Merge branch 'sycl' of https://github.com/intel/llvm into benchmarking-workflow

Author: ianayl

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

@@ -363,7 +363,9 @@ jobs:
         cts_exclude_filter=""
         # If CTS_TESTS_TO_BUILD is null - use filter
         if [ -z "$CTS_TESTS_TO_BUILD" ]; then
-          if [ "${{ contains(inputs.target_devices, 'opencl:cpu')  }}" = "true" ]; then
+          if [ "${{ contains(inputs.cts_testing_mode, 'build-only')  }}" = "true" ]; then
+            cts_exclude_filter=$PWD/devops/cts_exclude_filter_compfails
+          elif [ "${{ contains(inputs.target_devices, 'opencl:cpu')  }}" = "true" ]; then
             cts_exclude_filter=$PWD/devops/cts_exclude_filter_OCL_CPU
           elif [ "${{ contains(inputs.target_devices, 'level_zero:gpu')  }}" = "true" ]; then
             cts_exclude_filter=$PWD/devops/cts_exclude_filter_L0_GPU

.github/workflows/sycl-nightly.yml

@@ -242,7 +242,7 @@ jobs:
         body: "Daily build ${{ steps.tag.outputs.TAG }}"
         target_commitish: ${{ github.sha }}
 
-  ubuntu2204_docker_build_push:
+  docker_build_push:
     if: github.repository == 'intel/llvm'
     runs-on: [Linux, build]
     permissions:
@@ -254,42 +254,16 @@ jobs:
       with:
         name: sycl_linux_default
         path: devops/
-    - name: Build and Push Container (with drivers)
+    - name: Build and Push Container
       uses: ./devops/actions/build_container
       with:
         push: ${{ github.ref_name == 'sycl' }}
-        file: ubuntu2204_preinstalled
+        file: nightly
         username: ${{ github.repository_owner }}
         password: ${{ secrets.GITHUB_TOKEN }}
         build-args: |
           base_image=ghcr.io/intel/llvm/ubuntu2404_intel_drivers
-          base_tag=latest
+          base_tag=alldeps
         tags: |
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:${{ github.sha }}
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:latest
-    - name: Build and Push Container (no drivers)
-      uses: ./devops/actions/build_container
-      with:
-        push: ${{ github.ref_name == 'sycl' }}
-        file: ubuntu2204_preinstalled
-        username: ${{ github.repository_owner }}
-        password: ${{ secrets.GITHUB_TOKEN }}
-        build-args: |
-          base_image=ghcr.io/intel/llvm/ubuntu2204_base
-          base_tag=latest
-        tags: |
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:no-drivers-${{ github.sha }}
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:no-drivers
-    - name: Build and Push Container (Build image)
-      uses: ./devops/actions/build_container
-      with:
-        push: ${{ github.ref_name == 'sycl' }}
-        file: ubuntu2204_preinstalled
-        username: ${{ github.repository_owner }}
-        password: ${{ secrets.GITHUB_TOKEN }}
-        build-args: |
-          base_image=ghcr.io/intel/llvm/ubuntu2204_build
-          base_tag=latest
-        tags: |
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:build-${{ github.sha }}
-          ghcr.io/${{ github.repository }}/sycl_ubuntu2204_nightly:build
+          ghcr.io/${{ github.repository }}/sycl_ubuntu2404_nightly:${{ github.sha }}
+          ghcr.io/${{ github.repository }}/sycl_ubuntu2404_nightly:latest

devops/containers/ubuntu2204_preinstalled.Dockerfile renamed to devops/containers/nightly.Dockerfile

@@ -1,5 +1,5 @@
-ARG base_tag=latest
-ARG base_image=ghcr.io/intel/llvm/ubuntu2204_intel_drivers
+ARG base_tag=alldeps
+ARG base_image=ghcr.io/intel/llvm/ubuntu2404_intel_drivers
 
 FROM $base_image:$base_tag
 

devops/cts_exclude_filter_compfails

@@ -0,0 +1,2 @@
+# Please use "#" to add comments here.
+# Do not delete the file even if it's empty.

libdevice/nativecpu_utils.cpp

@@ -31,16 +31,7 @@ using __nativecpu_state = native_cpu::state;
 
 #define OCL_LOCAL __attribute__((opencl_local))
 #define OCL_GLOBAL __attribute__((opencl_global))
-
-DEVICE_EXTERNAL OCL_LOCAL void *
-__spirv_GenericCastToPtrExplicit_ToLocal(void *p, int) {
-  return (OCL_LOCAL void *)p;
-}
-
-DEVICE_EXTERNAL OCL_GLOBAL void *
-__spirv_GenericCastToPtrExplicit_ToGlobal(void *p, int) {
-  return (OCL_GLOBAL void *)p;
-}
+#define OCL_PRIVATE __attribute__((opencl_private))
 
 DEVICE_EXTERN_C void __mux_work_group_barrier(uint32_t id, uint32_t scope,
                                               uint32_t semantics);
@@ -61,6 +52,23 @@ __spirv_MemoryBarrier(uint32_t Memory, uint32_t Semantics) {
 // Turning clang format off here because it reorders macro invocations
 // making the following code very difficult to read.
 // clang-format off
+
+#define DefGenericCastToPtrExplImpl(sfx, asp, cv)\
+DEVICE_EXTERNAL cv asp void *\
+__spirv_GenericCastToPtrExplicit_##sfx(cv void *p ,int) {\
+  return (cv asp void *)p;\
+}
+
+#define DefGenericCastToPtrExpl(sfx, asp)\
+  DefGenericCastToPtrExplImpl(sfx, asp, )\
+  DefGenericCastToPtrExplImpl(sfx, asp, const)\
+  DefGenericCastToPtrExplImpl(sfx, asp, volatile)\
+  DefGenericCastToPtrExplImpl(sfx, asp, const volatile)
+
+DefGenericCastToPtrExpl(ToPrivate, OCL_PRIVATE)
+DefGenericCastToPtrExpl(ToLocal, OCL_LOCAL)
+DefGenericCastToPtrExpl(ToGlobal, OCL_GLOBAL)
+
 #define DefSubgroupBlockINTEL1(Type, PType)                                    \
   template <>                                                                  \
   __SYCL_CONVERGENT__ DEVICE_EXTERNAL Type                                     \

llvm/lib/SYCLLowerIR/SYCLJointMatrixTransform.cpp

@@ -21,6 +21,7 @@ namespace {
 
 static constexpr char ACCESS_CHAIN[] = "_Z19__spirv_AccessChain";
 static constexpr char MATRIX_TYPE[] = "spirv.CooperativeMatrixKHR";
+static constexpr char MATRIX_LAYOUT[] = "joint_matrix_layout_to_spv";
 
 Type *getInnermostType(Type *Ty) {
   while (auto *ArrayTy = dyn_cast<ArrayType>(Ty))
@@ -184,17 +185,99 @@ bool transformAccessChain(Function *F) {
   }
   return ModuleChanged;
 }
+
+StoreInst *findLastStoreBeforeLoad(Value *Ptr, Instruction *Load) {
+  BasicBlock::iterator It(Load);
+  while (It != Load->getParent()->begin()) {
+    --It;
+    if (auto *Store = dyn_cast<StoreInst>(&*It))
+      if (Store->getPointerOperand() == Ptr)
+        return Store;
+  }
+  return nullptr;
+}
+
+// Per SPIR-V specification Layout of a matrix must be a constant instruction
+// aka a constexpr or specialization constant. Meanwhile in SYCL headers
+// layout is passed as a parameter to joint_matrix_load function, so even if
+// that layout is a constant expression in the user's code - it's not possible
+// to prove that to the compiler, so constant propagation will happen only
+// after inlining, not in AST. That means, that with O0 layout would remain
+// to be a runtime variable in LLVM IR.
+// SYCL matrix layout is being mapped on SPIR-V matrix layout by
+// joint_matrix_layout_to_spv function. The following routine finds calls to
+// this function and replaces them with the found constant.
+// This function also cleans up code, that becomes dead. Pattern of the dead
+// code is stable, as user's code doesn't affect it.
+bool propagateConstexprLayout(Function *F) {
+  llvm::SmallVector<Instruction *, 8> ToErase;
+  for (auto I = F->user_begin(), E = F->user_end(); I != E;) {
+    User *U = *I++;
+    auto *CI = dyn_cast<CallInst>(U);
+    if (!CI)
+      continue;
+    auto *Op = dyn_cast<Instruction>(CI->getArgOperand(0));
+    if (!Op || !isa<LoadInst>(Op))
+      continue;
+    auto *Ptr = dyn_cast<Instruction>(cast<LoadInst>(Op)->getPointerOperand());
+    if (!Ptr)
+      continue;
+
+    ConstantInt *ConstLayout = nullptr;
+    StoreInst *SI = findLastStoreBeforeLoad(Ptr, Op);
+    if (!SI)
+      continue;
+    ConstLayout = dyn_cast<ConstantInt>(SI->getValueOperand());
+    if (ConstLayout) {
+      CI->replaceAllUsesWith(ConstLayout);
+      ToErase.push_back(CI);
+      ToErase.push_back(SI);
+      ToErase.push_back(Op);
+      ToErase.push_back(Ptr);
+      if (auto *Cast = dyn_cast<AddrSpaceCastInst>(Ptr)) {
+        auto *OrigPtr = Cast->getPointerOperand();
+        if (auto *AI = dyn_cast<AllocaInst>(OrigPtr))
+          ToErase.push_back(AI);
+      }
+    }
+  }
+
+  // There are possible cases, when a single instruction result is used multiple
+  // times. For this case we have to use a vector to store such instructions
+  // and keep track if we have removed them before to avoid double free().
+  SmallPtrSet<Instruction *, 8> Erased;
+  for (Instruction *II : ToErase) {
+    if (!II->use_empty())
+      continue;
+    if (Erased.contains(II))
+      continue;
+    II->dropAllReferences();
+    II->eraseFromParent();
+    Erased.insert(II);
+  }
+  return !ToErase.empty();
+}
 } // namespace
 
 PreservedAnalyses
 SYCLJointMatrixTransformPass::run(Module &M, ModuleAnalysisManager &MAM) {
   bool ModuleChanged = false;
+  llvm::SmallVector<Function *, 1> ToErase;
   for (Function &F : M) {
-    if (!F.isDeclaration())
-      continue;
+    if (!F.isDeclaration()) {
+      if (F.getName() == MATRIX_LAYOUT) {
+        ModuleChanged |= propagateConstexprLayout(&F);
+        ToErase.push_back(&F);
+      } else
+        continue;
+    }
     if (F.getName().starts_with(ACCESS_CHAIN))
       ModuleChanged |= transformAccessChain(&F);
   }
 
+  for (auto *F : ToErase)
+    if (F->users().empty())
+      F->eraseFromParent();
+
   return ModuleChanged ? PreservedAnalyses::none() : PreservedAnalyses::all();
 }

llvm/test/SYCLLowerIR/JointMatrixTransform/constexpr-layout.ll

@@ -0,0 +1,90 @@
+; The test checks, that users of the call to joint_matrix_layout_to_spv matrix
+; are replaced with the layout constant.
+
+; RUN: opt -passes=sycl-joint-matrix-transform < %s -S | FileCheck %s
+
+; ModuleID = 'test.bc'
+source_filename = "test.cpp"
+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-G1"
+target triple = "spir64-unknown-unknown"
+
+$joint_matrix_layout_to_spv = comdat any
+
+; CHECK: define weak_odr dso_local spir_kernel void @test
+; CHECK-NEXT: entry:
+; CHECK-NEXT: %{{.*}} = call spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHR{{.*}}(ptr addrspace(1){{.*}}, i32 noundef 0, i64 noundef{{.*}}
+; CHECK-NEXT: %{{.*}} = call spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHR{{.*}}(ptr addrspace(1){{.*}}, i32 noundef 1, i64 noundef{{.*}}
+; CHECK-NEXT: %{{.*}} = call spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHR{{.*}}(ptr addrspace(1){{.*}}, i32 noundef 1, i64 noundef{{.*}}
+; CHECK-NEXT: %{{.*}} = call spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHR{{.*}}(ptr addrspace(1){{.*}}, i32 noundef 2, i64 noundef{{.*}}
+; CHECK-NEXT: ret void
+
+; CHECK-NOT: joint_matrix_layout_to_spv
+
+define weak_odr dso_local spir_kernel void @test(ptr addrspace(1) %matrix.1, ptr addrspace(1) %matrix.2, i64 noundef %stride) {
+entry:
+  %layout.1 = alloca i32, align 4
+  %layout.2 = alloca i32, align 4
+  %layout.ascast.1 = addrspacecast ptr %layout.1 to ptr addrspace(4)
+  %layout.ascast.2 = addrspacecast ptr %layout.2 to ptr addrspace(4)
+  store i32 0, ptr addrspace(4) %layout.ascast.1, align 4
+  store i32 1, ptr addrspace(4) %layout.ascast.2, align 4
+
+  %layout.val.1 = load i32, ptr addrspace(4) %layout.ascast.1, align 4
+  %layout.spv.1 = call spir_func noundef i32 @joint_matrix_layout_to_spv(i32 noundef %layout.val.1)
+  %mload.1 = call spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ffLm16ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1) noundef %matrix.1, i32 noundef %layout.spv.1, i64 noundef %stride, i32 noundef 0)
+
+  %layout.val.2 = load i32, ptr addrspace(4) %layout.ascast.2, align 4
+  %layout.spv.2 = call spir_func noundef i32 @joint_matrix_layout_to_spv(i32 noundef %layout.val.2)
+  %mload.2 = call spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ffLm16ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1) noundef %matrix.2, i32 noundef %layout.spv.2, i64 noundef %stride, i32 noundef 0)
+
+  %layout.spv.3 = call spir_func noundef i32 @joint_matrix_layout_to_spv(i32 noundef %layout.val.2)
+  %mload.3 = call spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ffLm16ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1) noundef %matrix.2, i32 noundef %layout.spv.3, i64 noundef %stride, i32 noundef 0)
+
+  store i32 2, ptr addrspace(4) %layout.ascast.2, align 4
+  %layout.val.4 = load i32, ptr addrspace(4) %layout.ascast.2, align 4
+  %layout.spv.4 = call spir_func noundef i32 @joint_matrix_layout_to_spv(i32 noundef %layout.val.4)
+  %mload.4 = call spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ffLm16ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1) noundef %matrix.2, i32 noundef %layout.spv.4, i64 noundef %stride, i32 noundef 0)
+  ret void
+}
+
+declare dso_local spir_func noundef target("spirv.CooperativeMatrixKHR", float, 3, 16, 16, 2) @_Z32__spirv_CooperativeMatrixLoadKHRIU3AS1ffLm16ELm16ELN5__spv9MatrixUseE2ELNS1_12MatrixLayoutE3ELNS1_5Scope4FlagE3EEPNS1_28__spirv_CooperativeMatrixKHRIT0_XT5_EXT1_EXT2_EXT3_EEEPT_S3_mi(ptr addrspace(1) noundef, i32 noundef, i64 noundef, i32 noundef)
+
+define linkonce_odr dso_local spir_func noundef i32 @joint_matrix_layout_to_spv(i32 noundef %Layout) comdat {
+entry:
+  %retval = alloca i32, align 4
+  %Layout.addr = alloca i32, align 4
+  %retval.ascast = addrspacecast ptr %retval to ptr addrspace(4)
+  %Layout.addr.ascast = addrspacecast ptr %Layout.addr to ptr addrspace(4)
+  store i32 %Layout, ptr addrspace(4) %Layout.addr.ascast, align 4
+  %0 = load i32, ptr addrspace(4) %Layout.addr.ascast, align 4
+  switch i32 %0, label %sw.epilog [
+    i32 0, label %sw.bb
+    i32 1, label %sw.bb1
+    i32 2, label %sw.bb2
+    i32 3, label %sw.bb3
+  ]
+
+sw.bb:                                            ; preds = %entry
+  store i32 0, ptr addrspace(4) %retval.ascast, align 4
+  br label %return
+
+sw.bb1:                                           ; preds = %entry
+  store i32 1, ptr addrspace(4) %retval.ascast, align 4
+  br label %return
+
+sw.bb2:                                           ; preds = %entry
+  store i32 2, ptr addrspace(4) %retval.ascast, align 4
+  br label %return
+
+sw.bb3:                                           ; preds = %entry
+  store i32 3, ptr addrspace(4) %retval.ascast, align 4
+  br label %return
+
+sw.epilog:                                        ; preds = %entry
+  call void @llvm.trap()
+  unreachable
+
+return:                                           ; preds = %sw.bb3, %sw.bb2, %sw.bb1, %sw.bb
+  %1 = load i32, ptr addrspace(4) %retval.ascast, align 4
+  ret i32 %1
+}

sycl/cmake/modules/UnifiedRuntimeTag.cmake

@@ -1,7 +1,7 @@
-# commit 222e4b1d51536bb38e03e2000a79679af0a44a6d
-# Merge: 30d183a0 28108a7e
-# Author: Kenneth Benzie (Benie) <k.benzie@codeplay.com>
-# Date:   Fri Jan 17 10:28:34 2025 +0000
-#     Merge pull request #2561 from Bensuo/ben/cmd-buffer-l0-fence
-#     [L0][CMDBUF] Optimize fence/event waits during update
-set(UNIFIED_RUNTIME_TAG 222e4b1d51536bb38e03e2000a79679af0a44a6d)
+# commit 8b7a99578966eb691a961d9620ea38d235196b2f
+# Merge: ed095412 7b0e3b19
+# Author: Martin Grant <martin.morrisongrant@codeplay.com>
+# Date:   Mon Jan 20 09:27:22 2025 +0000
+#     Merge pull request #2582 from przemektmalon/przemek/intel-host-usm-support
+#     Enable Host USM backed images on Level Zero
+set(UNIFIED_RUNTIME_TAG 8b7a99578966eb691a961d9620ea38d235196b2f)

sycl/doc/developer/DockerBKMs.md

@@ -53,10 +53,6 @@ development containers:
    NVidia/AMD and can be used for building DPC++
    compiler from source with all backends enabled or for end-to-end testing
    with HIP/CUDA on machines with corresponding GPUs available.
-  - `devops/containers/sycl_ubuntu2204_nightly`: contains the latest successfully
-   built nightly build of DPC++ compiler. The Dockerfile comes in three flavors:
-   with pre-installed Intel drivers (`latest`), without them (`no-drivers`) and
-   with development kits installed (`build`).
 
 ### Ubuntu 24.04-based Dockerfiles
 
@@ -78,7 +74,8 @@ development containers:
    NVidia/AMD and can be used for building DPC++
    compiler from source with all backends enabled or for end-to-end testing
    with HIP/CUDA on machines with corresponding GPUs available.  
-
+ - `devops/containers/nightly`: contains the latest successfully
+   built nightly build of DPC++ compiler.
 
 ## Running Docker container interactively
 
@@ -199,7 +196,7 @@ Docker containers can be built with the following command:
 docker build -f path/to/devops/containers/file.Dockerfile path/to/devops/
 ```
 
-The `ubuntu2204_preinstalled.Dockerfile` script expects `llvm_sycl.tar.xz` file
+The `nightly.Dockerfile` script expects `llvm_sycl.tar.xz` file
 to be present in `devops/` directory.
 
 Containers other than base provide several configurable arguments, the most