Merge branch 'sycl' into georgi/test-hip-usm-copy2d

Author: aarongreig

.github/workflows/sycl-linux-run-tests.yml

@@ -307,6 +307,9 @@ jobs:
       if: inputs.tests_selector == 'cts'
       env:
         ONEAPI_DEVICE_SELECTOR: ${{ inputs.target_devices }}
+      # This job takes ~100min usually. But sometimes some test isn't
+      # responding, so the job reaches the 360min limit. Setting a lower one.
+      timeout-minutes: 150
       # By-default GitHub actions execute the "run" shell script with -e option,
       # so the execution terminates if any command returns a non-zero status.
       # Since we're using a loop to run all test-binaries separately, some test
@@ -340,6 +343,7 @@ jobs:
         done
         if [ -n "$status" ]; then
           echo "Failed suite(s): $failed_suites"
+          echo "Failed suite(s): $failed_suites" >> $GITHUB_STEP_SUMMARY
           exit 1
         fi
         exit 0

clang-tools-extra/clangd/test/sycl.test

@@ -25,11 +25,11 @@
 # CHECK-NEXT:        "range": {
 # CHECK-NEXT:          "end": {
 # CHECK-NEXT:            "character": 16,
-# CHECK-NEXT:            "line": 124
+# CHECK-NEXT:            "line": 116
 # CHECK-NEXT:          },
 # CHECK-NEXT:          "start": {
 # CHECK-NEXT:            "character": 11,
-# CHECK-NEXT:            "line": 124
+# CHECK-NEXT:            "line": 116
 # CHECK-NEXT:          }
 # CHECK-NEXT:        },
 # CHECK-NEXT:        "uri": "file://{{.*}}/include/sycl/queue.hpp"

clang/lib/Driver/ToolChains/MSVC.cpp

@@ -135,23 +135,18 @@ void visualstudio::Linker::ConstructJob(Compilation &C, const JobAction &JA,
     CmdArgs.push_back("-defaultlib:oldnames");
   }
 
-  if ((Args.hasArg(options::OPT_fsycl) &&
+  if ((!C.getDriver().IsCLMode() && Args.hasArg(options::OPT_fsycl) &&
        !Args.hasArg(options::OPT_nolibsycl)) ||
       Args.hasArg(options::OPT_fsycl_host_compiler_EQ)) {
     CmdArgs.push_back(Args.MakeArgString(std::string("-libpath:") +
                                          TC.getDriver().Dir + "/../lib"));
-    if (!Args.hasArg(options::OPT__SLASH_MDd) &&
-        !isDependentLibAdded(Args, "msvcrtd")) {
+    // When msvcrtd is added via --dependent-lib, we add the sycld
+    // equivalent.  Do not add the -defaultlib as it conflicts.
+    if (!isDependentLibAdded(Args, "msvcrtd")) {
       if (Args.hasArg(options::OPT_fpreview_breaking_changes))
         CmdArgs.push_back("-defaultlib:sycl" SYCL_MAJOR_VERSION "-preview.lib");
       else
         CmdArgs.push_back("-defaultlib:sycl" SYCL_MAJOR_VERSION ".lib");
-    } else {
-      if (Args.hasArg(options::OPT_fpreview_breaking_changes))
-        CmdArgs.push_back("-defaultlib:sycl" SYCL_MAJOR_VERSION
-                          "-previewd.lib");
-      else
-        CmdArgs.push_back("-defaultlib:sycl" SYCL_MAJOR_VERSION "d.lib");
     }
     CmdArgs.push_back("-defaultlib:sycl-devicelib-host.lib");
   }

clang/test/Driver/sycl-offload.c

@@ -486,7 +486,7 @@
 // RUN: %clang -fsycl -target x86_64-unknown-windows-msvc %s -o %t -### 2>&1 | FileCheck -check-prefix=CHECK-LINK-SYCL %s
 // RUN: %clang_cl -fsycl %s -o %t -### 2>&1 | FileCheck -check-prefix=CHECK-LINK-SYCL-CL %s
 // CHECK-LINK-SYCL-CL: "--dependent-lib=sycl{{[0-9]*}}"
-// CHECK-LINK-SYCL-CL: "-defaultlib:sycl{{[0-9]*}}.lib"
+// CHECK-LINK-SYCL-CL-NOT: "-defaultlib:sycl{{[0-9]*}}.lib"
 // CHECK-LINK-SYCL: "-defaultlib:sycl{{[0-9]*}}.lib"
 
 /// Check no SYCL runtime is linked with -nolibsycl
@@ -837,16 +837,3 @@
 // FSYCL-PREVIEW-BREAKING-CHANGES-DEBUG-CHECK: --dependent-lib=sycl{{[0-9]*}}-previewd
 // FSYCL-PREVIEW-BREAKING-CHANGES-DEBUG-CHECK-NOT: -defaultlib:sycl{{[0-9]*}}.lib
 // FSYCL-PREVIEW-BREAKING-CHANGES-DEBUG-CHECK-NOT: -defaultlib:sycl{{[0-9]*}}-preview.lib
-
-
-/// Check that at link step of "clang-cl -fsycl" we pull in sycl.lib even if at the compilation step sycl libraries were not provided (this is possible if user compiles manually without -fsycl by provided paths to the headers).
-// RUN: %clang_cl -### -fsycl -nolibsycl -target x86_64-unknown-windows-msvc -c %s 2>&1 | FileCheck -check-prefix FSYCL-CL-COMPILE-NOLIBS-CHECK %s
-// RUN: %clang_cl -### -fsycl %s 2>&1 | FileCheck -check-prefix FSYCL-CL-LINK-CHECK %s
-// FSYCL-CL-COMPILE-NOLIBS-CHECK-NOT: "--dependent-lib=sycl{{[0-9]*}}"
-// FSYCL-CL-LINK-CHECK: "-defaultlib:sycl{{[0-9]*}}.lib"
-
-/// Check that at link step of "clang-cl -fsycl /MDd" we pull in sycld.lib even if at the compilation step sycl libraries were not provided (this is possible if user compiles manually without -fsycl by provided paths to the headers).
-// RUN: %clang_cl -### -fsycl -nolibsycl /MDd -target x86_64-unknown-windows-msvc -c %s 2>&1 | FileCheck -check-prefix FSYCL-CL-COMPILE-NOLIBS-MDd-CHECK %s
-// RUN: %clang_cl -### -fsycl /MDd %s 2>&1 | FileCheck -check-prefix FSYCL-CL-LINK--MDd-CHECK %s
-// FSYCL-CL-COMPILE-NOLIBS-MDd-CHECK-NOT: "--dependent-lib=sycl{{[0-9]*}}d"
-// FSYCL-CL-LINK--MDd-CHECK: "-defaultlib:sycl{{[0-9]*}}d.lib"

devops/cts_exclude_filter

@@ -1,5 +1,5 @@
 # These two take too much time
 kernel_bundle
 marray
-# https://github.com/intel/llvm/issues/12926
-h_item
+# https://github.com/intel/llvm/issues/13332
+hierarchical

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

@@ -1328,7 +1328,7 @@ static bool isUnsupportedSPIRAccess(Value *Addr, Function *Func) {
   }
 
   // All the rest address spaces: skip SPIR-V built-in varibles
-  auto *OrigValue = Addr->stripPointerCasts();
+  auto *OrigValue = Addr->stripInBoundsOffsets();
   return OrigValue->getName().starts_with("__spirv_BuiltIn");
 }
 

llvm/test/Instrumentation/AddressSanitizer/spir_skip_builtin.ll

@@ -0,0 +1,17 @@
+; RUN: opt < %s -passes=asan -asan-instrumentation-with-call-threshold=0 -asan-stack=0 -asan-globals=0 -S | FileCheck %s
+
+target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-n8:16:32:64"
+target triple = "spir64-unknown-unknown"
+
+@__spirv_BuiltInGlobalInvocationId = external addrspace(1) constant <3 x i64>
+
+; Function Attrs: sanitize_address
+define spir_kernel void @_ZTSN4sycl3_V16detail19__pf_kernel_wrapperIZZ4mainENKUlRNS0_7handlerEE_clES4_E19bufferByRange2_initEE() #0 {
+entry:
+  ; CHECK-NOT: {{ __asan_load8.*__spirv_BuiltInGlobalInvocationId }}
+  %0 = load i64, ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) @__spirv_BuiltInGlobalInvocationId, i64 8), align 8
+  ret void
+}
+
+attributes #0 = { sanitize_address }
+

sycl/plugins/unified_runtime/CMakeLists.txt

@@ -110,17 +110,23 @@ if(SYCL_PI_UR_USE_FETCH_CONTENT)
 
   fetch_adapter_source(opencl
     ${UNIFIED_RUNTIME_REPO}
-    # Merge: e60c3c22 9287547e
+    # commit 85b7559033e329389452cb87615ad42fbb644d59 (HEAD -> main, origin/main, origin/HEAD)
+    # Merge: a7c202b4 5c300799
     # Author: Kenneth Benzie (Benie) <[email protected]>
-    # Date:   Thu Apr 4 10:23:33 2024 +0200
-    #     Merge pull request #1448 from steffenlarsen/steffen/make_ext_func_fail_unsupported
-    #     [OpenCL] Make extension function lookup return unusupported error
-    3609afc7f8781f2eae5de74deaf50ba52b1bb344
+    # Date:   Tue Apr 9 14:06:49 2024 +0100
+    #     Merge pull request #1496 from steffenlarsen/steffen/revert_memory_lookup_error
+    #     [OpenCL] Revert urMemBufferCreate extension function lookup error
+    85b7559033e329389452cb87615ad42fbb644d59
   )
 
   fetch_adapter_source(cuda
     ${UNIFIED_RUNTIME_REPO}
-    ${UNIFIED_RUNTIME_TAG}
+    # Merge: ecdd159f c02d137e
+    # Author: Kenneth Benzie (Benie) <[email protected]>
+    # Date:   Wed Apr 10 11:04:28 2024 +0100
+    #     Merge pull request #1181 from DuncanMcBain/cuda-mem-size-fix
+    #     Return device total global memory for MaxAllocSize
+    05b5899202f7e3b3f12f4c5b8fc8f01451d08105
   )
 
   fetch_adapter_source(hip

sycl/test-e2e/ESIMD/srnd.cpp

@@ -14,6 +14,7 @@
 #include <sycl/ext/intel/esimd.hpp>
 #include <sycl/sycl.hpp>
 
+#include <cmath>
 #include <iostream>
 
 using namespace sycl;

sycl/test-e2e/Matrix/element_wise_all_ops.cpp

@@ -6,9 +6,6 @@
 //
 //===----------------------------------------------------------------------===//
 // REQUIRES: aspect-ext_intel_matrix
-// Test is flaky/timeouts on some variants of DG2 and temporary disabled. Needs
-// to be investigated.
-// UNSUPPORTED: gpu-intel-dg2
 
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out

sycl/test-e2e/Matrix/joint_matrix_apply_two_matrices_impl.hpp

@@ -28,7 +28,7 @@ bool apply_verify(Tc *C, Tc *D, Ta *A, Ta *Ar) {
   return true;
 }
 template <typename Tc, typename Ta, size_t TM, size_t TN, size_t TK, size_t M,
-          size_t N, class kernel_name>
+          size_t N, size_t K, class kernel_name>
 bool apply_two_matrices(Tc *C, Tc *D, Ta *A, Ta *Ar, queue q) {
   size_t NDRangeM = M / TM;
   size_t NDRangeN = N / TN;
@@ -76,13 +76,13 @@ bool apply_two_matrices(Tc *C, Tc *D, Ta *A, Ta *Ar, queue q) {
                sg, sub_d, pD + (sg_startx * TM) * N + sg_starty / sg_size * TN,
                N, layout::row_major);
            joint_matrix_load(
-               sg, sub_a, pA + (sg_startx * TM) * N + sg_starty / sg_size * TK,
-               N);
+               sg, sub_a, pA + (sg_startx * TM) * K + sg_starty / sg_size * TK,
+               K);
            joint_matrix_apply(sg, sub_a, sub_ar,
                               [](const Ta &x, Ta &y) { y = x + 42; });
            ext::intel::experimental::matrix::joint_matrix_store(
                sg, sub_ar,
-               pAr + (sg_startx * TM) * N + sg_starty / sg_size * TK, N);
+               pAr + (sg_startx * TM) * K + sg_starty / sg_size * TK, K);
          }); // parallel for
    }).wait();
   return apply_verify<Tc, Ta, M, N>(C, D, A, Ar);
@@ -91,27 +91,27 @@ bool apply_two_matrices(Tc *C, Tc *D, Ta *A, Ta *Ar, queue q) {
 template <typename Ta, typename Tc, size_t TM, size_t TN, size_t TK,
           class kernel_name>
 bool test() {
-
   static constexpr size_t M = TM * 2;
   static constexpr size_t N = TN * 2;
+  static constexpr size_t K = TK * 2;
   queue q;
 
   Tc *C = malloc_shared<Tc>(M * N, q);
   Tc *D = malloc_shared<Tc>(M * N, q);
-  Ta *A = malloc_shared<Ta>(M * N, q);
-  Ta *Ar = malloc_shared<Ta>(M * N, q);
+  Ta *A = malloc_shared<Ta>(M * K, q);
+  Ta *Ar = malloc_shared<Ta>(M * K, q);
 
   matrix_rand(M, N, (Tc *)C, (Tc)100);
-  matrix_rand(M, N, (Ta *)A, (Ta)100);
+  matrix_rand(M, K, (Ta *)A, (Ta)100);
 
-  bool res =
-      apply_two_matrices<Tc, Ta, TM, TN, TK, M, N, kernel_name>(C, D, A, Ar, q);
+  bool res = apply_two_matrices<Tc, Ta, TM, TN, TK, M, N, K, kernel_name>(
+      C, D, A, Ar, q);
 
   if constexpr (std::is_same_v<Ta, bfloat16>)
-    std::cout << "bfloat16 " << TM << "x" << TN << ": "
+    std::cout << "bfloat16 " << TM << "x" << TN << "x" << TK << ": "
               << (res ? "passed" : "failed") << std::endl;
   else if constexpr (std::is_same_v<Ta, int8_t>)
-    std::cout << "int8_t " << TM << "x" << TN << ": "
+    std::cout << "int8_t " << TM << "x" << TN << "x" << TK << ": "
               << (res ? "passed" : "failed") << std::endl;
   return res;
 }
@@ -126,8 +126,8 @@ int main() {
   bool passed = true;
   for (unsigned int i = 0; i < combinations.size(); i++) {
     if (combinations[i].nsize == 0) { // Intel AMX
-      passed &= test<int8_t, int32_t, 8, 16, 32, class amx_int_8x16x32>();
-      passed &= test<bfloat16, float, 8, 16, 32, class amx_bf16_8x16x32>();
+      passed &= test<int8_t, int32_t, 16, 16, 64, class amx_int_16x16x64>();
+      passed &= test<bfloat16, float, 16, 16, 32, class amx_bf16_16x16x32>();
       break;
     }