| /aosp_15_r20/external/ComputeLibrary/src/gpu/cl/kernels/ |
| H A D | ClIm2ColKernel.cpp | 58 unsigned int num_elems_processed_per_iteration{}; member
96 …unsigned int num_elems_processed_per_iteration, bool is_padding_required_nchw, unsigned int num_gr… in validate_and_configure_window() argument
117 win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); in validate_and_configure_window()
125 … Steps(num_elems_processed_per_iteration * conv_info.stride().first, conv_info.stride().second)); in validate_and_configure_window()
129 …eil_to_multiple(input_width + border.right, kernel_dims.width * num_elems_processed_per_iteration), in validate_and_configure_window()
164 unsigned int num_elems_processed_per_iteration = 1; in configure_opencl_kernel() local
191 num_elems_processed_per_iteration = std::min(2U, input_channel); in configure_opencl_kernel()
213 const unsigned int vec_size = num_elems_processed_per_iteration; in configure_opencl_kernel()
234 num_elems_processed_per_iteration = 4; in configure_opencl_kernel()
240 num_elems_processed_per_iteration = 1; in configure_opencl_kernel()
[all …]
|
| H A D | ClWidthConcatenate4TensorsKernel.cpp | 88 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(8, min_dimension); in configure() local
89 …gned int vec_size_leftover = dst->dimension(0) % num_elems_processed_per_iteration; in configure()
94 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
97 …t::cpp11::to_string((src1->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration)); in configure()
98 …src1->dimension(0) + src2->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration)); in configure()
99 …src2->dimension(0) + src3->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration)); in configure()
136 Window win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClWidthConcatenate2TensorsKernel.cpp | 84 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(8, min_dimension); in configure() local
85 …gned int vec_size_leftover = dst->dimension(0) % num_elems_processed_per_iteration; in configure()
90 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
93 …t::cpp11::to_string((src1->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration)); in configure()
123 Window win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClHeightConcatenateKernel.cpp | 86 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, src->dimension(0)); in configure() local
90 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
92 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
116 Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClWidthConcatenateKernel.cpp | 82 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16, src->dimension(0)); in configure() local
87 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
88 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
111 Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClDepthConcatenateKernel.cpp | 76 …const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), s… in configure() local
81 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
82 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
103 auto win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel.cpp | 97 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, src->dimension(0)); in configure() local
103 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
104 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
124 auto win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClGemmLowpQuantizeDownInt32ScaleByFloatKernel.cpp | 100 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, src->dimension(0)); in configure() local
107 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
108 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
125 Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClBatchConcatenateKernel.cpp | 76 …const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), s… in configure() local
81 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
82 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
103 auto win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClGemmLowpQuantizeDownInt32ScaleKernel.cpp | 97 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, src->dimension(0)); in configure() local
103 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
104 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
124 Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClCastKernel.cpp | 96 …const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), s… in configure() local
100 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
101 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
114 Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClElementwiseKernel.cpp | 188 …const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / d… in generate_build_options_with_arithmetic_rules() local
191 …IN1=" + support::cpp11::to_string(src1.dimension(0) == 1 ? 1 : num_elems_processed_per_iteration)); in generate_build_options_with_arithmetic_rules()
192 …IN2=" + support::cpp11::to_string(src2.dimension(0) == 1 ? 1 : num_elems_processed_per_iteration)); in generate_build_options_with_arithmetic_rules()
193 …_opts.add_option("-DVEC_SIZE_OUT=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in generate_build_options_with_arithmetic_rules()
194 …SIZE_LEFTOVER=" + support::cpp11::to_string(dst.dimension(0) % num_elems_processed_per_iteration)); in generate_build_options_with_arithmetic_rules()
222 …const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / d… in configure_window_arithmetic_common() local
223 … = calculate_max_window(dst, Steps(num_elems_processed_per_iteration)); in configure_window_arithmetic_common()
|
| H A D | ClGemmLowpOffsetContributionKernel.cpp | 123 …const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, mm_result->dimension(0)); in configure() local
127 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
128 …FTOVER=" + support::cpp11::to_string(mm_result->dimension(0) % num_elems_processed_per_iteration)); in configure()
152 Window win = calculate_max_window(*mm_result, Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | ClActivationKernel.cpp | 110 …const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), s… in configure() local
130 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
131 …IZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); in configure()
209 Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); in configure()
|
| /aosp_15_r20/external/ComputeLibrary/src/core/CL/kernels/ |
| H A D | CLNormalizePlanarYUVLayerKernel.cpp | 69 const unsigned int num_elems_processed_per_iteration = 16 / input->element_size(); in validate_and_configure_window_nchw() local
71 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); in validate_and_configure_window_nchw()
73 AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); in validate_and_configure_window_nchw()
74 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); in validate_and_configure_window_nchw()
113 …const unsigned int num_elems_processed_per_iteration = (data_layout == DataLayout::NHWC) ? adjust_… in configure() local
122 …ld_opts.add_option(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration))); in configure()
123 …R=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_processed_per_iteration))); in configure()
142 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | CLBatchNormalizationLayerKernel.cpp | 86 …const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->element_size(),… in validate_and_configure_window_nchw() local
89 …indow win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); in validate_and_configure_window_nchw()
90 AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); in validate_and_configure_window_nchw()
95 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); in validate_and_configure_window_nchw()
142 …unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->info()->element_size(… in configure() local
147 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
148 …ER=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_processed_per_iteration)); in configure()
181 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); in configure()
|
| H A D | CLComparisonKernel.cpp | 79 …const unsigned int num_elems_processed_per_iteration = calculate_num_elems_processed_per_iteration… in validate_and_configure_window() local
84 Window win = calculate_max_window(out_shape, Steps(num_elems_processed_per_iteration)); in validate_and_configure_window()
88 AccessWindowHorizontal input1_access(&input1, 0, num_elems_processed_per_iteration); in validate_and_configure_window()
89 AccessWindowHorizontal input2_access(&input2, 0, num_elems_processed_per_iteration); in validate_and_configure_window()
90 AccessWindowHorizontal output_access(&output, 0, num_elems_processed_per_iteration); in validate_and_configure_window()
220 …const int num_elems_processed_per_iteration = calculate_num_elems_processed_per_iteration(*_input1… in border_size() local
223 …const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, … in border_size()
|
| H A D | CLRangeKernel.cpp | 85 …unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / output-… in configure() local
86 … = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); in configure()
100 …d_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
101 …R=" + support::cpp11::to_string(output->info()->dimension(0) % num_elems_processed_per_iteration)); in configure()
|
| H A D | CLQLSTMLayerNormalizationKernel.cpp | 50 …const uint32_t num_elems_processed_per_iteration = (input->dimension(0) < temp_num_elems_processed… in validate_and_configure_window() local
53 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); in validate_and_configure_window()
100 …const uint32_t num_elems_processed_per_iteration = max_cl_vector_width / input->info()->element_si… in configure() local
111 …uild_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); in configure()
|
| /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/ |
| H A D | NEBitwiseOrKernel.cpp | 79 constexpr unsigned int num_elems_processed_per_iteration = 16; in configure() local
82 … win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration)); in configure()
83 AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); in configure()
86 … AccessWindowHorizontal(input1->info(), 0, num_elems_processed_per_iteration), in configure()
87 … AccessWindowHorizontal(input2->info(), 0, num_elems_processed_per_iteration), in configure()
|
| H A D | NEBitwiseXorKernel.cpp | 79 constexpr unsigned int num_elems_processed_per_iteration = 16; in configure() local
81 … win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration)); in configure()
82 AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); in configure()
84 …ndow_and_padding(win, AccessWindowHorizontal(input1->info(), 0, num_elems_processed_per_iteration), in configure()
85 … AccessWindowHorizontal(input2->info(), 0, num_elems_processed_per_iteration), output_access); in configure()
|
| H A D | NEBitwiseAndKernel.cpp | 83 constexpr unsigned int num_elems_processed_per_iteration = 16; in configure() local
86 … win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration)); in configure()
87 AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); in configure()
90 … AccessWindowHorizontal(input1->info(), 0, num_elems_processed_per_iteration), in configure()
91 … AccessWindowHorizontal(input2->info(), 0, num_elems_processed_per_iteration), in configure()
|
| H A D | NEBitwiseNotKernel.cpp | 75 constexpr unsigned int num_elems_processed_per_iteration = 16; in configure() local
78 … win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); in configure()
79 AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); in configure()
80 …w_and_padding(win, AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration), ou… in configure()
|
| /aosp_15_r20/external/ComputeLibrary/src/core/CL/ |
| H A D | ICLSimpleKernel.cpp | 38 …ure(const ICLTensor *input, ICLTensor *output, unsigned int num_elems_processed_per_iteration, boo… in configure() argument
44 … win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), bo… in configure()
45 AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); in configure()
48 … AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration), in configure()
|
| /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/ |
| H A D | CpuPool2dKernel.cpp | 206 … unsigned int &num_elems_processed_per_iteration, in validate_and_configure_window() argument
232 num_elems_processed_per_iteration = 1; in validate_and_configure_window()
243 num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15; in validate_and_configure_window()
246 num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14; in validate_and_configure_window()
254 num_elems_processed_per_iteration = 1; in validate_and_configure_window()
258 num_elems_processed_per_iteration = 1; in validate_and_configure_window()
273 win = calculate_max_window(dst_info, Steps(num_elems_processed_per_iteration)); in validate_and_configure_window()
330 unsigned int num_elems_processed_per_iteration = 0; in validate() local
344 … (indices) ? indices->clone().get() : nullptr, pool_info, num_elems_processed_per_iteration, in validate()
|