[SYCL][ROCm] Setup lit tests for ROCm plugin

npmiller · npmiller · commit 84424ae07a29 · 2021-08-03T11:03:21.000+01:00
diff --git a/sycl/doc/GetStartedGuide.md b/sycl/doc/GetStartedGuide.md
@@ -424,6 +424,10 @@ skipped.
 If CUDA support has been built, it is tested only if there are CUDA devices
 available.
 
+If testing with ROCm for AMD make sure to specify the GPU being used
+by adding `-mcpu=<target>` to the CMake variable
+`SYCL_CLANG_EXTRA_FLAGS`.
+
 #### Run DPC++ E2E test suite
 
 Follow instructions from the link below to build and run tests:
diff --git a/sycl/test/CMakeLists.txt b/sycl/test/CMakeLists.txt
@@ -75,6 +75,34 @@ if(SYCL_BUILD_PI_CUDA)
   add_custom_target(check-sycl-cuda)
   add_dependencies(check-sycl-cuda check-sycl-ptx)
   add_dependencies(check-sycl check-sycl-cuda)
+endif()
+
+if(SYCL_BUILD_PI_ROCM)
+  add_custom_target(check-sycl-rocm)
+  if("${SYCL_BUILD_PI_ROCM_PLATFORM}" STREQUAL "NVIDIA")
+    add_lit_testsuite(check-sycl-rocm-ptx "Running device-agnostic SYCL regression tests for ROCm NVidia PTX"
+      ${CMAKE_CURRENT_BINARY_DIR}
+      ARGS ${RT_TEST_ARGS}
+      PARAMS "SYCL_TRIPLE=nvptx64-nvidia-cuda-sycldevice;SYCL_PLUGIN=rocm"
+      DEPENDS ${SYCL_TEST_DEPS}
+      EXCLUDE_FROM_CHECK_ALL
+    )
+
+    add_dependencies(check-sycl-rocm check-sycl-rocm-ptx)
+  elseif("${SYCL_BUILD_PI_ROCM_PLATFORM}" STREQUAL "AMD")
+    add_lit_testsuite(check-sycl-rocm-gcn "Running device-agnostic SYCL regression tests for ROCm AMDGCN"
+      ${CMAKE_CURRENT_BINARY_DIR}
+      ARGS ${RT_TEST_ARGS}
+      PARAMS "SYCL_TRIPLE=amdgcn-amd-amdhsa-sycldevice;SYCL_PLUGIN=rocm"
+      DEPENDS ${SYCL_TEST_DEPS}
+      EXCLUDE_FROM_CHECK_ALL
+    )
+
+    add_dependencies(check-sycl-rocm check-sycl-rocm-gcn)
+  else()
+    message(FATAL_ERROR "SYCL_BUILD_PI_ROCM_PLATFORM must be set to either 'AMD' or 'NVIDIA' (set to: '${SYCL_BUILD_PI_ROCM_PLATFORM}')")
+  endif()
 
+  add_dependencies(check-sycl check-sycl-rocm)
 endif()
 add_subdirectory(on-device)
diff --git a/sycl/test/atomic_ref/accessor.cpp b/sycl/test/atomic_ref/accessor.cpp
@@ -1,5 +1,8 @@
 // RUN: %clangxx -fsycl -fsycl-unnamed-lambda -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %RUN_ON_HOST %t.out
+//
+// Missing __clc_atomic_store_local_4_unordered on AMD
+// XFAIL: rocm_amd
 
 #include <CL/sycl.hpp>
 #include <algorithm>
diff --git a/sycl/test/atomic_ref/add.cpp b/sycl/test/atomic_ref/add.cpp
@@ -4,6 +4,10 @@
 // RUN:  -fsycl-device-only -S %s -o - | FileCheck %s --check-prefix=CHECK-LLVM-EMU
 // RUN: %clangxx -fsycl -fsycl-unnamed-lambda -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %RUN_ON_HOST %t.out
+//
+// Missing __clc__atomic_uload_global_4_unordered,
+// __clc__atomic_uload_global_8_unordered with AMD:
+// XFAIL: rocm_amd
 
 #include <CL/sycl.hpp>
 #include <algorithm>
diff --git a/sycl/test/atomic_ref/load.cpp b/sycl/test/atomic_ref/load.cpp
@@ -2,6 +2,12 @@
 // RUN: | FileCheck %s --check-prefix=CHECK-LLVM
 // RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %RUN_ON_HOST %t.out
+//
+// Missing __clc__atomic_load_global_4_unordered,
+// __clc__atomic_uload_global_4_unordered,
+// __clc__atomic_load_global_8_unordered and
+// __clc__atomic_uload_global_8_unordereud on AMD:
+// XFAIL: rocm_amd
 
 #include <CL/sycl.hpp>
 #include <algorithm>
diff --git a/sycl/test/atomic_ref/max.cpp b/sycl/test/atomic_ref/max.cpp
@@ -4,6 +4,11 @@
 // RUN:  -fsycl-device-only -S %s -o - | FileCheck %s --check-prefix=CHECK-LLVM-EMU
 // RUN: %clangxx -fsycl -fsycl-unnamed-lambda -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %RUN_ON_HOST %t.out
+//
+// Missing __clc__sync_fetch_and_max_global_8,
+// __clc__sync_fetch_and_umax_global_8, __clc__atomic_uload_global_4_unordered,
+// __clc__atomic_uload_global_8_unordered with AMD:
+// XFAIL: rocm_amd
 
 #include <CL/sycl.hpp>
 #include <algorithm>
diff --git a/sycl/test/atomic_ref/min.cpp b/sycl/test/atomic_ref/min.cpp
@@ -4,6 +4,11 @@
 // RUN:  -fsycl-device-only -S %s -o - | FileCheck %s --check-prefix=CHECK-LLVM-EMU
 // RUN: %clangxx -fsycl -fsycl-unnamed-lambda -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %RUN_ON_HOST %t.out
+//
+// Missing __clc__sync_fetch_and_min_global_8,
+// __clc__sync_fetch_and_umin_global_8, __clc__atomic_uload_global_4_unordered,
+// __clc__atomic_uload_global_8_unordered with AMD:
+// XFAIL: rocm_amd
 
 #include <CL/sycl.hpp>
 #include <algorithm>
diff --git a/sycl/test/atomic_ref/store.cpp b/sycl/test/atomic_ref/store.cpp
@@ -2,6 +2,12 @@
 // RUN: | FileCheck %s --check-prefix=CHECK-LLVM
 // RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %RUN_ON_HOST %t.out
+//
+// Missing __clc__atomic_store_global_4_unordered,
+// __clc__atomic_ustore_global_4_unordered,
+// __clc__atomic_store_global_8_unordered and
+// __clc__atomic_ustore_global_8_unordereud on AMD:
+// XFAIL: rocm_amd
 
 #include <CL/sycl.hpp>
 #include <algorithm>
diff --git a/sycl/test/atomic_ref/sub.cpp b/sycl/test/atomic_ref/sub.cpp
@@ -4,6 +4,10 @@
 // RUN:  -fsycl-device-only -S %s -o - | FileCheck %s --check-prefix=CHECK-LLVM-EMU
 // RUN: %clangxx -fsycl -fsycl-unnamed-lambda -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %RUN_ON_HOST %t.out
+//
+// Missing __clc__atomic_uload_global_4_unordered,
+// __clc__atomic_uload_global_8_unordered with AMD:
+// XFAIL: rocm_amd
 
 #include <CL/sycl.hpp>
 #include <algorithm>
diff --git a/sycl/test/basic_tests/built-ins.cpp b/sycl/test/basic_tests/built-ins.cpp
@@ -3,6 +3,10 @@
 
 // CUDA does not support printf.
 // UNSUPPORTED: cuda
+//
+// Hits an assertion with AMD:
+// XFAIL: rocm_amd
+
 #include <CL/sycl.hpp>
 
 #include <cassert>
diff --git a/sycl/test/esimd/odr.cpp b/sycl/test/esimd/odr.cpp
@@ -11,6 +11,9 @@
 // Cuda does not support intrinsics generated by the ESIMD compilation path:
 // UNSUPPORTED: cuda
 //
+// Linking issues with AMD:
+// XFAIL: rocm_amd
+
 #include <CL/sycl.hpp>
 #include <iostream>
 #include <sycl/ext/intel/experimental/esimd.hpp>
diff --git a/sycl/test/extensions/group-algorithm.cpp b/sycl/test/extensions/group-algorithm.cpp
@@ -1,6 +1,10 @@
 // RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -I . -o %t.out
 // Group operations are not supported on host device. The test checks that
 // compilation succeeded.
+//
+// Missing __spirv_GroupIAdd, __spirv_GroupAll, __spirv_GroupBroadcast,
+// __spirv_GroupAny, __spirv_GroupSMin on AMD:
+// XFAIL: rocm_amd
 
 // TODO: enable compile+runtime checks for operations defined in SPIR-V 1.3.
 // That requires either adding a switch to clang (-spirv-max-version=1.3) or
diff --git a/sycl/test/lit.cfg.py b/sycl/test/lit.cfg.py
@@ -103,6 +103,14 @@
 if triple == 'nvptx64-nvidia-cuda-sycldevice':
     config.available_features.add('cuda')
 
+if triple == 'amdgcn-amd-amdhsa-sycldevice':
+    config.available_features.add('rocm_amd')
+    # For AMD the specific GPU has to be specified with -mcpu
+    if not re.match('.*-mcpu.*', config.sycl_clang_extra_flags):
+        raise Exception("Error: missing -mcpu flag when trying to run lit "  \
+                        "tests for AMD GPU, please add `-mcpu=<target>` to " \
+                        "the CMake variable SYCL_CLANG_EXTRA_FLAGS")
+
 # Set timeout for test = 10 mins
 try:
     import psutil
diff --git a/sycl/test/on-device/CMakeLists.txt b/sycl/test/on-device/CMakeLists.txt
@@ -40,3 +40,17 @@ if(SYCL_BUILD_PI_CUDA)
     add_dependencies(check-sycl-cuda check-sycl-cuda-on-device)
   endif()
 endif()
+
+if(SYCL_BUILD_PI_ROCM)
+  add_lit_testsuite(check-sycl-rocm-on-device "Running the SYCL regression tests for ROCm"
+    ${CMAKE_CURRENT_BINARY_DIR}
+    ARGS ${RT_TEST_ARGS}
+    PARAMS "SYCL_PLUGIN=rocm"
+    DEPENDS ${SYCL_TEST_DEPS}
+    EXCLUDE_FROM_CHECK_ALL
+  )
+  set_target_properties(check-sycl-rocm-on-device PROPERTIES FOLDER "SYCL ROCm device tests")
+  if(TARGET check-sycl-rocm)
+    add_dependencies(check-sycl-rocm check-sycl-rocm-on-device)
+  endif()
+endif()
diff --git a/sycl/test/on-device/back_to_back_collectives.cpp b/sycl/test/on-device/back_to_back_collectives.cpp
@@ -3,6 +3,9 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
+//
+// Missing __spirv_GroupIAdd on AMD:
+// XFAIL: rocm_amd
 
 #include <CL/sycl.hpp>
 #include <numeric>
diff --git a/sycl/test/on-device/basic_tests/specialization_constants/host_apis.cpp b/sycl/test/on-device/basic_tests/specialization_constants/host_apis.cpp
@@ -2,6 +2,8 @@
 // RUN: %t.out
 
 // UNSUPPORTED: cuda
+// UNSUPPORTED: rocm_nvidia
+// UNSUPPORTED: rocm_amd
 
 #include <sycl/sycl.hpp>
 
diff --git a/sycl/test/on-device/basic_tests/specialization_constants/kernel_lambda_with_kernel_handler_arg.cpp b/sycl/test/on-device/basic_tests/specialization_constants/kernel_lambda_with_kernel_handler_arg.cpp
@@ -1,5 +1,8 @@
 // RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
+//
+// Hits an assert in the Lower Work Group Scope Code pass on AMD:
+// XFAIL: rocm_amd
 
 // This test checks all possible scenarios of running single_task, parallel_for
 // and parallel_for_work_group to verify that this code compiles and runs
diff --git a/sycl/test/on-device/basic_tests/specialization_constants/non_native/aot_w_kernel_handler_wo_spec_consts.cpp b/sycl/test/on-device/basic_tests/specialization_constants/non_native/aot_w_kernel_handler_wo_spec_consts.cpp
@@ -1,5 +1,8 @@
 // RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
+//
+// Hits an assert in the Lower Work Group Scope Code pass on AMD:
+// XFAIL: rocm_amd
 
 // This test checks correctness of compiling and running of application with
 // kernel lambdas containing kernel_handler arguments and w/o usage of
diff --git a/sycl/test/on-device/basic_tests/specialization_constants/non_native/gpu.cpp b/sycl/test/on-device/basic_tests/specialization_constants/non_native/gpu.cpp
@@ -1,6 +1,10 @@
 // REQUIRES: ocloc, gpu
 // UNSUPPORTED: cuda
 // CUDA is not compatible with SPIR.
+//
+// UNSUPPORTED: rocm_nvidia
+// UNSUPPORTED: rocm_amd
+// ROCm is not compatible with SPIR.
 
 // RUN: %clangxx -fsycl -fsycl-targets=spir64_gen-unknown-unknown-sycldevice -Xsycl-target-backend=spir64_gen-unknown-unknown-sycldevice "-device *" %S/Inputs/common.cpp -o %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
diff --git a/sycl/test/on-device/extensions/intel-ext-device.cpp b/sycl/test/on-device/extensions/intel-ext-device.cpp
@@ -4,6 +4,8 @@
 //
 // REQUIRES: gpu
 // UNSUPPORTED: cuda
+// UNSUPPORTED: rocm_nvidia
+// UNSUPPORTED: rocm_amd
 
 //==--------- intel-ext-device.cpp - SYCL device test ------------==//
 //
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/all_of.cpp b/sycl/test/on-device/group_algorithms_sycl2020/all_of.cpp
@@ -2,6 +2,9 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
+//
+// Missing __spirv_GroupAll on AMD:
+// XFAIL: rocm_amd
 
 #include "support.h"
 #include <CL/sycl.hpp>
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/any_of.cpp b/sycl/test/on-device/group_algorithms_sycl2020/any_of.cpp
@@ -2,6 +2,9 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
+//
+// Missing __spirv_GroupAny on AMD:
+// XFAIL: rocm_amd
 
 #include "support.h"
 #include <CL/sycl.hpp>
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/exclusive_scan.cpp b/sycl/test/on-device/group_algorithms_sycl2020/exclusive_scan.cpp
@@ -2,6 +2,10 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
+//
+// Missing __spirv_GroupIAdd, __spirv_GroupBroadcast, __spirv_GroupSMin and
+// __spirv_GroupSMax on AMD:
+// XFAIL: rocm_amd
 
 // TODO: enable compile+runtime checks for operations defined in SPIR-V 1.3.
 // That requires either adding a switch to clang (-spirv-max-version=1.3) or
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/group_broadcast.cpp b/sycl/test/on-device/group_algorithms_sycl2020/group_broadcast.cpp
@@ -2,6 +2,9 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
+//
+// Missing __spirv_GroupBroadcast on AMD:
+// XFAIL: rocm_amd
 
 #include "support.h"
 #include <CL/sycl.hpp>
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/inclusive_scan.cpp b/sycl/test/on-device/group_algorithms_sycl2020/inclusive_scan.cpp
@@ -2,6 +2,10 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
+//
+// Missing __spirv_GroupIAdd, __spirv_GroupBroadcast, __spirv_GroupSMin and
+// __spirv_GroupSMax on AMD:
+// XFAIL: rocm_amd
 
 // TODO: enable compile+runtime checks for operations defined in SPIR-V 1.3.
 // That requires either adding a switch to clang (-spirv-max-version=1.3) or
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/none_of.cpp b/sycl/test/on-device/group_algorithms_sycl2020/none_of.cpp
@@ -2,6 +2,9 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
+//
+// Missing __spirv_GroupAll and __spirv_GroupAny on AMD:
+// XFAIL: rocm_amd
 
 #include "support.h"
 #include <CL/sycl.hpp>
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/permute_select.cpp b/sycl/test/on-device/group_algorithms_sycl2020/permute_select.cpp
@@ -3,6 +3,9 @@
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
 //
+// Missing __spirv_SubgroupId, __spirv_SubgroupMaxSize, __spirv_SubgroupShuffle* on AMD:
+// XFAIL: rocm_amd
+//
 //==------------ permute_select.cpp -*- C++ -*-----------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/reduce.cpp b/sycl/test/on-device/group_algorithms_sycl2020/reduce.cpp
@@ -2,6 +2,9 @@
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
+//
+// Missinsg __spirv_GroupIAdd, __spirv_GroupSMin and __spirv_GroupSMax on AMD:
+// XFAIL: rocm_amd
 
 // TODO: enable compile+runtime checks for operations defined in SPIR-V 1.3.
 // That requires either adding a switch to clang (-spirv-max-version=1.3) or
diff --git a/sycl/test/on-device/group_algorithms_sycl2020/shift_left_right.cpp b/sycl/test/on-device/group_algorithms_sycl2020/shift_left_right.cpp
@@ -3,6 +3,9 @@
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
 // RUN: %ACC_RUN_PLACEHOLDER %t.out
 //
+// Missing __spirv_SubgroupId, __spirv_SubgroupMaxSize, __spirv_SubgroupShuffle* on AMD:
+// XFAIL: rocm_amd
+//
 //==------------ shift_left_right.cpp -*- C++ -*----------------------------==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
diff --git a/sycl/test/on-device/lit.cfg.py b/sycl/test/on-device/lit.cfg.py
diff --git a/sycl/test/on-device/srgb/srgba-read.cpp b/sycl/test/on-device/srgb/srgba-read.cpp
diff --git a/sycl/test/regression/group.cpp b/sycl/test/regression/group.cpp
diff --git a/sycl/test/regression/implicit_atomic_conversion.cpp b/sycl/test/regression/implicit_atomic_conversion.cpp
diff --git a/sycl/test/separate-compile/sycl-external.cpp b/sycl/test/separate-compile/sycl-external.cpp
diff --git a/sycl/unittests/SYCL2020/KernelBundle.cpp b/sycl/unittests/SYCL2020/KernelBundle.cpp
diff --git a/sycl/unittests/SYCL2020/SpecConstDefaultValues.cpp b/sycl/unittests/SYCL2020/SpecConstDefaultValues.cpp
diff --git a/sycl/unittests/misc/KernelBuildOptions.cpp b/sycl/unittests/misc/KernelBuildOptions.cpp
diff --git a/sycl/unittests/pi/BackendString.hpp b/sycl/unittests/pi/BackendString.hpp
diff --git a/sycl/unittests/pi/EnqueueMemTest.cpp b/sycl/unittests/pi/EnqueueMemTest.cpp
diff --git a/sycl/unittests/program_manager/itt_annotations.cpp b/sycl/unittests/program_manager/itt_annotations.cpp
diff --git a/sycl/unittests/scheduler/RequiredWGSize.cpp b/sycl/unittests/scheduler/RequiredWGSize.cpp
