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 26#if defined(DATA_TYPE) && defined(INITIAL_VALUE) 27#define VEC_TYPE(VEC_SIZE) VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) 28 29#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) 30#define VEC_FLOAT(VEC_SIZE) VEC_DATA_TYPE(float, VEC_SIZE) 31#define VEC_INT(VEC_SIZE) VEC_DATA_TYPE(int, VEC_SIZE) 32#define CONVERT_RTE(x, type) (convert_##type##_rte((x))) 33#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type) 34#define REQUANTIZE(VEC_SIZE, input, in_offset, out_offset, in_scale, out_scale, res) \ 35 { \ 36 const VEC_FLOAT(VEC_SIZE) in_f32 = (CONVERT(input, VEC_FLOAT(VEC_SIZE)) - (VEC_FLOAT(VEC_SIZE))((float)in_offset)) * (VEC_FLOAT(VEC_SIZE))((float)in_scale); \ 37 const VEC_FLOAT(VEC_SIZE) out_f32 = in_f32 / ((VEC_FLOAT(VEC_SIZE))(float)out_scale) + ((VEC_FLOAT(VEC_SIZE))((float)out_offset)); \ 38 res = CONVERT_SAT(CONVERT_DOWN(out_f32, VEC_INT(VEC_SIZE)), VEC_TYPE(VEC_SIZE)); \ 39 } 40#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */ 41 42#if defined(POOL_AVG) 43#define POOL_OP(x, y) ((x) + (y)) 44#else /* defined(POOL_AVG) */ 45#define POOL_OP(x, y) (max((x), (y))) 46#endif /* defined(POOL_AVG) */ 47 48#define DIV_OP(x, y) (x * (1.f / y)) 49 50#if defined(POOL_L2) 51#error "L2 pooling is not supported" 52#endif /* defined(POOL_L2) */ 53 54#if defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE) 55/** Performs pooling layer of size equal to MxN. This OpenCL kernel can perform the following pooling types: 56 * -# max, -DPOOL_MAX must be passed at compile time 57 * -# average, -DPOOL_AVG must be passed at compile time. If padding has to be expluded, -DEXCLUDE_PADDING should be passed at compile time 58 * 59 * @note Datatype must be passed at compile type using -DDATA_TYPE e.g. -DDATA_TYPE=uchar. Supported data types are QASYMM8/QASYMM8_SIGNED 60 * @note Accumulation data type must be passed at compile time using -DACC_DATA_TYPE e.g. -DACC_DATA_TYPE=int 61 * @note Pool size must be passed at compile time using -DPOOL_SIZE_X and -DPOOL_SIZE_Y. e.g. -DPOOL_SIZE_X=4, -DPOOL_SIZE_Y=4 62 * @note Input tensor width and height must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT 63 * @note Output tensor height, channels and batch size must be passed at compile time using -DDST_HEIGHT, -DDST_CHANNELS and -DDST_BATCH_SIZE 64 * @note Pool strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions 65 * @note Pool pads must be passed at compile time using -DPAD_X and -DPAD_Y 66 * @note Vector size must be passed at compile time using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 67 * @note Leftover vector size must be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE 68 * @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0 69 * @note If the output has be requantized, -DOFFSET_IN1, -DOFFSET_OUT, -DSCALE_IN1 and -DSCALE_OUT muste be passed at compile time 70 * 71 * @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8/QASYMM8_SIGNED 72 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) 73 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 74 * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) 75 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 76 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 77 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 78 * @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes) 79 * @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes) 80 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image 81 * @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr 82 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) 83 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) 84 * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) 85 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) 86 * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) 87 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) 88 * @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes) 89 * @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes) 90 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image 91 */ 92__kernel void pooling_layer_MxN_quantized_nhwc( 93 TENSOR4D_DECLARATION(input), 94 TENSOR4D_DECLARATION(output)) 95{ 96 // Note: If C is not multiple of VEC_SIZE, we shift back of VEC_SIZE_LEFTOVER elements to compute the leftover elements for get_global_id(0) == 0 97 // Note: If C is less than VEC_SIZE, VEC_SIZE should be SHRINKED to the closest smaller VEC_SIZE. This operation is performed on the host side 98 int offset_c = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0) * sizeof(DATA_TYPE); 99 int idx_out_w = get_global_id(1); 100#if DST_BATCH_SIZE != 1 101 // If batch size != 1, the batch size dimension is collapsed over the height dimension 102 int idx_out_h = get_global_id(2) % DST_HEIGHT; 103 int idx_out_n = get_global_id(2) / DST_HEIGHT; 104#else //DST_BATCH_SIZE != 1 105 int idx_out_h = get_global_id(2); 106 int idx_out_n = 0; 107#endif // DST_BATCH_SIZE != 1 108 109 int idx_in_w = idx_out_w * STRIDE_X - PAD_X; 110 int idx_in_h = idx_out_h * STRIDE_Y - PAD_Y; 111 112 __global unsigned char *in_base_ptr = input_ptr + input_offset_first_element_in_bytes + offset_c + idx_out_n * input_stride_w; 113 114 __global unsigned char *out_base_ptr = output_ptr + output_offset_first_element_in_bytes + offset_c + idx_out_w * output_stride_y + idx_out_h * output_stride_z + idx_out_n * output_stride_w; 115 116 int pool_x_s = max((int)0, -idx_in_w); 117 int pool_x_e = min((int)POOL_SIZE_X, (int)SRC_WIDTH - idx_in_w); 118 int pool_y_s = max((int)0, -idx_in_h); 119 int pool_y_e = min((int)POOL_SIZE_Y, (int)SRC_HEIGHT - idx_in_h); 120 121#if defined(POOL_AVG) && defined(EXCLUDE_PADDING) 122 int filter_size = 0; 123#elif defined(POOL_AVG) && !defined(EXCLUDE_PADDING) // defined(POOL_AVG) && defined(EXCLUDE_PADDING) 124 int filter_size = POOL_SIZE_X * POOL_SIZE_Y; 125#endif // defined(POOL_AVG) && !defined(EXCLUDE_PADDING) 126 127 VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE) 128 res0 = INITIAL_VALUE; 129 130 for(int y = pool_y_s; y < pool_y_e; ++y) 131 { 132 for(int x = pool_x_s; x < pool_x_e; ++x) 133 { 134 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) 135 data; 136 VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE) 137 data0; 138 139 data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + (x + idx_in_w) * input_stride_y + (y + idx_in_h) * input_stride_z)); 140 data0 = CONVERT(data, VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)); 141 142 res0 = POOL_OP(res0, data0); 143 144#if defined(POOL_AVG) && defined(EXCLUDE_PADDING) 145 filter_size++; 146#endif // defined(POOL_AVG) && defined(EXCLUDE_PADDING) 147 } 148 } 149 150#if defined(POOL_AVG) 151 res0 = (res0 + (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))(filter_size >> 1)) / filter_size; 152#endif // defined(POOL_AVG) 153 154 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) 155 out_q0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)); 156#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) 157 REQUANTIZE(VEC_SIZE, out_q0, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT, out_q0); 158#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */ 159 160 // Store result 161 STORE_VECTOR_SELECT(out_q, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, ((VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0)); 162} 163#endif // defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE) 164#endif // defined(DATA_TYPE) && defined(INITIAL_VALUE)