Michele Di Giorgio | 56dd726 | 2017-07-27 09:53:49 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 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(WIDTH) && defined(HEIGHT) && defined(DEPTH) |
| 27 | /** This function identifies the min and maximum value of an input 3D tensor. |
| 28 | * |
| 29 | * @note The width, height and depth of the input tensor must be provided at compile time using -DWIDTH, -DHEIGHT and -DDEPTH (e.g. -DWIDTH=320, -DHEIGHT=240, -DDEPTH=3) |
| 30 | * |
| 31 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 32 | * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) |
| 33 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 34 | * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) |
| 35 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 36 | * @param[in] src_stride_z Stride of the source image in Z dimension (in bytes) |
| 37 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 38 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image |
| 39 | * @param[in] dst_ptr Pointer to the min/max vector. Minimum value in position 0, maximum value in position 1. Supported data types: F32. |
| 40 | * @param[in] dst_stride_x Stride of the min/max vector in X dimension (in bytes) |
| 41 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 42 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the min/max vector |
| 43 | */ |
| 44 | __kernel void minmax_layer( |
| 45 | TENSOR3D_DECLARATION(src), |
| 46 | VECTOR_DECLARATION(dst)) |
| 47 | { |
| 48 | Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); |
| 49 | Vector dst = CONVERT_TO_VECTOR_STRUCT(dst); |
| 50 | |
| 51 | float4 min_value = (float4)FLT_MAX; |
| 52 | float4 max_value = (float4) - FLT_MAX; |
| 53 | float2 min_max_value = (float2)(FLT_MAX, -FLT_MAX); |
| 54 | |
| 55 | for(int z = 0; z < DEPTH; ++z) |
| 56 | { |
| 57 | for(int y = 0; y < HEIGHT; ++y) |
| 58 | { |
| 59 | int x = 0; |
| 60 | __global float *src_addr = (__global float *)(src.ptr + y * src_stride_y + z * src_stride_z); |
| 61 | |
| 62 | for(; x <= (int)(WIDTH - 8); x += 8) |
| 63 | { |
| 64 | float8 value = *(src_addr + x); |
| 65 | |
| 66 | min_value = select(value.s0123, min_value, min_value < value.s0123); |
| 67 | min_value = select(value.s4567, min_value, min_value < value.s4567); |
| 68 | |
| 69 | max_value = select(value.s0123, max_value, max_value > value.s0123); |
| 70 | max_value = select(value.s4567, max_value, max_value > value.s4567); |
| 71 | } |
| 72 | |
| 73 | for(; x < WIDTH; ++x) |
| 74 | { |
| 75 | float value = *(src_addr + x); |
| 76 | |
| 77 | min_max_value.s0 = min(min_max_value.s0, value); |
| 78 | min_max_value.s1 = max(min_max_value.s1, value); |
| 79 | } |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | // Perform min/max reduction |
| 84 | min_value.s01 = min(min_value.s01, min_value.s23); |
| 85 | min_value.s0 = min(min_value.s0, min_value.s1); |
| 86 | max_value.s01 = max(max_value.s01, max_value.s23); |
| 87 | max_value.s0 = max(max_value.s0, max_value.s1); |
| 88 | |
| 89 | min_max_value.s0 = min(min_max_value.s0, min_value.s0); |
| 90 | min_max_value.s1 = max(min_max_value.s1, max_value.s0); |
| 91 | |
| 92 | if(min_max_value.s0 == min_max_value.s1) |
| 93 | { |
| 94 | min_max_value.s0 = 0.0f; |
| 95 | min_max_value.s1 = 1.0f; |
| 96 | } |
| 97 | |
| 98 | // Store min and max |
| 99 | vstore2(min_max_value, 0, (__global float *)dst.ptr); |
| 100 | } |
| 101 | #endif // defined(WIDTH) && defined(HEIGHT) && defined(DEPTH) |