[CTS] Add more USM testing

* Expand USMMemcpy2D testing to test all USM allocation types
* Add USM linked list kernel execution test

The linked list test is based on USM/smemll.cpp from the SYCL-E2E tests
and checks that USM allocations are actually usable from a kernel.

test/conformance/device_code/CMakeLists.txt

@@ -53,6 +53,7 @@ add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/foo.cpp)
 add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/image_copy.cpp)
 add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/mean.cpp)
 add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/spec_constant.cpp)
+add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/usm_ll.cpp)
 
 set(KERNEL_HEADER ${UR_CONFORMANCE_DEVICE_BINARIES_DIR}/kernel_entry_points.h)
 add_custom_command(OUTPUT ${KERNEL_HEADER}

test/conformance/device_code/usm_ll.cpp

@@ -0,0 +1,68 @@
+// Copyright (C) 2023 Intel Corporation
+// Part of the Unified-Runtime Project, under the Apache License v2.0 with LLVM Exceptions.
+// See LICENSE.TXT
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#include <sycl/sycl.hpp>
+
+using namespace sycl;
+
+int numNodes = 4;
+
+struct Node {
+    Node() : pNext(nullptr), Num(0xDEADBEEF) {}
+
+    Node *pNext;
+    uint32_t Num;
+};
+
+int main() {
+    queue q;
+    auto dev = q.get_device();
+    auto ctxt = q.get_context();
+
+    if (!dev.get_info<info::device::usm_shared_allocations>()) {
+        return 0;
+    }
+
+    Node *s_head =
+        (Node *)aligned_alloc_shared(alignof(Node), sizeof(Node), dev, ctxt);
+    if (s_head == nullptr) {
+        return -1;
+    }
+    Node *s_cur = s_head;
+
+    for (int i = 0; i < numNodes; i++) {
+        s_cur->Num = i * 2;
+
+        if (i != (numNodes - 1)) {
+            s_cur->pNext = (Node *)aligned_alloc_shared(
+                alignof(Node), sizeof(Node), dev, ctxt);
+        } else {
+            s_cur->pNext = nullptr;
+        }
+
+        s_cur = s_cur->pNext;
+    }
+
+    auto e1 = q.submit([=](handler &cgh) {
+        cgh.single_task<class linkedlist>([=]() {
+            Node *pHead = s_head;
+            while (pHead) {
+                pHead->Num = pHead->Num * 2 + 1;
+                pHead = pHead->pNext;
+            }
+        });
+    });
+
+    e1.wait();
+
+    s_cur = s_head;
+    for (int i = 0; i < numNodes; i++) {
+        Node *old = s_cur;
+        s_cur = s_cur->pNext;
+        free(old, ctxt);
+    }
+
+    return 0;
+}

test/conformance/enqueue/enqueue_adapter_opencl.match

@@ -33,3 +33,4 @@
 {{OPT}}urEnqueueUSMMemcpy2DNegativeTest.InvalidSize/Intel_R__OpenCL___{{.*}}
 {{OPT}}urEnqueueUSMMemcpy2DNegativeTest.InvalidEventWaitList/Intel_R__OpenCL___{{.*}}
 {{OPT}}urEnqueueUSMPrefetchTest.InvalidSizeTooLarge/Intel_R__OpenCL___{{.*}}
+{{OPT}}urEnqueueKernelLaunchUSMLinkedList.Success/Intel_R__OpenCL___{{.*}}_UsePoolEnabled

test/conformance/enqueue/helpers.h

@@ -66,17 +66,33 @@ struct TestParameters2D {
     size_t height;
 };
 
+inline std::string USMKindToString(USMKind kind) {
+    switch (kind) {
+    case USMKind::Device:
+        return "Device";
+    case USMKind::Host:
+        return "Host";
+    case USMKind::Shared:
+    default:
+        return "Shared";
+    }
+}
+
 template <typename T>
 inline std::string
 print2DTestString(const testing::TestParamInfo<typename T::ParamType> &info) {
     const auto device_handle = std::get<0>(info.param);
     const auto platform_device_name =
         uur::GetPlatformAndDeviceName(device_handle);
     std::stringstream test_name;
+    auto src_kind = std::get<1>(std::get<1>(info.param));
+    auto dst_kind = std::get<2>(std::get<1>(info.param));
     test_name << platform_device_name << "__pitch__"
-              << std::get<1>(info.param).pitch << "__width__"
-              << std::get<1>(info.param).width << "__height__"
-              << std::get<1>(info.param).height;
+              << std::get<0>(std::get<1>(info.param)).pitch << "__width__"
+              << std::get<0>(std::get<1>(info.param)).width << "__height__"
+              << std::get<0>(std::get<1>(info.param)).height << "__src__"
+              << USMKindToString(src_kind) << "__dst__"
+              << USMKindToString(dst_kind);
     return test_name.str();
 }
 

test/conformance/enqueue/urEnqueueKernelLaunch.cpp

@@ -262,3 +262,102 @@ TEST_P(urEnqueueKernelLaunchMultiDeviceTest, KernelLaunchReadDifferentQueues) {
         ASSERT_EQ(val, output) << "Result on queue " << i << " did not match!";
     }
 }
+
+struct urEnqueueKernelLaunchUSMLinkedList
+    : uur::urKernelTestWithParam<uur::BoolTestParam> {
+    struct Node {
+        Node() : next(nullptr), num(0xDEADBEEF) {}
+
+        Node *next;
+        uint32_t num;
+    };
+
+    void SetUp() override {
+        program_name = "usm_ll";
+        UUR_RETURN_ON_FATAL_FAILURE(
+            uur::urKernelTestWithParam<uur::BoolTestParam>::SetUp());
+
+        use_pool = getParam().value;
+        ASSERT_SUCCESS(urQueueCreate(context, device, 0, &queue));
+        ur_usm_pool_desc_t pool_desc{UR_STRUCTURE_TYPE_USM_POOL_DESC, nullptr,
+                                     0};
+        if (use_pool) {
+            ASSERT_SUCCESS(urUSMPoolCreate(this->context, &pool_desc, &pool));
+        }
+    }
+
+    void TearDown() override {
+        auto *list_cur = list_head;
+        while (list_cur) {
+            auto *list_next = list_cur->next;
+            ASSERT_SUCCESS(urUSMFree(context, list_cur));
+            list_cur = list_next;
+        }
+
+        if (queue) {
+            ASSERT_SUCCESS(urQueueRelease(queue));
+        }
+
+        if (pool) {
+            ASSERT_SUCCESS(urUSMPoolRelease(pool));
+        }
+
+        UUR_RETURN_ON_FATAL_FAILURE(
+            uur::urKernelTestWithParam<uur::BoolTestParam>::TearDown());
+    }
+
+    size_t global_size = 1;
+    size_t global_offset = 0;
+    Node *list_head = nullptr;
+    const int num_nodes = 4;
+    bool use_pool = false;
+    ur_usm_pool_handle_t pool = nullptr;
+    ur_queue_handle_t queue;
+};
+
+UUR_TEST_SUITE_P(
+    urEnqueueKernelLaunchUSMLinkedList,
+    testing::ValuesIn(uur::BoolTestParam::makeBoolParam("UsePool")),
+    uur::deviceTestWithParamPrinter<uur::BoolTestParam>);
+
+TEST_P(urEnqueueKernelLaunchUSMLinkedList, Success) {
+    ur_device_usm_access_capability_flags_t shared_usm_flags = 0;
+    ASSERT_SUCCESS(
+        uur::GetDeviceUSMSingleSharedSupport(device, shared_usm_flags));
+    if (!(shared_usm_flags & UR_DEVICE_USM_ACCESS_CAPABILITY_FLAG_ACCESS)) {
+        GTEST_SKIP() << "Shared USM is not supported.";
+    }
+
+    // Build linked list with USM allocations
+    ASSERT_SUCCESS(urUSMSharedAlloc(context, device, nullptr, pool,
+                                    sizeof(Node),
+                                    reinterpret_cast<void **>(&list_head)));
+    ASSERT_NE(list_head, nullptr);
+    Node *list_cur = list_head;
+    for (int i = 0; i < num_nodes; i++) {
+        list_cur->num = i * 2;
+        if (i < num_nodes - 1) {
+            ASSERT_SUCCESS(
+                urUSMSharedAlloc(context, device, nullptr, pool, sizeof(Node),
+                                 reinterpret_cast<void **>(&list_cur->next)));
+            ASSERT_NE(list_cur->next, nullptr);
+        } else {
+            list_cur->next = nullptr;
+        }
+        list_cur = list_cur->next;
+    }
+
+    // Run kernel which will iterate the list and modify the values
+    ASSERT_SUCCESS(urKernelSetArgPointer(kernel, 0, nullptr, &list_head));
+    ASSERT_SUCCESS(urEnqueueKernelLaunch(queue, kernel, 1, &global_offset,
+                                         &global_size, nullptr, 0, nullptr,
+                                         nullptr));
+    ASSERT_SUCCESS(urQueueFinish(queue));
+
+    // Verify values
+    list_cur = list_head;
+    for (int i = 0; i < num_nodes; i++) {
+        ASSERT_EQ(list_cur->num, i * 4 + 1);
+        list_cur = list_cur->next;
+    }
+}

test/conformance/enqueue/urEnqueueUSMMemcpy2D.cpp

@@ -6,14 +6,24 @@
 #include "helpers.h"
 #include <uur/fixtures.h>
 
+using TestParametersMemcpy2D =
+    std::tuple<uur::TestParameters2D, uur::USMKind, uur::USMKind>;
+
 struct urEnqueueUSMMemcpy2DTestWithParam
-    : uur::urQueueTestWithParam<uur::TestParameters2D> {
+    : uur::urQueueTestWithParam<TestParametersMemcpy2D> {
     void SetUp() override {
         UUR_RETURN_ON_FATAL_FAILURE(
-            uur::urQueueTestWithParam<uur::TestParameters2D>::SetUp());
+            uur::urQueueTestWithParam<TestParametersMemcpy2D>::SetUp());
+
+        const auto [in2DParams, inSrcKind, inDstKind] = getParam();
+        std::tie(pitch, width, height, src_kind, dst_kind) =
+            std::make_tuple(in2DParams.pitch, in2DParams.width,
+                            in2DParams.height, inSrcKind, inDstKind);
+
         ur_device_usm_access_capability_flags_t device_usm = 0;
         ASSERT_SUCCESS(uur::GetDeviceUSMDeviceSupport(device, device_usm));
-        if (!device_usm) {
+        if (!device_usm && (src_kind == uur::USMKind::Device ||
+                            dst_kind == uur::USMKind::Device)) {
             GTEST_SKIP() << "Device USM is not supported";
         }
 
@@ -25,15 +35,13 @@ struct urEnqueueUSMMemcpy2DTestWithParam
             GTEST_SKIP() << "2D USM memcpy is not supported";
         }
 
-        const auto [inPitch, inWidth, inHeight] = getParam();
-        std::tie(pitch, width, height) =
-            std::make_tuple(inPitch, inWidth, inHeight);
-
         const size_t num_elements = pitch * height;
-        ASSERT_SUCCESS(urUSMDeviceAlloc(context, device, nullptr, nullptr,
-                                        num_elements, &pSrc));
-        ASSERT_SUCCESS(urUSMDeviceAlloc(context, device, nullptr, nullptr,
-                                        num_elements, &pDst));
+        ASSERT_SUCCESS(uur::MakeUSMAllocationByType(
+            src_kind, context, device, nullptr, nullptr, num_elements, &pSrc));
+
+        ASSERT_SUCCESS(uur::MakeUSMAllocationByType(
+            dst_kind, context, device, nullptr, nullptr, num_elements, &pDst));
+
         ur_event_handle_t memset_event = nullptr;
 
         ASSERT_SUCCESS(urEnqueueUSMFill(queue, pSrc, sizeof(memset_value),
@@ -52,17 +60,22 @@ struct urEnqueueUSMMemcpy2DTestWithParam
         if (pDst) {
             ASSERT_SUCCESS(urUSMFree(context, pDst));
         }
-        uur::urQueueTestWithParam<uur::TestParameters2D>::TearDown();
+        uur::urQueueTestWithParam<TestParametersMemcpy2D>::TearDown();
     }
 
     void verifyMemcpySucceeded() {
         std::vector<uint8_t> host_mem(pitch * height);
-        ASSERT_SUCCESS(urEnqueueUSMMemcpy2D(queue, true, host_mem.data(), pitch,
-                                            pDst, pitch, width, height, 0,
-                                            nullptr, nullptr));
+        const uint8_t *host_ptr = nullptr;
+        if (dst_kind == uur::USMKind::Device) {
+            ASSERT_SUCCESS(urEnqueueUSMMemcpy2D(queue, true, host_mem.data(),
+                                                pitch, pDst, pitch, width,
+                                                height, 0, nullptr, nullptr));
+            host_ptr = host_mem.data();
+        } else {
+            host_ptr = static_cast<const uint8_t *>(pDst);
+        }
         for (size_t w = 0; w < width; ++w) {
             for (size_t h = 0; h < height; ++h) {
-                const auto *host_ptr = host_mem.data();
                 const size_t index = (pitch * h) + w;
                 ASSERT_TRUE(*(host_ptr + index) == memset_value);
             }
@@ -75,9 +88,11 @@ struct urEnqueueUSMMemcpy2DTestWithParam
     size_t pitch = 0;
     size_t width = 0;
     size_t height = 0;
+    uur::USMKind src_kind;
+    uur::USMKind dst_kind;
 };
 
-static std::vector<uur::TestParameters2D> test_cases{
+static std::vector<uur::TestParameters2D> test_sizes{
     /* Everything set to 1 */
     {1, 1, 1},
     /* Height == 1 && Pitch > width */
@@ -92,7 +107,13 @@ static std::vector<uur::TestParameters2D> test_cases{
     {234, 233, 1}};
 
 UUR_TEST_SUITE_P(urEnqueueUSMMemcpy2DTestWithParam,
-                 ::testing::ValuesIn(test_cases),
+                 ::testing::Combine(::testing::ValuesIn(test_sizes),
+                                    ::testing::Values(uur::USMKind::Device,
+                                                      uur::USMKind::Host,
+                                                      uur::USMKind::Shared),
+                                    ::testing::Values(uur::USMKind::Device,
+                                                      uur::USMKind::Host,
+                                                      uur::USMKind::Shared)),
                  uur::print2DTestString<urEnqueueUSMMemcpy2DTestWithParam>);
 
 TEST_P(urEnqueueUSMMemcpy2DTestWithParam, SuccessBlocking) {
@@ -119,7 +140,8 @@ TEST_P(urEnqueueUSMMemcpy2DTestWithParam, SuccessNonBlocking) {
 
 using urEnqueueUSMMemcpy2DNegativeTest = urEnqueueUSMMemcpy2DTestWithParam;
 UUR_TEST_SUITE_P(urEnqueueUSMMemcpy2DNegativeTest,
-                 ::testing::Values(uur::TestParameters2D{1, 1, 1}),
+                 ::testing::Values(TestParametersMemcpy2D{
+                     {1, 1, 1}, uur::USMKind::Device, uur::USMKind::Device}),
                  uur::print2DTestString<urEnqueueUSMMemcpy2DTestWithParam>);
 
 TEST_P(urEnqueueUSMMemcpy2DNegativeTest, InvalidNullHandleQueue) {

test/conformance/testing/include/uur/fixtures.h

@@ -1071,7 +1071,8 @@ struct urProgramTest : urQueueTest {
 template <class T> struct urProgramTestWithParam : urContextTestWithParam<T> {
     void SetUp() override {
         UUR_RETURN_ON_FATAL_FAILURE(urContextTestWithParam<T>::SetUp());
-        uur::KernelsEnvironment::instance->LoadSource("foo", 0, il_binary);
+        uur::KernelsEnvironment::instance->LoadSource(program_name, 0,
+                                                      il_binary);
         ASSERT_SUCCESS(urProgramCreateWithIL(this->context, il_binary->data(),
                                              il_binary->size(), nullptr,
                                              &program));

test/conformance/testing/include/uur/utils.h

@@ -378,6 +378,18 @@ ur_device_partition_property_t makePartitionEquallyDesc(uint32_t cu_per_device);
 ur_device_partition_property_t
 makePartitionByAffinityDomain(ur_device_affinity_domain_flags_t aff_domain);
 
+enum class USMKind {
+    Device,
+    Host,
+    Shared,
+};
+
+ur_result_t MakeUSMAllocationByType(USMKind kind, ur_context_handle_t hContext,
+                                    ur_device_handle_t hDevice,
+                                    const ur_usm_desc_t *pUSMDesc,
+                                    ur_usm_pool_handle_t hPool, size_t size,
+                                    void **ppMem);
+
 } // namespace uur
 
 #endif // UR_CONFORMANCE_INCLUDE_UTILS_H_INCLUDED

test/conformance/testing/source/utils.cpp

@@ -658,4 +658,22 @@ makePartitionByAffinityDomain(ur_device_affinity_domain_flags_t aff_domain) {
     return desc;
 }
 
+ur_result_t MakeUSMAllocationByType(USMKind kind, ur_context_handle_t hContext,
+                                    ur_device_handle_t hDevice,
+                                    const ur_usm_desc_t *pUSMDesc,
+                                    ur_usm_pool_handle_t hPool, size_t size,
+                                    void **ppMem) {
+    switch (kind) {
+    case USMKind::Device:
+        return urUSMDeviceAlloc(hContext, hDevice, pUSMDesc, hPool, size,
+                                ppMem);
+    case USMKind::Host:
+        return urUSMHostAlloc(hContext, pUSMDesc, hPool, size, ppMem);
+    default:
+    case USMKind::Shared:
+        return urUSMSharedAlloc(hContext, hDevice, pUSMDesc, hPool, size,
+                                ppMem);
+    }
+}
+
 } // namespace uur