Compute Library
 23.11
roi_pooling_layer.cl
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2021 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 #include "helpers_asymm.h"
26 
27 #if DATA_SIZE == 32
28 #define VEC_SIZE 4
29 #define VEC_MAX vec4_max
30 #elif DATA_SIZE == 16
31 #define VEC_SIZE 8
32 #define VEC_MAX vec8_max
33 #elif DATA_SIZE == 8
34 #define VEC_SIZE 16
35 #define VEC_MAX vec16_max
36 #else /* DATA_SIZE not equals 8, 16, 32 */
37 #error "Unsupported data size"
38 #endif /* DATA_SIZE == 32 */
39 
40 // Define whether to use max (Quantized datatype) or fmax (Float) functions
41 #if defined(OFFSET_OUT) && defined(SCALE_OUT)
42 #define MAX(x, y) max(x, y)
43 #else // !(defined(OFFSET_OUT) && defined(SCALE_OUT)
44 #define MAX(x, y) fmax(x, y)
45 #endif // defined(OFFSET_OUT) && defined(SCALE_OUT)
46 
47 inline DATA_TYPE vec4_max(VEC_DATA_TYPE(DATA_TYPE, 4) vec)
48 {
49  VEC_DATA_TYPE(DATA_TYPE, 2)
50  temp = MAX(vec.lo, vec.hi);
51  return MAX(temp.x, temp.y);
52 }
53 
54 inline DATA_TYPE vec8_max(VEC_DATA_TYPE(DATA_TYPE, 8) vec)
55 {
56  VEC_DATA_TYPE(DATA_TYPE, 4)
57  temp = MAX(vec.lo, vec.hi);
58  return vec4_max(temp);
59 }
60 
61 inline DATA_TYPE vec16_max(VEC_DATA_TYPE(DATA_TYPE, 16) vec)
62 {
63  VEC_DATA_TYPE(DATA_TYPE, 8)
64  temp = MAX(vec.lo, vec.hi);
65  return vec8_max(temp);
66 }
67 
68 /** Performs a roi pooling on a single output pixel.
69  *
70  * @param[in] input Pointer to input Tensor3D struct.
71  * @param[in] region_start_x Start x index projected onto the input tensor.
72  * @param[in] region_end_x End x index projected onto the input tensor.
73  * @param[in] region_start_y Start y index projected onto the input tensor.
74  * @param[in] region_end_y End y index projected onto the input tensor.
75  * @param[in] pz z index of the input tensor.
76  *
77  * @return A max pooled value from the region specified in the input tensor.
78  */
79 inline DATA_TYPE roi_pool_1x1(const Tensor3D *input, int region_start_x, int region_end_x, int region_start_y, int region_end_y, int pz)
80 {
81  // Iterate through the pooling region
82  if((region_end_x <= region_start_x) || (region_end_y <= region_start_y))
83  {
84  return (DATA_TYPE)0;
85  }
86  else
87  {
88  int num_iter = (int)((region_end_x - region_start_x) / VEC_SIZE);
89  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
90  curr_max = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(MIN_VALUE);
91 
92  for(int j = region_start_y; j < region_end_y; ++j)
93  {
94  int i = region_start_x;
95  for(; i < region_start_x + num_iter * VEC_SIZE; i += VEC_SIZE)
96  {
97  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
98  val = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(input, i, j, pz));
99  curr_max = MAX(val, curr_max);
100  }
101  for(; i < region_end_x; ++i)
102  {
103  DATA_TYPE val = *(__global DATA_TYPE *)tensor3D_offset(input, i, j, pz);
104  curr_max = MAX(curr_max, val);
105  }
106  }
107 
108  const DATA_TYPE temp = (DATA_TYPE)VEC_MAX(curr_max);
109 
110 #if defined(OFFSET_OUT) && defined(SCALE_OUT)
111  return QUANTIZE(temp, OFFSET_OUT, SCALE_OUT, DATA_TYPE, 1);
112 #endif /* if quantized, requantize and return */
113 
114  return temp;
115  }
116 }
117 
118 /** Performs a roi pooling function.
119  *
120  * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32, QASYMM8;
121  * @note Datasize must be passed using -DDATA_SIZE e.g. -DDATA_SIZE=32;
122  * @note Input dimensions must be passed using -DMAX_DIM_X, -DMAX_DIM_Y and -DMAX_DIM_Z;
123  * @note Pooled region dimensions must be passed using -DPOOLED_DIM_X and -DPOOLED_DIM_Y;
124  * @note Spatial scale must be passed using -DSPATIAL_SCALE;
125  *
126  * @param[in] input_ptr Pointer to the source image. Supported data types: F16, F32, QASYMM8
127  * @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
128  * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
129  * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
130  * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
131  * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
132  * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
133  * @param[in] input_offset_first_element_in_bytes The offset of the first element in the pooled region of the source image as specifed by ROI
134  * @param[in] rois_ptr Pointer to the ROIs tensor. Layout: { batch_index, x1, y1, x2, y2 }. Supported data types: same as @p input_ptr
135  * @param[in] rois_stride_x Stride of the ROIs tensor in X dimension (in bytes)
136  * @param[in] rois_step_x Step of the ROIs tensor in X dimension (in bytes)
137  * @param[in] rois_stride_y Stride of the ROIs tensor in Y dimension (in bytes)
138  * @param[in] rois_step_y Step of the ROIs tensor in Y dimension (in bytes)
139  * @param[in] rois_offset_first_element_in_bytes The offset of the first element in the ROIs tensor
140  * @param[out] output_ptr Pointer to the destination image. Supported data types: same as input
141  * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes)
142  * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
143  * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes)
144  * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
145  * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
146  * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
147  * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image
148  * @param[in] input_stride_w Stride of the source image in W dimension (in bytes)
149  * @param[in] output_stride_w Stride of the destination image in W dimension (in bytes)
150  */
151 __kernel void roi_pooling_layer(
153  IMAGE_DECLARATION(rois),
154  TENSOR3D_DECLARATION(output),
155  unsigned int input_stride_w, unsigned int output_stride_w)
156 {
157  // Get pixels pointer
161 
162  const int px = get_global_id(0);
163  const int py = get_global_id(1);
164  const int pw = get_global_id(2);
165 
166  // Load roi parameters
167  // roi is laid out as follows { batch_index, x1, y1, x2, y2 }
168  const ushort roi_batch = (ushort) * ((__global ushort *)offset(&rois, 0, pw));
169  const VEC_DATA_TYPE(ushort, 4)
170  roi = vload4(0, (__global ushort *)offset(&rois, 1, pw));
171  const int2 roi_anchor = convert_int2_sat(round(convert_float2(roi.s01) * (float)SPATIAL_SCALE));
172  const int2 roi_dims = convert_int2_sat(fmax(round(convert_float2(roi.s23 - roi.s01) * (float)SPATIAL_SCALE), 1.f));
173 
174  // Calculate pooled region start and end
175  const float2 spatial_indx = (float2)(px, py);
176  const float2 pooled_dims = (float2)(POOLED_DIM_X, POOLED_DIM_Y);
177  const int2 max_spatial_dims = (int2)(MAX_DIM_X, MAX_DIM_Y);
178  int2 region_start = convert_int2_sat(floor(spatial_indx / pooled_dims * convert_float2(roi_dims))) + roi_anchor;
179  int2 region_end = convert_int2_sat(floor((spatial_indx + 1) / pooled_dims * convert_float2(roi_dims))) + roi_anchor;
180 
181  region_start = clamp(region_start, 0, max_spatial_dims);
182  region_end = clamp(region_end, 0, max_spatial_dims);
183 
184  // Move input and output pointer across the fourth dimension
185  input.ptr += roi_batch * input_stride_w;
186  output.ptr += pw * output_stride_w;
187 
188  for(int pz = 0; pz < MAX_DIM_Z; ++pz)
189  {
190  *(__global DATA_TYPE *)tensor3D_offset(&output, px, py, pz) = (__global DATA_TYPE)roi_pool_1x1(&input,
191  region_start.x,
192  region_end.x,
193  region_start.y,
194  region_end.y, pz);
195  }
196 }
CONVERT_TO_IMAGE_STRUCT_NO_STEP
#define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name)
Definition: helpers.h:876
vec4_max
DATA_TYPE vec4_max(DATA_TYPE4 vec)
Definition: roi_pooling_layer.cl:47
MAX
#define MAX(x, y)
Definition: roi_pooling_layer.cl:44
VEC_SIZE
#define VEC_SIZE
Definition: qlstm_layer_normalization.cl:54
VEC_DATA_TYPE
#define VEC_DATA_TYPE(type, size)
Definition: helpers.h:764
Image
Structure to hold Image information.
Definition: helpers.h:923
QUANTIZE
#define QUANTIZE(input, offset, scale, type, size)
Definition: helpers_asymm.h:398
Tensor3D
Structure to hold 3D tensor information.
Definition: helpers.h:932
arm_compute::utility::clamp
DataType clamp(const DataType &n, const DataType &lower=std::numeric_limits< RangeType >::lowest(), const DataType &upper=std::numeric_limits< RangeType >::max())
Performs clamping among a lower and upper value.
Definition: Utility.h:102
VLOAD
#define VLOAD(size)
Definition: helpers.h:204
roi_pool_1x1
DATA_TYPE roi_pool_1x1(const Tensor3D *input, int region_start_x, int region_end_x, int region_start_y, int region_end_y, int pz)
Performs a roi pooling on a single output pixel.
Definition: roi_pooling_layer.cl:79
TENSOR3D_DECLARATION
#define TENSOR3D_DECLARATION(name)
Definition: helpers.h:852
helpers_asymm.h
arm_compute::round
int round(float x, RoundingPolicy rounding_policy)
Return a rounded value of x.
Definition: Rounding.cpp:36
offset
__global uchar * offset(const Image *img, int x, int y)
Get the pointer position of a Image.
Definition: helpers.h:1128
MIN_VALUE
#define MIN_VALUE
Definition: softmax_layer.cl:34
Tensor3D::ptr
__global uchar * ptr
Pointer to the starting postion of the buffer.
Definition: helpers.h:934
roi_pooling_layer
__kernel void roi_pooling_layer(__global uchar *input_ptr, uint input_stride_x, uint input_step_x, uint input_stride_y, uint input_step_y, uint input_stride_z, uint input_step_z, uint input_offset_first_element_in_bytes, __global uchar *rois_ptr, uint rois_stride_x, uint rois_step_x, uint rois_stride_y, uint rois_step_y, uint rois_offset_first_element_in_bytes, __global uchar *output_ptr, uint output_stride_x, uint output_step_x, uint output_stride_y, uint output_step_y, uint output_stride_z, uint output_step_z, uint output_offset_first_element_in_bytes, unsigned int input_stride_w, unsigned int output_stride_w)
Performs a roi pooling function.
Definition: roi_pooling_layer.cl:151
IMAGE_DECLARATION
#define IMAGE_DECLARATION(name)
Definition: helpers.h:848
CONVERT_TO_TENSOR3D_STRUCT_NO_STEP
#define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name)
Definition: helpers.h:897
vec16_max
DATA_TYPE vec16_max(DATA_TYPE16 vec)
Definition: roi_pooling_layer.cl:61
tensor3D_offset
const __global uchar * tensor3D_offset(const Tensor3D *tensor, int x, int y, int z)
Get the pointer position of a Tensor3D.
Definition: helpers.h:1140
arm_compute::test::validation::input
auto input
Definition: LSTMLayerQuantized.cpp:486
vec8_max
DATA_TYPE vec8_max(DATA_TYPE8 vec)
Definition: roi_pooling_layer.cl:54
helpers.h