|
| 1 | +// Copyright (C) 2024 Intel Corporation |
| 2 | +// Part of the Unified-Runtime Project, under the Apache License v2.0 with LLVM Exceptions. |
| 3 | +// See LICENSE.TXT |
| 4 | +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 5 | + |
| 6 | +#include "ur_print.hpp" |
| 7 | +#include "uur/fixtures.h" |
| 8 | +#include "uur/raii.h" |
| 9 | + |
| 10 | +#include <map> |
| 11 | +#include <string> |
| 12 | + |
| 13 | +template <typename... Args> auto combineFlags(std::tuple<Args...> tuple) { |
| 14 | + return std::apply([](auto... args) { return (... |= args); }, tuple); |
| 15 | +} |
| 16 | + |
| 17 | +extern std::map<std::string, int> *ZeCallCount; |
| 18 | + |
| 19 | +using FlagsTupleType = std::tuple<ur_queue_flags_t, ur_queue_flags_t, |
| 20 | + ur_queue_flags_t, ur_queue_flags_t>; |
| 21 | + |
| 22 | +struct urEventCacheTest : uur::urContextTestWithParam<FlagsTupleType> { |
| 23 | + void SetUp() override { |
| 24 | + UUR_RETURN_ON_FATAL_FAILURE(urContextTestWithParam::SetUp()); |
| 25 | + |
| 26 | + flags = combineFlags(getParam()); |
| 27 | + |
| 28 | + ur_queue_properties_t props; |
| 29 | + props.flags = flags; |
| 30 | + ASSERT_SUCCESS(urQueueCreate(context, device, &props, &queue)); |
| 31 | + ASSERT_NE(queue, nullptr); |
| 32 | + |
| 33 | + ASSERT_SUCCESS(urMemBufferCreate(context, UR_MEM_FLAG_WRITE_ONLY, size, |
| 34 | + nullptr, &buffer)); |
| 35 | + |
| 36 | + (*ZeCallCount)["zeEventCreate"] = 0; |
| 37 | + (*ZeCallCount)["zeEventDestroy"] = 0; |
| 38 | + } |
| 39 | + |
| 40 | + void TearDown() override { |
| 41 | + if (buffer) { |
| 42 | + EXPECT_SUCCESS(urMemRelease(buffer)); |
| 43 | + } |
| 44 | + if (queue) { |
| 45 | + UUR_ASSERT_SUCCESS_OR_UNSUPPORTED(urQueueRelease(queue)); |
| 46 | + } |
| 47 | + UUR_RETURN_ON_FATAL_FAILURE(urContextTestWithParam::TearDown()); |
| 48 | + } |
| 49 | + |
| 50 | + auto enqueueWork(ur_event_handle_t *hEvent, int data) { |
| 51 | + input.assign(count, data); |
| 52 | + UUR_ASSERT_SUCCESS_OR_UNSUPPORTED(urEnqueueMemBufferWrite( |
| 53 | + queue, buffer, false, 0, size, input.data(), 0, nullptr, hEvent)); |
| 54 | + } |
| 55 | + |
| 56 | + void verifyData() { |
| 57 | + std::vector<uint32_t> output(count, 1); |
| 58 | + UUR_ASSERT_SUCCESS_OR_UNSUPPORTED(urEnqueueMemBufferRead( |
| 59 | + queue, buffer, true, 0, size, output.data(), 0, nullptr, nullptr)); |
| 60 | + |
| 61 | + if (!(flags & UR_QUEUE_FLAG_OUT_OF_ORDER_EXEC_MODE_ENABLE)) { |
| 62 | + ASSERT_EQ(input, output); |
| 63 | + } |
| 64 | + } |
| 65 | + |
| 66 | + const size_t count = 1024; |
| 67 | + const size_t size = sizeof(uint32_t) * count; |
| 68 | + ur_mem_handle_t buffer = nullptr; |
| 69 | + ur_queue_handle_t queue = nullptr; |
| 70 | + std::vector<uint32_t> input; |
| 71 | + ur_queue_flags_t flags; |
| 72 | +}; |
| 73 | + |
| 74 | +TEST_P(urEventCacheTest, eventsReuseNoVisibleEvent) { |
| 75 | + static constexpr int numIters = 16; |
| 76 | + static constexpr int numEnqueues = 128; |
| 77 | + |
| 78 | + for (int i = 0; i < numIters; i++) { |
| 79 | + for (int j = 0; j < numEnqueues; j++) { |
| 80 | + enqueueWork(nullptr, i * numEnqueues + j); |
| 81 | + } |
| 82 | + UUR_ASSERT_SUCCESS_OR_UNSUPPORTED(urQueueFinish(queue)); |
| 83 | + verifyData(); |
| 84 | + } |
| 85 | + |
| 86 | + // TODO: why events are not reused for UR_QUEUE_FLAG_OUT_OF_ORDER_EXEC_MODE_ENABLE? |
| 87 | + if ((flags & UR_QUEUE_FLAG_DISCARD_EVENTS) && |
| 88 | + !(flags & UR_QUEUE_FLAG_OUT_OF_ORDER_EXEC_MODE_ENABLE)) { |
| 89 | + ASSERT_EQ((*ZeCallCount)["zeEventCreate"], 2); |
| 90 | + } else { |
| 91 | + ASSERT_GE((*ZeCallCount)["zeEventCreate"], numIters * numEnqueues); |
| 92 | + } |
| 93 | +} |
| 94 | + |
| 95 | +TEST_P(urEventCacheTest, eventsReuseWithVisibleEvent) { |
| 96 | + static constexpr int numIters = 16; |
| 97 | + static constexpr int numEnqueues = 128; |
| 98 | + |
| 99 | + for (int i = 0; i < numIters; i++) { |
| 100 | + std::vector<uur::raii::Event> events(numEnqueues); |
| 101 | + for (int j = 0; j < numEnqueues; j++) { |
| 102 | + enqueueWork(events[j].ptr(), i * numEnqueues + j); |
| 103 | + } |
| 104 | + UUR_ASSERT_SUCCESS_OR_UNSUPPORTED(urQueueFinish(queue)); |
| 105 | + verifyData(); |
| 106 | + } |
| 107 | + |
| 108 | + ASSERT_LT((*ZeCallCount)["zeEventCreate"], numIters * numEnqueues); |
| 109 | +} |
| 110 | + |
| 111 | +TEST_P(urEventCacheTest, eventsReuseWithVisibleEventAndWait) { |
| 112 | + static constexpr int numIters = 16; |
| 113 | + static constexpr int numEnqueues = 128; |
| 114 | + static constexpr int waitEveryN = 16; |
| 115 | + |
| 116 | + for (int i = 0; i < numIters; i++) { |
| 117 | + std::vector<uur::raii::Event> events; |
| 118 | + for (int j = 0; j < numEnqueues; j++) { |
| 119 | + events.emplace_back(); |
| 120 | + enqueueWork(events.back().ptr(), i * numEnqueues + j); |
| 121 | + |
| 122 | + if (j > 0 && j % waitEveryN == 0) { |
| 123 | + ASSERT_SUCCESS(urEventWait(waitEveryN, |
| 124 | + (ur_event_handle_t *)events.data())); |
| 125 | + verifyData(); |
| 126 | + events.clear(); |
| 127 | + } |
| 128 | + } |
| 129 | + UUR_ASSERT_SUCCESS_OR_UNSUPPORTED(urQueueFinish(queue)); |
| 130 | + } |
| 131 | + |
| 132 | + ASSERT_GE((*ZeCallCount)["zeEventCreate"], waitEveryN); |
| 133 | + // TODO: why there are more events than this? |
| 134 | + // ASSERT_LE((*ZeCallCount)["zeEventCreate"], waitEveryN * 2 + 2); |
| 135 | +} |
| 136 | + |
| 137 | +template <typename T> |
| 138 | +inline std::string |
| 139 | +printFlags(const testing::TestParamInfo<typename T::ParamType> &info) { |
| 140 | + const auto device_handle = std::get<0>(info.param); |
| 141 | + const auto platform_device_name = |
| 142 | + uur::GetPlatformAndDeviceName(device_handle); |
| 143 | + auto flags = combineFlags(std::get<1>(info.param)); |
| 144 | + |
| 145 | + std::stringstream ss; |
| 146 | + ur::details::printFlag<ur_queue_flag_t>(ss, flags); |
| 147 | + |
| 148 | + auto str = ss.str(); |
| 149 | + std::replace(str.begin(), str.end(), ' ', '_'); |
| 150 | + std::replace(str.begin(), str.end(), '|', '_'); |
| 151 | + return platform_device_name + "__" + str; |
| 152 | +} |
| 153 | + |
| 154 | +UUR_TEST_SUITE_P( |
| 155 | + urEventCacheTest, |
| 156 | + ::testing::Combine( |
| 157 | + testing::Values(0, UR_QUEUE_FLAG_DISCARD_EVENTS), |
| 158 | + testing::Values(0, UR_QUEUE_FLAG_OUT_OF_ORDER_EXEC_MODE_ENABLE), |
| 159 | + // TODO: why the test fails with UR_QUEUE_FLAG_SUBMISSION_BATCHED? |
| 160 | + testing::Values( |
| 161 | + UR_QUEUE_FLAG_SUBMISSION_IMMEDIATE /*, UR_QUEUE_FLAG_SUBMISSION_BATCHED */), |
| 162 | + testing::Values(0, UR_QUEUE_FLAG_PROFILING_ENABLE)), |
| 163 | + printFlags<urEventCacheTest>); |
0 commit comments