Michalis Spyrou | d7e8281 | 2017-06-20 15:00:14 +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 "CL/CLAccessor.h" |
| 25 | #include "Globals.h" |
| 26 | #include "TensorLibrary.h" |
| 27 | #include "TypePrinter.h" |
| 28 | #include "Utils.h" |
| 29 | #include "validation/Datasets.h" |
| 30 | #include "validation/Reference.h" |
| 31 | #include "validation/Validation.h" |
| 32 | |
| 33 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 34 | #include "arm_compute/core/CL/ICLKernel.h" |
| 35 | #include "arm_compute/core/CL/OpenCL.h" |
| 36 | #include "arm_compute/core/Helpers.h" |
| 37 | #include "arm_compute/core/Types.h" |
| 38 | #include "arm_compute/core/Window.h" |
| 39 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 40 | #include "arm_compute/runtime/CL/CLSubTensor.h" |
| 41 | #include "arm_compute/runtime/CL/CLTensor.h" |
| 42 | #include "arm_compute/runtime/CL/CLTensorAllocator.h" |
| 43 | |
| 44 | #include "arm_compute/core/CL/ICLTensor.h" |
| 45 | |
| 46 | #include "boost_wrapper.h" |
| 47 | |
| 48 | #include <random> |
| 49 | #include <string> |
| 50 | |
| 51 | using namespace arm_compute; |
| 52 | using namespace arm_compute::test; |
| 53 | using namespace arm_compute::test::cl; |
| 54 | using namespace arm_compute::test::validation; |
| 55 | |
| 56 | namespace |
| 57 | { |
| 58 | const float tolerance_exp = 1.0f; /**< Tolerance value for comparing reference's output against implementation's output (exponential)*/ |
| 59 | const float tolerance_invsqrt = 4.0f; /**< Tolerance value for comparing reference's output against implementation's output (inverse square-root) */ |
| 60 | const float tolerance_log = 5.0f; /**< Tolerance value for comparing reference's output against implementation's output (logarithm) */ |
| 61 | |
| 62 | /** Compute Neon fixed point operation for signed 8bit fixed point. |
| 63 | * |
| 64 | * @param[in] shape Shape of the input and output tensors. |
| 65 | * |
| 66 | * @return Computed output tensor. |
| 67 | */ |
| 68 | CLTensor compute_fixed_point_op(const TensorShape &shape, int fixed_point_position, FixedPointOp op) |
| 69 | { |
| 70 | std::string fixed_point_operation_kernel; |
| 71 | #ifndef EMBEDDED_KERNELS |
| 72 | fixed_point_operation_kernel += "#include \"fixed_point.h\"\n"; |
| 73 | #endif |
| 74 | fixed_point_operation_kernel += |
| 75 | "__kernel void fixed_point_operation_qs8( \n" |
| 76 | " __global char* src, \n" |
| 77 | " __global char* dst) \n" |
| 78 | "{ \n" |
| 79 | " char16 in = vload16(0, src + get_global_id(0) * 16); \n" |
| 80 | " if(FIXED_POINT_OP == 0) \n" |
| 81 | " { \n" |
| 82 | " vstore16(EXP_OP_EXPAND(in, DATA_TYPE, 16, FIXED_POINT_POS), 0, dst + get_global_id(0) * 16); \n" |
| 83 | " } \n" |
| 84 | " else if(FIXED_POINT_OP == 1) \n" |
| 85 | " { \n" |
| 86 | " vstore16(INVSQRT_OP_EXPAND(in, DATA_TYPE, 16, FIXED_POINT_POS), 0, dst + get_global_id(0) * 16); \n" |
| 87 | " } \n" |
| 88 | " else \n" |
| 89 | " { \n" |
| 90 | " vstore16(LOG_OP_EXPAND(in, DATA_TYPE, 16, FIXED_POINT_POS), 0, dst + get_global_id(0) * 16); \n" |
| 91 | " } \n" |
| 92 | "} \n" |
| 93 | "\n"; |
| 94 | |
| 95 | // Create tensors |
| 96 | CLTensor src = create_tensor<CLTensor>(shape, DataType::QS8, 1, fixed_point_position); |
| 97 | CLTensor dst = create_tensor<CLTensor>(shape, DataType::QS8, 1, fixed_point_position); |
| 98 | |
| 99 | // Allocate tensors |
| 100 | src.allocator()->allocate(); |
| 101 | dst.allocator()->allocate(); |
| 102 | |
| 103 | BOOST_TEST(!src.info()->is_resizable()); |
| 104 | BOOST_TEST(!dst.info()->is_resizable()); |
| 105 | |
| 106 | // Set build options |
| 107 | std::string build_opts = "-DFIXED_POINT_POS=" + val_to_string<int>(fixed_point_position); |
| 108 | build_opts += " -DDATA_TYPE=qs8"; |
| 109 | |
| 110 | // Fill tensors. |
| 111 | int min = 0; |
| 112 | int max = 0; |
| 113 | switch(op) |
| 114 | { |
| 115 | case FixedPointOp::EXP: |
| 116 | min = -(1 << (fixed_point_position - 1)); |
| 117 | max = (1 << (fixed_point_position - 1)); |
| 118 | build_opts += " -DFIXED_POINT_OP=0"; |
| 119 | break; |
| 120 | case FixedPointOp::INV_SQRT: |
| 121 | min = 1; |
| 122 | max = 0x7F; |
| 123 | build_opts += " -DFIXED_POINT_OP=1"; |
| 124 | break; |
| 125 | case FixedPointOp::LOG: |
| 126 | min = (1 << (fixed_point_position - 1)); |
| 127 | max = 0x3F; |
| 128 | build_opts += " -DFIXED_POINT_OP=2"; |
| 129 | break; |
| 130 | default: |
| 131 | ARM_COMPUTE_ERROR("Operation not supported"); |
| 132 | } |
| 133 | |
| 134 | std::uniform_int_distribution<> distribution(min, max); |
| 135 | library->fill(CLAccessor(src), distribution, 0); |
| 136 | |
| 137 | std::vector<std::string> sources; |
| 138 | |
| 139 | #ifndef EMBEDDED_KERNELS |
| 140 | build_opts += " -I" + CLKernelLibrary::get().get_kernel_path(); |
| 141 | #else |
| 142 | sources.push_back(CLKernelLibrary::get().get_program_source("fixed_point.h")); |
| 143 | #endif /* EMBEDDED_KERNELS */ |
| 144 | |
| 145 | sources.push_back(fixed_point_operation_kernel); |
| 146 | |
| 147 | // Create program |
| 148 | ::cl::Program program = ::cl::Program(sources); |
| 149 | |
| 150 | // Build program |
| 151 | program.build(build_opts.c_str()); |
| 152 | |
| 153 | ::cl::Kernel kernel = ::cl::Kernel(program, "fixed_point_operation_qs8", nullptr); |
| 154 | |
| 155 | unsigned int idx = 0; |
| 156 | kernel.setArg(idx++, src.cl_buffer()); |
| 157 | kernel.setArg(idx++, dst.cl_buffer()); |
| 158 | |
| 159 | ::cl::NDRange gws(shape[0] / 16, 1, 1); |
| 160 | CLScheduler::get().queue().enqueueNDRangeKernel(kernel, 0, gws); |
| 161 | |
| 162 | return dst; |
| 163 | } |
| 164 | } // namespace |
| 165 | |
| 166 | #ifndef DOXYGEN_SKIP_THIS |
| 167 | BOOST_AUTO_TEST_SUITE(CL) |
| 168 | BOOST_AUTO_TEST_SUITE(FixedPoint) |
| 169 | BOOST_AUTO_TEST_SUITE(QS8) |
| 170 | |
| 171 | BOOST_AUTO_TEST_SUITE(Exp) |
| 172 | |
| 173 | BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) |
| 174 | BOOST_DATA_TEST_CASE(RunSmall, Small1DShape() * boost::unit_test::data::xrange(1, 6), shape, fixed_point_position) |
| 175 | { |
| 176 | // Compute function |
| 177 | CLTensor dst = compute_fixed_point_op(shape, fixed_point_position, FixedPointOp::EXP); |
| 178 | |
| 179 | // Compute reference |
| 180 | RawTensor ref_dst = Reference::compute_reference_fixed_point_operation(shape, DataType::QS8, DataType::QS8, FixedPointOp::EXP, fixed_point_position); |
| 181 | |
| 182 | // Validate output |
| 183 | validate(CLAccessor(dst), ref_dst, tolerance_exp); |
| 184 | } |
| 185 | |
| 186 | BOOST_AUTO_TEST_SUITE_END() |
| 187 | |
| 188 | BOOST_AUTO_TEST_SUITE(Log) |
| 189 | |
| 190 | BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) |
| 191 | BOOST_DATA_TEST_CASE(RunSmall, Small1DShape() * boost::unit_test::data::xrange(3, 6), shape, fixed_point_position) |
| 192 | { |
| 193 | // Compute function |
| 194 | CLTensor dst = compute_fixed_point_op(shape, fixed_point_position, FixedPointOp::LOG); |
| 195 | |
| 196 | // Compute reference |
| 197 | RawTensor ref_dst = Reference::compute_reference_fixed_point_operation(shape, DataType::QS8, DataType::QS8, FixedPointOp::LOG, fixed_point_position); |
| 198 | |
| 199 | // Validate output |
| 200 | validate(CLAccessor(dst), ref_dst, tolerance_log); |
| 201 | } |
| 202 | |
| 203 | BOOST_AUTO_TEST_SUITE_END() |
| 204 | |
| 205 | BOOST_AUTO_TEST_SUITE(Invsqrt) |
| 206 | |
| 207 | BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) |
| 208 | BOOST_DATA_TEST_CASE(RunSmall, Small1DShape() * boost::unit_test::data::xrange(1, 6), shape, fixed_point_position) |
| 209 | { |
| 210 | // Compute function |
| 211 | CLTensor dst = compute_fixed_point_op(shape, fixed_point_position, FixedPointOp::INV_SQRT); |
| 212 | |
| 213 | // Compute reference |
| 214 | RawTensor ref_dst = Reference::compute_reference_fixed_point_operation(shape, DataType::QS8, DataType::QS8, FixedPointOp::INV_SQRT, fixed_point_position); |
| 215 | |
| 216 | // Validate output |
| 217 | validate(CLAccessor(dst), ref_dst, tolerance_invsqrt); |
| 218 | } |
| 219 | |
| 220 | BOOST_AUTO_TEST_SUITE_END() |
| 221 | |
| 222 | BOOST_AUTO_TEST_SUITE_END() |
| 223 | BOOST_AUTO_TEST_SUITE_END() |
| 224 | BOOST_AUTO_TEST_SUITE_END() |
| 225 | #endif |