Merge from 'sycl' to 'sycl-web'

iclsrc · iclsrc · commit a887b7055a30 · 2022-04-27T07:45:07.000-07:00
diff --git a/clang/lib/CodeGen/CGExprScalar.cpp b/clang/lib/CodeGen/CGExprScalar.cpp
@@ -2059,10 +2059,9 @@ Value *ScalarExprEmitter::VisitCastExpr(CastExpr *CE) {
     llvm::Type *SrcTy = Src->getType();
     llvm::Type *DstTy = ConvertType(DestTy);
     if (SrcTy->isPtrOrPtrVectorTy() && DstTy->isPtrOrPtrVectorTy() &&
-        SrcTy->getPointerAddressSpace() != DstTy->getPointerAddressSpace()) {
-      llvm_unreachable("wrong cast for pointers in different address spaces"
-                       "(must be an address space cast)!");
-    }
+        SrcTy->getPointerAddressSpace() != DstTy->getPointerAddressSpace())
+      Src = Builder.CreateAddrSpaceCast(
+          Src, llvm::PointerType::get(SrcTy, DstTy->getPointerAddressSpace()));
 
     if (CGF.SanOpts.has(SanitizerKind::CFIUnrelatedCast)) {
       if (auto *PT = DestTy->getAs<PointerType>()) {
diff --git a/clang/test/CodeGenSYCL/address-space-builtin-alloca.cpp b/clang/test/CodeGenSYCL/address-space-builtin-alloca.cpp
@@ -0,0 +1,13 @@
+// RUN: %clang_cc1 -triple spir64-unknown-linux -fsycl-is-device -disable-llvm-passes -emit-llvm -x c++ %s -o - | FileCheck %s
+
+// Test to verify that address space cast is generated correctly for __builtin_alloca
+
+__attribute__((sycl_device)) void foo() {
+  // CHECK: %TestVar = alloca i32 addrspace(4)*, align 8
+  // CHECK: %TestVar.ascast = addrspacecast i32 addrspace(4)** %TestVar to i32 addrspace(4)* addrspace(4)*
+  // CHECK: %[[ALLOCA:[0-9]+]] = alloca i8, i64 1, align 8
+  // CHECK: %[[ADDRSPCAST:[0-9]+]] = addrspacecast i8* %[[ALLOCA]] to i8* addrspace(4)*
+  // CHECK: %[[BITCAST:[0-9]+]] = bitcast i8* addrspace(4)* %[[ADDRSPCAST]] to i32 addrspace(4)*
+  // CHECK: store i32 addrspace(4)* %[[BITCAST]], i32 addrspace(4)* addrspace(4)* %TestVar.ascast, align 8
+  int *TestVar = (int *)__builtin_alloca(1);
+}
diff --git a/libdevice/cmake/modules/SYCLLibdevice.cmake b/libdevice/cmake/modules/SYCLLibdevice.cmake
diff --git a/libdevice/cmath_wrapper_fp64.cpp b/libdevice/cmath_wrapper_fp64.cpp
@@ -1,4 +1,5 @@
-//==--- cmath_wrapper.cpp - wrappers for C math library functions ----------==//
+//==--- cmath_wrapper_fp64.cpp - wrappers for double precision C math library
+// functions ----------==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
diff --git a/libdevice/complex_wrapper_fp64.cpp b/libdevice/complex_wrapper_fp64.cpp
@@ -1,4 +1,5 @@
-//==--- complex_wrapper.cpp - wrappers for C99 complex math functions ------==//
+//==--- complex_wrapper_fp64.cpp - wrappers for double precision C99 complex
+// math functions ------==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
diff --git a/libdevice/fallback-cmath-fp64.cpp b/libdevice/fallback-cmath-fp64.cpp
@@ -1,4 +1,5 @@
-//==--- fallback-cmath.cpp - fallback implementation of math functions -----==//
+//==--- fallback-cmath-fp64.cpp - fallback implementation of double precision
+// math functions -----==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
diff --git a/libdevice/fallback-complex-fp64.cpp b/libdevice/fallback-complex-fp64.cpp
@@ -1,4 +1,5 @@
-//==----- fallback-complex.cpp - complex math functions for SPIR-V device --==//
+//==----- fallback-complex-fp64.cpp - double precision complex math functions
+// for SPIR-V device --==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
@@ -7,7 +8,6 @@
 //===----------------------------------------------------------------------===//
 
 #include "device_complex.h"
-#include "device_math.h"
 
 #ifdef __SPIR__
 #include <cmath>
diff --git a/libdevice/fallback-complex.cpp b/libdevice/fallback-complex.cpp
@@ -7,7 +7,6 @@
 //===----------------------------------------------------------------------===//
 
 #include "device_complex.h"
-#include "device_math.h"
 
 #ifdef __SPIR__
 #include <cmath>
diff --git a/sycl/plugins/level_zero/CMakeLists.txt b/sycl/plugins/level_zero/CMakeLists.txt
@@ -103,6 +103,8 @@ if (SYCL_ENABLE_XPTI_TRACING)
   set(XPTI_PROXY_SRC "${CMAKE_SOURCE_DIR}/../xpti/src/xpti_proxy.cpp")
 endif()
 
+find_package(Threads REQUIRED)
+
 add_sycl_plugin(level_zero
   SOURCES
     "${sycl_inc_dir}/CL/sycl/detail/pi.h"
diff --git a/sycl/plugins/level_zero/pi_level_zero.cpp b/sycl/plugins/level_zero/pi_level_zero.cpp
@@ -4636,18 +4636,31 @@ pi_result piKernelCreate(pi_program Program, const char *KernelName,
     return PI_ERROR_UNKNOWN;
   }
 
-  // Update the refcount of the program and context to show it's used by this
-  // kernel.
+  PI_CALL((*RetKernel)->initialize());
+  return PI_SUCCESS;
+}
+
+pi_result _pi_kernel::initialize() {
+  // Retain the program and context to show it's used by this kernel.
   PI_CALL(piProgramRetain(Program));
   if (IndirectAccessTrackingEnabled)
     // TODO: do piContextRetain without the guard
     PI_CALL(piContextRetain(Program->Context));
 
   // Set up how to obtain kernel properties when needed.
-  (*RetKernel)->ZeKernelProperties.Compute =
-      [ZeKernel](ze_kernel_properties_t &Properties) {
-        ZE_CALL_NOCHECK(zeKernelGetProperties, (ZeKernel, &Properties));
-      };
+  ZeKernelProperties.Compute = [this](ze_kernel_properties_t &Properties) {
+    ZE_CALL_NOCHECK(zeKernelGetProperties, (ZeKernel, &Properties));
+  };
+
+  // Cache kernel name.
+  ZeKernelName.Compute = [this](std::string &Name) {
+    size_t Size = 0;
+    ZE_CALL_NOCHECK(zeKernelGetName, (ZeKernel, &Size, nullptr));
+    char *KernelName = new char[Size];
+    ZE_CALL_NOCHECK(zeKernelGetName, (ZeKernel, &Size, KernelName));
+    Name = KernelName;
+    delete[] KernelName;
+  };
 
   return PI_SUCCESS;
 }
@@ -4727,13 +4740,8 @@ pi_result piKernelGetInfo(pi_kernel Kernel, pi_kernel_info ParamName,
     return ReturnValue(pi_program{Kernel->Program});
   case PI_KERNEL_INFO_FUNCTION_NAME:
     try {
-      size_t Size = 0;
-      ZE_CALL(zeKernelGetName, (Kernel->ZeKernel, &Size, nullptr));
-      char *KernelName = new char[Size];
-      ZE_CALL(zeKernelGetName, (Kernel->ZeKernel, &Size, KernelName));
-      pi_result Res = ReturnValue(static_cast<const char *>(KernelName));
-      delete[] KernelName;
-      return Res;
+      std::string &KernelName = *Kernel->ZeKernelName.operator->();
+      return ReturnValue(static_cast<const char *>(KernelName.c_str()));
     } catch (const std::bad_alloc &) {
       return PI_OUT_OF_HOST_MEMORY;
     } catch (...) {
@@ -5086,20 +5094,7 @@ pi_result piextKernelCreateWithNativeHandle(pi_native_handle NativeHandle,
 
   auto ZeKernel = pi_cast<ze_kernel_handle_t>(NativeHandle);
   *Kernel = new _pi_kernel(ZeKernel, OwnNativeHandle, Program);
-
-  // Update the refcount of the program and context to show it's used by this
-  // kernel.
-  PI_CALL(piProgramRetain(Program));
-  if (IndirectAccessTrackingEnabled)
-    // TODO: do piContextRetain without the guard
-    PI_CALL(piContextRetain(Program->Context));
-
-  // Set up how to obtain kernel properties when needed.
-  (*Kernel)->ZeKernelProperties.Compute =
-      [ZeKernel](ze_kernel_properties_t &Properties) {
-        ZE_CALL_NOCHECK(zeKernelGetProperties, (ZeKernel, &Properties));
-      };
-
+  PI_CALL((*Kernel)->initialize());
   return PI_SUCCESS;
 }
 
diff --git a/sycl/plugins/level_zero/pi_level_zero.hpp b/sycl/plugins/level_zero/pi_level_zero.hpp
@@ -185,7 +185,7 @@ template <class T> struct ZeCache : private T {
   // order to disallow access other than through "->".
   //
   typedef std::function<void(T &)> InitFunctionType;
-  InitFunctionType Compute;
+  InitFunctionType Compute{nullptr};
   bool Computed{false};
 
   ZeCache() : T{} {}
@@ -1379,6 +1379,9 @@ struct _pi_kernel : _pi_object {
       : ZeKernel{Kernel}, OwnZeKernel{OwnZeKernel}, Program{Program},
         MemAllocs{}, SubmissionsCount{0} {}
 
+  // Completed initialization of PI kernel. Must be called after construction.
+  pi_result initialize();
+
   // Returns true if kernel has indirect access, false otherwise.
   bool hasIndirectAccess() {
     // Currently indirect access flag is set for all kernels and there is no API
@@ -1445,6 +1448,7 @@ struct _pi_kernel : _pi_object {
 
   // Cache of the kernel properties.
   ZeCache<ZeStruct<ze_kernel_properties_t>> ZeKernelProperties;
+  ZeCache<std::string> ZeKernelName;
 };
 
 struct _pi_sampler : _pi_object {
diff --git a/sycl/source/detail/program_manager/program_manager.cpp b/sycl/source/detail/program_manager/program_manager.cpp
@@ -913,9 +913,6 @@ ProgramManager::getDeviceImage(OSModuleHandle M, KernelSetId KSId,
     std::cerr << "selected device image: " << &Img->getRawData() << "\n";
     Img->print();
   }
-
-  if (std::getenv("SYCL_DUMP_IMAGES") && !m_UseSpvFile)
-    dumpImage(*Img, KSId);
   return *Img;
 }
 
@@ -1094,6 +1091,7 @@ bool ProgramManager::kernelUsesAssert(OSModuleHandle M,
 
 void ProgramManager::addImages(pi_device_binaries DeviceBinary) {
   std::lock_guard<std::mutex> Guard(Sync::getGlobalLock());
+  const bool DumpImages = std::getenv("SYCL_DUMP_IMAGES") && !m_UseSpvFile;
 
   for (int I = 0; I < DeviceBinary->NumDeviceBinaries; I++) {
     pi_device_binary RawImg = &(DeviceBinary->DeviceBinaries[I]);
@@ -1230,10 +1228,12 @@ void ProgramManager::addImages(pi_device_binaries DeviceBinary) {
         }
       }
       m_DeviceImages[KSId].reset(new std::vector<RTDeviceBinaryImageUPtr>());
-
       cacheKernelUsesAssertInfo(M, *Img);
 
+      if (DumpImages)
+        dumpImage(*Img, KSId);
       m_DeviceImages[KSId]->push_back(std::move(Img));
+
       continue;
     }
     // Otherwise assume that the image contains all kernels associated with the
@@ -1248,6 +1248,8 @@ void ProgramManager::addImages(pi_device_binaries DeviceBinary) {
 
     cacheKernelUsesAssertInfo(M, *Img);
 
+    if (DumpImages)
+      dumpImage(*Img, KSId);
     Imgs->push_back(std::move(Img));
   }
 }
diff --git a/sycl/tools/sycl-trace/CMakeLists.txt b/sycl/tools/sycl-trace/CMakeLists.txt
@@ -31,18 +31,12 @@ add_custom_target(pi-pretty-printers
 # To get L0 loader
 add_dependencies(sycl_pi_trace_collector pi_level_zero)
 
-if (NOT DEFINED LEVEL_ZERO_INCLUDE_DIR)
-  set(LEVEL_ZERO_INCLUDE_DIR
-    "${CMAKE_CURRENT_BINARY_DIR}/../../plugins/level_zero/level_zero_loader/include")
-endif()
-target_include_directories(sycl_pi_trace_collector PRIVATE
-  ${LEVEL_ZERO_INCLUDE_DIR}
-)
+target_link_libraries(sycl_pi_trace_collector PRIVATE LevelZeroLoader::Headers)
 
 add_custom_target(ze-pretty-printers
   COMMAND ${Python3_EXECUTABLE}
   ${CMAKE_CURRENT_SOURCE_DIR}/generate_ze_pretty_printers.py
-  ${LEVEL_ZERO_INCLUDE_DIR}/level_zero/ze_api.h
+  ${SYCL_INCLUDE_BUILD_DIR}/sycl/level_zero/ze_api.h
   DEPENDS pi_level_zero
   BYPRODUCTS
     ${CMAKE_CURRENT_BINARY_DIR}/ze_printers.def
