSTORE(BLOCK_SIZE) |
| 111 | (u7, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 7, z)); |
| 112 | #if BLOCK_SIZE == 16 |
| 113 | VSTORE(BLOCK_SIZE) |
| 114 | (u8, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 8, z)); |
| 115 | VSTORE(BLOCK_SIZE) |
| 116 | (u9, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 9, z)); |
| 117 | VSTORE(BLOCK_SIZE) |
| 118 | (u10, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 10, z)); |
| 119 | VSTORE(BLOCK_SIZE) |
| 120 | (u11, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 11, z)); |
| 121 | VSTORE(BLOCK_SIZE) |
| 122 | (u12, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 12, z)); |
| 123 | VSTORE(BLOCK_SIZE) |
| 124 | (u13, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 13, z)); |
| 125 | VSTORE(BLOCK_SIZE) |
| 126 | (u14, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 14, z)); |
| 127 | VSTORE(BLOCK_SIZE) |
| 128 | (u15, 0, (__global DATA_TYPE *)tensor3D_offset(&dst, x, y + 15, z)); |
| 129 | #endif /* BLOCK_SIZE == 16 */ |
| 130 | #endif /* BLOCK_SIZE > 4 */ |
| 131 | } |
| 132 | #endif /* defined(DATA_TYPE) && defined(BLOCK_SIZE) && defined(NUM_GROUPS) && defined(K) */ |