|
| 1 | +/* |
| 2 | + * Copyright (C) 2024-2025 Intel Corporation |
| 3 | + * |
| 4 | + * SPDX-License-Identifier: MIT |
| 5 | + * |
| 6 | + */ |
| 7 | + |
| 8 | +#include "framework/test_case/register_test_case.h" |
| 9 | +#include "framework/ur/error.h" |
| 10 | +#include "framework/ur/ur.h" |
| 11 | +#include "framework/utility/file_helper.h" |
| 12 | +#include "framework/utility/timer.h" |
| 13 | + |
| 14 | +#include "definitions/submit_graph.h" |
| 15 | + |
| 16 | +#include <iostream> |
| 17 | +#include <ur_api.h> |
| 18 | + |
| 19 | +static constexpr size_t n_dimensions = 3; |
| 20 | +static constexpr size_t global_size[] = {1, 1, 1}; |
| 21 | +static constexpr size_t local_size[] = {1, 1, 1}; |
| 22 | +static constexpr size_t global_offset = 0; |
| 23 | + |
| 24 | +static TestResult run([[maybe_unused]] const SubmitGraphArguments &arguments, Statistics &statistics) { |
| 25 | + |
| 26 | + MeasurementFields typeSelector(MeasurementUnit::Microseconds, MeasurementType::Cpu); |
| 27 | + |
| 28 | + if (isNoopRun()) { |
| 29 | + statistics.pushUnitAndType(typeSelector.getUnit(), typeSelector.getType()); |
| 30 | + return TestResult::Nooped; |
| 31 | + } |
| 32 | + |
| 33 | + UrState ur; |
| 34 | + |
| 35 | + // check if the device supports command buffers |
| 36 | + ur_bool_t command_buffer_support = false; |
| 37 | + EXPECT_UR_RESULT_SUCCESS(urDeviceGetInfo( |
| 38 | + ur.device, UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP, |
| 39 | + sizeof(command_buffer_support), &command_buffer_support, nullptr)); |
| 40 | + if (!command_buffer_support) { |
| 41 | + return TestResult::DeviceNotCapable; |
| 42 | + } |
| 43 | + |
| 44 | + Timer timer; |
| 45 | + |
| 46 | + // Create kernel |
| 47 | + auto spirvModule = FileHelper::loadBinaryFile("api_overhead_benchmark_eat_time.spv"); |
| 48 | + if (spirvModule.size() == 0) |
| 49 | + return TestResult::KernelNotFound; |
| 50 | + |
| 51 | + ur_program_handle_t program; |
| 52 | + EXPECT_UR_RESULT_SUCCESS(urProgramCreateWithIL(ur.context, spirvModule.data(), |
| 53 | + spirvModule.size(), nullptr, &program)); |
| 54 | + |
| 55 | + EXPECT_UR_RESULT_SUCCESS(urProgramBuild(ur.context, program, nullptr)); |
| 56 | + |
| 57 | + const char *kernelName = "eat_time"; |
| 58 | + |
| 59 | + ur_kernel_handle_t kernel; |
| 60 | + EXPECT_UR_RESULT_SUCCESS(urKernelCreate(program, kernelName, &kernel)); |
| 61 | + |
| 62 | + ur_queue_handle_t queue; |
| 63 | + ur_queue_properties_t queueProperties = {}; |
| 64 | + |
| 65 | + auto inOrder = arguments.inOrderQueue; |
| 66 | + if (!inOrder) { |
| 67 | + queueProperties.flags = UR_QUEUE_FLAG_OUT_OF_ORDER_EXEC_MODE_ENABLE; |
| 68 | + } |
| 69 | + |
| 70 | + EXPECT_UR_RESULT_SUCCESS(urQueueCreate(ur.context, ur.device, |
| 71 | + &queueProperties, &queue)); |
| 72 | + |
| 73 | + ur_exp_command_buffer_handle_t cmdBuffer = nullptr; |
| 74 | + |
| 75 | + // Finalize the graph |
| 76 | + ur_exp_command_buffer_desc_t cmdBufferDesc = { |
| 77 | + UR_STRUCTURE_TYPE_EXP_COMMAND_BUFFER_DESC, |
| 78 | + nullptr, // pNext |
| 79 | + false, // isUpdatable |
| 80 | + inOrder, // isInOrder |
| 81 | + false // enableProfiling |
| 82 | + }; |
| 83 | + |
| 84 | + EXPECT_UR_RESULT_SUCCESS(urCommandBufferCreateExp( |
| 85 | + ur.context, ur.device, &cmdBufferDesc, &cmdBuffer)); |
| 86 | + for (auto i = 0u; i < arguments.numKernels; i++) { |
| 87 | + EXPECT_UR_RESULT_SUCCESS(urCommandBufferAppendKernelLaunchExp( |
| 88 | + cmdBuffer, kernel, n_dimensions, &global_offset, global_size, |
| 89 | + nullptr, 0, nullptr, 0, nullptr, 0, nullptr, |
| 90 | + nullptr, nullptr, nullptr)); |
| 91 | + } |
| 92 | + |
| 93 | + EXPECT_UR_RESULT_SUCCESS(urCommandBufferFinalizeExp( |
| 94 | + cmdBuffer)); |
| 95 | + |
| 96 | + // Warmup |
| 97 | + EXPECT_UR_RESULT_SUCCESS( |
| 98 | + urEnqueueCommandBufferExp(queue, cmdBuffer, 0, nullptr, nullptr)); |
| 99 | + EXPECT_UR_RESULT_SUCCESS(urQueueFinish(queue)); |
| 100 | + |
| 101 | + // Benchmark |
| 102 | + for (auto i = 0u; i < arguments.iterations; i++) { |
| 103 | + timer.measureStart(); |
| 104 | + |
| 105 | + EXPECT_UR_RESULT_SUCCESS( |
| 106 | + urEnqueueCommandBufferExp(queue, cmdBuffer, 0, nullptr, nullptr)); |
| 107 | + |
| 108 | + if (!arguments.measureCompletionTime) { |
| 109 | + timer.measureEnd(); |
| 110 | + } |
| 111 | + |
| 112 | + EXPECT_UR_RESULT_SUCCESS(urQueueFinish(queue)); |
| 113 | + |
| 114 | + if (arguments.measureCompletionTime) { |
| 115 | + timer.measureEnd(); |
| 116 | + } |
| 117 | + statistics.pushValue(timer.get(), typeSelector.getUnit(), typeSelector.getType()); |
| 118 | + } |
| 119 | + |
| 120 | + EXPECT_UR_RESULT_SUCCESS(urCommandBufferReleaseExp(cmdBuffer)); |
| 121 | + EXPECT_UR_RESULT_SUCCESS(urQueueRelease(queue)); |
| 122 | + EXPECT_UR_RESULT_SUCCESS(urKernelRelease(kernel)); |
| 123 | + EXPECT_UR_RESULT_SUCCESS(urProgramRelease(program)); |
| 124 | + |
| 125 | + return TestResult::Success; |
| 126 | +} |
| 127 | + |
| 128 | +[[maybe_unused]] static RegisterTestCaseImplementation<SubmitGraph> registerTestCase(run, Api::UR); |
0 commit comments